Patent Application
20230071394
The disclosed implementations relate generally to cyber-physical systems and more specifically to systems and methods for cyber-fault detection in cyber-physical systems.
Performance of traditional cyber-fault detection systems for industrial assets depend on availability of high definition simulation models and/or attack data. Conventional detection methods for cyber-faults in industrial assets cast the detection problem as a two-class or multi-class classification problem. Such systems use significant amount of normal and attack data generated from high definition simulation models of the asset to train the classifier to achieve high prediction accuracy. However, these techniques have limited use when the attack data is limited or unavailable, or when no simulation model is available to generate attack data.
Accordingly, there is a need for systems and methods for detection of cyber-faults (cyber-attacks and system faults) with high accuracy in industrial assets in such scenarios. In one aspect, some implementations include a computer-implemented method for implementing a one-class classifier to detect cyber-faults. The one-class classifier may be trained only using normal simulation data, normal historical field data, or a combination of both. In some implementations, to boost the detection accuracy of the one-class system, an ensemble of detection models for different operating regimes or boundary conditions may be used along with an adaptive decision threshold based on the confidence of prediction.
In one aspect, some implementations include a computer-implemented method for detecting cyber-faults in industrial assets. The method may include obtaining an input dataset from a plurality of nodes (e.g., sensors, actuators, or controller parameters) of industrial assets. The nodes may be physically co-located or connected through a wired or wireless network (in the context of IoT over 5G, 6G or Wi-Fi 6). The nodes need not be collocated for applying the techniques described herein. The method may also include predicting a fault node in the plurality of nodes by inputting the input dataset to a one-class classifier. The one-class classifier may be trained on normal operation data (e.g., historical field data or simulation data) obtained during normal operations (e.g., no cyber-attacks) of the industrial assets. The method may further include computing a confidence level of cyber fault detection for the input dataset using the one-class classifier. The method may also include adjusting a decision threshold based on the confidence level for categorizing the input dataset as normal or including a cyber-fault. The method may further include detecting the cyber-fault in the plurality of nodes of the industrial assets based on the predicted fault node and the adjusted decision threshold.
In another aspect, a system configured to perform any of the methods described in this disclosure is provided, according to some implementations.
In another aspect, a non-transitory computer-readable storage medium has one or more processors and memory storing one or more programs executable by the one or more processors. The one or more programs include instructions for performing any of the methods described in this disclosure.
For a better understanding of the various described implementations, reference should be made to the Description of Implementations below, in conjunction with the following drawings in which like reference numerals refer to corresponding parts throughout the figures.
Reference will now be made in detail to implementations, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the various described implementations. However, it will be apparent to one of ordinary skill in the art that the various described implementations may be practiced without these specific details. In other instances, well-known methods, procedures, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the implementations.
It will also be understood that, although the terms first, second, etc. are, in some instances, used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first electronic device could be termed a second electronic device, and, similarly, a second electronic device could be termed a first electronic device, without departing from the scope of the various described implementations. The first electronic device and the second electronic device are both electronic devices, but they are not necessarily the same electronic device.
The terminology used in the description of the various described implementations herein is for the purpose of describing particular implementations only and is not intended to be limiting. As used in the description of the various described implementations and the appended claims, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term “and/or” as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items. It will be further understood that the terms “includes,” “including,” “comprises,” and/or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof
As used herein, the term “if” is, optionally, construed to mean “when” or “upon” or “in response to determining” or “in response to detecting” or “in accordance with a determination that,” depending on the context. Similarly, the phrase “if it is determined” or “if [a stated condition or event] is detected” is, optionally, construed to mean “upon determining” or “in response to determining” or “upon detecting [the stated condition or event]” or “in response to detecting [the stated condition or event]” or “in accordance with a determination that [a stated condition or event] is detected,” depending on the context.
Cyber-fault attack data is rare in field. On top of that, generating abnormal dataset of cyber-attacks and system/component faults is a slow and expensive process requiring advanced simulation capabilities for the system of interest and a lot of domain knowledge. Therefore, it is essential to develop methodologies for cyber-fault detection and localization without abnormal dataset generation or simulation data altogether. For the description herein, normal data is data collected during operation of the asset that is considered ‘normal’, and attack data is data in which one or more node is manipulated. High definition simulation models are models that capture details of the nonlinear physics involved. Typically, the execution of these models may be slower than real time execution. Techniques described herein can be used to implement detection systems that are trained only on historical field data thereby eliminating dependence on availability of high definition simulation model and/or substantial amount of attack data. Another use case is when there is a high definition simulation model available, but generation of attack data is expensive both in terms of time and money. In such scenarios, if a model has to be deployed quickly, some implementations may generate a limited set of normal data to start with, and upgrade the detector as time progresses.
