Patent Application
20230342633
This application is the National Stage of International Application No. PCT/EP2020/072740, filed Aug. 13, 2020. The entire contents of this document is hereby incorporated herein by reference.
A manufacturing system is a collection or arrangement of operations and processes used to make a desired product or component. The manufacturing system includes the actual equipment for composing the processes and the arrangement of those processes. In a manufacturing system, if there is a change or disturbance in the system, the system may accommodate or adjust itself and continue to function efficiently.
Simulation in manufacturing systems provides for the use of software to make computer models of manufacturing systems, to analyze the computer models, and thereby obtain useful information about the operational behavior of the system and of the material flow in the system. A schematic representation of such a system is shown in
Physical simulation models of automated factories and plants contain all kinds of useful information about the operation behavior.
One State-of-the-art approach to employ simulations to improve the performance of machine learning approaches is depicted in
Simulation models are routinely used during the engineering phase (e.g., to determine the optimal design or parameterization of drive controllers). The simulation models are also used to produce training data for condition monitoring and failure prediction algorithms.
It is already known for condition monitoring and predictive maintenance to provide a combination of real sensor data and data from simulations during the training phase of the underlying machine learning (ML) model.
In machine learning, a feature is an individual measurable property or characteristic of a phenomenon being observed. Choosing informative, discriminating, and independent feature information is a step for effective algorithms in pattern recognition, classification, and regression. Feature information is often numeric, as in the chosen examples later.
The state of the art is schematically depicted in
An example of a Machine Learning algorithm is a Gradient Boosted Decision Tree. It is already known to the expert in the field to use simulation methods to provide training data for decision tree models, also referred to as Decision Tree Analysis, DTA.
Decision tree learning is one of the predictive modelling approaches used in statistics, data mining, and machine learning. Decision tree learning uses a decision tree (e.g., as a predictive model) to go from observations about an item (e.g., represented in the branches) to conclusions about the item's target value (e.g., represented in the leaves). https://en.wikipedia.org/wiki/Decision_tree_learning. Publication CN 109241649 A describes such a method for composite material detection, where training data is provided by finite elements simulations. CN 109239585 A uses circuit simulation data to train decision models for failure detection in electrical circuits.
However, during runtime, sensor data is typically analyzed independent of the simulation models. Such a procedure wastes valuable information and therefore is compromising the performance of the condition monitoring system.
The scope of the present invention is defined solely by the appended claims and is not affected to any degree by the statements within this summary.
The present embodiments may obviate one or more of the drawbacks or limitations in the related art. For example, a method, computer program product, and system to overcome the described disadvantages of the known method and provide the possibility to enter additional information from the model as described are provided.
A method for an augmented decision tree analysis in a Machine Learning algorithm for a manufacturing system includes inputting of input data containing data acquired during operation, amending of input data with feature information, and applying the input data in a decision tree analytics model with each leaf of the decision tree representing a machine state associated with a label giving information about feature values and operational conditions of the manufacturing system. Branches of the decision tree represent conjunctions of feature information that lead to those states and labels. There is at least one Simulation Model that shows dependencies between the label and the input data, and at least one of the Simulations models is replacing at least one part of at least one of the branches of the tree.
The proposed approach differs substantially from the described state of the art.
A computer system as depicted schematically in
The simulation models and analytical models are combined on a semantic, syntactic, and lexical level. Different available simulation models for data analytics are represented in
As a precondition for the procedure, physical factory, plant, machine, or device models that are used to set up factory, plant, or machine devices are built. Further, process data with labels from factories, plants, machines, or devices during operation has been acquired, and models have been parameterized to fit the acquired data for the labeled conditions. Then, the procedure is carried out to support data analysis with goals such as anomaly detection, root cause analysis, condition monitoring, prediction, or optimization based on acquired data and physical model. An overview of the analytics process for condition monitoring is shown in the diagram in
In the case of anomaly detection, a simulation model SM0, SM1, SM2 is used to generate example anomaly data that is compared to the acquired data, or acquired data is input to the simulation model while comparing the output to other parts (e.g., other measurement channels or time frames) of the acquired data. A comparison is carried out by an error calculation or correlation. In case enough overlap of the simulated and acquired data is detected, a simulation model specific anomaly is notified.
In the case of root cause analysis, a simulation model is used to determine causes of anomalies by identifying signals that influence anomalies.
In the case of prediction, a simulation model is used to simulate future behavior and predict expected values.
In the case of optimization, parameters and input values of a simulation model are varied in order to find optimum output values.
In the case of condition monitoring, a simulation model is used to generate example data for a number of conditions, define feature information, explain which input data is relevant for which condition, and generalize and enhance analytics models.
