Monitoring and early warning device for identifying thermal runaway of battery and operation method of monitoring and early warning device

Abstract

The invention discloses a monitoring and early warning device for identifying thermal runaway of a battery and an operation method thereof, belongs to the technical field of battery monitoring and early warning, and realizes high-precision battery anomaly detection by combining an advanced machine learning algorithm, a deep learning technology and a signal processing method. The reliability and the expandability of the system in practical application are ensured by the modular design and the comprehensive evaluation framework of the system, and the system can adapt to continuously changing data modes through continuous optimization and monitoring, so that powerful support is provided for battery quality control and performance optimization, and the system is suitable for popularization and application. Therefore, the thermal runaway safety risk of the batteries used by an energy storage system and an electric vehicle is accurately monitored and warned at an early stage, the thermal runaway spreading among the batteries is effectively prevented, and the occurrence probability and the loss degree of fire and explosion accidents caused by thermal runaway of the batteries are greatly reduced; therefore, quality improvement and efficiency improvement of safety risk prevention and control work in the new energy industry are facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of battery monitoring and early warning, and in particular to a monitoring and early warning device for identifying battery thermal runaway and an operating method thereof. Background Art

[0002] With the rapid development of the new energy industry, the scale of electrochemical energy storage installations and the number of electric vehicles in the market have increased rapidly year by year. New electrochemical energy storage batteries such as lithium ion and sodium ion are widely used in energy storage systems and electric vehicles due to their high energy density, light weight, and long life. However, in recent years, new energy fire and explosion accidents have occurred frequently, exposing the serious safety risks of new electrochemical batteries. Once the battery has thermal runaway, it will instantly release a large amount of flammable and explosive high-temperature gases such as hydrogen, carbon monoxide and methane. Once flammable, it will cause violent combustion or explosion, resulting in serious personal injury and property loss. It has become an unfavorable factor hindering the healthy and rapid development of the new energy industry in the future.

[0003] Therefore, it is an urgent practical problem to carry out extremely early warning and efficient disposal of battery thermal runaway before disaster. The thermal runaway warning technology is used for online real-time monitoring during the normal cycle charging and discharging of the battery. Once the battery shows the risk characteristics of thermal runaway in advance, the necessary emergency disposal measures are implemented in time, which will effectively suppress the scale of thermal runaway disasters and prevent the expansion of thermal runaway disaster losses.

[0004] The battery thermal runaway process is accompanied by changes in multiple physical parameters from the early stage to the late stage. The current technologies used to monitor battery thermal runaway include sound, voltage, internal resistance, gas composition, gas pressure, temperature and other methods. However, some of the above methods are only suitable for late-stage thermal runaway monitoring and cannot achieve early warning. Some of them are due to the complex actual on-site working conditions, resulting in noise interference and poor warning reliability.

[0005] The early warning of battery thermal runaway in real engineering projects needs to meet the requirements of early warning and high precision and reliability at the same time. Therefore, using the shell deformation characteristics to monitor the battery thermal runaway state in the early stage is a very promising early warning method. Unlike other technologies, the battery shell will deform significantly in the early stage of the thermal runaway safety risk. The battery module is filled with dozens of battery cells and is electrically connected in series and parallel, neatly arranged and pre-tightened with a restraining rolling strip. The deformation of the battery thermal runaway shell will cause the normal restraint pre-tightening force distribution between the batteries to change significantly. Using the pressure change characteristics between batteries for thermal runaway safety risk early warning can solve the two difficult problems of early monitoring and reliable warning at the same time, and has good practicality and promotion value.

[0006] At present, some research institutions also use pressure films for monitoring and early warning, but the pressure characteristics they select are single, and they often use the absolute value and change rate of the total pressure on the battery contact surface for feature judgment. In fact, the distribution of battery deformation and stress on the battery contact surface is uneven. Using a single data of total pressure for safety risk early warning often loses the information on the pressure distribution changes at different positions on the contact surface between batteries, and cannot use the pressure differential distribution changes and coupled change characteristics for early warning. The early warning effect cannot achieve the best, which reduces the utilization rate of the pressure data array at multiple points of the pressure film, resulting in a high cost of using single channel data;

[0007] In addition, since the battery shell itself is not very flat and the thickness of ordinary pressure film is relatively thin, when it is set between batteries for pressure measurement, the pressure film cannot achieve close contact with adjacent batteries, resulting in inaccurate pressure monitoring values, which seriously affects the reliability of battery thermal runaway warning. Moreover, ordinary pressure films have poor thermal insulation and heat resistance and cannot block the spread of heat from thermal runaway batteries to surrounding batteries, resulting in chain thermal runaway between batteries in the battery module due to heat conduction. Therefore, a monitoring and early warning device for identifying battery thermal runaway and an operating method thereof are proposed. Summary of the invention

[0008] The purpose of the present invention is to provide a monitoring and early warning device for identifying battery thermal runaway and an operating method thereof, so as to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions: a monitoring and early warning device for identifying thermal runaway of a battery, comprising a battery pack and a pressure sensing assembly, wherein the pressure sensing assembly is electrically connected to a pressure collector via a wire, wherein the pressure collector is electrically connected to a pressure monitoring and early warning module via a wire, and wherein the pressure monitoring and early warning module is electrically connected to a computer group via a wire;

[0010] The battery pack includes a plurality of batteries, and the pressure sensing assembly is installed on one side adjacent to two batteries;

[0011] The pressure sensing assembly comprises a pressure sensing film and two flexible thermal insulation sheets, and the pressure sensing film is installed on one adjacent side of the two flexible thermal insulation sheets.

[0012] As a further preferred embodiment of the present technical solution: the pressure sensing film adopts a pressure sensing area array with a row number n not greater than 200 and a column number m not greater than 100, and the flexible insulation sheet is made of aerogel felt or glass fiber wool, and the thickness of the flexible insulation sheet is not greater than 5 mm.

[0013] As a further preferred embodiment of the present technical solution: the pressure sensing component and the pressure collector are connected by independent wiring or common wiring, and the pressure collector and the pressure monitoring and early warning module communicate by CAN, RS485 or RS232 bus.

[0014] A method for operating a monitoring and early warning device for identifying battery thermal runaway comprises the following steps:

[0015] S1. The pressure sensing film in the pressure sensing component senses the external pressure given by the battery and converts it into an electrical signal output, thereby achieving pressure measurement and monitoring;

[0016] S2. The two flexible heat-insulating sheets are used to improve the heat dissipation efficiency and heat preservation effect of the pressure sensing film, and play the role of shock absorption, insulation and sealing for the pressure sensing film. At the same time, the gaps between the battery packs can be filled;

[0017] S3, monitor and record pressure data in real time through the set pressure collector, so as to facilitate monitoring and management;

[0018] S4. The set pressure monitoring and early warning module analyzes and processes the collected information data by running an algorithm;

[0019] S5. Finally, the analysis results are input into the set computer group.

