An intelligent fault diagnosis system for industrial systems and a method thereof

Abstract

The application discloses an industrial intelligent fault diagnosis system, comprising the following steps: obtaining multivariate time series fragments by collecting sensor data signals; constructing a correlation graph data set for each group of multivariate time series fragments by a KNN algorithm; filtering the correlation graph data set by a soft mask algorithm to obtain a causal estimation graph and a non-causal estimation graph; performing feature extraction on the causal estimation graph and the non-causal estimation graph by a causal decoupling algorithm to obtain a causal correlation graph and a non-causal correlation graph; adjusting the causal correlation graph and the non-causal correlation graph to obtain a causal intervention graph representation; checking causal features in the causal correlation graph to obtain a first fault category, checking non-causal features in the non-causal correlation graph to obtain a second fault category; checking causal intervention features in the causal intervention graph to obtain a third fault category; and outputting a fault category of an industrial system; the application can quickly and accurately diagnose industrial faults and ensure the completion of production tasks.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of industrial big data, PHM and machine learning, and particularly relates to an intelligent fault diagnosis method for an industrial system. BACKGROUND

[0002] Fault diagnosis (FD) technology generally refers to analyzing and judging whether a fault occurs according to the running state of an industrial device by using certain technical means, so as to determine the location, type and cause of the fault, i.e. fault detection and classification, fault positioning and tracing. As an important direction in the field of fault diagnosis, the research on fault detection and classification technology is not only the basis for fault positioning and tracing, but also has important theoretical and engineering significance in the industrial field. It will provide reference information for on-site personnel to accurately detect faults, judge fault categories and quickly propose remedial measures, so as to avoid some disastrous accidents caused by unexpected faults.

[0003] Fault diagnosis plays an important role in exploring the relationship between monitoring data and machine health status, and is an important problem in machine health management (PHM). Traditionally, this relationship is determined by the rich experience and rich professional knowledge of engineers. For example, experienced engineers can diagnose engine faults according to abnormal sound, or use advanced signal processing methods to analyze vibration signals to locate bearing faults. However, in engineering scenarios, users want an automated method to shorten the maintenance cycle and improve diagnostic accuracy. In particular, with the help of artificial intelligence, the fault diagnosis process is expected to be intelligent enough to automatically detect and identify the health status of machines.

[0004] Intelligent fault diagnosis (IFD) refers to applying machine learning theories such as artificial neural networks (ANN), support vector machines (SVM), deep neural networks (DNN) to machine fault diagnosis, which is expected to achieve the above purpose. Instead of using the experience and knowledge of engineers, this method uses machine learning theory to learn the diagnostic knowledge of machines from the collected data. IFD aims to build an end-to-end fault diagnosis model that can automatically judge the health status of machines from the collected data.

[0005] In addition, the emergence of traditional machine learning-based methods such as ANN, SVM, KNN, and other traditional machine learning methods in the prior art has promoted the emergence of IFD, including expert system-based methods [1], artificial neural network-based methods [2], support vector machine-based methods [3], and other intelligent methods [4]. In these methods, fault features are manually extracted from collected data. On this basis, sensitive features are selected to train a diagnostic model that can automatically identify the health status of the machine. With the help of traditional machine learning, the diagnostic model begins to establish a relationship between the selected features and the health status of the machine, thereby reducing human contribution to machine fault diagnosis and pushing it into the era of artificial intelligence. Although IFD based on traditional machine learning can identify the health status of the machine rather than human fault detection, manual feature extraction still relies mainly on human labor. In addition, traditional machine learning theory cannot be applied to growing data due to its low generalization performance, thereby reducing the accuracy of diagnosis. Deep learning is a new direction in the field of machine learning, and since Hinton used the greedy layer-wise pre-training strategy to train deep belief networks [5] (Deep Belief Network, DBN) in 2006, it has received widespread attention. In addition, a series of breakthroughs have been made in Convolutional Neural Networks (CNN), such as AlexNet [6], ResNet [7], etc. These theories have further stimulated the development of IFD and given birth to a series of deep learning-based fault diagnosis methods, such as DBN-based methods, CNN-based methods, and ResNet-based methods.

[0006] In these methods, deep learning helps to automatically learn fault features from collected data rather than manually extracting features in the past IFD. They attempt to provide an end-to-end diagnostic model when dealing with growing data. These models will directly link raw monitoring data to the corresponding machine health status, thereby further releasing human resources in IFD.

