Battery anomaly detection method based on deep support vector data scale description network

Abstract

The invention discloses a battery anomaly detection method based on a deep support vector data scale description network. The method specifically comprises the steps of 1, collecting electric vehicle battery sensor data; 2, preprocessing the battery sensor data acquired in the step 1 to obtain processed battery sensor data; step 3, constructing an SL-DSVDD anomaly detection model; and step 4, dividing the processed battery sensor data into a training set and a test set, inputting the training set into the SL-DSVDD anomaly detection model constructed in the step 3 for training to obtain a trained SL-DSVDD anomaly detection model, and inputting the test set into the trained SL-DSVDD anomaly detection model to obtain an anomaly detection result of the electric vehicle battery. The method solves the problem that an existing detection method is poor in detection efficiency and accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric vehicle (EV) power battery anomaly detection, and specifically relates to a battery anomaly detection method based on a deep support vector data scale description network. Background Art

[0002] With the widespread use of electric vehicles, the safety and stability of power batteries have become key factors in the overall performance and reliability of electric vehicles. However, during long-term operation, power batteries will inevitably age and degrade, leading to functional failures. These failures may lead to serious consequences, such as battery overheating, capacity decay, and even fire hazards. Therefore, effective status monitoring and fault warning of electric vehicle power batteries are particularly important.

[0003] Current battery anomaly detection methods can be mainly divided into two categories: physical model-based and data-driven models. Physical models perform fault diagnosis by establishing physical and chemical models of batteries, which has certain accuracy and reliability. However, when faced with complex and changeable battery data, their applicability and accuracy are often limited. Data-driven models better cope with the complexity and variability of battery data by extracting and transforming features in input data. Such models show high flexibility in feature extraction and anomaly detection, but common construction models have poor detection efficiency and accuracy when processing complex data sets. Summary of the invention

[0004] The purpose of the present invention is to provide a battery anomaly detection method based on a deep support vector data scale description network to solve the problems of poor detection efficiency and accuracy of existing detection methods.

[0005] The technical solution adopted by the present invention is a battery anomaly detection method based on a deep support vector data scale description network, which is specifically implemented according to the following steps: Step 1, collecting electric vehicle battery sensor data; Step 2, preprocessing the battery sensor data collected in step 1 to obtain processed battery sensor data; Step 3, construct the SL-DSVDD anomaly detection model; Step 4: Divide the processed battery sensor data into a training set and a test set, input the training set into the SL-DSVDD anomaly detection model constructed in step 3 for training, and obtain the trained SL-DSVDD anomaly detection model; input the test set into the trained SL-DSVDD anomaly detection model to obtain the anomaly detection result of the electric vehicle battery. The present invention is also characterized in that: In step 2, the preprocessing includes denoising, missing value processing, normalization, and time window processing.

[0006] In step 3, the SL-DSVDD anomaly detection model consists of a feature learning module based on scale learning and a DSVDD network; The feature learning module based on scale learning includes a supervision signal generation module, a multi-layer perceptron ; DSVDD network includes deep network and projection network , and multilayer perceptron Integrated in the DSVDD network and located in the deep network In the previous figure, the center of the hypersphere is the position of all normal samples in the feature space after passing through these three layers of networks.

[0007] Multilayer Perceptron The number of input features is 128, the number of output features is 10, the dimension of the hidden layer is 100, and the neurons in each layer are connected by the nonlinear activation function Sigmoid.

[0008] Deep Web It consists of 5 fully connected layers. The number of neurons in the first to fifth layers are 10, 20, 10, 10, and 15 respectively. Each layer is connected by a nonlinear activation function Sigmoid. It consists of 5 fully connected layers. The number of neurons in the 1st to 5th layers are 15, 10, 10, 20, and 15 respectively. Each layer is connected by the nonlinear activation function Sigmoid.

