Lithium ion battery internal short circuit detection method based on CAE-BPNN cooperation

Abstract

The invention relates to the technical field of lithium ion battery safety monitoring, in particular to a lithium ion battery internal short circuit detection method based on CAE-BPNN cooperation. The method comprises the steps of firstly collecting electrochemical impedance spectroscopy (EIS) data of different frequency points of the lithium ion battery for preprocessing and normalization, then unsupervised mining internal short circuit related deep nonlinear features through a convolutional auto-encoder (CAE), and inputting extracted low-dimensional effective features into a back propagation neural network (BPNN) to complete classification diagnosis. The convolutional auto-encoder (CAE) can efficiently remove redundancy, the back-propagation neural network (BPNN) has strong nonlinear mapping capability, and through the synergistic interaction of the convolutional auto-encoder and the back-propagation neural network, the method can effectively break through the limitation of the traditional technology, improve the sensitivity and generalization capability of early recognition of the tiny internal short circuit, and improve the accuracy of early recognition of the tiny internal short circuit. The method is suitable for safety monitoring of new energy automobile power batteries, energy storage power station battery packs, portable electronic equipment batteries and other scenes, the thermal runaway risk is reduced, and the safety and reliability of battery use are improved.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery safety monitoring technology, and particularly relates to a lithium-ion battery internal short-circuit detection technology. Its core is to accurately capture deep features related to internal short circuits in batteries through a collaborative mechanism of unsupervised feature mining and nonlinear classification. It is applicable to early fault identification and safety monitoring of various lithium-ion batteries, and can be widely used in scenarios such as power batteries for new energy vehicles, battery packs for energy storage power stations, and batteries for portable electronic devices. This provides technical support for improving battery safety, reducing the risk of thermal runaway, and ensuring the operational reliability of related equipment and systems. Background Technology

[0002] Lithium-ion batteries, with their high energy density and long cycle life, have become core energy supply components in new energy vehicles, energy storage power stations, and portable electronic devices. Their operational safety is directly related to the reliable operation of equipment and the safety of personnel and property. However, lithium-ion batteries are prone to internal short circuits under the influence of long-term charge-discharge cycles, extreme operating conditions, or manufacturing defects. The initial signal of this fault is weak but develops rapidly. If it is not detected in time, it can easily lead to thermal runaway, causing serious safety accidents such as fires and explosions. Therefore, early detection of internal short circuits is a key technical challenge in the field of battery safety monitoring.

[0003] Electrochemical impedance spectroscopy (EIS) is widely used for battery health monitoring because it can accurately reflect subtle changes in the internal electrochemical processes of batteries through multi-frequency response characteristics. However, traditional EIS-based detection methods have significant limitations: on the one hand, they rely heavily on manually designed features to process impedance data, which is not only highly subjective and inefficient, but also makes it difficult to capture deep nonlinear features related to internal short circuits and to fully explore the dynamic correlation patterns of impedance changes at different frequencies; on the other hand, existing methods often focus on feature extraction at single or a few frequency points, ignoring the synergistic information between multi-frequency responses, resulting in insufficient feature representation capabilities and thus affecting detection accuracy.

[0004] In recent years, the application of deep learning technology in the field of fault detection has been gradually advancing. Some detection methods based on neural networks have attempted to improve recognition performance, but they still face many bottlenecks: either the feature extraction process lacks specificity and fails to effectively remove redundant information and noise from EIS data; or the model structure design is unreasonable and fails to achieve synergistic optimization of feature redundancy removal and classification diagnosis, resulting in insufficient sensitivity to the identification of small internal short circuits, and poor generalization ability and insufficient adaptability when facing different types and batches of lithium-ion batteries and complex working environments.

