Detecting anomalies in engineering systems by monitoring with multiple sensors
By determining the paired condition probability distribution of sensors in the engineering system and utilizing machine learning models, combining physical interaction and time correlation, the problem of distinguishing exceptions and interference in complex systems is solved, and efficient monitoring of the status of the engineering system and a small number of them are achieved. Model training under the labeled example improves monitoring accuracy and adaptability.
Patent Information
- Application Number
- CN202380040461.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-05-17
- Filing Date
- 2023-05-03
- Publication Date
- 2024-12-27
AI Technical Summary
Complex engineering systems face difficulties in distinguishing abnormalities and interference in sensing monitoring, resulting in the measurement signal including noise and time delay. As the number of sensors increases, individual sensor functions are interfered with, making it difficult to identify abnormalities based on specific rules and ignore the impact. Causes information loss.
By determining its paired conditional probability distribution with other sensors for each sensor, a feature variable tensor is formed, used by machine learning models to identify normal or abnormal behaviors of engineering systems, and to monitor system state using physical interaction and temporal correlation. And train the model by comparing learning and supervising training methods, self-monitoring and a small number of labeled examples.
Effectively identify and classify exceptions, reduce information loss, improve monitoring accuracy of engineering system status, and be able to train models under a small number of labeled examples to adapt to noise and interference in the sensor network, and achieve accurate identification and processing of exceptions.
Smart Images
Figure CN119213380A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to sensory monitoring of engineering systems for identifying known and yet unknown anomalies. Background Art
[0002] Complex engineering systems, such as vehicles or industrial plants, are monitored by a large number of sensors in order to identify operating anomalies. It tends to be difficult to distinguish anomalies of the monitored engineering system from the inevitable disturbances of the sensor monitoring. The measurement signals detected by the sensors therefore often contain noise, and the transmission of the measurement signals via the network is subject to time delays. As the number of sensors increases, the function of individual sensors is increasingly likely to be disturbed. In order to identify anomalies based on specific rules and ignore the mentioned influences, the measurement data must be very strongly smoothed, with a large amount of information being lost. Summary of the invention
[0003] The invention provides a method for identifying anomalies in an engineering system, the behavior of which is monitored by means of an arrangement consisting of n sensors.
[0004] In the context of this method, for each sensor k=1, ..., n, a time series of N observations of this sensor at the time points t=1, ..., N is detected. For each time point t=1,…,N and for the indices i=1,…,n and j=1,…,n, determine the characteristic variable The feature variables represent the pairwise conditional probability distribution P of the observations of sensors i and j D (i|j). Therefore, P D (i|j) describes the probability that sensor i provides a particular specific observation given a situation in which sensor j provides a specific observation. For example, can be determined as the mean of the distribution of observations of sensor i for all time steps at which the observation of sensor j has the same value as in time step t. Since the observed values are real values, no two observations may be exactly the same, so that the "same value" requirement can be softened, for example, to "within an interval of ±ε around the value in time step t". Alternatively or in combination with this, the characteristic variable It can also involve the pairwise conditional probability distribution P D Any other characteristic of (i|j), and for example including P D The descriptive statistics of the characteristic variables of (i|j), such as if P D If (i|j) is a Gaussian distribution, then it is the standard deviation. The distribution P can also be determined or approximated in any other way. D (i|j), and then from this the characteristic variable can be determined
[0005] Therefore, all feature variables The tensor K * It is a “fingerprint” of the behavior of an engineered system that detects the inherent correlations between the various sensor modalities.
[0006] The trained machine learning model will be the “fingerprint” K * Mapped to a classification of whether the behavior of the engineering system is normal or abnormal. The classification can have any form. For example, the classification can be binary or contain one or more real-valued scores related to a specific aspect of the operation.