[0009] In step 3, the processing process of the SL-DSVDD anomaly detection model is: The Pearson product-moment correlation coefficient is used to calculate the correlation between the feature columns of different battery sensor data to obtain the PPMCC matrix. The PPMCC expression is: (1) In the formula, and Two specific column vectors representing processed battery sensor data, express The mean of express The mean of, N represents the total number of input battery sensor data; Given the PPMCC matrix , the PPMCC value corresponding to each feature for: (2) In the formula, A specific column vector representing the input processed battery sensor data, Indicates the corresponding ordinal number , k is the number of column features of the processed battery sensor data instance; For each processed battery sensor data instance input, the sampling is repeated c times, and then random sampling is performed γ times to generate the sampled data , according to the PPMCC matrix and sampling data, the monitoring signal is obtained ,Right now: (3) In the formula, Indicates that the sampled data Vectors filled with zeros to equal length; is a random number less than k, representing the number of selected processed battery sensor data instance column features; Vector Sent as input to the projection mesh In the process, an intermediate step vector is generated. , where h is the length of the third dimension after projection; (4) The intermediate step vector u Multilayer Perceptron Convert to construct feature space vector ; (5) In the formula, ; The feature space vector will be constructed With monitoring signal Compare and calculate the loss L1; (6) The constructed feature space vector output by the feature learning module based on scale learning As the input of the DSVDD network, the anomaly score is determined by calculating the distance between the sample and the center of the hypersphere. The output is the anomaly score based on the distance from the data instance to the center of the hypersphere. When the distance of the sample exceeds the preset radius, the sample is judged to be abnormal, which is the anomaly detection result of the electric vehicle battery.

[0010] Projection Network It is a single-layer linear projection network with 20 input features and 128 output features.

[0011] The loss function of the DSVDD network is: (8)

[0012] In the formula, Represents the construction of feature space vector, i.e., deep network Input, Represents the projected instance, i.e., the projection network The output, is the center of the DSVDD hypersphere, R is the maximum tolerance radius of the hypersphere, is a hyperparameter that determines the weight of the radius penalty term, Indicates the number of samples; The total loss of the SL-DSVDD anomaly detection model is: (9) In the formula, and are hyperparameters that determine the weights of the two loss components, which are 1 and 0.1 respectively.

[0013] In step 4, the training process of the SL-DSVDD anomaly detection model is as follows: the input data is extracted from the high-level distribution features through the feature learning module based on scale learning, and is passed to the DSVDD network to calculate the anomaly score of each sample point; based on the result of the forward propagation, the total loss value of the SL-DSVDD anomaly detection model is calculated according to formulas (6), (8) and (9); the gradient of the loss value to the network parameters is calculated through the back propagation algorithm, and the parameters are adjusted according to the optimization algorithm; the hyperparameters are automatically adjusted through grid search or Bayesian optimization method to obtain the best model performance.

[0014] The beneficial effects of the present invention are: (1) The feature learning module based on scale learning proposed by the method of the present invention automatically adjusts the importance of features by modeling and aligning the correlation between feature columns, reduces the influence of irrelevant features, and enables the model to capture more useful distribution information in high-dimensional space, thus avoiding the occurrence of data concentration phenomenon; (2) The method of the present invention can maintain strong robustness in a complex data distribution environment by constructing an SL-DSVDD network. It can not only process high-dimensional features, but also accurately identify anomalies under complex distribution. Through deep feature extraction and high-level distribution analysis, the reliability of detection is greatly improved. (3) The SL-DSVDD network constructed by the method of the present invention adopts a multi-layer deep support vector data description network. By learning the distribution rules in high-dimensional space to generate anomaly scores, it can not only significantly improve the accuracy of anomaly detection, but also accelerate the convergence speed of the model and improve the computing efficiency. At the same time, it can also maintain stability and efficiency in high-dimensional data sets, thereby greatly improving the actual application effect of electric vehicle power battery anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a battery anomaly detection method based on a deep support vector data scale description network of the present invention; Figure 2 The anomaly detection flow chart of the SL-DSVDD anomaly detection model constructed by the present invention; Figure 3 A network framework diagram of a feature learning module based on scale learning constructed in the method of the present invention; Figure 4 It is a framework diagram of the DSVDD network constructed in the method of the present invention. DETAILED DESCRIPTION

[0016] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1 The present invention is based on a battery anomaly detection method of a deep support vector data scale description network, such as Figure 1 As shown, the specific implementation steps are as follows: Step 1, collecting electric vehicle battery sensor data; Step 2, preprocessing the battery sensor data collected in step 1 to obtain processed battery sensor data; Step 3, construct the SL-DSVDD anomaly detection model; Step 4: divide the processed battery sensor data into a training set and a test set, input the training set into the SL-DSVDD anomaly detection model constructed in step 3 for training, and obtain the trained SL-DSVDD anomaly detection model; input the test set into the trained SL-DSVDD anomaly detection model to obtain the anomaly detection result of the electric vehicle battery.

