Patent Application
20180046149
The present embodiments relate generally to building automation systems.
Building automation systems include heating, ventilation and air conditioning (HVAC) systems, security systems, fire systems, or other systems. The systems are typically formed from distributed components wired together. HVAC systems may be formed with one, two, or three separate tiers or architectural levels. In a three-tier system, a floor level network provides general control for a particular floor or zone of a building. Controllers of the floor level network provide process controls based on sensor inputs to operate actuators. For example, an adjustment of a damper, heating element, cooling element, or other actuator is determined based on a set point and a measured temperature. Other control functions may be provided. The building level network integrates multiple floor level networks to provide consistent control between various zones within a building. Panels or other controllers control distribution systems, such as pumps, fans or other central plants for cooling and heating. Building level controllers may communicate among themselves and access floor level controllers for obtaining data. The management level network integrates control of the building level networks to provide a high-level control process of the overall building environment and equipment.
Each building is run separately. Data from the different levels is used to identify faults or diagnose problems for a given building. This data for a given building may not accurately reflect a problem or the influence of the building automation on a business.
Degraded or other performance may be predicted with a machine-learnt classifier. Based on operation of many building automation systems, machine learning is applied. The machine learning creates a predictor. The machine-learnt predictor is applied to the operation data of any building automation system to predict future failure or other event, providing prognostics that may be used to plan maintenance and/or schedule remedial action. Machine learning uses big data in the form of data from many building automation systems to learn to automatically predict and/or perform prognostics for other building automations systems.
In one further embodiment, clustering and/or other machine learning may be used to (1) identify the building automation systems in need of prognostics and/or a fault and/or (2) extract information used for predicting. The output of the clustering and/or other machine learning is used as an input to the prediction.
In one aspect, a method of building automation predication is provided in a building management system. First data related to a plurality of buildings is accessed by a building analytics system of the building management system. The first data includes building management system data from different times and enterprise data different than building management system data. The enterprise data is for an enterprise associated with the buildings of the plurality and the building management system data being for the buildings of the plurality. The building analytics system applies the first data from a first sub-set of the different times to machine learning with the first data from a second sub-set of the different times as a ground truth for learning prediction of the building management system. Because of the applying, a machine-learnt predictor of operation of the building automation systems generates an output.
In a second aspect, a building management system is provided for building automation prediction. Building automation systems for heating, ventilation, and air conditioning for multiple buildings are configured to output operational input and output data. A building processor is configured to predict future degradation of a part of at least one of the building automation systems based on the input and output data being applied to a machine-learnt predictor trained from times series data from multiple examples. A display is configured to output the prediction of the future degradation.
In a third aspect, a method is provided for building automation prognostics in a building management system. A building analytics system of the building management system accesses first time-series data related to operation a first building automation system over time. A first machine-learnt classifier predicts failure of a part of the first building automation system based on the first time-series data. The first machine-learnt classifier was trained based on second time-series data related to operation of other building automation systems over time. Results of the predicting are presented on a display of the building analytics system. The results include the failure and the part.
In a fourth aspect, a method is provided or building automation prediction in a building management system. A building or distribution in the building is identified as having undesired performance based on unsupervised clustering data from a plurality of buildings including the building. A machine-learnt cerebellar model articulation controller identifies a fault within a building system of the building. A recurrent neural network is applied to time series data for the fault. The application predicts occurrence of the fault within the building system.
Other systems, methods, and/or features of the present embodiments will become apparent to one with skill in the art upon examination of the following FIGS. and detailed description. It is intended that all such additional systems, methods, features, and advantages be included within this description, be within the scope of the invention, and be protected by the accompanying claims. Additional features of the disclosed embodiments are described in, and will be apparent from, the following detailed description and the FIGS.
The components in the FIGS. are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the embodiments. In the FIGS., like reference numerals designate corresponding parts throughout the different views.
Analytics are used in building automation. Embodiments disclosed herein provide improvements for building automation systems employing analytics. Analytics is the systematic use of physical data and related business insights developed through applied analytical disciplines (e.g. statistical, contextual, quantitative, predictive, cognitive, or other emerging models) to drive fact-based decision making for planning, management, measurement, and learning. Analytics may be descriptive, predictive, or prescriptive. For a non-building automation system example, a system may use Twitter® data to accurately predict rates of heart disease by region (e.g., county). The data analytics from twitter information mirrors heart disease rates from death certificates.
For building automation, performance analytics are applied within a single building and all its systems. Big data analytics are applied across multiple buildings that belong to and/or are controlled by a given enterprise. For example, the operation of tens or hundreds of branches, franchises, or facilities are analyzed. The analytics are designed to detect operational performance deficiency, such as fault detection and diagnostics—proactively and adaptively. At an enterprise level, performance analytics of a particular building automation system and big data analytics of the enterprise in which the building automation systems are employed are combined by a management system operable to analyze and perform processes based on the combination as further described herein.
