A method and system for early warning of thermal runaway of a battery pack

Abstract

The present application belongs to the technical field of battery safety monitoring, and specifically relates to a thermal runaway space-time dynamic early warning method and system for a battery pack, which comprises the following steps: collecting space-time data of the battery pack; enhancing small sample data by using a strategy of combining transfer learning and a generative adversarial network; extracting spatial correlation features and time dynamic features by using 3D-CNN and LSTM respectively, and performing weighted fusion through an attention mechanism to obtain a space-time coupling feature vector; dynamically adjusting a warning threshold based on a sliding window and an adaptive coefficient, triggering a warning and outputting a warning level according to the coupling feature vector; and predicting the diffusion source, diffusion direction, diffusion rate, coverage range and risk level of thermal runaway based on spatial diffusion topology modeling and time dynamic evolution analysis. The present application solves the problems of scarce thermal runaway experimental data, split space-time features, lagging early warning and poor dynamic adaptability, and realizes early and accurate early warning and diffusion trend prediction in a small sample scenario.

Description

Technical Field

[0001] This invention relates to the field of battery safety monitoring technology, and in particular to a spatiotemporal dynamic early warning method and system for battery pack thermal runaway. Background Technology

[0002] Lithium-ion batteries are widely used in new energy vehicles, energy storage power stations, and other fields due to their high energy density and long cycle life. However, batteries are prone to thermal runaway under abuse conditions such as overcharging, short circuits, and overheating, which can lead to high temperatures, smoke, fires, and even explosions, seriously threatening life and property safety. Therefore, achieving early and accurate warning of thermal runaway is crucial for ensuring the safety of battery systems.

[0003] Existing thermal runaway early warning methods are mainly divided into two categories: one is the threshold method based on a single physical quantity (such as temperature or voltage), which is simple but prone to false alarms and missed alarms, and the early warning is delayed; the other is the intelligent early warning method based on machine learning, which has better performance, but still has the following bottlenecks: 1) Strong data dependence: thermal runaway experiments are costly and dangerous, and it is difficult to obtain a large number of labeled samples, resulting in poor generalization ability of the model in small sample scenarios; 2) Spatiotemporal characteristics are fragmented: existing methods mostly process time series or spatial data separately, ignoring the dynamic diffusion nature of thermal runaway coupled with time and space; 3) Weak dynamic adaptability: fixed models are difficult to adapt to feature drift caused by battery aging and environmental changes; 4) Lack of diffusion prediction: it can only issue an alarm, but cannot predict the diffusion trend of thermal runaway, which is not conducive to the formulation of accurate emergency response strategies.

[0004] Thermal runaway is essentially a spatiotemporally coupled process of "local heat accumulation → heat conduction in adjacent regions → global thermal runaway." Its spread is governed by two main factors: spatially, it depends on the physical layout and heat conduction paths of the individual battery cells; temporally, it depends on the thermal evolution rate of the initial heat source and the response delays of adjacent cells. Therefore, developing a method that can utilize limited samples, integrate spatiotemporal features, achieve early warning, and predict spread trends is of great significance. Summary of the Invention

[0005] This invention proposes a spatiotemporal dynamic early warning method and system for battery pack thermal runaway, aiming to solve the technical problems existing in the prior art, such as how to overcome the fragmentation of spatiotemporal coupling characteristics and the lag caused by fixed early warning thresholds in small sample scenarios where real labeled samples of thermal runaway are extremely scarce (only 10-50 sets).

[0006] In a first aspect, embodiments of the present invention provide a spatiotemporal dynamic early warning method for thermal runaway of a battery pack, comprising:

[0007] S1, acquire the time series data of each cell in the battery pack to form a spatiotemporal raw dataset;

[0008] S2, after preprocessing the original spatiotemporal dataset, an enhancement strategy combining transfer learning and generative adversarial networks is adopted to generate virtual thermal runaway samples to expand the data and obtain an expanded dataset.

[0009] S3, extract the spatial correlation features and temporal dynamic features of the data in the expanded dataset, and perform weighted fusion of the spatial correlation features and temporal dynamic features through an attention mechanism to obtain a spatiotemporal coupling feature vector;

[0010] S4. Calculate the statistical distribution of the spatiotemporal coupling feature vector in real time based on a sliding window, adaptively adjust the warning threshold, and output the thermal runaway warning level according to the relationship between the spatiotemporal coupling feature vector and the warning threshold.

[0011] S5. Based on the spatiotemporal coupling feature vector, perform spatial diffusion topology modeling and temporal dynamic evolution prediction, and output prediction results including diffusion source location, diffusion direction, diffusion speed, coverage area and diffusion risk level.

