An engineering machinery intelligent fault diagnosis method based on self-learning graph convolution network

Abstract

The hydraulic pump abnormality detection method relates to an engineering machinery intelligent fault diagnosis method based on a self-learning graph convolution network, which is proposed on the basis of analyzing a recurrent graph convolutional network algorithm (RGCN). The working condition type is the working condition state of the hydraulic pump under the single action of the walking and rotating of the excavator and the compound action of the boom, stick and bucket linkage. On this basis, the adjacency matrix in the RGCN is optimized, the model hyperparameters are adjusted, the correlation of multi-source time series data in time and space is analyzed in combination with the LSTM, the feature information is encoded by using the VAE, and a RGCN-LSTM-VAE algorithm fusion model is constructed. The model is used for hydraulic pump abnormality detection, and the results show that the RGCN-LSTM-VAE algorithm fusion model based on multi-element time series data can achieve good abnormality detection effect in hydraulic pump abnormality detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the abnormal detection of hydraulic pumps in engineering practice, and proposes a self-learning graph convolutional network hydraulic pump abnormal detection method combining a recurrent graph convolutional network algorithm (RGCN), a long short-term memory network (LSTM), and a variational auto-encoder (VAE). BACKGROUND

[0002] As one of the important power elements of mechanical equipment, a hydraulic pump is an important component of a hydraulic system, and mainly functions as a conversion device for converting mechanical energy input by a prime mover into hydraulic energy. The hydraulic pump is widely used in modern industrial equipment such as mining metallurgy, national defense construction, power systems, agricultural machinery, chemical industry and petroleum, and shipbuilding. However, due to the complex structure of the hydraulic pump and the harsh working conditions, the hydraulic pump is prone to failure. According to statistics, in the hydraulic system of mechanical equipment, the hydraulic pump is the most prone to failure, with a damage rate of up to 95%. Therefore, it is of great practical significance and application value to study the hydraulic pump abnormal detection method, conduct state monitoring of the hydraulic pump, discover early faults of the hydraulic pump in advance, and take timely operation and maintenance measures to reduce personnel casualties, reduce economic losses, and ensure the normal operation of mechanical equipment.

[0003] With the continuous optimization and upgrading of the manufacturing industry, the production level continues to improve, and mechanical equipment is becoming increasingly large-scale, integrated, automated, and intelligent, gradually realizing a high-precision and high-efficiency working mode. However, this also leads to more and more factors affecting the normal working state of mechanical equipment. Therefore, the application of traditional fault diagnosis technology in intelligent equipment has certain limitations, and intelligent abnormal detection and fault diagnosis technology has emerged as the times require. In view of the faults of the hydraulic pump, domestic and foreign researchers have carried out a large number of in-depth researches, and have formed many abnormal detection technologies and fault diagnosis methods. For example, common abnormal detection methods based on statistics, abnormal detection methods based on pattern prediction, abnormal detection methods based on Bayesian inference, fault diagnosis methods based on time-frequency analysis, fault diagnosis methods based on clustering analysis, and fault diagnosis methods based on deep learning. In engineering practice, due to the complexity of the working conditions and the harshness of the working environment, it is necessary to improve and optimize the fault diagnosis method from multiple dimensions to improve the correctness and rapidity of early fault detection. SUMMARY

[0004] To solve the above problems, the hydraulic pump abnormality detection method of the self-learning graph convolution network is proposed to improve the correctness and rapidity of early fault detection. The method mainly embodies the diversification of working conditions and variables, wherein the diversification of working conditions is embodied in mining multiple working condition types under the single action of the excavator executing walking and rotating and the composite action of the boom, stick and bucket linkage, and the diversification of variables is embodied in that the extracted variable signals include power source signals such as hydraulic pump pressure, engine speed and torque percentage, and pressure signals of execution components such as the boom, stick and bucket. Compared with the traditional hydraulic pump abnormality method, the proposed hydraulic pump abnormality detection method based on the recursive graph neural network mainly realizes the detection of the working state of the hydraulic pump from multiple dimensions. By fusing RGCN, LSTM and VAE multiple machine learning algorithms, the interference of a large number of human factors and external environment is reduced, and the efficiency and accuracy of abnormality detection are improved.

