Power battery pack safety protection method, device and computer equipment

Abstract

The present invention relates to the technical field of battery safety protection, and discloses a safety protection method, device and computer equipment for a power battery pack, wherein the method comprises: collecting voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix; performing incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector; performing feature dependency calculation on the single cell feature vector to obtain a global feature vector of the battery pack; performing local and global fault detection and health score calculation to obtain a probability distribution of battery pack fault types and a single cell health status score; and limiting and adjusting the charge and discharge parameters of the power battery pack to obtain a charge and discharge protection control instruction. The method effectively solves the problem that the traditional method ignores the interaction between batteries, and improves the accuracy and effectiveness of the protection measures.

Description

Technical Field

[0001] The present invention relates to the technical field of battery safety protection, and in particular to a safety protection method, device and computer equipment for a power battery pack. Background Art

[0002] The power battery pack is composed of multiple single cells connected in series and parallel. There are inherent inconsistencies between the single cells. The performance degradation rate and mode of each single cell during the charging and discharging process are also different, which makes it difficult to accurately grasp and effectively monitor the internal state of the battery pack.

[0003] Current battery pack safety protection methods mainly rely on threshold judgments of voltage, current, temperature and other characterization parameters, and lack in-depth analysis of the interaction between individual cells within the battery pack. Traditional protection strategies based on a single parameter or simple rules cannot effectively identify and prevent potential safety hazards, especially under fast charging and discharging conditions, which are prone to local overcharging, over-discharging and other dangerous situations. Summary of the invention

[0004] The present invention provides a safety protection method, device and computer equipment for a power battery pack, which effectively solves the problem that traditional methods ignore the interaction between batteries and improves the accuracy and effectiveness of protection measures.

[0005] In a first aspect, the present invention provides a safety protection method for a power battery pack, the safety protection method for a power battery pack comprising:

[0006] Collect voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix;

[0007] Performing incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector;

[0008] Calculating the characteristic dependency relationship of the single cell characteristic vector to obtain a global characteristic vector of the battery pack;

[0009] Performing local and global fault detection and health score calculation on the single cell feature vector and the battery pack global feature vector to obtain a probability distribution of battery pack fault types and a single cell health status score;

[0010] According to the probability distribution of the battery pack fault type and the single cell health status score, the charge and discharge parameters of the power battery pack are restricted and adjusted to obtain a charge and discharge protection control instruction.

[0011] In a second aspect, the present invention provides a safety protection device for a power battery pack, the safety protection device for the power battery pack comprising:

[0012] The data acquisition module is used to collect voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix;

[0013] A feature extraction module, used for performing incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector;

[0014] A relationship calculation module, used for calculating the characteristic dependency relationship of the single cell characteristic vector to obtain a global characteristic vector of the battery pack;

[0015] A status scoring module, used to perform local and global fault detection and health score calculation on the single cell feature vector and the battery pack global feature vector, to obtain a probability distribution of battery pack fault types and a single cell health status score;

[0016] The restriction adjustment module is used to restrict and adjust the charge and discharge parameters of the power battery pack according to the probability distribution of the battery pack fault type and the single cell health status score to obtain a charge and discharge protection control instruction.

[0017] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned power battery pack safety protection method.

[0018] In the technical solution provided by the present invention, the internal state of the battery pack is accurately characterized by extracting multi-scale features of the voltage data of the single cell and analyzing the incremental capacity curve, thereby improving the accuracy of state monitoring. The feature fusion network based on the attention mechanism can adaptively capture the dependencies between the single cells, effectively solving the problem that the traditional method ignores the interaction between the cells. The deep learning is innovatively combined with the electrochemical feature analysis to construct an end-to-end fault diagnosis model, which greatly improves the accuracy and real-time performance of fault detection. The gradient-based interpretability analysis method is introduced to intuitively display the key areas where the fault occurs through the heat map, making the fault diagnosis results interpretable. A multi-level adaptive charge and discharge protection strategy is designed to dynamically adjust the control parameters according to the health status and fault risk level of the single cell, significantly improving the accuracy and effectiveness of the protection measures. The Bayesian optimization framework is used for online parameter tuning, which takes into account the charging efficiency while ensuring safety, and achieves a good balance between safety and performance.

[0019] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of an embodiment of a safety protection method for a power battery pack in an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of an embodiment of a safety protection device for a power battery pack in an embodiment of the present invention;

[0023] Figure 3 FIG. 1 is a schematic diagram of an embodiment of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0026] To facilitate understanding of this embodiment, a safety protection method for a power battery pack disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0027] 101. Collect voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix;

[0028] It is understandable that the execution subject of the present invention may be a safety protection device of a power battery pack, or may be a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0029] Specifically, a voltage acquisition module is provided in the power battery pack, and the module includes a high-precision analog-to-digital converter (ADC), a data buffer storage unit, and a data transmission interface. In order to ensure the real-time and accuracy of voltage data acquisition, the voltage acquisition module measures the voltage of each single cell in the power battery pack according to a preset sampling frequency. The sampling frequency comprehensively considers the power battery operating conditions, battery charging and discharging characteristics, and data storage and transmission capabilities, and ensures that the sampling interval is short enough to capture the dynamic changes of the battery voltage, while avoiding data redundancy leading to excessive storage and calculation burdens. Through this step, the real-time voltage value of each single cell at different time points is obtained to form a basic voltage data set. The data is arranged in time sequence according to the acquisition time to construct the voltage sampling sequence of each single cell. The voltage measurement value of each single cell corresponds to a specific sampling time, and the measurement values ​​of all single cells are arranged in time order to obtain a voltage sampling sequence. The sampled data is synchronously processed to ensure that the voltage data of all single cells are analyzed under the same time reference. The voltage sampling sequence of each single cell is synchronously time-stamped to form a voltage data sequence with a timestamp. The timestamp is set based on a unified clock signal, such as using the clock source provided by the battery management system or the GPS clock synchronization mechanism to ensure that the voltage data of all batteries are aligned with a unified time axis. The timestamped voltage data sequence is arranged in a matrix according to the serial number of the single cell to construct an initial voltage data matrix. The rows of the matrix correspond to different time points, while the columns correspond to different single cells. The elements in the matrix are the voltage values ​​of a single cell at a certain time point. In this way, the voltage change of the power battery pack over a period of time is represented. The initial voltage data matrix is ​​segmented and stored to obtain a segmented voltage matrix, and the segmented data is transmitted and integrated through the CAN bus to form the final single cell voltage data matrix. The segmented storage method is divided according to the time window, for example, the data is stored at a fixed time interval, and the data of different time periods are stored in different storage units. At the same time, the CAN bus is used for data transmission to ensure the stability and reliability of data transmission. Since the CAN bus has strong anti-interference ability and supports multi-node communication, it is suitable for complex systems such as battery packs that contain a large number of single cells. During the data transmission process, each segmented voltage matrix is ​​integrated in chronological order to restore the complete single cell voltage data matrix, and finally form the power battery voltage monitoring data.

