Abnormal detection method for power battery production process fluctuation based on spatiotemporal graph model

Abstract

The present invention proposes a method for detecting fluctuation anomalies in the production process of power batteries based on a spatiotemporal graph model. The method includes: using the key parameters of the battery production process for model training, using a temporal convolutional network with different convolution kernel sizes to obtain the temporal model of the sequence; constructing a connection matrix, using a graph convolutional neural network to obtain the spatial information of the time series; constructing a spatiotemporal graph convolution block with a gating mechanism to filter the obtained information to obtain effective temporal and spatial dependencies; aggregating the output information of all gated spatiotemporal graph convolutional networks to perform a single-step prediction on the input sliding time window; applying the observed value and predicted value of the data to calculate the prediction error, and then applying the prediction error to calculate the threshold of the anomaly score; if the prediction error in the test data is greater than the threshold, it is judged as an anomaly and an alarm is issued.

Description

Technical Field

[0001] The present invention relates to the field of battery production process fluctuation analysis, and in particular to a battery production process abnormal fluctuation detection method based on a spatiotemporal graph model. Background Art

[0002] The environmental protection and energy-saving characteristics of new energy vehicles have led to their rapid popularization and development, and the future prospects are positive. As the core component of new energy locomotives, the automated monitoring of core process fluctuations in the production process of power batteries is conducive to reducing costs and improving efficiency. Looking at the world, the current factory battery management system design and intelligent manufacturing plan management module design are in the simulation or closed testing stage, and many core technologies are in urgent need of breakthroughs, including battery production equipment diagnosis, battery production process fluctuation analysis, and battery production process capability analysis. The present invention focuses on battery production process fluctuation analysis anomaly detection. The battery production process includes many core processes, such as mixing, coating, rolling, slitting, lamination, injection, formation, etc., which need to be monitored in a timely and accurate manner. At the same time, each process can be recorded by sensors in a timely and effective manner to obtain a multi-dimensional time series. These parameters (current, voltage, temperature, etc.) may hide abnormal models of battery production processes, and abnormality detection in the production process has become a very important issue. However, abnormal patterns come in various forms, and it is very difficult and costly to obtain labels for all types of abnormalities. At the same time, during the operation of battery production equipment, it is in normal state most of the time, resulting in a huge imbalance between abnormal and normal categories. In the field of deep learning, many popular supervised learning classification algorithms cannot work effectively, which increases the difficulty of detecting abnormal fluctuations in the production process of power batteries.

[0003] Taking into account the accuracy, robustness and efficiency of anomaly detection in battery production process fluctuation analysis, the unsupervised learning method is the most appropriate method. Through the deep learning model, the potential spatiotemporal relationship in the data is learned, the key parameters are predicted, and the anomalies are detected based on the difference between the observed value and the predicted value. However, in the battery production process, the stability of the parameters and the linear and nonlinear relationship between the parameters are very complex and change dynamically over time. The existing unsupervised anomaly detection algorithm cannot handle the temporal and spatial dependencies of the power battery production process, and there is a lot of room for improvement.

[0004] In summary, in the problem of power battery production process fluctuation analysis, finding a new multidimensional time series anomaly detection method that can model the time information and spatial patterns of the core production process parameters of the battery has become a problem that needs to be solved urgently. Summary of the invention

[0005] This paper proposes an anomaly detection algorithm for battery production process fluctuations, called the GateSpatial-Temporal Graph Model for Anomaly Detection (GSTAD). The algorithm considers the time dependency and spatial pattern of the core parameters of the battery production process, and can achieve efficient anomaly detection of battery production process fluctuations.

[0006] The technical solution of the present invention comprises the following steps:

[0007] 1) Preprocess the key parameters of the battery production process so that parameters with different attributes and different ranges have the same measurement.

[0008] 2) The preprocessed data is divided into three subsets, where the N1 subset is used for model training, which contains only normal data sets; the N2 subset is used for adjusting the model's hyperparameters and selecting thresholds; and the N3 subset is used to test the performance of the model. Both N2 and N3 contain normal and abnormal data. The input of the model is a sliding time window X W ∈R k×L , where k represents the characteristic dimension of the multidimensional time series, and L represents the size of the sliding time window;

[0009] 3) Construct a gated spatiotemporal graph model;

[0010] Firstly, we construct a temporal convolution network with convolution kernels of different sizes and a convolution kernel of 1×1 to capture the time dependency of key parameters of the power battery production process.