[0007] It has been recognized that the physical configuration of many engineering systems defines physical interactions between the measured variables detected by different sensors, which in turn produce time correlations between the measured variables. For example, if the flow rate of a fluid medium is recorded at one end of a tubular line, this is correlated with the changing temperature recorded later at the other end of the tubular line. The presence of such a correlation can then be evaluated as a signal whether the engineering system is functioning properly. For example, if a tubular line leaks or tears and the medium flowing through at one end does not reach the other end, but instead flows into the factory floor, the correlation of the changing temperature with the flow rate suddenly disappears.
[0008] It is therefore advantageous to select at least two sensors for measured variables between which the physical configuration of the engineering system in the standard state of the system brings about a physical interaction.
[0009] One measured variable can be, in particular, a measure of the amount of energy delivered to an engineering system or present therein, and another measured variable can be a measure of the amount of energy present or energy output by the engineering system. In the example mentioned with a tubular line, the hot fluid delivered at one end of the tubular line introduces energy into the tubular line, and when the tubular line is heated thereby, the tubular line radiates energy in turn. For example, an increase in the motor current of an electric motor is also manifested in an increase in the vibration amplitude of the electric motor. Since energy is a physically conserved quantity, in many engineering systems there is a time correlation between the measured values provided by different sensors. The correlation does not have to be explicitly analyzed and expressed in order to monitor the normal functioning of the engineering system. In normal operation, it is sufficient for the correlation to simply exist physically so that the correlation can be learned by a machine learning model.
[0010] In a further advantageous embodiment, at least two further sensors are selected for further measured variables, between which the physical configuration of the engineering system excludes physical interactions in the standard state of the system. Thus, for example, the pressures in areas or containers that are not connected to one another should not correlate with one another. However, if such a correlation exists, this can indicate that an undesired leak exists between the areas or containers.
[0011] Thus, in particular, for example, further sensors can be arranged on different sides of a barrier which, in the standard state of the engineering system, prohibits a physical interaction between further measured variables.
[0012] In particular, for example, a machine learning module can learn the normal behavior of an engineering system and then classify as an anomaly everything that is "different to some extent", that is to say even if the anomaly is present only in a small signal share of one or more measurement signals detected by means of the sensor. This is somewhat similar to the fact that a burglar alarm does not trigger an alarm if the legal occupant of the house hammers a nail into the wall, whereas it simultaneously recognizes the operating noise caused by manual operation of the door lock by means of an unlocking tool as an attempted break-in and triggers the alarm.
[0013] For example, the engineering system can be, in particular, a vehicle or an industrial facility that processes one or more reactants into one or more products through one or more processing steps. In vehicles and industrial facilities, a large number of measured values are currently detected by sensing, and possible anomalies can be better evaluated with the help of the method proposed here. In particular, vehicles facilitate a large number of correlations that can be used to identify anomalies via their body, wherein sensors are distributed in a relatively narrow space in or at the body. In industrial facilities, such correlations are facilitated, for example, by the material flow of reactants and / or products through the facility. In addition, control devices in vehicles and "plant historians" in industrial facilities automatically record a large number of measured values. For example, the records can be used to obtain labeled training data of normal and abnormal states in combination with knowledge of the presence or absence of anomalies at a specific time. With the help of the labeled training data, machine learning models can be trained in a supervised manner. In particular, for example, the paired conditional probability distribution P can be determined and / or approximated with the help of training data. D (i|j). How far back in time a machine learning model can look is determined by the chosen architectural parameters, such as the width of the filter kernels in convolutional layers.
[0014] In another advantageous embodiment, the classification provided by the machine learning model forms a control signal. The engineering system is controlled with the aid of the control signal. For example, in response to an abnormality being detected in the vehicle, the maximum speed can be reduced, a certain dangerous driving maneuver (e.g., an overtaking maneuver) can be prohibited, or the vehicle can be stopped on a pre-planned emergency stop trajectory. For example, an industrial facility can be switched to a safety mode in which the throughput of the facility is reduced to a minimum. In the safety mode, the facility can also be manually started, for example in the event of a failure of the automatic control.