[0007] With the advent of intelligent manufacturing in the era of Industry 4.0, modern industrial production processes (such as aerospace, equipment manufacturing, etc.) are developing towards intelligence and integration. In order to reduce maintenance costs and ensure the safe operation of industrial systems, an effective and reliable health monitoring system becomes increasingly important. Fault diagnosis of industrial processes is a classification problem using multivariate time series signals. Due to the complex interaction and close relationship between components, once a fault occurs at a certain location, the readings of multiple components will be abnormal, affecting the operation of the entire process

[13] . In addition, different faults will cause different component readings to be abnormal, and the relationship between components is difficult to discover. In order to perform fault diagnosis, it is necessary to mine the interaction between multiple components and learn the hidden information in time series signals.

[0008] Existing IFD methods mostly focus on signals of multiple independent components in industrial systems, ignoring the interactions between different components. In general, there are complex interactions between components. For example, if a part of the system fails, multiple components related to it will produce abnormal signals. By fusing the data of multiple components, better fault diagnosis can be achieved. In order to mine the complex interactions between components, the graph data of the topology structure can complete this task. Therefore, a structural attributed graph is used to describe the industrial process data, where each component corresponds to a node, and the edges between nodes can be learned according to the similarity of component signals. The component signals are different under different fault modes, and the learned edges are also different, so that different topological graph structures are obtained. By learning such graph structures, hidden fault information can be mined to identify fault types

[14] , which converts the task into a graph recognition problem.

[0009] In order to distinguish the topological graph structures of different fault modes, it is very important to obtain a graph construction method that can represent the signal information of components and the interactions between components. The interactions between components can be explored by constructing a graph from the similarity between components, and the K-neighbor graph

[15] method can be used to obtain the edges connecting the components. A graph classification algorithm is also needed to learn a fault-oriented graph representation while maintaining the specificity of the graph, so as to distinguish the topological graphs of different faults. The graph neural network

[16] algorithm applied in various fields has shown excellent graph classification performance and can be used to learn error-oriented graph representation, which is an effective way to fuse information by information transmission mechanism to obtain information from multiple components. However, due to the complex structure of the graph and the large amount of information, it is very important to find the key part of the input graph and filter out irrelevant parts with the help of the powerful representation learning ability of GNN

[17] . For example, it is very important to find normal signals and abnormal signals in multiple components. Abnormal signals belong to the key part, and different subgraphs are divided into corresponding fault types according to the key part. The widely used attention

[18] and pooling

[19] learning methods adopt the "learning relevant" rule, which mines the hidden mutual information between the focused graph and the true label, leading to the use of shortcut features for decision-making. These shortcut features are mostly caused by the contingency of sample selection or learned environmental noise features, which are non-causal, so they are also called non-causal features.

[0010] However, due to factors such as equipment working conditions and operating environment, the data generated during the operation of mechanical equipment will be different, resulting in that the learned non-causal features can only perform well on the training set and have poor generalization ability, which hinders their practical application in industrial systems. Therefore, it is necessary to avoid learning non-causal features and find essential features that affect fault diagnosis, that is, causal features. A graph neural network strategy based on causal attention is proposed. The strategy promotes the learning of causal features of the input by the attention graph, alleviates the interference of non-causal features, and maximizes their causal influence on the prediction label, thereby solving the above problems. Specifically, first, an attention module is added to the input graph to generate estimates of causal and shortcut features. Then, according to the causal relationship, the backdoor adjustment formula is parameterized, and the causal estimate is combined with the non-causal estimate for stable prediction.

[0011] In summary, the present scheme solves the problems of causal relationship mining and fault identification of industrial systems for the task of industrial process fault diagnosis. The main key technical problems are summarized as follows:

[0012] (1) Causal feature learning based on causal attention

[0013] (2) Fault identification based on graph neural network SUMMARY

[0014] In view of the technical problems existing in the prior art, the present application proposes an industrial intelligent fault diagnosis system and method. The present application adds an attention module to the input graph to generate estimates of causal and shortcut features. Then, according to the causal relationship, the backdoor adjustment formula is parameterized, and the causal estimate is combined with the non-causal estimate for stable prediction.

