Battery cluster fault diagnosis method and system and medium

Abstract

The invention discloses a battery cluster fault diagnosis method and system and a medium, and belongs to the technical field of energy storage system safety. Comprising the following steps: acquiring multi-modal time sequence data of a battery cluster module in a first diagnosis period, inputting the multi-modal time sequence data into an A-ConvNeXt model cascaded with a channel attention module through depth separable convolution for preliminary diagnosis, and outputting a preliminary diagnosis result, confidence and a feature vector; fusing the feature vectors of a plurality of continuous initial judgment periods of the battery cluster to form longer second diagnosis period data, inputting the longer second diagnosis period data into a TCN-TF fusion network formed by connecting a time sequence convolutional network and a Transform encoder in parallel, and outputting a final diagnosis result and confidence after feature splicing and fusion through parallel extraction of local features and long-range dependence; when the final confidence exceeds a preset threshold value, a real fault is judged, the EMS is immediately triggered to execute a power reduction or shutdown instruction, and safe closed-loop control is achieved; according to the invention, accurate, rapid and adaptive diagnosis of the fault of the energy storage battery cluster is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, in particular to a battery cluster fault diagnosis method, system and medium. BACKGROUND

[0002] With the rapid development of renewable energy, energy storage systems play an increasingly important role in power peak shaving, frequency regulation and standby power supply. As the core component of energy storage systems, the operation state of battery clusters is directly related to the safety and reliability of the entire system. However, battery clusters are easily affected by internal and external factors during long-term operation, resulting in internal short circuit, thermal runaway, accelerated aging and other faults. If not diagnosed and handled in time, it may trigger a chain reaction, even leading to system paralysis or safety accidents.

[0003] Currently, the fault diagnosis method for battery clusters mainly relies on threshold judgment of single modal data (such as voltage or temperature) or traditional machine learning models. This kind of method often has low diagnosis accuracy, high false alarm rate and delayed response when dealing with multi-modal, nonlinear and time-varying dynamic fault characteristics. Especially in the face of complex working conditions and multi-fault coupling scenarios, traditional methods are difficult to effectively capture the early features and evolution rules of faults, limiting their application in engineering practice.

[0004] In addition, most existing diagnosis models use centralized training or general model deployment, lack of adaptive ability to individual battery cluster operation characteristics, and are difficult to achieve efficient and accurate real-time diagnosis on the edge side. Therefore, there is an urgent need for a battery cluster fault diagnosis method that can integrate multi-modal time series data, have long and short range feature extraction capability, and support individualized model optimization, to improve the intelligent operation and maintenance level and operation safety of energy storage systems. SUMMARY

[0005] The purpose of the present application is to provide a battery cluster fault diagnosis method, which realizes accurate, rapid and adaptive diagnosis of energy storage battery cluster faults.

[0006] To solve the above technical problems, the embodiments of the present application provide a battery cluster fault diagnosis system, comprising: a battery cluster interface board module, configured to acquire multi-modal time series data of each battery module in the battery cluster within a first diagnosis period, the multi-modal time series data including voltage, current, temperature and acoustic emission signals; The end-side computing node module is configured to process the multi-modal time series data in the first diagnosis period by an A-ConvNeXt model to obtain a diagnosis result and a confidence level of the first diagnosis period; if the confidence level of the first diagnosis period is greater than a preset threshold, a first diagnosis period feature vector corresponding to the diagnosis result is output; the A-ConvNeXt model replaces a depth separable convolution module of an original ConvNeXt model with a cascaded structure of a depth separable convolution and a channel attention module; the channel attention module adaptively adjusts the weight distribution of the multi-modal time series data according to the feature importance of different modalities in the multi-modal time series data and the relevance to the fault mode; The edge gateway module is configured to fuse the continuous first diagnosis period feature vectors corresponding to the battery cluster along a time axis to form second diagnosis period data; and process the second diagnosis period data corresponding to the battery cluster by a TCN-TF fusion network to obtain a diagnosis result and a confidence level of the second diagnosis period; the TCN-TF fusion network is composed of a time series convolution network and a Transformer encoder; the time series convolution network extracts local features and mutation information of the second diagnosis period data; the Transformer encoder captures long-range time dependence of the second diagnosis period data; the local features, the mutation information and the long-range time dependence are spliced and nonlinearly fused by a full connection layer to output the diagnosis result and the confidence level of the second diagnosis period. The industrial communication interface module is configured to determine that the battery cluster has a real fault if the confidence level of the second diagnosis period is greater than the preset threshold, and trigger the EMS to issue a control instruction to reduce the power of the battery cluster or stop the battery cluster.