[0020] As a further preferred embodiment of the present technical solution: in S4, the running algorithm includes a data preprocessing module, a support vector machine model, a long short-term memory network model, a model evaluation index module, an integrated learning module, a feature engineering module, a model interpretation module and a model deployment and monitoring module;

[0021] The data preprocessing module includes a time series creation unit, a standardization unit and a data enhancement unit;

[0022] The time series creation unit converts the original information data into a time series of fixed length. Given an original data sequence X with a length of T = {x1, x2, ..., x T}, where each x t ∈R d is a d-dimensional feature vector, and a sliding window method is used to create a sequence sample of length L. The specific formula is:

[0023] S i ={x i , x i+1 , ..., x i+L-1}, i=1, 2,..., T-L+1(1)

[0024] This gives us a three-dimensional tensor:

[0025] x∈R N×L×d (2)

[0026] In formula (2), N = T-L + 1 is the number of samples;

[0027] Among them, the algorithm complexity of the time series creation unit is O(NLd);

[0028] Among them, for the support vector machine model, the standardization unit first flattens the 3D data into 2D and then standardizes it through StandardScaler. The specific formula is:

[0029]

[0030] In formula (3), μ j and σ j are the mean and standard deviation of the jth feature respectively, and the specific algorithms are:

[0031]

[0032] The purpose of the standardization unit is to eliminate the scale differences between different features so that all features contribute equally to the model;

[0033] The data enhancement unit is used to improve the robustness of the model. For time series data, the following algorithm is used:

[0034] A1. Time distortion:

[0035] x′ t =x t + δ(t) (6)

[0036] In formula (6), δ(t) is a smooth distortion function;

[0037] A2. Amplitude scaling:

[0038] x′ t =(1+α)x t (7)

[0039] In formula (7), α~N(0,0.1) is a small random disturbance;

[0040] A3, Additive Gaussian noise:

[0041] x′ t =x t +∈ t (8)

[0042] In formula (8), ∈ t ~N(0,σ2 ), σ is the noise intensity;

[0043] A1-A3 data augmentation is used to assist model learning and improve robust feature representation, thereby improving the generalization ability of the model.

[0044] As a further preferred embodiment of the present technical solution: the support vector machine model includes a linear support vector machine, a soft margin support vector machine, a kernel skill unit and a probability output unit of a support vector machine;

[0045] The linear support vector machine is used to process the linearly separable case, and the hyperplane w that satisfies the following conditions T x+b=0, and combined with the following algorithm:

[0046] y i (w T x i +b)≥1, i=1,...,N(9)

[0047] In formula (9), w is the weight vector, b is the bias term, and y is i ∈{-1,1} is the category label;

[0048] Among them, maximizing the interval is equivalent to solving the following optimization problem:

[0049]

[0050] Combining formula (9) and formula (10), it is a convex quadratic programming problem, which is transformed into a dual problem through Lagrangian duality. The specific formula is:

[0051]

[0052] In formula (11) and formula (12), α i is the Lagrange multiplier, and solving formula (11) and formula (12) gives the optimal solution to the original problem;

[0053] The soft margin support vector machine is used to handle the nonlinear separable situation, and the relaxation variable ξ is introduced i And penalty parameter C, the specific formula is:

[0054]

[0055] y i (w T x i +b)≥1-ξ i ,ξ i ≥0, i=1, ..., N(14)

[0056] In formula (13), parameter C controls the degree of penalty for misclassified samples. A larger C will lead to a stricter classification boundary, while a smaller C will allow more misclassifications.

[0057] The kernel technique unit adopts the radial basis function kernel, and the specific formula is:

[0058] K(x i , x j )=exp(-γx i -x j 2 )(15)

[0059] In formula (15), γ is the kernel parameter, which controls the width of the Gaussian function. A larger γ will lead to a more complex decision boundary. When running through the kernel trick unit, the decision function algorithm is converted to:

[0060]

[0061] Among them, the radial basis function kernel can handle nonlinear decision boundaries and has higher performance for high-dimensional data, and it has only one parameter γ that needs to be adjusted;

[0062] The support vector machine model in the probability output unit of the support vector machine is a judgment model, and the probability output is obtained by Platt scaling. The output f(x) of the support vector machine model is mapped to the interval [0, 1] by the sigmoid function. The specific formula is:

[0063]

[0064] In formula (17), A and B are parameters estimated by maximizing the following likelihood function;

[0065] In formula (18), is the converted label, and p i is the probability predicted by the model.

[0066] As a further preferred embodiment of the present technical solution: the long short-term memory network model includes a long short-term memory network unit, a long short-term memory network structure unit, a loss function unit, an optimization algorithm unit and a regularization technology unit;

[0067] The long short-term memory network unit is used to solve the gradient vanishing problem of the traditional recurrent neural network, and the long short-term memory network unit includes a memory unit c t and three gating mechanisms, the three gating mechanisms are input gate i t 、Forget gate t and output gate o t , the specific formulas are:

[0068] B1, input gate i t :

[0069] i t =σ(W i [h t-1 , x t ]+b i )(19)

[0070] B2, forget gate f t :

[0071] f t =σ(W f [h t-1 , x t ]+b f )(20)

[0072] B3, output gate o t :

[0073] o t =σ(W o [h t-1 , x t ]+b o )(twenty one)

[0074] B4. Candidate memory cells

[0075]

[0076] B5, memory unit update c t :

[0077]

[0078] B6, hidden state h t :

[0079] h t =o t Θtanh(c t )(twenty four)

[0080] In formula (20) and formula (21), σ is the sigmoid function, and in formula (23) and formula (24), Θ is element-wise multiplication;

[0081] Among them, the input gate i t Control the addition of new information, forget gate f t Allow the network to selectively forget past information, output gate o t Output used to determine the current state;

[0082] The long short-term memory network structure unit includes two long short-term memory network layers, and each of the two long short-term memory network layers is followed by a Dropout layer, and one is a Dense layer with L2 regularization;

[0083] The loss function unit adopts binary cross entropy, and the specific formula is:

[0084]

[0085] In formula (25), y i is the true label, The probability predicted by the model;

[0086] Among them, binary cross entropy is a natural choice for binary classification problems and is particularly sensitive to low-probability events. At the same time, its gradient properties are conducive to optimization.

