Data-driven power battery fault early warning and safety risk assessment method

Abstract

The application discloses a kind of based on data driving power battery fault early warning and safety risk assessment method, comprising: data acquisition;The data collected are handled;The data after processing are divided, and the correlation analysis of power battery parameter is carried out, and data set is made;Data set is divided into training set and test set, so that NARX neural network structure model is trained on training set, so that NARX neural network structure model is tested on test set, according to the performance of test set, the hyperparameter adjustment of NARX neural network structure model is carried out, and voltage fault early warning model is established;Temperature fault early warning model is established according to probe temperature;According to voltage fault early warning model and temperature fault early warning model, establish power battery safety risk assessment system.The power battery fault early warning and safety risk assessment method disclosed in the application has the advantages of small amount of calculation, fast calculation speed, early fault early warning, can be more accurately early warning and evaluation to power battery and the like.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle power battery safety, and in particular to a data-driven method for power battery fault early warning and safety risk assessment. Background Technology

[0002] With the booming development of the electric vehicle industry, a series of problems have become increasingly prominent, especially battery safety issues, such as fire accidents.

[0003] For example, lithium-ion batteries are currently the mainstream power source for electric vehicles. They are characterized by highly nonlinear capacity and lifespan, aging characteristics spanning multiple time and spatial scales, and the inability to directly measure state of charge and state of health, making fault diagnosis and early warning extremely difficult. On-board lithium-ion power batteries are complex electrochemical systems with intricate failure mechanisms and numerous fault types. Furthermore, their harsh operating conditions and environments significantly increase the probability of failure. Therefore, accurate fault warning and safety risk assessment of power batteries are crucial for improving vehicle safety and reducing property damage. Summary of the Invention

[0004] In view of the difficulties and shortcomings of the existing technology, the purpose of this invention is to provide a data-driven method for power battery fault early warning and safety risk assessment.

[0005] This application discloses a data-driven method for power battery fault early warning and safety risk assessment, including:

[0006] S1: Data acquisition, including acquisition time, vehicle status, charging status, operating mode, vehicle speed, total voltage, total current, cumulative mileage, SOC, highest voltage battery subsystem number, highest single-cell voltage value, highest voltage battery single-cell code, lowest voltage battery single-cell code, lowest voltage battery subsystem number, lowest single-cell voltage value, highest temperature value, highest temperature subsystem number, highest temperature probe number, highest alarm level, probe temperature, and general alarm flags;

[0007] S2: Process the data collected by S1;

[0008] S3: Based on the vehicle status and battery status, divide the data processed in S2, perform correlation analysis on the power battery parameters, and create a dataset based on the results of the correlation analysis.

[0009] S4: Divide the dataset into a training set and a test set, iteratively train the NARX neural network structure model on the training set, and test the NARX neural network structure model on the test set. Based on the performance on the test set, adjust the hyperparameters of the NARX neural network structure model to evaluate the model and establish a voltage fault early warning model.

[0010] S5: Establish a temperature fault early warning model based on information entropy according to the probe temperature;

[0011] S6: Establish a power battery safety risk assessment system based on voltage fault early warning model and temperature fault early warning model.

[0012] This application presents a power battery fault early warning and safety risk assessment method comprising S1 to S6, which features low computational load, fast calculation speed, and relatively accurate early warning and assessment of power batteries, showing promising application prospects. The NARX neural network structure model exhibits good dynamics and anti-interference capabilities for power batteries, enabling high-precision and relatively accurate prediction of battery voltage. Furthermore, the NARX neural network structure model for battery voltage prediction does not require high precision in data acquisition, allowing the use of low-frequency data and effectively reducing data storage costs.

[0013] As some preferred embodiments of this application, S2 is further selectively made to include:

[0014] S21: Arrange the data collected in S1 in chronological order;

[0015] S22: Clean and complete the data arranged in chronological order;

[0016] S23: Perform data separation of individual power battery cells and identification of general alarm signs.

[0017] This application cleans the data collected in S1, removing (deleting) some severely missing data. Since the dataset collected in S1 is large, a small amount of missing data has little impact on the training of the NARX neural network structure model, thus avoiding the impact of missing data on the accuracy and reliability of the calculation results. In addition, this application completes the data collected in S1 by supplementing the less missing data, thereby perfecting and optimizing the dataset and avoiding the impact of missing data, especially important data, on the accuracy and reliability of the calculation results.

[0018] As some preferred embodiments of this application, S3 is further selectively made to include:

[0019] S31: Based on the vehicle status, it is divided into driving status and off status; based on the battery status, it is divided into charging status, discharging status, and stationary status.

[0020] S32: Based on the vehicle status and battery status, filter the data processed by S2 and divide it into the first dataset of driving conditions, charging conditions and stationary conditions.

[0021] S33: Based on the first dataset divided by S32, calculate the correlation coefficients between the parameters of individual power battery cells in the first dataset;

[0022] S34: Based on the correlation coefficients calculated in S33, determine the variables that are strongly correlated with the voltage of the power battery cells;

[0023] S35: Filter out the data corresponding to variables that are strongly correlated with the voltage of the power battery cells, and create a second dataset under driving conditions, charging conditions, and stationary conditions.

[0024] This application identifies and filters variables that are strongly correlated with voltage and uses them as the basis for establishing a voltage fault early warning model. This not only allows the creation of second datasets under various conditions to discard more weakly correlated data to reduce computational load, but also makes the datasets for iterative training of the NARX neural network structure model more targeted.