[0015] Therefore, the present application solves the problems of causal relationship mining and fault identification of industrial systems for the task of industrial process fault diagnosis. The main key technical problems are summarized as follows:

[0016] (1) Causal feature learning based on causal attention

[0017] (2) Fault identification based on graph neural network

[0018] In order to solve the technical problems, the present application adopts the following technical scheme:

[0019] An industrial intelligent fault diagnosis system, the fault diagnosis system comprises a data preprocessing module, a graph neural network, an attention module, a graph convolution module, a first classifier, a second classifier, a backdoor adjustment module, a third classifier and a fault identification module;

[0020] The data preprocessing module obtains multivariate time series fragments by normalizing and segmenting the industrial system sensor data signals collected;

[0021] The graph neural network constructs a relevant graph data set G = {A, X} through a KNN algorithm for each group of multivariate time series fragments;

[0022] Wherein, A records detailed structure information of the graph, and X is a matrix used to describe node features

[0023] The attention module filters the relevant graph data set through a soft mask algorithm to obtain a causal estimation graph and a non-causal estimation graph, respectively;

[0024] The graph convolution module extracts features from the causal estimation graph and the non-causal estimation graph through a causal decoupling algorithm to obtain a causal correlation graph and a non-causal correlation graph, respectively That is,

[0025]

[0026]

[0027] In the formula, G is a class representation of the causal correlation graph G c ; is a class representation of the non-causal correlation graph G t ; GConv c is a feature representation of the causal correlation graph, and GConv t is a feature representation obtained for the non-causal correlation graph;

[0028] The backdoor adjustment module adjusts the causal correlation graph and the non-causal correlation graph through the following formula to obtain a causal intervention graph representation:

[0029]

[0030] In the formula, h Gc is a representation of the causal correlation graph G c ; is a hierarchical representation of the non-causal correlation graph G t′ ;

[0031] The first classifier verifies the causal features in the causal correlation graph through the following formula to obtain a first fault class, that is:

[0032]

[0033] In the formula, L is a cross-entropy loss of the training data , ​is the true label of G;

[0034] The second classifier verifies the non-causal correlation graph in G to obtain the second fault category:

[0035]

[0036] wherein: KL is the Kullback-Leibler divergence, used to measure the similarity of two probability distributions, y unif is uniformly distributed;

[0037] The third classifier verifies the causal intervention graph in G to obtain the third fault category:

[0038]

[0039] wherein: τ' is the estimated set of non-causal correlation graphs, which collects the non-causal features appearing from the training data;

[0040] The fault identification module outputs the fault category of the industrial system by aggregating the first fault category, the second fault category and the third fault category according to the following formula:

[0041]

[0042] wherein: λ1 and λ2 are constants for controlling the degree of decoupling and causal intervention, which are adjustable hyperparameters.

[0043] Further, the attention module filters the correlation graph dataset to obtain the causal estimation graph and the non-causal estimation graph respectively by a soft mask algorithm, including the following steps:

[0044] G = {A, X} represents the industrial sensor correlation graph dataset as the input graph;

[0045] The multi-layer perception MLP node (·) and MLP edge (·) estimate the correlation graph dataset from two angles of nodes and edges respectively; for the node v i and edge(v i ,v j ) respectively, the following can be obtained:

[0046]

[0047]

[0048] wherein: α ci ,β cij are the nodes vi Attention scores at the node level in the causal correlation graph, while edge(v) i ,v j )and For edge-level attention scores of non-causal correlation graphs; and have

[0049] Using attention scores Create a soft mask M node , M edge ,

[0050] According to soft mask M node , M edge , Establish causal estimation graph G c Non-causal estimation graph G t ,Right now:

[0051] G c ={A⊙M edge ,X⊙M node}

[0052]

[0053] This invention also includes the following technical solutions:

[0054] A method for industrial intelligent fault diagnosis includes the following steps:

[0055] Multivariate time series segments are obtained by collecting sensor data signals from industrial systems, normalizing and segmenting them;

[0056] Each set of multivariate time series segments was used to construct a correlation graph dataset using the KNN algorithm;

[0057] The correlation graph dataset is filtered using a soft masking algorithm to obtain causal estimation graphs and non-causal estimation graphs respectively.