[0005] In summary, existing internal short-circuit detection technologies for lithium-ion batteries generally suffer from core defects such as insufficient capture of deep nonlinear features, low sensitivity in identifying minute internal short circuits, and weak generalization ability, failing to meet the high precision and high reliability requirements of the new energy industry for battery safety monitoring. Therefore, developing a novel internal short-circuit detection method that can automatically mine deep effective features and balance detection sensitivity and generalization ability has significant technical value and practical application significance. Summary of the Invention

[0006] In view of this, the present invention provides a method for detecting internal short circuits in lithium-ion batteries based on CAE-BPNN collaboration. While effectively overcoming the limitations of traditional technologies and improving the early identification capability of micro internal short circuits, it achieves accurate and efficient detection of battery safety status and is suitable for monitoring scenarios in complex working environments.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for detecting internal short circuits in lithium-ion batteries based on CAE-BPNN collaboration includes the following steps:

[0009] S1: Collect electrochemical impedance spectroscopy (EIS) data of lithium-ion batteries at different frequency points;

[0010] S2: Preprocess and segment normalize the collected EIS data to generate standardized input data;

[0011] S3: Construct a 1D convolutional autoencoder (CAE) model, mine deep nonlinear features related to internal short circuits through unsupervised learning, and output effective features;

[0012] S4: Perform PCA dimensionality reduction on the features extracted by CAE to obtain low-dimensional feature vectors;

[0013] S5: Based on the features extracted by CAE after PCA dimensionality reduction, a backpropagation neural network (BPNN) classifier is constructed. The low-dimensional effective features extracted by CAE are input into the BPNN for training to build a complete detection model and realize the classification and diagnosis of battery status.

[0014] Furthermore, step S1 specifically involves: injecting a small-amplitude sinusoidal disturbance signal of voltage or current into the battery in normal operating condition, and collecting the battery's current or voltage response data, performing frequency domain transformation on the obtained data, and extracting the basic characteristics such as the real part of impedance, the imaginary part of impedance, the impedance magnitude, and the impedance phase.

[0015] Further, step S2 specifically involves: firstly, outlier processing is performed on the original EIS feature data, using the 3σ principle to ensure data validity. The mean μ and standard deviation σ of the data samples are calculated, and data exceeding the range [μ-3σ, μ+3σ] are identified as outliers. The median of the feature column is then used to replace the outliers. Logarithmic compression is then applied to the original EIS feature data, using the following compression formula:

[0016]

[0017] in, For the data after logarithmic processing, For the original data, It can be the real part, imaginary part, magnitude, and phase data of the original EIS features. This avoids logarithmic calculations resulting in infinity. Then, the data is segmented by frequency: data with frequencies less than 1 Hz are classified as low-frequency, data with frequencies greater than or equal to 1 Hz but less than or equal to 100 Hz as mid-frequency, and data greater than 100 Hz as high-frequency. Each frequency band is then independently normalized using the Min-Max normalization method, as shown in the following formula:

[0018]

[0019] in, For the first Normalized data, It is the first One set of data that needs to be processed. It is a set of electrochemical impedance data; the features corresponding to multiple frequency points are flattened and integrated to form a high-dimensional frequency response feature vector; after standardizing the feature vector and mapping it to a reasonable range, it is randomly divided into training set, validation set and test set according to proportion to obtain standardized feature data to meet the model training requirements; finally, the standardized feature data is converted into classification label format.

[0020] Furthermore, step S3 specifically involves: constructing a 1D CAE model, the model structure of which includes an input layer, an encoder, a decoder, and an output layer; mining deep nonlinear features related to internal short circuits through unsupervised learning to remove redundant data information and extract features.

[0021] Further, step S4 specifically involves: performing principal component analysis (PCA) based on the low-dimensional effective features of the training set output by CAE, calculating the principal component coefficients, feature scores, and the variance explained proportion of each principal component; setting a cumulative variance explained proportion threshold, and selecting the top... The minimum number of principal components that satisfy the cumulative variance explanation ratio threshold is used to obtain the key feature vectors after PCA dimensionality reduction. The PCA principal component coefficients obtained from the training set are used to perform dimensionality reduction transformation on the effective features of CAE in the training set, validation set, and test set, respectively.

