Power battery short circuit risk quantitative evaluation method fusing physical mechanism and deep learning

Abstract

The invention relates to the field of battery health monitoring, and discloses a physical mechanism and deep learning fused power battery short circuit risk quantitative evaluation method, which comprises the following steps: acquiring real vehicle operation data from a data cloud platform, cleaning the real vehicle operation data to construct an equivalent circuit model, fitting model key parameters by means of a recursive least square method, and calculating the short circuit risk of a power battery. The processed data is converted into a time sequence format and standardized, a time window is divided to generate a target label, a time convolution network model containing structures of causal convolution, expansion convolution and the like is constructed, the model is trained by using a cross entropy loss function, evaluation is carried out by using indexes such as accuracy, and the trained model can receive multi-dimensional data of the battery in real time. The short-circuit risk probability is predicted, an alarm is triggered when a threshold value is exceeded, high-precision and real-time quantitative evaluation of the short-circuit risk of the power battery is realized, and the operation safety of the battery is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of battery health monitoring, and in particular to a method for quantitatively assessing the short-circuit risk of power batteries by integrating physical mechanisms and deep learning. Background Art

[0002] As a critical component in electric vehicles, energy storage systems, and other fields, power batteries' safety is directly related to the equipment's service life and operational stability. Short-circuit failures are a common safety hazard for power batteries, leading to performance degradation and even serious safety incidents such as fires and explosions. Therefore, accurately assessing and predicting power battery short-circuit risks is crucial for improving battery safety and ensuring stable equipment operation.

[0003] At present, the assessment method of battery short-circuit risk often relies on traditional physical mechanism models. Traditional physical mechanism models are usually simulated and calculated based on parameters such as the internal structure, current, voltage, and temperature of the battery. The short-circuit risk is predicted by assuming electrochemical reactions, thermal effects, and current changes inside the battery. The model is highly complex and it is difficult to accurately describe the behavior of the battery in a complex environment. At the same time, the computational cost is high, and computational efficiency becomes a bottleneck. Since the environmental conditions (such as temperature, current, etc.) are assumed to be fixed, but the battery will be affected by different working environments during actual use, the traditional model is difficult to cope with dynamic and nonlinear working conditions.

[0004] With the rapid development of big data technology and machine learning methods, data-driven approaches are widely used to assess the short-circuit risk of power batteries. These approaches collect sensor data (such as voltage, current, and temperature) from batteries under different operating conditions and utilize machine learning algorithms (such as neural networks and support vector machines) to generate models and predictions. However, data-driven approaches require large amounts of high-quality data. In practical applications, data collection can be subject to missing data, noise, or bias, which can affect the model's predictive accuracy. While complex algorithms such as deep learning can provide highly accurate predictions, their black-box nature makes it difficult to interpret the model's internal mechanisms, making battery short-circuit risk diagnosis and fault analysis challenging. Although machine learning methods excel at pattern recognition, relying solely on data-driven approaches struggles to capture the complex physical and chemical processes of batteries, leading to a decline in model accuracy under some extreme operating conditions. Summary of the Invention

[0005] The present invention aims to provide a quantitative assessment method for power battery short-circuit risk that integrates physical mechanisms and deep learning. By integrating physical mechanisms and deep learning, high-precision and real-time quantitative assessment of power battery short-circuit risk can be achieved under complex data and model conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for quantitatively assessing the short-circuit risk of power batteries that integrates physical mechanisms and deep learning, including: S1, selects real vehicle operation data from the data cloud platform and performs data cleaning; S2, build an equivalent circuit model and identify key parameters based on actual vehicle operation data; S3, using recursive least squares method to fit the key parameters of the equivalent circuit model and obtain the optimal parameters; S4, combining the cleaned actual vehicle operation data with the optimal parameters and converting them into a time series format, and performing normalization and standardization on the time series data; S5, divides the processed time series data into multiple time windows, and generates a target label for each time window: whether a short circuit occurs; S6, build a temporal convolutional network model, which includes an input layer, multiple TCN layers, a global pooling layer, a fully connected layer, and an output layer; S7, training the temporal convolutional network model using a cross entropy loss function, adjusting the learning rate and accelerating convergence; S8, uses accuracy, precision and recall to evaluate the trained temporal convolutional network model; S9 uses the trained temporal convolutional network model to receive multi-dimensional data of the battery in real time, predicts the short circuit risk probability of the battery, and generates an alarm when the risk probability exceeds the set threshold.