[0058] A causal correlation graph is obtained by extracting features from the causal estimation graph and the non-causal estimation graph separately using a causal decoupling algorithm. Non-causal correlation graph Right now:

[0059]

[0060]

[0061] In the formula: It is a causal relationship graph G c Category representation; Causal correlation graph Gt GConv c is a feature representation of the causal correlation graph, GConv t is a feature representation of the non-causal correlation graph;

[0062] The causal intervention graph representation is obtained by adjusting the causal correlation graph and the non-causal correlation graph according to the following formula:

[0063]

[0064] wherein h Gc is a representation of the causal correlation graph G c ; is a hierarchical representation of the non-causal correlation graph G t′ ;

[0065] The first fault category is obtained by checking the causal features in the causal correlation graph G according to the following formula:

[0066]

[0067] wherein: is the cross-entropy loss of the training data , and is the true label of G;

[0068] The second fault category is obtained by checking the non-causal features in the non-causal correlation graph G according to the following formula:

[0069]

[0070] wherein KL is the Kullback-Leibler divergence used to measure the similarity of two probability distributions, y unif is a uniform distribution;

[0071] The third fault category is obtained by checking the causal intervention features in the causal intervention graph G according to the following formula:

[0072]

[0073] wherein τ' is an estimated set of non-causal features collected from the training data;

[0074] The fault categories of the industrial system are output by summarizing the first fault category, the second fault category and the third fault category according to the following formula:

[0075]

[0076] In the formula: λ1 and λ2 are constants for controlling the degree of decoupling and causal intervention, which are adjustable hyperparameters.

[0077] Advantages

[0078] The application converts the fault diagnosis problem of an industrial system into a graph classification problem, and obtains a causal correlation graph and an acyclic correlation graph through an attention mechanism and a causal decoupling method, thereby obtaining causal features and acyclic features.

[0079] In order to obtain a complete graph from the causal correlation graph, the correlation graph is intervened causally through a backdoor adjustment formula to obtain a causal intervention graph.

[0080] The application uses causal features with invariance and stability for prediction, and excludes the interference of acyclic features on prediction. The application is suitable for the fault diagnosis problem of an industrial process, and promotes the fault recognition effect. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 It is a structural causal model of graph recognition in the prior art;

[0082] Figure 2 It is a flow block diagram of an industrial intelligent fault diagnosis method of the application. DETAILED DESCRIPTION

[0083] The application will be described below in conjunction with the drawings:

[0084] The application relates to a graph neural network, which is proposed to process inductive bias related to functional dependence, and therefore uses a non-Euclidean representation. The GNN provides an architecture inspired by deep structures, and provides a suitable method to process these information-rich structures. Compared with a deep neural network, the GNN can process data with complex spatiotemporal relationship features.

[0085] In the real world, a fault of a module in an industrial system may affect the normal operation of other modules and cause the change of the dependence relationship between the metrics. Based on this understanding, the GNN is used for a fault diagnosis task. For example, a method such as a graph convolutional network (GCN) [8], a recurrent graph neural network (RecGNN) [9], a graph attention network (GAT)

[10] , and a spatial temporal graph neural network (STGNN)

[11] .

[0086] Meanwhile, the present application relates to the causal model in the graph neural network, which is according to the process of GNN model, and can build the relationship between variables as a structural causal model

[12] , as shown in the figure. In this figure, the arrow represents the causal relationship, and the model can clearly show the causal relationship between the five variables. The explanation of SCM is as follows: Figure 1

[0087] · C←G→S variable S represents a non-causal feature. It is usually caused by the contingency of sample selection or the learned environmental noise feature. Variable C represents a causal feature. It can reflect the essential properties of graph G, and the causal relationship is established due to the coexistence of non-causal feature S and causal feature C.

[0088] · C→R←S variable R is used to represent the given graph data G. Usually, in order to obtain the graph representation R, the non-causal feature S and the causal feature C can be used at the same time. Then the input of GNN learning strategy is obtained.

[0089] · R→Y. After obtaining the graph representation R, taking it as the basis for learning the classifier, we can use the classifier to classify the input graph, represented by variable Y.

[0090] Through learning the structural causal model of GNN, it is found that there is a backdoor path between C and Y, that is, C←G→S→R→Y. Without a doubt, S is the confounding factor between C and Y. Due to the existence of this backdoor path, C and Y form a false relationship, and the use of non-causal feature S needs to be avoided. In order to make the model classify the graph according to the causal feature C and obtain accurate classification results, it is crucial to block the backdoor path.