[0011] The key parameters of different power battery production processes are expressed in the form of a graph, which is represented by G = (V, E), where V is a set of nodes and E is a set of edges; k is used to represent the number of nodes in the graph. Let v∈V represent a node, e=(v,u)∈E represents an edge from u to v; the adjacency matrix of the directed graph of the constructed battery production process parameter system is represented by A∈R k×k , if (v i , v j )∈E, then A ij =1, if Then A ij =0; In the battery production process, there are complex relationships between different parameters, such as the front and rear stretching tensions of rollers, the slitting speed and working power of slitting machines, and the stirring speed and stirring temperature in the homogenization process. The linear and nonlinear relationships of different production processes are captured by the structure of the adaptive graph. Specifically, two methods are used to learn the adjacency matrix. First, according to the correlation of the process parameters in the battery production process, the adjacency matrix is ​​constructed. Second, the model learns the adaptive adjacency matrix A through gradient descentapt , by randomly initializing the embedding representations of the start node and the target node as learnable parameters E1,E2∈R k×m , where m represents the embedding dimension of each node, and the adaptive connection matrix of the target is expressed as Formula 3. The linear and nonlinear relationships of different production processes are captured by the structure of the adaptive graph, and the information of related process parameters is aggregated using the graph convolutional neural network. The graph convolutional neural network is expressed as Formula 4.

[0012]

[0013]

[0014] Among them A apt represents the initialized adaptive adjacency matrix. The ReLU activation function is used to remove some weak connections and increase the sparsity of the adjacency matrix. Softmax is the activation function. During the model training process, the optimal connection matrix that is conducive to prediction is learned. X represents the input of the graph convolutional neural network, W i1 and W i2 Represents the parameter matrix of the graph convolutional neural network model.

[0015] 4) Use gate structure to filter the obtained spatiotemporal information;

[0016] In the gated spatiotemporal graph model, an output gate g is used as shown in Formula 5. i , the output of each gate structure contains effective information of the time series. In formula 6, the output results of each gate structure are added together, and through two layers of fully connected layers, the spatiotemporal information of the multidimensional time series is obtained to predict the observation value of the next time step;

[0017] g i =h i1 (Θ1*X+a)⊙h i2 (Θ2*X+b) (5)

[0018]

[0019] where h i1 and hi2 represent the two operations of the tandem temporal convolution and graph convolutional neural network, Θ1 and Θ2 are the parameters, ⊙ represents the multiplication of the corresponding elements of the matrix, a, b, c, d are the bias terms, W1 and W2 are the parameter matrices of the fully connected layer, and Tanh and ReLU are the activation functions.

[0020] 5) Use normal data to train the model, use the automated threshold strategy, use the validation set to adjust the model's hyperparameters, and select the best threshold. Finally, test the performance of the model. Intuitively, when the test data set contains anomalies, a larger prediction error will be obtained, and observations with prediction errors greater than the threshold will be judged as abnormal.

[0021] Preferably, the parameters with different attributes and different ranges are made to have the same metric, specifically:

[0022] The data of each sensor is normalized using the following formula:

[0023]

[0024] where x∈R N Represents a metric parameter, min(x), max(x) represent the minimum and maximum values ​​of x respectively, x′ represents the normalized value of x, and the value of eps is set to 1e-8 to avoid division by zero.

[0025] Preferably, the key parameters of the power battery production process include rolling speed, slitting length, and injection temperature.

[0026] Preferably, the sliding time window size is set to 42.

[0027] Preferably, the expansion coefficient r is set to 2 or 3.

[0028] Preferably, the temporal convolutional network uses causal convolution and dilated convolution, as shown in Formula 2; causal convolution adds zero padding so that the output at the current moment is only related to its historical information and the input at the current moment, ensuring the causal and autoregressive characteristics of the sequence; dilated convolution increases the receptive field size of the convolution, capturing the temporal pattern of the time series under different receptive fields, and the size of the receptive field expands exponentially with the depth of the model, allowing the model to obtain more global information. At the same time, 1×1 convolution is added to enhance the model's ability to learn linear relationships;

[0029]

[0030] Where f represents the convolution kernel, K represents the length of the convolution kernel, r represents the dilation coefficient, x′ represents the preprocessed data, and t represents time.

