Lithium ion battery overcharge thermal runaway experiment device and experiment method

Abstract

The invention discloses a lithium ion battery overcharge thermal runaway experimental device and experimental method, and belongs to the field of lithium ion battery thermal runaway. The experimental device comprises a combustion cabin with an air exhaust system, an electric heating constant-temperature drying box arranged in the cabin, a battery testing system and a battery acquisition module, a lithium ion battery in the drying box is fixed through a support, and the battery testing system can control the battery to be overcharged to be in thermal runaway and acquire current, voltage and SOC data. The battery acquisition module monitors different surface temperatures of the battery through three thermocouples. The experimental method is based on the device, after sample data collection and cleaning, data set division and data standardization, a plurality of BP neural network models of different hidden layer structures and training devices are constructed, high-precision models are automatically screened through MAPE, RMSE and R2 regression evaluation indexes and set rules, accurate prediction of the lithium ion battery overcharge thermal runaway time interval is achieved, and the accuracy of the lithium ion battery overcharge thermal runaway time interval is improved. And reliable support is provided for battery safety early warning.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery thermal runaway, specifically relating to an experimental apparatus and method for lithium-ion battery overcharge thermal runaway. Background Technology

[0002] Lithium-ion batteries have become the mainstream power source in core fields such as new energy vehicles, energy storage systems, and portable electronic products due to their significant advantages such as high energy density, long cycle life, and low self-discharge rate.

[0003] However, lithium-ion batteries pose serious safety hazards under overcharge conditions, triggering a series of chain reactions: First, the positive electrode material is prone to irreversible crystal structure changes, leading to a decrease in lithium storage capacity; simultaneously, abnormal lithium metal precipitation occurs on the negative electrode surface, and this precipitated lithium gradually grows into dendrites. If the dendrites continue to grow and pierce the internal separator, it will directly cause an internal short circuit. Furthermore, the electrolyte undergoes violent decomposition under the high voltage environment generated by overcharging, releasing not only large amounts of flammable gases (such as carbon monoxide and methane) but also accompanied by a strong exothermic reaction. The combined effect of these side reactions causes a rapid increase in internal temperature and pressure within the battery. Without effective early warning and control measures, this will ultimately lead to thermal runaway, resulting in serious consequences such as battery fire and explosion.

[0004] To address this issue, current research utilizes backpropagation (BP) neural networks to analyze monitoring parameters of lithium batteries, enabling early warning of thermal runaway and predicting the time interval between thermal runaway events. For example, patent application number 202411314593.9, entitled "A Method for Recognizing Thermal Runaway Features of Lithium Batteries Based on BP Neural Networks," proposes using temperature data, gas release information, and related characteristic parameters during battery operation as input variables for the BP neural network. The network is trained by constructing a dataset (divided into training and testing sets) to ultimately obtain a thermal runaway feature recognition model.

[0005] However, this type of method has significant limitations, and the constructed model cannot be guaranteed to be the optimal model. On the one hand, the number of hidden layer units in a BP neural network directly affects the model's fitting and generalization abilities—too few units may lead to underfitting, failing to fully learn the data features; too many units can easily cause overfitting, reducing the model's adaptability to new data. On the other hand, different training methods (such as Bayesian regularization and conjugate gradient) have significantly different effects on the model's convergence speed, stability, and final prediction accuracy, making it impossible to determine the optimal model that achieves the highest prediction accuracy and stability. Summary of the Invention

[0006] This invention proposes an experimental apparatus and method for overcharge thermal runaway of lithium-ion batteries, which can construct a BP neural network model that accurately describes the thermal runaway time interval.