[0030] 102. Perform incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector;

[0031] Specifically, the voltage values ​​in the voltage data matrix of the single cell are divided into multiple voltage interval sequences according to preset intervals. The division is based on the battery operating voltage range, and the appropriate partition accuracy is selected to ensure that the voltage change is fully captured. In each voltage interval, the capacity integral value in the interval is calculated, and the capacity integral values ​​of all voltage intervals are accumulated to obtain the charge and discharge capacity sequence of each voltage interval. The capacity change and voltage change of adjacent voltage intervals in the charge and discharge capacity sequence are ratio-calculated, that is, the incremental capacity value is calculated to characterize the change trend of the battery in different voltage intervals. The data is smoothed by the third-order spline interpolation method to ensure the continuity and resolvability of the incremental capacity curve. The third-order spline interpolation approximates the incremental capacity data by constructing a piecewise cubic polynomial, so that the curve has better smoothness and stability at each sampling point, thereby reducing the error caused by the discreteness of the data, and obtaining an incremental capacity curve data that is more in line with the actual operating characteristics of the battery. The incremental capacity curve data is normalized, and the normalized curve data is mapped to a pixel matrix to generate an incremental capacity curve image. The incremental capacity curve image is input into the image feature extraction network, which contains 5 convolution blocks, each of which consists of 3 convolution layers with 3×3 convolution kernels and 1 ReLU activation function layer. The role of the convolution layer is to extract local features. The small 3×3 receptive field effectively captures the slight changes in the incremental capacity curve image, while the ReLU activation function layer improves the nonlinear expression ability of the model, so that the network can better learn the incremental capacity features under different battery states. Under the layer-by-layer processing of these 5 convolution blocks, the network extracts 5 groups of convolution feature maps of different scales. Each group of feature maps represents the spatial information of the incremental capacity curve at different scales. The lower-level convolution blocks capture local features, while the higher-level convolution blocks gradually aggregate global information, making the final feature expression more comprehensive. The 5 groups of convolution feature maps of different scales are reduced in dimension to eliminate redundant information and improve computational efficiency, and a multi-scale feature fusion matrix is ​​obtained. The dimensionality reduction method uses global average pooling or principal component analysis to ensure that the data after dimensionality reduction can still maintain complete feature expression. After the fused feature matrix is ​​obtained, it is linearly mapped through three layers of fully connected layers to extract and optimize the feature vector of the single battery. The function of the fully connected layer is to map the high-dimensional feature space to a lower-dimensional feature space to ensure that the final output feature vector has good discrimination and representation capabilities, so that batteries with different health states or failure modes show obvious distribution differences in the feature space. Through the above steps, the feature vector of the single battery is obtained.

[0032] 103. Calculate the characteristic dependency relationship of the single battery characteristic vector to obtain the global characteristic vector of the battery pack;

[0033] Specifically, the feature vectors of the single cells are transformed and mapped to different representation spaces, and query vectors, key vectors and value vectors are generated respectively, so that different single cells can establish connections through specific calculation methods, thereby capturing the global feature relationship within the battery pack. The query vector is used to measure the importance of a single cell in the entire battery pack, while the key vector is used to calculate the degree of association between different batteries, and the value vector is used for the final information aggregation and feature fusion. The query vector and the key vector are dot-producted and numerically normalized using a scaling factor to calculate the similarity score matrix between the single cells. The core function of the dot product operation is to measure the similarity of different single cells in the feature space, that is, if the features of two single cells are relatively close, their dot product values ​​are larger, and if there are obvious differences in the features, the dot product values ​​are smaller. The dot product operation results are normalized using a scaling factor to ensure that the calculated similarity score matrix remains stable in value, while preventing the calculation accuracy from decreasing due to the large value range in subsequent calculations. After the similarity calculation is completed, the similarity score matrix is ​​calculated by the Softmax function to normalize the weights between different single cells and make the sum of the similarity scores of all batteries equal to 1. The calculation of Softmax effectively quantifies the relationship between different single cells and ensures that the feature information is reasonably distributed according to the weight. The similarity score matrix normalized by Softmax is matrix multiplied with the value vector to obtain the initial feature fusion result that considers the relationship between the single cells. The initial feature fusion result is input into a feedforward network composed of four fully connected layers for nonlinear feature transformation. The feedforward network enhances the feature expression ability through a series of nonlinear transformations, so that the model can better capture the complex relationship of the battery pack. Through layer-by-layer full connection operations, redundant information is effectively removed and key features are highlighted, so that the final backbone feature vector reflects the state of the entire battery pack. The backbone feature vector is divided to describe the characteristics of the battery pack in a more fine-grained manner. The backbone feature vector is divided into N sub-feature vectors, and the N sub-feature vectors are parallelized for feature extraction. Different attributes of the battery pack are learned through different feature subspaces, thereby improving the expression ability of the model. In this process, each sub-feature vector represents a certain aspect of the battery pack, such as charge and discharge consistency, health status difference, or anomaly detection indicators, and through parallel processing, the efficiency of feature extraction is improved, making the final calculation more efficient and stable. After completing the parallel feature extraction, the N groups of independent attention features are spliced ​​and converted into a unified feature space through linear mapping, so that all sub-features are fused in the same dimension. Feature enhancement is performed through adaptive layer normalization to eliminate the scale differences between different features and improve the model's adaptability to different battery packs.The global feature vector of the battery pack is obtained, which is used for subsequent applications such as fault detection, health assessment, and charge and discharge strategy optimization.

[0034] 104. Perform local and global fault detection and health score calculation on the single cell feature vector and the battery pack global feature vector to obtain the battery pack fault type probability distribution and the single cell health status score;

[0035] Specifically, the single cell feature vector and the global feature vector of the battery pack are cross-attention weighted fused. Through the cross-attention mechanism, the features of the single cell are not limited to their own state, but can be connected with the state of the entire battery pack, so as to extract feature information with more global dependency. In the cross-attention calculation process, the single cell feature vector is used as the query vector, and the global feature vector is used as the key vector and value vector. The weight is obtained through dot product calculation, and combined with the Softmax normalization operation, the fused features can dynamically pay attention to the state changes of the entire battery pack. After the cross-attention weighted fusion, the conditional batch normalization layer is introduced. This normalization method adaptively adjusts the feature distribution according to different battery operating conditions to ensure that the fused features can maintain good expression ability under different battery pack states, and obtain the initial fused features containing local-global feature dependencies. The initial fused features are input into three parallel fully connected layers for multi-scale feature extraction, so that network layers of different scales can capture information at different levels respectively, so that the low-level network extracts fine-grained features, such as subtle anomalies of single cells, while the high-level network focuses on the overall trend of the battery pack. Through three-way parallel calculation, a set of multi-scale feature groups are obtained, which reflect the battery health status information at different levels. Adaptive feature aggregation is performed on the multi-scale feature groups, so that the features of different scales are combined according to the optimal weight to form the final aggregated feature vector. Adaptive feature aggregation is completed through the attention mechanism or weighted summation, so that the contribution of the features is dynamically adjusted with the different battery states, ensuring that the final feature vector can retain key information without being interfered by redundant features. Considering the time dependency of the data during the operation of the battery pack, the aggregated feature vector is input into the time series feature extraction layer composed of the bidirectional gated recurrent unit (BiGRU), and the time series dependency features are extracted in the forward and reverse sequences respectively. The introduction of BiGRU can effectively capture the changing trend of battery features in the time dimension, so that the model not only pays attention to the battery status at the current moment, but also combines the operation data of the past period of time to infer the future health change trend. In the calculation process of BiGRU, the forward sequence can extract historical information, while the reverse sequence focuses on future information. Through the two-way information fusion, the final time series features can more accurately describe the dynamic changes of the battery operation status. The extracted dynamic features are screened through the attention gating mechanism and recalibrated in combination with the channel attention module. The function of the attention gating mechanism is to screen out the most critical features for battery fault detection and health assessment, and to weaken the interference of noise or irrelevant information. The channel attention module calculates the importance weights of different feature channels, so that important features occupy a higher weight in the final representation, ensuring that the final feature vector has a strong discriminative ability. The discriminative features are input into the fault classification head and the health scoring head respectively to calculate the probability distribution of the fault type of the battery pack and the health status score of the single cell.The fault classification head classifies different types of faults based on discriminant features, so the Softmax classifier is used to calculate the probability distribution of different faults in the battery pack and provide fault diagnosis results. The health scoring head uses a quantitative perception network structure to output the health status score of the single battery. The characteristic of this network structure is that through quantitative mapping, the scoring result is more stable and accurately reflects the true health status of the battery.

