Abnormality detection method for multi-series-parallel battery module

Abstract

The invention belongs to the technical field of battery management systems, and particularly relates to a multi-series-parallel battery module anomaly detection method, which is characterized in that a micro artificial intelligence coprocessor is embedded in each battery module, and is integrated with a voltage sampling unit and a temperature sampling unit; the method comprises the following steps: synchronously acquiring voltage and temperature data of each battery cell in a module according to a preset sampling period of 1 millisecond, and carrying out moving average filtering and outlier elimination preprocessing on the acquired original data; a miniature artificial intelligence coprocessor is embedded into each battery module, and a lightweight long-short-term memory network model is deployed, so that localized real-time initial judgment of abnormal signals is realized, and data processing and transmission delay is remarkably reduced; through a cascade confirmation mechanism constructed by a local area network flexible data rate bus based on a high-speed controller, a plurality of modules can rapidly and cooperatively verify abnormity, real faults and instantaneous interference are effectively distinguished, and the false alarm rate and the missed detection risk of the system are greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing, specifically a method for detecting anomalies in multi-series and parallel battery modules. Background Technology

[0002] As the safety requirements for electric vehicles and energy storage systems continue to increase, the Battery Management System (BMS), as a core component ensuring battery reliability, is receiving increasing attention for its architecture design and response performance. Traditional centralized BMS architectures upload raw data such as voltage, temperature, and current from each battery module to a central control unit for anomaly detection and decision-making. While this facilitates global management, it exposes significant bottlenecks in high-concurrency, high-dynamic operating scenarios. Especially in sudden safety events such as thermal runaway, the system's anomaly response time requirements are extremely stringent, needing to meet the millisecond-level response threshold specified in ASIL-C level of the functional safety standard ISO 26262.

[0003] Among these, edge-collaboration-based real-time anomaly detection technology has become a key direction for improving BMS responsiveness. This technology aims to decentralize some intelligent decision-making capabilities to the battery module level, reducing reliance on central nodes through localized data processing, thereby reducing latency introduced by communication links. Its core lies in building a novel architecture capable of efficiently executing lightweight intelligent algorithms at the module level and supporting rapid collaborative verification among multiple modules, enabling the immediate capture and confirmation of early anomaly signals.

[0004] However, in existing centralized processing models, module data needs to be encapsulated through multiple protocols before being uploaded via CAN or similar buses, resulting in communication delays generally exceeding 200 milliseconds, far from meeting the time window for rapid intervention in the early stages of thermal runaway. Simultaneously, traditional BMS lacks the ability to deploy low-power, high-efficiency anomaly identification models at the module level, making it impossible to perform real-time preliminary judgments on key characteristics such as voltage drops and abnormal temperature rise rates. Furthermore, even when some systems introduce local threshold alarm mechanisms, the lack of cross-module dynamic collaborative verification capabilities makes it difficult to distinguish between real faults and transient interference, leading to high false alarm rates or increased risks of missed detections. Therefore, this paper proposes an anomaly detection method for multi-series and parallel battery modules to address these issues. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an anomaly detection method for multi-series and parallel battery modules, thereby solving the technical problems mentioned in the background art.

[0006] To address the above technical problems, the technical solution adopted by this invention is as follows:

[0007] A method for detecting anomalies in multi-series-parallel battery modules includes the following specific steps:

[0008] Step 1: Embed a micro artificial intelligence coprocessor in each battery module. The coprocessor integrates a voltage sampling unit and a temperature sampling unit. It synchronously collects the voltage and temperature data of each cell in the module at a preset sampling period of 1 millisecond, and performs moving average filtering and outlier removal preprocessing on the collected raw data.

[0009] Step 2: Deploy and run a pruned and optimized long short-term memory neural network model on the micro artificial intelligence coprocessor. The model takes the voltage and temperature sequences preprocessed in Step 1 as input, extracts time-series features through its internal gated recurrent unit structure, and outputs an abnormal probability score of the module's operating state at the current moment.