[0007] To achieve the above objectives, the present invention proposes the following technical content: An experimental apparatus for overcharge thermal runaway of lithium-ion batteries, comprising: Combustion chamber with exhaust system; An electric thermostatic drying oven is placed in the combustion chamber; a support is installed in the cavity of the electric thermostatic drying oven, and the support holds the lithium-ion battery used for thermal runaway experiments. The battery testing system is connected to the computer via electrical signals, and both the battery testing system and the computer are located outside the combustion chamber. The battery testing system's cable first passes through a perforation on the combustion chamber, and then connects to the lithium-ion battery through a through-hole on the top of the electric thermostatic drying oven. The battery testing system is controlled by the computer's control signals. It first charges the lithium battery to 100% SOC, and then overcharges it at a set rate until the lithium-ion battery experiences thermal runaway. It can also collect the current, voltage, and SOC data of the lithium-ion battery. The battery acquisition module is connected to the computer via electrical signals. The battery testing system is located outside the combustion chamber. The cable of the battery acquisition module first passes through another hole on the combustion chamber, and then through a through hole on the top of the electric thermostatic drying oven to connect to the lithium-ion battery. The battery acquisition module has three thermocouples, which are fixed to the center of the left, right, and large surfaces of the lithium-ion battery with heat-resistant insulating tape, respectively, to monitor the temperature of the corresponding surfaces of the lithium-ion battery during thermal runaway. The cables of the battery testing system and the battery acquisition module do not seal the through hole of the electric thermostatic drying oven, leaving a gap for the escape of thermal runaway gas.

[0008] Furthermore, the set multiplier includes 0.6C, 0.8C, or 1C.

[0009] Furthermore, the temperature of the electric heating constant temperature drying oven is 40°C or 50°C.

[0010] An experimental method for overcharge thermal runaway of lithium-ion batteries includes the following steps: S1: Construct an experimental setup for thermal runaway of a lithium-ion battery under overcharge conditions. Induce thermal runaway in the lithium-ion battery at a set charge / discharge rate. Assume that a total of [number] tests are conducted. m The overcharge thermal runaway experiment of the lithium battery monitored a total of G Sample data for each lithium-ion battery, including: current, voltage, ambient temperature, surface temperature, and thermal runaway time interval. T The surface temperature is the average of the temperatures monitored by three thermocouples. S2: Will GThe sample data was cleaned, erroneous sample data was deleted, and a total of [number] samples were set after cleaning. P One sample data; S3: Will P The sample data is divided into a training set and a test set; the training set is set as follows: n indivual; S4: Standardize the training set data; S5: Construct BP neural network models with different hidden layers and different trainers, and set the goal to construct X BP neural network models; S6: Will n The standardized data are input into X BP neural network models, which then predict the thermal runaway time interval. The process is repeated to train each BP neural network model. The MAPE, RMSE, and R-value of each BP neural network model are calculated. 2 Regression evaluation indicators; S7: Based on the set rules, automatically select a high-precision BP neural network model; after verifying each high-precision BP neural network model with a test set, select the best prediction model.

[0011] Furthermore, in step S5, The construction process of a BP neural network is as follows: If a single hidden layer is selected for the BP neural network model, the number of hidden layer neurons in the BP neural network can be selected as 12, 16 and 24 respectively. If the BP neural network model uses two hidden layers, the number of neurons in the first hidden layer can be 16 or 24, and the number of neurons in the second hidden layer can be 8 or 12. The choice of trainer is either Bayesian regularization or conjugate gradient; After using different combinations of hidden layers and different trainers, a total of X BP neural network models are constructed.

[0012] Further, step S4 includes the following steps: S4.1: Calculate the mean of all training set sample data. The formula is:

[0013] In the formula, x ji Indicates the first j The sample, the first i The true value of each feature; Indicates the number of training set samples. i The mean of each feature; n This represents the total number of samples in the training set; This represents the mean of all training set sample data, i.e. =[ , ; S4.2: Calculate the standard deviation of all training set sample data. ;

[0014] In the formula, Indicates the first i The standard deviation of each feature; The standard deviation of all training set sample data is represented by: =[ ]; S4.3: Standardize each feature data point using the following formula:

[0015] In the formula, z ji Indicates the first j The sample, the first i Standardization results on the first feature; the first j Each standardized feature data is represented as: z j =[ z j1 , z j2 , z j3 , z j4 ].