[0012] The technical effect of the spatiotemporal dynamic early warning method for thermal runaway of battery pack disclosed in this invention is that the method can fully explore and integrate the spatiotemporal coupling characteristics of battery state under the condition of scarce thermal runaway samples (small sample size), realize early and accurate dynamic early warning, and further predict the spread trend of thermal runaway, thus gaining valuable time for safety intervention.

[0013] Furthermore, S1 specifically includes: collecting time-series data of each battery cell by arranging a multi-type sensor array in the battery pack; the multi-type sensor array includes a temperature sensor arranged on the surface of the battery tabs, a voltage sensor connected to the positive and negative electrodes of the battery, and a gas sensor arranged inside the battery pack; the collected data includes the surface temperature, terminal voltage of each cell, and the concentration of H2, CO, and C2H4 gases in the battery pack, with a collection frequency of 1-10Hz.

[0014] Furthermore, S2 specifically includes:

[0015] S21, perform outlier removal and normalization on the original spatiotemporal dataset to obtain a standardized sample set;

[0016] S22 uses normal operating data of the same type of battery as the source domain and small samples of thermal runaway as the target domain, and aligns the feature distribution of the source domain and the target domain through a domain adaptive network.

[0017] S23. Using the aligned source domain data and small samples of the target domain as input, virtual spatiotemporal data is generated through a generative adversarial network. The similarity between the virtual samples and the real samples is verified by using dynamic time warping distance. Virtual samples with similarity higher than a preset threshold are selected and added to the dataset to obtain an expanded dataset.

[0018] Furthermore, S3 specifically includes:

[0019] 3D-CNN is used to extract spatial correlation features of the data in the dataset, and LSTM is used to extract temporal dynamic features.

[0020] The 3D-CNN has 3 input channels, corresponding to temperature, voltage, and gas concentration data respectively; the LSTM has an input sequence length of 10-30 time steps.

[0021] Furthermore, the weighted fusion formula in S3 is as follows:

[0022] ;

[0023] Among them, F space F is the spatial correlation feature vector output by the 3D-CNN. time α is the temporal dynamic feature vector output by the LSTM, and α is the attention weight adaptively calculated through the softmax function.

[0024] Furthermore, the specific process of adaptively adjusting the warning threshold in S4 is as follows:

[0025] Set a sliding window and calculate the mean μ and standard deviation σ of the spatiotemporal coupling feature vectors within the window;

[0026] The warning threshold T is calculated using the formula T=μ+k·σ, where k is an adaptive coefficient adjusted based on the number of battery cycles.

[0027] When the spatiotemporal coupling feature vectors of multiple consecutive time steps exceed the warning threshold T, a warning signal is triggered.

[0028] Furthermore, S5 specifically includes:

[0029] Calculate the spatial attention weight α for each battery i :

[0030] ;

[0031] Where Max(F) space [i]) is the maximum value of the spatial correlation feature of the i-th monomer, which is used to reflect the degree of monomer anomaly;

[0032] Each individual node is defined as a topology node. Adjacency relationships between topology nodes are defined based on their physical location, and diffusion weight A is initialized. i,j Construct the spatial topology matrix A;

[0033] Based on the spatial association features extracted by 3D-CNN, the spatial association strength S between any two topological nodes is calculated. i,j :

[0034] ;

[0035] Among them, F space [i] represents the spatial association feature of the i-th unit;

[0036] Combined with diffusion weight A i,j Spatial correlation strength S i,j And attention weight α, which reflects the degree of individual abnormality. i Construct a diffusion probability matrix P, where the diffusion probability P0 is... i,j =A i,j ·S i,j ·α i , used to predict the direction of thermal runaway propagation, T is the transpose symbol.

[0037] Furthermore, the time-dynamic evolution prediction in S5 includes:

[0038] The diffusion rate coefficient is calculated based on the thermal evolution-related feature dimensions extracted from the temporal dynamic features using LSTM. , used to represent the number of monomers affected by thermal runaway propagation per unit time;

[0039] ;

[0040] Among them, F time [m:n] represents the reference window for the change status of the m-th to n-th interval in the feature dimension, F time [0:m] represents the normal state observation window for the interval from 0 to m in the feature dimension. Here, is the adaptive coefficient, and mean is the average value.

[0041] Based on the diffusion rate coefficient Predict the coverage area N(t) of thermal runaway after t seconds;

[0042] ;

[0043] Where, N total This represents the total number of individual cells in the battery pack. This is for rounding to the nearest integer.

[0044] Furthermore, S5 specifically includes: based on the diffusion rate coefficient Based on the predicted coverage area N(t), the risk level of thermal runaway propagation is classified.

