Thermal runaway prediction method and device, electronic equipment and storage medium thereof

Abstract

The invention provides a prediction method and device for thermal runaway of a battery module, electronic equipment and a storage medium thereof. The prediction method comprises the following steps: acquiring multiple times, and the degree, the voltage, the internal resistance, the acoustic signal and the gas concentration fraction of each time; generating a code corresponding to the time; and generating an abnormal vector corresponding to the code, and processing the abnormal vector by using the trained thermal runaway probability prediction layer so as to obtain the probability of thermal runaway of the battery module in the future time. The prediction method can predict the thermal runaway of the battery module.

Description

Technical Field

[0001] The present invention relates to the field of battery technology, and in particular to a thermal runaway prediction method, device, electronic device, and storage medium thereof. Background Art

[0002] Clean energy sources, such as wind power and photovoltaics, are intermittent and volatile, requiring energy storage systems to achieve peak load regulation and smooth output. Traditional energy storage methods (such as pumped hydro and compressed air storage) are limited by terrain and cost constraints. Energy storage power stations incorporating battery modules offer advantages such as rapid operation, high energy density, and long lifespan, making them the best choice. However, inadequate protective facilities at energy storage power stations lead to frequent safety incidents, even fires and explosions, which pose serious risks to life and property.

[0003] Therefore, how to predict the thermal runaway of battery modules has become an urgent problem to be solved. Summary of the Invention

[0004] The object of the present invention is to provide a thermal runaway prediction method, device, electronic device and storage medium thereof.

[0005] In order to achieve one of the above-mentioned objects, an embodiment of the present invention provides a method for predicting thermal runaway of a battery module, comprising the following steps: obtaining a plurality of times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; obtain the battery module at time Time t Temperature T t , voltage V t , internal resistance R t , acoustic signal S t and gas concentration fraction G t , is the hydrogen gas volume fraction, is the oxygen volume fraction, Carbon dioxide volume fraction, C HC Hydrocarbon volume fraction; Generation time T The corresponding code H t ; Where t is a natural number, 1≤t≤N; Generate code H t The corresponding abnormal vector z t , use the trained thermal runaway probability prediction layer to predict the abnormal vectors z1, z2, ..., z N Processing to obtain in, for representing the probability of thermal runaway of the battery module at future time Time N+τ , where τ = 1, 2, …, k, k is a natural number.

[0006] As a further improvement of an embodiment of the present application, the “generating time Time t corresponding to the encoding H t ” specifically includes: EMA() means exponential moving average; DWT_Denoise() means discrete wavelet denoising; Z-score() means z-score; MahalaFilter() means Mahalanobis distance based outlier filtering; MelSpec() means mel-spectrogram, and L means sliding window length; where the square brackets “[]” function is to concatenate all parameters in the square brackets in order to form a multi-dimensional vector, μ(·) means to calculate the mean, σ(·) means to calculate the standard deviation, max(·) means to calculate the maximum value of all input parameters, skew(·) means to calculate the skewness, and corr(,) means to calculate the Pearson correlation coefficient; MelSpec 1:64 means to do short-time Fourier transform on the acoustic emission signal , map to 64 mel frequency bands and take the log energy logE1, …, logE 64 , and form a 64-dimensional spectral feature vector; Spectral Entropy means the spectral entropy of; where W T , W V , W R , W G and W S are trainable linear mappings, E type is a modality type encoding, and E time is a temporal encoding; H t = Encoder(e t-L+1 , e t-L+2 , …, e t ), where Encoder(·) is an encoder composed of multiple layers of Transformer Encoder.

[0007] As a further improvement of an embodiment of the present application, the “generating encoding H t corresponding to the anomaly vector at " specifically includes: d c = 3d emb , where σ is ReLU, σ(x) = max{0, x}, BN (Batch Normalization) is channel normalization, Conv1D k is one-dimensional convolution with kernel width k, k is the receptive field; w ch = σ(ReLU(Mean t (F t ))), Mean t is used to make global average in time dimension, compress the input tensor into a multi-dimensional vector, and aggregate the time sequence information; where Q t = F t W Q , K t = F t W K , V t = F t W V , W Q , W K and W V are trainable projection matrices, Q t is the query tensor, K t is the key tensor, V t is the value tensor, d a is the attention hidden dimension, and softmax is the normalized exponential function; where is the element-wise multiplication; a t = σ(W2ReLU(W1z t +b1)+b2), MaxPool is used to take the maximum value in the time dimension, W1 is the weight of the first fully connected layer, b1 is the bias of the first fully connected layer, W2 is the weight of the second fully connected layer, b2 is the bias of the second fully connected layer, and σ is the Sigmoid function.

[0008] As a further improvement of an embodiment of the present application, the thermal runaway probability prediction layer is used to: where the Attn() function is used for multi-head attention aggregation operation on query, key and value tensors; where w is the linear layer weight and b is the bias; where t+τ>N; the training loss of the thermal runaway probability prediction layer where q φFor inference network, BCE() is cross-entropy, KL() is divergence loss function, and p is weight value.

[0009] The embodiment of the application further provides a prediction device for thermal runaway of a battery module, comprising the following modules: an information acquisition module, configured to acquire a plurality of times Time1, Time2,..., TimeN, wherein i and N are natural numbers, and 1<=i<=N-1; a preprocessing module, configured to acquire temperature T1, voltage V1, internal resistance R1, acoustic signal S1 and gas concentration fraction G1 of the battery module at time Time1; a prediction module, configured to generate an encoding H1 corresponding to time Time1, and generate an abnormal vector z1 corresponding to the encoding H1 by using a trained thermal runaway probability prediction layer to process the abnormal vector z1, so as to obtain a probability P1 of thermal runaway of the battery module at future time Time1, wherein τ=1, 2,..., k, and k is a natural number. N , time Time i is earlier than time Time i+1 , wherein i and N are natural numbers, and 1<=i<=N-1; the preprocessing module is configured to acquire temperature T1, voltage V1, internal resistance R1, acoustic signal S1 and gas concentration fraction G1 of the battery module at time Time1; the prediction module is configured to generate an encoding H1 corresponding to time Time1, and generate an abnormal vector z1 corresponding to the encoding H1 by using a trained thermal runaway probability prediction layer to process the abnormal vector z1, so as to obtain a probability P1 of thermal runaway of the battery module at future time Time1, wherein τ=1, 2,..., k, and k is a natural number. t t t t t t , hydrogen volume fraction, oxygen volume fraction, carbon dioxide volume fraction, C HC hydrocarbon volume fraction; the prediction module is configured to generate an encoding H1 corresponding to time Time1, and generate an abnormal vector z1 corresponding to the encoding H1 by using a trained thermal runaway probability prediction layer to process the abnormal vector z1, so as to obtain a probability P1 of thermal runaway of the battery module at future time Time1, wherein τ=1, 2,..., k, and k is a natural number. t t ; wherein t is a natural number, and 1<=t<=N; the prediction module is configured to generate an encoding H1 corresponding to time Time1, and generate an abnormal vector z1 corresponding to the encoding H1 by using a trained thermal runaway probability prediction layer to process the abnormal vector z1, so as to obtain a probability P1 of thermal runaway of the battery module at future time Time1, wherein τ=1, 2,..., k, and k is a natural number. t t N wherein, is used to represent a probability of thermal runaway of the battery module at future time Time N+τ , wherein τ=1, 2,..., k, and k is a natural number.