[0007] In some optional embodiments, the cloud server module is further configured to train the A-ConvNeXt model of the end-side computing node module and the TCN-TF fusion network of the edge gateway module, and regularly update the model parameters of the end-side computing node module and the edge gateway module.

[0008] The embodiments of the present application also provide a battery cluster fault diagnosis method, comprising: The multi-modal time series data of each battery module in the battery cluster in a first diagnosis period is obtained, and the multi-modal time series data includes voltage, current, temperature and acoustic emission signals. The A-ConvNeXt model is used to process the multi-modal time series data in the first diagnosis period to obtain a diagnosis result and a confidence level of the first diagnosis period; if the confidence level of the first diagnosis period is greater than a preset threshold, a first diagnosis period feature vector corresponding to the diagnosis result is output; the A-ConvNeXt model replaces a depth separable convolution module of an original ConvNeXt model with a cascaded structure of a depth separable convolution and a channel attention module; the channel attention module adaptively adjusts the weight distribution of the multi-modal time series data according to the feature importance of different modalities in the multi-modal time series data and the relevance to the fault mode; The first diagnosis period feature vectors corresponding to the battery cluster are fused along a time axis to form second diagnosis period data; a TCN-TF fusion network is used to process the second diagnosis period data corresponding to the battery cluster to obtain a diagnosis result and a confidence level of the second diagnosis period; the TCN-TF fusion network is composed of a time convolution network and a Transformer encoder; the time convolution network extracts local features and mutation information of the second diagnosis period data; the Transformer encoder captures long-range time dependence of the second diagnosis period data; the local features, the mutation information, and the long-range time dependence are concatenated and nonlinearly fused through a fully connected layer to output the diagnosis result and the confidence level of the second diagnosis period; If the confidence level of the second diagnosis period is greater than the preset threshold, it is determined that the battery cluster has a real fault, and a control instruction for reducing the power of the battery cluster or stopping the battery cluster is triggered by the EMS.

[0009] In some optional embodiments, the training of the A-ConvNeXt model and the TCN-TF fusion network includes the following steps: The multi-modal time series data of each battery cluster in the energy storage power station is taken as an independent sub-training set, and each sub-training set corresponds to a battery cluster; the TCN-TF fusion network corresponding to each battery cluster is independently trained, and the TCN-TF fusion network parameters of the battery cluster are generated after the training; the TCN-TF fusion network parameters corresponding to each battery cluster are uploaded to a cloud server; after the cloud server collects the TCN-TF fusion network parameters of all battery clusters, the TCN-TF fusion network parameters are weighted and averaged according to the data amount proportion of each sub-training set to aggregate new global model parameters; the aggregated global model parameters are sent to the TCN-TF fusion network of each battery cluster to replace the original TCN-TF fusion network parameters; each battery cluster continues to train using the sub-training set based on the updated TCN-TF fusion network parameters; the above operations are repeated until the TCN-TF fusion network of all battery clusters converges; The TCN-TF fusion network is taken as a teacher model, and the A-ConvNeXt student model is guided to learn and correct the model parameters of the A-ConvNeXt model by the diagnostic knowledge output by the TCN-TF fusion network through a knowledge distillation method.

[0010] In some optional embodiments, the A-ConvNeXt model is further subjected to INT8 quantization processing.

[0011] Embodiments of the present application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the battery cluster fault diagnosis method described above.

