Lithium ion battery pack cross-domain fault diagnosis method based on multi-modal contrast learning

Abstract

The invention discloses a lithium ion battery pack cross-domain fault diagnosis method based on multi-modal contrast learning. The method comprises the following steps: synchronously acquiring a voltage time sequence signal, a temperature space matrix and an electrochemical impedance spectrum of a battery pack, and respectively extracting high-dimensional features of each mode by adopting a layered feature extraction network; on the basis of a comparative learning framework, intra-modal and cross-modal loss functions are constructed, and robustness of feature representation is enhanced; in combination with a dynamic domain self-adaptive module, the multi-modal alignment weight is adjusted according to the real-time working condition, and the cross-scene adaptability is improved; and meanwhile, an electro-thermal coupling model is utilized to generate a physically compliant virtual fault sample, and the generalization ability of the model is optimized. The system supports lightweight deployment of edge computing equipment, realizes low-delay real-time diagnosis, and can perform continuous model optimization through a cloud platform. The method is suitable for battery health management of electric vehicles, energy storage power stations and other scenes, and the sensitivity and reliability of fault detection are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of safety monitoring technology for power batteries of new energy vehicles, and specifically relates to a cross-domain fault diagnosis system and method for lithium-ion battery packs that integrates multimodal sensing data and comparative learning. Background Technology

[0002] Lithium-ion battery packs are widely used in electric vehicles, energy storage systems, and other fields, and their safety directly affects the reliability of equipment operation. Traditional battery fault diagnosis methods mainly rely on voltage monitoring, identifying abnormal states by analyzing voltage changes in individual battery cells. However, these methods have significant limitations:

[0003] Insufficient single-modal data: Relying solely on voltage signals makes it difficult to capture early fault characteristics. For example, an internal short circuit (ISC) in the battery may initially manifest as only a localized temperature rise or impedance change, while voltage anomalies may appear only after tens of minutes. Furthermore, faults such as sensor drift and loose connections are weakly represented by voltage signals, leading to high false alarm or false negative rates. Poor cross-condition adaptability: Existing deep learning-based diagnostic models are typically trained in laboratory environments, but in real-world applications, the generalization ability of these models drops significantly due to variations in battery usage scenarios (such as charge / discharge rate, ambient temperature, and SOC range). Traditional domain adaptation methods (such as adversarial training) are unstable on multimodal data and fail to effectively address the domain shift problem between different modes. Data scarcity contradicts physical laws: The scarcity of actual battery fault samples makes supervised learning models prone to overfitting. Although some studies use generative adversarial networks (GANs) or variational autoencoders (VAEs) to generate virtual fault data, these methods lack modeling of the electrochemical-thermodynamic coupling law of batteries. The generated samples may deviate from the real fault distribution, affecting the actual performance of the model. High computational complexity: Existing deep learning models that integrate multimodal data (such as 3D-CNN and Transformer) usually have high computational resource requirements and are difficult to run in real time on edge computing devices, which limits their application in vehicle battery management systems (BMS) or distributed energy storage monitoring systems.

[0004] In recent years, contrastive learning has shown advantages in the field of unsupervised feature learning, but its application in battery fault diagnosis still faces challenges:

[0005] Traditional contrastive learning methods are typically designed for single-modal data and struggle to effectively utilize the complementarity of multimodal data. Existing methods do not fully consider the dynamic evolution of battery faults, resulting in a mismatch between the contrastive learning objectives and real fault characteristics. Furthermore, the model transfer capability across multiple devices and scenarios still needs improvement.

[0006] Therefore, there is an urgent need for a new fault diagnosis method that can efficiently integrate multimodal sensor data and combine it with the physical laws of batteries to achieve high-precision, low-latency, and highly generalizable fault detection and early warning. Summary of the Invention

[0007] This invention aims to overcome the shortcomings of existing technologies and provide a battery fault diagnosis system based on multimodal contrastive learning. The core of this method includes the following four points. First, in a hierarchical feature fusion architecture, voltage transient features are extracted using 1D-ResNet, temperature spatial gradients are modeled using GNN, and impedance frequency domain characteristics are analyzed using complex networks to construct a spatiotemporally correlated joint embedding representation. Second, in the cross-modal contrastive learning mechanism, a dual contrastive loss within and between modes is designed, utilizing physical fault samples generated by digital twins to enhance the model's generalization ability. Third, in the dynamic domain adaptive module, the alignment weights of each mode are dynamically adjusted based on real-time operating conditions (vehicle speed, SOC). Finally, regarding the edge-cloud collaborative deployment, a lightweight model is used for real-time state diagnosis and continuous optimization of the global model.