[0036] After the fault detection is completed, the back propagation algorithm is used to calculate the gradient of the image feature extraction network. In particular, for the feature map of the last convolutional layer in the network, there is a nonlinear relationship between the response value at each position (i, j) and the probability of the fault type. Therefore, not only the first-order gradient change trend is calculated, but also the second-order derivative and the third-order derivative are needed to more accurately capture the influence of the fault type on the feature maps at different positions. By weighted combination of high-order gradient information, the importance coefficient of each position is obtained to quantify the contribution of different regions in fault identification. The position importance coefficient is weighted summed with the feature map activation value of the corresponding position to generate the category response feature map of each channel, highlighting the image area that is highly related to the specific fault type. The category response feature map is multi-scale fused with the shallow feature map through residual connection, and combined with the adaptive weight calculation mechanism to obtain a multi-level thermal response map. The shallow feature map contains finer-grained spatial information, while the category response feature map mainly captures deep semantic information. The residual connection effectively combines the advantages of the two, and the adaptive weight calculation ensures that the information transfer between features at different levels can be dynamically adjusted, thereby enhancing the model's ability to focus on the fault area. The multi-level thermal response map is upsampled by bicubic interpolation to adjust it to the same size as the incremental capacity curve image, ensuring that subsequent analysis can be performed on a unified spatial scale. The edge details are optimized by local contrast enhancement technology to improve the clarity of the fault area and obtain the original scale thermal map. The original scale thermal map is divided into two sub-regions, the charging process and the discharging process, according to the voltage range. This division method can effectively separate the fault characteristics under different operating modes and improve the pertinence of fault detection. Adaptive threshold segmentation is performed on each sub-region to extract the most significant fault response area, so that the activation map is more focused on the key area where the fault occurs. Adaptive threshold segmentation dynamically adjusts the threshold according to the pixel distribution of the thermal map to ensure that a stable area extraction effect can be obtained under different fault types. Morphological processing and connected domain analysis are performed on the hot spots in the regional fault activation map to optimize the extraction results of fault features. Morphological processing includes operations such as dilation and corrosion to remove noise and enhance the structural characteristics of the fault area, while connected domain analysis is used to detect and annotate the spatial distribution characteristics and intensity change trends of different fault areas to ensure that the fault area feature map describes the scope and extent of the fault. Adaptive color mapping is performed on the fault area feature map to convert the single-channel grayscale information into a three-channel pseudo-color image in order to distinguish the response degree of different fault areas. At the same time, in order to improve the visualization effect, the transparency channel is superimposed so that the transparency of the high-activation area is lower and the transparency of the low-activation area is higher, highlighting the most critical fault attention area. Combined with the background information of the incremental capacity curve image, a visualization result with a heat map of the fault attention area is generated, and the heat map of the fault attention area of ​​the incremental capacity curve image is obtained.

[0037] 105. According to the probability distribution of battery pack fault types and the health status score of the single battery, the charge and discharge parameters of the power battery pack are restricted and adjusted to obtain the charge and discharge protection control instructions.

[0038] Specifically, the health status score of the single cell is graded by multiple thresholds, and the score interval is divided into a high-risk interval, a warning interval, and a safe interval to determine the health level label of the single cell. This grading method is optimized based on a large amount of historical data, so that batteries in different health states can adopt different protection strategies during the charging and discharging process. According to the heat map of the fault attention area calculated by the probability distribution of the battery pack fault type, the integral intensity of the charging section and the discharging section is calculated to determine the probability of fault occurrence and the degree of impact at different stages. The heat map of the fault attention area is obtained by the aforementioned gradient back propagation method, which reflects the location and severity of the fault occurrence of the battery pack under different working conditions. During the charging and discharging cycle, the high response area of ​​the heat map is integrated and calculated to obtain the integral intensity value of the charging and discharging process, and this intensity value is used to determine the target stage of the fault occurrence and the danger level of the fault. If the fault integral intensity of a battery in the charging stage is high, it indicates that its main problem is related to overcharging, abnormal internal resistance, or enhanced polarization effect, while if the integral intensity in the discharge stage is high, it means that there is a short circuit, internal resistance mutation, or irreversible capacity loss. Through this calculation process, the risk assessment result of the charging and discharging process is formed. A multi-objective decision matrix is ​​constructed based on the health level label and risk assessment results, and the weight coefficient of each control parameter is calculated using the hierarchical analysis method to form a hierarchical protection strategy. The role of the hierarchical analysis method is to quantify the relative importance of different influencing factors so that the control parameters take into account both safety and service life. For example, if the health status score of a battery is low, but its charge and discharge failure risk is small, there is no need to adopt overly strict charge and discharge restrictions. If the health score of a battery is still in the warning range but the fault integral intensity of its discharge process is high, its discharge rate should be limited first. The hierarchical analysis method can formulate the optimal control strategy for batteries of different health levels based on a comprehensive consideration of health status and failure risk. On this basis, strict protection measures are taken for single cells in the high-risk range to reduce their charge rate to the first percentage of the nominal value, and at the same time reduce the discharge rate to the second percentage of the nominal value to slow down the further degradation of the battery. In order to prevent overcharging or over-discharging from causing additional damage to high-risk batteries, the charging cut-off voltage is reduced by the first gradient value, and the discharging cut-off voltage is increased by the second gradient value to narrow its operating voltage range, thereby reducing the probability of loss and abnormal electrochemical reactions, and obtaining high-risk control parameters. For single cells in the warning range, relatively mild protection measures are taken. Therefore, combined with the risk section in the heat map of the fault focus area, the charging and discharging rate is dynamically adjusted within the nominal value range to ensure that it can still meet certain power requirements while avoiding serious faults. In the warning range, the charging and discharging cut-off voltage is adjusted to the third gradient value to optimize its performance while ensuring the battery life, and obtain the control parameters of the warning level.The high-risk level control parameters and the warning level control parameters are integrated, and the control parameters are fine-tuned online through the Bayesian optimization framework. In the Bayesian optimization process, an objective function is constructed to comprehensively consider the cycle life, charge and discharge safety, and charging time constraints of the battery pack, and the Gaussian process regression model is used to predict the effects of different control parameters. In each optimization iteration, the Bayesian optimization algorithm selects the current optimal parameter combination and conducts experimental evaluation. If the new parameter combination can improve the overall efficiency of the system, it is updated to the new optimal solution. In this way, the optimal charge and discharge control strategy of the power battery pack is realized while ensuring safety, and finally the charge and discharge protection control instructions are generated to ensure that the power battery pack can maintain a stable operating state throughout its life cycle, while extending its service life and reducing the occurrence of safety accidents.