Some implementations use an ensemble of models for prediction of faulty nodes or nodes experiencing fault nodes depending on accuracy of different models (i) for different operating regimes (e.g., steady state, slow/fast transient, rising/falling transient and so on), and (ii) for different boundary conditions (e.g., environmental conditions such as temperature, pressure, humidity and so on). This technique boosts the true positive rate (TPR) of detection compared to that obtained with a single monolithic model.
In some implementations, as described in detail below, decision thresholds on residuals are adapted based on the confidence of prediction accuracy. Residuals are appropriate functions of the difference between ground truth and a predicted value. For a multi-variable case as in the instant case, an appropriate norm is chosen to get a simplified metric. A relatively high confidence would result in a more aggressive tuning of the decision thresholds whereas a lower confidence would adjust the tuning accordingly. This technique lowers the false positive rate (FPR) of detection by relaxing decision thresholds in region of lower confidence resulting either due to inherent lower local sensitivity of the model or due to extrapolation of boundary conditions (e.g., encountering a boundary condition which is either not within its training envelope or in a sparse region).
Some implementations use a decision playback capability that allows for reducing false alarms using persistence criteria, while feeding back the early decision to a neutralization module since the onset so that the control system is not drifted too far because of decision delay.
As stated above in the Summary section, conventional detection methods for cyber-faults in industrial assets deal with the problem as a two-class or multi-class classification problem. Significant amount of normal and attack data are generated from high definition simulation models of the asset to train the classifier to achieve high prediction accuracy. The paradigm, however, is not applicable when no/limited attack data is available, and no simulation model is available to generate enough attack data or when data generation is expensive for the problem at hand.
To circumvent this issue, the use of one class classifiers is described in this disclosure for detection of cyber-faults.
A decision threshold adjustment module 108 of the system 100 may feed suitable decision thresholds 118 to the comparator module 110, which may generate the attack/no attack decision 112 for each sample by comparing the decision thresholds 118 to the residuals 116. The nominal decision thresholds are decided based on the distribution of residuals of normal data which are then adapted in real time based on the confidence on reconstruction of that sample.
A confidence predictor module 106 may predict confidence in the accuracy of the decision 112. In some implementations, the confidence predictor module 106 makes the prediction based on the input sample from the nodes 102, the nodes' relative location with respect to the hyperspace spanned by the training data, local sensitivity function of the reconstruction model 104 and the neighborhood of the operating point. The following subsections describe each of the modules in more details.
In some implementations, the reconstruction model 104 is a map :
n×w→
n×w, which takes as input the windowed data-stream from the nodes X ∈
n×w, where n is the number of nodes and w is the window length, compresses them to a feature space
∈
m, m<<n×w, and then reconstructs the windowed input back to {tilde over (X)} ∈
n×w from the latent features ƒ ∈
.
may be a combination of a compression map
:
n×w→
m and a generative map
:
m→
n×w. During training,
exploits the features in the normal data to learn the most effective way to compress X to
and reconstruct {tilde over (X)} from
simultaneously by solving the optimization problem
Because the compression and generation may be learnt on normal data only, any sample whose feature correlation does not resemble that of the normal dataset would have a relatively high reconstruction error. Any mapping into the feature space that is reversable can be used within this framework. For example, models like deep autoencoder, GAN or a combination of PCA-inverse PCA may serve as the model with different degrees of accuracy. For small number of nodes and where the correlation between nodes are primarily linear, a PCA-inverse PCA may be used for quick training and deployment. Here nodes can be either sensor or actuators which have a data stream attached thereto. However, as the number of nodes increase and the correlation becomes more complex, a deep neural network-based model like an autoencoder or GAN may be used, especially when a lot of data is available. Autoencoder or GAN may also have the advantage of being amenable to automated machine learning for rapid training and deployment on high volume of data and scalable across number of nodes and/or assets.
Here, note that can either be a monolithic model or an ensemble model, where the constituent models would be trained on different suitable subsets of the normal data. The reconstruction in that case is given by {tilde over (X)}=Σj=1p αj
j(X), where
j are the respective constituent reconstruction models for j=1,2, . . . , p, and αj is the corresponding weighting factor. Note that the vector
During operation, a preprocessing module may determine the location of the input X with respect to the training subspaces of the constituent models, which in turn may decide the elements of the weighting vector for a monolithic model would benefit substantially by employing the ensemble technique appropriately.