[0010] As a further improvement of the embodiment of the application, the preprocessing module is further configured to: EMA() means exponential moving average; DWT_Denoise() is discrete wavelet denoising; Z-score() is z-score; MahalaFilter() is outlier filtering based on Mahalanobis distance; function MelSpec() is mel spectrum, and L is sliding window length; ​​​​​​​​​Where the role of the square brackets "[]" is to splice all the parameters in the input square brackets into a multi-dimensional vector in order, the meaning of μ(·) is to find the mean, the meaning of σ(·) is to find the standard deviation, the meaning of max(·) is to find the maximum value of all parameters input, The meaning of the first-order difference approximation is to find the first-order difference approximation, the meaning of skew(·) is to find the skewness, and the meaning of corr(,) is to find the Pearson correlation coefficient; MelSpec 1:64 The meaning of the short-time Fourier transform is to do short-time Fourier transform on the acoustic emission signal , map to 64 mel frequency bands and take the logarithmic energy logE1,..., logE 64 , form a 64-dimensional spectral feature vector; the meaning of Spectral Entropy is The spectral entropy of; Where W T , W V , W R , W G and W S are trainable linear mappings, E type is a modal type code, and E time is a time code; H t = Encoder(e t-L+1 , e t-L+2 ,..., e t ), where Encoder(·) is an encoder composed of multiple layers of Transformer Encoder.

[0011] As a further improvement of an embodiment of the present application, the prediction module is further used for:

[0012] d c = 3d emb , where σ is ReLU, σ(x) = max{0, x}, BN (Batch Normalization) is channel normalization, Conv1D k is a one-dimensional convolution with kernel width k, and k is the receptive field; w ch = σ(ReLU(Mean t (F t ))), Mean t is used to do global average in the time dimension, compress the input tensor into a multi-dimensional vector, and aggregate the time sequence information; Where Q t = F t W Q , K t = F t W K , V t=F t W V , W Q , W K and W V are trainable projection matrices, Q t is a Query tensor, K t is a Key tensor, V t is a Value tensor, d a is an attention hidden dimension, and softmax is a normalized exponential function. where is an element-wise multiplication. a t = sigma (W2 ReLU (W1z t +b1) +b2), MaxPool is used to take the maximum value in the time dimension, W1 is a full connection layer 1 weight, b1 is a full connection layer 1 bias, W2 is a full connection layer 2 weight, b2 is a full connection layer 2 bias, and sigma is a Sigmoid function.

[0013] As a further improvement of an embodiment of the present application, the thermal runaway probability prediction layer is used for: where the Attn() function is used for multi-head attention aggregation operation on the query, key and value tensors. where w is a linear layer weight and b is a bias. where t+tau>N; the training loss of the thermal runaway probability prediction layer where q φ is an inference network, BCE() is a cross-entropy, KL() is a divergence loss function, and p is a weight value.

[0014] The present application also provides an electronic device, which comprises a processor, a memory, and a program or instructions stored in the memory and executable on the processor, and the program or instructions are executed by the processor to implement the steps of the prediction method as described above.

[0015] The present application also provides a storage medium, which stores a program or instructions, and the program or instructions are executed by a processor to implement the steps of the prediction method as described above.

[0016] Compared to the prior art, the present invention provides a method, device, electronic device, and storage medium for predicting thermal runaway in a battery module. The method comprises the following steps: acquiring multiple times, along with the temperature, voltage, internal resistance, acoustic signal, and gas concentration fraction at each time; generating codes corresponding to the times; generating anomaly vectors corresponding to the codes, and processing the anomaly vectors using a trained thermal runaway probability prediction layer to determine the probability of thermal runaway occurring in the battery module at a future time. This method is capable of predicting thermal runaway in a battery module. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 4 is a flow chart of the prediction method in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be described in detail below with reference to the various embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0019] As used herein, terms indicating spatial relative positions such as "upper," "above," "lower," and "below" are used for ease of explanation to describe the relationship of one unit or feature relative to another unit or feature as shown in the accompanying drawings. Terms of spatial relative position may be intended to encompass different orientations of the device in use or operation in addition to the orientation shown in the drawings. For example, if the device in the drawings were turned over, units described as being "below" or "beneath" other units or features would be "above" the other units or features. Thus, the exemplary term "below" may encompass both above and below orientations. The device may be oriented in other ways (rotated 90 degrees or in other orientations), and the spatially relative descriptors used herein interpreted accordingly.

[0020] The first embodiment of the present invention provides a method for predicting thermal runaway of a battery module, such as Figure 1 As shown, the following steps are included:

[0021] Step 101: Get multiple times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1;

[0022] Step 102: Obtain the battery module at time Time t Temperature T t , voltage V t , internal resistance R t, acoustic signal S t , and gas concentration fraction G t , is a hydrogen gas volume fraction, is an oxygen gas volume fraction, a carbon dioxide volume fraction, C HC a hydrocarbon volume fraction; generation time Time t corresponding code H t ; wherein t is a natural number, 1≤t≤N;

[0023] Here, the unit of temperature T t may be ℃, the unit of voltage V t may be volts V, and the unit of internal resistance R j may be mΩ; wherein, is a hydrogen gas volume fraction, which is the most sensitive gas indicator in the early stage of thermal runaway; is an oxygen gas volume fraction, which is used to distinguish overcharge and oxygen evolution; a carbon dioxide volume fraction, which reflects the decomposition intensity of the electrolyte; C HC a hydrocarbon (e.g., methane, ethane, etc.) volume fraction, which is used to detect side reactions and leaks. Here, the "volume fraction" can be understood as a numerical value corresponding to the volume of a certain gas, and the larger the numerical value, the larger the volume fraction.

[0024] Acoustic emission (AE) is a natural phenomenon in which a material or structure releases part of the strain energy in the form of elastic waves while deforming or damaging under external or internal forces. It is a relaxation process that occurs when the material transitions from an unstable high-energy state to a stable low-energy state due to uneven stress distribution. Acoustic emission technology detects acoustic emission signals using specialized acoustic emission instruments and analyzes the detected signals or parameters to infer changes within the material.

[0025] Here, the acoustic emission sensor can collect transient waveforms at 100 kHz to 1 MHz, and the maximum amplitude in the original waveform is denoted as A max , which has units of mV or db. To synchronize with temperature, voltage, and other 1 Hz time series, A is calculated within each 1s sliding window, wherein n is the number of AE events captured within that 1s, and this processing is equivalent to converting "several high-frequency AE amplitudes" into "one energy indicator per second". It can be understood that the larger the value of S j , the higher the transient elastic wave energy released in that second, indicating that the material's internal micro-cracks or lithium precipitation mechanisms are more active.

[0026] Here, the time Time i corresponds to the time Time i+1The time interval between them can be a constant, for example, 1S, that is, the sampling period can be 1Hz.