[0087] The optimization algorithm unit uses the Adam optimizer to minimize the loss function. The Adam optimizer combines the advantages of the momentum method and RMSprop. The specific algorithm is:

[0088] C1, m t =β1m t-1 +(1-β1)g t (26)

[0089]

[0090] In formula (26) and formula (27), g t is the current gradient. In formula (26), m t is the first-order moment estimate of the gradient. In formula (27), υ t is the second-order moment estimate of the gradient, where β1 and β2 are decay rates in formula (26), formula (27), formula (28) and formula (29), and η is the learning rate in formula (30), and ∈ is a small constant to prevent division by zero;

[0091] The regularization technology unit adopts Dropout and L2 regularization, and the specific formula of L2 regularization is:

[0092]

[0093] In formula (31), λ=0.01 is the regularization coefficient, w j is the weight of the output layer;

[0094] Among them, L2 regularization prevents overfitting by penalizing large weight values, which encourages the model to learn smoother decision boundaries, thereby improving generalization ability.

[0095] As a further preferred embodiment of the present technical solution: the model evaluation index module includes a confusion matrix unit, an accuracy unit, a precision unit, a recall rate unit, an F1 score unit, a ROC curve and AUC unit and a logarithmic loss unit;

[0096] Among them, the specific formula of the confusion matrix unit is:

[0097]

[0098] In formula (32), TN is a true negative example, FP is a false positive example, FN is a false negative example, and TP is a true positive example;

[0099] The specific formula of the accuracy unit is:

[0100]

[0101] The accuracy unit is used to reflect the overall performance of the model, but it may be misleading in the case of class imbalance;

[0102] The specific formula of the accuracy unit is:

[0103]

[0104] The precision unit is used to reflect the proportion of true positive samples among all samples predicted to be positive;

[0105] The specific formula of the recall rate unit is:

[0106]

[0107] The recall unit is used to reflect the proportion of correctly predicted samples among all true positive samples;

[0108] The specific formula of the F1 score unit is:

[0109]

[0110] The F1 score unit is the harmonic mean of the precision unit and the recall unit, which provides a balanced performance metric to solve the problem of class imbalance;

[0111] Among them, the ROC curve and the ROC curve in the AUC unit plot the TPR and FPR under different thresholds. The specific formula is:

[0112]

[0113] The AUC in the ROC curve and AUC unit is the area under the ROC curve. The specific algorithm is:

[0114]

[0115] Its formula (39) uses a numerical integration method to calculate the approximate value of AUC, and n is the number of points on the ROC curve;

[0116] The advantage of AUC is that it is insensitive to class imbalance and provides the average performance of the model under all possible thresholds;

[0117] Among them, the specific formula of the logarithmic loss unit is:

[0118]

[0119] In formula (40), y i is the true label, p i is the probability of the positive class predicted by the model;

[0120] The logarithmic loss unit considers not only the correctness of the prediction but also the confidence of the prediction, so it provides a more fine-grained performance evaluation than the accuracy unit.

[0121] As a further preferred embodiment of the present technical solution: the integrated learning module includes a Bagging unit and a Boosting unit;

[0122] The bagging unit creates multiple subsets by sampling with replacement from the training set, and trains a basic model for each subset. Finally, the predictions of the basic models are combined by voting or averaging. The specific operation process is as follows:

[0123] D1. Create M bootstrap samples {B1, B2, ..., B M};

[0124] D2. In each sample B i Train a support vector machine model f on i ;

[0125] D3. The specific algorithm for final prediction through voting is:

[0126]

[0127] Among them, the Bagging unit effectively reduces variance and improves the generalization ability of the model;

[0128] The Boosting unit uses the AdaBoost algorithm to improve the performance of the long short-term memory network, that is, by adjusting the sample weights, the model focuses on samples that are difficult to classify. The specific operation process is as follows:

[0129] E1. Initialize sample weights

[0130] E2. For t=1, ..., T, then:

[0131] E2a, h of the LSTM network model trained using the current weights t ;

[0132] E2b. Calculate the weighted error rate:

[0133] E2c, calculate model weights:

[0134] E2d, update sample weight: w i ←w i exp(-α t y i h t (x i ))(44);

[0135] E2e, normalized weight;

[0136] E3. Final model:

[0137] As a further preferred embodiment of the present technical solution: the feature engineering module includes a statistical feature unit, a frequency domain feature unit and a wavelet feature unit;

[0138] The statistical feature unit is used to calculate the statistical features for each time window, and the specific algorithm is:

[0139] F1, mean μ:

[0140]

[0141] F2, standard deviation σ:

[0142]

[0143] F3, skewness s:

[0144]

[0145] F4, Kurtosis k:

[0146]

[0147] The F1-F4 statistical features can capture different aspects of data distribution and are used to identify abnormal patterns;

[0148] The frequency domain feature unit uses fast Fourier transform to convert the time domain signal into the frequency domain and extract the following features:

[0149] E1, main frequency: the frequency at which the signal energy is most concentrated;

[0150] E2, spectral entropy:

[0151]

[0152] In formula (50), p i is the normalized power spectral density;

[0153] E3, frequency band energy: the proportion of energy in a specific frequency band to the total energy;

[0154] The E1-E3 frequency domain features are used to capture periodic patterns and abnormal oscillations in battery signals;

[0155] The wavelet feature unit uses discrete wavelet transform to capture the features of the signal at different scales, and extracts the following features for each decomposition level:

[0156] F1, mean and variance of the approximation coefficient;

[0157] F2, mean and variance of detail coefficient;

[0158] F3, wavelet energy:

[0159]

[0160] In formula (51), d j，k is the detail coefficient of the jth layer;

[0161] Among them, the wavelet feature unit can effectively capture the multi-scale characteristics of the signal, and it can effectively detect local anomalies.

[0162] As a further preferred embodiment of the present technical solution: the model explanation module is used to improve the interpretability of the entire model, and the model explanation module includes a SHAP value unit and a LIME unit;

[0163] The SHAP value unit is based on the concept of Shapley value in game theory and is used to explain the contribution of each feature to model prediction. For feature i, the specific formula is:

[0164]

[0165] In formula (52), F is the set of all features, f s is a model that uses only feature subset S;

[0166] Among them, the SHAP value unit provides a consistent and locally accurate feature importance measure, which is used to understand the decision-making process of the model;

[0167] The LIME unit explains the prediction of the complex model by fitting a local interpretable model near the prediction point, and its objective function is:

[0168]

[0169] In formula (53), f is the original model, g is the interpretable model, and π (x) is the local weight, L is the loss function, and Ω is the complexity penalty term;

[0170] Among them, the LIME unit provides intuitive local explanations, which are used to understand the decision basis of the model on specific samples.