Customers or others using data analytics in building automation may benefit. A management system employing the analytics processes and structures as described herein may be used to provide quality safe, comfortable, and productive environment. Service interruptions, breakdowns, and turndown time may be reduced. Cost of ownership may be reduced, and equipment and/or systems service life may be increased. By analyzing at an enterprise level in accordance with processes employed in the embodiments of management systems disclosed herein, the focus for an enterprise or company shifts from repair and maintenance to prevention and prediction. Overall operating expense, operating expenditure, operational expense, operational expenditure (OPEX) may be reduced, allowing better utilization of capital expenditure (CAPEX).
Using machine learning, data from multiple buildings and/or meta data from the enterprise are used to create a predictor of degradation of building automation. For example, a recurrent neural network uses time series of operational data from many buildings to learn to predict degradation or other event at a future time. The learnt predictor is applied by the building analytics system to data of a building automation system to predict the future degradation of a part of the automation system. Faults are predicted to offer a prognostics of the building automation system. The prediction identifies faults ahead of time. Prognostics takes corrective actions to repair faults ahead of time. The prediction or prognostics are used to prevent downtime and associated costs, such as through scheduling maintenance before the predicted fault occurs, allowing better utilization of resources.
The machine learning for prediction is driven by data, so is suitable for big data or enterprise level building management. Rather than applying a programmed algorithm that may not be suitable for some situations, a more robust machine-trained predicator is used by the management system. The machine training is adaptable and able to automatically predict despite the volume and complexity of data.
In a further embodiment, different machine-learnt classifiers are provided by the building analytics system in a hybrid. The hybrid of different types of neural networks and/or clustering techniques may make prediction performance or failure and prognostics of failure in a building system hierarchically using supervised and unsupervised learning methods. The building automation systems with likely or possible problems and/or the problems are identified using clustering and/or a first machine-learnt classifier. For example, clustering or case based reasoning identifies commonalties among lower performance buildings. The machine-learnt predictor uses the output of the first machine-learnt classifier to predict whether or not degradation will occur and/or the time of degradation.
The management system 8 includes an enterprise 10 associated with any number of building automation systems 12 and/or a meta data database 14. A computer or building analytics system 17 with a building processor 16 and display 18 are part of the enterprise 10 or separate from the enterprise 10. Additional, different, or fewer components may be provided. For example, the building analytics system 17 may include a keyboard or mouse (not shown in the figures) that is operatively connected to the processor 16 via an interface 19 for receiving user inputs. The interface 19 may also include a network communications interface for enabling the processor 16 to communicate with building automation systems 12 and meta database 14.
The enterprise 10 is a company, organization, collective, government entity (e.g., city) or individual using an automated facility or building for business activities other than the automation of the facility. Building automation includes safety (e.g., fire alarm), environmental (e.g., HVAC), security, hazard, combinations thereof or other building systems. These automated building systems 12 provide a space for conducting business. The business is provided for other purposes than automating the building, such as sales of products or services. The enterprise 10 is in business for providing products or services, but operates in one or multiple buildings with automation. For example, a bank has hundreds of buildings for branches and/or headquarters. The enterprise 10 provides banking services, and the enterprise 10 is housed in buildings.
The enterprise 10 generates information or data. The data is business data, such as for the sales, service, human resources, information technology of operation of networks different than the building automation, accounting, budgets, or other business data. This business data is enterprise 10 level meta data different than data generated as part of or for operation of the building automation systems 12.
The enterprise or meta data is stored in a meta data database 14. The database 14 is one or more memories, such as hard drives, flash drives, tape drives, or other database. The database 14 is operated as one memory or multiple separate memories to store the various meta data.
Example meta data includes employee or student performance, such as test scores or review ratings. Other meta data may be budgets, employee attendance, staffing level, maintenance schedule or information, sales, elevator usage, or other data at the enterprise level. While the data is generated as part of the enterprise, the granularity of the data may be by regions, employee, or even building. The meta data may include the type of building, orientation of the building, location of the building, types of space usage in the building, or other building specific information not specifically used or output by the building automation system 12.
The building automation systems 12 includes safety (e.g., fire alarm), environmental (e.g., HVAC), security, hazard, combinations thereof, or other building systems. The automation is of a building, floor, room, or zone hosting part of the enterprise 10. In the example of
A given instance of a building automation system generates data, such as data from sensors, actuators, panels, or controllers. Sensors may include temperature, humidity, fire, smoke, occupancy, air quality, gas, CO2 or CO, or other now known or later developed sensors, such as an oxygen sensor for use in hospitals. Actuator may include a valve, relay, solenoid, speaker, bell, switch, motor, motor starter, damper, pneumatic device, combinations thereof, or other now known or later developed actuating devices for building automation. The controllers or panels interact with other building automation devices for establishing, setting, altering, instructing, reporting, or routing information for controlling building automation functions. The controller is a panel, processor, workstation, and/or server.