[0039] In the embodiment of the present invention, the internal state of the battery pack is accurately characterized by extracting multi-scale features of the voltage data of the single cell and analyzing the incremental capacity curve, thereby improving the accuracy of state monitoring. The feature fusion network based on the attention mechanism can adaptively capture the dependencies between the single cells, effectively solving the problem that the traditional method ignores the interaction between the cells. The deep learning is innovatively combined with the electrochemical feature analysis to construct an end-to-end fault diagnosis model, which greatly improves the accuracy and real-time performance of fault detection. The gradient-based interpretability analysis method is introduced to intuitively display the key areas where the fault occurs through the heat map, making the fault diagnosis results interpretable. A multi-level adaptive charge and discharge protection strategy is designed to dynamically adjust the control parameters according to the health status and fault risk level of the single cell, significantly improving the accuracy and effectiveness of the protection measures. The Bayesian optimization framework is used for online parameter tuning, which takes into account the charging efficiency while ensuring safety, and achieves a good balance between safety and performance.

[0040] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0041] A voltage acquisition module is set for each single cell in the power battery pack to collect voltage data at a preset sampling frequency to obtain the real-time voltage value of each single cell;

[0042] Arrange the real-time voltage values ​​of each single cell in time sequence according to the acquisition time to obtain the voltage sampling sequence of each single cell;

[0043] Performing synchronous time stamping on the voltage sampling sequence of each single cell to obtain a voltage data sequence with a time stamp;

[0044] Arrange the voltage data sequence with time stamps in a matrix according to the sequence number of the single battery cells to obtain an initial voltage data matrix;

[0045] The initial voltage data matrix is ​​stored in segments to obtain a segmented voltage matrix, and the segmented voltage matrix is ​​data-transmitted and integrated via a CAN bus to obtain a single cell voltage data matrix.

[0046] Specifically, a voltage acquisition module is configured for each single cell in the battery pack. The module consists of a high-precision analog-to-digital converter (ADC), a data buffer storage unit, a time synchronization circuit, and a communication interface, and performs voltage data acquisition at a preset sampling frequency to ensure continuous monitoring of the battery operating status. The sampling frequency is set to , then each single cell has a The voltage acquisition value at is expressed as:

[0047]

[0048] in, Representative The single cell is The voltage value at the sampling moment, (in is the sampling number), Indicates that the voltage acquisition module is at time About battery The measured value obtained by sampling. The collected voltage data is arranged in time sequence according to the collection time to construct the voltage sampling sequence of the single cell. Since the sampling of different single cells will be affected by communication delay or sensor response time, the data is synchronized in a unified time. Assume that the data sampling time of all single cells forms a time series:

[0049]

[0050] Then the voltage sampling sequence of each single cell is expressed as:

[0051]

[0052] in, Representative The voltage data set of each single cell is a collection of sequential voltage data. In order to improve the timing consistency of the data, the voltage sampling sequence of each single cell is synchronized with the time stamp so that all data have the same time base time stamp. Assuming that the global clock signal is provided by a high-precision clock module, each data point Need to match a timestamp , then the voltage data sequence with timestamp is expressed as:

[0053]

[0054] in, For the The voltage data with timestamps of the single cells are arranged in a matrix according to the cell numbers to form an initial voltage data matrix. The rows of the matrix correspond to the sampling time series, and the columns correspond to different single cells. The elements in the matrix are the voltage values ​​of each cell at different time points, so it can be expressed as:

[0055]

[0056] Each column represents the voltage change of a single cell during the entire sampling period, and each row represents the voltage data of all single cells at a certain moment. The initial voltage data matrix is ​​stored in segments to optimize data management and computational efficiency. Assume that the storage interval is , then the number of sampling points contained in each segment is:

[0057]

[0058] in, is the number of segments. Then the segmented voltage matrix obtained by segmented storage is expressed as:

[0059]

[0060] in, For the The voltage data matrix of the segment, The segmented voltage matrix is ​​transmitted and integrated through the CAN bus to finally form a complete single cell voltage data matrix. The CAN bus has high anti-interference ability and real-time performance, and is suitable for multi-node communication of the battery management system. During the data transmission process, each segment matrix As a data packet is sent sequentially, and the data is reconstructed at the receiving end to restore the complete voltage data matrix:

[0061]

[0062] A complete single cell voltage data matrix is ​​formed, which is used for subsequent battery health analysis, fault detection and energy management.

[0063] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0064] The voltage values ​​in the single cell voltage data matrix are divided into a voltage interval sequence according to a preset interval, and the capacity integral value of each voltage interval in the voltage interval sequence is accumulated and calculated to obtain a charge and discharge capacity sequence of each voltage interval;

[0065] The capacity change and voltage change in adjacent voltage intervals in the charge and discharge capacity sequence are ratio-calculated, and the incremental capacity curve data is obtained through third-order spline interpolation.

[0066] Normalizing and pixel matrix mapping the incremental capacity curve data to obtain an incremental capacity curve image;

[0067] The incremental capacity curve image is input into the five convolution blocks of the image feature extraction network. Each convolution block contains three convolution layers with 3×3 convolution kernels and one ReLU activation function layer, and five groups of convolution feature maps of different scales are obtained.

[0068] The dimensionality of 5 groups of convolutional feature maps of different scales is reduced to obtain a multi-scale feature fusion matrix, which is then linearly mapped through three fully connected layers to obtain the single cell feature vector.

[0069] Specifically, the voltage data is partitioned and the voltage values ​​of the single cells are divided into voltage interval sequences according to preset intervals. The voltage change of the battery is discretized for subsequent capacity integral calculation. Suppose the voltage range of a single cell in the power battery pack is , assuming the number of voltage intervals is , then the width of each voltage interval is expressed as:

[0070]

[0071] in, Represents the size of the voltage interval. For each single cell, The charge and discharge capacity is calculated by Coulomb integration:

[0072]

[0073] in, Representative The capacity of the voltage range, Represents the charge and discharge current of the battery, is the time node of this voltage interval. By accumulating the capacity integral values ​​of all voltage intervals, the complete charge and discharge capacity sequence is obtained:

[0074]

[0075] This sequence reflects the charge and discharge characteristics of the battery at different voltage levels. After obtaining the charge and discharge capacity sequence, calculate the capacity change in adjacent voltage intervals. The voltage change The ratio of , get the incremental capacity data:

[0076]

[0077] in, Represents the incremental capacity curve data, describing the charge and discharge characteristics of the battery in different voltage ranges. Since the actual measured data is discrete, third-order spline interpolation is used for smoothing to obtain a more continuous incremental capacity curve. Suppose the third-order spline interpolation function is:

[0078]

[0079] in, is the incremental capacity curve obtained by fitting, is the interpolation coefficient, which is obtained by solving the boundary conditions and interpolation point constraints. Through this step, the incremental capacity data is converted into a smooth incremental capacity curve for subsequent feature extraction. The incremental capacity data is normalized to ensure that the features of different batteries have a uniform value range. The maximum value of the incremental capacity curve is set to , the minimum value is , then the normalization process is expressed as:

[0080]