[0005] Mainly includes: based on the feature information of the collected multi-element time series data, the improvement of the RGCN algorithm rule, enhances the adaptability of RGCN, combines with LSTM to analyze the correlation of multi-element time series data in time and space, and through VAE, the extracted related features are encoded and decoded to reconstruct, finally applied to the abnormality detection of hydraulic pump, in order to obtain ideal abnormality detection effect. This is not only the research of theoretical method, but also the landing transformation of theoretical research results in engineering practice, which solves the practical problems encountered in practical application. This has very important significance for promoting the transformation of theoretical research into actual productivity and embodying the practical application value of theoretical research results.

[0006] The application adopts the following technical scheme:

[0007] An engineering machinery intelligent fault diagnosis method based on a self-learning graph convolution network comprises the following steps:

[0008] Step 1: signal data under actual working conditions of the hydraulic pump are collected as input sequence X;

[0009] Step 2: the working conditions are divided according to the working condition types and the collected multi-element time series data; single working condition sequence x a and composite working condition sequence x b are divided, and training set D train and test set D test are divided;

[0010] Step 3: the spatial information of the data is encoded by using the space-based RGCN; x a and x b are input into two RGCNs with different adjacency matrices but same node relationship and same weight corresponding to linear transformation, and the adaptive classifier d(x) is used to distinguish the spatial features and

[0011] Step 4: Spatial feature generated based on RGCN And Further encode the time feature information with LSTM to generate the time feature f t a And f t b , and combine them into a feature set F;

[0012] Step 5: Hypothesis feature set The prior of z is subject to a multivariate Gaussian distribution, i.e. p θ (z)=N(z;0,I); The approximate posterior is a multivariate Gaussian distribution, i.e. q φ (z|x a )=N(z;μ,σ 2 I), then z~q φ (z|x a ) is sampled using the reparameterization trick;

[0013] Step 6: Signal reconstruction, reconstruct p θ (x R |z) composed of LSTM and RGCN, the reconstructed sequence is represented as x R , and the corresponding loss function is calculated;

[0014] Step 7: Calculate the reconstruction error, if the reconstruction error of the timestamp t is greater than the defined threshold, mark the data point as abnormal.

[0015] Further, the signal data collected in step 1 mainly includes the data of main pump pressure signal, main pump flow control pressure, main pump flow control current, main pump pressure sensor voltage, boom lifting pilot pressure, boom lowering pilot pressure, arm swing pilot pressure, arm retracting pilot pressure.

[0016] Further, the specific steps of step 2 are:

[0017] Step 2.1: Divide the working condition sequence, divide the data under the single working condition and the composite working condition of the hydraulic pump to generate the single working condition sequence x a and the composite working condition sequence x b ;

[0018] Step 2.2: Divide the data training set and test set , wherein is the data recorded by N V sensors / actuators at timestamp t; Assign label y t ∈{0,1} to the detection result, 1 represents x tis an outlier data point, otherwise 0;

[0019] Step 2.3: Data preprocessing, extract M multivariate sequences from training dataset D train with window size ω and stride s, each multivariate sequence X n is defined as X n = [x (n-1)·s , x (n-1)·s+1 ,..., x (n-1)·s+ω-1 ], test data D test is processed in the same way, obtaining M multivariate sequences

[0020] Step 2.4: Calculate the adjacency matrix, given a graph G = (V, E), where V and E represent vertices / nodes and edges respectively, the structure information is represented by the adjacency matrix , where ω(i, j) is the weight between nodes i and j, N V is the number of nodes;

[0021] Step 2.5: Calculate the activation function, RGCN updates the graph through the same graph convolution layer, given the node features H l of the current l-th layer, the activation calculation of the next layer is where σ is the sigmoid activation function represented as W l is the weight to be learned, where I is an identity matrix of size N V , is the degree matrix of ;