[0016] Further, step S6 includes the following steps: S6.1: The first j The standardized data is input into a BP neural network model, and the BP neural network is calculated. k The hidden layer neurons pair with the first j The input of standardized data is given by the following formula:

[0017] In the formula, net kj Indicates the first k The hidden layer neuron, for the first j Given a set of standardized data inputs, the number of neurons in this BP neural network is set to be... H , k ∈[1, H ], and is an integer; w ji Indicates the first k The hidden layer neuron, for the firstj In the standardized sample data, the first i The weights of each feature; z ji Indicates the first j The sample at the th i Standardization results on each feature; Indicates the first k The bias of each hidden neuron; S6.2: Calculate the output value of the activation function, using the following formula:

[0018] In the formula, h kj Indicates the first k The hidden layer neurons pair with the first j The activation function output value of the standardized data; S6.3: The output layer outputs the predicted value, with the following formula:

[0019] In the formula, Indicates the neural network for the first... j The predicted output for each sample; Indicates the output layer connection to the first k The weights of each hidden layer neuron; h kj Indicates the first k The hidden layer neurons pair with the first j The activation function output value of the standardized data; Indicates the bias of the output layer; S6.4: Calculate the MAPE, RMSE, and R for each BP neural network model. 2 Regression evaluation indicators; Coefficient of determination R 2 The formula is:

[0020] In the formula, Indicates the first j The true value of each sample; Indicates that the neural network for the first j Predicted values ​​for each sample; express n The average thermal runaway time interval of a real sample data set; The formula for Mean Absolute Percentage Error (MAPE) is:

[0021] In the formula, Indicates the first j The true value of each sample; Indicates that the neural network for the first j Predicted values ​​for each sample; n This represents the total number of samples in the training set; The formula for the root mean square error (RMSE) is:

[0022] In the formula, Indicates the first j The true value of each sample; Indicates that the neural network for the first j Predicted values ​​for each sample; n This represents the total number of samples in the training set.

[0023] Step S7 includes the following steps: S7.1: Normalize the above three error indices RMSE, MAPE, and R2 using the following formula:

[0024]

[0025]

[0026] in The value is 3. The value is 1%. The value is 0.01; Calculate the evaluation index F:

[0027] In the formula, Indicates weight; S7.2: This invention uses the value of the comprehensive index F as a unified criterion for distinguishing between high-precision and low-precision models. When the BP neural network model satisfies: If the model reaches a "high-precision prediction level", then the F-values ​​of all models are sorted from smallest to largest, and the model with the smallest F-value is selected as the final best BP neural network model for predicting the overcharge thermal runaway time of lithium-ion batteries.

[0028] The beneficial effects that can be achieved by adopting the above technologies are: 1. The device collects current, voltage, and SOC data through a battery testing system, monitors different surface temperatures of the battery with three thermocouples, and provides a stable ambient temperature in an electric heating constant temperature drying oven. It can obtain multi-dimensional, high-quality thermal runaway sample data, laying a reliable data foundation for subsequent model training. 2. The method constructs multiple BP neural network models with different hidden layer structures (single / double hidden layers, containing various combinations of the number of neurons) and trainers (Bayesian regularization, conjugate gradient), and combines multiple indicators such as MAPE, RMSE, and R² for screening. The optimal model is determined by minimizing RMSE, which ensures the accuracy of predicting the thermal runaway time interval and meets the requirements of high-precision early warning.

