Lithium battery internal short circuit fault early warning and positioning method and system based on deep learning

Abstract

The application provides a kind of early warning and positioning method and system for lithium battery internal short circuit fault based on deep learning;Early warning and positioning method includes: step S10, multidimensional data semantic acquisition and working condition adaptive pretreatment;Step S20, cross-scale spatiotemporal feature collaborative extraction and alignment;Step S30, dynamic threshold self-adaptive internal short circuit early warning and feedback optimization;Step S40, topological perception fault accurate positioning and verification;Step S50, embedded collaborative optimization and self-diagnosis deployment: the application can find the early weak signal of internal short circuit fault in time, improve the efficiency and timeliness of battery safety monitoring;Improve the sensitivity of early warning;Enhance the robustness in high temperature, low temperature, fast charging and discharging environment;Can accurately identify the battery monomer and electrode position where fault occurs.

Description

Technical Field

[0001] This invention relates to the field of lithium battery fault detection technology, and in particular to a method and system for early warning and location of internal short circuit faults in lithium batteries based on deep learning. Background Technology

[0002] With the rapid development of new energy vehicles, energy storage power stations, and portable electronic devices, lithium batteries, as core energy supply units, directly impact equipment performance and user safety due to their safety and reliability. Internal short-circuit faults in lithium batteries are a major cause of thermal runaway, fire, and even explosion, posing a serious threat to the safe operation of battery systems. Therefore, accurately and promptly achieving early warning and precise location of internal short-circuit faults to reduce safety risks has become a key research direction in the field of lithium battery management.

[0003] Although existing lithium battery internal short-circuit detection technologies have made some progress in improving battery safety, significant shortcomings still exist:

[0004] Problem 1: Physical model-based detection methods analyze voltage, current, and temperature parameters using electrochemical or equivalent circuit models to determine the presence of internal short circuits. These methods rely on highly accurate model parameters, making them difficult to adapt to dynamic changes in complex operating conditions such as high and low temperatures and rapid charging and discharging. This results in weak early signals being masked by noise, low warning sensitivity, and high false alarm and false negative rates, making it difficult to achieve effective early fault detection.

[0005] Question 2: Detection methods based on characteristic parameters diagnose faults by monitoring voltage drops, capacity decay, or changes in internal resistance. However, the weak signals in the early stages of internal short circuits are often interfered with by environmental noise or changes in operating conditions, resulting in insufficient detection sensitivity. These methods typically analyze only single-mode data, lacking the fusion of multi-dimensional data such as voltage, current, temperature, and electrochemical impedance spectroscopy. This leads to poor robustness under complex operating conditions and fails to meet the reliability requirements of practical applications.

[0006] Thirdly, traditional machine learning methods, such as support vector machines or decision trees, classify data using manually designed features. However, feature extraction is complex, requires high data quality, and struggles to capture subtle feature changes in the early stages of internal short circuits. Furthermore, these methods lack modeling of the battery module topology, making it impossible to accurately locate the individual cells or electrodes where internal short circuits occur. This limits troubleshooting efficiency and fails to meet the demands for rapid response and precise maintenance. Summary of the Invention

[0007] To address at least one technical problem in the prior art, embodiments of the present invention provide a method and system for early warning and location of internal short-circuit faults in lithium batteries based on deep learning. This method can promptly detect early, weak signals of internal short-circuit faults, improving the efficiency and timeliness of battery safety monitoring; enhancing the sensitivity of early warning; strengthening robustness under high temperature, low temperature, and rapid charge / discharge environments; and accurately identifying the location of the faulty battery cell and electrode. To achieve the above technical objectives, the technical solution adopted in embodiments of the present invention is as follows:

[0008] In a first aspect, embodiments of the present invention provide a method for early warning and localization of internal short-circuit faults in lithium batteries based on deep learning, comprising the following steps:

[0009] Step S10, Semantic Acquisition of Multidimensional Data and Adaptive Preprocessing of Operating Conditions: Real-time acquisition of voltage, current, temperature and electrochemical impedance spectroscopy data of lithium batteries through sensor network, noise reduction using dynamic adaptive transformation based on multi-Besch wavelet, and extraction of time domain features and frequency domain features to generate a semantic feature set for subsequent analysis.

[0010] Step S20, Cross-scale spatiotemporal feature collaborative extraction and alignment: Construct a multi-scale convolutional neural network, integrate the temporal dynamic weighting mechanism and the cross-modal feature alignment strategy, extract the short-term transient, medium-term periodic and long-term trend features of multi-dimensional data, and generate a comprehensive spatiotemporal feature vector for subsequent analysis;

[0011] Step S30, Dynamic Threshold Adaptive Early Warning and Feedback Optimization for Internal Short Circuit: Long Short Time Memory Network is used to model time series data, which is the comprehensive spatiotemporal feature vector generated in step S20. Combined with a dynamic attention allocation mechanism, the key features related to internal short circuit are focused. The warning threshold is adaptively adjusted based on the battery operating conditions to generate internal short circuit warning results. The warning performance is optimized through closed-loop feedback.

[0012] Step S40, topology-aware fault location and verification: The battery module topology is modeled using a graph neural network, and integrated spatiotemporal feature vectors are incorporated. Through multi-layer message passing and dynamic updating of node features, the location of the individual cells and electrodes where the internal short circuit occurs is accurately located, and the reliability of the location results is verified.

[0013] Step S50, Embedded Collaborative Optimization and Self-Diagnosis Deployment: Optimize multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks through dynamic pruning and multi-precision quantization; deploy an edge-end collaborative computing framework and a system self-diagnosis mechanism to achieve low-power real-time processing and high-reliability operation of the embedded battery management system.

[0014] Further, step S10 specifically includes:

[0015] Voltage data of individual battery cells and battery modules are collected by voltage sensors at a set sampling frequency. The voltage sensors are installed at the positive and negative terminals of the individual battery cells and the total output terminal of the battery module. The voltage data is transmitted to the data acquisition module through a wired communication interface.