[0027] Step 103: generating the encoding H t The corresponding abnormal vector z t , using the trained thermal runaway probability prediction layer to process the abnormal vectors z1, z2,..., z N , so as to obtain Wherein, for representing the probability of thermal runaway of the battery module at a future time Time N+τ , wherein τ=1,2,…,k, k is a natural number.

[0028] The prediction method can predict the probability of thermal runaway of the battery module in a future period of time.

[0029] In this embodiment, the "generating time Time t corresponding to the encoding H T " specifically includes:

[0030] EMA() means exponential moving average; DWT_Denoise() is discrete wavelet denoising; Z-score() is z-score; MahalaFilter() is outlier filtering based on Mahalanobis distance; The function MelSpec() is the Mel frequency spectrum, and L is the sliding window length;

[0031]

[0032] Wherein, the square brackets "[]" function is to concatenate all parameters in the square brackets into a multi-dimensional vector in order, μ(·) means to calculate the mean, σ(·) means to calculate the standard deviation, max(·) means to calculate the maximum value of all input parameters, the meaning of the first-order difference approximation is to calculate the skewness, corr(,) means to calculate the Pearson correlation coefficient; MelSpec 1:64 The meaning of the short-time Fourier transform of the acoustic emission signal is to map to 64 Mel frequency bands and take the logarithmic energy logE1,…,logE 64 , forming a 64-dimensional spectral feature vector; the meaning of SpectralEntropy is the spectral entropy of

[0033] where W T , W V , W R , W G and W S are trainable linear mappings, E type is a modality type encoding, E time is a time encoding;

[0034] H t = Encoder(e t-l+1 , e t-L+2 ,..., e t ), where Encoder(·) is an encoder stacked by multiple layers of Transformer Encoder.

[0035] For any time Time t , we can generate a corresponding array x t = [T t , V t , R t , G t , S t ], the sliding window length is L (the value of L can be 120), when the model is inferred at time t, take the latest L observations to form the input tensor X t-L+1:t = {x t-L+1 , x t-L+2 ,..., x t}.

[0036] The following preprocessing is performed: T t-L+1:t = {T t-L+1 , T t-L+2 ,..., T t}, where the meaning of EMA() is exponential moving average, which plays a role in smoothing the temperature sequence and suppressing high-frequency noise. V t-L+1:t = {V t-L+1 , V t-L+2 ,..., V t}, the function DWT_Denoise() is discrete wavelet denoising, which first does multi-scale wavelet decomposition, then soft threshold filtering on high frequency coefficients, and finally reconstructs the voltage signal. R t-L+1:t = {R t-L+1 , R t-L+2 ,..., R t}, the function z-score() is z-score, also called standard score, which is the process of dividing the difference between a number and the average by the standard deviation. The standard score is a symbol of the number of standard deviations that an observed or data point is above the average of the observed or measured values. Gt-L+1:t = {G t-L+1 ,G t-L+2 ,…,G t}, the function MahalaFilter() is Mahalanobis distance based outlier filtering, that is, for any point G k , t-L+1≤k≤t, the following calculation is performed: If χ 2 2-distributed threshold is exceeded, it is replaced by linear interpolation, and finally S t-L+1:t = {S t-L+1 ,S t-L+2 ,…,S t}, the function MelSpec() is the Mel energy spectrum, which performs STFT on 1MHz level AE time series, maps it to 64-dimensional Mel frequency band, and takes the logarithmic energy to finally obtain the acoustic feature vector. After the above five steps, is obtained as the input of the subsequent convolution-attention network.

[0037] After that,

[0038] The role of the square bracket “[]” is to concatenate the N parameters in the square bracket in order to form an N-dimensional vector, for example, Therefore, the value of is a 4-dimensional column vector formed by concatenating the 4 parameters and in order. The meaning of μ(·) is to take the mean, the meaning of σ(·) is to take the standard deviation,

[0039] the meaning of max(·) is to take the maximum value of all input parameters. The meaning of is to take the first-order difference approximation, that is, the first-order difference is the difference between two consecutive items in the discrete function,

[0040] reflecting the rate of change. The meaning of skew(·) is to take the skewness When skew(x1,x2,...,x L )>0, it indicates right skewness; when skew(x1,x2,...,x L )<0, it indicates left skewness; when skew(x1,x2,...,x L )=0, it indicates symmetric distribution. The meaning of corr(x,y) is to take the Pearson correlation coefficient,

[0041] μ xis the average of x1, x2,..., x L y is the average of y1, y2,..., y L MelSpec 1:64 The meaning of MelSpec is to do Short-Time Fourier transform (STFT) on the acoustic emission signal , map to 64 mel frequency bands and take the log energy logE1,..., logE 64 , form a 64-dimensional spectral feature vector. The meaning of SpectralEntropy is spectral entropy, which is used to describe the relationship between power spectrum and entropy rate. The calculation process is as follows: first, let Then The more dispersed the energy distribution is, the greater the spectral entropy is, reflecting the complexity of the frequency spectrum.

[0042]

[0043] After that, where W T , W V , W R , W G and W S are trainable linear mappings that compress or increase the dimension of each modality feature to a unified dimension d emb . It can be understood that after the column splicing processing of the square bracket [·], the resulting dimension is 5d emb ; the dimension of E type is 5d emb , which is a modality type embedding, that is, a same-dimension linear projection after adding a One-Hot identifier to the five pieces of features and to explicitly distinguish the source sensors; the dimension of E time is 5d emb , which is a time position encoding, which can be: 1, a learnable vector; 2, sine and cosine encoding sin and cos, used to express “which second the sample is located in the sliding window”. The dimensions of the five pieces of features and are both d m . Here, if d m = 24 and d emb = 32, then W T ∈R 32×24 , and after splicing, e t ∈R 5×32 = R 160 ​Here, concatenation has the following advantages: 1. Dimension uniformity. Different modalities have different original dimensions, and after mapping, they can be easily fed into the same Transformer; 2. Information complementarity. A single long vector contains temperature rise, overpressure, gas, and acoustic signs, which helps the model capture cross-modal coupling features; 3. Interpretability. E type Let the attention be explicitly aware of the sensor type, E time Preserve the timing order information. The subsequent process will feed the concatenated vector e t into a 12-layer Transformer Encoder to output a hidden sequence H t for the anomaly detection and probability prediction modules.

[0044] After that, where e t-L+1:t is e t-L+1 , e t-L+2 ,..., e t ; Encoder(·) is an encoder composed of N layers of Transformer Encoder; L is the sliding window length, for example, 120s, etc.; d hid is the hidden dimension, for example, 32 or 64. The structure of a single-layer Encoder is Z (l) = MHA(Y (l-1) ) + Y (l-1) , Y (l) = FFN(LN(Z (l) )) + LN(Z (l) ), and the MHA formula (j-numbered head) is where Q = Y (l-1) W Q , K = Y (l-1) W K , and V = Y (l-1) W V . The output represents These hidden vectors will be further utilized by the anomaly detection and probability prediction modules. Typical parameters are L = 120, d emb = 32, d hid = 32, N = 6, and h = 4. On an A6000 GPU, the single-batch inference delay is approximately 4.8 ms.