[0010] Step 3: When the anomaly probability score of any battery module exceeds the first set threshold, the micro artificial intelligence coprocessor of the module immediately broadcasts a cascading confirmation request message to other adjacent battery modules in the system through the Controller Area Network Flexible Data Rate Bus. The request message contains the identifier of the triggering module and its anomaly probability score.

[0011] Step 4: The adjacent battery module that receives the cascaded confirmation request message performs collaborative reasoning calculation locally based on its own real-time operating data and through its deployed pruned and optimized long short-term memory neural network model, and replies to the triggering module with a confirmation response message within 2 milliseconds. This response message contains the abnormal confidence level calculated locally.

[0012] Step 5: After receiving confirmation response messages from all adjacent modules, the trigger module makes a decision based on the preset voting fusion strategy. If the number of modules with confirmed abnormalities exceeds the second preset threshold, a final abnormal alarm signal is generated and reported to the main control unit of the battery management system through the flexible data rate bus of the controller area network. At the same time, the preset local protection action is activated.

[0013] Step 6: After receiving the final abnormal alarm signal, the main control unit of the battery management system executes the corresponding global control commands in conjunction with the system-level operation strategy, including adjusting the charging and discharging power, activating the thermal management subsystem, or performing fault isolation.

[0014] Preferably, the micro AI coprocessor embedded in step 1 is a low-power system-on-a-chip with a 40-nanometer process. Its core computing unit is a dual-core ARM Cortex-M7 processor with a main frequency of 400 MHz, and it integrates a hardware floating-point unit and a direct memory access controller. The static power consumption of the coprocessor is less than 50 milliwatts, and the dynamic power consumption does not exceed 300 milliwatts when running at full load.

[0015] Preferably, in step 1, the voltage sampling unit has a sampling accuracy of ±1 millivolt, the temperature sampling unit uses a negative temperature coefficient thermistor with a temperature measurement accuracy of ±0.5 degrees Celsius, the sampling data is digitized by a 24-bit analog-to-digital converter, and transmitted to the coprocessor's internal memory via direct memory access.

[0016] Preferably, the pruning-optimized long short-term memory neural network model deployed in step 2 has an original network structure containing two hidden layers with 128 neurons in each layer. Through an amplitude-based weight pruning algorithm, the number of network parameters is compressed to 30% of the original number, the model inference latency is controlled within 5 milliseconds, and the accuracy loss is less than 2%.

[0017] Preferably, the input features of the long short-term memory neural network model in step 2 include: voltage data, temperature data, voltage change rate, temperature change rate, and the difference between the average voltage of the module and the average voltage of the system within a sliding window with a length of 50 sampling points. The output is an anomaly probability score between 0 and 1. If the score exceeds 0.85, it is judged as a preliminary anomaly.

[0018] Preferably, in step 3, the communication rate of the Controller Area Network Flexible Data Rate Bus is 2 megabits per second, and the cascade confirmation request message adopts the standard data frame format, containing an 11-bit identifier and a data field length of 8 bytes, used to carry the trigger module identifier, anomaly probability score and timestamp information.

[0019] Preferably, in the collaborative reasoning calculation process in step 4, after receiving a request, adjacent modules prioritize using local data from the most recent 10 sampling periods to perform model reasoning. If local data is insufficient, instantaneous data at the current moment is used for calculation to ensure a response is completed within 2 milliseconds.

[0020] Preferably, the voting fusion strategy in step 5 is as follows: the second set threshold is set to 50% of the total number of modules participating in collaborative confirmation, and the abnormal probability score of the triggering module itself must exceed the first set threshold for three consecutive sampling periods before the collaborative confirmation result can be judged as the final abnormality.

[0021] Preferably, the local protection action in step 5 includes: cutting off the charging and discharging circuit of the module through the relay built into the module, and activating the module-level thermal management device, such as a fan or a thermoelectric cooler, to perform forced heat dissipation.

[0022] Preferably, the global control instructions executed by the main control unit of the battery management system in step 6 also include recording detailed data of abnormal events, including trigger time, involved modules, abnormal type and collaborative confirmation results, and updating the system health status database for subsequent trend analysis and predictive maintenance.