Control processes are run on controllers, sensors, and actuators as appropriate for the particular operations of each device. The sensor reports information appropriate or specific to the sensor, such as reporting the result of a comparison of a measured value to a desired or set point value. Actuators use the output sensor data to provide a response appropriate for the actuator. Controllers monitor the process or action of sensors and actuators without control in one mode of operation. In another mode of operation, the controllers override or influence the sensor and/or actuators to alter processing based on a regional or larger area control process. For example, a controller implements a coordination control application for overriding, setting, adjusting or altering the operation of another building automation application. Alternatively, the controllers run processes to measure deviation from a set point and control the response.
Other building automation devices may include personal computers, panels, or monitors. For example, one building automation device is an actuator for controlling a building wide component, such as a chiller, boiler, building intake vent, or building airflow out take vent. Using the building automation devices, major or building wide equipment, individual spaces, or local input and output points are controlled. The sensors, actuators, and/or control may be for zones, rooms, distribution, and/or plant operation.
The building automation system 12 implements building automation applications for controlling building functions. The building automation applications are programmed with programmable powerful processing control language (PPCL) or other language.
The building automation systems 12 are configured by software and/or hardware to collect, store, and output operational data 13a, 13b, and/or 13c in
The operational data 13 includes input and output data. Input data is any data used to control the operation of the building automation system 12, such as sensor values. Example input values include chilled water supply and return temperatures, discharge air temperature (DAT), supply air flow (SA Flow), return air flow (RA Flow), and outside air temperature (OAT). Output data is any data measuring performance of the building automation system 12, such as energy usage, temperature variation, error signals, outside air damper % open (OAD %)), and chilled water valve % open (CHV %).
The operational data 13 is provided for different times. A time series of data is provided by the building automation system 12. At different times, such as periodically (e.g., every second, minute, hour, or day), the operational data is logged, measured, or recorded. Two or more repetitions provide the time series of data. The time series may extend for any amount of time, such as over hours, days, weeks, or years. The beginning may be from a last reset. Alternatively, a moving window is used where the beginning is a given amount of time from the current time.
This building management data, such as the operational data, is specific to the building automation system 12, so is different than the meta data stored in the database 14. The database 14 may also store the building operation data, or the building operation data is stored in other memories. Time series data from any number of building automation systems 12 is provided.
The building automation systems 12 store the operational data for access in response to a query. Alternatively, the building automation systems 12 push data to the processor 16 of the building analytics system 17 or another device. The interface 19 of the building analytics system 17 accesses the meta data database 14 and/or the building automation systems 12 to pull or collect data. Alternatively, the data is periodically pushed to the interface 19 by the respective building automation system 12. The interface 19 is a port, communications interface, or other interface for networking.
The operational data 13 and/or enterprise data (such as meta data 14 for the enterprise 10) are communicated using wired or wireless communications. A local area, wide area, Internet, or other computer network may be used to communicate the operational data to the processor 16. For within the building automaton system 12, the same or different network is used, such as an 802.15.4 network, token network, or Mesh network. Bluetooth, Wi-Fi, computer network, Ethernet, proprietary, or other standard communication protocols may be used. 802.15.4 and 802.11x provide medium access control and a physical interface to a wireless medium. Any now known or later developed network and transport algorithms may be used.
Any packet size or data format may be used. Different bandwidths for any given communications path may be provided, such as adapting a lower level network for small data packets transmitted over short distances as compared to a higher-level network adapted for larger data packets at higher rates and for longer distances.
In typical building automation, building performance is based on observed data from sensors and operation data from actuators. The enterprise 10 also generates enterprise level data.
In one embodiment, the data analytics employed by the building analytics system 17 includes correlating multiple variables represented in the data 13 and 14 with one or more performance criteria also represented in the data. Other sources of performance may be used. Any clustering or case based analysis may be used. By including data 13 from multiple building automation systems 12 and/or enterprise data 14, this unsupervised learning by the building analytics system 17 may indicate useful information for diagnosis, prognostics, planning, or operation. Unsupervised learning employed by the building analytics system 17 determines the relationship of input variables or values of the variables to any user selected performance criterion or criteria without prior training of a classifier. The unsupervised learning employed by the building analytics system 17 indicates relationships based on data currently available without prior modeling or simulation. Additional or different machine learning may be used by the building analytics system 17 to identify building automation systems 12 and/or parts associated with poor performance, such as using cerebellar model articulation controller (CMAC) for classifying a fault.