[0045] In a second aspect, embodiments of the present invention provide a spatiotemporal dynamic early warning system for battery pack thermal runaway, used to execute the method described above, the system comprising:

[0046] The data acquisition module is configured to acquire time-series data of each cell in the battery pack to form a spatiotemporal raw dataset;

[0047] The data augmentation module is configured to preprocess the original spatiotemporal dataset and then use an enhancement strategy that combines transfer learning and generative adversarial networks to generate virtual thermal runaway samples to augment the data and obtain an augmented dataset.

[0048] The feature coupling module is configured to extract the spatial correlation features and temporal dynamic features of the data in the expanded dataset, and to perform weighted fusion of the spatial correlation features and temporal dynamic features through an attention mechanism to obtain a spatiotemporal coupling feature vector;

[0049] The early warning adjustment module is configured to calculate the statistical distribution of the spatiotemporal coupling feature vector in real time based on a sliding window, adaptively adjust the early warning threshold, and output the thermal runaway early warning level according to the relationship between the spatiotemporal coupling feature vector and the early warning threshold.

[0050] The risk prediction module is configured to perform spatial diffusion topology modeling and temporal dynamic evolution prediction based on the spatiotemporal coupling feature vector, and output prediction results including diffusion source location, diffusion direction, diffusion speed, coverage area and diffusion risk level.

[0051] The technical effect of the spatiotemporal dynamic early warning system for thermal runaway of a battery pack disclosed in this invention is that, through the fusion enhancement of "transfer learning + GAN", a high-precision early warning model can be effectively trained under the condition of only 10-50 sets of real thermal runaway samples, which solves the data scarcity bottleneck caused by high risk and high cost of experiments in this field. Attached Figure Description

[0052] Figure 1 A schematic flowchart of a spatiotemporal dynamic early warning method for thermal runaway of a battery pack provided in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the multi-sensor layout of a battery pack provided in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram illustrating the principle of the data preprocessing and small sample data augmentation structure provided in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of spatiotemporal coupling feature extraction provided in an embodiment of the present invention;

[0056] Figure 5 This is a schematic diagram of the thermal runaway propagation trend prediction architecture provided in an embodiment of the present invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0058] The core technical problem to be solved by this invention is how to overcome the problems of traditional methods' fragmentation of spatiotemporal coupling characteristics and the lag caused by fixed warning thresholds in small sample scenarios where real-world thermal runaway labeled samples are extremely scarce (only 10-50 sets). The goal is to achieve early and accurate warnings by fusing spatiotemporal dynamic features and simultaneously predicting the spread trend of thermal runaway. Specifically, this is a spatiotemporal dynamic early warning method for battery pack thermal runaway, referencing... Figures 1 to 5 As shown, the specific steps include:

[0059] S1, multi-dimensional spatiotemporal data acquisition, involves arranging multiple types of sensor arrays in the battery pack to collect time-series data from each individual battery cell, forming a raw spatiotemporal dataset.

[0060] Experimental subject: 16-cell series-parallel ternary lithium battery pack (single cell capacity 200Ah, nominal voltage 3.7V, total voltage 59.2V), cells arranged in a 4×4 matrix. Figure 2 ).

[0061] Sensor array: K-type thermocouple (accuracy ±0.5℃), voltage acquisition module (accuracy ±0.01V), CO sensor (response time ≤1s, measurement range 0-1000ppm, accuracy ±5%).

[0062] Thermal runaway triggering and data acquisition: 20-50 sets of real thermal runaway samples (including time-series data of temperature, voltage, and gas concentration) were obtained through overcharge and overheating experiments. 500 sets of normal operating condition data of the same type of battery were also collected as source domain data for transfer learning.

[0063] Through the dense sensor layout and synchronous acquisition described above, a high-dimensional spatiotemporal raw dataset was constructed, laying the data foundation for subsequent capture of the subtle spatiotemporal evolution of thermal runaway.

[0064] S2, Small sample data augmentation: After preprocessing the original spatiotemporal dataset, an augmentation strategy that combines transfer learning and generative adversarial networks is adopted to generate virtual thermal runaway samples to expand the dataset.

[0065] Data preprocessing: The 3σ criterion is used to remove outliers (such as sudden pulse signals) from the temperature sensor, and noise is removed by db4 wavelet transform and normalized to the [0,1] interval;

[0066] Transfer learning domain alignment: Construct a DAN network, which includes a feature extractor (3-layer CNN) and a domain discriminator (2-layer fully connected). The training objective is to minimize the distribution difference loss between the source domain and the target domain (using MMD loss). After 500 training iterations, the feature distribution similarity between the source domain and the target domain reaches 0.85.