[0018] Example 2 Based on Example 1, the battery sensor data includes vehicle speed, vehicle mode, vehicle mileage, on-board energy storage device status, vehicle gear battery fault list, voltage number of a single battery, maximum battery voltage, minimum battery voltage, total voltage, total current, insulation resistance, maximum battery temperature, minimum battery temperature, number of battery temperature probes, maximum temperature probe number, minimum temperature probe number, maximum alarm level and specific alarm information; The battery sensor data is collected in real time through the vehicle controller (ECU), and the battery sensor data is set in the power battery system of the electric vehicle.

[0019] Example 3 Based on Example 2, the preprocessing includes denoising, missing value processing, normalization, and time window processing in sequence; Among them, the denoising process uses Kalman filtering or moving average filtering to filter the noise in the sensor data and eliminate the interference of the external environment on data collection; Missing value processing uses interpolation or filling mechanisms to handle missing values ​​in sensor data to ensure data continuity; The normalized expression is:

[0020] In the formula, X is the battery sensor data, X' is the normalized data, and X min is the minimum value of the battery sensor data, X max The maximum value of the battery sensor data; Time windowing is to divide the continuous sensor data into fixed time windows (such as 1 minute or 5 minutes) to ensure the timing characteristics of the data, while reducing the amount of data and improving computing efficiency.

[0021] Example 4 The present invention is based on a battery anomaly detection method of a deep support vector data scale description network, such as Figure 1 As shown, the specific implementation steps are as follows: Step 1, collecting electric vehicle battery sensor data; Battery sensor data includes vehicle speed, vehicle mode, vehicle mileage, onboard energy storage device status, vehicle gear battery fault list, voltage number of individual batteries, maximum battery voltage, minimum battery voltage, total voltage, total current, insulation resistance, maximum battery temperature, minimum battery temperature, number of battery temperature probes, maximum temperature probe number, minimum temperature probe number, maximum alarm level and specific alarm information; Step 2, preprocessing the battery sensor data collected in step 1 to obtain processed battery sensor data; Preprocessing includes denoising, missing value processing, normalization, and time window processing; Step 3, construct the SL-DSVDD anomaly detection model; like Figure 2 As shown, the SL-DSVDD anomaly detection model consists of a feature learning module based on scale learning and a DSVDD network (deep support vector data description network); like Figure 3 As shown in the figure, the feature learning module based on scale learning is used to extract features from the processed battery sensor data, generate scale supervision signals by calculating the correlation between the feature columns of each battery sensor data, and align the data through the scale supervision signals, thereby effectively alleviating the high-dimensional concentration problem and extracting high-level distribution; specifically: the feature learning module based on scale learning includes a supervision signal generation module, a multi-layer perceptron ; The supervisory signal generation module uses the Pearson product-moment correlation coefficient (PPMCC) to calculate the correlation between the feature columns of different battery sensor data and obtain the PPMCC matrix. The PPMCC expression is: (1) In the formula, and Two specific column vectors representing processed battery sensor data, express The mean of express The mean of, N represents the total number of input battery sensor data; The PPMCC matrix reflects the linear relationship between features. If the correlation between two features is high, they are considered to have a higher similarity. Features with lower correlation are considered to contribute less to the model and are therefore assigned lower weights. The correlation values ​​in the PPMCC matrix are used as the basis for weight assignment. Features with higher correlation are assigned higher weights, and features with lower correlation are assigned lower weights. This can effectively reduce reliance on noise features and prevent data dimensionality concentration problems.