[0031] The solution proposed in the present invention: A method for detecting anomalies in battery production process fluctuations based on a spatiotemporal graph convolutional neural network can successfully model the core parameter system of battery production process fluctuations, improve the accuracy of online anomaly detection, ensure the stability and efficiency of the battery production line, and reduce production costs and employee operating burdens. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a work flow chart of the present invention;

[0033] Figure 2 Diagram of the structure of bit-dilated causal convolution;

[0034] Figure 3 This is the block structure diagram of the temporal convolutional network;

[0035] Figure 4 This is the structural diagram of the graph convolutional neural network;

[0036] Figure 5 This is the structural diagram of the gated spatiotemporal graph convolutional neural network; DETAILED DESCRIPTION

[0037] In order to make the purpose, scheme and advantages of the embodiments of the present invention more clear, the technical scheme of the present invention will be described clearly and completely in conjunction with the accompanying drawings in the present invention. The abnormality detection method for battery production process fluctuation analysis implemented by the present invention is as follows: Figure 1 As shown, the method includes the following steps: data preprocessing, model training, threshold selection, and online anomaly detection, which are specifically divided into the following steps:

[0038] Step 1: Collect important parameter data of the battery production process at equal time intervals, such as rolling speed, slitting length, injection temperature, etc. These data have different attributes and ranges. Use the following formula to normalize the data of each sensor:

[0039]

[0040] Where x represents a metric parameter, min(x) and max(x) represent the minimum and maximum values ​​of x respectively, x′ represents the normalized value of x, and the value of eps is set to 1e-8 to avoid division by zero.

[0041] Step 2: Divide the data into three subsets, where the N1 subset is used for model training, which contains only normal data sets; the N2 subset is used to adjust the model's hyperparameters and select thresholds; and the N3 subset is used to test the performance of the model, with the model input being a sliding time window X. W ∈R k×L , where k represents the feature dimension of the multidimensional time series, L represents the size of the sliding time window, and the sliding time window size is set to 42; the output of the model is the predicted value of all features at the next moment;

[0042] Step 3: construct a gated spatiotemporal graph model;

[0043] First, we construct a temporal convolutional network with convolution kernel sizes of 1×3, 1×5, 1×7, and 1×9 and a convolution kernel of 1×1 to capture the time dependency of key parameters of the power battery production process; causal convolution and dilated convolution are used in the temporal convolutional network, as shown in Formula 2. Figure 2 As shown in the figure, causal convolution adds zero padding so that the output at the current moment is only related to its historical information and the input at the current moment, ensuring the causal and autoregressive characteristics of the sequence. Dilated convolution can increase the receptive field size of the convolution, where the dilation coefficient r is set to 2 or 3 to capture the temporal pattern of the time series under different receptive fields. The size of the receptive field expands exponentially with the depth of the model, allowing the model to obtain more global information. Weight normalization, ReLU activation function and Dropout regularization are also added to the temporal convolutional network, and finally residual connections are added to enable the model to avoid gradient disappearance in very deep networks. As shown in the figure, Figure 3 As shown in the figure, a temporal convolutional network with different convolution kernel sizes and a 1×1 convolution are used to construct a temporal convolutional block. The purpose of adding a 1×1 convolution is to improve the model's ability to learn linear relationships.

[0044]

[0045] Where f represents the convolution kernel, K represents the length of the convolution kernel, and r represents the dilation coefficient of the dilated convolution.

[0046] The key parameters of different power battery production processes are expressed in the form of a graph, which is represented by G = (V, E), where V is a set of nodes and E is a set of edges. Let k represent the number of nodes in the graph. Let v∈V represent a node and e=(v,u)∈E represent an edge from u to v. The adjacency matrix of the directed graph of the constructed battery production process parameter system is represented by A∈R k×k , if (v i , v j )∈E, then A ij =s, if Then A ij =0. In the battery production process, there are complex relationships between different parameters, such as the pre-stretching tension and post-stretching tension of the roller, the slitting speed and working power of the slitting machine, and the stirring speed and stirring temperature in the homogenization process. It is of great significance to capture the linear and nonlinear relationship of different production processes through the structure of the adaptive graph and predict the value of each key parameter at the next moment. Two methods are used to learn the adjacency matrix. First, the expert knowledge is used to construct the adjacency matrix through the correlation of the process parameters of the battery production process. Second, the model learns the adaptive adjacency matrix A through gradient descent apt , by randomly initializing the embedding representations of the start node and the target node as learnable parameters E1,E2∈R k×m, where m represents the dimension of each node embedding, and the adaptive connection matrix of the target is expressed as Formula 3. Figure 4 As shown in the figure, using the graph convolutional neural network, each node can aggregate the information of its related process parameters and update its own vector expression. The convolutional graph convolutional neural network is expressed as Formula 4. A two-layer graph convolutional neural network is used to form a spatial network block to learn the spatial dependency of multidimensional time series.