[0045] In this embodiment, the "generating an encoded H t corresponding to an anomaly vector a t " specifically includes:

[0046] d c = 3d embwhere, σ is ReLU, σ(x) = max{0, x}, BN (Batch Normalization) is channel normalization, Conv1D k is one-dimensional convolution with kernel width k, k is receptive field; w ch = σ(ReLU(Mean t (F t )00, Mean t is used to make global average in time dimension, compress input tensor into multi-dimensional vector and aggregate time sequence information; where, Q t = F t W Q , K t = F t W K , V t = F t W V , W Q , W K and W V are trainable projection matrices, Q t is Query tensor, K T is Key tensor, V T is Value tensor, d A is attention hidden dimension, and softmax is normalized exponential function; where, is element-wise multiplication; a T = σ(W2ReLU(W1z T +b1) + b2), MaxPool is used to take maximum value in time dimension, W1 is full connection layer 1 weight, b1 is full connection layer 1 bias, W2 is full connection layer 2 weight, b2 is full connection layer 2 bias, and σ is Sigmoid function.

[0047] Let the multi-modal embedding sequence be where, L is window length (can be 120s), d emb is single time fusion feature dimension. Multi-scale one-dimensional convolution uses three receptive fields: k ∈ {3, 5, 7}. Operation formula, d c = 3d emb where, σ is ReLU, σ(x) = max{0, x}, BN (Batch Normalization) is channel normalization, Conv1D k is one-dimensional convolution with kernel width k, that is, same padding with step length 1.

[0048] It can be understood that the multi-scale design adopts a convolution kernel width k = 3, 5, 7 to capture short-term (spike), medium-term (micro-trend), and long-term (slowly changing) features, respectively; the three-way features are concatenated and input into the next layer. The convolution Conv1D slides the window on the time axis to calculate the weighted sum and extracts the local pattern; same padding ensures that the output length is still L. Batch normalization BN normalizes each channel to zero mean and unit variance, stabilizes the gradient, and speeds up the convergence. ReLU activation σ introduces nonlinearity, suppresses negative values, and prevents gradient disappearance. The channel number changes after concatenation, and the channel dimension is d emB is expanded to d C = 3d EMB , providing rich features for subsequent attention modules. As shown in the following table:

[0049]

[0050]

[0051] For example, if d EMb = 32, L = 120:

[0052] SE Channel Attention (compressing L-dimensional to obtain channel weights), wherein, is the multi-scale convolution output. Mean t is the global average in the time dimension, which compresses the Lxd c tensor into a d c vector, aggregating the timing information. Excitation two-layer full connection (FC1 + ReLU + FC2 + Sigmoid) outputs weights w ch [j] ∈ (0, 1) for each channel. The scaling ratio r (usually 4 or 8) reduces the parameter amount. The final recalibration is multiplied by the channel in the subsequent feature , allowing the network to automatically focus on high-weight channels and suppress useless noise. For example, if d c = 96, r = 4: W1 ∈ R 96×24 , W2 ∈ R 24×96 , w ch ∈ R 96 . The maximum element of the channel weight is greater than 0.9, indicating that the convolution channel (corresponding to a specific receptive field + sensor combination) is most critical for anomaly detection.

[0053] Temporal Self Attention, wherein Q t = F t W Q , K t = F t WK ,

[0054] V t =F t W V . F t is the convolution output, W Q , W K and W V Both are trainable projection matrices, Q t is the Query tensor, K t is the Key tensor, V t is the Value tensor, d a is the attention hidden dimension, and softmax is a normalized exponential function.

[0055] As shown in the following table:

[0056]

[0057] Here, the linear projection will Mapped to a uniform dimension d a :Q t ,K t ,V t =F t W Q,K,V . Similarity calculation Represents the dot product similarity of any two rows of features, divided by Scaling stabilizes the gradient. Softmax normalization performs softmax on each row to obtain the attention weight matrix S t ∈[0,1] L×L , the weight of each row is 1. Weighted sum A t =S t V t Weight the information of different time steps to the current row.

[0058] The final fusion, Among them, ⊙ is element-by-element multiplication; is the channel attention weight, Expand to L×d by broadcasting a With A t Multiply to emphasize important channels and suppress noisy channels.

[0059] A t To adaptively weight the features of each moment in the “past L seconds” according to similarity and capture cross-time dependencies; ch To further scale by channel importance, we get Considering both “time correlation” and “channel importance” at the same time is to generate the abnormal probability vector a t Direct input.

[0060] Generate anomaly vectors and perform Max Pooling on the time dimension to obtain a single moment representation Then through two layers of full connection + Sigmoid: a t =σ(W2ReLU(W1z t +b1)+b2), MaxPool is used to maximize the time dimension, W1 is the weight of fully connected layer 1, b1 is the bias of fully connected layer 1, W2 is the weight of fully connected layer 2, b2 is the bias of fully connected layer 2, and σ is the Sigmoid function. Two layers of FC+Sigmoid are used for nonlinear combination to output an M-dimensional probability vector. M is the type of abnormal mode defined by the system (for example: lithium deposition, overcharging, thermal accumulation, SEI cracking, etc.). t [m] represents the confidence / activity of the mth type of anomaly. The closer it is to 1, the more abnormal it is.

[0061] As shown in the following table:

[0062]

[0063] What is an exception vector? t =[a t,1 ,a t,2 ,…,a t,M ] T ∈[0,1] M , 0≤a t,m ≤1.

[0064] The multi-scale convolution-attention network (MS-CAN) is used to collect the multimodal data X of the last L seconds. t-L+1:t After feature extraction, it is obtained through two layers of full connection + Sigmoid mapping. Dimension M: represents the M typical abnormal modes pre-defined by the system (for example: m = 1: lithium deposition m = 2: overcharge m = 3: heat accumulation m = 4: SEI cracking m = 5: gas leakage m = 6: sensor failure). Numerical meaning: Each element a t,m It is the "confidence / activity of the mth type of anomaly at time t", which can be regarded as "probability of occurrence" or "relative severity". The numerical range is: 0 means "no anomaly", 1 means "extremely strong anomaly". The empirical threshold is: a t,m >0.7 triggers emergency action for mode m; a t,m <0.2 is considered normal.

[0065] Example interpretation, a t =[0.12,0.83,0.05,0.03,0.01,0.02] T , the confidence level of the second dimension (overcharge) is 0.83>0.7 The judgment of "significant overcharge risk" triggers the priority strategy of "reduce charging current + local cooling". The rest of the dimensions <0.2, indicating that lithium analysis, heat aggregation, etc. are not active.

[0066] Example interpretation, a t =[0.05,0.82,0.10,0.03,0.02,0.01] T , the 2nd dimension (overcharge) confidence 0.82 indicates strong abnormality → immediately reduce the charging current and local cooling, the 3rd dimension (heat aggregation) confidence 0.10 indicates a slight trend, continue to monitor, and the rest of the dimensions <0.1, indicating normal noise. If a t [lithium analysis] ≈ 0.82, and then it can be determined that there is a significant lithium analysis trend and the current reduction and cooling strategy is triggered. If a t The elements are all below 0.2, the system enters the normal state channel, and the LLM returns "NoAction".