[0025] As some preferred embodiments of this application, the correlation coefficient between the parameters of the power battery cells is further selectively adopted using the Pearson correlation coefficient, and the specific calculation formula is as follows:

[0026]

[0027] Where x and y are the actual observed values ​​of the two relevant parameters involved.

[0028] As some preferred embodiments of this application, S4 is further selectively made to include:

[0029] S41: Divide the second dataset into a test set and a training set;

[0030] S42: Slide the time window with a set step size to perform rolling optimization on the NARX neural network structure model on the training set and obtain the optimization result;

[0031] S43: Evaluate the optimization results on the test set, adjust the parameters according to the evaluation results, and obtain the voltage prediction value output by the NARX neural network structure model;

[0032] S44: Calculate the voltage residual value e based on the actual voltage observation value and the voltage prediction value output by the NARX neural network structure model;

[0033] S45: Establish a voltage fault early warning model based on the voltage residual value e and the classified voltage fault levels. It should be noted that the formula for calculating the voltage residual value e is:

[0034]

[0035] in: U represents the voltage estimate from the NARX neural network model; U represents the observed voltage value.

[0036] As some preferred embodiments of this application, the input to the NARX neural network structure model is further selectively made as:

[0037]

[0038] Where: x(t) represents the input time series data at time t; y(t) represents the output time series data at time t; n x n y These are the delay orders for the input and output, respectively;

[0039] The hidden layers of the NARX neural network structure model are calculated using the following formula:

[0040]

[0041] Where: f(·) is a nonlinear mapping function; k (t) represents the output of the output layer of the k-th hidden layer neuron; w ik The weights representing the connections between the i-th exogenous input and the k-th hidden layer neuron; w jk The weights connecting the j-th feedback input and the k-th hidden layer neuron are represented by w. 0k Indicates the weights of the hidden layer and the output layer; b k Indicates bias;

[0042] The output of the NARX neural network architecture model is calculated using the following formula:

[0043]

[0044] Where: g(·) is a linear activation function; b0 represents the output value of the NARX neural network architecture model; b0 represents the bias.

[0045] The formula for calculating the objective function of the NARX neural network architecture model is as follows:

[0046]

[0047] Where g(x, y) is a linear activation function; y represents the model output value; y represents the actual observed value.

[0048] As some preferred embodiments of this application, MSE is further selectively used to evaluate the NARX neural network structure model, and the calculation formula is as follows:

[0049]

[0050] in: y represents the model output value; y represents the actual observed value.

[0051] As some preferred embodiments of this application, the voltage fault levels are further selectively classified, specifically, for example, the voltage fault levels are divided into 4 types, for example:

[0052] If |e|≤ε1, it indicates that there is currently no voltage fault;

[0053] If ε1<|e|≤ε2, the power battery voltage fault level is Level 1;

[0054] If ε2<|e|≤ε3, the power battery voltage fault level is level two;

[0055] If ε3<|e|, the power battery voltage fault level is level three;

[0056] The specific values ​​of ε1, ε2, and ε3 are determined based on experimental and statistical prior information.

[0057] As some preferred embodiments of this application, S5 is further selectively made to include:

[0058] S51: For the temperature information at each moment after processing by S2, multiple confidence intervals are divided using a Gaussian distribution;

[0059] S52: Calculate the temperature probability distribution based on the frequency of falling into each confidence interval;

[0060] S53: Calculate the information entropy at the current moment;

[0061] S54: Based on S51, S52 and S53, set the information entropy threshold and establish a temperature fault early warning model.

[0062] As some preferred embodiments of this application, further selective division of confidence intervals includes:

[0063] S511: Filter the temperature data. Data within the normal range are used for interval division. The filtering function is as follows:

[0064]

[0065] S512: Calculate the mean and variance of temperature per unit time. The calculation formula is as follows:

[0066]

[0067]

[0068] Where: μ represents the average temperature, σ2 Represents the variance of the temperature distribution, where i is the temperature probe data number;

[0069] S513: Divide the confidence interval using the σ principle of the Gaussian distribution:

[0070] B = [b1 b2 b3 b4 b5]

[0071] Among them: b1∈(μ-3σ, μ+3σ], b2∈(μ-4σ, μ-3σ], b3∈(μ+3σ, μ+4σ], b4∈(-∞, μ-4σ], b5∈(μ+4σ, +∞).

[0072] As some preferred embodiments of this application, after selectively dividing the confidence interval, the frequency of falling into the confidence interval at that time is counted, and the frequency calculated based on the frequency of each confidence interval is regarded as the probability of falling into the confidence interval at that time, forming a probability interval P, as follows:

[0073] P = [p1 p2 p3 p4 ……p j ]

[0074] Where, probability p j The calculation formula is:

[0075]

[0076] Where: j is the number of confidence intervals divided, b i This represents the frequency of probe temperature falling within the confidence interval.

[0077] As some preferred embodiments of this application, the following formula is selectively used when calculating the information entropy at the current moment:

[0078]

[0079] As some preferred embodiments of this application, S6 is further selectively made to include:

[0080] S61: Establish the fault levels of the voltage fault early warning model and the fault levels of the temperature fault early warning model.