[0029] 3. This approach combines three core variables—hidden layer structure (single / double hidden layers), number of neurons (multiple combinations such as 12 / 16 / 24), and trainer type (Bayesian regularization / conjugate gradient)—into dozens of differentiated models. This multi-variable collaborative optimization logic is not the traditional "single-dimensional trial and error" (as mentioned in the existing blog post on lithium battery remaining life prediction based on BP neural networks), but rather constructs a complete model optimization strategy. Moreover, in conjunction with the three regression indicators MAPE, RMSE, and R², it solves the difficulty in determining the optimal model caused by the ambiguity of the selection criteria in existing technologies. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the experimental setup; Figure 2 This is a schematic diagram of thermocouple monitoring on the battery; Figure 3 This is a logical flowchart of the experimental method.

[0031] 1. Electric heating constant temperature drying oven; 2. Air extraction system. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1: An experimental apparatus for overcharge thermal runaway of lithium-ion batteries, used to provide relevant data on thermal runaway for an experimental method of overcharge thermal runaway of lithium-ion batteries. The experimental apparatus includes: The combustion chamber is equipped with an exhaust system (not shown in the figure). The exhaust system of the combustion chamber discharges the harmful gases after the thermal runaway combustion of the lithium-ion battery into the outside through the exhaust pipe (the rate is adjustable from 1 to 5 L / min). An electrically heated constant-temperature drying oven is placed inside the combustion chamber. A support frame is installed within the oven cavity, and the lithium-ion batteries are fixedly mounted on the frame. The oven provides a stable temperature for the lithium-ion batteries, such as 40°C or 50°C. A through-hole is located at the top of the oven, allowing any flue gas from thermal runaway to be discharged from the oven into the combustion chamber and then extracted by the aforementioned extraction system. This through-hole also allows cables from the subsequent battery testing system and battery data acquisition module to pass through the oven for data collection.

[0034] The battery testing system is electrically connected to a computer located outside the combustion chamber. The battery testing system's cable first passes through a pre-set opening in the combustion chamber, then connects to the lithium-ion battery through a through-hole in the top of the thermostatic drying oven. Controlled by the computer's control signals, the battery testing system performs overcharge thermal runaway experiments on the lithium-ion battery and collects data such as current, voltage, and SOC during the thermal runaway process caused by overcharging. The specific installation method is as follows: clamps on the positive and negative cables of the battery testing system are inserted through the through-holes into the internal cavity of the thermostatic drying oven and clamped to the positive and negative terminals of the lithium-ion battery, respectively, establishing an electrical connection. Because the computer is located outside the combustion chamber, it is protected from interference during thermal runaway. The overcharge thermal runaway experiment is as follows: Under the control of the computer's control signals, the battery testing system first charges the lithium battery to 100% SOC, and then overcharges it at set rates (0.6C, 0.8C, 1C) until the lithium-ion battery experiences thermal runaway.

[0035] The battery acquisition module is connected to the computer via electrical signals, and the battery testing system is also located outside the combustion chamber. The battery acquisition module's cable first passes through another pre-installed opening on the combustion chamber, and then connects to the lithium-ion battery through a through-hole at the top of the thermostatic drying oven. The battery acquisition module includes three thermocouples, which are fixed to the center of the left, right, and large central surfaces of the lithium-ion battery using heat-resistant insulating tape (e.g., [missing information]). Figure 2 As shown in the figure, the collected temperatures are denoted as T1, T3, and T2, respectively. Multiple thermocouples have a temperature measurement range of 0-1000℃ and an error of 2%, used to monitor the temperature of corresponding surfaces of the lithium-ion battery during thermal runaway. Since open flames only exist at the end electrodes of the lithium battery during thermal runaway, the heat-resistant insulating tape will not exceed its heating limit and detach from the thermocouples during the process, ensuring that the thermocouples can always provide monitoring data. The cables of the aforementioned battery testing system and battery acquisition module, after being inserted into the through-hole, do not completely seal the through-hole, leaving gaps for the escape of fumes generated during thermal runaway.