[0067] GAN Virtual Sample Generation: A GAN network was constructed with a 4-layer fully connected generator (hidden layer dimensions 256, 128, 64, and 32), incorporating an attention mechanism to focus on the critical stage of thermal runaway (after temperature ≥ 80℃). The discriminator consisted of a 3-layer fully connected layer. Using aligned source domain data and 20 real thermal runaway samples as input, 80 virtual samples were generated. DTW distance validation showed that 75 virtual samples had a similarity ≥ 0.9 with real samples. The final expanded small sample dataset size was 95 sets (20 real sets + 75 virtual sets). Ultimately, the dataset used for model training contained 95 samples (20 real sets + 75 virtual sets), effectively solving the small sample problem.

[0068] The introduction of an attention mechanism into the GAN's generator aims to enable the generator to more accurately simulate the morphology of key stages (such as rapid temperature rise and gas production bursts) when constructing virtual thermal runaway sequences. One specific implementation method is to use "conditional gated attention":

[0069] (1) Input noise vector z: random noise, used as the starting point for generation; condition vector c: contains time step information and stage label. The stage label can be provided by a parallel, simple temperature judgment module.

[0070] (2) An attention layer is inserted at an intermediate layer of the generator (e.g., before reshaping the features into a spatiotemporal sequence). This layer takes the intermediate layer features and the “stage label” in the conditional vector c as input. Attention weights are computed through a small neural network (such as an MLP). This network uses the “stage label” as the key and the intermediate layer features as the query. For time steps labeled as “key stages” (T≥80°C), this mechanism computes higher attention weights.

[0071] (3) The calculated Attention_Weight will be used to weight the intermediate layer features (i.e., the product of features and weights) or to control the information flow (e.g., as an input gating signal in the LSTM unit).

[0072] Through this structured conditional attention mechanism, the virtual samples generated by GANs no longer learn the entire sequence on an average basis, but instead specifically enhance the simulation accuracy of dangerous evolution stages, significantly improving the quality and training effect of data after small sample augmentation.

[0073] S3, Spatiotemporal feature extraction and fusion: 3D-CNN is used to extract spatial correlation features of the data in the dataset, LSTM is used to extract temporal dynamic features, and the spatial correlation features and temporal dynamic features are weighted and fused through an attention mechanism to obtain a spatiotemporal coupled feature vector.

[0074] Temporal Feature Extraction: LSTM is used to extract features from the time-series data of each sensor, and temporal modeling is performed on the spatial feature sequence output by the 3D-CNN to capture the slow evolution trend in the early stages of thermal runaway (such as changes in the rate of temperature rise). LSTM Structure: The input sequence length is 10–30 time steps, with 2 hidden layers, 64 hidden units, and a dropout rate of 0.2. The temporal dynamic features include the rate of temperature change, the rate of voltage change, gas concentration changes, and mutation features (number of mutation points, mutation amplitude), forming a temporal feature vector.

[0075] Spatial Domain Feature Extraction: A sensor topology graph is constructed, with each sensor as a graph node. A 3D-CNN is used to extract features from the topology graph, capturing the correlation between sensor data at different spatial locations (such as temperature gradients, temperature differences, voltage differences, and conduction delays between battery temperatures on the module surface), forming spatial domain feature vectors. The 3D-CNN structure is as follows: input channels = 3 (temperature / voltage / gas concentration), convolutional kernel size 3×3×3, containing 3 convolutional layers (with 16, 32, and 64 kernels respectively), 2 pooling layers (max pooling with a stride of 2), and the activation function ReLU.

[0076] Spatiotemporal feature fusion: An attention-weighted fusion mechanism is employed. During training, the weights of temporal and spatial features are adaptively learned (focusing on features sensitive to thermal runaway, such as abrupt changes in battery temperature and voltage, and the spatial correlation between CO concentration and temperature). The fusion yields a spatiotemporally coupled feature vector comprehensively representing the thermal runaway state. The weighted fusion formula is as follows:

[0077] ;

[0078] Among them, F space F is the spatial correlation feature vector output by the 3D-CNN. time α is the temporal dynamic feature vector output by the LSTM, and α is the attention weight adaptively calculated through the softmax function.

[0079] S4, Dynamic Early Warning: Based on the sliding window, the statistical distribution of the spatiotemporal coupling feature vector is calculated in real time, the early warning threshold is adaptively adjusted, and the thermal runaway early warning level is output according to the relationship between the spatiotemporal coupling feature vector and the early warning threshold.