[0012] The battery cluster fault diagnosis method provided by the present application has at least the following beneficial effects: The present application realizes deep feature extraction and fusion of multi-modal time series data by constructing a two-stage diagnosis architecture of A-ConvNeXt and TCN-TF fusion network. The channel attention module in the A-ConvNeXt model can adaptively assign different weights to different modal data such as voltage, current, temperature and acoustic emission signals, focusing on the features most relevant to the fault, thereby improving the accuracy of preliminary diagnosis. Subsequently, the TCN-TF fusion network uses a time series convolution network (TCN) and a Transformer encoder in parallel, which can simultaneously capture local mutation details and long-range temporal dependencies of battery data, realize accurate modeling of the fault evolution process, and effectively reduce false positives through confidence judgment, ensuring that only in the case of high confidence can the EMS trigger power reduction or shutdown, realizing safe closed-loop control. BRIEF DESCRIPTION OF DRAWINGS

[0013] The accompanying drawings, which are included to provide a further understanding of the application, form a part of the application and, along with the detailed description, serve to explain the application. The detailed description and the specific embodiments described therein are intended for purposes of illustration only and are not intended to limit the scope of the application. In the drawings:

[0014] Figure 1 is a flowchart of a battery cluster fault diagnosis method according to an embodiment of the present application; Figure 2 is a schematic diagram of a battery cluster fault diagnosis system according to an embodiment of the present application; Figure 3 is a circuit schematic diagram of an interface board according to an embodiment of the present application; Figure 4 is a ConvNeXt model structure diagram according to an embodiment of the present application; Figure 5 is a flowchart of an end-side computing node implementation according to an embodiment of the present application; is a flowchart of an end-side computing node implementation according to an embodiment of the present application;Figure 6 is a secondary diagnosis flow chart of an edge gateway provided according to an embodiment of the present application; Figure 7 is an update timing diagram provided according to an embodiment of the present application; Figure 8 is an extended frame format and register mapping definition diagram of EMS communication provided according to an embodiment of the present application; Figure 9 is a cabinet arrangement and wiring diagram of a pool cluster fault diagnosis device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in detail with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0016] One embodiment of the present application relates to a battery cluster fault diagnosis method, and the implementation details of the battery cluster fault diagnosis method of the present embodiment will be described below. The following content is only provided for the implementation details for the convenience of understanding, and is not necessary for implementing the present solution.

[0017] The specific process of the battery cluster fault diagnosis method of the present embodiment can be as shown in Figure 1 , which includes: Step 101: Obtain the multi-modal time series data of each battery module in the battery cluster in a first diagnosis period, wherein the multi-modal time series data includes voltage, current, temperature and acoustic emission signals. The battery cluster fault diagnosis method is applied to a battery cluster fault diagnosis system, and the battery cluster fault diagnosis system is as shown in Figure 2As shown, the battery cluster contains multiple battery modules, each equipped with a battery voltage and current monitoring unit and interface, and the monitoring data is transmitted to the cluster-level node (ECU based on NXP CortexM core) at a rate of 8n 15Mbps. The cluster-level node sends data to the edge gateway through CAN (1Mbps), which uses a TI AM5728 processor and supports Modbus / RTU / RS485 protocols. It can also access the edge gateway of the energy Internet of Things platform as a backup OIA. Finally, the data is transmitted to the cloud server (Feeding Pan Gu Big Data Platform) through TLS1.3 encrypted Ethernet, realizing remote monitoring and data management of the battery cluster. The overall architecture ensures the whole process of battery cluster data from collection, transmission to cloud analysis, providing technical support for battery cluster state monitoring and management.

[0018] The multi-modal time series data includes voltage, current, temperature, and acoustic emission signals.