[0091] In order to eliminate the influence of non-causal factor S and make the model use causal features for classification, it is necessary to eliminate the backdoor path

[22] . The solution provided by causal theory can be used: do calculation is performed on the causal feature C to obtain P m (Y|C)=P(Y|do(C)) block the backdoor path. The shortcut feature will not change due to the blocking of the backdoor path, so the marginal probability P(S=s) is stable under intervention, that is, P(S)=P m (S). In addition, the causal relationship between C and S is independent of the reaction of Y to C and S, and they are independent of each other, so the conditional probability P(Y|C,s) is also constant, that is, P m (Y|C,s)=P(Y|C,s). Finally, after the causal intervention, the causal feature C is independent of the non-causal feature S, that is, P m (s|C)=P m (s). From the above formula, we can get:

[0092] ​

[0093]

[0094] Through the above analysis, the scheme proposes an industrial process fault diagnosis method, a causal attention graph neural network (CTA-GNN), which considers the relationship between industrial data and diagnosis results, mines causal features that have a decisive effect on diagnosis results, filters mixed features that interfere with diagnosis results, and fuses feature information using GNN. The scheme adds an attention module to the input graph to generate estimates of causal features and non-causal features. Then, according to the causal relationship, the backdoor adjustment formula is parameterized, the causal estimate and the non-causal estimate are combined, and stable fault diagnosis is performed.

[0095] Step 1: The collected industrial process sensor signals are normalized, and then the normalized data is segmented to obtain multivariate time series segments. Each set of multivariate time series segments forms a KNN connected graph, and a graph data set suitable for graph neural networks is constructed.

[0096] Step 2: Construct the CTA-GNN model, which includes three main parts: causal estimation module, causal intervention module and fault identification module. Use the training data set constructed in step 1 to learn the parameters of the CTA-GNN model. The learned model can be used for fault diagnosis of industrial processes.

[0097] The present application proposes a CTA-GNN model considering causal relationship, and converts the fault diagnosis problem of industrial system into a graph classification problem. CTA-GNN obtains causal correlation graph and non-causal correlation graph through attention mechanism and causal decoupling method, so as to obtain causal features and non-causal features. Then, in order to obtain a complete graph from the causal correlation graph, the correlation graph is intervened by the backdoor adjustment formula, and the causal intervention graph is obtained. The present application uses causal features with invariance and stability for prediction, and excludes the interference of non-causal features on prediction. Therefore, the present application is suitable for fault diagnosis problem of industrial process, and promotes fault identification effect.

[0098] Specific process

[0099] The present application proposes a CTA-GNN model for multivariate time series classification applied to industrial process fault diagnosis, and the overall framework is as follows Figure 2The framework is an end-to-end supervised graph neural network model, which generally includes three steps: causal estimation, causal feature extraction, and fault identification. In the first step, the collected industrial process sensor signals are constructed into a KNN connection graph using the KNN algorithm to obtain a graph dataset. The attention module is used to calculate the soft mask to obtain the causal estimation graph, which is the preliminary representation of the correlation graph. In the second step, the causal decoupling is used to obtain the causal correlation graph and the acausal correlation graph, and the causal features and the acausal features are extracted. In the third step, the causal intervention graph is obtained by using the random addition to the correlation graph through the backdoor adjustment formula. Finally, the class of the input graph is predicted by the readout function and the classifier, and the fault is classified.

[0100] First step: data preprocessing

[0101] The extraction of sensor signals of industrial systems is usually obtained by components. Different components are distributed in different positions of the industrial system, so the signals generated form n original measurement variables. In time t, the signal segment generated by the i-th component is However, due to the long running time of the industrial system, the obtained signal segment has a large span, which is usually difficult to process, so multiple signal segments are obtained by window sliding, which can be represented as Since the signal segment is stable in a short time and will not change much, KNN connection graphs are constructed from each multiple signal segment as the input of graph structure modeling, where each signal segment corresponds to a node on the graph. The initial feature of each node is the test value of the sensor sequence segment. Each KNN connection graph supports the training of the CTA-GNN model as a sample.

[0102] Second step: CTA-GNN model construction

[0103] (1) Correlation graph construction

[0104] When the input graph G = {V, E} is obtained, the soft mask is represented as The node feature is represented as Given a soft mask M, its complementary mask can be represented as If the graph is represented in another form G = {a, X}, where a records the detailed structure information of the graph, and X is a matrix used to describe the node features. Then a graph can be divided into two graphs: G1 = {A edge , X node} and After research

[21] , it can be considered that the class of the graph can often be derived from more basic causal features. Therefore, the participation graph that aggregates causal features is defined as the causal correlation graph G c , and the graph that aggregates acausal features is defined as the acausal correlation graph Gt .