[0015] As mentioned before, we can use the tensor K * However, specifically labeled training examples of anomalies are usually rare because most engineering systems work well so that anomalies are rare.
[0016] Therefore, the present invention provides a method for training a machine learning model for use in the method described above, which method is sufficient with only a small number of labeled training examples.
[0017] In the context of the method, a method for providing a characteristic variable The tensor K * training examples, the feature variables characterize the pairwise conditional probability distribution P of the observations of sensors i and j D (i|j). It is still not necessary to know whether the training examples involve normal or abnormal behavior of the engineering system.
[0018] A distribution of disturbances p is provided, for which it is known that they change the tensor K only in such a way that this can also occur in normal operation of the engineering system. * .
[0019] An example of such a disturbance p is additive normally distributed noise, which can indeed also be contained in the observations provided by the sensors in each case.
[0020] Another example is the tensor K * The off-diagonal elements of are randomly reset to zero, passing through the elements at level t+1 or t-1 Replace the element at level t and / or tensor K * The interference simulates the errors that may occur when sensing and transmitting measurement data via the network. Therefore, the elements on layer t and the elements at level t+1 or t-1 Mixing can be caused, for example, by inaccurate time synchronization between the sensors. For example, poor time synchronization or problems in the transmission of measurement data via the network can lead to the exchange of layers t, t'. Sporadic sensor errors can, for example, manifest themselves in the tensor K * The off-diagonal elements are missing.
[0021] The perturbations p allow so-called positive-negative pairs to be created for the contrastive learning process. Contrastive learning is self-supervised learning based on positive and negative examples that are known to be similar or dissimilar to each other.
[0022] By applying two perturbations p1 and p2 sampled from the distribution K to the same training example * To produce a variant and Opposite Therefore, the variation in the pair and Only disturbances are different, such as those that also occur during normal operation of the engineering system and sensor monitoring. Therefore, the machine learning model should take these variations into account. and Identified as being similar to each other.
[0023] By applying two perturbations p1 and p2 sampled from the distribution to two different training examples K *′ and K *″ To produce a variant and Negative pair Therefore, the variation in the pair and Not only is the usual interference different, but it also involves two completely different examples of variation. Therefore, the machine learning model should consider the variation and Identified as being dissimilar to each other.
[0024] In order to accurately train a machine learning model for this, the machine learning model to be trained is transformed from and and and The processed products are produced and This need not be the corresponding final output of the machine learning model that can identify the classification in terms of normal or abnormal state. Instead, the machine learning model may include, for example, *A feature extractor extracts features and a classification head maps the features onto the sought classification of the engineering system behavior. The feature extractor may, for example, specifically comprise a sequence of a plurality of convolutional layers which convert their respective inputs into feature maps of reduced dimension by applying one or more filter kernels. The processing product may then be formed by the feature extractor and
[0025] In the context of contrastive learning, we now optimize the parameters that characterize the behavior of a machine learning model with the goal of:
[0026] ·Maximization involves facing The processed products and similarity, and
[0027] Minimize the number of negative pairs involved The processed products and similarity.
[0028] That is, the feature extractor of the machine learning model is trained to attribute the positive pairs The processed products and At the same time, the feature extractor is trained to: in the latent space, the features attributed to the negative pairs The processed products and Therefore, the processing products can correspond to a point z in the latent space i , and the same positive treatment product can correspond to a point z in the latent space j . Measurement point z i and z j An exemplary cost function for whether or not two are close to each other is NT-Xent-Loss ("Normalized Temperature Scaled Cross Entropy"):
[0029]
[0030] where the function Ψ measures the similarity and τ is a temperature parameter. The temperature parameter τ can be varied, for example, in particular according to an “annealing schedule” as a function of the epoch number.
[0031] Therefore, a machine learning model can complete most of its training in this way self-supervised, without having to use labeled training examples for this.