[0023] Preferably, the method further includes an online model update mechanism, in which the battery management system master controller periodically, for example every 24 hours, summarizes the operating data and anomaly records of each module, retrains and optimizes the long short-term memory neural network model on the cloud or local server, and distributes the updated model parameters to the micro artificial intelligence coprocessors of each module through the controller area network flexible data rate bus.

[0024] Preferably, the micro AI coprocessor also integrates a hardware security module for encrypting and storing model parameters and key data during transmission, ensuring data integrity and tamper resistance.

[0025] Preferably, during the initialization phase, the method performs time synchronization on the micro AI coprocessors of all battery modules with a synchronization accuracy better than 10 microseconds, and establishes a topological connection table between modules, which is stored in the non-volatile memory of each coprocessor to guide the cascading confirmation process.

[0026] Preferably, the training data for the pruning-optimized long short-term memory neural network model comes from historical datasets covering various battery types, operating conditions, and anomaly patterns, with a dataset size exceeding 1 million samples. Data augmentation techniques are used to expand the samples under boundary conditions and noise interference.

[0027] Preferably, the method is applicable to a battery pack consisting of no less than 10 battery modules connected in series, each module containing no less than 12 cells, and the overall system anomaly detection response time from data acquisition to final alarm generation is ensured to be completed within 50 milliseconds.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] By embedding a micro-AI coprocessor in each battery module and deploying a lightweight long short-term memory network model, localized real-time initial judgment of abnormal signals is achieved, significantly reducing data processing and transmission latency. A cascading verification mechanism built on a flexible data rate bus of a local area network based on a high-speed controller enables multiple modules to quickly and collaboratively verify anomalies, effectively distinguishing between real faults and transient interference, and significantly reducing the system's false alarm rate and the risk of missed detections. The synergistic effect of these technical features compresses the response time of the entire battery management system to severe abnormal events such as thermal runaway to within 50 milliseconds, meeting high standards of functional safety. Furthermore, the local intelligent decision-making reduces reliance on the central controller, improving the robustness and scalability of the system architecture. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] In the attached diagram:

[0032] Figure 1 This is a schematic diagram of the overall technical solution architecture of the multi-series parallel battery module anomaly detection method proposed in this invention;

[0033] Figure 2 This is a schematic diagram of the core principle framework of the local anomaly initial judgment based on pruning and optimizing long short-term memory neural network in this invention;

[0034] Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow between battery modules based on cascaded confirmation requests and collaborative reasoning in this invention;

[0035] Figure 4 This is a flowchart illustrating the end-to-end real-time anomaly detection logic process from data acquisition to final alarm generation in this invention.

[0036] Figure 5 This is a schematic diagram of the supporting framework for online model updates and closed-loop management of system health status in this invention. Detailed Implementation

[0037] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0038] Specific implementation examples are given below.

[0039] Example

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0041] Currently, with the increasing safety requirements of electric vehicles and energy storage systems, battery management systems (BMS), as a core component ensuring battery reliability, are receiving growing attention for their architecture design and response performance. Traditional centralized BMS architectures upload raw data such as voltage, temperature, and current from each battery module to a central control unit for anomaly detection and decision-making. While this facilitates global management, it exposes significant bottlenecks in high-concurrency, high-dynamic operating scenarios. Especially in sudden safety events such as thermal runaway, the system's anomaly response time requirements are extremely stringent, needing to meet the millisecond-level response threshold specified in the ASIL-C level of the functional safety standard ISO 26262. Therefore, edge-collaboration-based real-time anomaly detection technology has become a key direction for improving the responsiveness of BMS. This technology aims to offload some intelligent decision-making capabilities to the battery module level, reducing reliance on central nodes through localized data processing, thereby reducing latency introduced by communication links. Its core lies in building a new architecture capable of efficiently executing lightweight intelligent algorithms at the module level and supporting rapid collaborative verification between multiple modules, enabling immediate capture and confirmation of early anomaly signals.

[0042] However, in the existing centralized processing mode, module data needs to be encapsulated through multiple levels of protocols and then uploaded through the controller LAN or similar bus. The communication delay is generally more than 200 milliseconds, which is far from meeting the time window for rapid intervention in the early stage of thermal runaway.