[0080] Dynamic warning threshold adjustment: a) Set the sliding window size to 30-60s, with a step size of 1s, and update the fused feature vector within the window in real time. Calculate the mean μ and standard deviation σ of the coupled feature vector within the window in real time. b) Calculate the warning threshold T = μ + k·σ, where k is an adaptive coefficient (adjusted according to the number of battery cycles, ranging from 1.5 to 2.5; k = 2.5 when the number of battery cycles ≤ 500, and k = 1.5 when the number of cycles ≥ 1000, to adapt to battery aging). c) When the coupled feature vector ≥ T for three consecutive time steps, trigger the warning signal.

[0081] Warning levels are categorized based on the probability of thermal runaway: ([0,0.3) no risk, [0.3,0.6) low risk, [0.6,0.8) medium risk, ≥0.8 high risk).

[0082] This dynamic mechanism effectively reduces false alarms caused by environmental noise or normal fluctuations and enables the quantitative classification of risks.

[0083] S5, thermal runaway diffusion trend prediction: Based on the spatiotemporal coupling feature vector, perform spatial diffusion topology modeling and temporal dynamic evolution prediction, and output prediction results including diffusion source location, diffusion direction, diffusion velocity, coverage area and diffusion risk level.

[0084] The coupling features extracted from the spatiotemporal fusion model are directly mapped to diffusion trend quantification indicators to predict the direction, rate, timing or coverage, and level of thermal runaway diffusion. This includes the following functional modules:

[0085] (1) Spatial diffusion topology modeling of battery pack. First, define the topology nodes and adjacency relationships, initialize the topology weights, and then construct the spatial topology matrix. (N is the number of battery cells) A mathematical model of "cell location-adjacency relationship" is established to provide a basis for diffusion path analysis.

[0086] ;

[0087] In the formula, Ai,j represents the initial diffusion weight from node i to node j.

[0088] Topology node definition: Each battery cell is a node, and the numbering corresponds one-to-one with the sensor layout (e.g., in a 4×4 matrix battery pack, nodes 1 to 16 correspond to...). Figure 2 Monomers 1-16).

[0089] Adjacency relationship definition: Direct adjacency (first-order neighborhood) refers to nodes that are in contact with each other vertically, horizontally, or front and back (e.g., the first-order neighbors of node 6 are nodes 2, 5, 7, and 10); Indirect adjacency (second-order neighborhood) refers to nodes that are separated by one node (e.g., the second-order neighbors of node 6 are nodes 1, 3, 9, 11, and 14).

[0090] Topological weight initialization: Initial weights are set based on the physical structure (direct adjacent weight = 1, indirect adjacent weight = 0.3, non-adjacent weight = 0), and subsequently dynamically corrected by the spatial features of the 3D-CNN. The specific process is as follows:

[0091] (a) Feature mapping: The high-level features output by the 3D-CNN for each pair of adjacent units (i,j) are mapped to a scalar C through a lightweight fully connected layer (or by directly calculating the correlation). ij This is called "data-driven correlation," with a value range of [0,1]. It can be achieved by calculating the cosine similarity or dot product of the feature vectors and then normalizing it using the Sigmoid function.

[0092] Formula example: C ij =σ( <CNN Feature(i) CNN Feature(j) >), where σ is the Sigmoid function, and <> is the dot product, CNN Feature(i) and CNN Feature(j) These are the i-th and j-th high-level features output by each adjacent unit (i,j) in each team.

[0093] (b) Dynamic fusion: The initial static weights A are fused together. initij Relevance to data-driven approach C ij Perform weighted fusion to generate dynamic topology weights A dynamicij .

[0094] Formula example: A dynamicij =β·A initij +(1-β)·C ij ;

[0095] Parameter explanation:

[0096] β is a learnable or preset fusion coefficient (e.g., β=0.5) used to balance the importance of prior knowledge and real-time data.

[0097] When C ij When the value is very high, even if i and j are only indirectly adjacent in physical terms (initial weight 0.3), A dynamicij It may also be close to 1, indicating that the data strongly suggests a strong correlation between the two (possibly due to uneven heat dissipation, internal short circuits, etc.).

[0098] (c) Triggering and Updating: This correction is performed online and continuously. C is calculated every time a new time window (e.g., 1 second) of data is processed by the 3D-CNN. ij and update A dynamicij The updated A dynamicij The matrix will be immediately used to calculate the diffusion probability matrix P at the current time.

[0099] This algorithm evolves the topology model from a static prior graph into a dynamic graph that reflects the real-time thermal-electric coupling state of the battery pack, thus making the diffusion path prediction more adaptable to specific fault scenarios and individual differences in battery packs.