[0036] Example 2: As Figure 3As shown, an experimental method for overcharge thermal runaway of lithium-ion batteries includes the following steps: S1: Construct an experimental setup for overcharge thermal runaway of a lithium-ion battery. Induce thermal runaway of the lithium-ion battery at a set charging rate. Assume that one overcharge thermal runaway experiment is conducted. During the time period from the start of overcharge to the thermal runaway of the lithium-ion battery, the following parameters are monitored: G Sample data for each lithium-ion battery, each sample data includes: current. I ,Voltage U Ambient temperature T a Surface temperature T s Among them, surface temperature T s The meaning is the average temperature of the aforementioned three thermocouples.

[0037] It also monitors the corresponding thermal runaway time intervals. T Thermal runaway time interval T Represented as: the time from the start of overcharge T Time from 0 to the occurrence of thermal runaway T The time period of 1.

[0038] S2: Will G The sample data was cleaned, erroneous sample data was deleted, and a total of [number] samples were set after cleaning. P One sample data; S3: Will P The sample data is divided into a training set and a test set in an 8:2 ratio; the training set consists of... n indivual; S4: Standardize the training set data to eliminate differences in data units. This includes the following steps: S4.1: Calculate the mean of all training set sample data. The formula is:

[0039] Mode middle, x ji Indicates the first j The sample, the first i The true value of each feature, j ∈[1, n ], i ∈[1,4], and is an integer; Indicates the number of training set samples. i The mean of each feature; n This represents the total number of samples in the training set. This represents the mean of all training set sample data, i.e. =[ , .

[0040] S4.2: Calculate the standard deviation of all training set sample data. ;

[0041] In equation (2), Indicates the first i The standard deviation of each feature. The standard deviation of all training set sample data is represented by: =[ ]; x ji Indicates the first j The sample, the first i The true value of each feature; Indicates the first i The mean of each feature.

[0042] S4.3: Standardize each feature data point using the following formula: (3) In equation (3), z ji Indicates the first j The sample, the first i Standardization results on each feature; x ji Indicates the first j The sample, the first i The true value of each feature; Indicates the first i The mean of each feature; Indicates the first i The standard deviation of the first feature; j Each standardized feature data is represented as: z j =[ z j1 , z j2 , z j3 , z j4 ].

[0043] S5: Construct a candidate set X of BP neural network models, consisting of various network structures, neuron numbers, and training algorithms. By training each combined model, multiple BP neural networks with differentiated structures are obtained. Then, based on a pre-determined high-precision evaluation system, each model is screened, and the BP neural network model with the best predictive performance is finally determined. This process ensures that model optimization shifts from empirical trial and error to a systematic search. Specifically, this invention introduces a three-dimensional combination strategy of "structure-scale-training algorithm" in the construction of BP neural networks, forming a systematic model optimization process as follows: (1) Hidden layer structure dimension: When the network structure adopts a single hidden layer, three different neuron scales of 12, 16 and 24 are set to cover three types of network structures with light, medium and high capacity.

[0044] (2) Number of neurons: When the network structure adopts a double hidden layer, the first layer is set with 16 or 24 neurons and the second layer is set with 8 or 12 neurons, forming a variety of layer capacity combinations to enhance the model's ability to learn complex data features.

[0045] (3) Training Algorithm Dimension: The trainer is set to two types of training algorithms with different convergence mechanisms: Bayesian regularization (trainbr) and conjugate gradient (trainscg), to cover the typical training paths of probabilistic convergence and gradient optimization. Through the full combination of the above three dimensions, multiple BP neural network models with different structures, sizes, and training algorithms can be generated, forming a candidate model set X. This combination strategy covers the main factors affecting the performance of BP networks and avoids missing potential optimal structures.

[0046] The core of this solution lies in its systematic combination of three key dimensions—hidden layer structure, neuron size, and training algorithm—to upgrade the traditional single-dimensional parameter tuning method to a comprehensive and exhaustive model optimization mechanism. This mechanism enables the formation of structurally differentiated BP candidate models, fundamentally solving the problem of traditional BP networks being "reliant on experience and difficult to determine the optimal structure," thus improving the determinism and reproducibility of model selection.