[0008] To solve the above technical problems, the present invention is implemented using the following technical solution:

[0009] A cross-domain fault diagnosis method for lithium-ion battery packs based on multimodal contrastive learning includes the following steps:

[0010] S1: Multimodal data synchronous acquisition: Real-time acquisition of battery pack voltage timing signals via BMS system. The spatial temperature matrix output by the temperature sensor array and impedance spectra measured by the electrochemical workstation ,in This refers to the number of individual battery cells. Let V be the voltage value (in V) of the i-th individual cell at time t. The coordinates of the battery pack in a two-dimensional arrangement. The scanning frequency (unit: Hz);

[0011] S2: Layered feature extraction of voltage, temperature, and impedance:

[0012] Voltage mode features: Millisecond-level transient features were extracted using 1D-ResNet with dilated convolution, as shown in (1) below:

[0013]

[0014] in, As a voltage characteristic dimension, It is a one-dimensional residual convolutional network;

[0015] Temperature modal characteristics: The temperature spatial gradient is processed by a graph neural network (GNN), as shown in (2) below:

[0016]

[0017] in, Temperature gradient tensor , Temperature is a characteristic dimension;

[0018] Impedance mode characteristics: The Nyquist spectrum was processed using a complex weighted network, as shown in (3) below:

[0019]

[0020] in, Let be a trainable real-valued weight matrix. For impedance characteristics;

[0021] S3: Cross-modal contrastive learning: The inter-modal contrastive loss function constructed by this model is shown in the following (4):

[0022]

[0023] in, For cosine similarity, For temperature coefficient, This represents the batch sample size. This is an indicator function (1 if i ≠ j, 0 otherwise);

[0024] S4: Dynamic Domain Adaptive Alignment: Working Condition Recognition Network The output dynamically adjusts the alignment weights as shown in (5):

[0025]

[0026] in, For the Sigmoid function, The battery is in its state of charge. For vehicle speed, Identify sub-networks based on operating conditions;

[0027] S5: Digital twin data augmentation: Virtual fault samples are generated through an electro-thermal coupling model as shown in (6):

[0028]

[0029] In the formula, Thermal diffusivity (unit: ), Time-varying internal resistance ( ), For operating current, Material density (unit: ), Specific heat capacity (unit: );

[0030] S6: Fault diagnosis decision: The probability distribution of the output fault type is shown in the following (7):

[0031]

[0032] in, The classifier weight matrix is... Label the fault type (including: Normal / ISC / Sensor Fault, etc.);

[0033] The fault propagation risk index is shown below (8):

[0034]

[0035] in, For fault type weighting coefficients, This indicates the battery's health status.

[0036] In a preferred embodiment, the intra-modal contrastive learning in step S3 employs an improved MoCo v3 framework, including:

[0037] S31: Query Encoder key encoder The parameter updates are as follows (9):

[0038]

[0039] Among them, momentum coefficient =0.999, For query encoder The parameters, For query encoder Parameters;

[0040] S32: Length of the dynamic dictionary queue =8192, temperature coefficient =0.07.

[0041] In the preferred technical solution, the working condition identification network in step S4 include:

[0042] S41: LSTM layer processes vehicle speed timing ;

[0043] S42: The importance weights of the SOC interval are calculated using the attention mechanism as shown in (10):

[0044]

[0045] in, For trainable weight matrix, This is the battery state-of-charge time vector.

[0046] The fault types include: internal short circuit (ISC); sensor drift. Inconsistency among individual units Precursors to thermal runaway .

[0047] An edge computing device for implementing the above method is characterized by comprising: a multi-source data synchronous acquisition interface (sampling rate ≥ 1 kHz); a lightweight feature extraction module (parameter size < 100 KB); and a real-time diagnostic result output interface (latency < 10 ms).

[0048] A battery health management cloud platform based on this method, deployed with the above method, achieves the following:

[0049] The Jacobian matrix constraints for cross-model transfer are shown in (11):

[0050]

[0051] in, For the model Jacobian matrix, It is the Frobenius norm;

[0052] The SHAP value for root cause analysis is calculated as shown in (12-13):

[0053] (12)

[0054] (13)

[0055] in, The SHAP contribution value for the i-th feature. For the complete feature set, This is the feature subset without considering feature i. Output function for the diagnostic model, To represent the voltage, temperature, and impedance mode characteristics in the input feature subsets respectively, For modal encoders (such as ResNet1D, GNN, etc.) defined in S2. This is the weight matrix for the classification layer.

[0056] Compared with the prior art, the advantages of the present invention are:

[0057] 1. By simultaneously analyzing voltage, temperature field, and impedance spectrum data, the problem of insufficient sensitivity of traditional single voltage signals to early faults (such as micro short circuits) is overcome, and the fault detection rate is significantly improved.

[0058] 2. The "intra-modal-inter-modal-cross-domain" three-layer contrastive learning framework is better able to capture the correlation of fault features under complex working conditions compared with existing single contrastive methods such as MoCo.