[0047]

[0048]

[0049] Among them A apt represents the initialized adaptive adjacency matrix. The ReLU activation function is used to remove some weak connections and increase the sparsity of the adjacency matrix. Softmax is the activation function. During the model training process, the optimal connection matrix that is conducive to prediction is learned. X represents the input of the graph convolutional neural network, W i1 and W i2 Represents the parameter matrix of the graph convolutional neural network model.

[0050] Step 4: Figure 5 As shown, the temporal convolution block in step 3 and the spatial convolution block in step 4 are used to establish a gated spatiotemporal graph convolutional neural network. Gating mechanisms are widely used in tasks that process sequences, such as long short-term temporal networks and gated recurrent units. In GSTAD, a gate structure is used to filter the obtained spatiotemporal information, and a simple output gate with a gate structure shown in Formula 5 is used to select and output information that is conducive to prediction. The output of each gate structure contains valid information about the time series. In Formula 6, the output results of each gate structure are added together, and the spatiotemporal information of the multidimensional time series is obtained through a fully connected layer to predict the observation value of the next time step.

[0051] g i =h i1 (Θ1*X+a)⊙h i2 (Θ2*X+b) (5)

[0052]

[0053] where h i1 and hi2 represent the two operations of the tandem temporal convolution and graph convolutional neural network, Θ1 and Θ2 are the parameters, ⊙ represents the multiplication of the corresponding elements, a, b, c, d are bias terms, W1 and W2 are the parameter matrices of the fully connected layer, and Tanh and ReLU are activation functions.

[0054] Step 5: Use normal data to train the model, use mean square error loss as the loss function, Adam as the optimizer, and use the gradient back propagation algorithm to update the model parameters. The trained model can make good predictions for normal states, but the prediction results for abnormal data are very wrong. The prediction error is used as the anomaly score, as shown in Formula 7. A non-parametric, dynamic, and unsupervised method is used to evaluate the residuals, which can select thresholds and identify extreme values ​​without making any assumptions. This method is suitable for data streams with different attributes and different ranges, and solves diversity, non-stationarity, and noise problems by automatically setting thresholds. Use the validation set to adjust the model's hyperparameters and select the best threshold.

[0055] The threshold comes from the set ε, which can be expressed by formula 8, and the optimal threshold comes from formula 9.

[0056]

[0057]

[0058] in:

[0059] Δμ(e)=μ(e)-μ({e∈e|e<ε})

[0060] Δσ(e)=σ(e)-σ({e∈e|e<ε})

[0061] e a ={e∈e|e>ε)

[0062] E seq =continous sequences of e a ∈e a (9)

[0063] where z is an ordered set of positive values ​​representing the number of standard deviations above the mean. The value of z depends on the context, but empirical results show that better results are obtained when z is between 3 and 13. This function also has a good effect on outliers with larger values ​​of e. a and sequence E seq Punishment is performed to avoid too many false positive behaviors, and finally the performance of the model is tested. The performance evaluation indicators of the model are precision (Precision, P), recall (Recall, R) and F1 score (F1 score), as shown in Formula 10. Since anomalies often appear in a continuous time period, if at least one abnormal observation is detected in an abnormal segment, the abnormal segment is considered to be correctly detected.

[0064]

[0065] TP is true positive, FP is false positive, and FN is false negative. Precision indicates the percentage of truly abnormal sequences among all sequences judged as abnormal by the model, Recall indicates the percentage of all abnormal sequences correctly identified as abnormal by the model, and F1 score is the harmonic mean of precision and recall, which is a more reasonable performance indicator. The maximum F1 score is obtained by adjusting the hyperparameters of the model to obtain the best model, which can realize the abnormal detection of battery production process fluctuations.

