Lithium battery health monitoring and early warning method based on multi-sensor data fusion

Abstract

The invention discloses a lithium battery health monitoring and early warning method based on multi-sensor data fusion, and the method comprises the steps: arranging a multi-measurement-point and multi-variable sensor in a lithium battery pack, collecting the data of a lithium battery in real time, carrying out the normalization processing, carrying out the feature selection through a random forest algorithm, and constructing a lithium battery thermal runaway health monitoring data set; a lithium battery thermal runaway health monitoring and early warning model MCCN-TimeMixer based on a multi-scale causal convolutional network MCCN and a TimeMixer model is constructed, local thermal state change features are extracted through the MCCN, and then a global relation between the features and a lithium battery thermal runaway health monitoring state is established through the TimeMixer; a hedera helix algorithm improved by a multivariate learning algorithm is adopted to optimize model hyper-parameters, finally error correction is performed through a fast deep stacked network extreme learning machine FDSNELM, and accurate prediction and early warning of the lithium battery thermal runaway health monitoring state are realized. The method can be applied to the modeling process of lithium battery thermal runaway health monitoring and early warning, and the accuracy of lithium battery thermal runaway health monitoring and early warning is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery thermal runaway health monitoring and early warning, and specifically relates to a lithium battery health monitoring and early warning method based on multi-sensor data fusion. Background Art

[0002] In recent years, in order to alleviate the energy crisis and environmental degradation, all sectors of society have begun to focus on the development of new energy. As an important representative, new energy vehicles have developed rapidly. The power source of new energy vehicles is mainly batteries, and the performance and status of batteries directly affect the safe and efficient operation of new energy vehicles. Among them, lithium batteries have been widely used in new energy vehicles and other fields due to their advantages such as high energy density and long life. However, there is a risk of thermal runaway in the use of lithium batteries, which may cause serious safety accidents. Therefore, it is very important to monitor and warn of thermal runaway of lithium batteries. At present, there are three main methods for studying the health monitoring of thermal runaway of lithium batteries: First, the traditional physical monitoring method. By monitoring the physical parameters of lithium batteries such as voltage, current, and temperature, it is determined whether it is in a thermal runaway state. However, this method can often only detect abnormalities when thermal runaway has occurred or is about to occur, and the timeliness of the warning is poor. Moreover, the monitoring of a single physical parameter is easily interfered by environmental factors and the accuracy is not high. Second, the model-based method. For example, the electrochemical model of lithium batteries describes the behavior of the battery by simulating the migration process of lithium ions from the positive electrode to the negative electrode and the accompanying electrochemical reactions and critical events, thereby predicting the possibility of thermal runaway. However, this method based on the electrochemical model is relatively complex, computationally intensive, and easily interfered by external dynamic factors, resulting in low accuracy and weak robustness. Third, data-driven algorithms. The hidden features related to thermal runaway of lithium batteries are discovered by processing sensory data through machine learning, deep learning and other methods. However, most data-driven models currently do not fully consider the temporal and spatial nature of lithium battery data. When faced with complex charging and discharging conditions, the prediction effect of the thermal runaway state of lithium batteries is not ideal. In addition, the existing thermal runaway health monitoring and early warning methods are not accurate enough in feature selection, and it is difficult to effectively capture the key features of thermal runaway health monitoring of lithium batteries. At the same time, the optimization of hyperparameters of the model is often not efficient enough, which affects the accuracy of monitoring and early warning. Moreover, the application scenarios of many methods are very limited, and online monitoring and early warning cannot be performed in actual application scenarios. They can only be monitored offline using complex equipment in laboratory environments.

[0003] In summary, in order to improve the accuracy, timeliness and reliability of lithium battery thermal runaway health monitoring and early warning, there is an urgent need for a new lithium battery health monitoring and early warning method based on multi-sensor data fusion, which can comprehensively and accurately extract the characteristics of lithium batteries, efficiently optimize model parameters, and realize real-time online monitoring and early warning in actual application scenarios. Summary of the invention

[0004] Purpose of the invention: In view of the problems of insufficient accuracy and inefficient hyperparameter optimization in existing lithium battery health monitoring and early warning methods, the present invention proposes a lithium battery health monitoring and early warning method based on multi-sensor data fusion to ensure the accuracy of lithium battery thermal runaway health monitoring and early warning.