[0105] However, in practical applications, the correlation graph with correct values cannot be directly used. Therefore, it is necessary to obtain two types of correlation graphs by learning masks: c = {A 0 M edge , X 0 M node} and In this scheme, the correlation graph representation is obtained by using two steps of soft mask calculation and causal decoupling.

[0106] 1) Soft mask calculation: first, an attention module is used to filter causal and acausal features. Then, according to the obtained features, causal estimation and acausal estimation are generated. Let f(·) represent a GNN-based encoder, and let G = {V, E} represent the input graph, where the nodes are represented as follows:

[0107] H = f(A, X) (2)

[0108] In order to obtain the attention score, two multilayer perceptrons (MLP) MLP node (·) and MLP edge (·) are used from the node and edge perspectives respectively. For node v i and edge(v i ,v j ), the following can be obtained:

[0109]

[0110]

[0111] α ci ,β cij are the node-level attention scores of node v i in the causal correlation graph, and edge(v i ,v j ) and are the edge-level attention scores of the acausal correlation graph. And there are The attention scores are used to construct soft masks M node , M edge ,

[0112] Finally, the initial representations of the causal correlation graph and the acausal correlation graph are obtained using the graphs G c = {A 0 M edge , X 0 M node} and .

[0113] 2) Causal Decoupling: An initial correlation map is created by computing a soft mask to capture causal and non-causal features separately. To obtain the causal and non-causal correlation maps, representations of the correlation maps can be obtained separately using a GNN encoder. Finally, the class of the input map is predicted using a readout function and a classifier.

[0114]

[0115]

[0116] This allows us to obtain the causal correlation graph G. c Category Non-causal correlation graph G t Category Among them GConv c It is a GNN module for obtaining causal correlation graphs, GConv t It is a GNN module for obtaining non-causal correlations. It is a global representation of the causal relationship graph. A global representation of a non-causal correlation graph.

[0117] The purpose of a causal correlation graph is to estimate the causal features that it represents and can classify as true labels. Accordingly, the supervised loss for the graph classification problem is defined as:

[0118]

[0119] in, Training data Cross-entropy loss, This is the true label of G. In contrast, non-causal correlation graphs are designed to approximate non-causal patterns. Therefore, for all known fault classes, non-causal correlation graphs can be predicted, and the unified loss on the graph classification problem is defined as:

[0120]

[0121] Where KL is the KL divergence, y unif To achieve a uniform distribution. By optimizing the above two objectives, causal features can be effectively distinguished from non-causal features. However, a previous study

[22] showed that real-world graph data is noisy, which undoubtedly leads to a greater correlation between the causal part and the label than between the whole graph and the label. Therefore, the proposed decoupling method cannot make the causal correlation graph converge to the complete graph (a noiseless complete graph is a special case), and therefore further optimization is needed.

[0122] (2) Causal Adjustment of Correlation Diagram

[0123] To finally converge the causal correlation graph to the complete graph, the backdoor adjustment is adopted in this scheme. By stratifying the non-causal factors, each layer of the target causal correlation graph is paired with the non-causal correlation graph to form an intervention graph, which can effectively weaken the confounding effect. Due to the irregularity of the graph data, it prevents us from intervening at the data level. Therefore, we intervene implicitly at the representation level and propose a loss guided by the backdoor adjustment:

[0124]

[0125]

[0126] where is the classification result of the implicit intervention graph G' in the classifier Φ. Gc is the representation of the causal correlation graph G c . is the stratified representation of the non-causal correlation graph G t′ obtained by formula (6). τ' is the estimated set of non-causal features collected from the training data.

[0127] In this invention, the process of formula (9) is intervened by random addition. Equation (15) is called causal intervention loss at the graph representation level

[23] . Due to the commonality of causal features, the intervention graph can make stable predictions in different stratifications. Finally, the learning goal of CTA-GNN is given, that is, the total loss:

[0128]

[0129] where λ1 and λ2 are constants that control the degree of decoupling and causal intervention, which are adjustable hyperparameters.

[0130] Summarize the above steps: First, data preprocessing is performed, and the sensor signals collected in the industrial process are constructed into a graph data set to obtain a preliminary graph data set. After that, the attention module is used to calculate the soft mask, and the node and edge representation is given. Then the causal correlation graph and the non-causal correlation graph are obtained through the causal loss function and the non-causal loss function. For the causal correlation graph, the causal intervention graph is obtained through the backdoor adjustment. Finally, the readout function and the classifier are used to predict the category of the input graph, and various faults are classified.