The information output by the clustering or other classification of the building analytics system 17 is used with or without the operating data of many building automation systems 12 to machine train a predictor of degraded or other operation or make a prognostics. The clustering or other classification may output a time series. The operation data is in a time series. The predictor is trained by the building analytics system 17 using time series data. In an alternative embodiment, the predictor is trained without the clustering or other classification and/or without time series data.
Once trained, the predictor is used by the building analytics system 17 to predict for any given building automation system 12, such as one of the building automation systems 12 used for training or a different building automation system 12 not used for training. The operation data 13, classification (e.g., clustering) outputs, and/or other input feature of the predictor is input to the predictor of the building analytics system 17.
Referring to
The building processor 16 is a computer, server, panel, workstation, general processor, digital signal processor, application specific integrated circuit, field programmable gate array, analog circuit, digital circuit, combinations thereof, or other now known or later developed device for processing big data and determining the relationship of big data to building automation or vice versa. The building processor 16 is a device for performing the data analytics, such as the machine learning.
The processor 16 is part of the enterprise 10. In one embodiment, the data analytics is performed by a management computer of a building automation system 12. Alternatively, the processor 16 is separate from the enterprise 10 to provide the data analytics as described herein to the enterprise by the building analytics system 17 as depicted in
The processor 16 of the building analytics system 17 is configured to analyze the data, such as the building automation operation data 13 and/or the enterprise data 14. The data represents various variables. Values are provided for the variables. The values may be measures of the variable over time, by location (e.g., value for each building automation system), constant, or combinations thereof. For classification and clustering, big data is pre-selected by the user or default big data is used. For prediction, bid data is pre-selected by the user or default big data is used. Output from other classification may additionally be used for prediction by the processor 16. The big data represents variables of the operational and/or enterprise data. The big data is used by the building analytics system 17 for the machine learning.
The building processor 16 of the building analytics system 17 performs machine training to relate input features to predictions of degradation, other event, or forecast. For example,
The processor 16 of the building analytics system 17 applies machine learning to the data. The machine learning as employed by the building analytics system 17 enables the processor 16 to determine a statistical or other relationship between the data. The time series data is used for training as well as the ground truth. A sub-set of times is used as input features. Another sub-set associated with the performance criteria (e.g., part failure) of the time series is used by the building analytics system 17 as the ground truth. The machine learning employed by the building analytics system 17 relates the input operation data for time prior to the times of degraded performance to predict the degraded performance.
Any now known or later developed machine learning may be used by the building analytics system 17, such as neural network. For example, the processor 16 of the building analytics system 17 uses a recurrent neural network or other machine learning based on a time series to predict the future time series and/or an event.
The machine learning employed by the building analytics system 17 determines the relationship between one or a set of variables or values to another one or set of variables or values. In one embodiment, one or more variables are selected by the building analytics system 17 as performance criterion or criteria. For example, time series showing degraded performance are identified. The machine training learns to distinguish degraded performance from non-degraded performance using the time series data. Time series data prior to eventual degraded performance is distinguished from time series data prior to eventual non-degraded performance. The values of the variables corresponding to degraded performance are used as a measure of performance. The variable representing performance may be from the operational data or may be from the enterprise data. In one example, degraded performance may be reflected through increased enterprise cost. The enterprise cost over time is used as a measure of performance. The machine training employed by the building analytics system 17 learns to distinguish between time series operational data that leads to increased enterprise cost from time series operational data that does not lead to the increased enterprise cost. As another example, the operation data is used as the measure of performance, such as the percentage open of a valve or damper. In other examples, the machine training learns to predict future operation without specifically identifying degradation. A separate analysis identifies degradation or other event from the predicted sets 62 of operation data.
In one embodiment, a hybrid system of analytics is trained by the building analytics system 17. An additional classifier or classifiers are trained, as represented in
The output of the classifiers 70, 72, once trained, is processed and output by the building analytics system 17 to use in training the predictor 64. Similarly, the outputs are used for predicting from the trained predictor 64. Any synthesis of the outputs from the classifiers 70, 72 may be used. For example, the outputs from the classifiers 70, 72 are used directly as input features for the predictor training or application. As another example, one or more values are calculated by the building analytics system 17 using one or more outputs from each of the classifiers 70, 72. The calculated values are used as part of the input feature for the predictor.
In an example of use of the clustering classifier 70, unsupervised machine learning is used by the building analytics system 17. Variables from the big data are used to cluster relative to any measure of performance in order to determine which variables or values of variables distinguish between good and bad performance. Combinations of variables and the associated values may by employed by the building analytics system 17 to distinguish or correlate more strongly with the performance. The good and bad performances are relative terms based on the range of values for the performance measure. A default or user selected delineation between good and bad performance may be used. Alternatively, the clustering or other unsupervised learning employed by the building analytics system 17 applies a standard deviation or other analysis to distinguish between good and bad performance.