[0081] S62: Determine the fault level of the joint fault warning model based on the fault level of the voltage fault warning model and the fault level of the temperature fault warning model;

[0082] When the fault levels of the voltage fault early warning model and the temperature fault early warning model are different, the fault level of the joint fault early warning model shall be the highest level among the fault levels of the voltage fault early warning model and the temperature fault early warning model, and the calculation formula is as follows:

[0083] level = max{Tlevel, Ulevel}

[0084] Wherein, level represents the fault level of the joint fault warning model, Tlevel represents the fault level of the temperature fault warning model based on information entropy, and Ulevel represents the fault level of the voltage fault warning model based on voltage residual.

[0085] When the fault level of the voltage fault warning model and the fault level of the temperature fault warning model are the same, the fault level of the joint fault warning model is upgraded.

[0086] Based on the frequency of occurrence of various fault levels within a set time period using a joint fault early warning model, the long-term risk level of the power battery is classified, and a long-term risk assessment system for the power battery is established. Based on the duration of continuous fault occurrence within a set time period, the short-term risk level of the power battery is classified, and a short-term risk assessment system for the power battery is established. Attached Figure Description

[0087] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0088] Figure 1 The flowchart shows a data-driven method for power battery fault early warning and safety risk assessment.

[0089] Figure 2 This is a set of illustrative calculation results for the correlation coefficients of individual battery cell parameters;

[0090] Figure 3 Diagram showing voltage fault levels;

[0091] Figure 4 This is a schematic diagram illustrating the levels of temperature distribution anomalies.

[0092] Figure 5 This is a schematic diagram of a long-term risk assessment system.

[0093] Figure 6 This is a schematic diagram of a short-term risk assessment system. Detailed Implementation

[0094] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0095] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also mean including the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof.

[0096] like Figure 1 As shown, a data-driven power battery fault early warning and safety risk assessment method according to an embodiment of the present invention includes:

[0097] S1: Data acquisition, including acquisition time, vehicle status, charging status, operating mode, vehicle speed, total voltage, total current, cumulative mileage, SOC, highest voltage battery subsystem number, highest single-cell voltage value, highest voltage battery single-cell code, lowest voltage battery single-cell code, lowest voltage battery subsystem number, lowest single-cell voltage value, highest temperature value, highest temperature subsystem number, highest temperature probe number, highest alarm level, probe temperature, and general alarm flags;

[0098] S2: Process the data collected by S1;

[0099] S3: Based on the vehicle status and battery status, divide the data processed in S2, perform correlation analysis on the power battery parameters, and create a dataset based on the results of the correlation analysis.

[0100] S4: Divide the dataset into a training set and a test set, iteratively train the NARX (Nonlinear autoregressive exogenous) neural network structure model on the training set, and test the NARX neural network structure model on the test set. Based on the performance on the test set, adjust the hyperparameters of the NARX neural network structure model to establish a voltage fault early warning model.

[0101] S5: Establish a temperature fault early warning model based on information entropy according to the probe temperature;

[0102] S6: Establish a power battery safety risk assessment system based on voltage fault early warning model and temperature fault early warning model.

[0103] It should be noted that S1, S2, S3, S4, S5, and S6 in this application do not necessarily have a sequential order; they are merely symbols representing each step. For example, S1 represents the "data acquisition" step; S2 represents the "processing the data acquired in S1" step; similarly, S3, S4, S5, and S6 are also merely symbols representing their respective steps. This writing style is used solely for ease of description. Specifically, there is no sequential order between steps S4 and S5. In practice, S4 can be performed first, followed by S5; alternatively, S5 can be performed first, followed by S4; or alternatively, S4 and S5 can be performed simultaneously.

[0104] It should be noted that the raw data collected by S1 may selectively originate from the new energy vehicle big data platform. Preferably, the raw format of the collected data conforms to GB / T 32960.3—2016 "Technical Specifications for Remote Service and Management System of Electric Vehicles Part 3: Communication Protocol and Data Format". Before applying the data, it should be calculated and processed according to the description in the national standard. For example, the description and requirements of some important data are shown in Table 1:

[0105] Table 1: Description and Requirements of Some Important Data

[0106]

[0107]

[0108]

[0109] It should be noted that the data collected by S1 is not limited to the above types of data. It can also collect other types of data, such as the current of the battery cells of the power battery. The specific selection can be made according to actual needs.

[0110] It should be noted that the power battery in this application can be selectively formed by multiple battery cells connected in series, parallel, or in a hybrid configuration. The term "hybrid configuration" refers to a connection form that includes both series and parallel connections.

[0111] This application presents a power battery fault early warning and safety risk assessment method comprising S1 to S6, which has low computational requirements and can accurately predict and assess power batteries, showing promising application prospects. The NARX neural network structure model exhibits good dynamics and anti-interference capabilities for power batteries, enabling high-precision and relatively accurate voltage prediction. Furthermore, the NARX neural network structure model for power battery voltage prediction does not require high data acquisition accuracy, allowing the use of low-frequency data and effectively reducing data storage costs.

[0112] As some preferred embodiments of this application, S2 is further selectively made to include:

[0113] S21: Arrange the data collected in S1 in chronological order;

[0114] S22: Clean and complete the data arranged in chronological order;

[0115] S23: Perform data separation of individual power battery cells and identification of general alarm signs.

[0116] Because the raw data collected by S1 has problems such as out-of-order data, duplication, and missing data, it is necessary to process the raw data to avoid causing significant problems for subsequent data analysis and model training. In practice, the processing of raw data includes deleting redundant data and data with many missing data, as well as completing some data.

[0117] For example, the original data may contain issues such as time disorder and identical times, as shown in Table 2.

[0118] Table 2 shows potential time-related issues with the original data.