[0022] Given the PPMCC matrix , where k is the number of features in the processed battery sensor data instance column, and the PPMCC value corresponding to each feature is for: (2) In the formula, A specific column vector representing the input processed battery sensor data, Indicates the corresponding ordinal number A specific column vector of ; For each processed battery sensor data instance input, the sampling is repeated c times, and then γ times of random sampling is performed to generate the sampled data , is less than k The random number represents the number of selected processed battery sensor data instance column features. According to the PPMCC matrix and the sampled data, the monitoring signal is obtained. ,Right now: (3) In the formula, Indicates that the sampled data Vectors filled with zeros to equal length; Vector Sent as input to the projection mesh , generates an intermediate step vector (uniform signal) , where h is the length of the third dimension after projection, to adjust the dimension and destroy the original distribution, thereby suppressing the influence of random noise and abnormal instance features. Projection Network It is a single-layer linear projection network with 20 input features and 128 output features; (4) The intermediate step vector u Multilayer Perceptron Convert to construct feature space vector , used to reconstruct instances to follow the distribution of the supervision signal and prepare for computing the scale loss; Multilayer Perceptron The input feature number of the input layer is 128, the output feature number is 10, the dimension of the hidden layer is 100, and the neurons in each layer are connected by the nonlinear activation function Sigmoid; (5) In the formula,

[0023] The feature space vector will be constructed With monitoring signal For comparison, calculate the loss L1, the expression is: (6) The whole process of generating loss L1 can be described as: (7) In the present invention, the multi-layer perceptron The input data has 128 features and the input dimension is 128×1. According to the number of input and output features and the dimension of the hidden layer, The weight matrix size of the first layer is 100×128, and the weight matrix size of the second layer is 10×128; the input is already passed through the projection network After the processed data, the output is the data vector with a dimension of 10×1 after being projected into the new feature space.

[0024] The input of the DSVDD network is the output of the feature learning module based on scale learning. The hypersphere is constructed. The output of the DSVDD network is the anomaly score based on the distance from the data instance to the center of the hypersphere. The DSVDD network is used to detect low-level and high-level distribution anomalies in the data and determine the anomaly score by calculating the distance between the sample and the center of the hypersphere. When the distance of the sample exceeds the preset radius, the sample is judged as abnormal.

[0025] like Figure 4 As shown, the DSVDD network includes a deep network and projection network , and multilayer perceptron Integrated in the DSVDD network and located in the deep network In the front; the center of the hypersphere is the position of all normal samples in the feature space after passing through these three layers of networks, and the radius of the hypersphere is the maximum distance between the normal samples and the center of the sphere.

[0026] Deep Web It consists of 5 fully connected layers. The number of neurons in the 1st to 5th layers are 10, 20, 10, 10, and 15 respectively. Each layer is connected by a nonlinear activation function Sigmoid. It is used to learn low-level distribution features and make them generally separable in space. The output of the feature learning module based on scale learning is a deep network. Input; Projection Network It consists of 5 fully connected layers. The number of neurons in the 1st to 5th layers are 15, 10, 10, 20, and 15 respectively. Each layer is connected by a nonlinear activation function Sigmoid. The data instance is projected and compressed into a hypersphere in the inverse distribution variation space for anomaly detection. Deep network The output is the projection network Input, projection network The output is the abnormality detection result of the electric vehicle battery.

[0027] The loss function of the DSVDD network is: (8)

[0028] In the formula, Represents the construction of feature space vector, i.e., deep network Input, Represents the projected instance, i.e., the projection network The output, is the center of the DSVDD hypersphere, R is the maximum tolerance radius of the hypersphere, is a hyperparameter that determines the weight of the radius penalty term, Indicates the number of samples; The total loss of the SL-DSVDD anomaly detection model is: (9) In the formula, and are hyperparameters that determine the weights of the two loss components, which are 1 and 0.1 respectively.

[0029] The deep network in the SL-DSVDD anomaly detection model of the present invention and multilayer perceptron The SL-DSVDD anomaly detection model will not be affected by the collapse of the hypersphere for the following reasons: Assumption 1: In a DSVDD network with sigmoid as the activation function, when the hypersphere collapses, if the neurons in a certain layer of each subnetwork collapse to zero, the neurons in the remaining layers must also collapse to 0; Proof 1: Assume that only one layer of the neural network collapses to zero. Since bias is usually not used in DSVDD networks, we consider the unbiased case. In addition, we assume that the activation function is S-shaped, which is helpful in preventing the hypersphere from collapsing in our experiments. make represents the weight matrix of each layer, represents the learning rate, represents the input matrix of each layer (the output of the last layer), represents the derivative of the activation function, represents the propagation error matrix, represents the intermediate error matrix of each layer, △w i represents the weight error; Then we have:

[0030]

[0031]

[0032] Since the center of the hypersphere has changed, the error of the final output layer will not be zero and will change at different times; Therefore, the propagated error matrix of each layer after the collapse will not be a zero matrix, and the updated weight value will not be zero. This update process will be extended to the folded layer. Obviously, the collapsed layer will not be updated with the theoretical zero gradient. Therefore, the previous assumption 1 is wrong, so there must be a folded layer after any non-output layer.