The confidence of reconstruction (e.g., using the reconstruction model 104), which is essentially an indication of its accuracy, may vary depending on various cases even in normal conditions. Accordingly, it may be important to adjust decision thresholds (used in deciding whether a datapoint is normal or anomalous) accordingly so that an optimum balance between FPR and TPR are maintained. Most common reasons for variation in confidence may include local model sensitivity, model uncertainty, and extrapolation, discussed below. The following subsections describe how some implementations tackle each of these cases. In some implementations, hardened sensors (if available) are used as an additional source of confidence. Hardened sensors are sensors that are physically made secure by using additional redundant hardware.
Local model sensitivity: In some implementations in which the reconstruction model 104 is a highly nonlinear model, the sensitivity of the model will vary based on its operating point. Assuming a stationary output noise, higher sensitivity regions would be more capable in resolving a smaller difference, thus making the reconstructions more accurate. The sensitivity of the model as a function of input space can be computed beforehand or online and may be an indicator of the reconstruction confidence.
Model uncertainty: Depending on sparsity of training data in certain regions, the accuracy of reconstruction may vary. Based on the training set, the uncertainty may be precomputed and serve as a second indicator of the reconstruction confidence.
Extrapolation: During deployment, the reconstruction model 104 may see data points which fall outside the training boundary. The reconstruction accuracy is expected to be lower in those regions and a suitable metric denoting the statistical distance of such a datapoint from the training boundary may serve as a confidence metric or another indicator of the reconstruction confidence.
Some implementations designate boundary conditions and/or hardened sensors to decide the location of the sample with respect to the training set. In absence of that, all attacks would likely be classified as a sparse region/extrapolation from training set. If most of the attacks are accompanied by lower confidence predictions, they would be evaluated against relaxed thresholds, leading to a lower TPR. Some implementations design the confidence metric to avoid this undesirable scenario.
The decision thresholds 118 are an important component in the whole system to categorize a sample as a normal datapoint or an attack (or cyber fault) datapoint. If the decision thresholds 118 are set too low, then the FPR would be high as some of the noise in the normal data would be categorized as attacks. Conversely, a high decision threshold would amount to missing certain attacks of small magnitudes. Thus, tuning the decision thresholds 118 for optimal TPR/FPR metric may provide more accurate decisions.
The nominal decision threshold vector
In various implementations, the threshold adaptation vector
Depending on the usage scenario, FPR can have a varied requirement. If the end goal is to raise an alarm/flag to alert an operator, some delay can be tolerated between the attack and decision to keep the false alarm rate low. On the other hand, if the decision is to be fed back to a cyber-fault neutralization systems, then a delay in decision communication may jeopardize the stability of the whole system. In such cases, it might be beneficial to start feeding back the decisions 112 as they come in even at the expense of a slightly higher FPR so that the automated downstream system is engaged. Suppose a first tier relays decisions based on single samples. This may have a higher FPR, but a lower detection delay. A second tier may relay decisions after a persistence window. This will help reduce the FPR of the first tier, while appropriately letting mechanisms engage without delay. If the second tier confirms the decision at the end of the persistence period, the downstream system would remain engaged with probably an additional visual alarm/flag (thus enabling playback in the past) and disengage otherwise.
The techniques described above are amenable to AutoML paradigm, making it easier and faster to train, update and deploy the reconstruction models. The scalable architecture makes it suitable for both unit level and fleet level deployment. As described above, the model is trained only on field data (no simulation model needed) which in turn makes it suitable to be deployed on assets from other manufacturers.
The computer 306 typically includes one or more processor(s) 322, a memory 308, a power supply 324, an input/output (I/O) subsystem 326, and a communication bus 328 for interconnecting these components. The processor(s) 322 execute modules, programs and/or instructions stored in the memory 308 and thereby perform processing operations, including the methods described herein.