[0081] The normalized data is mapped into a pixel matrix to generate an incremental capacity curve image. Assume that the image resolution is , then the pixel mapping function is expressed as:

[0082]

[0083] in, Represents pixel The grayscale value at , the value range is between [0,255]. The incremental capacity curve image is input into a deep convolutional neural network containing five convolutional blocks for feature extraction. Each convolutional block contains three layers Convolutional layer of convolution kernel and a ReLU activation function layer. Assume the input image After the convolution layer The output feature map after processing is , then the convolution operation is expressed as:

[0084]

[0085] in, is the convolution kernel, is the bias term, is the ReLU activation function, and * represents the convolution operation. Through the layer-by-layer extraction of five convolution blocks, five groups of convolution feature maps of different scales are obtained, and each group of feature maps represents the incremental capacity features of different levels. The five groups of convolution feature maps of different scales are reduced in dimension to reduce the computational overhead and extract key features. The dimension reduction method uses global average pooling, that is:

[0086]

[0087] in, and is the height and width of the feature map, Represents the multi-scale feature fusion matrix after dimensionality reduction. The multi-scale feature fusion matrix is ​​input into a neural network composed of three fully connected layers for linear mapping to obtain the single cell feature vector. Assume that the weight of the fully connected layer is , the bias term is , then the calculation formula of the eigenvector is:

[0088]

[0089] in, Represents the final single cell feature vector, which is used for subsequent battery health assessment and fault detection.

[0090] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0091] Perform feature transformation on the single battery feature vector to generate a query vector, a key vector and a value vector respectively;

[0092] Perform dot product operation on the query vector and the key vector, and normalize the values ​​by the scaling factor to obtain the similarity score matrix between the cells;

[0093] The similarity score matrix is ​​calculated by the Softmax function, and the matrix multiplication operation is performed with the value vector to obtain the initial feature fusion result that takes into account the relationship between the single cells;

[0094] The initial feature fusion result is input into a feed-forward network consisting of four fully connected layers for nonlinear feature transformation to obtain the backbone feature vector;

[0095] The backbone feature vector is divided into N sub-feature vectors, and the N sub-feature vectors are subjected to parallel feature extraction to obtain N groups of independent attention features.

[0096] N groups of independent attention features are concatenated and linearly mapped, and feature enhancement is performed through adaptive layer normalization to obtain the global feature vector of the battery pack.

[0097] Specifically, the feature vector of each single battery is mapped to generate a query vector, a key vector, and a value vector. Assume that the power battery pack contains The characteristic vector of each single cell is expressed as ,in represents the dimension of the feature vector, and these features are projected into the query key and value space through linear transformation:

[0098]

[0099] in, is the mapping matrix, Represents the feature dimension after mapping, select To reduce the computational complexity and make the calculation more efficient. The features of each single battery are converted into new query, key and value representations to calculate the correlation between batteries. By calculating the query vector With key vector The dot product between them gives the similarity score matrix between the monomer batteries:

[0100]

[0101] in, Representative The battery and The similarity of the batteries, is a scaling factor to prevent the dot product result from being too large and affecting the stability of the calculation. The final similarity score matrix The size is , which is used to describe the relationship between the internal characteristics of the battery pack. Perform exponential normalization to obtain the normalized weight matrix:

[0102]

[0103] in, Representative Battery for the first The attention weights of the batteries, and the sum of all attention weights is 1. With value vector Perform matrix multiplication to obtain fused features:

[0104]

[0105] in, It is The fusion features of the battery not only contain its own information, but also combine the relevant features of other batteries to form a representation of global feature dependence. The fusion features are transformed nonlinearly to extract deeper feature information. It is fed into a feedforward calculation structure consisting of four fully connected layers. The calculation formula of each layer is as follows:

[0106]

[0107] in, and They are The weight matrices and biases of the layers, represents a nonlinear activation function, using piecewise linear transformation. After 4 layers of calculation, the backbone feature vector is obtained , this vector contains the global feature information of the single battery. To obtain a more fine-grained sub-feature vector. conduct Group partitioning, the dimension of each sub-feature vector is:

[0108]

[0109] And expressed as:

[0110]

[0111] in, Representative The battery Group characteristics. The sub-features are calculated independently to extract feature information of different scales. Adaptive transformation via separate computational modules:

[0112]

[0113] in, Represents the weight matrix used to calculate attention. After calculation, each sub-feature has independent information representation. The independent features are concatenated to obtain the final global representation:

[0114]

[0115] Using the mapping matrix To do the conversion:

[0116]

[0117] in, and are the weights and biases of the final mapping. In order to ensure the stability of the final features, Perform adaptive normalization:

[0118]

[0119] in, and are the mean and standard deviation of the feature vector, respectively, to ensure that the features remain consistent when calculated in different batches and improve the generalization ability of the model.

[0120] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0121] The single cell feature vector and the battery pack global feature vector are cross-attention weighted fused, and feature enhancement is performed through a conditional batch normalization layer to obtain the initial fused feature containing local-global feature dependencies;

[0122] Input the initial fusion features into three parallel fully connected layers to extract multi-scale features to obtain a multi-scale feature group, and perform adaptive feature aggregation on the multi-scale feature group to obtain an aggregated feature vector;

[0123] The aggregated feature vector is input into the temporal feature extraction layer composed of bidirectional gated recurrent units, and the temporal dependency features are extracted on the forward and reverse sequences respectively to obtain the dynamic features considering the temporal relationship;

[0124] The dynamic features are screened through the attention gating mechanism, and the channel attention module is combined to perform feature recalibration to obtain discriminant features;

[0125] The discriminant features are input into the fault classification head and the health scoring head respectively. The fault classification head is used to output the probability distribution of battery pack fault types, and the health scoring head uses a quantitative perception network structure to output the health status score of the single cell.

[0126] Specifically, we construct an attention mechanism that integrates local information and global information. Assume that the battery pack contains The characteristic vector of each single cell is expressed as ,in represents the dimension of the feature vector, and the global feature vector of the battery pack is expressed as In order to establish the relationship between the single cell and the overall features, the query vector, key vector and value vector are calculated:

[0127]

[0128] in, is the mapping matrix, represents the feature dimension after mapping, To reduce the computational complexity and make the computation more efficient. and key vector Perform a dot product operation and normalize by the scaling factor to get the similarity score between the single cell and the global feature:

[0129]

[0130] in, Representative The similarity score between each single cell and the global feature, As a scaling factor, it can prevent the value from being too large and affecting the stability of the calculation. The similarity score is exponentially normalized:

[0131]

[0132] in, Represents the normalized attention weight, ensuring that the sum of all attention scores is 1. With value vector Perform weighted summation to obtain fusion features:

[0133]

[0134] In order to enhance the feature representation capability, the fused features are enhanced through the conditional batch normalization layer. Assume that the mean and standard deviation of the fused features are and , then the normalization process is:

[0135]

[0136] This processing method ensures that the features remain consistent between different batches and improves the generalization ability of the model. Input three parallel fully connected layers for multi-scale feature extraction. Each fully connected layer extracts features at different levels, so that the low layer captures local features and the high layer captures global features. Assume that the weight matrices of the three fully connected layers are , then the three-way calculation formulas are:

[0137]

[0138] The three sets of features are adaptively aggregated to obtain the final aggregated feature vector:

[0139]