[0007] The principles and advantages of this solution are as follows: In practical applications, during the data processing phase, after extracting real-world vehicle operating data from a data cloud platform, data cleaning is performed by removing outliers, filling missing data (using mean filling or interpolation), and deleting irrelevant variables to provide high-quality data for subsequent analysis. The cleaned data is then converted to a time series format with time correlation and normalized using the optimal parameter combination obtained by fitting an equivalent circuit model. During the modeling phase, an equivalent circuit model is constructed to describe the battery's electrical behavior, and key model parameters are fitted using recursive least squares. The current parameter estimates of the output equation are first linearized, and the parameters and covariance matrix are initialized. The covariance matrix is ​​continuously updated to fit the optimal parameters by calculating the prediction error and gain vector, simultaneously identifying the battery's state of health (SOH). During the deep learning phase, a temporal convolutional network model is constructed. The input layer receives the processed time series data, and multiple TCN layers perform convolution to capture data features. The model is trained using a cross-entropy loss function, and convergence is accelerated by adjusting the learning rate. Finally, model performance is evaluated using accuracy, precision, and recall. The trained model can receive multi-dimensional battery data in real time, predict the probability of short-circuit risk, and generate an alert when the threshold is exceeded. This solution combines an equivalent circuit model based on physical mechanisms with a temporal convolutional network model based on deep learning. This approach leverages both the physical characteristics of the battery and the data processing capabilities of deep learning to improve assessment accuracy. Recursive least squares is used to quickly fit the optimal parameters of the equivalent circuit model, improving the model's description of the battery's electrical behavior. The trained model can predict the probability of short-circuit risk in real time and generate timely alerts when the threshold is exceeded, ensuring the safe operation of the power battery.

[0008] Preferably, as an improvement, the data cleaning includes removing outliers, filling missing data, and deleting irrelevant variables. The method of filling missing data is mean filling or interpolation. The formula for mean filling is:

[0009] in, is the value after filling, is the existing data of the same feature column, is the number of data points.

[0010] Technical effect: It is easy to ensure data integrity and accuracy, improve the reliability of subsequent analysis and model training, and provide a high-quality data basis for accurate assessment of short-circuit risks.

[0011] Preferably, as an improvement, the equivalent circuit model is:

[0012] Where R is the internal resistance, C is the capacitance, is the open circuit voltage of the battery, and I is the current.

[0013] Technical effects: facilitate the description of the electrical behavior of the battery, provide a theoretical basis for understanding the working principle and parameter identification of the battery, and make the model better reflect the actual running state of the battery.

[0014] Preferably, as an improvement, in S3, the recursive least squares method is used to fit the key parameters of the equivalent circuit model, including: S31, linearize the current parameter estimation of the output equation; S32, initialize the parameters and initialize the covariance matrix, calculate the prediction error and gain vector, and update the covariance matrix based on the prediction error and gain vector and fit the optimal parameters; S33, identify the battery state of health SOH through the capacity degradation and internal resistance increase of the battery.

[0015] Technical effects: the recursive least squares method can quickly and accurately fit the optimal parameters of the equivalent circuit model, while identifying the battery state of health SOH, improving the model accuracy and comprehensive understanding of the battery state.

[0016] Preferably, as an improvement, the prediction error is:

[0017] Wherein, is the current parameter prediction, is the current battery voltage.

[0018] Technical effects: facilitate the calculation of the difference between the current parameter prediction and the actual voltage, and provide a basis for parameter updating.

[0019] Preferably, as an improvement, the gain vector is:

[0020] Wherein, is the gradient vector, is the forgetting factor, is the covariance matrix.

[0021] Technical effects: facilitate the combination of gradient vector, forgetting factor and covariance matrix, reasonable adjustment of parameter update amplitude, and guarantee the accuracy and stability of parameter fitting.

[0022] Preferably, as an improvement, the time series format is:

[0023] Wherein, V is the voltage, I is the current, T is the temperature, S is the charge and discharge time, R is the internal resistance, C is the capacitance, is the open circuit voltage of the battery, and SOH is the state of health of the battery.

[0024] Technical effects: for the processing of time series data, it is convenient to ensure that the data input has time correlation, which is suitable for subsequent deep learning model training.

[0025] Preferably, as an improvement, in S6, the time convolution network model comprises causal convolution and dilated convolution, and the dilated convolution operation comprises:

[0026] wherein, is the size of the convolution kernel, is the weight of the convolution kernel, is the dilation factor.