[0119]

[0120] In practice, the data is first sorted in ascending order to overcome the problems caused by out-of-order time data in subsequent analysis and training. For data with the same time, a set of data can be selectively deleted. If one of the data sets with repeated time has missing values ​​and the other does not, it is preferable to keep the data set with no missing values; otherwise, keep the first row of data from the repeated time sets.

[0121] Furthermore, the original data may also contain partial missing data, complete missing data, or data mutations, as shown in Table 3.

[0122] Table 3 shows potential missing or abrupt issues in the original data.

[0123]

[0124] In practice, for the problem of completely missing data, considering the large dataset and the minimal impact of a small number of missing data points on model training, the option is to directly delete that row of data. For partially missing data, interpolation is used to fill in the gaps, selectively using a previous, unmissing value to replace the missing one.

[0125] To address data mutations, a sliding window can be selectively used to detect local outliers, which are then replaced by linear interpolation. Specifically, the sliding window has two parameters: window size and sliding step. The window size refers to the amount of data contained within the window, and the sliding step refers to the amount of data that needs to be traversed. For example, in a dataset with 1000 data points, the sliding window size is 100, and the sliding step is 10. Initially, the first 100 data points are in the initial sliding window. After one slide, data points 11-110 are in the sliding window. Traversing the entire dataset requires 100 slides. This application uses a sliding window to detect local outliers and further employs linear interpolation to replace the mutated local outliers, thereby processing the mutated data and facilitating subsequent data processing, avoiding difficulties in data analysis and model training.

[0126] It is important to note that in practical implementation, data mutations are preferably applied to parameters that change relatively slowly, such as SOC, voltage, and temperature, while parameters that change rapidly, such as current, are not subject to data mutation processing. Local outliers are defined as elements whose mean differs from the mean of the data within the sliding window by more than three standard deviations. This method has the advantages of fast calculation speed and relatively accurate outlier identification. The formulas for calculating the mean and standard deviation of the data within the sliding window, and the outlier replacement method, are as follows:

[0127] Let the size of the sliding window be m, then the data within the sliding window is as follows:

[0128] X = [x1 x2 … x m-1 x m ]

[0129] Where: X represents a battery parameter in the sliding window; x i This represents the data elements in the sliding window, i = 1, 2, ..., m.

[0130] The formula for calculating the mean of data in a sliding window is as follows:

[0131]

[0132] Where: m represents the number of data items in the sliding window; x i This represents the i-th data element in the sliding window.

[0133] The standard deviation of the data in the sliding window is as follows:

[0134]

[0135] The formula for determining whether an element within a sliding window is an outlier is as follows:

[0136]

[0137] Linear interpolation is used to replace outliers, and the calculation formula is as follows:

[0138]

[0139] In the above formula, x t For outliers, x t-1 x t+1 These represent the values ​​of the outlier at the previous and next time points, respectively.

[0140] Regarding the separation of power battery cell data and the identification of general alarm flags, it should be noted that the separation of power battery cell data includes the separation of battery cell voltage, probe temperature, and current, and the identification of general alarm flags complies with GB / T32960.3—2016 "Technical Specifications for Remote Service and Management System of Electric Vehicles Part 3: Communication Protocol and Data Format".

[0141] This application cleans the data collected in S1, removing (deleting) some severely missing data. Since the dataset collected in S1 is large, a small amount of missing data has little impact on the training of the NARX neural network structure model, thus avoiding the impact of missing data on the accuracy and reliability of the calculation results. In addition, this application completes the data collected in S1 by supplementing the less missing data, thereby perfecting and optimizing the dataset and avoiding the impact of missing data, especially important data, on the accuracy and reliability of the calculation results.

[0142] As some preferred embodiments of this application, S3 is further selectively made to include:

[0143] S31: Based on the vehicle status, it is divided into driving status and off status; based on the battery status, it is divided into charging status, discharging status, and stationary status.

[0144] S32: Based on the vehicle status and battery status, filter the data processed by S2 and divide it into the first dataset of driving conditions, charging conditions and stationary conditions.

[0145] S33: Based on the first dataset divided by S32, calculate the correlation coefficients between the parameters of individual power battery cells in the first dataset;

[0146] S34: Based on the correlation coefficients calculated in S33, determine the variables that are strongly correlated with the voltage of the power battery cells;

[0147] S35: Filter out the data corresponding to variables that are strongly correlated with the voltage of the power battery cells, and create a second dataset under driving conditions, charging conditions, and stationary conditions.

[0148] It should be noted that, in this application, the charging state may selectively include charging while parked and charging while driving; the discharging state refers to the state where the vehicle is not charging and the total current of the power battery is not zero; the stationary state refers to the state where the vehicle is turned off and is not charging.

[0149] This application identifies and filters variables that are strongly correlated with voltage and uses them as the basis for establishing a voltage fault early warning model. This not only allows the creation of second datasets under various conditions to discard more weakly correlated data to reduce computational load, but also makes the datasets for iterative training of the NARX neural network structure model more targeted.

[0150] As some preferred embodiments of this application, the correlation coefficient between the parameters of the power battery cells is further selectively adopted using the Pearson correlation coefficient, and the specific calculation formula is as follows:

[0151]

[0152] Where x and y are the actual observed values ​​of the two related parameters. For example, when calculating the correlation between voltage and temperature, one of x and y is the voltage parameter, and the other is the temperature parameter. Similarly, when calculating the correlation coefficient between SOC and voltage, one of x and y is the SOC, and the other is the voltage parameter.

