A new energy vehicle power battery fault classification method based on ShuffleDarkNet37-SE

Abstract

This paper proposes a new energy vehicle power battery fault classification method based on ShuffleDarkNet37-SE. This ShuffleDarkNet37-SE method includes: preprocessing initial data from new energy vehicles after acquiring it from vehicle sensors; inputting the preprocessed data into a ShuffleDarkNet37-SE classification network; and issuing a corresponding battery fault alarm after the classification network outputs the results. This ShuffleDarkNet37-SE method can determine the fault type of new energy vehicle batteries, provide timely warnings, and reduce vehicle battery failures.

Description

Technical Field

[0001] The present invention belongs to the field of fault monitoring and fault classification, and relates to a ShuffleDarkNet37-SE new energy vehicle power battery fault classification method, which is suitable for classifying new energy vehicle power battery faults. Background Art

[0002] New energy vehicle battery failure classification prediction methods are mainly divided into the following three categories: statistical analysis methods, physical model methods, and machine learning methods. Statistical analysis methods perform statistical analysis on historical failure data to determine the accuracy of classification predictions for different failure types.

[0003] Statistical methods are only applicable to scenarios with statistical or periodic laws. However, the driving environment of new energy vehicles, the battery usage environment, and the material properties of battery packs are not periodic or similar. They can only be used to model a single brand under a similar operating environment. The efficiency is extremely low and the generalization is too poor. It can be seen that statistical methods are not suitable for the classification and prediction of new energy vehicle battery failures.

[0004] The physical model approach achieves classification and prediction of different fault types by modeling the battery's internal chemical reactions, heat conduction, and current distribution. This approach requires a precise internal battery model with accurate parameters, but this can lead to individual deviations as each cell is used differently, and the accuracy of the classification prediction cannot be guaranteed.

[0005] Machine learning methods include time series methods, image convolutional neural networks, and deep neural networks. Time series methods, including recurrent neural networks, long short-term memory networks, bidirectional long short-term memory networks, and gated networks, are primarily applicable to prediction problems with strong time series characteristics. Deep neural network methods are primarily suitable for classification and prediction problems with low-dimensional data that are less prone to overfitting. Image convolutional neural networks, due to their deep convolutional modules, are therefore applied to image-based classification and prediction problems with high-dimensional data.

[0006] Currently, there's no method that can achieve both rapid training and high accuracy, while also being applicable to a wide range of datasets. For each different dataset, the network needs to be tested and adapted repeatedly. Inspired by ShuffleNet and DarkNet53, we combine the results of these two methods using deep, fully connected layers to improve the accuracy of new energy power battery fault classification. Summary of the Invention

[0007] The present invention proposes a ShuffleDarkNet37-SE new energy vehicle power battery fault classification method, which combines ShuffleNet, DarkNet37 and SENet for the classification of new energy vehicle power battery faults, improves the accuracy and efficiency of new energy vehicle battery fault classification, and reduces vehicle battery failures; the steps in the use process are:

[0008] A. After the vehicle sensor obtains the initial data of the new energy vehicle, pre-process the initial data of the new energy vehicle;

[0009] The step A comprises:

[0010] (1) Clean the original data and delete the meaningless data and outlier data in the original data;

[0011] (2) When the number of data samples is unbalanced, the linear interpolation method is used to expand the data. The linear interpolation method is:

[0012] (1)

[0013] Where, For newly generated sample data; and They are the two original sample data before and after respectively; The number of new data inserted into the original sample; The location generated for new data;

[0014] (3) If the magnitude difference of data is too large, it will cause a large error in the accuracy of new energy vehicle fault classification. The maximum and minimum normalization method is used to normalize the data to eliminate the magnitude difference between the data of each dimension. The maximum and minimum normalization method is:

[0015] (2)

[0016] Where, is the normalized data; is the original data; is the maximum value of the column where the sample data is located; is the minimum value of the column where the sample data is located;