[0067] Here, 1, MaxPool focuses on the maximum abnormal peak value, which is not diluted by averaging; 2, double-layer FC provides nonlinear feature fusion, and the output dimension is the abnormal category; 3, Sigmoid probability, convenient for threshold judgment and subsequent decision; 4, the number of categories M can be expanded, only need to add labels and corresponding row parameters.

[0068] In this embodiment, the thermal runaway probability prediction layer is configured to: Wherein, the Attn() function is used for multi-head attention aggregation operation on query, key, value three tensors; Wherein, w is the linear layer weight, and b is the bias; Wherein, t+τ>N; the training loss of the thermal runaway probability prediction layer Wherein, q φ is the inference network, BCE() is the cross entropy, KL() is the divergence loss function, and p is the weight value.

[0069] Here, q φ is the inference network, the input H1, H2,..., H N , the output mean μ φ and variance In practice, z t can also be sampled, so as to compress historical information and bring a small amount of random disturbance, avoiding model memorization. Attn can be understood as a similarity weight, like the "guess you like" algorithm in short videos, features with high relevance are weighted more. Rollout τ-1 z t"Push" the future by τ-1 steps, if the difference between consecutive time probabilities is <0.01, the system keeps the last result and does not repeat the alarm. For example, the output KL can limit the sampling distribution from drifting too far; the weight p can be 0.1. The confidence calibration of the thermal runaway probability prediction layer T is the temperature, searched on the validation set to minimize the ECE (calibration error). T>1 means overall cooling (probability closer to 0.5), T<1 means more extreme probabilities.

[0070] Here, the goal of the thermal runaway probability prediction layer is to combine historical data + anomaly features to predict the probability of thermal runaway in the next 5, 10, 15 minutes. If the probability exceeds the threshold η=0.60, an early warning is triggered.

[0071] Generally includes the following steps:

[0072] Step 1, send "past 120 seconds of multi-modal data" into a GRU (recurrent network); the last cell output of the GRU is connected to an MLP, which outputs two vectors: mean μ t , variance σ t . Use these two sets of parameters to randomly sample a hidden vector z t from a Gaussian distribution. This step is called "variational sampling": give the model a little randomness to prevent memorization.

[0073] Step 2: z t is fused with global features using self-attention Attn: let z t "pay attention" to the part of H t that is most relevant to itself, and get the reinforced

[0074] Step 3: calculate the 5-minute risk, into a linear layer (weight w, bias b), and then pass through Sigmoid to squeeze the value to 0-1: that is, the "5-minute thermal runaway probability".

[0075] Step 4: roll to 10, 15 minutes, Rollout: run z t and Decoder again, which is equivalent to "fast forward the time further", get (predict 6-10 minutes). Go through the same Attn + linear + Sigmoid to get the 10-minute risk. Fast forward again → calculate the 15-minute risk. If the difference between the probability calculated after rolling and the last round is <0.01, the system keeps the last round result and does not repeat the alarm.

[0076] Triggering logic, as long as one of the 5 / 10 / 15-minute probabilities is >0.60 → immediately enter the disposal strategy retrieval. At the same time, the 95% confidence interval of the probability Report monitoring interface. Here σ MC Variance from multiple Dropout samples.

[0077] What "reward" to use during training to make it learn? BCE (cross-entropy): the closer the prediction to the true label, the higher the score. KL divergence: constrain the hidden vector from getting too far off; weights only 10% to prevent over-restricting.

[0078] In reinforcement learning, Rollout refers to the process of simulating and collecting sample data during training, where the agent interacts with the environment according to the current policy. In each Rollout, the agent observes the current state from the environment, then selects an action according to the chosen policy. Next, the agent interacts with the environment, executes the action, and observes the next state and the reward obtained. After that, the agent updates its policy according to the new state and selects the next action in the next step. This process continues until a certain stopping condition is met, such as reaching the maximum number of steps or reaching the terminal state. The purpose of Rollout is to generate sample data through interaction with the environment, which is used for policy optimization and value function estimation. These sample data will be used to update policy parameters or fit the value function to improve the performance of the agent. Usually, Rollout is performed multiple times during training to collect enough sample data to train and optimize the agent. The length of Rollout can be adjusted according to the specific task, which can be fixed or variable. In some continuous control tasks, Rollout may last for hundreds or thousands of time steps. However, in some discrete decision-making tasks, Rollout may be completed within a limited number of steps. In summary, Rollout is a process in reinforcement learning that involves interacting with the environment and collecting sample data to train the agent's policy and value function. Here, Rollout τ-1 For z t "Roll out" τ-1 steps into the future.

[0079] Example data (1000 validation samples), probability and true hit rate before and after temperature calibration (T=1.3 optimal), average prediction after calibration is closer to the true occurrence frequency, ECE decreases from 0.061 to 0.024.

[0080]

[0081]

[0082] Symbol table:

[0083]

[0084] Take the CLS vector H t Quantized to token, concatenated with system prompt word, sent to GPTDecoder. A special token is preposed to the input sequence of the TransformerEncoder <cls>After N-layer encoding, the hidden vector H T is regarded as the global semantic representation of the whole sequence.

[0085] where Q(·) is a vector quantizer with Product Quantization+Codebook; K is the Codebook size (e.g. 4096); m is the number of shards, which divides H t into m segments, each of which is independently looked up. The quantization result z t = [c1, …, c m is immediately available as an "exceptional context token".

[0086] System Prompt, example template - fixed content + real-time injected fields: <sys>You are a safety expert for energy storage systems. Please generate no more than 3 JSON treatment recommendations based on the input anomaly vector and historical fault knowledge graph.

[0087] Anomaly vector a t Where a t is replaced at runtime with a t,1 ,…,a t,M .

[0088] Token concatenation order, [System Prompt tokens‖z t ‖ <eos>→ GPT Decoder, where || denotes concatenate, <eos>The end mark.

[0089] The GPT Decoder's architecture is the same as the autoregressive Transformer decoder in GPT-2 / 3, enhanced with instruction fine-tuning, graph alignment, and RLAIF. The input is the concatenated token sequence from the previous step.

[0090] Output: token stream, obtained after de-decoding [

[0092] {"action":"local cooling","power":2kW,"duration":180s},

[0093] {"action":"Reduce charging current","ratio":0.5},

[0094] {"action":"Isolation gas spray","flow_rate":3L / min} ]

[0096] Generation strategy: Top-p sampling (p = 0.8) + temperature T = 0.7, maximum length 256 tokens.

[0097] Overall timing, The above process usually completes an inference within 30ms, meeting the real-time safety control requirements of the energy storage system.

[0098] Risk probability inference, like (Default θ = 0.42), triggering policy retrieval.

[0099] Main symbols in risk probability inference formula

[0100]

[0101] Trigger logic And θ=0.42, when the trigger logic is satisfied, the strategy will be retrieved.