[0140] in, is an adaptive weight coefficient, which is dynamically adjusted during the training process to ensure the optimal feature combination. Perform time series feature extraction to capture the changes in battery status over time. The input is a temporal feature extraction layer composed of a bidirectional gated recurrent unit (GRU). Indicates The hidden state of the battery in the forward and reverse GRU is calculated as follows:

[0141]

[0142]

[0143] The final time series feature vector is:

[0144]

[0145] In order to filter the most critical feature information, the attention gating mechanism is used to filter the temporal features, and the channel attention module is combined to perform feature recalibration. Assume that the attention weight matrix is ,but:

[0146]

[0147] Introduce the channel attention mechanism to calculate the importance of features in different channels and recalibrate them:

[0148]

[0149] in, is the weight matrix of channel attention. Input the fault classification head and health score head respectively to calculate the probability distribution of the fault type of the battery pack and the health status score of the single battery. Assume that the weight of the fault classification head is ,but:

[0150]

[0151] in, Represents the probability distribution of battery pack failure types. At the same time, the health score head uses a quantized perception network structure for calculation. Assume that the quantized mapping matrix of the health score is ,but:

[0152]

[0153] in, Represents the health score of a single battery.

[0154] In a specific embodiment, the safety protection method for the power battery pack further includes the following steps:

[0155] Based on the probability distribution of battery pack fault types, gradient back propagation is performed on the image feature extraction network. The second-order derivative and third-order derivative of the feature map at each position (i, j) in the last convolutional layer of the image feature extraction network with respect to the fault type probability are calculated respectively. The position importance coefficient is obtained by weighted combination. The position importance coefficient is then weightedly summed with the feature map activation value at the corresponding position to obtain the category response feature map of each channel.

[0156] The category response feature map is multi-scale fused with the shallow feature map through residual connection and adaptive weight calculation to obtain a multi-level thermal response map;

[0157] The multi-level thermal response map is upsampled to the same size as the incremental capacity curve image by bicubic interpolation, and the edge details are optimized by local contrast enhancement to obtain the original scale thermal map;

[0158] The original scale heat map is divided into two sub-regions, the charging process and the discharging process, according to the voltage range, and each sub-region is segmented by adaptive threshold to obtain the regional fault activation map;

[0159] Morphological processing and connected domain analysis are performed on the hot spots in the fault activation map of each region to extract the spatial distribution characteristics and intensity distribution characteristics of the key fault feature areas and obtain the fault area feature map.

[0160] The fault area feature map is converted into a three-channel pseudo-color image through adaptive color mapping, and the transparency channel is superimposed for visualization enhancement to obtain the fault attention area heat map of the incremental capacity curve image.

[0161] Specifically, the last convolutional layer of the image feature extraction network is calculated. Suppose the output category probability of the network is ,in Represents the probability of a certain fault type, and calculates the probability relative to each position in the last convolutional layer Feature map The second and third derivatives of :

[0162]

[0163] in, Representative feature map The second-order derivative of the fault type probability reflects the sensitivity of the characteristic graph at that location, while Represents the third-order derivative, capturing a more nonlinear relationship. The second-order and third-order derivatives are weighted together to obtain the position importance coefficient:

[0164]

[0165] in, and is the adaptive weight coefficient. The feature map activation value at the corresponding position Perform a weighted sum to obtain a class response feature map:

[0166]

[0167] In order to enhance the expressiveness of the response map, the residual connection is used with the shallow feature map. Multi-scale fusion is performed so that the final multi-level thermal response map contains both the global information of deep features and the local details of shallow features:

[0168]

[0169] in, and is the adaptive weight coefficient. Perform bicubic interpolation upsampling to keep the size consistent with the incremental capacity curve image:

[0170]

[0171] in, Represents the upsampling multiple. Optimize edge details and use local contrast enhancement technology to make the thermal map clearly highlight the contour of the fault area. According to the voltage range, it is divided into two sub-areas: charging process and discharging process. Assume the voltage range is , according to the segmentation threshold To divide the area:

[0172]

[0173]

[0174] The heat map of each sub-region is segmented by adaptive threshold to extract the most significant fault area, so that the activation map is more focused on the key area where the fault occurs. The mean and standard deviation of the heat map are and , then the adaptive threshold is calculated as:

[0175]

[0176] in, is the adaptive coefficient. The regional fault activation map is obtained by binarization operation:

[0177]

[0178] right Morphological processing, including dilation, erosion, and connected domain analysis, is performed to eliminate isolated noise points and enhance continuity. Suppose the morphological operation matrix is , then the expansion operation and corrosion operation are:

[0179]

[0180] Finally, the optimized fault area feature map is obtained Calculate the spatial distribution characteristics of the fault area, including connected domain analysis, and calculate the area of ​​different fault areas and intensity distribution characteristics :

[0181]

[0182] Fault area feature map Adaptive color mapping is performed to convert it into a three-channel pseudo-color image, and the transparency channel is superimposed so that the transparency of the high-activation area is lower and the transparency of the low-activation area is higher. Finally, a heat map of the fault attention area of ​​the incremental capacity curve image is obtained.

[0183] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0184] Perform multi-threshold grading on the health status score of the single battery, divide the score range into high-risk range, warning range and safe range, and obtain the health level label of the single battery;

[0185] Based on the heat map of the fault attention area obtained by calculating the probability distribution of the battery pack fault type, the integral intensity of the charging and discharging sections is calculated to obtain the integral intensity value. The target stage and danger level of the fault are determined based on the integral intensity value to obtain the risk assessment result of the charging and discharging process.

[0186] A multi-objective decision matrix is ​​constructed based on the health grade labels and risk assessment results, and the weight coefficients of each control parameter are calculated through the hierarchical analysis method to obtain a hierarchical protection strategy;

[0187] For a single battery in a high-risk range, the charging rate of the single battery is reduced to a first percentage of the nominal value, the discharging rate is reduced to a second percentage of the nominal value, the charging cut-off voltage is reduced by a first gradient value, and the discharging cut-off voltage is increased by a second gradient value, to obtain a high-risk level control parameter;

[0188] For single cells in the warning range, combined with the risk section in the heat map of the fault attention area, the charge and discharge rate is dynamically adjusted within the nominal value range, and the charge and discharge cut-off voltage is adjusted to the third gradient value to obtain the warning level control parameters;

[0189] The high-risk level control parameters and the warning level control parameters are integrated, and the control parameters are fine-tuned online through the Bayesian optimization framework. At the same time, the constraints of the battery pack cycle life and charging time are considered to obtain the charge and discharge protection control instructions.