[0016] Charge and discharge current data are collected by a current sensor at a set sampling frequency. The current sensor is installed in the main current circuit of the battery module and integrated into the embedded battery management system. The current data is transmitted to the data acquisition module via a CAN bus.

[0017] The surface temperature data of the battery cell is collected by thermocouples at a set sampling frequency. The thermocouples are evenly installed at the center of the surface of the battery cell shell. The temperature data is transmitted to the data acquisition module through a wired interface.

[0018] Electrochemical impedance spectroscopy data are acquired using an electrochemical workstation at a set frequency range. The electrode probes of the electrochemical workstation are connected to the positive and negative electrodes of the battery cell, and the electrochemical impedance spectroscopy data are transmitted to the data acquisition module via a serial communication interface.

[0019] A dynamic adaptive transform based on multi-Besch wavelets is adopted to dynamically adjust the wavelet decomposition level according to the battery operating conditions, denoise the voltage, current and temperature data and extract time-domain features; the electrochemical impedance spectroscopy data is normalized and the real and imaginary parts of the Nyquist plot are extracted to generate frequency-domain features.

[0020] An adaptive data calibration module for operating conditions is constructed. Based on the battery operating conditions, the Kalman filter algorithm is used to calibrate the offset and noise of voltage, current and temperature data, and amplitude normalization is used to correct the spectral deviation of electrochemical impedance spectroscopy data.

[0021] A cross-modal semantic association analysis module is constructed. Based on the Pearson correlation coefficient, the intrinsic correlation of voltage, current, temperature and electrochemical impedance spectroscopy data is analyzed through the correlation matrix between features. By aligning the distribution between modes, multimodal features are fused based on minimizing the maximum mean difference to generate a semantic feature set.

[0022] Further, step S20 specifically includes:

[0023] A multi-scale convolutional neural network is constructed, which contains three parallel convolutional branches: the first branch uses a 3×3 convolutional kernel to extract short-term transient features and detect rapidly changing anomalous data; the second branch uses a 5×5 convolutional kernel to extract medium-term periodic features and analyze periodic fluctuations; and the third branch uses a 7×7 convolutional kernel to extract long-term trend features and capture gradual anomalies.

[0024] A time-series dynamic weighting mechanism is introduced to dynamically adjust the weights of each convolutional branch according to the battery operating conditions, and a comprehensive spatiotemporal feature vector is generated by weighted summation;

[0025] A cross-modal feature alignment module is constructed. Through inter-modal cosine similarity calculation and feature space projection, the distribution differences of voltage, current, temperature and electrochemical impedance spectral features are calibrated to generate an aligned integrated spatiotemporal feature vector.

[0026] Further, step S30 specifically includes:

[0027] The comprehensive spatiotemporal feature vector across scales is input into the long short-term memory network to model the time series data; the time series data refers to the comprehensive spatiotemporal feature vector generated in step S20, which includes the short-term transient, medium-term periodic and long-term trend features of voltage, current, temperature and electrochemical impedance spectroscopy, arranged in chronological order;

[0028] A dynamic attention allocation mechanism is adopted, and the weights of each time step are calculated through the Softmax function to focus on key features related to internal short circuits and generate a weighted feature sequence.

[0029] Based on the battery operating conditions, the warning threshold is dynamically adjusted. Combining a fully connected layer and a sigmoid activation function, the internal short circuit warning probability, i.e. the warning result, is output and transmitted to the human-machine interface for display.

[0030] A closed-loop feedback optimization module is constructed to collect comparison data between early warning results and actual faults, dynamically update the parameters of the weights and attention allocation mechanism of the long short-term memory network, and optimize early warning performance.

[0031] Further, step S40 specifically includes:

[0032] A graph neural network is constructed to represent the battery module as a topological graph structure, where nodes represent individual battery cells, edges represent electrical connections, and edge weights are dynamically updated based on electrochemical impedance spectroscopy data.

[0033] The multi-scale spatiotemporal feature vector is used as the initial feature input of the node into the graph neural network. Through multi-layer graph convolution operation, the feature information of each node is passed to the neighboring nodes through the edge in the battery module topology graph. The feature information is aggregated and updated layer by layer, and the node features are dynamically updated to generate feature vectors representing the internal short circuit location.

[0034] The updated feature vector is input into the classifier to predict the specific cell number and electrode location of the internal short circuit, and the location result is transmitted to the human-computer interaction interface for display.

[0035] A fault location result verification module is constructed. By comparing historical fault location data and real-time electrochemical impedance spectroscopy data, the reliability of the location results is verified, and a verification report is generated and transmitted to the human-computer interaction interface for display.

[0036] Further, step S50 specifically includes:

[0037] Through dynamic pruning, connections in multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks whose absolute weight values ​​are less than a set weight threshold are dynamically removed based on the computing resources of the embedded battery management system.

[0038] Multi-precision quantization is used to quantize the weights in the multi-scale convolutional neural network, long short-term memory network, and graph neural network into 8-bit and 4-bit integers; the 8-bit integers are used for the weights of the time steps of the long short-term memory network in step S30 and the edge weights of the graph neural network in step S40; the 4-bit integers are used for the branch weights of the convolutional neural network in step S20 and the weights of the fully connected layers of the classifier in step S40.

[0039] Deploy an edge-end collaborative computing framework, which includes an edge end and a device end. The edge end is an edge computing node, and the device end is an embedded battery management system. The edge end and the device end work together through a message queue telemetry transmission protocol. Through the edge-end collaborative computing framework and a dynamic task allocation mechanism, based on a load balancing algorithm, the computing task ratio between the device end and the edge end is optimized.

[0040] A system self-diagnosis module is built to monitor the operating status of the sensor network, data acquisition module, and embedded battery management system in real time, detect hardware and software anomalies, generate self-diagnosis reports, and transmit them to the human-machine interface for display.