[0019] This step is completed by the battery cluster interface board module. The interface board is used to synchronously collect the voltage, current, temperature, and acoustic emission signals of each battery module, with a sampling frequency not less than 10 kHz, ensuring the synchronization and high fidelity of the data. The circuit diagram of the interface board is as follows Figure 3As shown in the figure, the multi-modal signal synchronous acquisition and high isolation degree hardware architecture of the battery cluster interface board are clearly shown. Specifically, the voltage channel linearly attenuates the battery string high voltage signal to the ±10 V range through a high-precision, low-temperature-drift voltage dividing resistor network (R1-R4), and is converted into digital by a 24-bit Σ-Δ ADC (U1, model ADS131E08S), which guarantees the accuracy and dynamic range of voltage sampling; the current channel uses a Hall sensor and a sampling resistor R5 to convert the current signal into a ±10 V differential voltage, which is connected to the differential input end of the same ADC, realizing strict synchronous sampling of voltage and current; the temperature channel uses a T-type thermocouple combined with a cold end compensation chip U2 (MAX31850) to directly output a digital signal, which is sent to the MCU through a One-Wire bus, simplifying the temperature measurement circuit and improving the anti-interference performance; the acoustic emission channel integrates an IEPE constant current source U4 (LM334) to provide a 2 mA constant current excitation for the acoustic emission sensor, and the high-frequency acoustic emission signal output by the sensor is synchronously acquired by the ADC after AC coupling. To achieve high reliability of the system, the circuit uses a DC-DC module U5 (B0505S-1W) to isolate the analog and digital power supplies, with a withstand voltage of 2.5 kV, and the SPI communication bus is isolated by a digital isolator U6 (ADUM1401), which forms a "signal-power" double isolation architecture, significantly improving the anti-interference ability and operation stability of the interface board in complex electromagnetic environments. This design fully meets the requirements of the right claim for synchronous sampling frequency of multi-modal signals not less than 10 kHz and electrical safety requirements for industrial device type tests.

[0020] In step 102, the multi-modal time series data in the first diagnosis period is processed by an A-ConvNeXt model to obtain a diagnosis result and a confidence level of the first diagnosis period. If the confidence level of the first diagnosis period is greater than a preset threshold, a first diagnosis period feature vector corresponding to the diagnosis result is output. The A-ConvNeXt model replaces the depth separable convolution module of the original ConvNeXt model with a cascaded structure of depth separable convolution and channel attention modules. The channel attention module adaptively adjusts the weight distribution of the multi-modal time series data according to the feature importance of different modalities in the multi-modal time series data and the relevance to the fault mode. The A-ConvNeXt model is subjected to INT8 quantization processing.

[0021] The A-ConvNeXt model structure diagram is as follows Figure 4As shown in the figure, the complete architecture of the A-ConvNeXt lightweight diagnostic model after INT8 quantization is clearly shown. The model takes a 512-byte filtered data frame as input, and then passes through N cascaded A-ConvNeXt modules for feature extraction. The core of each module is composed of a depth separable convolution, a channel attention module, and a GELU activation function in sequence, and the gradient flow is enhanced through a residual connection. After feature extraction, the global features are aggregated by the global average pooling layer and input into the dual-output diagnostic head: the classification head outputs the suspected fault code through a fully connected layer; the embedding head outputs a 128-byte feature vector through another fully connected layer for secondary diagnosis by the edge gateway. This structure ensures the high-precision preliminary judgment capability of the end side under low latency and low resource consumption.

[0022] This step is implemented by the end-side computing node module, and the implementation flowchart of the end-side computing node is as shown in Figure 5 As shown in the figure, the complete software and hardware workflow of the end-side computing node and its key performance indicators are shown. The hardware platform core uses an STM32H743 industrial-grade microcontroller with a 480 MHz main frequency, not more than 320 kB of RAM, and 1.5 MB of Flash storage resources, which can stably operate in a wide temperature range of -40 ℃ to +85 ℃, meeting the harsh environmental requirements of energy storage sites.