[0043] Meanwhile, traditional battery management systems lack the ability to deploy low-power, high-efficiency anomaly identification models at the module level, making it impossible to perform real-time preliminary judgments on key characteristics such as voltage drops and abnormal temperature rise rates. Furthermore, even when some systems introduce local threshold alarm mechanisms, the lack of cross-module dynamic collaborative verification capabilities makes it difficult to distinguish between real faults and transient interference, leading to high false alarm rates or increased risks of missed detections. Therefore, there is an urgent need for an anomaly detection method for multi-series and parallel battery modules to reduce the anomaly response time to less than 50 milliseconds while ensuring functional safety.

[0044] To address the aforementioned technical challenges, this invention proposes a method for localized, real-time initial assessment of abnormal signals by embedding a micro-AI coprocessor in each battery module and deploying a lightweight long short-term memory network model. This significantly reduces data processing and transmission latency. Furthermore, a cascaded verification mechanism built upon a flexible data rate bus on a local area network based on a high-speed controller enables multiple modules to quickly and collaboratively verify anomalies, effectively distinguishing between real faults and transient interference, and drastically reducing the system's false alarm rate and the risk of missed detection. The synergistic effect of these technical features reduces the response time of the entire battery management system to severe abnormal events such as thermal runaway to less than 50 milliseconds, meeting high standards for functional safety. Simultaneously, the local intelligent decision-making reduces reliance on the central controller, improving the robustness and scalability of the system architecture. This method can be applied to anomaly detection methods for multi-series and parallel battery modules.

[0045] refer to Figure 1 The schematic diagram of the overall technical architecture of the multi-series-parallel battery module anomaly detection method proposed in this invention illustrates the core components of the system. This architecture includes multiple battery modules, each integrating a micro-AI coprocessor, a voltage sampling unit, a temperature sampling unit, a relay, a thermal management device, and a Controller Area Network (CAN) flexible data rate bus interface. All modules are interconnected via the CAN bus and communicate with the main control unit of the battery management system. (Reference) Figure 4 The end-to-end real-time anomaly detection logic flowchart in this invention, from data acquisition to final alarm generation, details the execution sequence of the method. The entire process strictly follows six preset core steps, ensuring that the entire process from raw data acquisition to final alarm generation is completed within 50 milliseconds.

[0046] In the above-mentioned method for detecting anomalies in multi-series and parallel battery modules, step 1 involves embedding a micro artificial intelligence coprocessor in each battery module. The coprocessor integrates a voltage sampling unit and a temperature sampling unit, and synchronously collects the voltage and temperature data of each cell in the module at a preset sampling period of 1 millisecond. The collected raw data is then preprocessed by moving average filtering and outlier removal.

[0047] Specifically, the micro AI coprocessor uses a low-power system-on-a-chip (SoC) manufactured using a 40nm process. Its core computing unit is a dual-core ARM Cortex-M7 processor with a clock speed of 400 MHz, and it integrates a hardware floating-point unit and a direct memory access controller. The coprocessor's static power consumption is less than 50 milliwatts, and its dynamic power consumption does not exceed 300 milliwatts under full load. This coprocessor is physically encapsulated within the protective casing of each battery module and is directly electrically connected to 12 or more battery cells within the module. The voltage sampling unit has a sampling accuracy of ±1 millivolt, and the temperature sampling unit uses a negative temperature coefficient thermistor with a temperature measurement accuracy of ±0.5 degrees Celsius. The sampled data is digitized via a 24-bit analog-to-digital converter and transmitted to the coprocessor's internal memory via direct memory access. During the initialization phase, the micro AI coprocessors of all battery modules execute a precise time synchronization protocol through the Controller Area Network (CAN) flexible data rate bus, with a synchronization accuracy better than 10 microseconds, ensuring that the data acquisition from all modules is strictly aligned on the timeline.

[0048] The preprocessing process consists of two consecutive sub-operations:

[0049] First, a moving average filter of length 5 is applied to the original voltage and temperature sequences to smooth high-frequency noise;

[0050] Secondly, outlier removal is performed by calculating the deviation between the current sample point and the mean of the previous 5 sample points. If the absolute value of the deviation exceeds 3 times the standard deviation, the point is replaced with the mean of the sliding window. The preprocessed data is stored in a circular buffer inside the coprocessor, with a depth of 100 sample points, providing a stable data source for subsequent model inference.