[0041] Secondly, embodiments of the present invention provide a deep learning-based early warning and location system for internal short-circuit faults in lithium batteries, used to implement the deep learning-based early warning and location method for internal short-circuit faults in lithium batteries as described above, including:

[0042] The data semantic acquisition and operating condition adaptive preprocessing unit is used to execute step S10, which involves multi-dimensional data semantic acquisition and operating condition adaptive preprocessing: real-time acquisition of voltage, current, temperature and electrochemical impedance spectroscopy data of lithium battery through sensor network, noise reduction using dynamic adaptive transform based on multi-Besch wavelet, and extraction of time domain features and frequency domain features to generate a semantic feature set for subsequent analysis.

[0043] The cross-scale spatiotemporal feature collaborative extraction and alignment unit is used to execute step S20. Cross-scale spatiotemporal feature collaborative extraction and alignment: construct a multi-scale convolutional neural network, integrate the temporal dynamic weighting mechanism and the cross-modal feature alignment strategy, extract the short-term transient, medium-term periodic and long-term trend features of multi-dimensional data, and generate a comprehensive spatiotemporal feature vector for subsequent analysis.

[0044] The dynamic threshold adaptive early warning and feedback optimization unit is used to execute step S30, which is a dynamic threshold adaptive early warning and feedback optimization of internal short circuit: a long short time memory network is used to model time series data, the time series data is the comprehensive spatiotemporal feature vector generated in step S20, and a dynamic attention allocation mechanism is used to focus on key features related to internal short circuit. The warning threshold is adaptively adjusted based on the battery operating conditions to generate internal short circuit warning results, and the warning performance is optimized through closed-loop feedback.

[0045] The topology-aware localization and verification unit is used to execute step S40, which is a topology-aware fault accurate localization and verification: the battery module topology is modeled using a graph neural network, and integrated spatiotemporal feature vectors are incorporated. Through multi-layer message passing and dynamic updating of node features, the location of the individual cells and electrodes where the internal short circuit occurs is accurately located, and the reliability of the localization results is verified.

[0046] The embedded collaborative optimization and self-diagnosis unit is used to execute step S50, which involves deploying embedded collaborative optimization and self-diagnosis: optimizing multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks through dynamic pruning and multi-precision quantization; deploying an edge-end collaborative computing framework and a system self-diagnosis mechanism to achieve low-power real-time processing and high-reliability operation of the embedded battery management system.

[0047] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows:

[0048] 1) By real-time acquisition of lithium battery voltage, current, temperature, and electrochemical impedance spectroscopy data, combined with deep learning models such as multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks, early weak signals of internal short-circuit faults can be detected in a timely manner, effectively avoiding the risk of thermal runaway due to detection delays and improving the efficiency and timeliness of battery safety monitoring. Adaptive calibration under operating conditions and cross-modal semantic correlation analysis are employed to automatically extract complex features from multi-dimensional data, significantly improving the sensitivity and accuracy of early warnings. Compared to the limitations of traditional physical models and feature parameter methods that rely on fixed thresholds or single-modal data, multi-dimensional data fusion and dynamic feature extraction overcome noise interference under complex operating conditions, enhancing robustness in high-temperature, low-temperature, and rapid charge-discharge environments.

[0049] 2) By modeling the battery module topology using graph neural networks and incorporating cross-scale spatiotemporal features, precise localization of internal short-circuit faults can be achieved, accurately identifying the location of the faulty battery cell and electrode. It not only provides fault detection but also generates more reliable location reports through topology-aware localization and verification mechanisms, significantly improving fault diagnosis and maintenance efficiency while reducing repair time and costs. Compared to the lack of localization capabilities in traditional machine learning methods, this approach solves the challenge of fault localization in complex battery module topologies through multi-layer message passing, dynamic node feature updates, and real-time analysis of electrochemical impedance spectroscopy data.

[0050] 3) By using dynamic pruning, multi-precision quantization, and deploying an edge-end collaborative computing framework, we can ensure low-power real-time processing of the embedded battery management system, which solves the limitations of traditional deep learning methods such as large computational load and difficulty in embedded deployment. Attached Figure Description

[0051] Figure 1 This is a flowchart of the early warning and positioning method in an embodiment of the present invention.

[0052] Figure 2 This is a schematic diagram of the system composition framework in an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0054] In the description of the embodiments of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0055] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can also refer to the internal connection of two components; and they can refer to a wireless connection or a wired connection. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0056] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0057] This invention proposes a deep learning-based method for early warning and localization of internal short-circuit faults in lithium batteries. By real-time acquisition of voltage, current, temperature, and electrochemical impedance spectroscopy data of lithium batteries, and utilizing deep learning models such as multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks, combined with innovative mechanisms such as operating condition adaptive calibration, cross-modal semantic association analysis, dynamic threshold adaptation, closed-loop feedback optimization, and edge-end collaborative computing framework, this method achieves highly sensitive early warning and accurate localization of internal short-circuit faults. It is adapted to the low-power real-time processing requirements of embedded battery management systems and ensures robustness under operating conditions such as high temperature, low temperature, and rapid charge and discharge.

[0058] A deep learning-based method for early warning and localization of internal short-circuit faults in lithium batteries includes the following steps:

[0059] Step S10, Semantic Acquisition of Multidimensional Data and Adaptive Preprocessing under Operating Conditions: Real-time acquisition of voltage, current, temperature and electrochemical impedance spectroscopy data of lithium batteries through sensor network, noise reduction using dynamic adaptive transform based on multi-Besch wavelet, and extraction of time-domain and frequency-domain features to generate a semantic feature set for subsequent analysis.