In one example, clustering is used by the building analytics system 17 to measure building performance. The operational data 13 of the building automation systems and/or enterprise data 14 are clustered by the building analytics system 17 to determine whether the building automation systems 12 are operating as desired. In another example, data analytics are used by the building analytics system 17 to measure performance of the enterprise, business unit, employee, customer, or other enterprise-related group. The operational data of the building automation systems and business data from a business controlling the multiple buildings are clustered by the building analytics system 17 to determine whether the building automation systems 12 impact the enterprise.
The machine learning employed by the building analytics system 17 finds patterns, behavior, family, clustering, classifications, or other grouping of factors correlating with the performance. In one example, the enterprise 10 is a school system with many buildings for schools. In this example, student performance is used as the measure of performance. This enterprise data may be test scores, grades, or other information available as meta data 14 for access by the building analytics system 17. Any or all of the operational variables of the building automation systems 12 for this school enterprise 10 may be analyzed by the building analytics system 17 to determine correlation with or degree of influence on the performance measure. In this example, the clustering as identified by the building analytics system 17 in accordance with embodiments disclosed herein may indicate that the classroom ventilation directly impacts student performance given other variables remaining the same. The group of buildings with poorer ventilation may be identified by the building analytics system 17 as a cluster.
The other variables or values may impact performance, but to a lesser degree, as determined by the building analytics system 17. Based on the identified clusters, the building analytics system 17 is able to determine whether one group of variables or values impacting performance more substantially than others. Based on pre-selected criteria, such as correlation ranking, the cluster results are ranked by the building analytics system 17 for the user to choose and use. For example, the level of influence, correlation coefficient, or relative impact is used the building analytics system 17 to distinguish between the variables or value range influence on the performance.
In one embodiment, different types of unsupervised learning are applied by the building analytics system 17 in the management system 8 to the same data with the same performance criterion or criteria. For example, different types of clustering are applied by the building analytics system 17 such that the results from the different types of clustering (e.g., correlation coefficients of each variable to a given performance criterion) are averaged, weighted averaged, or otherwise combined by the building analytics system 17. Probability distributions may also be combined. In other embodiments, the results from the different types of clustering are automatically selected by the building analytics system 17 based on a pre-defined ranking. For example, the user pre-selects a ranking criterion or criteria, such as correlation ranking. The results from the different types of clustering are ranked by the building analytics system 17 for the user to choose and/or use. A processor automatically selects the higher N ranked results, where N is an integer of 1 or higher.
In one embodiment, the processor 16 of the building analytics system 17 applies unsupervised learning to identify sub-sets of building automation systems 12. The sub-set may be of underperforming systems 12 or systems 12 with optimal or sub-optimal performance. For example, in this embodiment, the processor 16 is able to identify a correlation of the operational and/or enterprise data with a measure of building automation performance to then identify both the buildings and variables for those buildings associated with the poor performance. In a banking enterprise example, the building analytics system 17 in accordance with disclosed embodiments, may identify one chiller or chillers in the banking enterprise 10 not performing equally across climatic regions. Chiller operation and location may be identified by the building analytics system 17 in a cluster of the poor performing buildings within the enterprise 10. As a result of performing clustering as disclosed herein, the banking enterprise may alter the design of the chillers in some regions of the enterprise 10 without suffering the cost of replacing chillers in all regions.
Enterprise data (e.g., meta data, service records, utility data, business data, and/or budget information), building data (e.g., age and/or location), systems data (e.g., type of distribution system—water and/or air), application data (e.g., building sensor and/or operations data), and/or other types of data are analyzed by the building analytics system 17 in accordance with disclosed embodiments, such as analyzed for building performance, enterprise performance, or other factor. The different buildings being controlled in an enterprise 10 may be a respective building automation system 12 in communication with the building analytics system 17 may be rated by the building analytics system 17 for performance using different criteria and/or sources of data. The data is used by the building analytics system 17 to find insight into the performance and/or control to optimize performance or diagnose building automation or enterprise performance. The enterprise data is used as input variables and values related to performance and/or as the performance.
The clustering is used by the building analytics system 17 to identify groupings or other information used by the predictor 64. In machine training the predicator 64 by the building processor 16, the cluster information may be used. The relationship of any cluster distinctions to forecast operation and/or degradation is learned by the building analytics system 17. Alternatively, the clustering is used by the building analytics system 17 to identify which building automation systems 12 the predictor should learn from and/or be applied to once learnt.