[0051] In the above-mentioned method for detecting anomalies in multi-series and parallel battery modules, step 2 involves deploying and running a pruned and optimized long short-term memory neural network model on a micro artificial intelligence coprocessor. This model takes the voltage and temperature sequences preprocessed in step 1 as inputs, extracts temporal features through its internal gated recurrent unit structure, and outputs an anomaly probability score of the module's operating state at the current moment.

[0052] Specifically, the deployed pruned and optimized Long Short-Term Memory (LSTM) neural network model has an original network structure containing two hidden layers, each with 128 neurons. Through an amplitude-based weight pruning algorithm, the number of network parameters is compressed to 30% of the original number, the model inference latency is controlled within 5 milliseconds, and the accuracy loss is less than 2%. The model's input features include: voltage data, temperature data, voltage change rate, temperature change rate, and the difference between the module's average voltage and the system's average voltage within a sliding window of 50 sampling points.

[0053] The voltage and temperature change rates are calculated by first-order difference between adjacent sampling points within a sliding window. The system average voltage is periodically (e.g., every 100 milliseconds) broadcast to each module by the battery management system's main control unit. The model output is an anomaly probability score between 0 and 1; a score exceeding 0.85 is considered a preliminary anomaly.

[0054] Prior to deployment, the model was trained on a historical dataset containing over 1 million samples, covering various battery types, operating conditions, and anomaly patterns. Data augmentation techniques were used to expand the dataset to include samples under boundary conditions and noise interference. The model is stored as quantized integers in the coprocessor's non-volatile memory, and the inference process is accelerated by a hardware floating-point unit. Each inference operation begins in the latter half of a 1-millisecond sampling period, ensuring computation is completed before the start of the next sampling period, thus maintaining strict real-time performance.

[0055] In the above-mentioned method for detecting anomalies in multiple series and parallel battery modules, in step 3, when the anomaly probability score of any battery module exceeds the first set threshold, the micro artificial intelligence coprocessor of the module immediately broadcasts a cascading confirmation request message to other adjacent battery modules in the system through the Controller Area Network Flexible Data Rate Bus. The request message contains the identifier of the triggering module and its anomaly probability score.

[0056] Specifically, the communication rate of the Controller Area Network (CAN) Flexible Data Rate Bus (FDBB) is 2 Mbps. The cascading acknowledgment request message adopts the standard data frame format, containing an 11-bit identifier and an 8-byte data field length, used to carry the trigger module identifier, anomaly probability score, and timestamp information. The trigger module identifier is its unique physical address in the battery pack, ranging from 1 to N (N is the total number of modules, not less than 10). The anomaly probability score is encoded as an 8-bit unsigned integer, ranging from 0 to 255, corresponding to a floating-point value from 0.0 to 1.0. The timestamp information is a microsecond-level count value at the trigger moment, used by the receiver for data alignment. The broadcast behavior is driven by the coprocessor's interrupt service routine. Once the output of step 2 exceeds the threshold of 0.85, a message is immediately constructed and sent through the CANBB interface.

[0057] refer to Figure 3 The multi-level interaction relationship and data flow diagram between battery modules based on cascaded confirmation requests and collaborative reasoning in this invention clearly illustrates this broadcast process. The topology connection table between modules established during the initialization phase is stored in the non-volatile memory of each coprocessor. This table defines the list of adjacent modules for each module, ensuring that broadcasts are only sent to physically adjacent modules, rather than being broadcast across the entire network, thereby optimizing bus bandwidth utilization.

[0058] In the above-mentioned method for detecting anomalies in multiple series and parallel battery modules, in step 4, the adjacent battery module that receives the cascaded confirmation request message performs collaborative reasoning calculation locally based on its own real-time operating data through its deployed pruned and optimized long short-term memory neural network model, and replies to the triggering module with a confirmation response message within 2 milliseconds. This response message contains the anomaly confidence level calculated locally.