[0060] The sensor network includes voltage sensors, current sensors, thermocouples, and an electrochemical workstation. Voltage data from individual battery cells and the battery module is acquired at a sampling frequency of 1Hz using voltage sensors installed at the positive and negative terminals of the individual battery cells and the overall output terminal of the battery module. The voltage data is transmitted to the data acquisition module via an RS-485 wired communication interface. The data acquisition module, equipped with an ARM Cortex-M4 processor and a 16-bit analog-to-digital converter, receives voltage data and stores it in an internal buffer for detecting voltage anomalies and reflecting voltage fluctuations caused by internal short circuits, serving as the primary input for feature extraction. Charge and discharge current data is acquired at a sampling frequency of 1Hz using Hall effect current sensors installed in the main current loop of the battery module and integrated into the embedded battery management system. Current data is transmitted to the data acquisition module via a CAN bus and used to analyze current anomalies during charging and discharging, serving as input for feature extraction. Surface temperature data from individual battery cells is acquired at a sampling frequency of 0.1Hz using thermocouples uniformly installed at the center of the battery cell casing surface and transmitted to the data acquisition module via an I2C wired interface. This data is used to monitor changes in battery surface temperature and reflect thermal anomalies caused by internal short circuits. Electrochemical impedance spectroscopy (EIS) data is acquired using an electrochemical workstation with a frequency range of 0.1 Hz to 10 kHz. The electrode probes of the workstation are connected to the positive and negative electrodes of the battery cells. The EIS data is transmitted to the data acquisition module via a UART serial communication interface. After receiving the EIS data, impedance spectral features are extracted to reflect the electrochemical characteristics of internal short circuits. A dynamic adaptive transform based on multi-Besch wavelets is employed, dynamically adjusting the wavelet decomposition level to 3 to 5 layers according to the battery operating conditions. This denoising process is used to extract time-domain features from the voltage, current, and temperature data. The EIS data is then normalized using min-max normalization to extract the real and imaginary parts of the Nyquist plot, generating frequency-domain features, namely impedance modulus and phase angle. An adaptive data calibration module is constructed to calibrate the offset and noise of voltage, current, and temperature data using a Kalman filter algorithm based on the battery operating conditions. Amplitude normalization is used to correct the spectral deviation of the EIS data, ensuring data consistency under different operating conditions. Battery operating conditions, including temperature and charge / discharge rate, can be obtained from the embedded battery management system. A cross-modal semantic association analysis module is constructed. Based on the Pearson correlation coefficient, the intrinsic correlation of voltage, current, temperature and electrochemical impedance spectroscopy data is analyzed through the correlation matrix between features. By aligning the distribution between modes and fusing multimodal features based on minimizing the maximum mean difference, a semantic feature set is generated and stored in the memory of the data acquisition module for subsequent analysis.

[0061] Step S20, cross-scale spatiotemporal feature collaborative extraction and alignment: Construct a multi-scale convolutional neural network, integrate the temporal dynamic weighting mechanism and the cross-modal feature alignment strategy, extract the short-term transient, medium-term periodic and long-term trend features of multi-dimensional data, and generate a comprehensive spatiotemporal feature vector for subsequent analysis, i.e., for internal short-circuit early warning and location.

[0062] The multi-scale convolutional neural network comprises three parallel convolutional branches: the first branch uses a 3×3 convolutional kernel with 32 filters to extract short-term transient features and detect rapidly changing anomalous data, such as voltage transient drops of less than 50 milliseconds; the second branch uses a 5×5 convolutional kernel with 64 filters to extract medium-term periodic features and analyze periodic fluctuations, such as periodic current oscillations; the third branch uses a 7×7 convolutional kernel with 128 filters to extract long-term trend features and capture gradual anomalies, such as temperature increases exceeding 1 degree Celsius per minute. A temporal dynamic weighting mechanism is introduced to dynamically adjust the weights of each convolutional branch according to the battery operating conditions. The weight coefficients are calculated using the softmax function, and a comprehensive spatiotemporal feature vector is generated through weighted summation. The battery operating conditions, including temperature and charge / discharge rate, can be obtained from the embedded battery management system. A cross-modal feature alignment module is constructed. The cross-modal feature alignment module calculates the cosine similarity between modes with a threshold of 0.8. The feature space projection is based on principal component analysis, retaining 95% of the variance. The distribution differences of voltage, current, temperature and electrochemical impedance spectroscopy features are calibrated to generate an aligned comprehensive spatiotemporal feature vector, which is stored in the memory of the data acquisition module for subsequent analysis.

[0063] Step S30, Dynamic Threshold Adaptive Early Warning and Feedback Optimization for Internal Short Circuit: Long Short-Term Memory Network is used to model time series data, which is the comprehensive spatiotemporal feature vector generated in step S20. Combined with a dynamic attention allocation mechanism, the key features related to internal short circuit are focused, and the warning threshold is adaptively adjusted based on the battery operating conditions to generate internal short circuit warning results. The warning performance is optimized through closed-loop feedback.

[0064] The comprehensive spatiotemporal feature vector across scales is input into the Long Short-Term Memory (LSTM) network to model time-series data. Here, the time-series data refers to the comprehensive spatiotemporal feature vector generated in step S20, which includes short-term transient, medium-term periodic, and long-term trend features of voltage, current, temperature, and electrochemical impedance spectroscopy, arranged chronologically. The comprehensive spatiotemporal feature vector has a dimension of 256. The LSM network contains two layers with 128 hidden units and tanh activation. A dynamic attention allocation mechanism is employed, using the Softmax function to calculate the weights at each time step, ranging from 0 to 1 with a step size of 0.01, focusing on key features related to internal short circuits, such as voltage drops and impedance abrupt changes, to generate a weighted feature sequence. The warning threshold is dynamically adjusted based on battery operating conditions. Specifically, a linear regression model can be used, with training data derived from historical operating conditions. The regression coefficient R² is greater than 0.9, and the warning threshold ranges from 0.5 to 0.9. Combining a fully connected layer and a sigmoid activation function, the internal short-circuit warning probability, i.e., the warning result, is output and displayed on the human-machine interface. Battery operating conditions, including temperature and charge / discharge rate, can be obtained from the embedded battery management system. A closed-loop feedback optimization module is constructed to collect comparison data between early warning results and actual faults. Using the gradient descent algorithm with a learning rate of 0.001, the parameters of the weights of the long short-term memory network and the attention allocation mechanism are dynamically updated to optimize the early warning performance.