[0100] (2) Location of diffusion source (identification of initial thermal runaway cell, diffusion source location, or fault source cell). Based on the spatial attention weight α of each cell. i (That is, α is decomposed into each individual cell, where i represents the i-th cell); Define α i >θ α (θ) α A cell with a value of 0.3 (determined by statistical analysis of the enhanced sample set) is considered the initial thermal runaway cell (diffusion source). If multiple cells satisfy the condition, α is taken as... i The two largest cells act as a joint diffusion source.

[0101] ;

[0102] Max(F space [i]) represents the maximum value of the spatial correlation characteristic of the i-th cell, used to reflect the degree of cell anomaly. For example, in a 4×4 battery pack, if the α6 of the central cell 6 is 0.42 (maximum), then the diffusion source is cell 6.

[0103] (3) Construction of the diffusion probability matrix for spatial dimension prediction, i.e., prediction of the direction of thermal runaway diffusion. Based on the spatial correlation features extracted by 3D-CNN (output features of the last convolutional kernel, reflecting the thermal conduction correlation between different monomers), the spatial correlation strength between monomers i and j is calculated:

[0104] ;

[0105] Among them, F space [i] represents the spatial association feature of the i-th unit. The cosine similarity is calculated by vector dot product, where Si,j∈[0,1], and the larger the value, the greater the probability of diffusion from i to j.

[0106] Diffusion probability P i,j =A i,j ·S i,j ·α i This is used to predict the direction of thermal runaway propagation. Normalization is applied to ensure that the sum of elements in each row of matrix P is 1, thus guaranteeing P's propagation direction.i,j To meet the probability distribution requirements, the diffusion probability between nodes is quantified by calculating the azimuth angles of the three adjacent cells with the highest probability, using the diffusion source as the origin. For example, the azimuth angles of the diffusion source cell 6 and its adjacent cell 5 are... 6,5 =1, A 6,5 =0.85, α6=0.42, then P 6,5 =1×0.85×0.42=0.357 (the highest among all adjacent cells), indicating that cell 6 has the highest probability of diffusing to cell 5. Therefore, the diffusion direction is: central cell 6 → cell 5 due west.

[0107] (4) Diffusion rate prediction. Based on the output time feature vector F time The focus is on characteristic dimensions related to thermal evolution (such as the rate of temperature rise, rate of voltage change, slope of voltage fall, and rate of increase of gas concentration), and a diffusion rate coefficient is defined. , which represents the number of monomers diffusing per unit time, and is used as the output for diffusion rate prediction.

[0108] ;

[0109] Among them, F time [m:n] represents the reference window for the change status of the m-th to n-th interval in the feature dimension, F time [0:m] represents the normal state observation window for the interval from 0 to m in the feature dimension. Here, is the adaptive coefficient, and mean is the average value. M is the dividing point, separating the normal state from the changing state; n is the end point of the second window, defining the time length for observing the changing state. For example:

[0110] ;

[0111] Among them, F time [10:20] represents the temperature rise rate correlation characteristic, F time [0:10] represents the normal time series feature vector.

[0112] (5) Coverage prediction. Based on the temporal feature evolution within the sliding window, the diffusion coverage N(t) within the next t seconds is predicted:

[0113] ;

[0114] Where, N total This represents the total number of individual cells in the battery pack. To round to the nearest integer. N(t) can be understood as the number of thermal runaway cells after t seconds (initially one diffusion source). For example, if v = 0.35 cells / s, when t = 2 seconds, N(2) = round(0.35 × 2) + 1 = 2 (covering the diffusion source and one adjacent cell, cell 6 → cell 5); when t = 5 seconds, N(5) = round(0.35 × 5) + 1 = 3 (covering 3 cells, cell 6 → cells 5 and 7).

[0115] (6) Prediction of thermal runaway propagation level. Based on The thermal runaway propagation risk level is determined by N(t) (Level 0: no propagation; Level 1-2: slow propagation; Level 3-4: rapid propagation; Level 5: rapid propagation). Taking a battery pack with 16 cells as an example, the thermal runaway propagation level is determined and output according to the table below.

[0116]

[0117] S6, Early Warning Test. Five sets of real thermal runaway samples (overcharge triggered) that were not used in training were selected for testing. Data was input in real time, and the model output the corresponding thermal runaway early warning level and spread trend. Comparison with real results shows that the method of this invention is superior to traditional methods (such as LSTM model (temporal features only) and CNN model (spatial features only) in small sample scenarios in terms of early warning accuracy, early warning time, and robustness.

[0118] In the actual implementation process, the installation position and number of sensors can be adjusted according to the actual battery pack structure and application scenario, appropriate outlier detection and normalization methods can be selected, and the parameters of time series analysis and models can be optimized to improve the accuracy and reliability of early warning.