[0033] Based on proof 1, we derive hypothesis 2; Assumption 2: The activation function is Sigmoid, and the neurons in the DSVDD network with non-constant center cannot all collapse to zero; Proof 2: Assuming Deep Net All neurons in the fold collapse to zero; Then we have:

[0034] However, the center of the hypersphere is not constant, which shows that:

[0035] Therefore, deep net Will not collapse. Now, suppose that the multilayer perceptron is zero, according to hypothesis 1, the deep network will also collapse to 0, but this is not true. Therefore, the deep network and multilayer perceptron None of them collapse, which shows that our model is not affected by the collapse of the hypersphere.

[0036] Step 4, dividing the processed battery sensor data into a training set and a test set, inputting the training set into the SL-DSVDD anomaly detection model constructed in step 3 for training to obtain a trained SL-DSVDD anomaly detection model, and inputting the test set into the trained SL-DSVDD anomaly detection model to obtain an anomaly detection result of the electric vehicle battery; The training process of the SL-DSVDD anomaly detection model is as follows: Step 4.1, forward propagation: The input data is passed through the scale-based feature learning module to extract high-level distribution features and is passed to the DSVDD network to calculate the anomaly score of each sample point; Step 4.2, loss calculation: According to the result of forward propagation, the total loss value of the SL-DSVDD anomaly detection model is calculated according to formula (6), formula (8) and formula (9); Step 4.3, back propagation: The gradient of the loss value to the network parameters is calculated through the back propagation algorithm, and the parameters are adjusted according to the optimization algorithm (such as Adam or SGD). The Adam optimizer performs well when processing high-dimensional data and can adaptively adjust the learning rate to accelerate convergence; Step 4.4, hyperparameter adjustment: In order to avoid overfitting problems in the SL-DSVDD anomaly detection model during training, the best hyperparameter combination is selected by cross-validation during model training, including learning rate, batch size, and weight coefficient in the loss function. and . These hyperparameters are automatically tuned through methods such as grid search or Bayesian optimization to obtain the best model performance.

[0037] Example 5 After model training is completed, the SL-DSVDD model will be rigorously verified and evaluated on multiple electric vehicle battery data sets. Common performance evaluation indicators include: AUC (Area Under the Curve): Used to evaluate the classification ability of the model. The closer the AUC value is to 1, the stronger the model's ability to distinguish normal from abnormal samples.

[0038] F1-Score: Measures the balance between the accuracy and recall of model classification, which is particularly suitable for anomaly detection tasks because the number of abnormal data is usually small.

[0039] Accuracy: Evaluates the overall performance of the model across all data.

[0040] Recall: Pay special attention to the performance of the model in detecting abnormal samples, ensuring that as few important anomalies as possible are missed in practical applications.

[0041] Through extensive experiments on electric vehicle battery datasets, the SL-DSVDD model has demonstrated higher robustness and detection accuracy than traditional detection algorithms. Especially when processing high-dimensional complex data, SL-DSVDD can effectively capture subtle anomalies in the data to ensure the safety and reliability of the battery.

[0042] The dataset used for the experiment includes standardized data of 10 electric vehicles collected by the National New Energy Vehicle Big Data Alliance (NDANEV). There are 5,509,708 data instances in the dataset, including approximately 4 million driving state data, 1 million charging state data, and 650,000 stationary data. As shown in Table 1, for each data instance, 29 features are provided, including the gear position of the vehicle, the maximum temperature of the battery, and the total current of the battery.

[0043] Table 1 Basic information of the new energy vehicle big data open laboratory dataset

[0044] The proportion of anomalies in the NDANEV dataset accounts for only a small fraction of the total instances, 0.2% on average. All different anomalies are listed in Table 2.