[0065] Step S40, topology-aware fault location and verification: The battery module topology is modeled using a graph neural network, and integrated spatiotemporal feature vectors are incorporated. Through multi-layer message passing and dynamic updating of node features, the location of the individual cells and electrodes where the internal short circuit occurs is accurately located, and the reliability of the location results is verified.

[0066] Graph neural networks represent battery modules as a topological graph structure. Nodes represent individual battery cells, with a typical battery module containing 12 to 24 nodes. Edges represent electrical connections based on physical wiring. Edge weights are dynamically updated based on the real part of the electrochemical impedance spectroscopy data at a rate greater than 0.1 ohms per second. A comprehensive spatiotemporal feature vector across scales serves as the initial feature input to the graph neural network. Multi-layer message passing refers to the graph neural network using multi-layer graph convolution operations to pass the feature information of each node to its neighboring nodes through edges in the battery module topology graph, aggregating and updating layer by layer to dynamically update node features. This generates feature vectors representing the location of internal short circuits. For example, a three-layer graph convolution operation with 64, 32, and 16 channels, ReLU activation, aggregates neighbor node information, dynamically updates node features, and generates feature vectors representing the location of internal short circuits. The updated feature vectors are input to a classifier containing two fully connected layers with 32 and 8 neurons respectively. The fully connected layers have weights set and are activated by Softmax to predict the specific battery cell number and electrode location where the internal short circuit occurs. The location results are then transmitted to a human-computer interaction interface for display. A location result verification module is constructed. By comparing historical fault location data and real-time electrochemical impedance spectroscopy data, based on Euclidean distance and a threshold of 0.05, the reliability of the location result is verified, a verification report is generated, and the data is transmitted to the human-machine interface for display via a serial peripheral interface.

[0067] Step S50, Embedded Collaborative Optimization and Self-Diagnosis Deployment: Optimize multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks through dynamic pruning and multi-precision quantization; deploy an edge-end collaborative computing framework and a system self-diagnosis mechanism to achieve low-power real-time processing and high-reliability operation of the embedded battery management system.

[0068] Deep learning models can be optimized through dynamic pruning and multi-precision quantization. Dynamic pruning removes connections with absolute weight values ​​less than 0.01 from multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks based on the computational resources of the embedded battery management system, reducing model parameters by approximately 30%. Multi-precision quantization quantizes the weights in multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks into 8-bit and 4-bit integers. The 8-bit integers are used for the time-step weights of the long short-term memory network in step S30 and the edge weights of the graph neural network in step S40, prioritizing high-precision tasks such as time-series modeling and topology localization. The 4-bit integers are used for the branch weights of the convolutional neural network in step S20 and the weights of the fully connected layers of the classifier in step S40, optimizing storage and computational efficiency and reducing storage requirements by approximately 50%. The edge-end collaborative computing framework comprises an edge terminal and a device terminal. The edge terminal consists of edge computing nodes, such as industrial-grade single-board computers, while the device terminal is an embedded battery management system. The edge terminal and device terminal collaborate via a message queue telemetry transmission protocol, reducing the computational burden on the device terminal. Through the edge-end collaborative computing framework and dynamic task allocation mechanism, based on a load balancing algorithm, the ratio of computational tasks between the device terminal and the edge terminal is optimized. The device terminal can deploy lightweight feature extraction and initial screening models, while the edge terminal can perform complex feature fusion and early warning / localization tasks. The adjustment range for the ratio of computational tasks between the device terminal and the edge terminal is 0.3 to 0.7. A system self-diagnosis module is constructed to monitor the real-time operating status of the sensor network, data acquisition module, and embedded battery management system. For example, it monitors whether the voltage data from voltage sensors is within the specified range, and detects whether the embedded battery management system is working properly using heartbeat signals. It detects hardware and software anomalies and generates self-diagnostic reports, including fault codes and timestamps, which are then transmitted to the human-machine interface for display.

[0069] The human-machine interface displays the probability of warnings, location results, and self-diagnostic reports, providing users with intuitive monitoring of fault status. The hardware uses an LCD screen.

[0070] This invention also proposes a deep learning-based early warning and location system for internal short circuit faults in lithium batteries, used to implement the deep learning-based early warning and location method for internal short circuit faults in lithium batteries described above, including:

[0071] The data semantic acquisition and operating condition adaptive preprocessing unit is used to execute step S10, which involves multi-dimensional data semantic acquisition and operating condition adaptive preprocessing: real-time acquisition of voltage, current, temperature and electrochemical impedance spectroscopy data of lithium battery through sensor network, noise reduction using dynamic adaptive transform based on multi-Besch wavelet, and extraction of time domain features and frequency domain features to generate a semantic feature set for subsequent analysis.

[0072] The cross-scale spatiotemporal feature collaborative extraction and alignment unit is used to execute step S20. Cross-scale spatiotemporal feature collaborative extraction and alignment: construct a multi-scale convolutional neural network, integrate the temporal dynamic weighting mechanism and the cross-modal feature alignment strategy, extract the short-term transient, medium-term periodic and long-term trend features of multi-dimensional data, and generate a comprehensive spatiotemporal feature vector for subsequent analysis.

[0073] The dynamic threshold adaptive early warning and feedback optimization unit is used to execute step S30, which is a dynamic threshold adaptive early warning and feedback optimization of internal short circuit: a long short time memory network is used to model time series data, the time series data is the comprehensive spatiotemporal feature vector generated in step S20, and a dynamic attention allocation mechanism is used to focus on key features related to internal short circuit. The warning threshold is adaptively adjusted based on the battery operating conditions to generate internal short circuit warning results, and the warning performance is optimized through closed-loop feedback.