[0059] 3. The alignment weight is automatically adjusted based on real-time operating conditions (vehicle speed / SOC / temperature), which solves the problem of performance drop when the traditional fixed weight method changes vehicle model.

[0060] 4. Virtual fault samples with physical constraints are generated by an electro-thermal coupling model, which alleviates the overfitting problem caused by the scarcity of actual fault samples.

[0061] 5. The lightweight edge computing module ensures the real-time performance of the model while also meeting the needs of low latency and big data analysis. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating a cross-domain fault diagnosis method for lithium-ion battery packs based on multimodal contrastive learning, provided as an embodiment of the present invention. Detailed Implementation

[0063] 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 and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0064] The invention will now be further described with reference to the accompanying drawings.

[0065] See Figure 1 This invention provides a cross-domain fault diagnosis method for lithium-ion battery packs based on multimodal contrastive learning, specifically including the following steps:

[0066] S1: Multimodal data synchronous acquisition: Real-time acquisition of battery pack voltage timing signals via BMS system. The spatial temperature matrix output by the temperature sensor array and impedance spectra measured by the electrochemical workstation ,in This refers to the number of individual battery cells. Let V be the voltage value (in V) of the i-th individual cell at time t. The coordinates of the battery pack in a two-dimensional arrangement. The scanning frequency (unit: Hz);

[0067] S2: Layered feature extraction of voltage, temperature, and impedance:

[0068] Voltage mode features: Millisecond-level transient features were extracted using 1D-ResNet with dilated convolution, as shown in (1) below:

[0069]

[0070] in, As a voltage characteristic dimension, It is a one-dimensional residual convolutional network;

[0071] Temperature modal characteristics: The temperature spatial gradient is processed by a graph neural network (GNN), as shown in (2) below:

[0072]

[0073] in, Temperature gradient tensor , Temperature is a characteristic dimension;

[0074] Impedance mode characteristics: The Nyquist spectrum was processed using a complex weighted network, as shown in (3) below:

[0075]

[0076] in, Let be a trainable real-valued weight matrix. For impedance characteristics;

[0077] S3: Cross-modal contrastive learning: The inter-modal contrastive loss function constructed by this model is shown in the following (4):

[0078]

[0079] in, For cosine similarity, For temperature coefficient, This represents the batch sample size. This is an indicator function (1 if i ≠ j, 0 otherwise);

[0080] S4: Dynamic Domain Adaptive Alignment: Working Condition Recognition Network The output dynamically adjusts the alignment weights as shown in (5):

[0081]

[0082] in, For the Sigmoid function, The battery is in its state of charge. For vehicle speed, Identify sub-networks based on operating conditions;

[0083] S5: Digital twin data augmentation: Virtual fault samples are generated through an electro-thermal coupling model as shown in (6):

[0084]

[0085] In the formula, Thermal diffusivity (unit: ), Time-varying internal resistance ( ), For operating current, Material density (unit: ), Specific heat capacity (unit: );

[0086] S6: Fault diagnosis decision: The probability distribution of the output fault type is shown in the following (7):

[0087]

[0088] in, The classifier weight matrix is... Label the fault type (including: Normal / ISC / Sensor Fault, etc.);

[0089] The fault propagation risk index is shown below (8):

[0090]

[0091] in, For fault type weighting coefficients, This indicates the battery's health status.

[0092] The intra-modal contrastive learning in step S3 employs an improved MoCo v3 framework, including:

[0093] S31: Query Encoder key encoder The parameter updates are as follows (9):

[0094]

[0095] Among them, momentum coefficient =0.999, For query encoder The parameters, For query encoder Parameters;

[0096] S32: Length of the dynamic dictionary queue =8192, temperature coefficient =0.07.

[0097] The working condition identification network in step S4 include:

[0098] S41: LSTM layer processes vehicle speed timing ;

[0099] S42: The importance weights of the SOC interval are calculated using the attention mechanism as shown in (10):

[0100]

[0101] in, For trainable weight matrix, This is the battery state-of-charge time vector.

[0102] The fault types include: internal short circuit (ISC); sensor drift. Inconsistency among individual units Precursors to thermal runaway .

[0103] An edge computing device for implementing the above method is characterized by comprising: a multi-source data synchronous acquisition interface (sampling rate ≥ 1 kHz); a lightweight feature extraction module (parameter size < 100 KB); and a real-time diagnostic result output interface (latency < 10 ms).