[0047] S6: Input the n standardized data points sequentially into all BP candidate models in the model set X to obtain the thermal runaway time prediction results for each model. Then, based on a rigorous high-precision criterion composed of MAPE, RMSE, and R², screen all candidate models and select a subset of models that meet the high-precision requirements. Finally, validate this subset of models using a test set, and determine the optimal BP model using the minimum RMSE as the sole criterion, achieving high-precision prediction of thermal runaway time.

[0048] S6.1: Assume that the first... j The standardized data is input into a BP neural network model, and the BP neural network is calculated. k The hidden layer neurons pair with the first j The input of standardized data is given by the following formula:

[0049] In the formula, net kj Indicates the first k The hidden layer neuron, for the first j Given a set of standardized data inputs, the number of neurons in this BP neural network is set to be... H , k ∈[1, H ], and is an integer; w ji Indicates the first k The hidden layer neuron, for the first j In the standardized sample data, the first i The weights of each feature; z ji Indicates the first j The sample at the th i Standardization results on each feature; Indicates the first k The bias of each hidden neuron.

[0050] S6.2: The output value of the activation function tansig (hyperbolic tangent sigmoid function), the formula is:

[0051] In the formula, h kj Indicates the first k The hidden layer neurons pair with the first j The activation function output value of the standardized data; S6.3: Output layer outputs predicted values:

[0052] In the formula, Indicates the neural network for the first... j The predicted output of each sample (i.e., the predicted thermal runaway time interval). Indicates the output layer connection to the first k The weights of each hidden layer neuron; h kj Indicates the first k The hidden layer neurons pair with the first j The activation function output value of the standardized data; This indicates the bias of the output layer.

[0053] S6.4: Calculate the MAPE, RMSE, and R for each BP neural network model. 2 Regression evaluation indicators; Coefficient of determination R 2 The formula is:

[0054] In the formula, Indicates the first j The true value of the nth sample (i.e., the nth sample) j (the actual thermal runaway time interval for each sample); Indicates that the neural network for the first j The predicted value of the nth sample (i.e., the nth sample) j (Predicted thermal runaway time interval for each sample); express n The average thermal runaway time interval of a real sample data set; The formula for Mean Absolute Percentage Error (MAPE) is:

[0055] In the formula, Indicates the first j The true value of the nth sample (i.e., the nth sample) j (the actual thermal runaway time interval for each sample); Indicates that the neural network for the first j The predicted value of the nth sample (i.e., the nth sample) j (Predicted thermal runaway time interval for each sample); n This represents the total number of samples in the training set; The formula for root mean square error (RMSE) is:

[0056] In the formula, Indicates the first j The true value of the nth sample (i.e., the nth sample) j (the actual thermal runaway time interval for each sample); Indicates that the neural network for the first j The predicted value of the nth sample (i.e., the nth sample) j (Predicted thermal runaway time interval for each sample); n This represents the total number of samples in the training set; S7: To achieve a unified quantitative evaluation of the prediction performance of multiple BP neural network models, this invention proposes a comprehensive evaluation index construction method based on multi-index fusion. This method unifies, normalizes, and weights three evaluation indices—RMSE, MAPE, and R²—to form a single dimensionless comprehensive error index F, which is used to automatically select high-precision BP neural network models. Specifically, it includes the following steps: S7.1: Normalize the above three error indices, RMSE, MAPE, and R², to obtain the following formula:

[0057]

[0058]

[0059] in The value is 3. The value is 1%. The value is 0.01 ( ≥0.99). To achieve single-value quantification of the overall model performance, this invention uses a linear weighting method to fuse the three normalized error indices into a unified comprehensive evaluation index F:

[0060] The weights satisfy:

[0061] In this embodiment, unless otherwise specified, equal weights are used by default:

[0062] The weights can be adjusted according to specific application scenarios, and this invention does not impose any limitations on this. The resulting comprehensive evaluation index F is a dimensionless quantity, and the smaller the value, the better the overall predictive performance of the model.