[0022] Step 2.6: Optimize the adjacency matrix, adopt an adaptive way to automatically learn the graph structure, that is, the adjacency matrix is initialized as where σ is the sigmoid operation, W A' is the weight to learn the adjacency matrix, and is randomly initialized from a uniform distribution U(0, 1), A' is both positive semi-definite and symmetric, and each element A' ij ∈(0, 1);

[0023] Step 2.7: Add constraints to A', add sparsity constraint L L1-norm and uncertainty constraint L entropy , Apply sparsity constraint (L L1-norm ) on the adjacency matrix to make A' sparse and keep contribution relevance; for uncertainty constraint L entropy ​​, quantifies and reduces the uncertainty of the predicted probability by entropy loss.

[0024] Further, the specific steps of step 3 are:

[0025] Step 3.1: encode the spatial information of x a and x b into spatial features and

[0026] Step 3.2: distinguish spatial features and

[0027] Step 3.3: calculate the d(x) loss function, where L y is the cross-entropy loss function, and is the generated feature of the nth sequence, whose class label is 0 or 1.

[0028] Further, the specific steps of step 6 are:

[0029] Step 6.1: signal reconstruction, reconstruct p θ (x R |z) composed of LSTM and RGCN, and the reconstructed sequence is represented as x R .

[0030] Step 6.2: calculate the loss function, the loss function of the variational autoencoder is L VAE = L rec + L KLD , where L rec is the reconstruction loss;

[0031] Step 6.3: calculate the reconstruction loss, where MSE represents the mean square error; L rec takes into account the reconstruction loss of RGCN input x a and LSTM input ;

[0032] Step 6.4: calculate the KL divergence, where N z is the dimension of the latent variable z;

[0033] Step 6.5: calculate the overall loss function, L = λ1(L rec + L KLD ) + λ2L discr + λ3L entr o py + L L1-normwherein λ1, λ2 and λ3 are hyperparameters for balancing these loss terms; set L rec and the factor of L KLD to be the same, and set the factor of L L1-norm to be 1;

[0034] Step 6.6: Hyperparameter adjustment, adjust the hyperparameters λ1, λ2 and λ3 with dynamic weighted loss scheme.

[0035] Further, the specific steps of step 7 are:

[0036] Step 7.1: Calculate the reconstruction error, define the reconstruction error or anomaly score of each test data point as wherein and are the corresponding data points in x t ; The reconstruction error of x t in all multivariate data containing x t is accumulated and normalized;

[0037] Step 7.2: If the reconstruction error of the timestamp t is greater than the defined threshold, mark the data point as abnormal; otherwise, it is a normal data point.

[0038] Advantages of the present application:

[0039] 1. The present application proposes a hydraulic pump anomaly detection method based on the fusion of multivariate time series data RGCN-LSTM-VAE algorithm, in order to improve the correctness and rapidity of early fault detection. Compared with the traditional hydraulic pump anomaly method, the proposed multivariate time series data anomaly detection method mainly realizes the detection of the working state of the hydraulic pump from multiple dimensions. By fusing RGCN, LSTM and VAE machine learning algorithms, a large number of human factors and external environmental disturbances are reduced. Through the improvement of the rules of RGCN algorithm, the adaptability of RGCN is enhanced, the correlation of multivariate time series data in time and space is analyzed combined with LSTM, and the relevant features extracted are encoded and decoded for reconstruction by VAE, which is applied to the anomaly detection of multivariate time series data of hydraulic pump, so as to achieve ideal anomaly detection effect.