In another embodiment reflected in
As shown in
To use or predict, the machine-learnt predictor 64 of the building analytics system 17 receives an input vector. For prediction, the input vector is values of variables used for training. In the example of
The learned performance being predicted and/or the forecast sets 68 of input and output data indicate degradation of the building automation system 12. The particular combination of predicted data for a time in one set 68 or variation of data across times (e.g., across forecast sets 68) indicates degradation of a particular part of the building automation system 12 or degradation in general. Alternatively, the prediction is part specific, such as different machine-learnt predictors 64 of the building analytics system 17 trained to predict degradation of different parts. In yet other alternatives, the machine-learnt predictor 64 of the building analytics system 17 predicts degradation of the overall building automation system 12 regardless of failure of any particular part.
In one example, an enterprise 10 owns 200 buildings and observes higher energy consumption of chillers installed at the same time and similar buildings. Upon further investigation, the cause is gradual degradation of a cooling coil valve and outdoor air (OA) damper with the air-handling units of the building automation systems 12 as the damper and valves seem to open to the maximum position. As a result, the chillers of the building automation systems 12 have to deliver more energy. Using any number of examples of this occurrence in combination with examples where the fault does not occur, machine learning is used by the building analytics system 17 to create the predictor 64 from the times series of both proper and improper operation. Big data in the form of many examples allows the building analytics system 17 to learn. This machine-learnt predictor 64 of the building analytics system 17 may then be applied to any of the same or different building automation systems 12. The time series of input and output data for a given building automation system 12 is input to the machine-learnt predictor 64 of the building analytics system 17. The predictor 64 then predicts whether and/or when this undesired performance or degradation will occur in the future. The faults in the building automation system 12, if any, may be predicted by the building analytics system 17 so that the faults may be prevented ahead of time.
The machine-learnt predictor 64 is applied by the building analytics system 17 to data from any of the building automation systems 12. The predictor 64 may be trained for a particular building automation system arrangement, such as duplicated building automation systems 12 for similar sized buildings of an enterprise 10. In other embodiments, the predictor 64 is trained on building automation systems 12 with any amount of variation in design, such as different arrangements for different size buildings. The predictor 64 is trained by the building analytics system 17 for a particular enterprise or sub-set of buildings of a particular enterprise 10. Different predictors 64 are trained for different enterprises. Different predictors 64 may be trained for different sets of building automation systems 12 in a same enterprise 10. Different predictors 64 may be trained for predicting different information, such as predicting degraded performance for different parts of building automations systems 12. In other embodiments, a predictor 64 is trained for predicting across more than one enterprise 10.
Once trained, the predictor 64 is applied by the building analytics system 17 to any building automation system 12. Data from many building automation systems 12 is used to train the predictor 64, which is then use to predict for a given building automation system 12. The same predictor 64 may be applied by the building analytics system 17 to different building automation systems 12. The data for each building automation system 12 is input by the building processor 16 separately to predict future operation of the given building automation system 12.
Other classifiers may be used by the building analytics system 17 to determine which building automation systems 12 and/or for what type of degradation to test. Classification is used to select the data to input and/or the predictor 64 to use. For example, clustering using unsupervised learning for an enterprise 10 identifies poor performing building automation systems 12 and/or a source of poor performance (e.g., damper). A predictor 64 of the building analytics system 17 predicts future degradation for the system 12 or source for any building automation systems 12 identified by the clustering. In another hybrid approach, outputs from CMAC and clustering for many building automation systems 12 of an enterprise 10 are further synthesized and processed to provide inputs for further prediction and prognostics using a trained neural network. The trained neural network may handle large big data, such as data with the same measures made over time. The prediction is automated or semi-automated. The machine-learnt predictor 64 allows use of large amounts of data for a particular building automation system 12.
The machine-learnt predictor 64 of the building analytics system 17 may predict future degradation or other event. Any threshold for degraded verses non-degraded operation may be used. The prediction by the building analytics system 17 may be of failure or creation of improper control loop in the building automation system 12. The prediction may be of a likely time or range of likely times for the degradation to occur. The prediction may alternatively or additionally be of what part will have degraded performance or a source of degraded performance. In other embodiments, the prediction is a forecast of performance, and degradation is derived from automated or programmed analysis of the forecast performance.
Other types of machine learning may be used by the building analytics system 17 in addition to or instead of machine learning a predictor 64. Rather than clustering or case-based reasoning, a machine learnt classifier employed by the building analytics system 17 may be trained to diagnose operation of the enterprise and/or building automation system using both building automation data and enterprise data.
Exogenous data, building management system data, other third party data, and/or other data is analyzed by the building analytics system 17 for performance 91. This data analytics by the building analytics system 17 may yield an ideal or desired performance 98, such as using clustering to identify the characteristics (e.g., values) for correlated variables of buildings with better performance. The building analytics system 17 compares this desired performance 98 with the actual performance 92. Using predictive, prognostic, and/or prescriptive analytics 96, the comparison by the building analytics system 17 may trigger an upgrade, change, or retraining of the online predictor or trained classifier.