[0119] Compared with the prior art, the beneficial effects of the present invention are:

[0120] Small sample size and high adaptability: Creatively combining transfer learning and GAN, the model can be effectively trained in small sample scenarios with only 10-50 sets of real thermal runaway samples, solving the core pain point of scarce industry data.

[0121] Deep fusion of spatiotemporal features: Through the architecture of 3D-CNN+LSTM+attention mechanism, the temporal dynamics and spatial correlation features are extracted synchronously and collaboratively, the feature representation is more comprehensive, and the ability to capture early weak anomalous signals is significantly improved.

[0122] Dynamic and accurate early warning: The model adopts a dynamic threshold mechanism based on sliding window and adaptive coefficients, which enables the model to adapt to feature changes caused by battery aging, etc., resulting in high early warning accuracy and long lead time.

[0123] Integrated diffusion trend prediction: While providing early warning, it seamlessly outputs a complete diffusion trend prediction, including diffusion source, direction, speed, range and risk level, based on the same set of spatiotemporal coupling characteristics, providing a direct decision-making basis for graded and zoned emergency response.

[0124] Highly practical: The sensor layout is reasonable and the algorithm modules are integrated.

[0125] Based on the same inventive concept, embodiments of the present invention provide a spatiotemporal dynamic early warning system for battery pack thermal runaway, used to execute the method described herein, the system comprising:

[0126] The data acquisition module is configured to acquire time-series data of each cell in the battery pack to form a spatiotemporal raw dataset;

[0127] The data augmentation module is configured to preprocess the original spatiotemporal dataset and then use an enhancement strategy that combines transfer learning and generative adversarial networks to generate virtual thermal runaway samples to augment the data and obtain an augmented dataset.

[0128] The feature coupling module is configured to extract the spatial correlation features and temporal dynamic features of the data in the expanded dataset, and to perform weighted fusion of the spatial correlation features and temporal dynamic features through an attention mechanism to obtain a spatiotemporal coupling feature vector;

[0129] The early warning adjustment module is configured to calculate the statistical distribution of the spatiotemporal coupling feature vector in real time based on a sliding window, adaptively adjust the early warning threshold, and output the thermal runaway early warning level according to the relationship between the spatiotemporal coupling feature vector and the early warning threshold.

[0130] The risk prediction module is configured to perform spatial diffusion topology modeling and temporal dynamic evolution prediction based on the spatiotemporal coupling feature vector, and output prediction results including diffusion source location, diffusion direction, diffusion speed, coverage area and diffusion risk level.

[0131] The system improves response efficiency through a multimodal sensing-computing-decision integrated architecture. Each module is connected through standardized interfaces, which can be deployed as a whole on a high-performance BMS platform or tailored according to computing resources (such as deploying only the first four modules for low-cost models), supporting flexible deployment and online upgrades.

[0132] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A spatiotemporal dynamic early warning method for thermal runaway of a battery pack, characterized in that, include: S1, acquire the time series data of each cell in the battery pack to form a spatiotemporal raw dataset; S2, after preprocessing the original spatiotemporal dataset, an enhancement strategy combining transfer learning and generative adversarial networks is adopted to generate virtual thermal runaway samples to expand the data and obtain an expanded dataset. S3, extract the spatial correlation features and temporal dynamic features of the data in the expanded dataset, and perform weighted fusion of the spatial correlation features and temporal dynamic features through an attention mechanism to obtain a spatiotemporal coupling feature vector; S4. Calculate the statistical distribution of the spatiotemporal coupling feature vector in real time based on a sliding window, adaptively adjust the warning threshold, and output the thermal runaway warning level according to the relationship between the spatiotemporal coupling feature vector and the warning threshold. S5. Based on the spatiotemporal coupling feature vector, perform spatial diffusion topology modeling and temporal dynamic evolution prediction, and output prediction results including diffusion source location, diffusion direction, diffusion speed, coverage area and diffusion risk level.

2. The method according to claim 1, characterized in that, S1 specifically includes: collecting time-series data of each battery cell by arranging a multi-type sensor array in the battery pack; the multi-type sensor array includes a temperature sensor arranged on the surface of the battery tabs, a voltage sensor connected to the positive and negative electrodes of the battery, and a gas sensor arranged inside the battery pack; the collected data includes the surface temperature, terminal voltage of each cell, and the concentration of H2, CO, and C2H4 gases in the battery pack, with a collection frequency of 1-10Hz.