[0022] Furthermore, step S5 specifically involves: constructing a BPNN classifier based on the features extracted by the CAE after PCA dimensionality reduction; during detection, the EIS data of the battery to be detected is preprocessed in step S2 and then input into the trained CAE model to extract features, and then the features are input into the BPNN classifier to determine the battery status based on the output results.

[0023] The beneficial effects of this invention are as follows: By deeply integrating EIS and CAE, this invention automatically mines deep nonlinear features related to internal short circuits using an unsupervised learning model, thus eliminating the dependence of traditional technologies on manually designed feature selection. It can comprehensively capture impedance variation patterns at different frequencies. At the same time, it constructs a complete collaborative technology link of CAE-BPNN, giving full play to the synergistic effect of CAE's efficient redundancy removal and BPNN's strong nonlinear mapping, effectively improving the detection sensitivity of small internal short circuit faults and the generalization ability across batches and operating conditions, thus making up for the shortcomings of existing technologies in early fault identification. Attached Figure Description

[0024] Figure 1 Flowchart of Internal Short Circuit Detection in Lithium-ion Batteries Based on CAE-BPNN Collaboration

[0025] Figure 2 Schematic diagram of convolutional autoencoder structure

[0026] Figure 3 Schematic diagram of backpropagation neural network structure Detailed Implementation

[0027] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following describes in detail the CAE-BPNN collaborative lithium-ion battery internal short circuit detection method of this invention, combined with specific experimental scenarios, equipment parameters, and operating procedures. Those skilled in the art can reproduce the technical effects of this invention based on this embodiment.

[0028] Please see Figure 1 The present invention provides a method for detecting internal short circuits in a lithium battery, comprising:

[0029] Step S101: Collect EIS data at different frequency points.

[0030] In one embodiment, a small-amplitude voltage or current disturbance signal less than 5% of the battery's rated voltage or current is injected, and the peak-to-peak value of the disturbance signal is set much lower than the battery's rated voltage; the battery's current or voltage response data is collected, and the obtained data is subjected to frequency domain transformation; the basic characteristics such as the real part of impedance, the imaginary part of impedance, the impedance magnitude, and the impedance phase are extracted respectively.

[0031] Step S102: Data preprocessing and normalization.

[0032] In one embodiment, step S102 involves outlier processing of the EIS signal. The 3σ principle is used to ensure data validity. The mean μ and standard deviation σ of the data samples are calculated, and data exceeding the range [μ-3σ, μ+3σ] are identified as outliers. The median of the feature column is used to replace the outliers. Piecewise normalization is then applied to logarithmically compress the original EIS feature data. The specific compression formula is as follows:

[0033]

[0034] in, For the data after logarithmic processing, For the original data, It can be the real part, imaginary part, magnitude, and phase data of the original EIS features. This avoids logarithmic calculations resulting in infinity. Then, the data is segmented by frequency: data with frequencies less than 1 Hz are classified as low-frequency, data with frequencies greater than or equal to 1 Hz but less than or equal to 100 Hz as mid-frequency, and data greater than 100 Hz as high-frequency. Each frequency band is then independently normalized using the Min-Max normalization method, as shown in the following formula:

[0035]

[0036] in, For the first Normalized data, It is the first One set of data that needs to be processed. It is a set of electrochemical impedance data; the features corresponding to multiple frequency points are flattened and integrated to form a high-dimensional frequency response feature vector; after standardizing the feature vector and mapping it to a reasonable range, it is randomly divided into training set, validation set and test set in a ratio of 7:1.5:1.5 to obtain standardized feature data, which are adapted to the model training requirements; finally, the standardized feature data is converted into a classification label format, where "0" indicates that the battery is in normal state and "1" indicates that the battery has an internal short circuit fault.

[0037] Step S103: Construct a convolutional autoencoder.

[0038] The schematic diagram of the convolutional autoencoder structure is shown below. Figure 2 As shown.