[0045] Table 2 Basic anomalies of the new energy vehicle big data open laboratory data set

[0046] Example 6 For the following single-car experiments, data for four car models, LB24, LB42, LB47, and LB74, are provided. LB24 and LB47 are selected because they have a relatively large proportion of abnormal instances and a variety of abnormal categories (as shown in Table 3), which can effectively test the performance of each algorithm. LB42 and LB74, which have a limited number of anomalies and abnormal categories, are selected as comparisons to demonstrate the effectiveness of the experimental setup. A mixed multi-car dataset is formed based on all 10 types of cars to explore the performance of the proposed model in relatively complex situations. The multi-car dataset is obtained by undersampling the normal instances of each type of vehicle and forms a 50% abnormality rate dataset with 20,688 samples and 28 abnormal conditions.

[0047] Table 3 Introduction to the bicycle dataset

[0048] The performance of the SL-DSVDD model of the present invention was verified by experiments on electric vehicle battery datasets and comparisons with other algorithms. AUC, F1 score, recall, and precision were incorporated to evaluate the efficacy of SL-DSVDD, on which the SL-DSVDD model of the present invention was compared with DSVDD, DIF, RCA, ICL, REPEN, and DevNet. Experiments were conducted on PyTorch 1.13.1 using the Adam optimizer and NVIDIA GeForce GTX 4060. The learning rate was set to 0.00006 and the batch size was set to 128. Considering the large proportion of normal data, the normal data of the four single-vehicle datasets (LB24, LB42, LB47, and LB74) were downsampled before the experiment, which enabled us to obtain smaller datasets to shorten the training time and verify the high level of feature extraction efficiency of the tested algorithm. At the same time, we eliminated 9 features with low correlation with other features and abnormal states in all experimental datasets to combat the curse of dimensionality.

[0049] Table 4

[0050] As shown in Table 4, the AUC of the SL-DSVDD model, DSVDD and DevNet of the present invention on all four single vehicle datasets exceeded 90%, and exceeded 95% on at least three datasets; their F1, AP and BP indicators are also generally relatively high, showing their effectiveness in battery anomaly detection. However, the AUC of DSVDD on LB47 and the AUC of DevNet on LB24 both dropped by more than 5%, and their F1 scores also dropped by about 10%. In contrast, the four indicators of the SL-DSVDD model of the present invention on all four datasets remained relatively stable and were above 99.9% in most cases. On LB24, LB47 and LB74, the SL-DSVDD model of the present invention gave the highest AUC, BR and F1-Score. On the simplest single vehicle dataset LB42, the indicators of the SL-DSVDD model of the present invention are not the highest, but very close to the optimal value. This shows the superiority, reliability, stability and noise resistance of the SL-DSVDD model of the present invention. It is worth noting that most algorithms achieve satisfactory performance on the simpler datasets LB42 and LB74, which can demonstrate the effectiveness and fairness of our experimental settings.

[0051] The outstanding performance of the SL-DSVDD model of the present invention is mainly attributed to its scale learning mechanism, which enables it to fully learn the high-level features of the data and reduce the impact of inter-class similarity.

[0052] The experimental performance of the proposed SL-DSVDD framework demonstrates reliability, accuracy, and robustness on various single vehicle datasets. This demonstrates the effectiveness of the combined scale learning mechanism and shows the advancement of our algorithm in the field of EV battery anomaly detection and its adaptability to different vehicle models.

Claims

1. A battery anomaly detection method based on a deep support vector data scale description network, characterized in that: Follow the steps below to implement it: Step 1, collecting electric vehicle battery sensor data; Step 2, preprocessing the battery sensor data collected in step 1 to obtain processed battery sensor data; Step 3, construct the SL-DSVDD anomaly detection model; Step 4: divide the processed battery sensor data into a training set and a test set, input the training set into the SL-DSVDD anomaly detection model constructed in step 3 for training, and obtain the trained SL-DSVDD anomaly detection model; input the test set into the trained SL-DSVDD anomaly detection model to obtain the anomaly detection result of the electric vehicle battery.

2. The battery anomaly detection method based on deep support vector data scale description network according to claim 1 is characterized in that: In step 2, the preprocessing includes denoising, missing value processing, normalization, and time window processing.