[0040] 2. The present application proposes an engineering machinery intelligent fault diagnosis method based on self-learning graph convolution network, and based on the actual working condition data, it fuses RGCN, LSTM and VAE three algorithms, which has high precision, fast speed and easy modeling, can greatly improve the research efficiency and avoid a lot of repetitive labor, and lays a foundation for the health maintenance of equipment and the early fault diagnosis research of other key components. BRIEF DESCRIPTION OF DRAWINGS ​

[0041] Figure 1 The flowchart of the RGCN-LSTM-VAE algorithm fusion on which the present application is based. DETAILED DESCRIPTION

[0042] The present application will be further described in detail below in conjunction with the drawings and technical solutions, and the embodiments of the present application will be described in detail through preferred embodiments, but the embodiments of the present application are not limited thereto.

[0043] A self-learning graph convolution network-based intelligent fault diagnosis method for construction machinery, in this embodiment, the hydraulic pump of the excavator is taken as an example, and the type is a plunger pump. The working principle is as follows: the cylinder contains a plurality of plungers and is uniformly distributed, the sliding shoe on the plunger is pressed on the swash plate of the cylinder, and the swash plate and the surface of the cylinder present a certain angle. When the cylinder and the plunger rotate together, the plunger rotates to the farthest distance between the swash plate and the cylinder, and the volume between the plunger and the cylinder hole will be evacuated. When the distance between the swash plate and the cylinder is the closest, the volume between the plunger and the cylinder hole is the smallest. As long as the oil suction port and the oil outlet port of the distribution disc are configured to the farthest and closest positions of the swash plate from the cylinder, as long as the main pump rotates, the plunger performs such reciprocating action between the holes to realize oil suction and discharge. Moreover, the hydraulic pump is the power source of the entire excavator, and supports the actions of the vehicle walking, rotating, boom and arm, etc. Therefore, the signals related to the hydraulic pump can be directly or indirectly collected from multiple angles, and the abnormal detection research of the hydraulic pump based on multi-element time series data can be carried out.

[0044] The data used in the present application are all multi-element time series data collected from vehicles under actual working conditions. For non-specific working conditions of the hydraulic pump, the data of more than ten kinds of signals related to the hydraulic pump, such as pressure signals, sensor voltage signals and control current signals, are directly or indirectly collected for abnormal detection of the hydraulic pump.

[0045] In the present application, the self-learning graph convolution network-based hydraulic pump abnormal detection method is used to perform abnormal detection analysis on the more than ten kinds of signal data collected from the hydraulic pump under actual working conditions. The experimental results show that the self-learning graph convolution network-based abnormal detection method can achieve ideal diagnosis effect in the abnormal detection of the hydraulic pump.

[0046] Referring to Figure 1 A self-learning graph convolution network-based intelligent fault diagnosis method for construction machinery, which specifically comprises the following steps:

[0047] Step 1: The signal data of the excavator hydraulic pump under actual working conditions is collected, and the variable signals mainly include ten kinds of signal data such as main pump pressure signal, main pump flow control pressure, main pump flow control current, main pump pressure sensor voltage, boom lifting pilot pressure, boom lowering pilot pressure, arm swing pilot pressure, arm retracting pilot pressure, etc., as multivariate time series X.

[0048] Step 2: According to the working condition type and the collected multivariate time series data, the sequence x under single working conditions such as single walking and single rotation is divided a and the sequence x under compound working conditions such as linkage of boom, arm and bucket b is divided, and the training set D train and the test set D test are divided.

[0049] Specifically:

[0050] Step 2.1: Divide the working condition sequence. Since the hydraulic pump presents different working states under different working conditions, from the input multivariate time series X, according to the execution of each action, the data of the hydraulic pump under single working conditions such as walking and rotation and compound working conditions such as linkage of boom, arm and bucket are divided to generate single working condition sequence x a and compound working condition sequence x b .

[0051] Step 2.2: Divide the data training set and the test set wherein is the data recorded by n v sensors / actuators at time stamp t. The result of detection is assigned a label y t ∈{0,1}, 1 represents x t is an abnormal data point, otherwise 0.