[0131] REFERENCES

[0132] [1] S. Liao, Expert system methodologies and applications-A decade review from to 2004, Expert Syst. Appl. 28 (2005) (1995) 93-103.

[0133] [2] G. M. Knapp, H. P. Wang, Machine fault classification - A neural network approach, Int. J. Prod. Res. 30 (1992) 811-823.

[0134] [3] A. Widodo, B. S. Yang, Support vector machine in machine condition monitoring and fault diagnosis, Mech. Syst. Signal Process. 21 (2007) 2560-2574.

[0135] [4] R. Liu, B. Yang, E. Zio, X. Chen, Artificial intelligence for fault diagnosis of rotating machinery: A review, Mech. Syst. Signal Process. 108 (2018) 33-47.

[0136] [5] G. E. Hinton, R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks, Science 313 (2006) 504-507.

[0137] [6] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Commun. ACM 60 (2017) 84-90.

[0138] [7] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: IEEE Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2016, pp. 770-778.

[0139] [8] Kipf, T. N, Welling, M. Semi-supervised classification with graph convolutional networks. Proceedings of ICLR. 2017.

[0140] [9] You, J, Liu, B, Ying, Z, et al. Graph convolutional policy network for goal-directed molecular graph generation. Proceedings of NeurIPS 6410-6421.

[0141]

[10] L. Ruiz, F. Gama and A. Ribeiro. Gated Graph Recurrent Neural Networks. IEEE Transactions on Signal Processing, vol. 68, pp. 6303-6318, 2020.

[0142]

[11] Xiaoyang Wang, Yao Ma, et al. Traffic Flow Prediction via Spatial Temporal Graph Neural Network. In Proceedings of The Web Conference 2020. Association for Computing Machinery, New York, NY, USA, 1082-1092.

[0143]

[12] Adib, Riddhiman, et al. “A causally formulated hazard ratio estimation through backdoor adjustment on structural causal model.” Machine Learning for Healthcare Conference. PMLR, 2020.

[0144]

[13] L. Ma, J. Dong, K. Peng, and C. Zhang, “Hierarchical monitoring and root-cause diagnosis framework for key performance indicator-related multiple faults in process industries,” IEEE Trans. Ind. Informat., vol. 15, no. 4, pp. 2091-2100, Apr. 2019.

[0145]

[14] K. Zhong, M. Han, T. Qiu, and B. Han, “Fault diagnosis of complex processes using sparse kernel local Fisher discriminant analysis,” IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 5, pp. 1581-1591, May 2020.

[0146]

[15] A. Boutet, A.-M. Kermarrec, N. Mittal and F. Taiani, “Being prepared in a sparse world: The case of KNN graph construction,” 2016 IEEE 32nd International Conference on Data Engineering (ICDE), 2016, pp. 241-252, doi: 10.1109 / ICDE.2016.7498244.

[0147]

[16] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner and G. Mon-fardini, “The Graph Neural Network Model,” in IEEE Transactions on Neural Networks, vol. 20, no. 1, pp. 61-80, Jan. 2009, doi: 10.1109 / TNN.2008.2005605.

[0148]

[17] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang and P. S. Yu, “A Comprehensive Survey on Graph Neural Networks,” in IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 1, pp. 4-24, Jan. 2021, doi: 10.1109 / TNNLS.2020.2978386.

[0149]

[18] Shaked Brody, Uri Alon, and Eran Yahav. 2022. How Attentive are Graph Attention Networks. In ICLR.

[0150]

[19] Muhan Zhang, Zhicheng Cui, Marion Neumann, and Yixin Chen. 2018. An end-to-end deep learning architecture for graph classification. In AAAI.

[0151]

[20] Judea Pearl. 2014. Interpretation and identification of causal mediation. Psychological methods 19, 4 (2014), 459.

[0152]

[21] T. Matsuo, “A reusable model of causal graph in qualitative reasoning,” 2009 IEEE International Conference on Information Reuse & Integration, 2009, pp. 246-251, doi: 10.1109 / IRI.2009.5211559.

[0153]

[22] Wanyu Lin, Hao Lan, and Baochun Li. 2021. Generative causal explanations for graph neural networks. In ICML. 6666–6679.

[0154]

[23] W. Huang et al., “Causal Intervention for Object Detection,” 2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI), 2021, pp. 770-774, doi: 10.1109 / ICTAI52525.2021.00122.