[0074] The topology-aware localization and verification unit is used to execute step S40, which is a topology-aware fault accurate localization and verification: the battery module topology is modeled using a graph neural network, and integrated spatiotemporal feature vectors are incorporated. Through multi-layer message passing and dynamic updating of node features, the location of the individual cells and electrodes where the internal short circuit occurs is accurately located, and the reliability of the localization results is verified.

[0075] The embedded collaborative optimization and self-diagnosis unit is used to execute step S50, which involves deploying embedded collaborative optimization and self-diagnosis: optimizing multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks through dynamic pruning and multi-precision quantization; deploying an edge-end collaborative computing framework and a system self-diagnosis mechanism to achieve low-power real-time processing and high-reliability operation of the embedded battery management system.

[0076] Example 1: Internal short circuit warning and location of electric vehicle battery module under high temperature conditions;

[0077] Scenario Description: The battery module of an electric vehicle contains 24 lithium battery cells, operates in a high-temperature environment (45 degrees Celsius), has a fast charging rate of 2C, and is used for urban commuting. It is necessary to monitor internal short-circuit faults in real time to ensure driving safety and battery life.

[0078] The method for early warning and location of internal short circuit faults in lithium batteries based on deep learning proposed in this application can detect an internal short circuit in the positive electrode of the 8th battery cell within 30 seconds. The warning accuracy is 96%, the location accuracy is 98%, the total delay is 45 milliseconds, the total power consumption is 720 milliwatts, and it is suitable for high temperature of 45 degrees Celsius and fast charging 2C conditions.

[0079] Analysis: Adaptive calibration under operating conditions effectively handles voltage drift caused by high temperature, cross-modal semantic association enhances feature consistency, dynamic threshold adaptation improves early warning sensitivity, graph neural network achieves accurate positioning, and edge-end collaborative optimization ensures real-time performance, meeting the safety requirements of electric vehicles.

[0080] In Comparative Example 1, for the same scenario described above, a traditional static threshold monitoring method is applied. A voltage sensor collects the voltage at the output terminals of individual battery cells and the battery module at a frequency of 1 Hz, while a temperature sensor collects the surface temperature of individual battery cells at a frequency of 0.1 Hz. The data is transmitted to the data acquisition module via an internal integrated circuit bus. The data acquisition module uses fixed thresholds to trigger an alarm when the voltage is below 3.5 volts or the temperature is above 50 degrees Celsius, without any noise reduction or operating condition calibration. An anomaly is detected after 60 seconds, with an alarm accuracy of 80%, but it cannot pinpoint the specific battery cell or electrode location. The total delay is 100 milliseconds, and the total power consumption is 500 milliwatts.

[0081] Comparative analysis: Traditional methods rely on fixed thresholds, lack adaptive calibration for operating conditions, cannot cope with signal drift caused by high temperatures, resulting in high false alarm rates, long warning delays, lack of multimodal data fusion and graph neural networks, and missing positioning functions.

[0082] Example 2: Internal short circuit warning and location of battery modules in energy storage power stations under low temperature conditions;

[0083] Scenario Description: A battery module of an energy storage power station contains 16 lithium battery cells, operates in a low-temperature environment with an ambient temperature of -10 degrees Celsius and a slow discharge rate of 0.5C, and is used for grid peak shaving. It is necessary to ensure long-term stable operation and rapid fault response.

[0084] The method for early warning and location of internal short circuit faults in lithium batteries based on deep learning proposed in this application can detect the internal short circuit of the negative electrode of the fifth battery cell within 25 seconds. The warning accuracy is 95%, the location accuracy is 97%, the total delay is 40 milliseconds, the total power consumption is 700 milliwatts, and it is suitable for low temperature conditions of -10 degrees Celsius.

[0085] Analysis: Adaptive calibration of operating conditions handles impedance fluctuations caused by low temperatures, cross-modal semantic association enhances feature robustness, closed-loop feedback optimizes model performance, and edge-end collaboration reduces latency, significantly improving early warning accuracy and functional integrity, and meeting the long-term stable operation requirements of energy storage power stations.

[0086] Comparative Example 2: For the same scenario described above, a traditional simple feature splicing monitoring method is applied. The voltage sensor collects voltage data at a frequency of 1 Hz, the current sensor collects current data at a frequency of 1 Hz, and the temperature sensor collects temperature data at a frequency of 0.1 Hz. These data are transmitted to the data acquisition module via the internal integrated circuit bus and the controller local area network bus. The data acquisition module directly splices the voltage, current, and temperature data without operating condition calibration or cross-modal analysis, generates a feature vector, inputs it into a fixed single-layer neural network model with a threshold of 0.5, and outputs a warning signal. An anomaly is detected after 70 seconds, with a warning accuracy of 75%, but the fault location cannot be pinpointed. The total delay is 120 milliseconds, and the total power consumption is 550 milliwatts.

[0087] Comparative analysis: Traditional methods rely on simple feature splicing, lack condition adaptation and cross-modal fusion, have poor feature consistency under low temperature conditions, resulting in a high false alarm rate, lack graph neural network localization function, and have a long latency.

[0088] Example 3: Internal short circuit warning and location of portable device batteries under rapid charging and discharging conditions;

[0089] Scenario Description: A single lithium battery cell for a portable device is operating under rapid charge / discharge conditions, with a charging rate of 3C and an ambient temperature of 25 degrees Celsius. It is used in a mobile communication device and requires rapid detection of internal short circuits to prevent thermal runaway.

[0090] The method for early warning and location of internal short circuit faults in lithium batteries based on deep learning proposed in this application can detect internal short circuits in the positive electrode within 20 seconds, with a warning accuracy of 97%, a location accuracy of 99%, a total delay of 35 milliseconds, and a total power consumption of 680 milliwatts, making it suitable for fast charging and discharging 3C operating conditions.

[0091] Analysis: Dynamic threshold adaptation and graph neural networks capture weak anomalies under rapid charging and discharging, while edge-end collaborative optimization reduces latency, significantly improving early warning accuracy and positioning precision, meeting the high real-time requirements of portable devices.