[0190] Specifically, calculate the health score of a single battery The score is calculated based on historical operating data and current operating conditions, and the value range is set to To classify the battery health status, multiple thresholds are defined The scoring interval is divided into high-risk interval, warning interval and safe interval, among which The specific classification method is as follows:

[0191]

[0192] in, Representative After completing the health grading, the heat map of the fault attention area is calculated by combining the probability distribution of the fault type. , perform integral calculations on the charging and discharging sections to determine the stage and degree of danger of the fault. Assume that the integral intensity of the high response area in the thermal map is:

[0193]

[0194] in, and The fault activation diagrams are respectively for the charging stage and the discharging stage. and , determine the main stages of the failure:

[0195]

[0196] in, This means that the faults in the charging stage are dominant. It means that the fault is dominant in the discharge stage. The degree of danger is calculated by normalization:

[0197]

[0198] in, The value range is [0,1]. The larger the value, the higher the risk of failure. A multi-objective decision matrix is ​​constructed based on the health level label and risk assessment results, and the weight coefficient of each control parameter is calculated by the hierarchical analysis method to generate a hierarchical protection strategy. Assume that the control parameters include the charging rate , discharge rate , charging cut-off voltage and discharge cut-off voltage , then construct the decision matrix:

[0199]

[0200] Calculate weights using analytic hierarchy process So that:

[0201]

[0202] The weight coefficient is calculated based on the impact of different parameters on safety and is obtained through normalization. Determine the charging and discharging strategies for different batteries. , reduce the charge and discharge rate to reduce load stress, and adjust the voltage threshold to reduce damage to the battery under extreme conditions. and the nominal discharge rate is , the high-risk level control parameters are calculated as follows:

[0203]

[0204]

[0205] in, For the charge and discharge rate reduction ratio, For the voltage gradient adjustment value. The battery is combined with the risk section of the fault heat map to dynamically adjust the charge and discharge rate:

[0206]

[0207]

[0208] in, Represents the third gradient value, which is used to fine-tune the charge and discharge threshold of the warning level. The high-risk level control parameters and the warning level control parameters are integrated and fine-tuned online through the Bayesian optimization framework to optimize the cycle life and charging time while ensuring the safety of the battery pack. Suppose the optimization objective function is:

[0209]

[0210] in, is the cycle life, For charging time, is the weight coefficient. The optimal parameters are calculated iteratively through Bayesian optimization:

[0211]

[0212] Finally, the charge and discharge protection control instructions are obtained to realize the intelligent charge and discharge management of the battery pack.

[0213] In this embodiment, the control parameters are fine-tuned online through the Bayesian optimization framework, and the constraints of the battery pack cycle life and charging time are considered to obtain the charge and discharge protection control instructions, including: dividing the power battery pack into M sub-modules according to the series-parallel topology structure, configuring an intelligent agent unit for each sub-module, and building a communication network between the agent units based on the physical connection relationship within the battery pack to obtain a distributed control network; assigning a local objective function to each intelligent agent unit, taking the health status score and fault type probability distribution of the single battery in the sub-module as the optimization target, and taking the charge and discharge rate and cut-off voltage as the control variables to obtain a local control optimization problem; preliminarily optimizing the local objective function of each intelligent agent unit based on the first-order gradient descent method to obtain the initial control parameter solution; and obtaining the initial control parameter solution through the communication network. Local control parameter information is exchanged between adjacent agent units, and the second-order Newton method is used for iterative optimization to obtain the optimized control parameters that take into account the spatial coupling effect; the optimized control parameters are subjected to a global consistency constraint check to ensure that the control strategy of each submodule meets the overall safety boundary conditions of the battery pack, and a collaboratively optimized control scheme is obtained; the collaboratively optimized control scheme is input into the Bayesian optimizer for online parameter fine-tuning to dynamically balance the cycle life and charging efficiency of the battery pack and obtain the final charge and discharge protection control instructions; a communication blocking test is performed on the charge and discharge protection control instructions, and the control reliability in the event of communication interruption is ensured through a preset backup control strategy, and a control instruction with fault-tolerant capability is obtained; the control instruction with fault-tolerant capability is distributed to each intelligent agent unit to realize distributed collaborative control of the battery pack charging and discharging process.

[0214] The above describes the safety protection method of the power battery pack in the embodiment of the present invention. The following describes the safety protection device of the power battery pack in the embodiment of the present invention. Figure 2 , an embodiment of the safety protection device of the power battery pack in the embodiment of the present invention includes:

[0215] The data acquisition module 201 is used to collect voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix;

[0216] The feature extraction module 202 is used to perform incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector;

[0217] The relationship calculation module 203 is used to calculate the characteristic dependency relationship of the single battery characteristic vector to obtain the global characteristic vector of the battery pack;

[0218] The status scoring module 204 is used to perform local and global fault detection and health score calculation on the single cell feature vector and the battery pack global feature vector to obtain the battery pack fault type probability distribution and the single cell health status score;

[0219] The restriction adjustment module 205 is used to restrict and adjust the charge and discharge parameters of the power battery pack according to the probability distribution of the battery pack fault type and the single cell health status score, and obtain the charge and discharge protection control instruction.

[0220] Through the collaboration of the above components, the internal state of the battery pack is accurately characterized by extracting multi-scale features of the single cell voltage data and analyzing the incremental capacity curve, which improves the accuracy of state monitoring. The feature fusion network based on the attention mechanism can adaptively capture the dependencies between single cells, effectively solving the problem that the traditional method ignores the interaction between cells. The deep learning is innovatively combined with electrochemical feature analysis to build an end-to-end fault diagnosis model, which greatly improves the accuracy and real-time performance of fault detection. The gradient-based interpretability analysis method is introduced to intuitively display the key areas where the fault occurs through the heat map, making the fault diagnosis results interpretable. A multi-level adaptive charge and discharge protection strategy is designed to dynamically adjust the control parameters according to the health status and fault risk level of the single cell, significantly improving the accuracy and effectiveness of the protection measures. The Bayesian optimization framework is used for online parameter tuning, which takes into account the charging efficiency while ensuring safety, and achieves a good balance between safety and performance.

[0221] above Figure 2 The safety protection device of the power battery pack in the embodiment of the present invention is described in detail from the perspective of modular functional entities, and the computer device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0222] Figure 3It is a structural schematic diagram of a computer device provided in an embodiment of the present invention. The computer device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the computer device 300. Furthermore, the processor 310 can be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the computer device 300 to implement the steps of the above-mentioned power battery pack safety protection method.

[0223] The computer device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The illustrated computer device structure does not constitute a limitation on the computer device provided by the present invention, and may include more or fewer components than illustrated, or a combination of certain components, or a different arrangement of components.

[0224] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0225] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0226] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A safety protection method for a power battery pack, characterized in that: The method comprises: Collect voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix; Performing incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector; Calculating the feature dependency of the single cell feature vector to obtain a global feature vector of the battery pack; specifically comprising: performing feature transformation on the single cell feature vector to generate a query vector, a key vector and a value vector respectively; performing a dot product operation on the query vector and the key vector, and performing numerical normalization through a scaling factor to obtain a similarity score matrix between the single cells; performing a Softmax function calculation on the similarity score matrix, and performing a matrix multiplication operation with the value vector to obtain an initial feature fusion result that takes into account the relationship between the single cells; inputting the initial feature fusion result into a feedforward network composed of four fully connected layers to perform nonlinear feature transformation to obtain a backbone feature vector; dividing the backbone feature vector into N sub-feature vectors, and performing parallel feature extraction on the N sub-feature vectors respectively to obtain N groups of independent attention features; splicing and linear mapping the N groups of independent attention features, and performing feature enhancement through adaptive layer normalization to obtain a global feature vector of the battery pack; Performing local and global fault detection and health score calculation on the single cell feature vector and the battery pack global feature vector to obtain a probability distribution of battery pack fault types and a single cell health status score; According to the probability distribution of the battery pack fault type and the single cell health status score, the charge and discharge parameters of the power battery pack are restricted and adjusted to obtain a charge and discharge protection control instruction.