[0153] When calculating the correlation coefficient, the Pearson correlation coefficient can be used. For detailed calculation results, please refer to [link to relevant documentation]. Figure 2 The larger the data for the two battery parameters, the stronger the linear correlation between them. For example... Figure 2 In the data, the correlation coefficient between individual unit voltage and total voltage is 1, and the correlation coefficient between individual unit voltage and SOC is 0.955, indicating a large relationship between individual unit voltage and total voltage, and between individual unit voltage and SOC. However, the correlation coefficient between individual unit voltage and internal resistance is 0.031, indicating a small relationship between the two.

[0154] As some preferred embodiments of this application, S4 is further selectively made to include:

[0155] S41: Divide the second dataset into a test set and a training set;

[0156] S42: Slide the time window with a set step size to perform rolling optimization on the NARX neural network structure model on the training set and obtain the optimization result;

[0157] S43: Evaluate the optimization results on the test set, adjust the parameters according to the evaluation results, and obtain the voltage prediction value output by the NARX neural network structure model;

[0158] S44: Calculate the voltage residual value e based on the actual voltage observation value and the voltage prediction value output by the NARX neural network structure model;

[0159] S45: Establish a voltage fault early warning model based on the voltage residual value e and the classified voltage fault levels. It should be noted that the formula for calculating the voltage residual value e is:

[0160]

[0161] in: U represents the voltage estimate from the NARX neural network model; U represents the observed voltage value.

[0162] It should be noted that the temporal order of the data must be maintained when partitioning the dataset. When dividing the dataset into training and test sets, the top 75% and bottom 25% of the dataset can be used respectively, i.e., the top 75% as the training set and the bottom 25% as the test set. The specific proportions can be adjusted appropriately based on engineering experience. For example, the top 70% and bottom 30% could be used as the training set and the bottom 30% as the test set; or the top 71% and bottom 29% as the training set and the bottom 29% as the test set; or the top 72% and bottom 28% as the training set and the bottom 28% as the test set; or the top 73% and bottom 27% as the training set and the bottom 27% as the test set; or the top 74% and bottom 26% as the training set and the bottom 26% as the test set; and so on.

[0163] As some preferred embodiments of this application, the input to the NARX neural network structure model is further selectively made as:

[0164]

[0165] Where: x(t) represents the input time series data at time t, specifically the variables strongly correlated with the unit voltage selected by S35; y(t) represents the output time series data at time t; n x n y These are the delay orders for the input and output, respectively;

[0166] The hidden layers of the NARX neural network structure model are calculated using the following formula:

[0167]

[0168] Where: f(·) is a nonlinear mapping function, and "·" refers to the expression within parentheses in the calculation formula; H k (t) represents the output of the output layer of the k-th hidden layer neuron; w ik The weights representing the connections between the i-th exogenous input and the k-th hidden layer neuron; w jk The weights connecting the j-th feedback input and the k-th hidden layer neuron are represented by w. 0k Indicates the weights of the hidden layer and the output layer; b k Indicates bias;

[0169] The output of the NARX neural network architecture model is calculated using the following formula:

[0170]

[0171] Where: g(·) is a linear activation function, and "·" refers to the expression within parentheses in the calculation formula; b0 represents the output value of the NARX neural network architecture model; b0 represents the bias.

[0172] The formula for calculating the objective function of the NARX neural network architecture model is as follows:

[0173]

[0174] Where g(x, y) is a linear activation function; y represents the model output value; y represents the actual observed value.

[0175] As some preferred embodiments of this application, MSE (Mean Squared Error) is further selectively used to evaluate the NARX neural network structure model, and the calculation formula is as follows:

[0176]

[0177] in: y represents the model output value; y represents the actual observed value.

[0178] It's important to note that a smaller MSE indicates better model performance on the test set, and hyperparameter tuning should be based on MSE. In the models described above, parameters that need to be manually selected are called hyperparameters. For example, in the NARX neural network model, the number of hidden layers and nodes per layer, the size of the constant in the regularization term, etc., these parameters need to be specified beforehand. Inappropriate hyperparameter selection can lead to underfitting or overfitting. There are two approaches to selecting hyperparameters: one is to fine-tune them based on experience, and the other is to select parameters of different sizes, input them into the model, and choose the one that performs best.

[0179] It should be noted that when calculating voltage residuals, the output of the NARX neural network model is the predicted value of the individual cell voltage. The formula for calculating voltage residuals is as follows:

[0180]

[0181] Where: e is the voltage residual; U represents the voltage estimate from the NARX neural network model; U represents the voltage observation.

[0182] The NARX neural network model exhibits good dynamics and anti-interference capabilities for nonlinear systems, and demonstrates high accuracy in predicting single-cell voltage. Validation shows that the maximum MSE is only approximately 1x10⁻⁶. -4 Approximately. Furthermore, the NARX neural network model for voltage prediction does not require high data acquisition accuracy, allowing the use of low-frequency data, significantly reducing data storage costs and processing volume. In practical implementation, a sliding window is further applied to update the neural network parameters, thereby reducing training time and improving real-time prediction efficiency.

[0183] As some preferred embodiments of this application, the fault levels of the voltage fault early warning model are further selectively classified. For example, the voltage fault levels are divided into three types, specifically:

[0184] If |e|≤ε1, it indicates that there is currently no voltage fault;

[0185] If ε1<|e|≤ε2, the power battery voltage fault level is Level 1;

[0186] If ε2 < |e| ≤ ε3, the power battery voltage fault level is level two;

[0187] If ε3 < |e|, the power battery voltage fault level is level three;

[0188] The specific values ​​of ε1, ε2, and ε3 are determined based on experimental and statistical prior information, and the voltage fault level can be represented as follows: Figure 3 As shown.