[0039] The convolutional autoencoder structure includes an input layer, an encoder, a decoder, and an output layer. In one embodiment, step S103 adopts a 1D convolutional autoencoder architecture, which consists of an input layer, an encoder, a decoder, and an output layer connected sequentially. The input layer is set as a sequence input layer with an input dimension of 1 and a minimum length of 122, used to adapt the flattened high-dimensional EIS feature vector. The encoder part consists of three consecutive stacked groups in the combination of "1D convolutional layer → batch normalization layer → LeakyReLU activation layer → Dropout layer". The first convolutional layer has a kernel size of 5×1 and a number of 64, the second convolutional layer has a kernel size of 3×1 and a number of 96, and the third convolutional layer has a kernel size of 3×1 and a number of 128. All convolutional layers have a stride of 1 and a padding method of "same". Then, a max pooling 1D layer is connected, with a pooling kernel size of 2×1 and a stride of 2, forming a complete feature extraction and dimensionality reduction chain. The decoder part is adapted to the encoder structure. It is stacked in three consecutive groups according to the combination of "transposed 1D convolutional layer → batch normalization layer → LeakyReLU activation layer". The first layer has a transposed convolutional kernel size of 3×1, a number of 96, a stride of 2, and a pruning method [0 1]. The second layer has a transposed convolutional kernel size of 3×1, a number of 64, a stride of 1, and a pruning method [1 1]. The third layer has a transposed convolutional kernel size of 5×1, a number of 1, a stride of 1, and a pruning method [2 2]. The last transposed convolutional layer is connected to the batch normalization layer and the Sigmoid activation layer. The output layer is set as a regression layer to complete the reconstruction of the input features.

[0040] Step S104: Extract low-dimensional effective features from EIS.

[0041] In one embodiment, step S104 in the encoder part sets up 3 convolutional layers and activation functions. The first convolutional layer has a kernel size of 5×1 and a number of 64 kernels; the second convolutional layer has a kernel size of 3×1 and a number of 96 kernels; and the third convolutional layer has a kernel size of 3×1 and a number of 128 kernels. All convolutional layers have a stride of 1 and a padding method of "same". The activation function used in all layers is LeakyReLU. Each convolutional layer is followed by a batch normalization layer and a dropout layer with a dropout probability of 0.1. Finally, a max pooling 1D layer is connected for downsampling to form the bottleneck layer feature output, thereby achieving feature dimensionality reduction. The feature mapping process can be represented by the following formula:

[0042]

[0043] in, For standardized high-dimensional input features, For encoder convolution kernel weights, For bias terms, This represents the convolution operation. It is a non-linear activation function. To output low-dimensional bottleneck features, three transposed convolutional layers and activation functions are set in the decoder part. The first transposed convolutional kernel size is 3×1, the number is 96, the stride is 2, and the pruning method is [0 1]. The second transposed convolutional kernel size is 3×1, the number is 64, the stride is 1, and the pruning method is [1 1]. The third transposed convolutional kernel size is 5×1, the number is 1, the stride is 1, and the pruning method is [2 2]. The activation function of all transposed convolutional layers is LeakyReLU. Each layer is followed by a batch normalization layer to reconstruct the bottleneck layer features and restore the key information of the input features. The formula for the reconstruction process is expressed as follows:

[0044]

[0045] in, The weights are the transposed convolutional kernel weights for the decoder. For bias terms, This indicates the transpose convolution operation. The high-dimensional features are reconstructed; in the output layer: with the optimization objective of minimizing the error between the input features and the reconstructed features, the L2 norm reconstruction loss function is adopted, and its formula is:

[0046]

[0047] in, For the sample size, For the first The original input features of each sample For the corresponding reconstructed features, Represents the L2 norm;

[0048] PCA dimensionality reduction is performed on the low-dimensional effective features extracted by CAE. Based on the low-dimensional effective features of the training set, principal component coefficients, feature scores, and the variance explained proportion of each principal component are calculated using principal component analysis. A cumulative variance explained proportion threshold of 90% is set, and the top [features] are selected. Principal components ( To meet the minimum number of principal components required to explain ≥90% of the cumulative variance, ensuring maximum retention of core discriminative information while eliminating redundancy, the PCA principal component coefficients obtained from the training set are used to perform dimensionality reduction transformation on the CAE low-dimensional effective features of the training, validation, and test sets, respectively, to obtain the key feature vectors after PCA dimensionality reduction, which are then used for the training and detection of the subsequent BPNN classifier.