[0171] As a further preferred embodiment of the present technical solution: the model deployment and monitoring module includes a model serialization unit, an online learning unit and a model monitoring unit;

[0172] The model serialization unit uses TensorFlow's SavedModel format to serialize the long short-term memory network model and uses joblib to serialize the support vector machine model, thereby ensuring that the model can be consistently loaded and used in different environments;

[0173] The online learning unit is used to adapt to potential changes in data distribution, and a variant of stochastic gradient descent is used for incremental learning for the support vector machine model, while small batch updates are used to fine-tune model parameters for the long short-term memory network model;

[0174] The model monitoring unit is used to monitor indicators and track model performance. Specific monitoring indicators include:

[0175] G1, Prediction distribution shift: Use Kolmogorov-Smirnov test to detect and predict changes in distribution;

[0176] G2, feature distribution shift: Use Jensen-Shannon divergence to quantify the change in feature distribution;

[0177] G3, Model performance indicators: Regularly calculate the indicators in the model evaluation indicator module and set alarm thresholds;

[0178] Among them, the model monitoring unit is used to promptly detect problems with model performance degradation to ensure the reliability of the system.

[0179] Compared with the prior art, the present invention has the following beneficial effects:

[0180] The present invention achieves high-precision battery anomaly detection by combining advanced machine learning algorithms, deep learning technology and signal processing methods. The modular design of the system and the comprehensive evaluation framework ensure its reliability and scalability in practical applications. Through continuous optimization and monitoring, the system can adapt to changing data patterns, thereby providing strong support for battery quality control and performance optimization. Therefore, early and accurate monitoring and early warning of thermal runaway safety risks of batteries used in energy storage systems and electric vehicles can be carried out, and the spread of thermal runaway between batteries can be effectively prevented, which greatly reduces the probability and degree of loss of fire and explosion accidents caused by thermal runaway of batteries, thereby helping to improve the quality and efficiency of safety risk prevention and control in the new energy industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0181] Figure 1 This is a structural schematic diagram of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0182] Figure 2 This is an operation flow chart of an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0183] Figure 3 A schematic diagram of the architecture of an operating algorithm in an operating method of a monitoring and early warning device for identifying thermal runaway of a battery according to the present invention;

[0184] Figure 4 It is a schematic diagram of the architecture of a data preprocessing module in an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0185] Figure 5 A schematic diagram of the architecture of a support vector machine model in an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0186] Figure 6 A schematic diagram of the architecture of a long short-term memory network model in an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0187] Figure 7 It is a schematic diagram of the architecture of a model evaluation index module in an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0188] Figure 8 It is a schematic diagram of the architecture of an integrated learning module in an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0189] Fig. 9 It is a schematic diagram of the architecture of a feature engineering module in an operation method of a monitoring and early warning device for identifying battery thermal runaway according to the present invention;

[0190] Fig.10It is a schematic diagram of the architecture of a model interpretation module in an operation method of a monitoring and early warning device for identifying thermal runaway of a battery according to the present invention;

[0191] Fig.11 The present invention is a schematic diagram of the architecture of the model deployment and monitoring module in an operation method of a monitoring and early warning device for identifying battery thermal runaway. DETAILED DESCRIPTION

[0192] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0193] Example 1

[0194] See also Figure 1 The present invention provides a technical solution: a monitoring and early warning device for identifying thermal runaway of a battery, comprising a battery pack 1 and a pressure sensing assembly 2, the pressure sensing assembly 2 being electrically connected to a pressure collector 3 through a wire, the pressure collector 3 being electrically connected to a pressure monitoring and early warning module 4 through a wire, and the pressure monitoring and early warning module 4 being electrically connected to a computer group 5 through a wire;

[0195] The battery pack 1 includes a plurality of batteries, and the pressure sensing assembly 2 is installed on one side adjacent to two batteries;

[0196] The pressure sensing assembly 2 includes a pressure sensing film 21 and two flexible heat-insulating sheets 22 . The pressure sensing film 21 is installed on one side adjacent to the two flexible heat-insulating sheets 22 .

[0197] In this embodiment, specifically: the pressure sensing assembly 2 is provided with a pressure sensing film 21 with a pressure sensing area array row number n not greater than 200 and a column number m not greater than 100, and the pressure sensing assembly 2 is provided with a length and width consistent with that of each battery in the battery pack 1;

[0198] The pressure sensing component 2 is provided with a flexible heat-insulating sheet 22 with a thickness not greater than 5 mm, and the manufacturing material of the flexible heat-insulating sheet 22 includes but is not limited to aerogel felt or glass fiber wool.

[0199] In this embodiment, specifically: the pressure sensing component 2 and the pressure collector 3 are connected by an independent flat cable or a common flat cable, and the number of the flat cable is not less than 2 pins;

[0200] The pressure collector 3 and the pressure monitoring and early warning module 4 communicate with each other via CAN, RS485 or RS232 bus.

[0201] Example 2

[0202] See also Figure 2-Figure 11 , an operating method of a monitoring and early warning device for identifying battery thermal runaway, comprising the following steps:

[0203] S1, the pressure sensing film 21 in the pressure sensing assembly 2 senses the external pressure given by the battery and converts it into an electrical signal output, thereby achieving pressure measurement and monitoring;

[0204] S2. The two flexible heat-insulating sheets 22 are used to improve the heat dissipation efficiency and heat preservation effect of the pressure sensing film 21, and play the role of shock absorption, insulation and sealing for the pressure sensing film 21. At the same time, the gaps between the battery packs 1 can be filled;

[0205] S3, monitor and record pressure data in real time through the set pressure collector 3, so as to facilitate monitoring and management work;

[0206] S4, the set pressure monitoring and early warning module 4 analyzes and processes the collected information data by running an algorithm;

[0207] S5. Finally, the analysis and operation results are input into the set computer group 5.

[0208] In this embodiment, specifically: in S4, the running algorithm includes a data preprocessing module, a support vector machine model, a long short-term memory network model, a model evaluation index module, an integrated learning module, a feature engineering module, a model interpretation module, and a model deployment and monitoring module;

[0209] The data preprocessing module includes a time series creation unit, a standardization unit, and a data enhancement unit;

[0210] The time series creation unit converts the original information data into a time series of fixed length. Given an original data sequence X with a length of T = {x1, x2, ..., x T}, where each x t ∈R d is a d-dimensional feature vector, and a sliding window method is used to create a sequence sample of length L. The specific formula is:

[0211] S i ={x i , x i+1 , ..., x i+L-1}, i=1, 2,..., T-L+1(1)

[0212] This gives us a three-dimensional tensor:

[0213] x∈R N×L×d (2)

[0214] In formula (2), N = T-L + 1 is the number of samples;

[0215] Among them, the algorithm complexity of the time series creation unit is O(NLd), and according to the actual usage, the efficient operation of numpy can be used to reduce the calculation overhead;

[0216] Among them, for the support vector machine model, the standardization unit first flattens the 3D data into 2D and then standardizes it through StandardScaler. The specific formula is:

[0217]

[0218] In formula (3), μ j and σ j are the mean and standard deviation of the jth feature respectively, and the specific algorithms are:

[0219]

[0220] The purpose of the normalization unit is to eliminate the scale differences between different features so that all features contribute equally to the model, which has an important effect on distance-based algorithms.