In some implementations, the memory 308 stores one or more programs (e.g., sets of instructions), and/or data structures, collectively referred to as “modules” herein. In some implementations, the memory 308, or the non-transitory computer readable storage medium of the memory 308, stores the following programs, modules, and data structures, or a subset or superset thereof:
Details of operations of the above modules are described above in reference to
The above identified modules (e.g., data structures, and/or programs including sets of instructions) need not be implemented as separate software programs, procedures, or modules, and thus various subsets of these modules may be combined or otherwise re-arranged in various implementations. In some implementations, the memory 308 stores a subset of the modules identified above. In some implementations, a database 330 (e.g., a local database and/or a remote database) stores one or more modules identified above and data (e.g., decisions 112) associated with the modules. Furthermore, the memory 308 may store additional modules not described above. In some implementations, the modules stored in the memory 308, or a non-transitory computer readable storage medium of the memory 308, provide instructions for implementing respective operations in the methods described below. In some implementations, some or all of these modules may be implemented with specialized hardware circuits that subsume part or all of the module functionality. One or more of the above identified elements may be executed by the one or more of processor(s) 322.
The I/O subsystem 326 communicatively couples the computer 306 to any device(s), such as servers (e.g., servers that generate reports), and user devices (e.g., mobile devices that generate alerts), via a local and/or wide area communications network (e.g., the Internet) via a wired and/or wireless connection. Each user device may request access to content (e.g., a webpage hosted by the servers, a report, or an alert), via an application, such as a browser. In some implementations, output of the computer 306 (e.g., decision 112 generated by the decision threshold comparator module 110) is communicated to a control system that controls the nodes 102 of the industrial assets 302.
The communication bus 328 optionally includes circuitry (sometimes called a chipset) that interconnects and controls communications between system components.
In some implementations, the method further includes computing reconstruction residuals (e.g., using the reconstruction model 104) for the input dataset such that the residual is low if the input dataset resembles the normal operation data, and high if the input dataset does not resemble the historical field data or simulation data. Detecting cyber-faults in the plurality of nodes includes comparing the decision thresholds to the reconstruction residuals (e.g., using the decision threshold comparator module 110) to determine if a datapoint in the input dataset is normal or anomalous.
In some implementations, the one-class classifier is a reconstruction model (e.g., a deep autoencoder, a GAN, or a combination or PCA-inverse PCA, depending on the number of nodes) configured to reconstruct nodes of the industrial assets from the input dataset, using (i) a compression map that compresses the input dataset to a feature space, and (ii) a generative map that reconstructs the nodes from latent features of the feature space. In some implementations, the reconstruction model is a map :
n×w→
n×w that obtains windowed data-stream from the nodes X ∈
n×w. n is the number of nodes and w is the window length. n can be a few nodes to several hundred nodes depending on the asset; for w, depending on the asset dynamics and sampling rate, it can be a few tens to a few thousands. The compression map is a map
:
n×w→
m that compresses the windowed data-stream to a feature space
∈
m, m<<n×w, where m is the latent space, and the generative map is a map
:
m→
n×w that reconstructs the windowed input back to {tilde over (X)} ∈
n×w from the latent features ƒ ∈
. In some implementations, the reconstruction model
compresses X to
and reconstruct {tilde over (X)} from
simultaneously by solving the optimization problem
n is the number of nodes. Latent features are a projection of the dataset to a lower dimensional space. Typically, this also includes an inverse projection to reconstruct the dataset from the latent space. A simple example of latent space is the eigenvectors of a matrix. PCA/f-PCA is another example of a linear projection to latent space. Autoencoder/GAN are examples of nonlinear projections to latent space. Since latent space dimension m<<n×w, any projection that satisfies this constraint will compress the n×w, dataset to m dimensions.
In some implementations, the one-class classifier (or a suitably designed or adapted anomaly detector) is an ensemble of reconstruction models, and each reconstruction model of the ensemble is trained on different operating regimes or boundary conditions of the input dataset. The confidence prediction and other methods to improve the accuracy of the classifier is not limited to one-class classifiers, and can be applied to traditional two-class or multi-class methods as well. In some implementations, the reconstruction is computed using the equation {tilde over (X)}Σj=1p αjj(X).
j are the respective constituent reconstruction models for j=1,2, . . . , p, αj is the corresponding weighting factor, and the vector
for a mono-lithic model would benefit substantially by employing the ensemble technique appropriately). Assets with significant variations include any asset that has very different transient signatures from steady state signatures. There might be further classifications of transients (rising/falling). In some implementations, the operating regimes are determined based on physical characteristics of the industrial assets or using data driven methods. In some implementations, the physical characteristics are used for training separate models for the steady state or different kinds of steady states and transients or different kinds of transients (e.g., fast rising, slow rising, fast falling, slow falling, or in general by separating transients by thresholding the slew rates) in order to ensure reconstruction error for each constituent model remains below a predetermined threshold. In some implementations, the data driven methods computes clusters of reconstruction errors (e.g., computed using different unsupervised techniques like GMM, k_means, DBSCAN) for normal operating conditions and uses the clusters to iteratively partition the input space (i.e., all possible inputs) until all the clusters have reconstruction errors below a predetermined threshold (e.g., a key performance indicator or KPI of the particular system).