[0059] Specifically, in the collaborative inference computation process, after receiving a request, adjacent modules prioritize using local data from the most recent 10 sampling periods to perform model inference. If local data is insufficient, they use instantaneous data from the current moment for computation, ensuring a response is completed within 2 milliseconds. Upon receiving the cascading confirmation request message, the receiving module's coprocessor immediately parses out the triggering module identifier and timestamp.

[0060] Subsequently, data from the corresponding time and the 10 sampling periods before and after it are extracted from its circular buffer based on the timestamp, forming a new 50-point sliding window (filled with the latest data if insufficient), and input into the same pruned and optimized long short-term memory neural network model deployed locally for inference. This process reuses the model from step 2, but the input data is the receiver's own runtime data. The calculated anomaly confidence score is also a value between 0 and 1, encoded as an 8-bit unsigned integer. The acknowledgment response message is unicast back to the trigger module via the Controller Area Network Flexible Data Rate Bus, and its data field contains the receiver module identifier and the calculated anomaly confidence score. The entire link of receiving, processing, calculating, and responding must be completed within 2 milliseconds, which places extremely high demands on the computing power of the coprocessor and the efficiency of the communication protocol stack.

[0061] In the above-mentioned method for detecting anomalies in multiple series and parallel battery modules, in step 5, after the trigger module receives the confirmation response messages from all adjacent modules, it makes a decision based on the preset voting fusion strategy. If the number of modules confirmed to have anomalies exceeds the second preset threshold, a final anomaly alarm signal is generated and reported to the main control unit of the battery management system through the flexible data rate bus of the controller area network, and at the same time, the preset local protection action is activated.

[0062] Specifically, the voting fusion strategy is as follows: a second threshold is set at 50% of the total number of modules participating in collaborative confirmation, and the triggering module's own anomaly probability score must exceed the first threshold for three consecutive sampling periods before the collaborative confirmation result can be determined as a final anomaly. The triggering module internally maintains a timer and a counter. The timer ensures its own anomaly state lasts for at least 3 milliseconds (i.e., 3 sampling periods); the counter counts the number of modules in the reply confirmation message with an anomaly confidence score exceeding 0.85. Only when the timer times out and the counter value is greater than or equal to 50% of the total number of adjacent modules is it determined as a final anomaly.

[0063] Once the determination is successful, the trigger module immediately performs two operations:

[0064] On the one hand, the final abnormal alarm signal message is constructed and reported to the main control unit of the battery management system through the flexible data rate bus of the controller area network;

[0065] On the other hand, preset local protection actions are activated. These local protection actions include: cutting off the module's charging and discharging circuit via a relay built into the module, and activating module-level thermal management devices, such as fans or thermoelectric coolers, for forced heat dissipation. These actions are directly controlled by the coprocessor's general purpose input / output pins, with response delays in the microsecond range.

[0066] In the above-mentioned method for detecting anomalies in multi-series and parallel battery modules, in step 6, after receiving the final anomaly alarm signal, the main control unit of the battery management system executes corresponding global control commands in conjunction with the system-level operation strategy, including adjusting the charging and discharging power, activating the thermal management subsystem, or performing fault isolation.

[0067] Specifically, the global control commands executed by the battery management system's main control unit also include recording detailed data on abnormal events, including trigger time, involved modules, abnormality type, and collaborative confirmation results, and updating the system health status database for subsequent trend analysis and predictive maintenance. Upon receiving an alarm signal, the main control unit first parses the message content to obtain precise information about the abnormal module.

[0068] Subsequently, based on the preset system-level operation strategy, various measures may be taken: for example, sending a request to the vehicle controller to reduce the motor output power; or activating the battery pack-level liquid cooling thermal management system to cool the entire battery pack.

[0069] Alternatively, by controlling a high-voltage relay, the faulty module can be physically isolated from the battery string to prevent the fault from spreading.

[0070] Simultaneously, the main control unit packages and stores all relevant data from this anomaly, including complete raw sampling data fragments, collaborative confirmation results from each module, and decision-making process logs, into a system health status database in non-volatile memory. This database provides the data foundation for subsequent offline analysis, model retraining, and predictive maintenance.