[0221] Among them, the data enhancement unit is used to improve the robustness of the model. For time series data, the following algorithm is used:

[0222] A1. Time distortion:

[0223] x′ t =x t + δ(t) (6)

[0224] In formula (6), δ(t) is a smooth distortion function;

[0225] A2. Amplitude scaling:

[0226] x′ t =(1+α)x t (7)

[0227] In formula (7), α~N(0,0.1) is a small random disturbance;

[0228] A3, Additive Gaussian noise:

[0229] x′ t =x t +∈ t (8)

[0230] In formula (8), ∈ t ~N(0,σ 2 ), σ is the noise intensity;

[0231] A1-A3 data augmentation is used to assist model learning and improve robust feature representation, thereby improving the generalization ability of the model.

[0232] In this embodiment, specifically: the support vector machine model includes a linear support vector machine, a soft margin support vector machine, a kernel skill unit and a probability output unit of a support vector machine;

[0233] Among them, the linear support vector machine is used to deal with the linearly separable case, and the hyperplane w that satisfies the following conditions T x+b=0, and combined with the following algorithm:

[0234] y i (w T x i +b)≥1, i=1,...,N(9)

[0235] In formula (9), w is the weight vector, b is the bias term, and y is i ∈{-1,1} is the category label;

[0236] Among them, maximizing the interval is equivalent to solving the following optimization problem:

[0237]

[0238] Combining formula (9) and formula (10), it is a convex quadratic programming problem, which is transformed into a dual problem through Lagrangian duality. The specific formula is:

[0239]

[0240] In formula (11) and formula (12), α i is the Lagrange multiplier, and solving formula (11) and formula (12) gives the optimal solution to the original problem;

[0241] Among them, the soft margin support vector machine is used to deal with the nonlinear separable situation, and the relaxation variable ξ is introduced i And penalty parameter C, the specific formula is:

[0242]

[0243] y i (w T x i +b)≥1-ξ i ,ξ i ≥0, i=1, ..., N(14)

[0244] In formula (13), parameter C controls the degree of penalty for misclassified samples. A larger C will lead to a stricter classification boundary, while a smaller C will allow more misclassifications.

[0245] Among them, the kernel technique unit adopts the radial basis function kernel, and the specific formula is:

[0246] K(x i , x j )=exp(-γx i -x j 2 )(15)

[0247] In formula (15), γ is the kernel parameter, which controls the width of the Gaussian function. A larger γ will lead to a more complex decision boundary. When running through the kernel trick unit, the decision function algorithm is converted to:

[0248]

[0249] Among them, the radial basis function kernel can handle nonlinear decision boundaries and has higher performance for high-dimensional data. It only has one parameter γ that needs to be adjusted, so it is convenient and simple to use.

[0250] Among them, the support vector machine model in the probability output unit of the support vector machine is a judgment model. The probability output is obtained through Platt scaling, and the output f(x) of the support vector machine model is mapped to the [0, 1] interval through the sigmoid function. The specific formula is:

[0251]

[0252] In formula (17), A and B are parameters estimated by maximizing the following likelihood function;

[0253] In formula (18), is the converted label, and p i is the probability predicted by the model.

[0254] In this embodiment, specifically: the long short-term memory network model includes a long short-term memory network unit, a long short-term memory network structure unit, a loss function unit, an optimization algorithm unit and a regularization technology unit;

[0255] Among them, the long short-term memory network unit is used to solve the gradient disappearance problem of the traditional recurrent neural network, and the long short-term memory network unit includes a memory unit c t and three gating mechanisms, the three gating mechanisms are input gate i t 、Forget gate t and output gate o t , the specific formulas are:

[0256] B1, input gate i t :

[0257] i t=σ(W i [h t-1 , x t ]+b i )(19)

[0258] B2, forget gate f t :

[0259] f t =σ(W f [h t-1 , x t ]+b f )(20)

[0260] B3, output gate o t :

[0261] o t =σ(W o [h t-1 , x t ]+b o )(twenty one)

[0262] B4. Candidate memory cells

[0263]

[0264] B5, memory unit update c t :

[0265]

[0266] B6, hidden state h t :

[0267] h t =o t Θtanh(c t )(twenty four)

[0268] In formula (20) and formula (21), σ is the sigmoid function, and in formula (23) and formula (24), Θ is element-wise multiplication;

[0269] Among them, the input gate i t Control the addition of new information, forget gate f t Allow the network to selectively forget past information, output gate o t It is used to determine the output of the current state. Through the above operation, the long short-term memory network model can learn long-term dependencies;

[0270] The LSTM network structure unit includes two LSTM network layers, each of which is followed by a Dropout layer and a Dense layer with L2 regularization.

[0271] The specific settings are:

[0272] H1, LSTM (50units, activation='relu', return_sequences=True);

[0273] H2, Dropout (0.2);

[0274] H3, LSTM(50untis, activation='relu');

[0275] H4, Dropout(0.2);

[0276] H5, Dense(1, activation='sigmoid', kernel_regularizer=l2(0.01));

[0277] The above settings enable the two-layer LSTM network to allow the model to learn more complex time dependencies, and the Dropout layer prevents overfitting, the ReLU activation function accelerates the training process and effectively alleviates the problem of gradient disappearance, and the L2 regularized Dense layer further prevents overfitting.

[0278] Among them, the loss function unit adopts binary cross entropy, and the specific formula is:

[0279]

[0280] In formula (25), y i is the true label, The probability predicted by the model;

[0281] Among them, the binary cross entropy is a natural choice for binary classification problems and is particularly sensitive to low-probability events. At the same time, its gradient properties are conducive to optimization.