[0017] (4) Data redundancy will affect the fault classification performance of the model. The Pearson correlation coefficient method is used to reduce the dimension of the data. The Pearson correlation coefficient method is:

[0018] (3)

[0019] Where, for 、 The correlation coefficient value of two columns of data; for 、 The covariance between and Respectively 、 respective standard deviations;

[0020] (5) After reducing the data dimension, the data samples are converted into 40×40×1 images as the final input of the classification model;

[0021] (6) Divide the data into training set and test set, with the ratio of training set to test set being 2:1;

[0022] B. After data preprocessing, the data is input into the ShuffleDarkNet37-SE classification network;

[0023] The classification network includes:

[0024] (1) Reduce 16 convolutional layers from DarkNet53 to reduce the network's computational parameters and make the classification model lightweight. The network is named DarkNet37.

[0025] (2) The attention channel can focus on important features, and the channel attention module is integrated into the DarkNet37 network to form a DarkNet37-SE network;

[0026] (3) The results of ShuffleNet and DarkNet37-SE are combined through a fully connected layer network, and the classification probability is output through a normalized index to form the ShuffleDarkNet37-SE classification method. This method absorbs the advantages of the two networks and improves the classification accuracy. The normalized index is:

[0027] (4)

[0028] Where, exp is a natural constant; Indicates that the predicted result category belongs to The probability of Total number of representative categories;

[0029] C. Output the new energy vehicle battery fault classification results for early warning.

[0030] DarkNet37 consists of a two-dimensional convolutional layer, one 208×208×64 residual convolutional layer, two 104×104×128 residual convolutional layers, five 52×52×256 residual convolutional layers, five 26×26×512 residual convolutional layers, and two 13×13×1024 residual convolutional layers.

[0031] DarkNet37-SE consists of a CBL layer, a 208×208×64 residual convolution layer, two 104×104×128 residual convolution layers, a SENet layer, five 52×52×256 residual convolution layers, a SENet layer, five 26×26×512 residual convolution layers, a SENet layer, two 13×13×1024 residual convolution layers, an average output layer, a fully connected layer, and a Softmax layer.

[0032] The present invention has the following advantages and effects compared to the prior art:

[0033] (1) Compared with other methods, the present invention constructs a deep learning network model for the classification of new energy vehicle power battery faults. This method does not require tedious manual feature extraction. The network model preset by the present invention learns a large amount of labeled data through supervised learning, and then substitutes the new energy vehicle battery data to be classified into the trained network model to output the classification results. Therefore, the method of the present invention can effectively classify new energy vehicle power battery faults, which is conducive to earlier detection of battery faults and preventing losses or even personal safety hazards caused by untimely detection of new energy vehicle battery faults.

[0034] (2) This paper improves DarkNet53 by reducing 16 convolutional layers to reduce the number of computational parameters. The channel attention mechanism is integrated into the improved DarkNet37 network. The channel attention mechanism module enables the network model to learn important information of interest from feature channels and ignore irrelevant information, thereby improving the classification accuracy and robustness of the network model.

[0035] (3) This paper aggregates ShuffleNet, DarkNet37, and the channel attention mechanism for analysis, and finally uses full connectivity and Softmax to output the fault classification probability. This method, which comprehensively considers multiple network models, can achieve higher classification accuracy while being lightweight. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is an overall flow chart of the method of the present invention.

[0037] Figure 2 This is the overall structure diagram of ShuffleDarkNet37-SE in the method of the present invention.

[0038] Figure 3 This is the overall structure diagram of DarkNet37-SE in the method of the present invention. DETAILED DESCRIPTION

[0039] The present invention proposes a ShuffleDarkNet37-SE new energy vehicle power battery fault classification method, which is described in detail with reference to the accompanying drawings as follows:

[0040] Figure 1 The following is a flowchart of the overall method of the present invention. First, data is acquired. Then, data preprocessing is performed. Data preprocessing includes data cleaning, data expansion, data normalization, and correlation analysis, and finally, the final image sample is generated. The dataset is then partitioned. The partitioned dataset is input into ShuffleNet and DarkNet37-SE. The outputs of ShuffleNet and DarkNet37-SE are then combined using a fully connected neural network. Finally, the prediction result is output.