[0071] In addition, the method also includes an online model update mechanism. The battery management system master controller periodically, for example every 24 hours, summarizes the operating data and anomaly records of each module, retrains and optimizes the long short-term memory neural network model on the cloud or local server, and distributes the updated model parameters to the micro artificial intelligence coprocessors of each module through the controller area network flexible data rate bus.

[0072] refer to Figure 5 The schematic diagram of the supporting framework for online model updates and closed-loop management of system health status in this invention illustrates this closed-loop process. The main control unit periodically collects anonymized operational features and abnormal samples from each module and uploads them to the cloud training platform. The training platform uses incremental learning algorithms to fine-tune the original model, generating a new model with superior performance. The new model parameters, after being encrypted and signed, are distributed via the flexible data rate bus of the controller area network. Upon receiving the update package, the micro-AI coprocessor of each module verifies the signature using its integrated hardware security module and completes the atomic replacement of the model in a secure environment, ensuring that the system continues to operate normally during the update process.

[0073] Furthermore, the miniature AI coprocessor integrates a hardware security module for encrypted storage and transmission of model parameters and critical data, ensuring data integrity and tamper resistance. This hardware security module supports encryption algorithms such as AES-128, and all model parameters and system configurations stored in non-volatile memory are encrypted. All critical messages on the Controller Area Network's Flexible Data Rate Bus, such as cascading acknowledgment requests, acknowledgment responses, and model update packets, include HMAC-based message authentication codes to prevent man-in-the-middle attacks and data tampering.

[0074] To illustrate the implementation effect of this invention, a battery pack application scenario consisting of 12 battery modules connected in series is constructed. Each module contains 16 cells. After system initialization, the micro-AI coprocessors of all modules complete time synchronization with a synchronization accuracy of 8 microseconds and establish their respective adjacent module lists (each module has 2 adjacent modules, except for the first and last modules).

[0075] Under normal operating conditions, each module collects data and performs local inference at 1-millisecond intervals, resulting in a stable system load. Assume that at t=1000 milliseconds, a micro-short circuit in module 6 causes a sudden drop of 50 millivolts in the voltage of its cell 8 within 2 milliseconds, while the temperature rises at a rate of 10 degrees Celsius per second. The coprocessor of module 6 completes inference at t=1001 milliseconds, outputting an anomaly probability score of 0.92, exceeding the 0.85 threshold. At t=1001.1 milliseconds, it immediately broadcasts a cascaded acknowledgment request to its neighboring modules 5 and 7. Modules 5 and 7 receive the request at t=1001.5 milliseconds, perform collaborative inference using their local data, calculate anomaly confidence scores of 0.88 and 0.90 respectively, and reply with an acknowledgment response before t=1002.8 milliseconds. Module 6 received all responses at t=1003 milliseconds. At this time, its abnormal state had lasted for 3 cycles (t=1001, 1002, 1003), and 2 out of 2 adjacent modules (100%) confirmed the abnormality, exceeding the 50% threshold.

[0076] Therefore, at t=1003.1 milliseconds, module 6 generates the final abnormal alarm signal and reports it to the main control unit, while simultaneously cutting off its own circuit and starting the fan. The main control unit receives the alarm at t=1003.5 milliseconds, immediately sends a power reduction command to the vehicle controller, and activates the liquid cooling system. From the occurrence of the abnormality (t=1000) to the generation of the final alarm (t=1003.1), the total delay is 3.1 milliseconds, far below the requirement of 50 milliseconds, fully verifying the effectiveness of the invention.

[0077] In the description of this invention, it should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions provided in this disclosure can be achieved, and no limitation is imposed herein.