[0005] Technical solution: The present invention proposes a lithium battery health monitoring and early warning method based on multi-sensor data fusion, comprising the following steps:

[0006] (1) Arrange multiple measuring points and multi-variable sensors inside the lithium battery pack to collect lithium battery data in real time;

[0007] (2) Normalize the lithium battery data, use the random forest algorithm for feature selection, and reconstruct the data to form a lithium battery thermal runaway health monitoring data set;

[0008] (3) Construct a lithium battery thermal runaway health monitoring and early warning model MCCN-TimeMixer based on the MCCN and TimeMixer models; extract the local thermal state change characteristics of the spatial parameter sequence composed of multi-dimensional parameter data of lithium batteries through the multi-scale causal convolutional network MCCN; input the extracted features into TimeMixer to establish the change connection between the spatial parameter sequence and the global lithium battery thermal runaway health monitoring state;

[0009] (4) Introducing a multivariate learning algorithm to optimize the propagation and evolution phases and the survivor selection phase of the Ivy algorithm, we obtain the optimized algorithm MIVYA. MIVYA is used to optimize the hyperparameters in the MCCN-TimeMixer model to obtain the corresponding optimal parameters.

[0010] (5) The battery status is predicted using the optimized lithium battery thermal runaway health monitoring and early warning model MCCN-TimeMixer. Finally, error correction is performed using the fast deep stacked network extreme learning machine FDSNELM to obtain the lithium battery thermal runaway health monitoring and early warning results.

[0011] Furthermore, the lithium battery data described in step (1) includes temperature, voltage, current, gas concentration, and pressure.

[0012] Furthermore, the implementation process of step (2) is as follows:

[0013] Convert each value of each feature to the interval [0,1] to obtain a data set containing different features;

[0014]

[0015] Where X is the original data point, Xmin is the minimum value of the feature in the data set, which is used to ensure that the minimum value after normalization is 0. max is the maximum value of the feature in the data set, which is used to ensure that the maximum value after normalization is 1. norm is the normalized data point;

[0016] The model is trained using the random forest algorithm to generate multiple decision trees. During training, each tree randomly selects samples from the data set and randomly selects features. For each split node in each tree, the reduction in the Gini coefficient caused by splitting each feature is calculated. The reduction in the Gini coefficient of each feature in all trees is accumulated to obtain the contribution of the feature to the model. The total contribution of each feature is divided by the total number of trees to obtain the average contribution of the feature. Based on the average contribution of the feature, a feature subset with high contribution is selected to form a new data set.

[0017] Furthermore, the reduction in the Gini coefficient is:

[0018]

[0019] Where M is the number of categories in node m, Gini m is the Gini coefficient of node m, is the weighted average Gini coefficient of the two child nodes after segmentation using feature j;

[0020] The average Gini coefficient reduction for each feature j across all trees in the random forest is:

[0021]

[0022] in, is the reduction in the Gini coefficient of feature j in the i-th tree, n is the number of trees in the random forest; the importance of the feature When performing feature selection, the higher the feature importance, the greater the contribution of the feature to the predictive ability of the model.

[0023] Furthermore, the multi-scale causal convolutional network MCCN in step (3) includes a multi-scale feature extraction module and a causal convolutional network.

[0024] Furthermore, the implementation process of extracting the local thermal state change characteristics of the spatial parameter sequence composed of multi-dimensional parameter data of the lithium battery through the multi-scale causal convolutional network MCCN in step (3) is as follows:

[0025] Construct a multi-scale feature extraction module and use convolution kernels of different sizes to extract features of different scales:

[0026] Feature m =Conv m(Input) (4)

[0027] Among them, Feature m Represents the features of the mth scale, Conv m represents the convolution operation of the mth scale, and the input is the multi-dimensional parameter data of the lithium battery;

[0028] Construct a causal convolutional network and use it to extract features from time series data to maintain the causality of time, which is achieved through the following formula:

[0029] S(i,j)=∑ m ∑ n A(im,jn)·K(m,n) (5)

[0030] Among them, A is the input data, K is the convolution kernel, and S is the feature map;

[0031] After the convolution, a bias is added and a nonlinear activation function is introduced to capture more complex data features:

[0032] z i,j =h(S(i,j)+b) (6)

[0033] Among them, b is the bias, h is the activation function, and z is the output of the convolutional layer;

[0034] The features extracted at different scales are fused by splicing or weighting to integrate information at different scales and improve the expressiveness of features:

[0035] FusedFeature=Fuse(Feature1,Feature2,…,Feature M ) (7)

[0036] Among them, FusedFeature is the fused feature, and Fuse represents the fusion operation.