[0023] At the software level, after the system is powered on, it first completes the initialization of the clock, CAN-FD, and SPI communication interfaces, and then receives a frame of multi-modal data transmitted from the battery cluster interface board in real time in an interrupt mode. To improve signal quality, the system calls a 32-order FIR hardware filter in the embedded CMSIS-DSP library to filter out high-frequency noise from the voltage, current, temperature, and acoustic emission signals. The preprocessed data is sent to the lightweight A-ConvNeXt diagnostic model quantized by INT8. This model replaces the original ConvNeXt depth separable convolution module with a cascaded structure of "depth separable convolution + channel attention module", and uses GELU activation function, which significantly compresses the model volume to about 1.2 MB while maintaining high accuracy, allowing it to be completely stored in the microcontroller Flash, and achieving a single-step inference delay of not more than 15 ms in real-time performance.

[0024] In the decision and communication link, if the model output confidence exceeds the preset threshold (e.g., 60%), the node will package the suspected fault code (2 Byte) and the related feature vector (128 Byte) and send them to the edge gateway through the 1 Mbps baud rate CAN-FD bus for further diagnosis; if the confidence does not reach the threshold, the node can switch to a low-power mode for energy-saving operation. The entire process is executed in a fixed cycle (e.g., 10 ms), ensuring the system's ability to make millisecond-level fault preliminary judgments and respond in real time at the end side.

[0025] Step 103, fuse the corresponding battery cluster's continuous multiple first diagnostic period feature vectors along the time axis to form second diagnostic period data; process the corresponding battery cluster's second diagnostic period data through a TCN-TF fusion network to obtain a diagnostic result and a confidence level of the second diagnostic period; the TCN-TF fusion network is composed of a time series convolution network and a Transformer encoder fusion, the time series convolution network extracts local features and mutation information of the second diagnostic period data, and the Transformer encoder captures long-range time dependence of the second diagnostic period data, and the local features, mutation information and long-range time dependence are spliced and nonlinearly fused through a fully connected layer to output the diagnostic result and the confidence level of the second diagnostic period; Each battery cluster in the energy storage power station has multiple modal time series data as an independent sub-training set, and each sub-training set corresponds to a battery cluster; each sub-training set independently trains a TCN-TF fusion network corresponding to the battery cluster, and generates TCN-TF fusion network parameters exclusive to the battery cluster after training; upload the TCN-TF fusion network parameters corresponding to each battery cluster to the cloud server, and after the cloud server collects the TCN-TF fusion network parameters of all battery clusters, perform weighted averaging according to the data amount proportion of each sub-training set to aggregate new global model parameters; distribute the aggregated global model parameters to the TCN-TF fusion network exclusive to each battery cluster to overwrite the original TCN-TF fusion network parameters; each battery cluster continues to train using the sub-training set based on the updated TCN-TF fusion network parameters; repeat the above operations until the TCN-TF fusion networks of all battery clusters converge; Through a knowledge distillation method, the TCN-TF fusion network is used as a teacher model, and the diagnostic knowledge output by the TCN-TF fusion network is used to guide the learning of an A-ConvNeXt student model to correct the A-ConvNeXt model parameters.

[0026] This step is dominated by the edge gateway module. The secondary diagnosis flowchart of the edge gateway is shown in Figure 6 The figure shows the core processing flow of the edge gateway side for secondary fault confirmation and closed-loop control. Specifically, the edge gateway receives CAN-FD data frames uploaded from multiple battery cabin inner end side computing nodes through the SocketCAN interface; to eliminate clock drift between different nodes, the integrated FPGA coprocessing unit (such as Xilinx Artix-7) performs high-precision timestamp alignment on multiple data, with an accuracy of 1 microsecond, laying a foundation for subsequent accurate time series analysis.

[0027] In the data preprocessing stage, the system carries out Min-Max normalization on the four types of modal signals of voltage, current, temperature and acoustic emission respectively, and splices them into a 4*128-dimensional time series tensor. The tensor is input into the TCN-TF fusion diagnosis network, which adopts 3 layers of time convolution (convolution kernel size is 3, expansion coefficient is 1, 2, 4 respectively) to extract local features and mutation information of different time scales, and is connected in parallel with a 4-head attention mechanism, a 64-dimensional hidden layer Transformer encoder to capture long-range time dependence, finally realizing the deep fusion of multi-modal time series features and accurate classification of faults.