[0189] As some preferred embodiments of this application, S5 may be further selectively included as follows:

[0190] S51: For the temperature information at each moment after processing by S2, multiple confidence intervals are divided using a Gaussian distribution;

[0191] S52: Calculate the temperature probability distribution based on the frequency of falling into each confidence interval;

[0192] S53: Calculate the information entropy at the current moment;

[0193] S54: Based on S51, S52 and S53, set the information entropy threshold and establish a temperature fault early warning model.

[0194] As some preferred embodiments of this application, further selective division of confidence intervals includes:

[0195] S511: Filter the temperature data. Data within the normal range are used for interval division. The filtering function is as follows:

[0196]

[0197] S512: Calculate the mean and variance of temperature per unit time. The calculation formula is as follows:

[0198]

[0199]

[0200] Where: μ represents the average temperature, σ 2 Represents the variance of the temperature distribution, where i is the temperature probe data number;

[0201] S513: Divide the confidence interval using the σ principle of the Gaussian distribution:

[0202] B = [b1 b2 b3 b4 b5]

[0203] Among them: b1∈(μ-3σ, μ+3σ], b2∈(μ-4σ, μ-3σ], b3∈(μ+3σ, μ+4σ], b4∈(-∞, μ-4σ], b5∈(μ+4σ, +∞).

[0204] It should be noted that the values ​​of A℃ and B℃ in the function are determined based on the actual operating conditions. For example, A℃ can be set to -25℃ and B℃ to 55℃. As an alternative implementation, A℃ can be set to -20℃ and B℃ to 50℃, and these values ​​can be selectively set according to the actual situation.

[0205] As some preferred embodiments of this application, the confidence interval is further selectively divided, and the frequency of falling into the confidence interval at a given time is counted. The frequency calculated based on the frequency of each confidence interval is taken as the probability of falling into the confidence interval at that time, forming a probability interval P, as follows:

[0206] P = [p1 p2 p3 p4 …… p] j ]

[0207] Where, probability p j The calculation formula is:

[0208]

[0209] Where: j is the number of confidence intervals divided, b i This represents the frequency of probe temperature falling within the confidence interval.

[0210] As some preferred embodiments of this application, the following formula is selectively used when calculating the information entropy at the current moment:

[0211]

[0212] As some preferred embodiments of this application, an information entropy threshold is further selectively set based on experiments and mathematical statistics to determine the fault level of the temperature fault early warning model; in specific implementations, the temperature fault level is further selectively divided into three types, including:

[0213] If |E t When |≤ξ1, it indicates that there is currently no temperature fault;

[0214] If ξ1 < |E t When |≤ξ2, the power battery temperature fault level is level one;

[0215] If ξ2<|E t When |≤ξ3, the power battery temperature fault level is level two;

[0216] If ξ3<|E t At that time, the power battery temperature fault level was level three;

[0217] The specific values ​​of ξ1, ξ2, and ξ3 are determined based on experimental and statistical prior information. Temperature fault level classification can be as follows: Figure 5 As shown. This application's information entropy calculation based on temperature distribution does not require high data acquisition accuracy, enabling the identification of battery faults using low-frequency data and significantly reducing data storage costs. Furthermore, the information entropy calculation utilizes data information at each moment without comparing data over a time series, resulting in low time complexity and real-time application. Additionally, the information entropy-based temperature fault early warning model considers the overall battery temperature distribution and addresses battery inconsistencies, enabling early diagnosis and warning of battery faults and preventing potential failures.

[0218] As some preferred embodiments of this application, in S61: establish the fault level of the voltage fault early warning model, and establish the fault level of the temperature fault early warning model;

[0219] S62: Determine the fault level of the joint fault warning model based on the fault level of the voltage fault warning model and the fault level of the temperature fault warning model;

[0220] When the fault levels of the voltage fault early warning model and the temperature fault early warning model are different, the fault level of the joint fault early warning model shall be the highest level among the fault levels of the voltage fault early warning model and the temperature fault early warning model, and the calculation formula is as follows:

[0221] level = max{Tlevel, Ulevel}

[0222] Wherein, level represents the fault level of the joint fault warning model, Tlevel represents the fault level of the temperature fault warning model based on information entropy, and Ulevel represents the fault level of the voltage fault warning model based on voltage residual.

[0223] When the fault level of the voltage fault warning model and the fault level of the temperature fault warning model are the same, the fault level of the joint fault warning model is upgraded.

[0224] Specifically, such as Figure 1 As shown in the fault level table of the joint fault early warning model established by S6, when both temperature and voltage are normal values, the fault level of the joint fault early warning model is normal; when one of temperature and voltage corresponds to a fault level of level one and the other corresponds to a fault level of level two, the fault level of the joint fault early warning model is level two; when both temperature and voltage correspond to level one fault, the fault level of the joint fault early warning model is level two.

[0225] As some preferred embodiments of this application, the power battery safety risk assessment system established in S6 is further selectively configured to include a long-term risk assessment system and a short-term risk assessment system. In specific implementation, the long-term risk level of the power battery is further selectively classified based on the frequency of occurrence of various fault levels in the joint fault early warning model within a set period of time, and a long-term risk assessment system for the power battery is established; the short-term risk level of the power battery is classified based on the duration of continuous fault occurrence within a set period, and a short-term risk assessment system for the power battery is established.