Once trained, the machine-learnt classifier 90 employed in the building analytics system 17 receives the input feature vector from the enterprise, utility, or other data to predict performance 94. The predictive performance 94 may be compared by the building analytics system 17 to actual performance 92 for use in other various analytics 96. The output of the machine-learnt classifier 90 may be used in clustering processes performed by the building analytics system 17, such as relating predicted performance 94 of the energy or operation of the building automation to an enterprise performance variable. Clustering as employed by the building analytics system 17 may be used to derive an input for the input features vector of the machine-learnt classifier 90.
In another embodiment represented in
The data to be used for training and inversion may be determined by clustering employed by the building analytics system 17 in accordance with the embodiments described herein. The variables most determinative of the desired operation or energy performance 104 are determined by clustering pre-process by the building analytics system 17.
Returning to
In other embodiments, the processor 16 transmits results for use in control or other uses. The building automation systems 12 may be controlled to increase performance. The results may be transmitted to a manager or service to schedule maintenance to avoid failure or degradation. The prediction is used to avoid downtime or costs associated with improper operation. The building processor 16 outputs the prediction to the technician on the display 18, outputs a message to a supervisor, or outputs a calendar event for training.
Additional, different or fewer acts may be provided than shown in
In act 50 for learning, a building processor or other part of the building analytics system accesses data related to a plurality of buildings through an interface 19 or from memory. The access is by receipt of information, request of information, or loading information. Multiple memories may be mined by the processor in the management system 10 related to multiple building automation systems in an enterprise. In alternative embodiments, data related to a single building is accessed.
The data includes building management system or building automation system operational data with or without enterprise data different than the building management system or building automation system data. The building automation systems in the enterprise and in communication with the building management system generate data specific to the building automation. For example, actuator settings, sensor readings, set points, meter information, weather, utility information, measures of performance, or other input or output data for the daily operation of the respective building automation system are accessed. The building management system includes automation for heating, cooling, ventilation, fire safety, or combinations thereof data.
The enterprise generates data specific to the business of the enterprise. The business of the enterprise is not automation of the buildings. Instead, budget, maintenance, employee complaint, or human resources data of the enterprise is accessed.
The enterprise data is accessed by the processor from an enterprise database 14. The enterprise database 14 is one or more memories organized as one database or as separate data structures. The enterprise data representing one or more variables is accessed. The values for a given variable may be the same or different across the multiple buildings. For example, the maintenance budget for the building is associated with the multiple buildings but may or may not be different for different buildings. The amount of deviation from the budget is more likely to be different for different buildings.
The accessed data may be all or a default sub-set of all available data. Alternatively, a user indicates the data to access. The user configures the analysis by indicating the prediction to be made, such as the user indicating a prediction of whether valve or damper operation will degrade. This input may indicate specific data to access, such as data likely to be used by machine learning to predict the operation. In other embodiments, a specific part is not indicated. Instead, the operation of the automation system in general, all parts, or default ones of the parts of the building automation are to be predicted. The machine learning may indicate which variables and corresponding data correlate with the prediction and which do not. Less than all of the originally selected data is used for the trained predictor 64, such as just using the determinative variables.
Some of the accessed data includes a time series. Data from different times is accessed to be used for learning to predict. Values for the same variables of the same devices (e.g., building automation component parts) are provided at different times. For example, the sets 60 of
In one embodiment, data is also accessed by the processor for classification other than prediction. For example, data to machine train a classifier and/or data for unsupervised (e.g., clustering) machine learning is accessed. The classifier is trained to output desired information, such as identify a source of fault and/or identify poor performing building automation systems 12. The role of classification is to broadly classify good performers from bad performers. The result is then fed into CMAC learning to identify a source of the fault. This may be used to limit the training of the predictor 64 to particular building automation systems 12, parts, or faults. For example, clustering is used by the processor to distinguish good and poor performing building automation systems 12. Prediction is learned from data from the poor performing building automation systems 12, such as data during good performance as well as bad performance. Data from the good performing buildings is not used. Alternatively, data from both sets of buildings is used, but used differently in learning to predict
In act 52 for learning, a building processor 16 of the building analytics system 17 applies machine learning to the accessed data. For learning to predict, some of the data is used as input to the learning and other of the data is used as a measure of performance or the ground truth of the predicted operation. For example and as represented in
The machine learning employed by the processor is a neural network, a recurrent neural network, or other machine learning for dealing with a time series or prediction. The machine learning learns to statistically relate the input values to the ground truth. For a neural network, layers of nodes, weights for the nodes, and connections between the nodes are learned to predict or forecast the output sets 62 representing future operation from the input vector (i.e., input horizon sets 60). Other data that is not time series may be used as an input, such as an enterprise variable that is predictive of future fault or operation.