[0037] Furthermore, in step (3), the extracted features are input into TimeMixer to establish the connection between the spatial parameter sequence and the change of the global lithium battery thermal runaway health monitoring state. The implementation process is as follows:

[0038] The TimeMixer model uses a multi-scale mixing architecture to process the changes in time series data at different time scales; it is achieved through multi-scale observations and past decomposable mixing:

[0039] Generate a collection of multi-scale observation sequences by averaging downsampling in

[0040] The PDM module is used to extract key information from the past and mix seasonal and trend components at different scales separately:

[0041]

[0042] Among them, S-Mix and T-Mix represent seasonal and trend mixes, respectively;

[0043] Use the future multi-predictor hybrid module FMM to predict future sequences:

[0044]

[0045] in, It is the multi-scale past information processed by the PDM module. is the final predicted future sequence; according to the prediction results of the TimeMixer model, the final prediction results are output.

[0046] Furthermore, the implementation process of introducing the multivariate learning algorithm to optimize the propagation and evolution phases and the survivor selection phase of the Ivy algorithm in step (4) is as follows:

[0047] At the beginning of the algorithm, the initial position of the IVYA population in the search space is randomly determined using the following formula:

[0048] I i =I min +rand(1,D)⊙(I max -I min ) (10)

[0049] Among them, I i is the position of the ith individual, I min and I max are the lower and upper bounds of the search space, respectively, and D is the number of dimensions;

[0050] Coordinated and orderly population growth, simulating the growth rate of ivy, is modeled by differential equations. The differential equation for growth rate is as follows:

[0051]

[0052] Among them, the vector and They represent the growth rates of the discrete-time system at time t and time (t+1), respectively. rand is a random real number in the interval [0,1]. N(1,D) represents a random vector of dimension D, which is a random number in the standard Gaussian normal distribution.

[0053] Get sunlight source growth, member I i The formulas for climbing and logical movement in the direction of the light source are as follows:

[0054]

[0055] Where ||N(1,D)|| is a vector whose components are the absolute values ​​of the components of vector N(1,D);

[0056] Propagation and evolution, after members roam globally in the search space to the nearest and most important neighbor, they try to directly follow the best member of the entire population, the formula is as follows:

[0057]

[0058] in, is the updated position;

[0059] In the propagation and evolution stages, a weight adjustment mechanism based on multiple intelligence is introduced:

[0060] w i =w i +α·(f best -f i ) (14)

[0061] Among them, w i is the weight of the ith individual, f best is the best fitness in the current population, f i is the fitness of the i-th individual, α is the learning rate;

[0062] In the survivor selection stage, multiple intelligence theory is used to evaluate and select individuals:

[0063]

[0064] Among them, P i is the probability of the i-th individual being selected, β is an adjustment parameter, and f avg is the average fitness of the population.

[0065] Furthermore, the implementation process of using MIVYA to optimize the hyperparameters in the MCCN-TimeMixer model in step (4) is as follows:

[0066] The hyperparameters of the MCCN-TimeMixer model include learning rate, batch size, and regularization. During the optimization process, an evaluation function is defined to evaluate the performance of the model under specific hyperparameter settings:

[0067] f(I i )=MCCN-TimeMixer(I i ) (16)

[0068] Among them, f(I i ) is the evaluation function, Ii is a combination of hyperparameters.

[0069] Furthermore, the implementation process of step (5) is as follows:

[0070] Calculate the error between the initial predicted value and the actual value of the MCCN-TimeMixer model:

[0071]

[0072] Among them, e is the error, y is the actual value, is the initial prediction value of the MCCN-TimeMixer model;

[0073] The error is corrected using the fast deep stacked network extreme learning machine FDSNELM model:

[0074]

[0075] in, is the error prediction value, H is the hidden layer output of the FDSNELM model, β is the output weight of the FDSNELM model, and g is the activation function of the FDSNELM model;

[0076] The FDSNELM model first calculates the hidden layer output H, and then calculates the error prediction value through the activation function g and the output weight β

[0077] Add the initial prediction value of the MCCN-TimeMixer model to the error prediction value of the FDSNELM model to get the final prediction value:

[0078]

[0079] in, is the final predicted value, is the initial prediction value of the MCCN-TimeMixer model, is the error prediction value of the FDSNELM model;

[0080] Output thermal runaway health monitoring warning based on the prediction results. The warning formula is as follows:

[0081] W=g(P) (20)

[0082] Among them, g represents the early warning logic function, and W represents the final early warning result.