[0032] For example, one can then use K training examples labeled with the target class *The classification head is trained in a supervised manner. Since the classification head only accounts for a small part of the machine learning model, in particular with regard to the number of parameters to be optimized, a relatively small number of labeled training examples is sufficient for the training. In this context, it is also particularly advantageous that the classification head receives as input processing products that have already been well preclassified as a result of self-supervised training. As a result, the classification head does not have to invest, for example, in increased training effort to correct what was previously missed.
[0033] The present invention also provides another method for training a machine learning model for use in the method described at the beginning. Unlike the contrastive learning described above, the method uses normal supervised training.
[0034] In the context of the method, a method for providing a characteristic variable The tensor K * training examples, the feature variables characterize the pairwise conditional probability distribution P of the observations of sensors i and j D (i|j). The training example K * Concerning the normal behavior of engineering systems.
[0035] Now, providing interference p * distribution of disturbances for which it is known that they change the tensor K only in a way that would not be expected in normal operation of the engineering system. * An example of such interference is uniformly distributed noise.
[0036] By sampling the disturbance p from the distribution * Applied to training examples K * To produce a variant On the one hand, we use the training examples K of the normal behavior of the engineering system * and on the other hand by using variations as training examples of abnormal behavior of engineering systems Train machine learning models in a supervised manner.
[0037] By means of the method described, as long as the disturbance p * If the appropriate distribution of is available for the corresponding application, the machine learning model can be trained directly on the desired classification task in one step. * If such a distribution is not available or difficult to model, the above scheme on contrastive learning can be used. An important advantage of contrastive learning is that it does not rely on the interference p * Modeling.
[0038] Optionally, within the scope of the method, the training examples K can be transformed into * Enhancements are made to further variants which can then be used as further training examples of normal behavior within the scope of supervised learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Further measures for improving the invention are shown in more detail below together with the description of preferred exemplary embodiments of the invention with reference to the drawings.
[0040] Example
[0041] in:
[0042] Figure 1 An embodiment of a method 100 for identifying an anomaly in an engineering system 1 is shown;
[0043] Figure 2 Show feature variables The tensor K * Description;
[0044] Figure 3 An embodiment of a method 200 for training a machine learning model 2 is shown;
[0045] Figure 4 An embodiment of a method 300 for training a machine learning model 2 is shown. DETAILED DESCRIPTION
[0046] Figure 1 is a schematic flow chart of an exemplary embodiment of a method 100 for detecting an anomaly in an engineering system 1. The behavior of the system 1 is monitored by means of an arrangement of n sensors.
[0047] In step 110, for each sensor k=1, ..., n, a time series of N observations of the sensor at time points t=1, ..., N is detected.
[0048] Here, according to block 111, at least two sensors can be selected for measured variables, between which the physical configuration of the engineering system 1 in the standard state of the system 1 enables a physical interaction. Then, according to block 111a, at least two further sensors can also be optionally selected for further measured variables, between which the physical configuration of the engineering system 1 in the standard state of the system 1 excludes a physical interaction.
[0049] In step 120, for each time point t=1, ..., N and for the indices i=1, ..., n and j=1, ..., n, the characteristic variables are determined (120) The feature variables represent the pairwise conditional probability distribution P of the observations of sensors i and j D (i|j).
[0050] In step 130, all feature variables The tensor K *Mapped from the trained machine learning model 2 to the classification 3 of whether the behavior of the engineering system 1 is normal or abnormal
[0051] In step 140 , the classification 3 provided by the machine learning model 2 forms the control signal 4 .
[0052] In step 150 , the engineering system 1 is controlled using the control signal 4 .
[0053] Figure 2 The tensor K is illustrated by a simple example with three time-dependent observations x(t), y(t), and z(t). * For each time point t, the conditional probability of the values of all variables x(t), y(t) and z(t) can be described based on the existence of one of the three values x(t), y(t) and z(t). The characteristic variable characterizing the pairwise conditional probability distribution is is compiled into a two-dimensional matrix. All such matrices at all time points t together form the tensor K * .