A detailed description of the condition monitoring case contains the following acts a to j, depicted also as an overview in
An example of such a table looks like that:
The values of the table may be used for simulation Models M, as different possibilities are shown in
This process is automated and may be supported by a user interface (not shown in the figures) guiding the user through the analytics procedure.
The user interface may be used to map simulated to measured data automatically considering model labels (e.g., 1) and valid model regions. In one embodiment, mapping is carried out by scripts or standardized interfaces (e.g., Predictive Model Markup Language (PMML)—Functional Mock Up (FMU) mapper). The mapping is supported by similarity propositions.
The goal is to derive a decision tree that shows how the labeling classes depend on the acquired data so that based on acquired data, the existing condition class is automatically shown so that appropriate actions may be carried out by the maintenance staff or operator. In the automated case, the mapping procedure uses similarity scores that, in the user supported case, may also be proposed to the user.
Similarity scores are derived by finding word similarities of the input and output data descriptions and signal similarities by calculating signal correlations.
Similarity indicators to support the mapping between simulation and measured data for analytics are based on:
Word similarities:
Simulation model inputs and outputs are described in XML files such as FMU. For example,
Analytic model inputs and outputs are described in XML files such as PMML, ONNX. For example,
The method parses the XML description files and looks for similarities in hypertext and text descriptions. Similarity is defined comparing words and using semantic word nets, for example. Type name and simpleType name have a similarity score of 18/24=0.75; AngularVelocity and AngularVelocity have a similarity score of 1
Signal correlations: A cross correlation score between signals is considered to describe similarity. Additionally, anomalies are detected in signals, and two signals are considered similar if anomalies are recognized at similar time steps.
In the example shown in
Multiple similarity scores may be aggregated, for example, by calculating a mean similarity score for the input and output variables.
Based on parsing the description of input and output values in standardized XML files from, for example, Functional Mock Up (FMU) for simulation models or Predictive Model Markup Language (PMML) for analytics models similarity scores (e.g., known from text analysis or using semantic webs are calculated). Cross-correlations between signals are calculated as an indicator for signal similarity. Also, the relative amount of common time steps with anomalies is taken as a similarity score. The scores for each simulation input/output measured analytics input/output data pair are aggregated (e.g., by calculating a mean score). Optionally, the mean score may also be visualized to the user in the User interface together with more detailed information of the aggregation procedure.
Proceeding with the forgoing example table, valid region: I>10; O>20
then classification methods such as decision tree analysis DTA are used to learn an analytics model. First, relevant values and valid regions of a simulation model are identified. A valid model range is determined by simulating model outputs with interpolated and extrapolated input values and models associated with a given label, 115.
The input/output/label relation is checked whether it is consistent with the analytics model—if not, then a limit of the valid model range is reached.
Contradictions are tested for using a distance measure between simulations and measurements. For example,
ε=Σi,j|Iijmeas−Iijsim|+Σij|Oijmeas−Oijsim|
The measurement is used if there is a contradiction.
This is shown in an example in
Second, the amount of necessary simulation data is determined depending on the analytics method 114.
In the case of neural network analytics, the required amount of data increases with the number of neurons.
In the case of a decision tree analysis, the required amount of data increases with the minimum splitting value and depth of the tree.
Also, in the case of difficult conditions for analytics models such as unbalanced labels or too few data for the degrees of freedom of the analytics model, simulation models are used to increase the amount of available data. Further, for failure cases where a realization is cumbersome or even impossible, simulation models are used to fill this lack of data.
In act 1, the feature information F that may be represented and explained by simulation model SM will be defined (see table 700) based on Input data I and Output data O.
In act 2, the pre-trained feature information F that fits to label L and correlates to analytics model AM is defined.
The feature information that may be represented and explained by the physical model are calculated from the measured data and used to build an analytics classification model so that labels are associated with feature information values.
In a second act, model feature information outputs for certain labels are compared to measured feature information at a given input. The simulation model yielding best agreement based on the feature information values (e.g., smallest error ε between model feature information and measured feature outputs) is associated with the respective label.
Act 1: Define a feature information that may be represented and explained by model.
Act 2: Define pre-trained feature information that fits to label and correlates to model.
Error ε=F1meas−F1sim determines which simulation model describes the label with the respective feature information value.
[N0, N1] where N0 is the number of data sets with label
N1 number of data sets with label 1
In
In other words, a simulation model is used to provide relevant feature information that is related to simulation parameters (e.g., spring stiffness, speed, weight, torque, temperature, load, current, power), and thus to physical values and technical components of the manufacturing system, machine, or factory.
The processing of the data is depicted in the diagram 900, similar to that of 800, but without the Analytics Model AM, filtered for Label L=0. In this example, for each label, the measured output O values are compared to the output of each simulation model SM0, SM1 at the same input I. The simulation model with the smallest error is then associated with the label.