Claims

1. An industrial intelligent fault diagnosis system, characterized in that, The fault diagnosis system comprises a data preprocessing module, a graph neural network, an attention module, a graph convolution module, a first classifier, a second classifier, a backdoor adjustment module, a third classifier and a fault identification module. The data preprocessing module obtains multivariate time series segments through normalization and segmentation processing of collected industrial system sensor data signals; The graph neural network constructs a relevant graph data set G = {A, X} through KNN algorithm for each group of multivariate time series segments; Wherein: A records the detailed structure information of the graph, and X is a matrix used to describe the node features; The attention module filters the relevant graph data set through a soft mask algorithm to obtain a causal estimation graph and a non-causal estimation graph respectively; The graph convolution module respectively extracts features of the causal estimation graph and the acausal estimation graph by a causal decoupling algorithm to obtain a causal correlation graph representation and an acausal correlation graph representation ; that is:​ ； ； wherein: is a causal graph representation is a class representation of is an acausal graph representation is a class representation of is a GNN module to obtain a causal graph, is a GNN module to obtain an acausal graph; The backdoor adjustment module adjusts the causal correlation graph and the non-causal correlation graph through the following formula to obtain a causal intervention graph representation: ； wherein: is a non-causally dependent graph of hierarchical representations; The first classifier uses the following formula to analyze the causal correlation graph. The first fault category is obtained by verifying the causal features, namely: ； In the formula: Training data Cross-entropy loss, This is G's true label; The second classifier checks the non-causal features in the graph by the following formula obtains the second fault category. ； where: KL is the Kullback-Leibler divergence, used to measure the similarity of two probability distributions, is a uniform distribution; The third classifier verifies the causal intervention features by the following formula obtains a third fault category. ； where: is the estimated set of acyclically dependent graphs that collects the acyclically dependent features that occur from the training data; The fault identification module outputs the fault type of the industrial system by aggregating the first fault category, the second fault category and the third fault category through the following formula: ； wherein: and are constants that control the degree of decoupling and causal intervention, and are tunable hyperparameters.

2. The industrial intelligent fault diagnosis system of claim 1, wherein, The attention module filters the relevant graph data set through a soft mask algorithm to obtain a causal estimation graph and a non-causal estimation graph, and the process comprises the following steps: G = {A, X} represents the industrial sensor relevant graph data set as an input graph; Through a multilayer perceptron And Estimation is made from both node and edge perspectives for the relevant graph dataset; for node and edge( , ) respectively, we can get: ； ； wherein: are nodes the node-level attention score and the edge-level attention score in the causal correlation graph; and are nodes the node-level attention score and the edge-level attention score in the non-causal correlation graph; and have ； Utilizing attention scores , , , Soft mask , , , ; According to soft mask Establishing causal estimation maps And non-causal estimation maps That is: ； 。 3. The method of claim 1-2, wherein the system is characterized by, The process comprises the following steps: Obtain multivariate time series segments through normalization and segmentation processing of collected industrial system sensor data signals; Construct a relevant graph data set through KNN algorithm for each group of multivariate time series segments; Filter the relevant graph data set through a soft mask algorithm to obtain a causal estimation graph and a non-causal estimation graph respectively; obtaining causal and acausal graph representations by feature extraction on causal and acausal estimation graphs respectively and acausal graph representations ; i.e. ； ； wherein: is a causal graph representation of the class representation; is an acyclic graph representation of the class representation; is a GNN module to obtain a causal graph, is a GNN module to obtain an acyclic graph; The backdoor adjustment module adjusts the causal correlation graph and the non-causal correlation graph through the following formula to obtain a causal intervention graph representation: ； wherein: is a non-causally dependent graph hierarchical representation; The first fault category is obtained by checking the causal features in the causal correlation graph by the following formula: i.e. ； where: is the cross-entropy loss of the training data is the true label for G;​ The non-causal features in the non-causal correlation graph are verified by the following formula to obtain a second fault category: The non-causal features in the non-causal correlation graph are verified by the following formula to obtain a second fault category: ； where: KL is the Kullback-Leibler divergence, used to measure the similarity of two probability distributions, is a uniform distribution; The causal intervention graph is checked by the following formula The third fault category is obtained by checking the causal intervention feature ； where: is the estimated set of acyclically dependent graphs, which collects the acyclically dependent features that appear from the training data; Output the fault type of the industrial system by aggregating the first fault category, the second fault category and the third fault category through the following formula: ； wherein: and are constants that control the degree of decoupling and causal intervention, and are tunable hyperparameters.

Images