[0028] If the diagnostic confidence output by the model reaches or exceeds the preset threshold of 90%, the edge gateway performs a closed-loop control operation: writes the fault type (1 byte), confidence (1 byte) and timestamp (4 bytes) into the local SQLite database for persistent storage, and at the same time, through the RS485 interface, writes the above information into the specified holding register (address 40001-40008) of the energy management system (EMS) according to the Modbus-RTU protocol, directly triggers the power reduction or shutdown instruction, and completes the automatic closed loop from fault diagnosis to safe execution.

[0029] To realize continuous optimization of the model under the premise of protecting data privacy, the edge gateway desensitizes the gradient information generated by training (clipped to 1.0 using L2 norm), and uploads it to the cloud server through a TLS 1.3 encrypted channel, strictly controlling the daily upload traffic within 128 KB, fully meeting the strict safety specifications of the power system "data not outbound".

[0030] To realize continuous optimization of the model, the application adopts an end-edge-cloud collaborative training mechanism, and its update timing diagram is as follows: Figure 7As shown, the figure shows the complete closed-loop process of cloud model federation aggregation and secure OTA update. At the initial time T0, the edge gateway uploads the desensitized gradient G_t generated by local training to the cloud server; then at T1 stage, the cloud performs 20 rounds of iterative training based on the federated average algorithm (FedAvg) with a learning rate of 0.001, aggregates all edge node parameters to generate a new generation of global model W_{t+1}; enter T2 stage, in order to improve the efficiency of model deployment, W_{t+1} is sequentially executed based on amplitude-based weight pruning (sparsity 80%) and INT8 quantization compression, and finally a lightweight end-side model package (387 kB) and a high-performance side-side model package (2.1 MB) are generated; at T3 stage, the cloud issues model difference package through TLS 1.3 encrypted channel, and guarantees the anti-tampering process through ECDSA-P256 digital signature mechanism; to T4 stage, the edge gateway verifies the signature and writes the model into the new partition of eMMC storage, and calculates the return CRC32 check code; finally at T5 stage, the cloud checks the CRC32 without error and issues the "commit" instruction to complete the model switching, if the check fails, it will automatically roll back to the last version W_t, thereby realizing a high-reliability, tamper-proof model OTA update closed loop, fully meeting the technical requirements of the model online upgrade in the claims.

[0031] Step 104, if the second diagnostic cycle confidence is greater than the preset threshold, it is determined that the battery cluster has a real fault, and a control instruction of reducing the power of the battery cluster or stopping is triggered by the EMS.

[0032] This step realizes safe closed-loop control with the upper management system through industrial communication interface. The extended frame format and register mapping definition for communication with EMS are as follows: Figure 8 As shown, when the final diagnostic confidence exceeds the preset threshold (e.g. 90%), the edge gateway writes the final fault type (e.g. 0x0001 = thermal runaway, 0x0002 = internal short circuit, 0x0003 = jam, 0x0004 = internal leakage), confidence and Unix timestamp, etc. into the specific holding register (e.g. 40001-40008) of EMS according to the Figure 7 definition format. The EMS immediately triggers a control instruction to reduce the power of the battery cluster or stop, realizing safe closed-loop. The average recognition accuracy of the device for the above four types of faults is ≥97.8%, and the false positive rate is ≤0.3%.

[0033] The cabinet arrangement and wiring diagram of the battery cluster fault diagnosis device can be seen from Figure 9 , which clearly shows the physical size, installation method (such as DIN rail, 1U rack), wiring specification (such as power line diameter, bus terminal resistance, shielding layer ground) of interface board, end-side node and edge gateway, etc., ensuring the engineering implementability and environmental adaptability of the device.

[0034] Another embodiment of the present application relates to a battery cluster fault diagnosis system, and the implementation details of the battery cluster fault diagnosis system of the present embodiment are described below. The following implementation details are provided for the convenience of understanding and are not essential for implementing the present solution.