Claims

1. A method for detecting abnormal fluctuations in power battery production process based on a spatiotemporal graph model, characterized in that: The method specifically comprises the following steps: 1) Preprocess the key parameters of the battery production process so that parameters with different attributes and different ranges have the same measurement; 2) The preprocessed data is divided into three subsets, where the N1 subset is used for model training, which contains only normal data sets; the N2 subset is used for adjusting the model's hyperparameters and selecting thresholds; and the N3 subset is used to test the performance of the model. Both N2 and N3 contain normal and abnormal data. The input of the model is a sliding time window X W ∈R k×L , where k represents the characteristic dimension of the multidimensional time series, and L represents the size of the sliding time window; 3) Construct a gated spatiotemporal graph model; Firstly, we construct a temporal convolution network with convolution kernels of different sizes and a convolution kernel of 1×1 to capture the time dependency of key parameters of the power battery production process. The key parameters of different power battery production processes are expressed in the form of a graph, which is represented by G = (V, E), where V is a set of nodes and E is a set of edges; k is used to represent the number of nodes in the graph; let v∈V represent a node, e=(v,u)∈E represents an edge from u to v; the adjacency matrix of the directed graph of the constructed battery production process parameter system is represented by A∈R k×k , if (v i , v j )∈E, then A ij =1, if Then A ij =0; The linear and nonlinear relationships of different production processes are captured through the structure of the adaptive graph. Specifically, two methods are used to learn the adjacency matrix. First, the adjacency matrix is ​​constructed based on the correlation of the process parameters of the battery production process. Second, the model learns the adaptive adjacency matrix A through gradient descent apt , by randomly initializing the embedding representations of the start node and the target node as learnable parameters E1,E2∈R k×m , where m represents the embedding dimension of each node, and the adaptive connection matrix of the target is expressed as Formula 3; the linear and nonlinear relationship between different production processes is captured by the structure of the adaptive graph, and the information of related process parameters is aggregated using the graph convolutional neural network. The graph convolutional neural network is expressed as Formula 4, Among them A apt represents the initialized adaptive adjacency matrix, X represents the input of the graph convolutional neural network, and W i1 and W i2 Represents the parameter matrix of the graph convolutional neural network model; 4) Use gate structure to filter the obtained spatiotemporal information; In the gated spatiotemporal graph model, an output gate g is used as shown in Formula 5. i , the output of each gate structure contains effective information of the time series; in formula 6, the output results of each gate structure are added together, and through two layers of fully connected layers, the spatiotemporal information of the multidimensional time series is obtained to predict the observation value of the next time step; g i =h i1 (Θ1*X+a)⊙h i2 (Θ2*X+b) (5) where h i1 and hi2 represent the two operations of the tandem temporal convolution and graph convolutional neural network, Θ1 and Θ2 are the parameters, ⊙ represents the multiplication of the corresponding elements of the matrix, a, b, c, d are the bias terms, W1 and W2 are the parameter matrices of the fully connected layer, and Tanh and ReLU are the activation functions; 5) Use normal data to train the model, use an automated threshold strategy, use the validation set to adjust the model's hyperparameters, and select the best threshold; finally, test the performance of the model. Intuitively, when the test data set contains anomalies, observations with prediction errors greater than the threshold are judged as anomalies.

2. The method for detecting abnormal fluctuations in power battery production process based on a spatiotemporal graph model according to claim 1 is characterized in that: The method of making parameters with different attributes and different ranges have the same measurement is as follows: The data of each sensor is normalized using the following formula: where x∈R N Represents a metric parameter, min(x), max(x) represent the minimum and maximum values ​​of x respectively, x′ represents the normalized value of x, and the value of eps is set to 1e-8 to avoid division by zero.

3. The method for detecting abnormal fluctuations in power battery production process based on a spatiotemporal graph model according to claim 1 is characterized in that: The key parameters of the power battery production process include rolling speed, slitting length, and injection temperature.

4. The method for detecting abnormal fluctuations in power battery production process based on a spatiotemporal graph model according to claim 1, characterized in that: The sliding time window size is set to 42.

5. The method for detecting abnormal fluctuations in power battery production process based on a spatiotemporal graph model according to claim 1 is characterized in that: The expansion coefficient r is set to 2 or 3.

6. The method for detecting abnormal fluctuations in power battery production process based on a spatiotemporal graph model according to claim 1, characterized in that: The temporal convolutional network adopts causal convolution and dilated convolution, as shown in Formula 2; causal convolution adds zero padding so that the output at the current moment is only related to its historical information and the input at the current moment, ensuring the causal and autoregressive characteristics of the sequence; dilated convolution increases the receptive field size of the convolution, captures the temporal pattern of the time series under different receptive fields, and the size of the receptive field expands exponentially with the depth of the model, so that the model obtains more global information; at the same time, 1×1 convolution is added to improve the model's ability to learn linear relationships; Where f represents the convolution kernel, K represents the length of the convolution kernel, r represents the dilation coefficient, x′ represents the preprocessed data, and t represents time.

Images