[0052] Step 2.3: Data preprocessing. Extract multivariate sequences from the training data set D train to , the window size is ω and the step size is s. Each multivariate sequence X n is defined as X n =[x (n-1)·s ,x (n-1)·s+1 ,...,x (n-1)·s+ω-1 ], and the test data D test is processed in the same way to obtain M multivariate sequences

[0053] Step 2.4: Calculate the adjacency matrix. Given a graph G=(V,E), where V and E represent vertices / nodes and edges respectively. The structure information can be represented by the adjacency matrix , wherein ω(i,j) is the weight between nodes i and j, N V It is the number of nodes.

[0054] Step 2.5: Calculate the activation function. RGCN updates the graph through the same graph convolutional layers, given the node features H of the current l-th layer. l The activation calculation for the next layer is as follows: Where σ is represented as The sigmoid activation function, W l It is the weight to be learned. Where I is a value of size N V The identity matrix, for The degree matrix.

[0055] Step 2.6: Optimize the adjacency matrix. Since not all multivariate time series data have a definite graph structure, meaning the adjacency matrix A is often unavailable, an adaptive approach is proposed to automatically learn the graph structure. Specifically, the adjacency matrix is ​​initialized as follows: Where σ is the sigmoid operation, W A' The weights of the adjacency matrix are learned and randomly initialized from a uniform distribution U(0,1). A' is both positive semi-definite and symmetric, and each element A' ij ∈(0,1).

[0056] Step 2.7: Add constraints to A'. A sparsity constraint L was added. L1-norm and uncertainty constraints L entropy , Applying sparse constraints (L) to the adjacency matrix L1-norm This makes A' sparse while preserving contribution relevance. For uncertainty constraint L... entropy By quantifying and reducing the uncertainty of prediction probability through entropy loss, it aims to predict the possibility of a correlation between two nodes.

[0057] Step 3: Encode the spatial information of the data using spatial-based RGCN. a and x b The input will be fed into two RGCNs with different adjacency matrices but the same node relationships and the same weights corresponding to the linear transformations. Then, an adaptive classifier d(x) is used to distinguish nodes from x. a and x b Generated spatial features and

[0058] Specifically:

[0059] Step 3.1: Use space-based RGCN to analyze x a and xb Encoding spatial information to generate spatial features and

[0060] Step 3.2: Discriminate spatial features with adaptive classifier d(x) and

[0061] Step 3.3: Calculate d(x) loss function. where L y is the cross-entropy loss function, and is the generated feature of the nth sequence, whose class label is 0 or 1.

[0062] Step 4: In order to learn the latent representation with spatial and temporal information, the spatial features generated based on RGCN and are further encoded with LSTM to generate temporal features f t a and f t b , respectively, and combined into feature set F;

[0063] Step 5: Assuming that the prior of feature set obeys a multivariate Gaussian distribution, i.e. p θ (z) = N(z; 0, I); the approximate posterior is a multivariate Gaussian distribution, i.e. q φ (z | x a ) = N(z; μ, σ 2 I), where the mean μ and standard deviation σ are output by the encoding multilayer perception composed of multiple dense layers. z ~ q φ (z | x a ) is sampled using the reparameterization trick;

[0064] Step 6: Signal reconstruction. Reconstruct p θ (x R | z) composed of LSTM and RGCN, the reconstructed sequence representation is x R , and the corresponding loss function is calculated;

[0065] Specifically:

[0066] Step 6.1: Signal reconstruction. Reconstruct p θ (x R | z) composed of LSTM and RGCN, the reconstructed sequence representation is x R .

[0067] Step 6.2: Calculate loss function. The loss function of the variational autoencoder is LVAE = L rec + L KLD where L rec is the reconstruction loss.

[0068] Step 6.3: Calculate the reconstruction loss. where MSE denotes the mean square error. Thus, L rec The reconstruction loss of RGCN input x a and LSTM input is considered simultaneously, so that it can better encode spatial and temporal information.

[0069] Step 6.4: Calculate the KL divergence. L KLD is a regularizer, representing the KL divergence between the encoder q φ (z | x a ) and the prior p θ (z), where N z is the dimension of the latent variable z.