In a decision tree analysis, the models are hence already implicitly associated to branches of the tree 901 on a syntax level.
In the decision tree 1001, it is noted that on the left side with V<6, O<10, and I<6, the Simulation Model M1 fits, but on the other side, with V<6, O>=10, and I>=20, M1 does not fit.
Feature information values are added as inputs if the feature information values are less correlated to the existing inputs but more correlated to outputs, and are added as outputs if the feature information values are less correlated to the existing outputs but more correlated to inputs.
In a second act, simulation model parameters are adapted to reproduce all feature information that is used by the analytics. In a fault classification example with time series input that is preprocessed by a Fourier transformation into frequency space, the analytic decision tree analytics model has shown the frequency feature columns of importance (e.g., 0 and 20) that may be used to classify the data into label 0 (good) and label 1 (faulty) conditions.
Hence, the simulation model used to generate more data was designed to contain a resonance frequency of 4.5 Hz corresponding to feature column 20, as shown in
The simulations module is improved to contain a resonance frequency at 4.5 Hz corresponding to feature information 20 that is used by the decision tree to distinguish classes with label 0 and 1.
In an application example, an analysis leads to two decision trees that both give 100% accuracy. A simulation model, however, indicates that only feature information values 17 to 40 would give robust, physically explainable classification, so that the decision tree model building on feature information values in this range is chosen.
In that way, the simulation results also help to choose the decision tree model, such that the complexity and/or the depth of the tree, which is the main source of overfitting, is reduced.
The diagram in
The diagram in
Simulations indicate that features between 17 and 40 (e.g., corresponding to frequency ranges 4.5 to 10 Hz) are best indicators for separation, so that DTA 2 may be chosen.
Another example is shown in
An example pseudocode implementation for the example depicted in
That example of an augmented decision tree 1411 with Simulation Models M0, M1 and data table 1400 shows the combination of simulation and analytics model on the lexical level.
The method of the present embodiments shows the following advantages: Instead of using additional simulated data, simulation models are directly introduced in the decision tree. By including simulation models in branches of decision trees, the decision tree depth is reduced, and accuracy is enhanced at a given number of data sets.
As a result, overfitting is reduced because a simulation model describes a physical behavior that is more generally valid.
For each class, a separate model may be used. An additional class is introduced, which indicates the points that cannot be labeled correctly.
j) Finally, the combined trained machine learning/simulation model is used for continuous classification as shown in
In comparison to the prior art solution, the procedure of the present embodiments has further advantages: The physical meaning of data inputs increases the acceptance of analytics model by humans in the sense of explainable AI.
A generalization of analytics model to physically relevant findings makes analytics also more explainable.
Additionally, the support of the feature engineering process is improved by proposing physically relevant feature information.
Further, the support of the simulation model creation process by analytics models and simulation data leads to analytics models that are more robust to changes in the environment, and the overfitting to a limited number of data sets is reduced. Generation of simulation data over a wide range of conditions, especially for the fault cases, increases the amount of training data and thus helps to overcome overfitting.
The support of analytics model selection by single and multivariate simulation models improves precision of classification areas and reduces tree complexity.
The method offers support to operate analytics and simulation models simultaneously.
The following technical feature information contribute to the mentioned advantages: an optional user interface to map simulation model inputs/outputs to data columns on a semantic level using similarity scores for user guidance; the management of labeled models, simulation data, and improvements to analytics models; use of a simulation model to propose relevant feature information that is related to simulation parameters (e.g., spring stiffness), and thus to physical values and technical components of the machine/factory; proposed improvements to simulation models, based on relevant feature information and identified by a decision tree analysis; relevant decision trees are based on simulation results so that the complexity/depth of the tree, which is the main source of overfitting, is reduced; inclusion of a simulation model in a decision tree on a lexical and syntax level so that robustness and precision are increased; and support of continuous operation of decision tree models with included simulation models.
The present embodiments provide a combination of system simulation models and decision trees (e.g., the integration, such as replacement of tree-branches with a simulation model). Also, a proposition to map simulation inputs/outputs to measurement and analytics data columns based on a similarity measure of description, anomaly similarity, and correlation is provided.
The elements and features recited in the appended claims may be combined in different ways to produce new claims that likewise fall within the scope of the present invention. Thus, whereas the dependent claims appended below depend from only a single independent or dependent claim, it is to be understood that these dependent claims may, alternatively, be made to depend in the alternative from any preceding or following claim, whether independent or dependent. Such new combinations are to be understood as forming a part of the present specification.
While the present invention has been described above by reference to various embodiments, it should be understood that many changes and modifications can be made to the described embodiments. It is therefore intended that the foregoing description be regarded as illustrative rather than limiting, and that it be understood that all equivalents and/or combinations of embodiments are intended to be included in this description.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2020/072740 | 8/13/2020 | WO |