2. The safety protection method for a power battery pack according to claim 1, characterized in that: The voltage data of each single cell in the power battery pack is collected to obtain a single cell voltage data matrix, including: A voltage acquisition module is set for each single cell in the power battery pack to collect voltage data at a preset sampling frequency to obtain the real-time voltage value of each single cell; Arrange the real-time voltage values ​​of each single cell in time sequence according to the acquisition time to obtain the voltage sampling sequence of each single cell; Performing synchronous time stamping on the voltage sampling sequence of each single cell to obtain a voltage data sequence with a time stamp; Arranging the voltage data sequence with the timestamp in a matrix according to the sequence number of the single cells to obtain an initial voltage data matrix; The initial voltage data matrix is ​​stored in segments to obtain a segmented voltage matrix, and the segmented voltage matrix is ​​data-transmitted and integrated via a CAN bus to obtain a single cell voltage data matrix.

3. The safety protection method for a power battery pack according to claim 2, characterized in that: The step of performing incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector includes: The voltage values ​​in the single cell voltage data matrix are divided into a voltage interval sequence according to preset intervals, and the capacity integral value of each voltage interval in the voltage interval sequence is accumulated and calculated to obtain a charge and discharge capacity sequence of each voltage interval; Performing a ratio operation on the capacity change and the voltage change in adjacent voltage intervals in the charge-discharge capacity sequence, and obtaining incremental capacity curve data through third-order spline interpolation; Normalizing and pixel matrix mapping the incremental capacity curve data to obtain an incremental capacity curve image; Inputting the incremental capacity curve image into five convolution blocks of an image feature extraction network, each convolution block includes three convolution layers with 3×3 convolution kernels and one ReLU activation function layer, to obtain five groups of convolution feature maps of different scales; The dimensionality of five groups of convolutional feature maps of different scales is reduced to obtain a multi-scale feature fusion matrix, and the multi-scale feature fusion matrix is ​​linearly mapped through three fully connected layers to obtain a single cell feature vector.

4. The safety protection method for a power battery pack according to claim 3, characterized in that: The local and global fault detection and health score calculation are performed on the single cell feature vector and the battery pack global feature vector to obtain the battery pack fault type probability distribution and the single cell health status score, including: Performing cross-attention weighted fusion on the single cell feature vector and the battery pack global feature vector, and performing feature enhancement through a conditional batch normalization layer to obtain an initial fused feature containing a local-global feature dependency relationship; Inputting the initial fusion features into three parallel fully connected layers to perform multi-scale feature extraction to obtain a multi-scale feature group, and performing adaptive feature aggregation on the multi-scale feature group to obtain an aggregated feature vector; Inputting the aggregated feature vector into a temporal feature extraction layer composed of bidirectional gated recurrent units, and extracting temporal dependency features on the forward and reverse sequences respectively, to obtain dynamic features that consider the temporal relationship; The dynamic features are screened through an attention gating mechanism, and the features are recalibrated in combination with a channel attention module to obtain discriminant features; The discriminant features are respectively input into the fault classification head and the health scoring head. The fault classification head is used to output the probability distribution of battery pack fault types, and the health scoring head uses a quantized perception network structure to output the health status score of the single cell.

5. The safety protection method for a power battery pack according to claim 4, characterized in that: The safety protection method of the power battery pack also includes: Based on the probability distribution of the battery pack fault type, gradient back propagation is performed on the image feature extraction network, and the second-order derivative and the third-order derivative of the feature map at each position (i, j) in the last convolutional layer of the image feature extraction network with respect to the fault type probability are respectively calculated, and the position importance coefficient is obtained by weighted combination, and then the position importance coefficient is weightedly summed with the feature map activation value of the corresponding position to obtain the category response feature map of each channel; The category response feature map is multi-scale fused and adaptively weighted with the shallow feature map through a residual connection method to obtain a multi-level thermal response map; Upsampling the multi-level thermal response map to the same size as the incremental capacity curve image by bicubic interpolation, and optimizing edge details by local contrast enhancement to obtain an original scale thermal map; The original scale heat map is divided into two sub-regions of charging process and discharging process according to the voltage range, and each sub-region is segmented by adaptive threshold to obtain a sub-region fault activation map; Performing morphological processing and connected domain analysis on the hotspot area in the sub-region fault activation map, extracting the spatial distribution characteristics and intensity distribution characteristics of the key fault feature area, and obtaining a fault area feature map; The fault area feature map is converted into a three-channel pseudo-color image through adaptive color mapping, and the transparency channel is superimposed for visualization enhancement to obtain a thermal map of the fault attention area of ​​the incremental capacity curve image.

6. The safety protection method for a power battery pack according to claim 5, characterized in that: The limiting and adjusting the charge and discharge parameters of the power battery pack according to the probability distribution of the battery pack fault type and the single cell health status score to obtain the charge and discharge protection control instruction includes: Performing multi-threshold grading on the health status score of the single cell battery, dividing the score interval into a high-risk interval, a warning interval and a safe interval, and obtaining a health grade label of the single cell battery; According to the heat map of the fault attention area obtained by calculating the probability distribution of the battery pack fault type, the integral intensity of the charging section and the discharging section is calculated to obtain the integral intensity value, and the target stage and danger level of the fault occurrence are determined according to the integral intensity value to obtain the risk assessment result of the charging and discharging process; A multi-objective decision matrix is ​​constructed based on the health grade label and the risk assessment result, and the weight coefficient of each control parameter is calculated by the hierarchical analysis method to obtain a hierarchical protection strategy; For a single battery in a high-risk range, the charging rate of the single battery is reduced to a first percentage of the nominal value, the discharging rate is reduced to a second percentage of the nominal value, the charging cut-off voltage is reduced by a first gradient value, and the discharging cut-off voltage is increased by a second gradient value, to obtain a high-risk level control parameter; For the single battery in the warning range, in combination with the risk section in the thermal map of the fault concern area, the charge and discharge rate is dynamically adjusted within the nominal value range, and the charge and discharge cut-off voltage is adjusted to a third gradient value to obtain a warning level control parameter; The high-risk level control parameters and the warning level control parameters are integrated, and the control parameters are fine-tuned online through a Bayesian optimization framework. At the same time, the constraints of the battery pack cycle life and charging time are considered to obtain the charge and discharge protection control instructions.

7. A safety protection device for a power battery pack, characterized in that: Used to execute the safety protection method of the power battery pack according to any one of claims 1 to 6, the safety protection device of the power battery pack comprises: The data acquisition module is used to collect voltage data of each single cell in the power battery pack to obtain a single cell voltage data matrix; A feature extraction module, used for performing incremental capacity curve conversion and single cell feature extraction on the single cell voltage data matrix to obtain a single cell feature vector; A relationship calculation module, used for calculating the characteristic dependency relationship of the single cell characteristic vector to obtain a global characteristic vector of the battery pack; A status scoring module, used to perform local and global fault detection and health score calculation on the single cell feature vector and the battery pack global feature vector, to obtain a probability distribution of battery pack fault types and a single cell health status score; The restriction adjustment module is used to restrict and adjust the charge and discharge parameters of the power battery pack according to the probability distribution of the battery pack fault type and the single cell health status score to obtain a charge and discharge protection control instruction.

8. A computer device, characterized in that: The computer device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory so that the computer device executes the safety protection method for the power battery pack according to any one of claims 1 to 6.

Images