[0083] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention constructs a lithium battery thermal runaway health monitoring and early warning model based on MCCN and TimeMixer models, giving full play to the ability of MCCN to extract the local thermal state change characteristics of the spatial parameter sequence, and the advantage of TimeMixer to establish the connection between the spatial parameter sequence and the global lithium battery thermal runaway health monitoring state change, to achieve comprehensive monitoring and early warning of the lithium battery thermal runaway health monitoring state, and improve the safety and reliability of lithium battery use; 2. In view of the shortcomings of traditional thermal runaway health monitoring methods such as low accuracy and limited adaptability, the present invention is a new model for thermal runaway health monitoring of lithium batteries. The improved Ivy algorithm (IVYA) is improved, and the improved multivariate learning is introduced into it to obtain the improved new algorithm MIVYA; using this algorithm to optimize the hyperparameters in the MCCN-TimeMixer model can improve the generalization ability and prediction accuracy of the model, ensuring that the thermal runaway state can be effectively monitored under different working conditions; 3. Error correction can reduce the deviation between the model prediction results and the actual thermal runaway state of the lithium battery; fine-tuning the model output through FDSNELM further improves the accuracy of the thermal runaway health monitoring and early warning results, reduces the risk of false alarms and missed alarms, and provides reliable protection for the safe operation of lithium batteries;

[0084] In summary, the present invention can be effectively applied to the modeling process of thermal runaway health monitoring and early warning of lithium battery, comprehensively considers the multi-dimensional parameter data of lithium batteries, realizes accurate prediction of thermal runaway health monitoring status, improves the prediction accuracy of the model, and safeguards the safe use of lithium batteries; in practical applications, it helps to timely discover thermal runaway risks, take corresponding prevention and control measures, and avoid the occurrence of safety accidents, which has important practical significance and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is a flow chart of the present invention;

[0086] Figure 2 A flow chart of feature selection performed by the random forest algorithm in the present invention;

[0087] Figure 3 This is a flow chart of error correction performed by FDSNELM in the present invention. DETAILED DESCRIPTION

[0088] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0089] like Figure 1 As shown, the present invention proposes a lithium battery health monitoring and early warning method based on multi-sensor data fusion, which specifically includes the following steps:

[0090] Step 1: Collect various data of lithium batteries in real time.

[0091] Multi-point and multi-variable sensors are arranged inside a lithium battery pack composed of a single lithium battery to collect various data of the lithium battery in real time, including temperature, voltage, current, gas concentration, pressure, etc.

[0092] Step 2: Normalize the data obtained in step (1), use the random forest algorithm for feature selection, reconstruct the data to form a lithium battery thermal runaway health monitoring data set, and divide it into training set, validation set and test set.

[0093] Step 2.1: Use the following formula to convert each value of each feature to the interval [0,1] to obtain a data set containing different features;

[0094]

[0095] Where X is the original data point, X min is the minimum value of the feature in the data set, which is used to ensure that the minimum value after normalization is 0. max is the maximum value of the feature in the data set, which is used to ensure that the maximum value after normalization is 1. norm are the normalized data points.

[0096] Step 2.2: If Figure 2 As shown, the random forest algorithm is used to train the model and generate multiple decision trees. During training, each tree randomly selects samples from the data set and randomly selects features. For each split node in each tree, the reduction in impurity (Gini coefficient) caused by splitting each feature is calculated. The reduction in the Gini coefficient of each feature in all trees is accumulated to obtain the contribution of the feature to the model. The total contribution of each feature is divided by the total number of trees to obtain the average contribution of the feature. According to the average contribution of the feature, the top 5 feature subsets with high contribution (large Gini coefficient reduction) are selected to form a new data set.

[0097] The calculation formula for the reduction of the Gini coefficient is:

[0098]

[0099] Where M is the number of categories in node m, Gini m is the Gini coefficient of node m, It is the weighted average Gini coefficient of the two child nodes after segmentation using feature j.