[0092] Comparative Example 3: For the same scenario described above, a traditional single-mode voltage monitoring method is applied. The voltage sensor collects voltage data at a frequency of 1 Hz and transmits it to the data acquisition module via an internal integrated circuit bus. The data acquisition module uses a fixed threshold of 3.5 volts to determine voltage anomalies, without analyzing other modal data, and directly transmits the data to the embedded battery management system. The embedded battery management system compares the voltage data with the threshold and generates an early warning signal. An anomaly is detected after 80 seconds, with an early warning accuracy of 70%, but the fault location cannot be pinpointed. The total delay is 150 milliseconds, and the total power consumption is 400 milliwatts.

[0093] Comparative analysis: Traditional methods rely solely on voltage data, lack multi-modal fusion and dynamic thresholds, have a high false alarm rate under rapid charging and discharging conditions, lack positioning capabilities, and have long delays.

[0094] The data in the above embodiments and comparative examples are derived from laboratory simulation tests and real-world scenario verification; the faults were generated by laboratory simulated short-circuit experiments. Test procedure: 100 independent tests were run under each operating condition, recording the warning accuracy, sensitivity, false positive rate, battery cell location accuracy, electrode location accuracy, and single inference time. Warning accuracy was calculated as the ratio of correctly predicted faults to the total number of predictions; sensitivity was calculated as the ratio of correctly identified fault samples to the total number of fault samples; the false positive rate was calculated as the ratio of incorrectly predicted fault normal samples to the total number of normal samples; location accuracy was calculated as the ratio of the number of times faulty battery cells or electrodes were correctly identified to the total number of faults; and single inference time was measured using an embedded device timer.

[0095] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early warning and localization of internal short-circuit faults in lithium batteries based on deep learning, characterized in that, Includes the following steps: Step S10, Semantic Acquisition of Multidimensional Data and Adaptive Preprocessing of Operating Conditions: Real-time acquisition of voltage, current, temperature and electrochemical impedance spectroscopy data of lithium batteries through sensor network, noise reduction using dynamic adaptive transformation based on multi-Besch wavelet, and extraction of time domain features and frequency domain features to generate a semantic feature set for subsequent analysis. Step S20, Cross-scale spatiotemporal feature collaborative extraction and alignment: Construct a multi-scale convolutional neural network, integrate the temporal dynamic weighting mechanism and the cross-modal feature alignment strategy, extract the short-term transient, medium-term periodic and long-term trend features of multi-dimensional data, and generate a comprehensive spatiotemporal feature vector for subsequent analysis; Step S30, Dynamic Threshold Adaptive Early Warning and Feedback Optimization for Internal Short Circuit: Long Short Time Memory Network is used to model time series data, which is the comprehensive spatiotemporal feature vector generated in step S20. Combined with a dynamic attention allocation mechanism, the key features related to internal short circuit are focused. The warning threshold is adaptively adjusted based on the battery operating conditions to generate internal short circuit warning results. The warning performance is optimized through closed-loop feedback. Step S40, topology-aware fault location and verification: The battery module topology is modeled using a graph neural network, and integrated spatiotemporal feature vectors are incorporated. Through multi-layer message passing and dynamic updating of node features, the location of the individual cells and electrodes where the internal short circuit occurs is accurately located, and the reliability of the location results is verified. Step S50, Embedded collaborative optimization and self-diagnosis deployment: Optimize multi-scale convolutional neural networks, long short-term memory networks and graph neural networks through dynamic pruning and multi-precision quantization; Deploy an edge-end collaborative computing framework and a system self-diagnosis mechanism to achieve low-power real-time processing and high-reliability operation of the embedded battery management system; Step S10 specifically includes: Voltage data of individual battery cells and battery modules are collected by voltage sensors at a set sampling frequency. The voltage sensors are installed at the positive and negative terminals of the individual battery cells and the total output terminal of the battery module. The voltage data is transmitted to the data acquisition module through a wired communication interface. Charge and discharge current data are collected by a current sensor at a set sampling frequency. The current sensor is installed in the main current circuit of the battery module and integrated into the embedded battery management system. The current data is transmitted to the data acquisition module via a CAN bus. The surface temperature data of the battery cell is collected by thermocouples at a set sampling frequency. The thermocouples are evenly installed at the center of the surface of the battery cell shell. The temperature data is transmitted to the data acquisition module through a wired interface. Electrochemical impedance spectroscopy data are acquired using an electrochemical workstation at a set frequency range. The electrode probes of the electrochemical workstation are connected to the positive and negative electrodes of the battery cell, and the electrochemical impedance spectroscopy data are transmitted to the data acquisition module via a serial communication interface. A dynamic adaptive transform based on multi-Besch wavelets is adopted to dynamically adjust the wavelet decomposition level according to the battery operating conditions, denoise the voltage, current and temperature data and extract time-domain features; the electrochemical impedance spectroscopy data is normalized and the real and imaginary parts of the Nyquist plot are extracted to generate frequency-domain features. An adaptive data calibration module for operating conditions is constructed. Based on the battery operating conditions, the Kalman filter algorithm is used to calibrate the offset and noise of voltage, current and temperature data, and amplitude normalization is used to correct the spectral deviation of electrochemical impedance spectroscopy data. A cross-modal semantic association analysis module is constructed. Based on the Pearson correlation coefficient, the intrinsic correlation of voltage, current, temperature and electrochemical impedance spectroscopy data is analyzed through the correlation matrix between features. By aligning the distribution between modes, multimodal features are fused based on minimizing the maximum mean difference to generate a semantic feature set. Step S20 specifically includes: A multi-scale convolutional neural network is constructed, which contains three parallel convolutional branches: the first branch uses a 3×3 convolutional kernel to extract short-term transient features and detect rapidly changing anomalous data; the second branch uses a 5×5 convolutional kernel to extract medium-term periodic features and analyze periodic fluctuations; and the third branch uses a 7×7 convolutional kernel to extract long-term trend features and capture gradual anomalies. A time-series dynamic weighting mechanism is introduced to dynamically adjust the weights of each convolutional branch according to the battery operating conditions, and a comprehensive spatiotemporal feature vector is generated by weighted summation; A cross-modal feature alignment module is constructed. Through inter-modal cosine similarity calculation and feature space projection, the distribution differences of voltage, current, temperature and electrochemical impedance spectral features are calibrated to generate an aligned integrated spatiotemporal feature vector. Step S30 specifically includes: The comprehensive spatiotemporal feature vector across scales is input into the long short-term memory network to model the time series data; the time series data refers to the comprehensive spatiotemporal feature vector generated in step S20, which includes the short-term transient, medium-term periodic and long-term trend features of voltage, current, temperature and electrochemical impedance spectroscopy, arranged in chronological order; A dynamic attention allocation mechanism is adopted, and the weights of each time step are calculated through the Softmax function to focus on key features related to internal short circuits and generate a weighted feature sequence. Based on the battery operating conditions, the warning threshold is dynamically adjusted. Combining a fully connected layer and a sigmoid activation function, the internal short circuit warning probability, i.e. the warning result, is output and transmitted to the human-machine interface for display. A closed-loop feedback optimization module is constructed to collect comparison data between early warning results and actual faults, dynamically update the parameters of the weights and attention allocation mechanism of the long short-term memory network, and optimize early warning performance. Step S40 specifically includes: A graph neural network is constructed to represent the battery module as a topological graph structure, where nodes represent individual battery cells, edges represent electrical connections, and edge weights are dynamically updated based on electrochemical impedance spectroscopy data. The multi-scale spatiotemporal feature vector is used as the initial feature input of the node into the graph neural network. Through multi-layer graph convolution operation, the feature information of each node is passed to the neighboring nodes through the edge in the battery module topology graph. The feature information is aggregated and updated layer by layer, and the node features are dynamically updated to generate feature vectors representing the internal short circuit location. The updated feature vector is input into the classifier to predict the specific cell number and electrode location of the internal short circuit, and the location result is transmitted to the human-computer interaction interface for display. A fault location result verification module is constructed. By comparing historical fault location data and real-time electrochemical impedance spectroscopy data, the reliability of the location results is verified, and a verification report is generated and transmitted to the human-computer interaction interface for display. Step S50 specifically includes: Through dynamic pruning, connections in multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks whose absolute weight values ​​are less than a set weight threshold are dynamically removed based on the computing resources of the embedded battery management system. Multi-precision quantization is used to quantize the weights in the multi-scale convolutional neural network, long short-term memory network, and graph neural network into 8-bit and 4-bit integers; the 8-bit integers are used for the weights of the time steps of the long short-term memory network in step S30 and the edge weights of the graph neural network in step S40; the 4-bit integers are used for the branch weights of the convolutional neural network in step S20 and the weights of the fully connected layers of the classifier in step S40. Deploy an edge-end collaborative computing framework, which includes an edge end and a device end. The edge end is an edge computing node, and the device end is an embedded battery management system. The edge end and the device end work together through a message queue telemetry transmission protocol. Through the edge-end collaborative computing framework and a dynamic task allocation mechanism, based on a load balancing algorithm, the computing task ratio between the device end and the edge end is optimized. A system self-diagnosis module is built to monitor the operating status of the sensor network, data acquisition module, and embedded battery management system in real time, detect hardware and software anomalies, generate self-diagnosis reports, and transmit them to the human-machine interface for display.