[0070] Step 6.5: Calculate the overall loss function. L = λ1(L rec + L KLD ) + λ2L discr + λ3L entropy + L L1-norm where λ1, λ2 and λ3 are hyperparameters used to balance these loss terms. The factor of L rec and L KLD is fixed to be the same, and the factor of L L1-norm is set to 1.

[0071] Step 6.6: Hyperparameter adjustment. λ1, λ2 and λ3 are three hyperparameters that need to be adjusted. Different configurations of hyperparameters can lead to different performance of the model. In order to find the best configuration, a dynamic weighted loss scheme is adopted, which aims to adaptively learn the relative weights of the losses of multiple tasks.

[0072] Step 7: Calculate the reconstruction error. During the detection process, if the reconstruction error of timestamp t is greater than the defined threshold, mark this data point as abnormal.

[0073] Specifically:

[0074] Step 7.1: Calculate the reconstruction error. The reconstruction error or anomaly score of each test data point is defined as where and are the corresponding data points of x t in and . All the multivariate sequences containing x t xt The reconstruction error (mean square error) of t is accumulated and normalized.

[0075] Step 7.2: If the reconstruction error of the timestamp t is greater than the defined threshold, the data point is marked as abnormal. Otherwise, it is a normal data point.

[0076] The abnormality detection method combines ten kinds of signal data of the hydraulic pump, and all the data are obtained based on actual working conditions, and has very important application value for abnormality detection of the hydraulic pump and health maintenance of the whole equipment.

[0077] The present application is based on the actual working condition of the hydraulic pump of the excavator, and the single action and the compound action working condition division of the single walking, single rotating, boom, stick and bucket linkage are carried out, and the corresponding single working condition data and compound working condition data are selected. Then, the RGCN and the LSTM are used to analyze the characteristics and correlation of the multivariate time signal in time and space, and the VAE is used to encode, decode and signal reconstruction, and finally the reconstruction error is calculated for abnormality detection. In order to improve the adaptability of the model, firstly, the adjacency matrix in RGCN is optimized. Since not all multivariate time series data have a clear graph structure, i.e. the adjacency matrix is not available most of the time, an adaptive method is proposed to automatically learn the graph structure to optimize the adjacency matrix. Secondly, the hyperparameters are adjusted. Different configurations of hyperparameters may lead to different performance of the model. In order to find the best configuration, a dynamic weighted loss scheme is adopted to adaptively learn the relative weights of the losses of multiple tasks. Therefore, based on the multivariate time series data, a self-learning graph convolution network hydraulic pump abnormality detection method is proposed, which enhances the adaptability of RGCN and combines the ability of VAE to learn the mode distribution to avoid overfitting.

[0078] The present application not only researches the theoretical method, but also realizes the landing transformation of the theoretical research results in engineering practice to solve the practical problems encountered in practical application, which has very important significance for promoting the transformation of theoretical research into actual productivity and embodying the practical application value of theoretical research results.

Claims

1. A self-learning graph convolution network-based intelligent fault diagnosis method for engineering machinery, characterized in that: The method comprises the following steps: Step 1: Multivariate time series data of hydraulic pump under actual working conditions are collected as input sequence ; Step 2: Divide the working conditions according to the working condition type and the collected multivariate time series data; divide into single working condition sequence and composite working condition sequence ; and divide the training set and test set ; Step 3: encode the spatial information of data with spatial-based RGCN; and input to two RGCNs with different adjacency matrices but same node relations and same weights corresponding to linear transformation distinguish spatial features and ; Step 4: Process the spatial features generated based on RGCN and LSTM is used to further encode temporal feature information to generate temporal features. and and combine them into a feature set. ; Step 5: Hypothesis Set , The prior is assumed to be a multivariate Gaussian distribution, i.e. The approximate posterior is a multivariate Gaussian distribution, i.e. The reparameterization trick is used to sample from ; Step 6: signal reconstruction, the LSTM and RGCN composition reconstruction, the reconstructed sequence is represented as and the corresponding loss function is calculated; Step 7: Calculate reconstruction error, if timestamp Reconstruction error is greater than defined threshold, mark the timestamp corresponding value as an anomaly.