3. The battery anomaly detection method based on deep support vector data scale description network according to claim 1 is characterized in that: In step 3, the SL-DSVDD anomaly detection model consists of a feature learning module based on scale learning and a DSVDD network; The feature learning module based on scale learning includes a supervision signal generation module, a multi-layer perceptron ; DSVDD network includes deep network and projection network , and multilayer perceptron Integrated in the DSVDD network and located in the deep network In the previous figure, the center of the hypersphere is the position of all normal samples in the feature space after passing through these three layers of networks.

4. The battery anomaly detection method based on deep support vector data scale description network according to claim 3 is characterized in that: Multilayer Perceptron The number of input features is 128, the number of output features is 10, the dimension of the hidden layer is 100, and the neurons in each layer are connected by the nonlinear activation function Sigmoid.

5. The battery anomaly detection method based on deep support vector data scale description network according to claim 3 is characterized in that: Deep Web It consists of 5 fully connected layers. The number of neurons in the first to fifth layers are 10, 20, 10, 10, and 15 respectively. Each layer is connected by a nonlinear activation function Sigmoid. It consists of 5 fully connected layers. The number of neurons in the 1st to 5th layers are 15, 10, 10, 20, and 15 respectively. Each layer is connected by the nonlinear activation function Sigmoid.

6. The battery anomaly detection method based on deep support vector data scale description network according to claim 3 is characterized in that: In step 3, the processing process of the SL-DSVDD anomaly detection model is: The Pearson product-moment correlation coefficient is used to calculate the correlation between the feature columns of different battery sensor data to obtain the PPMCC matrix. The PPMCC expression is: （1） In the formula, and Two specific column vectors representing processed battery sensor data, express The mean of express The mean of, N represents the total number of input battery sensor data; Given the PPMCC matrix , the PPMCC value corresponding to each feature for: （2） In the formula, A specific column vector representing the input processed battery sensor data, Indicates the corresponding ordinal number , k is the number of column features of the processed battery sensor data instance; For each processed battery sensor data instance input, the sampling is repeated c times, and then random sampling is performed γ times to generate the sampled data , according to the PPMCC matrix and sampling data, the monitoring signal is obtained ,Right now: （3） In the formula, Indicates that the sampled data Vectors filled with zeros to equal length; is less than k A random number representing the number of selected features of the processed battery sensor data instance column; Vector Sent as input to the projection mesh In the process, an intermediate step vector is generated. , where h is the length of the third dimension after projection; （4） The intermediate step vector u Multilayer Perceptron Convert to construct feature space vector ; （5） In the formula, ; The feature space vector will be constructed With monitoring signal Compare and calculate the loss L1; （6） The constructed feature space vector output by the feature learning module based on scale learning As the input of the DSVDD network, the anomaly score is determined by calculating the distance between the sample and the center of the hypersphere. The output is the anomaly score based on the distance from the data instance to the center of the hypersphere. When the distance of the sample exceeds the preset radius, the sample is judged to be abnormal, which is the anomaly detection result of the electric vehicle battery.

7. The battery anomaly detection method based on deep support vector data scale description network according to claim 6 is characterized in that: Projection Network It is a single-layer linear projection network with 20 input features and 128 output features.

8. The battery anomaly detection method based on deep support vector data scale description network according to claim 6 is characterized in that: The loss function of the DSVDD network is: （8） In the formula, Represents the construction of feature space vector, i.e., deep network Input, Represents the projected instance, i.e., the projection network The output, is the center of the DSVDD hypersphere, R is the maximum tolerance radius of the hypersphere, is a hyperparameter that determines the weight of the radius penalty term, Indicates the number of samples; The total loss of the SL-DSVDD anomaly detection model is: （9） In the formula, and are hyperparameters that determine the weights of the two loss components, which are 1 and 0.1 respectively.

9. The battery anomaly detection method based on deep support vector data scale description network according to claim 6 is characterized in that: In step 4, the training process of the SL-DSVDD anomaly detection model is as follows: the input data is extracted from the high-level distribution features through the feature learning module based on scale learning, and is passed to the DSVDD network to calculate the anomaly score of each sample point; based on the result of the forward propagation, the total loss value of the SL-DSVDD anomaly detection model is calculated according to formulas (6), (8) and (9); the gradient of the loss value to the network parameters is calculated through the back propagation algorithm, and the parameters are adjusted according to the optimization algorithm; the hyperparameters are automatically adjusted through grid search or Bayesian optimization method to obtain the best model performance.

Images