[0104] A battery health management cloud platform based on this method, deployed with the above method, achieves the following:

[0105] The Jacobian matrix constraints for cross-model transfer are shown in (11):

[0106]

[0107] in, For the model Jacobian matrix, It is the Frobenius norm;

[0108] The SHAP value for root cause analysis is calculated as shown in (12-13):

[0109] (12)

[0110] (13)

[0111] in, The SHAP contribution value for the i-th feature. For the complete feature set, This is the feature subset without considering feature i. Output function for the diagnostic model, To represent the voltage, temperature, and impedance mode characteristics in the input feature subsets respectively, For modal encoders (such as ResNet1D, GNN, etc.) defined in S2. This is the weight matrix for the classification layer.

[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cross-domain fault diagnosis method for lithium-ion battery packs based on multimodal contrastive learning, characterized in that, Includes the following steps: S1: Multimodal data synchronous acquisition: Real-time acquisition of battery pack voltage timing signals via BMS system. The spatial temperature matrix output by the temperature sensor array and impedance spectra measured by the electrochemical workstation ,in This refers to the number of individual battery cells. Let V be the voltage value of the i-th individual cell at time t, in V. The coordinates of the battery pack in a two-dimensional arrangement. Scan frequency, unit: Hz; S2: Layered feature extraction of voltage, temperature, and impedance: Voltage mode features: Millisecond-level transient features were extracted using 1D-ResNet with dilated convolution, as shown in (1) below: （1） in, As a voltage characteristic dimension, It is a one-dimensional residual convolutional network; Temperature modal characteristics: The temperature spatial gradient is processed by a graph neural network (GNN), as shown in (2) below: （2） in, Temperature gradient tensor , Temperature is a characteristic dimension; Impedance mode characteristics: The Nyquist spectrum was processed using a complex weighted network, as shown in (3) below: （3） in, Let be a trainable real-valued weight matrix. For impedance characteristics; S3: Cross-modal contrastive learning: The inter-modal contrastive loss function constructed by this model is shown in the following (4): （4） in, For cosine similarity, For temperature coefficient, This represents the batch sample size. This is an indicator function; it is 1 when i ≠ j, and 0 otherwise. S4: Dynamic Domain Adaptive Alignment: Working Condition Recognition Network The output dynamically adjusts the alignment weights as shown in (5): （5） in, For the Sigmoid function, The battery is in its state of charge. For vehicle speed, Identify sub-networks based on operating conditions; S5: Digital twin data augmentation: Virtual fault samples are generated through an electro-thermal coupling model as shown in (6): （6） In the formula, Thermal diffusivity, unit: , For time-varying internal resistance, , For operating current, Material density, unit: , Specific heat capacity, unit: ; S6: Fault diagnosis decision: The probability distribution of the output fault type is shown in the following (7): （7） in, The classifier weight matrix is... The fault type is labeled as follows: Normal / ISC / Sensor Fault; The fault propagation risk index is shown below (8): （8） in, For fault type weighting coefficients, This indicates the battery's health status.

2. The method for cross-domain fault diagnosis of lithium-ion battery packs based on multimodal contrastive learning according to claim 1, characterized in that, The intra-modal contrastive learning in step S3 employs an improved MoCo v3 framework: S31: Query Encoder key encoder The parameter updates are as follows (9): （9） Among them, momentum coefficient =0.999, For query encoder The parameters, For query encoder Parameters; S32: Length of the dynamic dictionary queue =8192, temperature coefficient =0.

07.

3. The method for cross-domain fault diagnosis of lithium-ion battery packs based on multimodal contrastive learning according to claim 1, characterized in that, Working condition identification network in step S4 Include: S41: LSTM layer processes vehicle speed timing ; S42: The importance weights of the SOC interval are calculated using the attention mechanism as shown in (10): （10） in, For trainable weight matrix, This is the battery state-of-charge time vector.

4. The method for cross-domain fault diagnosis of lithium-ion battery packs based on multimodal contrastive learning according to claim 1, characterized in that, The fault types include: internal short circuit (ISC); sensor drift; individual cell inconsistency; and precursors to thermal runaway.

5. An edge computing device, characterized in that, Implementing the method according to any one of claims 1-4, comprising: a multi-source data synchronous acquisition interface with a sampling rate ≥1kHz; a lightweight feature extraction module with a parameter size <100KB; and a real-time diagnostic result output interface with a latency <10ms.

6. A battery health management cloud platform, characterized in that, Deploying the method according to any one of claims 1-4, to achieve: The Jacobian matrix constraints for cross-model transfer are shown in (11): （11） in, For the model Jacobian matrix, It is the Frobenius norm; The SHAP value for root cause analysis is calculated as shown in (12-13): （12） （13） in, The SHAP contribution value for the i-th feature. For the complete feature set, This is the feature subset without considering feature i. Output function for the diagnostic model, To represent the voltage, temperature, and impedance mode characteristics in the input feature subsets respectively, The modal encoder as defined in claim 1S2 This is the weight matrix for the classification layer.

Images