[0282] Among them, the optimization algorithm unit uses the Adam optimizer to minimize the loss function. The Adam optimizer combines the advantages of the momentum method and RMSprop. Its specific algorithm is:

[0283] C1, m t =β1m t-1 +(1-β1)g t (26)

[0284]

[0285] In formula (26) and formula (27), g t is the current gradient. In formula (26), mt is the first-order moment estimate of the gradient. In formula (27), υ t is the second-order moment estimate of the gradient, where β1 and β2 are decay rates in formula (26), formula (27), formula (28) and formula (29), and η is the learning rate in formula (30), and ∈ is a small constant to prevent division by zero;

[0286] Among them, the regularization technology unit adopts Dropout and L2 regularization. The specific formula of L2 regularization is:

[0287]

[0288] In formula (31), λ=0.01 is the regularization coefficient, w j is the weight of the output layer;

[0289] Among them, L2 regularization prevents overfitting by penalizing large weight values, which encourages the model to learn smoother decision boundaries, thereby improving generalization ability.

[0290] In this embodiment, specifically: the model evaluation index module includes a confusion matrix unit, an accuracy unit, a precision unit, a recall unit, an F1 score unit, a ROC curve and AUC unit, and a logarithmic loss unit;

[0291] Among them, the specific formula of the confusion matrix unit is:

[0292]

[0293] In formula (32), TN is a true negative example, FP is a false positive example, FN is a false negative example, and TP is a true positive example;

[0294] Among them, the specific formula of the accuracy unit is:

[0295]

[0296] The accuracy unit is used to reflect the overall performance of the model, but it may be misleading in the case of class imbalance;

[0297] Among them, the specific formula of the precision unit is:

[0298]

[0299] The precision unit is used to reflect the proportion of true positive samples among all samples predicted to be positive;

[0300] Among them, the specific formula of the recall rate unit is:

[0301]

[0302] The recall unit is used to reflect the proportion of correctly predicted samples among all true positive samples;

[0303] Among them, the specific formula of the F1 score unit is:

[0304]

[0305] The F1 score unit is the harmonic mean of the precision unit and the recall unit, which provides a balanced performance metric to solve the problem of class imbalance;

[0306] Among them, the ROC curve and the ROC curve in the AUC unit plot the TPR and FPR under different thresholds. The specific formula is:

[0307]

[0308] In formula (37), TPR is the true positive rate and FPR is the false positive rate;

[0309] ROC curve and AUC unit AUC is the area under the ROC curve. The specific algorithm is:

[0310]

[0311] Its formula (39) uses a numerical integration method to calculate the approximate value of AUC, and n is the number of points on the ROC curve;

[0312] The advantage of AUC is that it is insensitive to class imbalance and provides the average performance of the model under all possible thresholds;

[0313] Among them, the specific formula of the logarithmic loss unit is:

[0314]

[0315] In formula (40), y i is the true label, p i is the probability of the positive class predicted by the model;

[0316] The logarithmic loss unit considers not only the correctness of the prediction but also the confidence of the prediction, so it provides a more fine-grained performance evaluation than the accuracy unit.

[0317] In this embodiment, specifically: the integrated learning module includes a Bagging unit and a Boosting unit;

[0318] The bagging unit creates multiple subsets by sampling with replacement from the training set, and trains a basic model for each subset. Finally, the predictions of the basic models are combined by voting or averaging. The specific operation process is as follows:

[0319] D1. Create M bootstrap samples {B1, B2, ..., B M};

[0320] D2. In each sample B i Train a support vector machine model f on i ;

[0321] D3. The specific algorithm for final prediction through voting is:

[0322]

[0323] Among them, the Bagging unit effectively reduces variance and improves the generalization ability of the model;

[0324] Among them, the Boosting unit uses the AdaBoost algorithm to improve the performance of the long short-term memory network, that is, by adjusting the sample weights, the model focuses on samples that are difficult to classify. The specific operation process is as follows:

[0325] E1. Initialize sample weights

[0326] E2. For t=1, ..., T, then:

[0327] E2a, h of the LSTM network model trained using the current weights t ;

[0328] E2b. Calculate the weighted error rate:

[0329] E2c, calculate model weights:

[0330] E2d, update sample weight: w i ←w i exp(-α t y i h t (x i ))(44);

[0331] E2e, normalized weights

[0332] E3. Final model:

[0333] Among them, the Boosting unit combines multiple weak learners into a strong learner, effectively improving the performance of the model.

[0334] In this embodiment, specifically: the feature engineering module includes a statistical feature unit, a frequency domain feature unit and a wavelet feature unit, and the feature engineering module creates multiple derived features for capturing more complex patterns;

[0335] Among them, the statistical feature unit is used to calculate the statistical features of each time window. The specific algorithm is:

[0336] F1, mean μ:

[0337]

[0338] F2, standard deviation σ:

[0339]

[0340] F3, skewness s:

[0341]

[0342] F4, Kurtosis k:

[0343]

[0344] The F1-F4 statistical features can capture different aspects of data distribution and are used to identify abnormal patterns;

[0345] Among them, the frequency domain feature unit uses fast Fourier transform to convert the time domain signal to the frequency domain and extract the following features:

[0346] E1, main frequency: the frequency at which the signal energy is most concentrated;

[0347] E2, spectral entropy:

[0348]

[0349] In formula (50), p i is the normalized power spectral density;

[0350] E3, frequency band energy: the proportion of energy in a specific frequency band to the total energy;

[0351] The E1-E3 frequency domain features are used to capture periodic patterns and abnormal oscillations in battery signals;

[0352] Among them, the wavelet feature unit uses discrete wavelet transform to capture the characteristics of the signal at different scales, and for each decomposition level, the following features are extracted:

[0353] F1, mean and variance of the approximation coefficient;

[0354] F2, mean and variance of detail coefficient;

[0355] F3, wavelet energy:

[0356]

[0357] In formula (51), dj，k is the detail coefficient of the jth layer;

[0358] Among them, the wavelet feature unit can effectively capture the multi-scale characteristics of the signal, and it can effectively detect local anomalies.

[0359] In this embodiment, specifically: the model explanation module is used to improve the interpretability of the entire model, and the model explanation module includes a SHAP value unit and a LIME unit;

[0360] The SHAP value unit is based on the concept of Shapley value in game theory and is used to explain the contribution of each feature to model prediction. For feature i, the specific formula is:

[0361]

[0362] In formula (52), F is the set of all features, f s is a model that uses only feature subset S;

[0363] Among them, the SHAP value unit provides a consistent and locally accurate feature importance measure, which is used to understand the decision-making process of the model;

[0364] Among them, the LIME unit explains the prediction of the complex model by fitting a local interpretable model near the prediction point, and its objective function is:

[0365]

[0366] In formula (53), f is the original model, g is the interpretable model, and π (x) is the local weight, L is the loss function, and Ω is the complexity penalty term;

[0367] Among them, the LIME unit provides intuitive local explanations, which are used to understand the decision basis of the model on specific samples.