[0100] The average Gini coefficient reduction for each feature j across all trees in the random forest is:

[0101]

[0102] in, is the reduction in the Gini coefficient of feature j in the ith tree, and n is the number of trees in the random forest. It can be used for feature selection. The higher the feature importance, the greater the contribution of the feature to the predictive ability of the model.

[0103] Step 2.3: After feature selection using the random forest algorithm, the new dataset is divided into training set, validation set, and test set in a ratio of 6:2:2.

[0104] Step 3: Construct the MCCN-TimeMixer model for thermal runaway health monitoring and early warning of lithium batteries based on the MCCN and TimeMixer models; extract the local thermal state change characteristics of the spatial parameter sequence composed of multi-dimensional parameter data of lithium batteries through the multi-scale causal convolutional network MCCN. Input the extracted features into TimeMixer to establish the connection between the spatial parameter sequence and the change of the global thermal runaway health monitoring state of lithium batteries.

[0105] Step 3.1: Construct a multi-scale feature extraction module. In this step, we use convolution kernels of different sizes to extract features of different scales to capture the local thermal state change characteristics in the lithium battery parameter data. This multi-scale feature extraction can help the model capture the subtle differences in local thermal state changes at different scales, thereby more accurately understanding the complexity and diversity of the data.

[0106] Feature m =Conv m (Input) (4)

[0107] Among them, Feature m Represents the features of the mth scale, Conv m Represents the convolution operation of the mth scale, and the Input is the multi-dimensional parameter data of the lithium battery.

[0108] Step 3.2: Construct a causal convolutional network and use the causal convolutional neural network (CCN) to extract features from time series data and maintain the causality of time. CCN ensures that the model does not use future information in time series prediction by using causal convolution, which is particularly important for the prediction of thermal runaway health monitoring status of lithium batteries because it can avoid logical errors in prediction. This can be achieved by the following formula:

[0109] S(i,j)=∑ m ∑ n A(im,jn)·K(m,n) (5)

[0110] Among them, A is the input data, K is the convolution kernel, and S is the feature map.

[0111] Step 3.3: Introduce nonlinear activation function. After convolution, a bias is usually added and a nonlinear activation function is introduced to increase the expressive power of the model so that it can capture more complex data features. This can be achieved by the following formula:

[0112] z i,j =h(S(i,j)+b) (6)

[0113] Among them, b is the bias, h is the activation function, and z is the output of the convolutional layer.

[0114] Step 3.4: Feature fusion, which combines features extracted at different scales. This can be done by concatenation or weighting to integrate information at different scales and improve the expressiveness of features. This can be achieved using the following formula:

[0115] FusedFeature=Fuse(Feature1,Feature2,…,Feature M ) (7)

[0116] Among them, FusedFeature is the fused feature, and Fuse represents the fusion operation.

[0117] Step 3.5: Input the extracted features into the TimeMixer model. The TimeMixer model handles the variations of time series data at different time scales through a multi-scale mixing architecture. This can be achieved through multi-scale observations and past decomposable mixtures (PDM).

[0118] Multi-scale observations: Generate a set of multi-scale observation sequences by averaging downsampling in

[0119] Past Decomposable Mixture (PDM): The PDM module is used to extract key information from the past and mix seasonal and trend components at different scales separately. This can be achieved using the following formula:

[0120]

[0121] Among them, S-Mix and T-Mix represent seasonal and trend mixes respectively.

[0122] Step 3.6: Use the Future Multi-Predictor Mixture (FMM) module to predict future sequences. This can be achieved by the following formula:

[0123]

[0124] in, It is the multi-scale past information processed by the PDM module. is the final predicted future sequence.

[0125] Step 3.7: Based on the prediction results of the TimeMixer model and the specific conditions of the lithium battery, output the final prediction results.

[0126] The training set is input into the MCCN-TimeMixer model to build and train the machine learning model so that it can recognize specific features and patterns. The validation set is used to adjust model parameters and prevent overfitting during the training process to ensure the generalization ability of the model. The test set is used to evaluate the performance of the model on unseen data, providing key indicators such as accuracy, precision, and recall, thereby verifying the actual application effect of the model.