[0063] S7.2: This invention uses the value of the comprehensive index F as a unified criterion for distinguishing between high-precision and low-precision models. When the BP neural network model satisfies:

[0064] The model is then judged to have reached a "high-precision prediction level". At this point, the model simultaneously possesses: 1) a small absolute error (RMSE); 2) a small relative error (MAPE); and 3) a high goodness of fit (…). Approximate 1) For all BP models that satisfy F≤1, further sort them by F value from smallest to largest, and select the one with the smallest F value as the final best BP neural network model for predicting the overcharge thermal runaway time of lithium-ion batteries.

[0065] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A device for testing thermal runaway of overcharged lithium-ion batteries, characterized in that, include: Combustion chamber with exhaust system; An electric thermostatic drying oven is placed inside the combustion chamber; The electric thermostatic drying oven is equipped with a support frame for placing lithium-ion batteries used in thermal runaway experiments. The battery testing system is connected to the computer via electrical signals. Both the battery testing system and the computer are located outside the combustion chamber. The cable of the battery testing system first passes through the perforation on the combustion chamber and then connects to the lithium-ion battery through the through hole on the top of the electric heating constant temperature drying oven. The battery testing system is controlled by the control signal of the computer. It first charges the lithium battery to 100% SOC and then overcharges it at a set rate until the lithium-ion battery experiences thermal runaway. It can also collect current, voltage, and SOC data of lithium-ion batteries; The battery acquisition module is connected to the computer via electrical signals. The battery testing system is located outside the combustion chamber. The cable of the battery acquisition module first passes through another hole on the combustion chamber, and then through a through hole on the top of the electric thermostatic drying oven to connect to the lithium-ion battery. The battery acquisition module has three thermocouples, which are fixed to the center of the left, right, and large surfaces of the lithium-ion battery with heat-resistant insulating tape, respectively, to monitor the temperature of the corresponding surfaces of the lithium-ion battery during thermal runaway. The cables of the battery testing system and the battery acquisition module do not seal the through hole of the electric thermostatic drying oven, leaving a gap for the escape of thermal runaway gas.

2. The device for testing thermal runaway of overcharged lithium ion battery according to claim 1, wherein, The set multiplier includes 0.6C, 0.8C, or 1C. 3.The lithium ion battery overcharge thermal runaway experiment device according to claim 1, characterized in that, The temperature of the electric thermostatic drying oven is 40℃ or 50℃.

4. A method for overcharge thermal runaway experiment of a lithium ion battery, based on the overcharge thermal runaway experiment device of claim 1 or 2, characterized in that, Includes the following steps: S1: build a kind of lithium ion battery overcharge thermal runaway experimental device, let lithium ion battery occur thermal runaway under the set charge-discharge rate, suppose a total of m times of lithium battery overcharge thermal runaway experiment, a total of G sample data of lithium ion battery are monitored, each sample data includes: current, voltage, ambient temperature, surface temperature, thermal runaway time interval T ; The surface temperature is the average of the temperatures monitored by the three thermocouples; S2: clean the sample data, delete the wrong sample data, set the total number of sample data after cleaning to be G P ​​ S3: divide the sample data into a training set and a test set; set the training set to P n ;​ S4: Standardize the training set data; S5: Construct BP neural network models with different hidden layers and different trainers, and set the goal to construct X BP neural network models; S6: input the standardized data into X BP neural network models respectively, predict the thermal runaway time interval by the X BP neural network models respectively, and continuously train each BP neural network model; n calculate the MAPE, RMSE, R 2 regression evaluation indexes of each BP neural network model respectively, S7: Based on the set rules, automatically select a high-precision BP neural network model; after verifying each high-precision BP neural network model with a test set, select the best prediction model.