Using many examples in the training in the management system, the machine learning by the processor may be more accurate. The many examples may make the machine learning more able to learn to predict given input values different than any of the training data.
In act 54 of the learning, the building processor 16 of the analytics system 17 outputs results of the application of machine learning to a display 18, network, memory, or other processor. The machine-learnt predictor 64 of the operation of the building automation systems 12 is output. For example, the learnt neural network, such as in the form of a matrix, for predicting degradation is output. More than one machine-learnt predictor 64 may be output, such as outputting predictors 64 trained to make different predictions. For example, the predictor 64 is trained to predict operation of a part or sub-system of the building automation systems 12. The predictor 64 may additionally or alternatively predict a time of occurrence of degraded or other performance.
The output machine-learnt predictor 64 is used or applied. Later-acquired data, relative to the training data, is input to the machine-learnt predictor 64. The machine-learnt predictor 64 forecasts the operation or occurrence of an event (e.g., degraded performance) based on the input data, but without the ground truth. Future operation is predicted.
The same or different building processor 16 and/or building analytic system 17 apply the learnt predictor 64. In other embodiments, a given building automation system 12 applies the predictor 64.
In act 50, the building analytics system 17, building processor 16, or other device of the building management system 8 accesses input data. The input data is from a given building automation system 12. Different building automation systems 12 are analyzed separately. Some of the same data may be used for different building automation systems 12, such as enterprise data associated with the different automation systems 12 or data output by other machine-learnt classification.
The same type of data used to train is accessed by the processor. The type of data used for ground truth is not accessed as the machine-learnt predictor 64 predicts the future or makes a forecast. In the example of
In act 52 of the application of the learned predictor 64 by the processor, the machine-learnt classifier (i.e., predictor 64) predicts operation of the building automation system 12 or part of the building automation system 12. The analytics system 17 or building processor 16 inputs the accessed data of the input vector into the machine-learnt predictor 64.
Based on the input data, the machine-learnt classifier employed by the processor outputs the forecast or prediction. For example, failure of a part is predicted based on the input times-series data with or without other data. As another example, a time of degraded performance (e.g., starting or range of times) is predicted. In other embodiments, the operation over time is forecast. The forecast operation is analyzed to identify the information of interest (e.g., identify failure using any criterion from the predicted operation data).
The machine-learnt classifier is applied by the processor to any building automation system 12. In one embodiment, the building automation system 12 to which the predictor 64 is to be applied is identified with a different machine-learnt classifier. For example, clustering is used to identify poorly performing building automation systems 12. Any data, such as the same or different data, is used for this initial classification. Data from all or many building automation systems 12 is input to classify membership or cluster. The predictor 64 is then applied to the building automation systems 12 with poor performance. The predictor 64 indicates if and/or when failure or further degradation will occur based on input data for the given building automation system 12. For example, the trained recurrent neural network is separately applied to operation time-series data from each poorly performing building automation system 12. The neural network outputs the forecast for each automation system 12. In other embodiments, the predictor 64 is applied to user selected, all, or other building automation systems 12.
In act 54 of the application of the machine-learnt predictor 64, the forecast, prediction, and/or prognostics are presented on the display 18 of the building analytics system 17 by the processor 16. The output may be instead transmitted and output on another device, such as printed or displayed remotely.
Any results, such as failure occurring, the building automation system 12 expected to fail, the part expected to fail, the time of expected failure, or other prediction of failure, is output. The result may be failure, degraded performance, enhanced performance, or other event associated with building automation. The forecast operation may be output. Alternatively, information derived from the forecast operation is output, such as a time and/or device operating in an undesired way. Probability information may be output, such as providing a range of times of expected degradation in performance with probabilities of the degradation provided for each of the times.
The output may be used to schedule maintenance, such as replacing a part before breakage and/or during an already scheduled shut down. The output may be used to identify commonality for altering design. The output may be used to alter operation of the enterprise 10, such as relocating employees before an expected shutdown. The output may be used to initiate analysis by a technician in an effort to identify a control problem.
While the invention has been described above by reference to various embodiments, it should be understood that many changes and modifications can be made without departing from the scope of the invention. It is therefore intended that the foregoing detailed description be regarded as illustrative rather than limiting, and that it be understood that it is the following claims, including all equivalents, that are intended to define the spirit and scope of this invention.
This application claims the benefit of the filing date of U.S. Provisional Patent Application 62/131,749, filed Mar. 11, 2015, which is hereby incorporated by reference to the extent permitted by law.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/US2016/020025 | 2/29/2016 | WO | 00 |
| Number | Date | Country | |
|---|---|---|---|
| 62131749 | Mar 2015 | US |