In some implementations, computing the confidence level of cyber fault detection (e.g., using the confidence prediction module 106) includes computing model sensitivity of the one-class classifier for the input dataset. In some implementations, the one-class classifier is a reconstruction model that is a nonlinear model. The model sensitivity varies based on operating points, and higher sensitivity regions are more capable than lower sensitivity regions in resolving a smaller difference, thereby making the reconstruction more accurate (as the reconstruction model is a highly nonlinear model, the sensitivity of the model will vary based on its operating point. Assuming a stationary output noise, higher sensitivity regions would be more capable in resolving a smaller difference, thus making the reconstructions more accurate). Higher sensitivity and lower sensitivity are relative terms and may be defined by the KPI of the system. For example, 1% may be small in one application, whereas the same value may be unacceptably large in another depending on the KPI.
In some implementations, computing the confidence level of cyber fault detection (e.g., using the confidence prediction module 106) includes computing model uncertainty of the one-class classifier for the input dataset based on sparsity of training dataset used to train the one-class classifier. Depending on sparsity of training data in certain regions, the accuracy of reconstruction may vary. Based on the training set, the uncertainty may be precomputed and serve as a second indicator of confidence predictor.
In some implementations, computing the confidence level of cyber fault detection (e.g., using the confidence prediction module 106) includes computing statistical distance or L2 distance in an n-space of the input dataset from a training dataset used to train the one-class classifier. For extrapolation, during deployment, the reconstruction model is bound to see data points which falls outside the training boundary. The reconstruction accuracy is expected to be lower in those regions and a suitable metric denoting the statistical distance of the said datapoint from the training boundary will serve as a confidence metric.
In some implementations, the method further includes: designating boundary conditions (e.g., ambient conditions) and/or hardened sensors to compute location of the input dataset with respect to a training dataset used to train the one-class classifier, for computing the confidence level of cyber fault detection using the one-class classifier. In absence of that, all attacks would likely be classified as a sparse region or extrapolation from training set. If most of the attacks are accompanied by lower confidence predictions, they would be evaluated against relaxed thresholds, leading to a lower TPR. As described above, hardened sensors are physically made secure by using additional redundant hardware. The probability that those sensors are attacked is very low. Some implementations determine the confidence metric so as to avoid this undesirable scenario.
In some implementations, the method further includes computing an adaptive decision threshold (e.g., using the decision threshold adjustment module 108) for each node of the plurality of nodes based on a predetermined percentile (e.g., the 99th percentile, or an appropriate percentile value depending on a KPI of the system) of a corresponding residual of the one-class classifier for normal data on the respective node. In some implementations, computing the adaptive decision threshold includes: computing a nominal decision threshold vector
In some implementations, the method further includes generating an alarm (e.g., using the decision threshold comparator module 110 or a separate module for generating alerts) that alerts an operator of the industrial assets based on the detected cyber-faults.
In some implementations, the method further includes transmitting (e.g., using the decision threshold comparator module 110) the detected cyber-faults to a cyber fault neutralization system configured to neutralize the detected cyber-faults in the industrial assets. In some implementations, the method further includes monitoring the industrial assets to determine if the detected cyber-faults persist after a predetermined time period; and in accordance with a determination that the detected cyber-faults persist after the predetermined time period, causing the cyber fault neutralization system to continue to neutralize the detected cyber-faults. The persistence period may be set based on a KPI of the system, and may determine the detection delay (e.g., 15 samples for a gas turbine). In some implementations, the method further includes in accordance with a determination that the detected cyber-faults persist after the predetermined time period, continuing to transmit the detected cyber-faults to a cyber-fault neutralization system, wherein the cyber-fault neutralization system is further configured to playback the transmitted detected cyber-faults and to determine if it is required to continue to neutralize the detected cyber-faults.
The foregoing description, for purpose of explanation, has been described with reference to specific implementations. However, the illustrative discussions above are not intended to be exhaustive or to limit the scope of the claims to the precise forms disclosed. Many modifications and variations are possible in view of the above teachings. The implementations are chosen in order to best explain the principles underlying the claims and their practical applications, to thereby enable others skilled in the art to best use the implementations with various modifications as are suited to the particular uses contemplated.