2. The method according to claim 1, characterized in that: The signal data collected in the step 1 comprises the data of the main pump pressure signal, the main pump flow control pressure, the main pump flow control current, the main pump pressure sensor voltage, the boom lifting pilot pressure, the boom lowering pilot pressure, the arm swing-out pilot pressure, and the arm swing-in pilot pressure.

3. The method of claim 1, wherein the method comprises: The specific steps of the step 2 are as follows: Step 2.1: Divide the working condition sequence, divide the data under the single working condition and the composite working condition of the hydraulic pump, and generate a single working condition sequence and a composite working condition sequence ; Step 2.2: Splitting the data training set and test set where is data recorded by the sensors / actuators at time stamp ; assigning a label to the result of the detection is an anomalous data point, otherwise 0; Step 2.3: Data pre-processing, extracting multivariate sequences from the training dataset with window size and stride , each multivariate sequence is defined as The test data is processed in the same way, resulting in multivariate sequences​​ Step 2.4: Compute the adjacency matrix, given a graph where V and E denote vertices / nodes and edges, respectively, and the structural information is represented by the adjacency matrix where , is the weight between nodes and , is the number of nodes; Step 2.5: Calculate the activation function. RGCN updates the graph through the same graph convolutional layer, given the current... Layer node features The activation calculation for the next layer is as follows: ,in It is represented as The sigmoid activation function, It is the weight to be learned. ,in It is the size of The identity matrix, for The degree matrix; Step 2.6: Optimize the adjacency matrix with an adaptive approach to automatically learn the graph structure, i.e., the adjacency matrix is initialized as where is the sigmoid operation, is the weight to learn the adjacency matrix, and is randomly initialized from a uniform distribution , is both positive semi-definite and symmetric, and each element ; Step 2.7: To add constraints, add sparsity constraints and uncertainty constraints , , , Apply sparsity constraints on the adjacency matrix to make sparse and keep contribution relevance; for uncertainty constraints , quantify and reduce the uncertainty of prediction probabilities through entropy loss.

4. The method according to claim 1, characterized in that: The specific steps of the step 3 are as follows: Step 3.1: encode the spatial information of and using a spatial-based RGCN to generate spatial features and ; Step 3.2: Utilizing an adaptive classifier Differentiating spatial features And ; Step 3.3: Compute Loss function, where is the cross-entropy loss function, and is the generated feature of the th sequence with its class label being 0 or 1.

5. The method of claim 1, wherein the method comprises: The specific steps of the step 6 are as follows: Step 6.1: Signal Reconstruction, for LSTM and RGCN composition reconstruction, the reconstructed sequence is denoted as ; Step 6.2: Compute the loss function, the loss function of the variational autoencoder is where is the reconstruction loss; Step 6.3: Compute reconstruction loss, where MSE denotes the mean squared error; while considering the reconstruction loss of RGCN input and LSTM input as well. Step 6.4: Compute the KL divergence, where is the dimensionality of the latent variable ; Step 6.5: calculating the overall loss function, ， wherein , and are hyperparameters used to balance these loss terms; fixing the factors of and to be the same and setting the factor of to 1 ; Step 6.6: Hyperparameter tuning, employing a dynamic weighted loss scheme for hyperparameter , and tuning.

6. The method of claim 1, wherein the method comprises: The specific steps of the step 7 are as follows: Step 7.1: Calculate reconstruction error, define the reconstruction error or anomaly score for each test data point as , where and are the corresponding data points in and ; the reconstruction errors of are accumulated and normalized across all multivariate data containing ; Step 7.2: If the reconstruction error of the timestamp is greater than a defined threshold, the timestamp corresponding numerical value is marked as an anomaly; Otherwise, it is a normal data point.

Images