[0226] As some preferred embodiments of this application, a power battery safety risk assessment system is further established based on the combined fault levels. A long-term risk assessment system based on fault frequency statistics is established by statistically analyzing the occurrence frequency (n) of various fault levels over a relatively long period to determine the long-term risk level of the power battery.

[0227] In specific implementation, the long-term risk level of power batteries is further selectively divided into three categories, including:

[0228] If the frequency of a first-level fault is n1≤N 11 This indicates that the current power battery has no long-term risk;

[0229] If the frequency of a first-level fault is N 11 <n1≤N 12 At that time, the long-term risk of current power batteries is low.

[0230] If the frequency of a first-level fault is N 12 When n < n1, the current long-term risk of power batteries is medium risk;

[0231] If the frequency of a level 2 fault is n2≤N 21 When this is the case, it indicates that there is no long-term risk at present;

[0232] If the frequency of a level 2 fault is N 21 <n2≤N 22 At that time, the long-term risk of power batteries was medium risk.

[0233] If the frequency of a level 2 fault is N 22 When n < n2, the current long-term risk of power batteries is high.

[0234] If the frequency of a level 2 fault is N 22 When n < n2, the current long-term risk of power batteries is high.

[0235] If the frequency of a level 3 fault is n3≤N 31 This indicates that the current power battery has no long-term risk;

[0236] If the frequency of a level 3 fault is N 31 When n < n3, the current long-term risk of power batteries is high.

[0237] Among them, the current long-term risk level of the power battery is the highest risk level determined by the Level 1, Level 2, and Level 3 fault assessment results; N 11 N 12 N 21 N 22 N 31 The specific values ​​are determined based on prior knowledge of big data and mathematical statistics. Figure 5 This is a schematic diagram of a long-term risk assessment system.

[0238] It should be noted that long-term risk assessment focuses on historical data, judging and escalating current safety risks based on the cumulative frequency of failures, corresponding to the potential long-term safety risks of the battery. For example, using one month's historical data, if more than 100 level-two failures occur, it is considered high-risk and requires timely vehicle repair. In practice, the specific threshold needs to be determined based on mathematical statistics.

[0239] In practical implementation, a short-term risk assessment system based on the continuous occurrence of faults is selectively established. The short-term risk level of the power battery is determined by the duration of consecutive fault occurrences within a short period. It should be noted that short-term risk focuses on instantaneous data. For example, if a Level 3 fault alarm occurs continuously for 20 seconds at a certain moment, it is considered high risk, corresponding to a sudden safety risk to the battery. The duration of consecutive occurrences is determined by statistical analysis.

[0240] In practice, the short-term risk level of power batteries is divided into three categories, including:

[0241] If the duration of a Level 1 fault is t1≤T1, it indicates that the current power battery has no short-term risk.

[0242] If the duration of a Level 1 fault is less than t1, the current short-term risk of the power battery is low.

[0243] If the duration of the secondary fault is t2≤T2, the current power battery has no short-term risk.

[0244] If the duration of a Level 2 fault is less than t2, the current short-term risk of the power battery is medium risk.

[0245] If the Level 3 fault continues to occur for a duration t3≤T3, the current power battery has no short-term risk.

[0246] If the duration of a Level 3 fault is less than t3, the current short-term risk of the power battery is high.

[0247] Among them, the current short-term risk level of the power battery is the highest risk level determined by the first, second and third level fault assessment results.

[0248] The specific values ​​of T1, T2, and T3 are determined based on prior knowledge of big data and mathematical statistics. A schematic diagram of the short-term risk assessment system is shown below. Figure 6 It should be noted that there is no necessary connection or size relationship between T1, T2, and T3, but in specific implementation, it is preferable to make T1 < T2 < T3.

[0249] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data-driven method for early warning and safety risk assessment of power battery faults, characterized in that, include: S1: Data acquisition, the data includes acquisition time, vehicle status, charging status, operating mode, vehicle speed, total voltage, total current, cumulative mileage, SOC, highest voltage battery subsystem number, highest battery cell voltage, highest voltage battery cell code, lowest voltage battery cell code, lowest voltage battery subsystem number, lowest battery cell voltage, highest temperature value, highest temperature subsystem number, highest temperature probe number, highest alarm level, probe temperature, and general alarm flags; S2: Process the data collected by S1; S3: Based on the vehicle status and battery status, divide the data processed in S2, perform correlation analysis on the power battery parameters, and create a dataset based on the results of the correlation analysis. S4: Divide the dataset into a training set and a test set, iteratively train the NARX neural network structure model on the training set, test the NARX neural network structure model on the test set, adjust the hyperparameters of the NARX neural network structure model based on the performance on the test set, evaluate the model, and establish a voltage fault early warning model. S5: Based on the probe temperature, establish a temperature fault early warning model based on information entropy; S6: Based on the voltage fault early warning model and the temperature fault early warning model, establish a power battery safety risk assessment system.

2. The data-driven method for early warning and safety risk assessment of power batteries according to claim 1, characterized in that, S2 includes: S21: Arrange the data collected in S1 in chronological order; S22: Clean and complete the data arranged in chronological order; S23: Perform data separation of individual power battery cells and identification of general alarm signs.