[0127] Step 4: Introduce the multivariate learning algorithm into the Ivy algorithm (IVYA) to improve the propagation and evolution stages and the survivor selection stage of the Ivy algorithm to obtain a new improved algorithm MIVYA. Use the improved algorithm to optimize the hyperparameters in the MCCN-TimeMixer model to obtain the corresponding optimal parameters.

[0128] Step 4.1: At the beginning of the algorithm, the initial position of the IVYA population in the search space is randomly determined using the following formula:

[0129] I i =I min +rand(1,D)⊙(I max -I min ) (10)

[0130] Among them, I i is the position of the ith individual, I min and I max are the lower and upper bounds of the search space respectively, and D is the number of dimensions.

[0131] Step 4.2: Coordinated and orderly population growth, simulate the growth rate of ivy, and model it through differential equations. The differential equation for growth rate is as follows:

[0132]

[0133] Among them, the vector and represents the growth rate of a discrete-time system (time t and time t+1), rand is a random real number in the interval [0,1], N(1,D) represents a random vector of dimension D, which is a random number from a standard Gaussian (normal) distribution.

[0134] Step 4.3: Get sunlight source to grow, member I i The formulas for climbing and logical movement in the direction of the light source are as follows:

[0135]

[0136] where ||N(1,D)|| is a vector whose components are the absolute values ​​of the components of vector N(1,D).

[0137] Step 4.3: Propagation and evolution, after members roam globally in the search space to the nearest and most important neighbor, they try to directly follow the best member of the entire population, the formula is as follows:

[0138]

[0139] in, is the updated location.

[0140] In the propagation and evolution stages, a weight adjustment mechanism based on multiple intelligence can be introduced:

[0141] w i =w i +α·(f best -f i ) (14)

[0142] Among them, w i is the weight of the ith individual, f best is the best fitness in the current population, f i is the fitness of the ith individual, and α is the learning rate.

[0143] Step 4.4: In the survivor selection stage, the theory of multiple intelligences can be combined to evaluate and select individuals:

[0144]

[0145] Among them, P i is the probability of the i-th individual being selected, β is an adjustment parameter, and f avg is the average fitness of the population.

[0146] Step 4.5: Use the improved IVYA to optimize the hyperparameters of the MCCN-TimeMixer model. This includes learning rate, batch size, regularization parameters, etc. During the optimization process, we need to define an evaluation function to evaluate the performance of the model under specific hyperparameter settings:

[0147] f(I i )=MCCN-TimeMixer(I i ) (16)

[0148] Among them, f(I i ) is the evaluation function, I i is a combination of hyperparameters.

[0149] Step 5: Use the optimized lithium battery thermal runaway health monitoring and early warning model MCCN-TimeMixer to predict the battery status, and finally perform error correction through the fast deep stacked network extreme learning machine FDSNELM to obtain the lithium battery thermal runaway health monitoring and early warning results, such as Figure 3 shown.

[0150] Step 5.1: Calculate the error, calculate the error between the initial prediction of the MCCN-TimeMixer model and the actual value. This step is the basis of the error correction process because it quantifies the degree of inaccuracy of the model's prediction. This can be achieved by the following formula:

[0151]

[0152] Among them, e is the error, y is the actual value, is the initial prediction value of the MCCN-TimeMixer model.

[0153] The error e is calculated by subtracting the initial predicted value from the actual value y This error value will be used to guide the FDSNELM model to perform error correction.

[0154] Step 5.2: Error correction using FDSNELM model We use the Fast Deep Stacked Network Extreme Learning Machine (FDSNELM) model to correct the errors. The FDSNELM model is specifically designed to correct the prediction errors of other models and improve the accuracy of the overall prediction. This can be achieved using the following formula:

[0155]

[0156] in, is the error prediction value, H is the hidden layer output of the FDSNELM model, β is the output weight of the FDSNELM model, and g is the activation function of the FDSNELM model.

[0157] The FDSNELM model first calculates the hidden layer output H, and then calculates the error prediction value through the activation function g and the output weight β This error prediction value is the correction result of the initial error e by the FDSNELM model.

[0158] Step 5.3: Calculate the final prediction value. We add the initial prediction value of the MCCN-TimeMixer model to the error prediction value of the FDSNELM model to get the final prediction value. This step combines the advantages of the two models to improve the accuracy of the prediction.

[0159]

[0160] in, is the final predicted value, is the initial prediction value of the MCCN-TimeMixer model, is the error prediction value of the FDSNELM model.