[0054] Figure 3 is used to train machine learning models 2 to combine Figure 1 A schematic flow chart of an embodiment of a method 200 used in the described method 100. The method 200 is based on contrastive learning.
[0055] In step 210, a feature variable is provided The tensor K * training examples, the feature variables characterize the pairwise conditional probability distribution P of the observations of sensors i and j D (i|j).
[0056] In step 220 , the distribution of disturbances p is provided, for which it is known that they change the tensor K only in such a way that this can also occur in normal operation of the engineering system. * .
[0057] In step 230, two perturbations p1 and p2 sampled from the distribution are applied to the same training example K * To produce a variant and Opposite The variant thus differs only by disturbances p which also occur in the normal behavior of the engineering system 1 and in the sensory data acquisition.
[0058] In step 240, two perturbations p1 and p2 sampled from the distribution are applied to two different training examples K *′ and K *″ To produce a variant and Negative pair The variants thus differ by more than just a "normal" disturbance p.
[0059] In step 250, the machine learning model 2 to be trained is used to transform and The processed products are produced and
[0060] According to box 251, the machine learning model 2 may include a feature extractor 21 and a classification head 22, wherein the feature extractor extracts the tensor K from the * Features are extracted and the classification head maps the features to the sought classification of the behavior of the engineering system.
[0061] In step 260, the parameters 2a characterizing the behavior of the machine learning model 2 are optimized in order to and The similarity between them is the smallest while and The similarity between them is the greatest. The parameters may in particular be parameters 21a of feature extractor 21. The optimized state of parameters 2a, 21a is denoted by reference numerals 2a*, 21a*.
[0062] In step 270, with the help of training examples K labeled with target classification * The classification head 22 is trained in a supervised manner. That is, the parameters 22a characterizing the behavior of the classification head 22 are optimized so that the training examples K * The optimized state of the parameters is indicated by reference numeral 22a*.
[0063] Figure 4 is used to train machine learning models 2 to combine Figure 1 A schematic flow chart of an embodiment of a method 300 used in the described method 100. Unlike the method 200, the method is based on supervised training.
[0064] In step 310, a feature variable is provided The tensor K * training examples, the feature variables characterize the pairwise conditional probability distribution P of the observations of sensors i and j D (i|j). The training example K * It concerns the normal behavior of the engineering system 1 .
[0065] In step 320, a disturbance p is provided *distribution of disturbances for which it is known that they change the tensor K only in a way that would not be expected in normal operation of the engineering system. * .
[0066] In step 330, the disturbance p sampled from the distribution is * Applied to training examples K * To produce a variant
[0067] In step 340 , on the one hand, with the aid of training examples K of normal behavior of the engineering system 1 * and on the other hand by means of variations as training examples of abnormal behavior of the engineering system 1 The machine learning model 2 is trained in a supervised manner. That is, the parameters 2a that characterize the behavior of the machine learning model 2 are directly optimized, with the goal of converting the training examples K * Map to normal classification and transform Mapped to anomaly classification.
Claims
1. A method (100) for detecting anomalies in an engineering system (1), the behavior of which is monitored by means of a device consisting of n sensors, the method comprising the following steps: For each sensor k=1, ..., n, detect (110) a time series of N observations of said sensor at time points t=1, ..., N For each time point t=1, ..., N and for the indices i=1, ..., n and j=1, ..., n, determine (120) the characteristic variables The feature variable represents the pairwise conditional probability distribution P of the observations of the sensors i and j D (i|j); All feature variables The tensor K * The trained machine learning model (2) is mapped (130) to a classification (3) as to whether the behavior of the engineering system (1) is normal or abnormal.
2. The method (100) according to claim 1, wherein at least two sensors are selected (111) for measured variables, and the physical configuration of the engineering system (1) causes a physical interaction between the measured variables in a standard state of the system (1).