[0102] The specific retrieval candidate action is: using the abnormal vector a t To query, retrieve the most similar k=20 historical disposal cases in the Milvus vector library The specific multi-objective screening is: for each candidate u i Calculate (P TR ,C asset ,C action ), and use the improved NSGA-II to find the Pareto frontier. The cost sorting is as follows: select the TR ≤θ under the premise of C asset +0.6C action Minimal action sequence u ★ . The specific execution is: u ★ is sent to the BMS / EMS, and the 1s convolution closed loop is monitored in real time. If the risk is not reduced, the suboptimal scheme is switched.

[0103] Numerical example, assume Then The system immediately enters the "strategy retrieval" state, extracts similar "local cooling + flow reduction" combinations from the database in the past, and filters the optimal disposal sequence according to the constraints. This step converts the Decoder implicit semantics into "future 5 / 10 / 15 minute thermal runaway probability". When the risk exceeds the threshold θ, it triggers Case-Based Retrieval + Multi-Objective Optimization, ensuring timely and cost-effective disposal actions.

[0104] The disposal knowledge evolution layer, LLM (fine-tuned GPT-4o), generates intervention sequences aligned with the "TR graph". GPT4o is fine-tuned and aligned with the TR graph.

[0105] Data resources and task definition

[0106] TR graph: node V = {v i} contains Failure-Mode, Sensor-Pattern, Mitigation-Action, Asset, Cost five categories; edge E = {(v i ,r,v j )} contains induces, measured_by, mitigated_by, affects four kinds of relationships.

[0107] Instruction data: composed of 24,000 triples → intervention steps, 11,000 dialogue questions and answers, and 30,000 expert annotation examples.

[0108] Objective: given the abnormal vector a t and the Top-k graph subgraph g t = (V t ,E t ), generate JSON sequence [{"action":..., "param":...},...], while ensuring that the action is within the graph feasible edge E t .

[0109] The "abnormal vector a t " and the thermal runaway probability of each time period have been generated. But what actions should be taken (reduce the charging current, local cooling, nitrogen injection, etc.). It also needs to refer to past experience and professional knowledge. These experiences are organized into a "knowledge graph" - the data resource to be introduced below.

[0110]

[0111] What is a TR (Thermal Runaway) graph? Simply put: a graph = a network of "failure—feature—action—cost" relationships. Before the model decides on an action, it first "walks the graph" to confirm feasibility.

[0112] This is a package of training examples that write "graph knowledge" as triples: 24k "Failure-Mode→Sensor-Pattern→Mitigation-Action"

[0113] Example: (Failure-Mode: overcharge, Sensor-Pattern: high voltage, Mitigation-Action: reduce charging current) → {"action": "reduce charging current", "ratio": 0.5}.

[0114] 1.1k Q&A pairs, e.g., "If Pack-1 temperature rises to 65°C and H2 concentration rises, what should be done?"

[0115] 0.3k expert annotations, where experts tell the model "this answer is more professional / safe" to fine-tune tone.

[0116] These data allow GPT-4o to learn: "understand the anomaly → find the action in the graph → answer in JSON".

[0117] The final goal of the model is to take in the current anomaly vector $\mathbf{a}_t$ (tell the model "which risk is the most prominent") and the Top-k subgraph $\mathcal{G}_t = (V_t, E_t)$ in the graph that is most relevant to the current situation.

[0118] Output: a list of JSON actions, such as [{"action": "local cooling", "power": 2, "unit": "kW", "duration": 180, "unitTime": "s"}, {"action": "reduce charging current", "ratio": 0.5}].

[0119] And it must satisfy two points: the action is legal - every action in the output can find a "Mitigation-Action" edge in the subgraph $E_t$, indicating that it can indeed solve the current failure mode; the parameters are feasible - power, duration, cost, etc. come from the "Cost" node to avoid giving instructions that machines cannot do or exceed the budget.

[0120] Its relationship with the previous text is: the anomaly detection layer → tells us "what risk is active". The probability prediction layer → tells us "how fast will it run out of control".

[0121] Atlas + GPT-4o (resources explained in this section) → translate the results of 1, 2 into specific, compliant, and cost-controllable disposal actions. In this way, a complete closed loop is formed: monitoring → prediction → decision → execution → re-monitoring.

[0122] Instruction Tuning: Minimize the Cross-Entropy loss using ∑ j logP θ (y j ∣y ＜j , prompt). Triple Alignment: Use Contrastive Triplet Loss, where h + , h - are positive and negative triplet vectors, and γ is the interval. Retrieval-Augmentation: At runtime, use the anomaly vector a t as a query to retrieve similar subgraphs from the vector library (200,000 entities) and concatenate them to the system prompt. RLAIF fine-tuning: Construct the reward r = αΔP TR - βCost - λLLM_length, fine-tune the GPT4o weights using PPO, reinforcing the strategy of "fast risk reduction, low cost, and concise response".

[0123] Inference stage interface, Prompt = <sys>: abnormal vector a t , subgraph g t ; <instruction>→ GPT-4o.

[0124] Model output [{"action": "local cooling", "power": 2 kW, "duration": 180 s}, {"action": "reduce charging current", "ratio": 50%}, {"action": "activate nitrogen isolation", "flow_rate": 3 L / min}]

[0125] Online evolutionary mechanism, collect ΔP after execution TR Cost action t resp and other indicators; if r < r min , write {output, r} to the Replay Buffer, and update the RLAIF offline every night.

[0126] 5. Ablation verification results (summary)

[0127] Model configurations F1@0.42 Avg. Cost Avg. ΔP TR ]] BaseGPT-4o (untuned) 0.67 1.00 0.21 + Instruction Tune 0.78 0.92 0.29 + Triplet Align 0.82 0.90 0.31 + Retrieval 0.85 0.88 0.33 + RLAIF (full) 0.89 0.83 0.37

[0128] The treatment knowledge evolution layer refines instructions, aligns triples, enhances retrieval, and uses reinforcement learning in four steps to enable GPT-4o to output efficient, low-cost, and interpretable intervention sequences under graph constraints, and relies on RLAIF for continuous self-evolution.

[0129] Target vector

[0130]

[0131] u sequence = a sequence of intervention actions, e.g., u = [local cooling 2 kW 180 s, reduce charging current 50%, nitrogen spray 3 L / min]. Each action has a cost, time consumption, and a risk reduction effect.

[0132] "Three-dimensional" and "single-target" gameplay, three-dimensional gameplay: consider risk, asset loss, and action cost simultaneously - this is called "multi-objective optimization". Pareto frontier: a set of "not worse than each other" solutions, left to the dispatcher to choose. Single-target gameplay: weight the three dimensions (α + β + γ = 1) to form a total score J. Closs = how much money you have to pay for not reducing risk; Cinterv = how much money you spend on actions; Ctime = additional risk due to slow response; business side only cares about one score for better decision-making.

[0133] Constraints, temperature rise rate: max +5℃ per minute (national standard GB / T36276-2023). maxtdT / dt → "temperature cannot rise too fast at any time". Cooling power: single cooling module ≤ 3 kW, grid capacity limit. Isolation cabin trigger: at most 2 times a day, to avoid frequent switching losses.

[0134]

[0135]

[0136] Symbol: pmut: probability of each gene being randomly altered; pmin / pmax: set minimum, maximum mutation rate; pmem: probability of participating in crossover for the case-memory scheme; PTRcur: current risk.