[0078] The above are merely preferred embodiments of the present invention and do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for detecting anomalies in multi-series and parallel battery modules, characterized in that, include: Step 1: Embed a micro AI coprocessor in each battery module. The coprocessor integrates a voltage sampling unit and a temperature sampling unit, and synchronously collects the voltage and temperature data of each cell in the module at a sampling period of 1 millisecond. It also performs moving average filtering and outlier removal preprocessing on the collected raw data. Step 2: Deploy and run a pruned and optimized long short-term memory neural network model on the micro artificial intelligence coprocessor. The model takes the voltage and temperature sequences preprocessed in Step 1 as input, extracts time-series features through its internal gated recurrent unit structure, and outputs an abnormal probability score of the module's operating state at the current moment. Step 3: When the anomaly probability score of any battery module exceeds the first set threshold, the micro artificial intelligence coprocessor of the module immediately broadcasts a cascading confirmation request message to other adjacent battery modules in the system through the Controller Area Network Flexible Data Rate Bus. The request message contains the identifier of the triggering module and its anomaly probability score. Step 4: The adjacent battery module that receives the cascaded confirmation request message performs collaborative reasoning calculation locally based on its own real-time operating data and through its deployed pruned and optimized long short-term memory neural network model, and replies to the triggering module with a confirmation response message within 2 milliseconds. This response message contains the abnormal confidence level calculated locally. Step 5: After receiving confirmation response messages from all adjacent modules, the trigger module makes a decision based on the preset voting fusion strategy. If the number of modules with confirmed abnormalities exceeds the second preset threshold, a final abnormal alarm signal is generated and reported to the main control unit of the battery management system through the flexible data rate bus of the controller area network. At the same time, the preset local protection action is activated. Step 6: After receiving the final abnormal alarm signal, the main control unit of the battery management system executes the corresponding global control commands in conjunction with the system-level operation strategy, including adjusting the charging and discharging power, activating the thermal management subsystem, or performing fault isolation.

2. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The micro AI coprocessor uses a low-power system-on-a-chip with a 40-nanometer process. Its core computing unit is a dual-core ARM Cortex-M7 processor with a main frequency of 400 MHz. It also integrates a hardware floating-point unit and a direct memory access controller. The coprocessor's static power consumption is less than 50 milliwatts, and its dynamic power consumption does not exceed 300 milliwatts when running at full load.

3. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The voltage sampling unit has a sampling accuracy of ±1 millivolt, and the temperature sampling unit uses a negative temperature coefficient thermistor with a temperature measurement accuracy of ±0.5 degrees Celsius. The sampled data is digitized by a 24-bit analog-to-digital converter and transmitted to the coprocessor's internal memory via direct memory access.

4. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The original network structure of the pruned and optimized long short-term memory neural network model contains two hidden layers with 128 neurons in each layer. The number of network parameters is compressed to 30% of the original number through an amplitude-based weight pruning algorithm. The model inference latency is controlled within 5 milliseconds, and the accuracy loss is less than 2%.

5. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The input features of the long short-term memory neural network model include: voltage data, temperature data, voltage change rate, temperature change rate, and the difference between the average voltage of the module and the average voltage of the system within a sliding window with a length of 50 sampling points. The output is an anomaly probability score between 0 and 1. A score exceeding 0.85 is considered a preliminary anomaly.

6. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The communication rate of the Controller Area Network Flexible Data Rate Bus is 2 megabits per second. The cascaded confirmation request message adopts the standard data frame format, which includes an 11-bit identifier and a data field length of 8 bytes, used to carry the trigger module identifier, anomaly probability score and timestamp information.

7. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: During the collaborative inference calculation process, after receiving a request, adjacent modules prioritize using local data from the most recent 10 sampling periods to perform model inference. If local data is insufficient, instantaneous data at the current moment is used for calculation to ensure a response is completed within 2 milliseconds.

8. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The voting fusion strategy sets a second threshold of 50% of the total number of modules participating in collaborative confirmation, and the abnormal probability score of the triggering module itself must exceed the first threshold for three consecutive sampling periods before the collaborative confirmation result can be judged as the final abnormality.

9. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The local protection actions include: cutting off the module's charging and discharging circuit via a relay built into the module, and activating the module-level thermal management device for forced heat dissipation.

10. The method for detecting anomalies in a multi-series parallel battery module according to claim 1, characterized in that: The method also includes an online model update mechanism, in which the main control unit of the battery management system periodically summarizes the operating data and anomaly records of each module, retrains and optimizes the long short-term memory neural network model on the cloud or local server, and distributes the updated model parameters to the micro artificial intelligence coprocessors of each module through the flexible data rate bus of the controller area network.

Images