3. The method according to claim 1, characterized in that, S2 specifically includes: S21, perform outlier removal and normalization on the original spatiotemporal dataset to obtain a standardized sample set; S22 uses normal operating data of the same type of battery as the source domain and small samples of thermal runaway as the target domain, and aligns the feature distribution of the source domain and the target domain through a domain adaptive network. S23. Using the aligned source domain data and small samples of the target domain as input, virtual spatiotemporal data is generated through a generative adversarial network. The similarity between the virtual samples and the real samples is verified by using dynamic time warping distance. Virtual samples with similarity higher than a preset threshold are selected and added to the dataset to obtain an expanded dataset.

4. The method according to claim 1, characterized in that, S3 specifically includes: 3D-CNN is used to extract spatial correlation features of the data in the dataset, and LSTM is used to extract temporal dynamic features. The 3D-CNN has 3 input channels, corresponding to temperature, voltage, and gas concentration data respectively; the LSTM has an input sequence length of 10-30 time steps.

5. The method according to claim 4, characterized in that, The weighted fusion formula in S3 is: ； Among them, F space F is the spatial correlation feature vector output by the 3D-CNN. time F is the temporal dynamic feature vector output by the LSTM, α is the attention weight adaptively calculated through the softmax function, and F is the time-series dynamic feature vector. fusion This is the fused spatiotemporal coupling feature vector.

6. The method according to claim 1, characterized in that, The specific process of adaptively adjusting the early warning threshold in S4 is as follows: Set a sliding window and calculate the mean μ and standard deviation σ of the spatiotemporal coupling feature vectors within the window; The warning threshold T is calculated using the formula T=μ+k·σ, where k is an adaptive coefficient adjusted based on the number of battery cycles. When the spatiotemporal coupling feature vectors of multiple consecutive time steps exceed the warning threshold T, a warning signal is triggered.

7. The method according to claim 1, characterized in that, S5 specifically includes: Calculate the spatial attention weight α for each battery i : ； Where Max(F) space [i]) is the maximum value of the spatial correlation feature of the i-th monomer, which is used to reflect the degree of monomer anomaly; Each individual node is defined as a topology node. Adjacency relationships between topology nodes are defined based on their physical location, and diffusion weight A is initialized. i,j Construct the spatial topology matrix A; Based on the spatial association features extracted by 3D-CNN, the spatial association strength S between any two topological nodes is calculated. i,j : ； Among them, F space [i] represents the spatial association feature of the i-th unit; Combined with diffusion weight A i,j Spatial correlation strength S i,j And attention weight α, which reflects the degree of individual abnormality. i Construct a diffusion probability matrix P, where the diffusion probability P0 is... i,j =A i,j ·S i,j ·α i , used to predict the direction of thermal runaway propagation, T is the transpose symbol.

8. The method according to claim 1, characterized in that, The time-dynamic evolution prediction in S5 includes: The diffusion rate coefficient is calculated based on the thermal evolution-related feature dimensions extracted from the temporal dynamic features using LSTM. , used to represent the number of monomers affected by thermal runaway propagation per unit time; ； Among them, F time [m:n] represents the reference window for the change status of the m-th to n-th interval in the feature dimension, F time [0:m] represents the normal state observation window for the interval from 0 to m in the feature dimension. Here, is the adaptive coefficient, and mean is the average value. Based on the diffusion rate coefficient Predict the coverage area N(t) of thermal runaway after t seconds; ； Where, N total This represents the total number of individual cells in the battery pack. This is for rounding to the nearest integer.

9. The method according to claim 8, characterized in that, S5 specifically includes: based on the diffusion rate coefficient Based on the predicted coverage area N(t), the risk level of thermal runaway propagation is classified.

10. A spatiotemporal dynamic early warning system for thermal runaway of a battery pack, characterized in that, The system for performing the method as described in any one of claims 1 to 9 includes: The data acquisition module is configured to acquire time-series data of each cell in the battery pack to form a spatiotemporal raw dataset; The data augmentation module is configured to preprocess the original spatiotemporal dataset and then use an enhancement strategy that combines transfer learning and generative adversarial networks to generate virtual thermal runaway samples to augment the data and obtain an augmented dataset. The feature coupling module is configured to extract the spatial correlation features and temporal dynamic features of the data in the expanded dataset, and to perform weighted fusion of the spatial correlation features and temporal dynamic features through an attention mechanism to obtain a spatiotemporal coupling feature vector; The early warning adjustment module is configured to calculate the statistical distribution of the spatiotemporal coupling feature vector in real time based on a sliding window, adaptively adjust the early warning threshold, and output the thermal runaway early warning level according to the relationship between the spatiotemporal coupling feature vector and the early warning threshold. The risk prediction module is configured to perform spatial diffusion topology modeling and temporal dynamic evolution prediction based on the spatiotemporal coupling feature vector, and output prediction results including diffusion source location, diffusion direction, diffusion speed, coverage area and diffusion risk level.

Images