[0041] Figure 2 This is the overall structure diagram of ShuffleDarkNet37-SE in the method of the present invention. ShuffleDarkNet37-SE consists of ShuffleNet, DarkNet37-SE and deep fully connected layers.

[0042] Figure 3 This is the overall structure of DarkNet37-SE in the method of the present invention. DarkNet37-SE consists of a CBL layer, one 208×208×64 residual convolution layer, two 104×104×128 residual convolution layers, an SENet layer, five 52×52×256 residual convolution layers, an SENet layer, five 26×26×512 residual convolution layers, an SENet layer, two 13×13×1024 residual convolution layers, an average output layer, a fully connected layer, and a softmax layer.

[0043] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A ShuffleDarkNet37-SE new energy vehicle power battery fault classification method, characterized in that: ShuffleNet, DarkNet37, and SENet are combined to classify power battery failures in new energy vehicles. The steps in use are as follows: A. After the vehicle sensor obtains the initial data of the new energy vehicle, pre-process the initial data of the new energy vehicle; The step A comprises: (1) Clean the original data and delete the meaningless data and outlier data in the original data; (2) Data expansion is performed using the linear interpolation method, which is: （1） Where, For newly generated sample data; and They are the two original sample data before and after respectively; The number of new data inserted into the original sample; The location generated for new data; (3) The maximum and minimum normalization method is used to normalize the data to eliminate the magnitude differences between the data of each dimension. The maximum and minimum normalization method is: （2） Where, is the normalized data; is the original data; is the maximum value of the column where the sample data is located; is the minimum value of the column where the sample data is located; (4) Use the Pearson correlation coefficient method to reduce the dimension of the data. The Pearson correlation coefficient method is: （3） Where, for 、 The correlation coefficient value of two columns of data; for 、 The covariance between and Respectively 、 respective standard deviations; (5) After reducing the data dimension, the data samples are converted into 40×40×1 images as the final input of the classification model; (6) Divide the data into training set and test set, with the ratio of training set to test set being 2:1; B. After data preprocessing, the data is input into the ShuffleDarkNet37-SE classification network; The classification network includes: (1) Reduce 16 convolutional layers from DarkNet53 and name the network DarkNet37; (2) Integrate the channel attention module into the DarkNet37 network to form the DarkNet37-SE network; (3) Combine the results of ShuffleNet and DarkNet37-SE through a fully connected layer network, and then output the classification probability through a normalized index to form the ShuffleDarkNet37-SE classification method; the normalized index is: （4） Where, exp is a natural constant; Indicates that the predicted result category belongs to The probability of Total number of representative categories; C. Output the new energy vehicle battery fault classification results for early warning.

2. The new energy vehicle power battery fault classification method of ShuffleDarkNet37-SE according to claim 1, characterized in that: DarkNet37 consists of a two-dimensional convolutional layer, one 208×208×64 residual convolutional layer, two 104×104×128 residual convolutional layers, five 52×52×256 residual convolutional layers, five 26×26×512 residual convolutional layers, and two 13×13×1024 residual convolutional layers.

3. The new energy vehicle power battery fault classification method of ShuffleDarkNet37-SE according to claim 1, characterized in that: DarkNet37-SE consists of a CBL layer, a 208×208×64 residual convolution layer, two 104×104×128 residual convolution layers, a SENet layer, five 52×52×256 residual convolution layers, a SENet layer, five 26×26×512 residual convolution layers, a SENet layer, two 13×13×1024 residual convolution layers, an average output layer, a fully connected layer, and a Softmax layer.

Images