[0035] Specifically, the battery cluster interface board module is configured to obtain multi-modal time series data of each battery module in the battery cluster within a first diagnosis period. The end-side computing node module is configured to process the multi-modal time series data of each battery module in the battery cluster by an A-ConvNeXt model to obtain a preliminary diagnosis result and a corresponding confidence. If the confidence is greater than a preset threshold, a feature vector corresponding to the preliminary diagnosis result is output. The A-ConvNeXt model replaces the depth separable convolution module of the original ConvNeXt model with a cascaded structure of a depth separable convolution and a channel attention module. The channel attention module adaptively adjusts the weight distribution of the multi-modal time series data based on the feature importance of different modalities in the multi-modal time series data and the relevance to the fault mode. The edge gateway module is configured to fuse the feature vector data of the corresponding battery cluster within a plurality of consecutive first diagnosis periods to form second diagnosis period data, wherein the second diagnosis period is greater than the first diagnosis period. The TCN-TF fusion network processes the second diagnosis period data of the corresponding battery cluster to obtain a final diagnosis result and a corresponding confidence. The TCN-TF fusion network is composed of a time convolution network and a Transformer encoder. The time convolution network extracts local features and mutation information of the second diagnosis period data, and the Transformer encoder captures long-range temporal dependencies of the second diagnosis period data. The local features, mutation information, and long-range temporal dependencies are concatenated as fusion features to input into a fully connected layer. The industrial communication interface module is configured to determine that the battery cluster has a real fault if the confidence is greater than the preset threshold, and trigger the EMS to issue a control instruction to reduce the power of the battery cluster or shut down.

[0036] The cloud server module is further configured to train the A-ConvNeXt model of the end-side computing node module and the TCN-TF fusion network of the edge gateway module, and regularly update the model parameters of the end-side computing node module and the edge gateway module.

[0037] The steps of the above methods are divided only for clarity, and can be combined into one step or split into multiple steps during implementation, as long as the same logical relationship is included, which is within the scope of the present application. Adding insignificant modifications or introducing insignificant designs in the algorithm or process does not change the core design of the algorithm and process, which is within the scope of the present application.

[0038] Another embodiment of the present application relates to a computer readable storage medium storing a computer program. The computer program, when executed by a processor, implements the method embodiments described above.

[0039] That is, those skilled in the art can understand that all or part of the steps of the method in the above embodiments can be completed by a program instructing the relevant hardware, the program is stored in a storage medium, and includes a plurality of instructions for causing a device (which can be a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0040] Those skilled in the art can understand that the above embodiments are specific embodiments for implementing the present application, and in actual applications, various changes can be made in form and details without departing from the spirit and scope of the present application.

Claims

1. A battery cluster fault diagnosis system, characterized in that, The system includes: The battery cluster interface board module is used to acquire multimodal timing data of each battery module in the battery cluster during the first diagnostic cycle. The multimodal timing data includes voltage, current, temperature and acoustic emission signals. The edge computing node module is used to process multimodal time-series data within the first diagnostic cycle using the A-ConvNeXt model to obtain the diagnostic results and confidence levels for the first diagnostic cycle. If the confidence level for the first diagnostic cycle is greater than a preset threshold, the module outputs the first diagnostic cycle feature vector corresponding to the diagnostic results. The A-ConvNeXt model replaces the depthwise separable convolution module of the original ConvNeXt model with a cascaded structure of depthwise separable convolution and channel attention modules. The channel attention module adaptively adjusts the weight allocation of the multimodal time-series data based on the feature importance of different modes in the multimodal time-series data and their correlation with the fault mode. The edge gateway module is used to fuse multiple consecutive first diagnostic cycle feature vectors of the corresponding battery cluster along the time axis to form second diagnostic cycle data; the second diagnostic cycle data of the corresponding battery cluster is processed by the TCN-TF fusion network to obtain the diagnostic results and confidence levels of the second diagnostic cycle; the TCN-TF fusion network is composed of a temporal convolutional network and a Transformer encoder. The temporal convolutional network extracts local features and mutation information of the second diagnostic cycle data, and the Transformer encoder captures the long-range time dependencies of the second diagnostic cycle data. The local features, mutation information, and long-range time dependencies are concatenated and nonlinearly fused through a fully connected layer to output the diagnostic results and confidence levels of the second diagnostic cycle. The industrial communication interface module is used to determine that the battery cluster has a real fault if the confidence level of the second diagnostic cycle is greater than a preset threshold, and to trigger the EMS to issue a control command to reduce the power of the battery cluster or shut down the system.