3. A method (100) according to claim 2, wherein one measured variable is a measure of the amount of energy supplied to the engineering system or present therein, and wherein another measured variable is a measure of the amount of energy present or energy output by the engineering system.
4. A method (100) according to any one of claims 2 to 3, wherein at least two additional sensors are selected (111a) for additional measurement variables, and the physical configuration of the engineering system (1) excludes physical interactions between the additional measurement variables in a standard state of the system (1).
5. The method (100) according to claim 4, wherein the further sensors are arranged on different sides of a barrier which, in a standard state of the engineering system (1), prohibits a physical interaction between the further measured variables.
6. The method (100) according to any one of claims 1 to 5, wherein a vehicle or an industrial facility for processing one or more reactants into one or more products by one or more processing steps is selected as the engineering system (1).
7. The method (100) according to any one of claims 1 to 6, wherein forming (140) a manipulation signal (4) from the classification (3) provided by the machine learning model (2), and The engineering system (1) is controlled (150) by means of the control signal (4).
8. A method (200) for training a machine learning model (2) for use in a method (100) according to any one of claims 1 to 7, the method (200) comprising the following steps: Provide (210) for feature variables The tensor K * training examples, the feature variables represent the pairwise conditional probability distribution P of the observations of the sensors i and j D (i|j); Providing (220) a distribution of disturbances p, for which it is known that they change the tensor K only in such a way that this can also occur in normal operation of the engineering system * ; By applying two perturbations p1 and p2 sampled from the distribution to the same training example K * To produce (230) variant and Opposite By applying two perturbations p1 and p2 sampled from the distribution to two different training examples K * ′ and K * ″ to produce (240) variant and Negative pair · From the variants by means of the machine learning model to be trained (2) and The (250) treatment products are produced respectively and Optimize (260) the parameters (2a) characterizing the behavior of the machine learning model (2) with the goal of oMaximization involves the The processed products and similarity, and oMinimize the negative pair The processed products and similarity.
9. The method (200) of claim 8, wherein The machine learning model (2) includes (251) a feature extractor (21) and a classification head (22), wherein the feature extractor extracts the information from the tensor K * extracting features from the classification head, the classification head mapping the features to the sought classification of the behavior of the engineering system, and The processing product is formed (252) by the feature extractor (21) and 10. The method (200) according to claim 9, wherein the training examples K labeled with the target classification are used. * The classification head (22) is trained (270) in a supervised manner.
11. The method (200) according to any one of claims 8 to 10, wherein the disturbance p comprises Additive normally distributed noise, and / or The tensor K * Randomly zero the off-diagonal elements of , and / or · Through the elements at level t+1 or t-1 Replace the element at level t and / or The tensor K * Exchange of layers t and t'.
12. A method (300) for training a machine learning model (2) for use in a method (100) according to any one of claims 1 to 7, the method (300) comprising the following steps: Provide (310) for feature variables The tensor K * training examples, the feature variables represent the pairwise conditional probability distribution P of the observations of the sensors i and j D (i|j), where the training example K * involving the normal behavior of the engineering system (1); Provide (320) interference p * distribution of the disturbance, for which it is known that the disturbance changes the tensor K only in a way that would not be expected in normal operation of the engineering system. * ; By taking the disturbance p sampled from the distribution * Applied to training examples K * To produce (330) variant On the one hand, with the help of training examples K of normal behavior of the engineering system (1) * and on the other hand by means of variations as training examples of abnormal behavior of the engineering system (1) The machine learning model (2) is trained (340) in a supervised manner.
13. A computer program comprising machine-readable instructions which, when executed on one or more computers and / or computing instances, cause the one or more computers or computing instances to perform a method (100) according to any one of claims 1 to 12.
14. A machine-readable data carrier and / or a download product having a computer program according to claim 13.
15. One or more computers and / or computing instances having a computer program according to claim 13 and / or having a machine-readable data carrier and / or a download product according to claim 14.