3. The data-driven method for early warning and safety risk assessment of power batteries according to claim 2, characterized in that, S3 includes: S31: Based on the vehicle status, it is divided into driving status and off status; based on the battery status, it is divided into charging status, discharging status, and stationary status. S32: Based on the vehicle status and the battery status, filter the data processed in S2 and divide it into a first dataset of driving conditions, charging conditions and stationary conditions. S33: Based on the first dataset divided in S32, calculate the correlation coefficient between the parameters of the individual power battery cells in the first dataset. The correlation coefficient between the parameters of the individual power battery cells is the Pearson correlation coefficient, and the specific calculation formula is as follows: Where: x and y are the actual observed values ​​of the two relevant parameters involved; S34: Based on the correlation coefficient calculated in S33, determine the variable that is strongly correlated with the voltage of the power battery cell; S35: Filter out the data corresponding to variables that are strongly correlated with the voltage of the power battery cell, and create a second dataset under driving conditions, charging conditions, and stationary conditions.

4. The data-driven method for early warning and safety risk assessment of power batteries according to claim 3, characterized in that, S4 includes: S41: Divide the second dataset into a test set and a training set; S42: Slide the time window with a set step size to perform rolling optimization on the NARX neural network structure model on the training set and obtain the optimization result; S43: Evaluate the optimization results on the test set, adjust the parameters according to the evaluation results, and obtain the voltage prediction value output by the NARX neural network structure model; S44: Calculate the voltage residual value e based on the actual voltage observation value and the voltage prediction value output by the NARX neural network structure model; S45: Based on the voltage residual value e and the classified voltage fault levels, establish a voltage fault early warning model.

5. The data-driven method for early warning and safety risk assessment of power batteries according to claim 4, characterized in that, The input to the NARX neural network structure model is: in: This represents the input time series data at time t; This indicates that the time series data will be output at time t. , These are the delay orders for the input and output, respectively; The hidden layers of the NARX neural network structure model are calculated using the following formula: in: It is a nonlinear mapping function; This represents the output of the output layer of the k-th hidden layer neuron; Indicates the connection of the first The exogenous input and the weights of the k-th hidden layer neuron; Indicates the connection of the first The feedback input and the weights of the k-th hidden layer neuron; This represents the weights of the hidden layer and the output layer; Indicates bias; The output of the NARX neural network structure model is calculated using the following formula: in: It is a linear activation function; This represents the output value of the NARX neural network architecture model; Indicates bias; The formula for calculating the objective function of the NARX neural network architecture model is as follows: in: It is a linear activation function; Indicates the model output value; This represents the actual observed value.

6. The data-driven method for early warning and safety risk assessment of power batteries according to claim 5, characterized in that, The NARX neural network architecture model is evaluated using the MSE (Mean Sequence of Evaluation), and the calculation formula is as follows: in: Indicates the model output value; This represents the actual observed value.

7. The data-driven method for early warning and safety risk assessment of power batteries according to claim 6, characterized in that, S51: For the temperature information at each moment after processing by S2, multiple confidence intervals are divided using a Gaussian distribution; S52: Calculate the temperature probability distribution based on the frequency of falling into each confidence interval; S53: Calculate the information entropy at the current moment; S54: Based on S51, S52 and S53, set the information entropy threshold and establish a temperature fault early warning model.

8. The data-driven method for early warning and safety risk assessment of power batteries according to claim 7, characterized in that, When dividing the confidence interval, the following steps are included: S511: Filter the temperature data. Data within the normal range are used for interval division. The filtering function is as follows: S512: The formulas for calculating the mean and variance of temperature per unit time are as follows: in: Represents the average temperature. Represents the variance of the temperature distribution. Number the temperature probe data; S513: Using Gaussian distribution The principle is to divide the confidence interval: in: ∈ ( , ], ∈ ( , ], ∈ ( , ], ∈ ( , ], ∈ ( , ).

9. The data-driven method for early warning and safety risk assessment of power batteries according to claim 8, characterized in that, After dividing the confidence interval, the frequency of each time point falling within that confidence interval is counted. The frequency calculated based on the frequency of each confidence interval is taken as the probability of falling within that confidence interval at that time, forming a probability interval P, as follows: Among them, probability The calculation formula is: Where: j is the number of confidence intervals divided. The frequency of the probe temperature falling within the confidence interval; The information entropy at the current moment is calculated according to the following formula: 。 10. The data-driven method for early warning and safety risk assessment of power batteries according to any one of claims 1 to 9, characterized in that, S6 includes: S61: Establish the fault levels of the voltage fault early warning model and the fault levels of the temperature fault early warning model. S62: Determine the fault level of the joint fault warning model based on the fault level of the voltage fault warning model and the fault level of the temperature fault warning model; When the fault levels of the voltage fault early warning model and the temperature fault early warning model are different, the fault level of the joint fault early warning model shall be the highest level among the fault levels of the voltage fault early warning model and the temperature fault early warning model, and the calculation formula is as follows: in, This indicates the fault level of the joint fault early warning model. This indicates the fault level of the temperature fault early warning model based on information entropy. This indicates the fault level in a voltage fault early warning model based on voltage residuals. When the fault level of the voltage fault warning model and the fault level of the temperature fault warning model are the same, the fault level of the joint fault warning model is upgraded. Based on the frequency of occurrence of various fault levels within a set time period using a joint fault early warning model, the long-term risk level of the power battery is classified, and a long-term risk assessment system for the power battery is established. Based on the duration of continuous fault occurrence within a set time period, the short-term risk level of the power battery is classified, and a short-term risk assessment system for the power battery is established.

Images