[0368] In this embodiment, specifically: the model deployment and monitoring module includes a model serialization unit, an online learning unit, and a model monitoring unit;

[0369] The model serialization unit uses TensorFlow's SavedModel format to serialize the long short-term memory network model and uses joblib to serialize the support vector machine model, thus ensuring that the model can be loaded and used consistently in different environments.

[0370] Among them, the online learning unit is used to adapt to potential changes in data distribution. For the support vector machine model, a variant of stochastic gradient descent is used for incremental learning, while for the long short-term memory network model, small batch updates are used to fine-tune the model parameters;

[0371] The model monitoring unit is used to monitor indicators and track model performance. Specific monitoring indicators include:

[0372] G1, Prediction distribution shift: Use Kolmogorov-Smirnov test to detect and predict changes in distribution;

[0373] G2, feature distribution shift: Use Jensen-Shannon divergence to quantify the change in feature distribution;

[0374] G3, Model performance indicators: Regularly calculate the indicators in the model evaluation indicator module and set alarm thresholds;

[0375] Among them, the model monitoring unit is used to promptly detect problems with model performance degradation to ensure the reliability of the system.

[0376] Working principle: When in use, the pressure sensing film 21 in the pressure sensing assembly 2 senses the external pressure given by the battery and converts it into an electrical signal output, thereby realizing the measurement and monitoring of the pressure, and the two flexible insulation sheets 22 are used to improve the heat dissipation efficiency and heat preservation effect of the pressure sensing film 21, and play the role of shock absorption, insulation and sealing for the pressure sensing film 21, and at the same time, can fill the gap between the battery packs 1;

[0377] The pressure collector 3 is used to monitor and record the pressure data in real time, so as to facilitate monitoring and management. The pressure monitoring and early warning module 4 is used to analyze and process the collected information data by running an algorithm, and finally, the analysis and operation results are input into the computer group 5.

[0378] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A monitoring and early warning device for identifying battery thermal runaway, comprising a battery pack (1) and a pressure sensing component (2), characterized in that: The pressure sensing component (2) is electrically connected to a pressure collector (3) via a wire, the pressure collector (3) is electrically connected to a pressure monitoring and early warning module (4) via a wire, and the pressure monitoring and early warning module (4) is electrically connected to a computer group (5) via a wire; The battery pack (1) comprises a plurality of batteries, and the pressure sensing assembly (2) is installed on one side adjacent to two batteries; The pressure sensing assembly (1) comprises a pressure sensing film (21) and two flexible thermal insulation sheets (22), and the pressure sensing film (21) is installed on one adjacent side of the two flexible thermal insulation sheets (22).

2. A method for operating a monitoring and early warning device for identifying battery thermal runaway, characterized in that: The following steps are involved: S1, sensing the external pressure given by the battery through the pressure sensing film (21) in the pressure sensing component (2), and converting it into an electrical signal output, thereby achieving pressure measurement and monitoring; S2. The two flexible heat-insulating sheets (22) are used to improve the heat dissipation efficiency and heat preservation effect of the pressure sensing film (21), and play the role of shock absorption, insulation and sealing for the pressure sensing film (21). At the same time, the gaps between the battery packs (1) can be filled; S3, real-time monitoring and recording of pressure data by means of a pressure collector (3) is provided, thereby facilitating monitoring and management work; S4, the set pressure monitoring and early warning module (4) analyzes and processes the collected information data by running an algorithm; S5. Finally, the analysis and operation results are input into the set computer group (5).

3. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 2 is characterized in that: In S4, the running algorithm includes data preprocessing module, support vector machine model, long short-term memory network model, model evaluation index module, ensemble learning module, feature engineering module, model interpretation module and model deployment and monitoring module; The data preprocessing module includes a time series creation unit, a standardization unit and a data enhancement unit; The time series creation unit converts the original information data into a time series of fixed length. Given an original data sequence X with a length of T = {x1, x2, ..., x T }, where each x t ∈R d is a d-dimensional feature vector, and a sliding window method is used to create a sequence sample of length L. The specific formula is: S i ={x i ,x i+1 ,...,x i+L-1 },i=1,2,...,T-L+1 (1) This gives us a three-dimensional tensor: x∈R N×L×d (2) In formula (2), N = T-L + 1 is the number of samples; Among them, for the support vector machine model, the standardization unit first flattens the 3D data into 2D and then standardizes it through StandardScaler. The specific formula is: In formula (3), μ j and σ j are the mean and standard deviation of the jth feature respectively, and the specific algorithms are: The purpose of the standardization unit is to eliminate the scale differences between different features so that all features contribute equally to the model; The data enhancement unit is used to improve the robustness of the model. For time series data, the following algorithm is used: A1. Time distortion: x′ t =x t +δ(t) (6) In formula (6), δ(t) is a smooth distortion function; A2. Amplitude scaling: x′ t =(1+α)x t (7) In formula (7), α~N(0,0.1) is a small random disturbance; A3, Additive Gaussian noise: x′ t =x t +∈ t (8) In formula (8), ∈ t ~N(0,σ 2 ), σ is the noise intensity; A1-A3 data augmentation is used to assist model learning and improve robust feature representation, thereby improving the generalization ability of the model.

4. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The support vector machine model includes a linear support vector machine, a soft margin support vector machine, a kernel skill unit and a probability output unit of a support vector machine; The linear support vector machine is used to process the linearly separable case, and the hyperplane w that satisfies the following conditions T x+b=0, and combined with the following algorithm: y i (w T x i +b)≥1,i=1,...,N (9) In formula (9), w is the weight vector, b is the bias term, and y is i ∈{-1,1} is the category label; Among them, maximizing the interval is equivalent to solving the following optimization problem: Combining formula (9) and formula (10), it is a convex quadratic programming problem, which is transformed into a dual problem through Lagrangian duality. The specific formula is: In formula (11) and formula (12), α i is the Lagrange multiplier, and solving formula (11) and formula (12) gives the optimal solution to the original problem; The soft margin support vector machine is used to handle the nonlinear separable situation, and the relaxation variable ξ is introduced i And penalty parameter C, the specific formula is: y i (w T x i +b)≥1-ξ i ,x i ≥0,i=1,...,N (14) In formula (13), parameter C controls the degree of penalty for misclassified samples. A larger C will lead to a stricter classification boundary, while a smaller C will allow more misclassifications. The kernel technique unit adopts the radial basis function kernel, and the specific formula is: K(x i ,x j )=exp(-γ||x i -x j || 2 ) (15) In formula (15), γ is the kernel parameter, which controls the width of the Gaussian function. A larger γ will lead to a more complex decision boundary. When running through the kernel trick unit, the decision function algorithm is converted to: The support vector machine model in the probability output unit of the support vector machine is a judgment model, and the probability output is obtained by Platt scaling. The output f(x) of the support vector machine model is mapped to the interval [0, 1] by the sigmoid function. The specific formula is: In formula (17), A and B are parameters estimated by maximizing the following likelihood function; In formula (18), is the converted label, and p i is the probability predicted by the model.

5. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The long short-term memory network model includes a long short-term memory network unit, a long short-term memory network structure unit, a loss function unit, an optimization algorithm unit and a regularization technology unit; The long short-term memory network unit is used to solve the gradient vanishing problem of the traditional recurrent neural network, and the long short-term memory network unit includes a memory unit c t and three gating mechanisms, the three gating mechanisms are input gate i t 、Forget gate t and output gate o t , the specific formulas are: B1, input gate i t : i t =σ(W i [h t-1 ,x t ]+b i )(19) B2, forget gate f t : f t =σ(W f [h t-1 ,x t ]+b f )(20) B3, output gate o t : the t =σ(W o [h t-1 ,x t ]+b o (21) B4. Candidate memory cells B5, memory unit update c t : B6, hidden state h t : h t =o t Θtanh(c t )(24) In formula (20) and formula (21), σ is the sigmoid function, and in formula (23) and formula (24), Θ is element-wise multiplication; The long short-term memory network structure unit includes two long short-term memory network layers, and each of the two long short-term memory network layers is followed by a Dropout layer, and one is a Dense layer with L2 regularization; The loss function unit adopts binary cross entropy, and the specific formula is: In formula (25), y i is the true label, The probability predicted by the model; The optimization algorithm unit uses the Adam optimizer to minimize the loss function. The Adam optimizer combines the advantages of the momentum method and RMSprop. The specific algorithm is: C1、m t =β1m t-1 +(1-β1)g t (26) C2、 C3、 C4、 C5、 In formula (26) and formula (27), g t is the current gradient. In formula (26), m t is the first-order moment estimate of the gradient. In formula (27), υ t is the second-order moment estimate of the gradient, where β1 and β2 are decay rates in formula (26), formula (27), formula (28) and formula (29), and η is the learning rate in formula (30), and ∈ is a small constant to prevent division by zero; The regularization technology unit adopts Dropout and L2 regularization, and the specific formula of L2 regularization is: In formula (31), λ=0.01 is the regularization coefficient, w j is the weight of the output layer.

6. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The model evaluation index module includes a confusion matrix unit, an accuracy unit, a precision unit, a recall rate unit, an F1 score unit, a ROC curve and AUC unit, and a logarithmic loss unit; Among them, the specific formula of the confusion matrix unit is: In formula (32), TN is a true negative example, FP is a false positive example, FN is a false negative example, and TP is a true positive example; The specific formula of the accuracy unit is: The specific formula of the accuracy unit is: The specific formula of the recall rate unit is: The specific formula of the F1 score unit is: Among them, the ROC curve and the ROC curve in the AUC unit plot the TPR and FPR under different thresholds. The specific formula is: The AUC in the ROC curve and AUC unit is the area under the ROC curve. The specific algorithm is: or Its formula (39) uses a numerical integration method to calculate the approximate value of AUC, and n is the number of points on the ROC curve; Among them, the specific formula of the logarithmic loss unit is: In formula (40), y i is the true label, p i is the probability of the positive class predicted by the model.

7. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The integrated learning module includes a Bagging unit and a Boosting unit; The bagging unit creates multiple subsets by sampling with replacement from the training set, and trains a basic model for each subset. Finally, the predictions of the basic models are combined by voting or averaging. The specific operation process is as follows: D1. Create M bootstrap samples {B1, B2, ..., B M }; D2. In each sample B i Train a support vector machine model f on i ; D3. The specific algorithm for final prediction through voting is: The Boosting unit uses the AdaBoost algorithm to improve the performance of the long short-term memory network, that is, by adjusting the sample weights, the model focuses on samples that are difficult to classify. The specific operation process is as follows: E1. Initialize sample weights E2. For t=1, ..., T, then: E2a, h of the LSTM network model trained using the current weights t ; E2b. Calculate the weighted error rate: E2c, calculate model weights: E2d, update sample weight: w i ←w i exp(-α t y i h t (x i ))(44); E2e, normalized weight; E3. Final model:

8. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The feature engineering module includes a statistical feature unit, a frequency domain feature unit and a wavelet feature unit; The statistical feature unit is used to calculate the statistical features for each time window, and the specific algorithm is: F1, mean μ: F2, standard deviation σ: F3, skewness s: F4, Kurtosis k: The frequency domain feature unit uses fast Fourier transform to convert the time domain signal into the frequency domain and extract the following features: E1, main frequency: the frequency at which the signal energy is most concentrated; E2, spectral entropy: In formula (50), p i is the normalized power spectral density; E3, frequency band energy: the proportion of energy in a specific frequency band to the total energy; The wavelet feature unit uses discrete wavelet transform to capture the features of the signal at different scales, and extracts the following features for each decomposition level: F1, mean and variance of the approximation coefficient; F2, mean and variance of detail coefficient; F3, wavelet energy: In formula (51), d j，k is the detail coefficient of the jth layer.

9. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The model explanation module is used to improve the interpretability of the entire model, and the model explanation module includes a SHAP value unit and a LIME unit; The SHAP value unit is based on the concept of Shapley value in game theory and is used to explain the contribution of each feature to model prediction. For feature i, the specific formula is: In formula (52), F is the set of all features, f s is a model that uses only feature subset S; The LIME unit explains the prediction of the complex model by fitting a local interpretable model near the prediction point, and its objective function is: In formula (53), f is the original model, g is the interpretable model, and π (x) is the local weight, L is the loss function, and Ω is the complexity penalty term.

10. The operating method of the monitoring and early warning device for identifying battery thermal runaway according to claim 3 is characterized in that: The model deployment and monitoring module includes a model serialization unit, an online learning unit and a model monitoring unit; The model serialization unit uses TensorFlow's SavedModel format to serialize the long short-term memory network model, and uses joblib to serialize the support vector machine model; The online learning unit is used to adapt to potential changes in data distribution, and a variant of stochastic gradient descent is used for incremental learning for the support vector machine model, while small batch updates are used to fine-tune model parameters for the long short-term memory network model; The model monitoring unit is used to monitor indicators and track model performance. Specific monitoring indicators include: G1, Prediction distribution shift: Use Kolmogorov-Smirnov test to detect and predict changes in distribution; G2, feature distribution shift: Use Jensen-Shannon divergence to quantify the change in feature distribution; G3. Model performance indicators: Regularly calculate the indicators in the model evaluation indicator module and set alarm thresholds.

Images