[0161] Final prediction value is to take the initial predicted value and the predicted error This method uses the error correction capability of the FDSNELM model to reduce the overall prediction error.

[0162] Step 5.4: Output thermal runaway health monitoring warning based on the prediction results. The warning formula is as follows:

[0163] W=g(P) (20)

[0164] Among them, g represents the early warning logic function, and W represents the final early warning result.

[0165] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A lithium battery health monitoring and early warning method based on multi-sensor data fusion, characterized in that: The following steps are involved: (1) Arrange multiple measuring points and multi-variable sensors inside the lithium battery pack to collect lithium battery data in real time; (2) Normalize the lithium battery data, use the random forest algorithm for feature selection, and reconstruct the data to form a lithium battery thermal runaway health monitoring data set; (3) Construct a lithium battery thermal runaway health monitoring and early warning model MCCN-TimeMixer based on the MCCN and TimeMixer models; extract the local thermal state change characteristics of the spatial parameter sequence composed of multi-dimensional parameter data of lithium batteries through the multi-scale causal convolutional network MCCN; input the extracted features into TimeMixer to establish the change connection between the spatial parameter sequence and the global lithium battery thermal runaway health monitoring state; (4) Introducing a multivariate learning algorithm to optimize the propagation and evolution phases and the survivor selection phase of the Ivy algorithm, we obtain the optimized algorithm MIVYA. MIVYA is used to optimize the hyperparameters in the MCCN-TimeMixer model to obtain the corresponding optimal parameters. (5) The battery status is predicted using the optimized lithium battery thermal runaway health monitoring and early warning model MCCN-TimeMixer. Finally, error correction is performed using the fast deep stacked network extreme learning machine FDSNELM to obtain the lithium battery thermal runaway health monitoring and early warning results.

2. A lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1, characterized in that: The lithium battery data described in step (1) includes temperature, voltage, current, gas concentration, and pressure.

3. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: The implementation process of step (2) is as follows: Convert each value of each feature to the interval [0,1] to obtain a data set containing different features; Where X is the original data point, X min is the minimum value of the feature in the data set, which is used to ensure that the minimum value after normalization is 0. max is the maximum value of the feature in the data set, which is used to ensure that the maximum value after normalization is 1. norm is the normalized data point; The model is trained using the random forest algorithm to generate multiple decision trees. During training, each tree randomly selects samples from the data set and randomly selects features. For each split node in each tree, the reduction in the Gini coefficient caused by splitting each feature is calculated. The reduction in the Gini coefficient of each feature in all trees is accumulated to obtain the contribution of the feature to the model. The total contribution of each feature is divided by the total number of trees to obtain the average contribution of the feature. Based on the average contribution of the feature, a feature subset with high contribution is selected to form a new data set.

4. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: The reduction in the Gini coefficient is: Where M is the number of categories in node m, Gini m is the Gini coefficient of node m, is the weighted average Gini coefficient of the two child nodes after segmentation using feature j; The average Gini coefficient reduction for each feature j across all trees in the random forest is: in, is the reduction in the Gini coefficient of feature j in the i-th tree, n is the number of trees in the random forest; the importance of the feature When performing feature selection, the higher the feature importance, the greater the contribution of the feature to the predictive ability of the model.

5. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: The multi-scale causal convolutional network MCCN in step (3) includes a multi-scale feature extraction module and a causal convolutional network.

6. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: The implementation process of extracting the local thermal state change characteristics of the spatial parameter sequence composed of multi-dimensional parameter data of the lithium battery through the multi-scale causal convolutional network MCCN in step (3) is as follows: Construct a multi-scale feature extraction module and use convolution kernels of different sizes to extract features of different scales: Feature m =Conv m (Input) (4) Among them, Feature m Represents the features of the mth scale, Conv m represents the convolution operation of the mth scale, and the input is the multi-dimensional parameter data of the lithium battery; Construct a causal convolutional network and use it to extract features from time series data to maintain the causality of time, which is achieved through the following formula: S(i,j)=∑ m ∑ n A(i-m,j-n)·K(m,n) (5) Among them, A is the input data, K is the convolution kernel, and S is the feature map; After the convolution, a bias is added and a nonlinear activation function is introduced to capture more complex data features: z i,j =h(S(i,j)+b) (6) Among them, b is the bias, h is the activation function, and z is the output of the convolutional layer; The features extracted at different scales are fused by splicing or weighting to integrate information at different scales and improve the expressiveness of features: FusedFeature=Fuse(Feature1,Feature2,…,Feature M ) (7) Among them, FusedFeature is the fused feature, and Fuse represents the fusion operation.

7. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: In step (3), the extracted features are input into TimeMixer to establish the connection between the spatial parameter sequence and the change of the global lithium battery thermal runaway health monitoring state. The implementation process is as follows: The TimeMixer model uses a multi-scale mixing architecture to process the changes in time series data at different time scales; it is achieved through multi-scale observations and past decomposable mixing: Generate a collection of multi-scale observation sequences by averaging downsampling in The PDM module is used to extract key information from the past and mix seasonal and trend components at different scales separately: Among them, S-Mix and T-Mix represent seasonal and trend mixes, respectively; Use the future multi-predictor hybrid module FMM to predict future sequences: in, It is the multi-scale past information processed by the PDM module. is the final predicted future sequence; according to the prediction results of the TimeMixer model, the final prediction results are output.

8. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: The implementation process of introducing the multivariate learning algorithm to optimize the propagation and evolution phases and the survivor selection phase of the Ivy Algorithm in step (4) is as follows: At the beginning of the algorithm, the initial position of the IVYA population in the search space is randomly determined using the following formula: I i =I min +rand(1,D)⊙(I max -I min ) (10) Among them, I i is the position of the ith individual, I min and I max are the lower and upper bounds of the search space, respectively, and D is the number of dimensions; Coordinated and orderly population growth, simulating the growth rate of ivy, is modeled by differential equations. The differential equation for growth rate is as follows: Among them, the vector and They represent the growth rates of the discrete-time system at time t and time (t+1), respectively. rand is a random real number in the interval [0,1]. N(1,D) represents a random vector of dimension D, which is a random number in the standard Gaussian normal distribution. Get sunlight source growth, member I i The formulas for climbing and logical movement in the direction of the light source are as follows: Where ||N(1,D)|| is a vector whose components are the absolute values ​​of the components of vector N(1,D); Propagation and evolution, after members roam globally in the search space to the nearest and most important neighbor, they try to directly follow the best member of the entire population, the formula is as follows: in, is the updated position; In the propagation and evolution stages, a weight adjustment mechanism based on multiple intelligence is introduced: w i =w i +α·(f best -f i ) (14) Among them, w i is the weight of the ith individual, f best is the best fitness in the current population, f i is the fitness of the i-th individual, α is the learning rate; In the survivor selection stage, multiple intelligence theory is used to evaluate and select individuals: Among them, P i is the probability of the i-th individual being selected, β is an adjustment parameter, and f avg is the average fitness of the population.

9. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1 is characterized in that: The implementation process of using MIVYA to optimize the hyperparameters in the MCCN-TimeMixer model in step (4) is as follows: The hyperparameters of the MCCN-TimeMixer model include learning rate, batch size, and regularization. During the optimization process, an evaluation function is defined to evaluate the performance of the model under specific hyperparameter settings: f(I i )=MCCN-TimeMixer(I i ) (16) Among them, f(I i ) is the evaluation function, I i is a combination of hyperparameters.

10. The lithium battery health monitoring and early warning method based on multi-sensor data fusion according to claim 1, characterized in that: The implementation process of step (5) is as follows: Calculate the error between the initial predicted value and the actual value of the MCCN-TimeMixer model: Among them, e is the error, y is the actual value, is the initial prediction value of the MCCN-TimeMixer model; The error is corrected using the fast deep stacked network extreme learning machine FDSNELM model: in, is the error prediction value, H is the hidden layer output of the FDSNELM model, β is the output weight of the FDSNELM model, and g is the activation function of the FDSNELM model; The FDSNELM model first calculates the hidden layer output H, and then calculates the error prediction value through the activation function g and the output weight β Add the initial prediction value of the MCCN-TimeMixer model to the error prediction value of the FDSNELM model to get the final prediction value: in, is the final predicted value, is the initial prediction value of the MCCN-TimeMixer model, is the error prediction value of the FDSNELM model; Output thermal runaway health monitoring warning based on the prediction results. The warning formula is as follows: W=g(P) (20) Among them, g represents the early warning logic function, and W represents the final early warning result.

Images