[0049] Step S105: Construct a backpropagation neural network.

[0050] The schematic diagram of the backpropagation neural network structure is shown below. Figure 3 As shown.

[0051] The backpropagation neural network structure includes an input layer, a hidden layer, and an output layer. In one embodiment, step S105 adopts a stacked architecture of "input layer-hidden layer-output layer", with the input layer set as the feature input layer, whose input dimension is the same as that after PCA dimensionality reduction. The dimensional features are consistent. The hidden layers consist of three layers. The first hidden layer is stacked in the following order: fully connected layer (64 neurons) → batch normalization layer → Leaky ReLU activation layer (negative slope 0.2) → Dropout layer (dropout probability 0.4). The second hidden layer is stacked in the following order: fully connected layer (32 neurons) → batch normalization layer → Leaky ReLU activation layer (negative slope 0.2) → Dropout layer (dropout probability 0.3). The third hidden layer is stacked in the following order: fully connected layer (16 neurons) → batch normalization layer → Leaky ReLU activation layer (negative slope 0.2) → Dropout layer (dropout probability 0.2). The output layer is set in the following order: fully connected layer (2 neurons) → Softmax activation layer → classification layer.

[0052] Step S106: Fault identification.

[0053] In one embodiment, the low-dimensional features extracted in step S104 are used to form a feature input layer in the input layer, with the input dimension corresponding to the low-dimensional features extracted by PCA; output neurons are set in the output layer to correspond to the two types of battery states, and a binary classification task is achieved through an activation function.

Claims

1. A method for detecting internal short circuits in lithium-ion batteries based on CAE-BPNN collaboration, characterized in that, Includes the following steps: S1: Collect electrochemical impedance spectroscopy (EIS) data of lithium-ion batteries at different frequency points; S2: Preprocess and segment normalize the collected EIS data to generate standardized input data; S3: Construct a 1D convolutional autoencoder (CAE) model, mine deep nonlinear features related to internal short circuits through unsupervised learning, and output features; S4: Perform PCA dimensionality reduction on the features extracted by CAE to obtain low-dimensional feature vectors; S5: Based on the features extracted by CAE after PCA dimensionality reduction, a backpropagation neural network (BPNN) classifier is constructed. The low-dimensional effective features extracted by CAE are input into the BPNN for training to build a complete detection model and realize the classification and diagnosis of battery status.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: processing outliers in the original EIS feature data using the 3σ principle; and performing logarithmic compression on the original EIS feature data, with the specific compression formula being: in, For the data after logarithmic processing, For the original data, It can be the real part, imaginary part, magnitude, and phase data of the original EIS features. This avoids logarithmic calculations from reaching infinity.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: dividing the data into a low-frequency band (less than 1 Hz), a mid-frequency band (1 Hz to 100 Hz), and a high-frequency band (greater than 100 Hz) according to frequency, and performing Min-Max normalization on the data of each frequency band.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: constructing a 1D CAE model, inputting the flattened standardized feature vector, implementing feature downsampling through a pooling layer, and extracting the features output by the encoder.

5. The method according to claim 1, characterized in that, Step S5 specifically includes: constructing a BPNN classifier; processing the EIS data of the battery to be detected in step S2, inputting it into the trained CAE model to extract features, obtaining low-dimensional features through PCA dimensionality reduction, inputting the low-dimensional features into the trained BPNN, and outputting the detection results of whether the battery is normal or internally short-circuited.

Images