[0137] Online closed-loop execution

[0138] Step 1, choose a case: select a case from the Pareto frontier that satisfies PTR < θ, then use the formula "a · (asset loss) + β · (action cost)" to do a secondary sorting. Get the optimal u * . F2 / F3: the two dimensions of "asset loss" and "action + time" in the previous three dimensions. Arg min u: pick the one with the lowest cost.

[0139] Step 2, dynamic monitoring: after the case is issued, recalculate the risk every 1 second: ΔPTR = old risk - new risk. ε: preset "effect threshold" (such as 0.02). ΔPTR < ε → indicates that it is not enough to change the optimal case; at most 3 times, to prevent frequent switching.

[0140] Write to the case memory (Case-Memory), archive <u * , old risk, new risk, action cost>: later initialization can use the "old gold case" as a reference to make the genetic algorithm find a good solution faster.

[0141] Small example

[0142]

[0143] Threshold θ = 0.40 → A, B are both available; secondary sorting: α = 0.6 (asset loss is more important), β = 0.4; A: 0.6 × 10000 + 0.4 × 300 = 6120; B: 0.6 × 9000 + 0.4 × 900 = 5760 ← B is the first choice. Issue B → check the risk every second, if after 10 seconds the risk only decreases from 0.35 to 0.34 (Δ = 0.01 < ∈) → change A. After execution, the risk decreases to 0.29, and this record is archived for future reference.

[0144]

[0145]

[0146] Main formula symbol interpretation

[0147]

[0148] A second embodiment of the present invention provides a device for predicting thermal runaway of a battery module, comprising the following modules:

[0149] Information acquisition module, used to obtain multiple time Time1, Time2, ..., Time n Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1;

[0150] The preprocessing module is used to obtain the battery module at time Time t Temperature T t , voltage V t , internal resistance R t , acoustic signal S t and gas concentration fraction G t , is the hydrogen gas volume fraction, is the volume fraction of oxygen, Carbon dioxide volume fraction, C HC Hydrocarbon volume fraction; Generation time t The corresponding code H t ; Wherein, t is a natural number, 1≤t≤N;

[0151] Prediction module, used to generate the code H t The corresponding abnormal vector z t , use the trained thermal runaway probability prediction layer to predict the abnormal vectors z1, z2, ..., z N Processing to obtain in, Used to indicate time in the future N+τ , the probability of thermal runaway of the battery module, where τ = 1, 2, …, k, and k is a natural number.

[0152] In this embodiment, the preprocessing module is further used to:

[0153] EMA() means exponential moving average; DWT_Denoise() is discrete wavelet denoising; Z-score() is the z score; MahalaFilter() is an outlier filter based on Mahalanobis distance; The function MelSpec() is the Mel spectrum, and L is the sliding window length;

[0154]

[0155] where the role of the square brackets "[]" is to concatenate all the parameters in the input square brackets into a multi-dimensional vector in order, μ(·) means to take the mean, σ(·) means to take the standard deviation, max(·) means to take the maximum value of all the input parameters, the meaning of the first-order difference approximation is to take the skewness, and the meaning of corr(,) is to take the Pearson correlation coefficient; MelSpec 1:64 The meaning of MelSpec is to do short-time Fourier transform on the acoustic emission signal , map to 64 mel frequency bands and take the log energy logE1,…,logE 64 , form a 64-dimensional spectral feature vector; The meaning of SpectralEntropy is the spectral entropy of;

[0156] where W T , W V , W R , W G and W S are trainable linear mappings, E type is a modal type code, and E time is a time code;

[0157] H t =Encoder(e t-L+1 ,e t-L+2 ,...,e t ), where Encoder(·) is an encoder composed of multiple layers of Transformer Encoder.

[0158] In this embodiment, the prediction module is further configured to:

[0159]

[0160] d c =3d emb , where σ is ReLU, σ(x)=max{0,x}, BN(BatchNormalization) is channel normalization, Conv1D k is a one-dimensional convolution with kernel width k, and k is the receptive field;

[0161] w ch =σ(ReLU(Mean t (F t ))), Mean t is used to take the global average in the time dimension, compress the input tensor into a multi-dimensional vector, and aggregate the time sequence information;

[0162] wherein Q t =F t W Q ,K t =F t W K ,V t =F t W V ,W Q ,W K and W V are trainable projection matrices, Q t is a Query tensor, K t is a Key tensor, V t is a Value tensor, d a is an attention hidden dimension, and softmax is a normalized exponential function. wherein is an element-wise multiplication.

[0163] a t = sigma (W2 ReLU (W1z t +b1) +b2), MaxPool is used to take the maximum value in the time dimension, W1 is a full connection layer 1 weight, b1 is a full connection layer 1 bias, W2 is a full connection layer 2 weight, b2 is a full connection layer 2 bias, and sigma is a Sigmoid function.

[0164] In the embodiment, the thermal runaway probability prediction layer is used for:

[0165] wherein the Attn() function is used for a multi-head attention aggregation operation on the query, key and value tensors. wherein w is a linear layer weight and b is a bias. wherein t+tau>N; the training loss of the thermal runaway probability prediction layer wherein q φ is an inference network, BCE() is a cross entropy, KL() is a divergence loss function, and p is a weight value.

[0166] Embodiment three of the present application provides an electronic device, which comprises a processor, a memory, and a program or instructions stored on the memory and executable on the processor, and the program or instructions are executed by the processor to implement the steps of the prediction method in embodiment one.

[0167] Embodiment four of the present application provides a storage medium, which stores a program or instructions, and the program or instructions are executed by a processor to implement the steps of the prediction method in embodiment one.

[0168] It should be understood that although the present specification is described in terms of embodiments, not every embodiment exhibits every characteristic or exhibits the described implementation to the full extent. The description herein of any particular embodiment is merely intended to describe the particular embodiment and is not intended to limit the scope of the application. Various modifications of the application can be made by persons of ordinary skill in the art without departing from the spirit or scope of the application. It is recognized that certain embodiments are described in terms of steps performed by individuals or groups of individuals. It should be understood that these descriptions are merely illustrative of the steps that can be performed by a computer, a group of computers, or other electronic devices capable of carrying out the steps.

[0169] The above detailed description merely describes certain particular embodiments of the application, and is not intended to limit the scope of the application. Various modifications of the application can be made by persons of ordinary skill in the art without departing from the spirit or scope of the application. It is recognized that certain embodiments are described in terms of steps performed by individuals or groups of individuals. It should be understood that these descriptions are merely illustrative of the steps that can be performed by a computer, a group of computers, or other electronic devices capable of carrying out the steps.< / instruction> < / sys> < / eos> < / eos> < / sys> < / cls>

Claims

1. A method for predicting thermal runaway of a battery module, characterized in that: The following steps are involved: Get multiple times Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; Get the battery module at time Time t Temperature T t , voltage V t , internal resistance R t , acoustic signal S t and gas concentration fraction G t , is the hydrogen gas volume fraction, is the oxygen volume fraction, Carbon dioxide volume fraction, C HC Hydrocarbon volume fraction; Generation time t The corresponding code H t ; Wherein, t is a natural number, 1≤t≤N; Generate code H t The corresponding abnormal vector z t , use the trained thermal runaway probability prediction layer to predict the abnormal vectors z1, z2, ..., z N Processing to obtain in, Used to indicate time in the future N+τ , the probability of thermal runaway of the battery module, where τ = 1, 2, …, k, and k is a natural number.