2. The battery cluster fault diagnosis system as described in claim 1, characterized in that, It also includes a cloud server module, which is used to train the A-ConvNeXt model of the edge computing node module and the TCN-TF fusion network of the edge gateway module, and to update the model parameters of the edge computing node module and the edge gateway module regularly.

3. A method for diagnosing battery cluster faults, characterized in that, The method includes: During the first diagnostic cycle, acquire multimodal timing data of each battery module in the battery cluster, including voltage, current, temperature and acoustic emission signals; The A-ConvNeXt model processes multimodal time-series data within the first diagnostic period to obtain the diagnostic results and confidence levels for the first diagnostic period. If the confidence level for the first diagnostic period is greater than a preset threshold, the corresponding feature vector for the first diagnostic period is output. The A-ConvNeXt model replaces the depthwise separable convolutional module of the original ConvNeXt model with a cascaded structure of depthwise separable convolution and channel attention modules. The channel attention module adaptively adjusts the weight allocation of the multimodal time-series data based on the feature importance of different modalities and their correlation with the fault mode. The feature vectors of multiple consecutive first diagnostic cycles for the corresponding battery cluster are fused along the time axis to form the second diagnostic cycle data. The second diagnostic cycle data of the corresponding battery cluster is processed by the TCN-TF fusion network to obtain the diagnostic results and confidence levels of the second diagnostic cycle. The TCN-TF fusion network is composed of a temporal convolutional network and a Transformer encoder. The temporal convolutional network extracts local features and mutation information of the second diagnostic cycle data, and the Transformer encoder captures the long-range time dependencies of the second diagnostic cycle data. The local features, mutation information, and long-range time dependencies are concatenated and nonlinearly fused through a fully connected layer to output the diagnostic results and confidence levels of the second diagnostic cycle. If the confidence level in the second diagnostic cycle is greater than the preset threshold, it is determined that there is a real fault in the battery cluster, triggering the EMS to issue a control command to reduce the power of the battery cluster or shut it down.

4. The battery cluster fault diagnosis method as described in claim 3, characterized in that, The training of the A-ConvNeXt model and the TCN-TF fusion network includes the following steps: The multimodal time-series data of each battery cluster in the energy storage power station are used as independent sub-training sets, with each sub-training set corresponding to one battery cluster. Each sub-training set independently trains the TCN-TF fusion network for its corresponding battery cluster, generating TCN-TF fusion network parameters specific to that battery cluster after training. The TCN-TF fusion network parameters corresponding to each battery cluster are uploaded to a cloud server. The cloud server collects the TCN-TF fusion network parameters of all battery clusters, performs a weighted average according to the data volume ratio of each sub-training set, and aggregates them to generate new global model parameters. The aggregated global model parameters are then distributed to the TCN-TF fusion network specific to each battery cluster, overwriting the original TCN-TF fusion network parameters. Each battery cluster continues to train using the sub-training sets based on the updated TCN-TF fusion network parameters. The above operations are repeated until the TCN-TF fusion networks of all battery clusters converge. Using the knowledge distillation method, the TCN-TF fusion network is used as the teacher model, and the diagnostic knowledge output by the TCN-TF fusion network guides the learning of the A-ConvNeXt student model, thereby correcting the parameters of the A-ConvNeXt model.

5. The battery cluster fault diagnosis method as described in claim 3, characterized in that, It also includes performing INT8 quantization on the A-ConvNeXt model.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, is capable of performing the battery cluster fault diagnosis method as defined in any one of claims 3 to 5.

Images