5. The method of claim 4, wherein the method further comprises: In step S5, The construction process of a BP neural network is as follows: If a single hidden layer is selected for the BP neural network model, the number of hidden layer neurons in the BP neural network can be selected as 12, 16 and 24 respectively. If the BP neural network model uses two hidden layers, the number of neurons in the first hidden layer can be 16 or 24, and the number of neurons in the second hidden layer can be 8 or 12. The choice of trainer is either Bayesian regularization or conjugate gradient; After using different combinations of hidden layers and different trainers, a total of X BP neural network models are constructed.

6. The method of claim 4, wherein the method further comprises: Step S4 includes the following steps: S4.1: Calculate the mean of all training set sample data The formula is: ； In the formula, x ji represents the true value of the first j sample, the first i feature; represents the mean value of the first i feature of all training set samples; n represents the total number of training set samples; represents the mean value of all training set sample data, that is, [ , ; S4.2: Calculate the standard deviation of all training set sample data ; ； wherein the standard deviation of the first i characteristic; the standard deviation of all training set sample data, i.e. [ ]. S4.3: Standardize each feature data point using the following formula: ； In the formula, z ji denotes the standardized result on the j th sample, the i th feature; the standardized feature data of the j th feature is expressed as: z j [ z j1 , z j2 , z j3 , z j4 ].

7. The method of claim 6, wherein the method further comprises: Step S6 includes the following steps: S6.1: input the standardized data into a BP neural network model, calculate the input of the first hidden layer neuron of the BP neural network to the standardized data, the formula is as follows: j k j ​​​ ； In the formula, net kj Indicates the first k The nth hidden layer neuron, for the nth... j Given a set of standardized data inputs, the number of neurons in this BP neural network is set to be... H , k ∈[1, H ], and is an integer; w ji Indicates the first k The nth hidden layer neuron, for the nth... j In the standardized sample data, the first i The weights of each feature; z ji Indicates the first j The sample at the th i Standardization results on each feature; Indicates the first k The bias of each hidden neuron; S6.2: Calculate the output value of the activation function, using the following formula: ； In the formula, h kj denotes the activation function output value of the k hidden layer neuron on the j standardized data; S6.3: The output layer outputs the predicted value, with the following formula: ； In the formula, represents the prediction output of the neural network for the i-th sample; j represents the weight of the output layer connecting the i-th hidden layer neuron; k h kj represents the activation function output value of the i-th hidden layer neuron for the i-th normalized data; k j represents the bias quantity of the output layer;​​​​ S6.4: Calculate the MAPE, RMSE, R 2 regression evaluation index; Coefficient of determination R 2 The formula is: ； In the formula, represents the true value of the first j sample; represents the predicted value of the first j sample by the neural network; represents the average value of the thermal runaway time interval of the first n real sample data; The formula for Mean Absolute Percentage Error (MAPE) is: ； In the formula, Indicates the first j The true value of each sample; Indicates that the neural network for the first... j Predicted values ​​for each sample; n This represents the total number of samples in the training set; The formula for the root mean square error (RMSE) is: ； wherein represents the true value of the j th sample; represents the predicted value of the j th sample by the neural network; n represents the total number of samples in the training set.

8. The method of claim 7, wherein the method further comprises: Step S7 includes the following steps: S7.1: Normalize the above three error indices, RMSE, MAPE, and R², using the following formula: ； ； ； wherein the value of a is taken as 3, the value of a is taken as 1%, the value of a is taken as 0.01; Calculate the evaluation index F: ； In the formula, denotes the weight; S7.2: The application takes the value of the comprehensive index F as a unified criterion for distinguishing high-precision and low-precision models. When the BP neural network model satisfies: then it is determined that the model reaches the "high-precision prediction level". The F values of all models are sorted from small to large, and the one with the smallest F value is selected as the final best BP neural network model for predicting the thermal runaway time of lithium-ion batteries.

Images