2. The prediction method according to claim 1, characterized in that The "Generation Time" t The corresponding code H t Specifically include: EMA() means exponential moving average; DWT_Denoise() is discrete wavelet denoising; Z-score() is the z score; MahalaFilter() is an outlier filter based on Mahalanobis distance; The function MelSpec() is the Mel spectrum, and L is the sliding window length; The square brackets "[]" are used to sequentially concatenate all the parameters in the input square brackets into a multidimensional vector. μ(·) means to find the mean, σ(·) means to find the standard deviation, and max(·) means to find the maximum value of all the input parameters. The meaning of is to find the first-order difference approximation, the meaning of skew(·) is to find the skewness, and the meaning of corr(,) is to find the Pearson correlation coefficient; MelSpec 1:64 The meaning of acoustic emission signal Perform short-time Fourier transform, map to 64 Mel-bands and take the logarithmic energy logE1,…,logE 64 , forming a 64-dimensional spectral feature vector; the meaning of Spectral Entropy is spectral entropy of ; Among them, W T , W V , W R , W G and W S Both are trainable linear mappings, E type is the modality type code, E time Code for time; H t =Encoder(e t-L+1 ,e t-L+2 ,...,e t ), where Encoder(·) is an encoder composed of multiple layers of Transformer Encoder.

3. The prediction method according to claim 2, characterized in that The "generated code H t The corresponding exception vector a t Specifically include: d c =3d emb , where σ is ReLU, σ(x) = max{0,x}, BN (BatchNormalization) is channel normalization, Conv1D k is a one-dimensional convolution with kernel width k, where k is the receptive field; w ch =σ(ReLU(Mean t (F t ))), Mean t Used to perform global averaging in the time dimension, compress the input tensor into a multi-dimensional vector, and aggregate time series information; Among them, Q t =F t W Q , K t =F t W K , V t =F t W V , W Q , W K and W V Both are trainable projection matrices, Q t is the Query tensor, K t is the Key tensor, V t is the Value tensor, d a is the attention hidden dimension, and softmax is the normalized exponential function; Among them, ⊙ is element-by-element multiplication; a t =σ(W2ReLU(W1z t +b1)+b2), MaxPool is used to maximize the time dimension, W1 is the weight of the fully connected layer 1, b1 is the bias of the fully connected layer 1, W2 is the weight of the fully connected layer 2, b2 is the bias of the fully connected layer 2, and σ is the Sigmoid function.

4. The prediction method according to claim 3, characterized in that The thermal runaway probability prediction layer is used to: Among them, the Attn() function is used to perform multi-head attention aggregation operations on the query, key, and value tensors; Among them, w is the linear layer weight and b is the bias; Where t+τ>N; the training loss of the thermal runaway probability prediction layer Among them, q φ To infer the network, BCE() is the cross entropy, KL() is the divergence loss function, and p is the weight value.

5. A device for predicting thermal runaway of a battery module, characterized in that: Includes the following modules: Information acquisition module, used to obtain multiple time Time1, Time2, ..., Time N Time i Earlier than Time i+1 , where i and N are both natural numbers, 1≤i≤N-1; The preprocessing module is used to obtain the battery module at time Time t Temperature T t , voltage V t , internal resistance R t , acoustic signal S t and gas concentration fraction G t , is the hydrogen gas volume fraction, is the oxygen volume fraction, Carbon dioxide volume fraction, C HC Hydrocarbon volume fraction; Generation time t The corresponding code H t ; Wherein, t is a natural number, 1≤t≤N; Prediction module, used to generate the code H t The corresponding abnormal vector z t , use the trained thermal runaway probability prediction layer to predict the abnormal vectors z1, z2, ..., z N Processing to obtain in, Used to indicate time in the future N+τ , the probability of thermal runaway of the battery module, where τ = 1, 2, …, k, and k is a natural number.

6. The prediction device according to claim 5, characterized in that The pre-processing module is further configured to: EMA() means exponential moving average; DWT_Denoise() is discrete wavelet denoising; Z-score() is the z score; MahalaFilter() is an outlier filter based on Mahalanobis distance; The function MelSpec() is the Mel spectrum, and L is the sliding window length; The square brackets "[]" are used to sequentially concatenate all the parameters in the input square brackets into a multidimensional vector. μ(·) means to find the mean, σ(·) means to find the standard deviation, and max(·) means to find the maximum value of all the input parameters. The meaning of is to find the first-order difference approximation, the meaning of skew(·) is to find the skewness, and the meaning of corr(,) is to find the Pearson correlation coefficient; MelSpec 1:64 The meaning of acoustic emission signal Perform short-time Fourier transform, map to 64 Mel-bands and take the logarithmic energy logE1,…,logE 64 , forming a 64-dimensional spectral feature vector; the meaning of SpectralEntropy is spectral entropy of ; Among them, W T , W V , W R , W G and W S Both are trainable linear mappings, E type is the modality type code, E time Code for time; H t =Encoder(e t-L+1 ,e t-L+2 ,...,e t ), where Encoder(·) is an encoder composed of multiple layers of Transformer Encoder.

7. The prediction device according to claim 6, characterized in that The prediction module is also used to: d c =3d emb , where σ is ReLU, σ(x) = max{0,x}, BN (BatchNormalization) is channel normalization, Conv1D k is a one-dimensional convolution with kernel width k, where k is the receptive field; w ch =σ(ReLU(Mean t (F t ))), Mean t Used to perform global averaging in the time dimension, compress the input tensor into a multi-dimensional vector, and aggregate time series information; Among them, Q t =F t W Q , K t =F t W K , V t =F t W V , W Q , W K and W V Both are trainable projection matrices, Q t is the Query tensor, K t is the Key tensor, V t is the Value tensor, d a is the attention hidden dimension, and softmax is the normalized exponential function; Among them, ⊙ is element-by-element multiplication; a t =σ(W2ReLU(W1z t +b1)+b2), MaxPool is used to maximize the time dimension, W1 is the weight of the fully connected layer 1, b1 is the bias of the fully connected layer 1, W2 is the weight of the fully connected layer 2, b2 is the bias of the fully connected layer 2, and σ is the Sigmoid function.

8. The prediction device according to claim 7, characterized in that The thermal runaway probability prediction layer is used to: Among them, the Attn() function is used to perform multi-head attention aggregation operations on the query, key, and value tensors; Among them, w is the linear layer weight and b is the bias; Where t+τ>N; the training loss of the thermal runaway probability prediction layer Among them, q φ To infer the network, BCE() is the cross entropy, KL() is the divergence loss function, and p is the weight value.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the prediction method according to any one of claims 1 to 4.

10. A storage medium, characterized in that: The storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the prediction method according to any one of claims 1 to 4 are implemented.

Images