2. A deep learning-based early warning and location system for internal short-circuit faults in lithium batteries, characterized in that, The method for implementing early warning and location of internal short-circuit faults in lithium batteries based on deep learning as described in claim 1 includes: The data semantic acquisition and operating condition adaptive preprocessing unit is used to execute step S10, which involves multi-dimensional data semantic acquisition and operating condition adaptive preprocessing: real-time acquisition of voltage, current, temperature and electrochemical impedance spectroscopy data of lithium battery through sensor network, noise reduction using dynamic adaptive transform based on multi-Besch wavelet, and extraction of time domain features and frequency domain features to generate a semantic feature set for subsequent analysis. The cross-scale spatiotemporal feature collaborative extraction and alignment unit is used to execute step S20. Cross-scale spatiotemporal feature collaborative extraction and alignment: construct a multi-scale convolutional neural network, integrate the temporal dynamic weighting mechanism and the cross-modal feature alignment strategy, extract the short-term transient, medium-term periodic and long-term trend features of multi-dimensional data, and generate a comprehensive spatiotemporal feature vector for subsequent analysis. The dynamic threshold adaptive early warning and feedback optimization unit is used to execute step S30, which is a dynamic threshold adaptive early warning and feedback optimization of internal short circuit: a long short time memory network is used to model time series data, the time series data is the comprehensive spatiotemporal feature vector generated in step S20, and a dynamic attention allocation mechanism is used to focus on key features related to internal short circuit. The warning threshold is adaptively adjusted based on the battery operating conditions to generate internal short circuit warning results, and the warning performance is optimized through closed-loop feedback. The topology-aware localization and verification unit is used to execute step S40, which is a topology-aware fault accurate localization and verification: the battery module topology is modeled using a graph neural network, and integrated spatiotemporal feature vectors are incorporated. Through multi-layer message passing and dynamic updating of node features, the location of the individual cells and electrodes where the internal short circuit occurs is accurately located, and the reliability of the localization results is verified. The embedded collaborative optimization and self-diagnosis unit is used to execute step S50, which involves deploying embedded collaborative optimization and self-diagnosis: optimizing multi-scale convolutional neural networks, long short-term memory networks, and graph neural networks through dynamic pruning and multi-precision quantization; deploying an edge-end collaborative computing framework and a system self-diagnosis mechanism to achieve low-power real-time processing and high-reliability operation of the embedded battery management system.

Images