[0027] Technical effects: through causal convolution, it is convenient to ensure that future information will not be used when predicting, and through dilated convolution, long-term dependencies can be captured.

[0028] Preferably, as an improvement, in the time convolution network model, the input layer receives time series data; each TCN layer in the plurality of TCN layers captures local features and long-term dependencies in the time series through dilated convolution and causal convolution, and adopts residual connection to train stability; the global pooling layer performs global average pooling on the data output by the TCN layer; the fully connected layer maps the global features to the output space, and outputs the final classification result; the output layer uses the Softmax activation function to directly output the binary classification result.

[0029] Technical effects: it is convenient to make the parts of the model work together, from data input to feature extraction, processing to final classification result output, to realize efficient and accurate short circuit risk assessment.

[0030] Preferably, as an improvement, the cross-entropy loss function is:

[0031] wherein, is the true label, is the short circuit risk probability predicted by the model.

[0032] Technical effects: through cross-entropy, it is convenient to effectively measure the difference between the predicted probability and the actual label, and to efficiently adjust the learning rate to accelerate model convergence. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a structural schematic diagram of the power battery short circuit risk quantitative evaluation method combining physical mechanism and deep learning. DETAILED DESCRIPTION

[0034] The following will be further described in detail through specific embodiments: The embodiments are basically as shown in the accompanying Figure 1 As shown: A quantitative assessment method for power battery short-circuit risk that integrates physical mechanisms and deep learning includes: S1. Select real vehicle operation data from the data cloud platform and perform data cleaning. The real vehicle operation data includes voltage, current, temperature, charge and discharge status, time, etc. Since the original data contains noise and missing data, these real vehicle operation data need to be cleaned, including removing outliers, filling missing data, and deleting irrelevant variables. The method of filling missing data is mean filling or interpolation. The formula for mean filling is:

[0035] in, is the value after filling, is the existing data of the same feature column, is the number of data points.

[0036] S2: Build an equivalent circuit model to describe the battery's electrical behavior and identify key parameters based on actual vehicle operation data. The equivalent circuit includes internal resistance, capacitance, and the battery's open-circuit voltage. Using battery voltage, current, and temperature data, parameter identification techniques are used to extract key battery parameters, including internal resistance, capacitance, open-circuit voltage, and battery state of health (SOH). The equivalent circuit model is:

[0037] Where R is the internal resistance, C is the capacitance, is the open circuit voltage of the battery, and I is the current. is the voltage across the capacitor at time t, which satisfies the dynamic equation and is expressed in Euler discrete format as follows: .

[0038] Over time The increase of equations, The equations form an overdetermined system.

[0039] S3, using recursive least squares method to fit the key parameters of the equivalent circuit model to obtain optimal parameters. In S3, using recursive least squares method to fit the key parameters of the equivalent circuit model includes: S31, linearize the current parameter estimation of the output equation; since the model is nonlinear, the output equation needs to be linearized at the current parameter estimation. Linearization is performed at .

[0040] The current parameter prediction function is:

[0041] in, is the open circuit voltage of the battery, R is the internal resistance, I is the current, The value of The state and parameters at the moment are estimated.

[0042] right exist Perform a first-order Taylor expansion at:

[0043] in, is the gradient vector:

[0044] S32, initialization parameters are , and initialize the covariance matrix as:

[0045] in, is the forgetting factor (usually ), is the identity matrix.

[0046] Calculate the prediction error and gain vector, update the covariance matrix based on the prediction error and gain vector, and fit the optimal parameters; The prediction error is:

[0047] in, is the current parameter prediction, is the battery voltage at the current moment.

[0048] The gain vector is:

[0049] in, is the gradient vector, For the forgetting factor, is the covariance matrix.

[0050] Update the covariance matrix to:

[0051] Update parameters:

[0052] The optimal parameters are fitted by recursive least squares method. .

[0053] S33, identifies the battery health status (SOH) through the battery's capacity decline and internal resistance increase. Battery health status (SOH) is an important indicator for evaluating battery aging and performance degradation. SOH is estimated using the following formula:

[0054] in, is the initial battery internal resistance, is the internal resistance of the battery at the current moment.

[0055] S4, the actual vehicle operation data after cleaning is combined with the optimal parameters and converted into a time series format; the time series format is:

[0056] Among them, V is voltage, I is current, T is temperature, S is charge and discharge time, R is internal resistance, C is capacitance, is the battery open circuit voltage, and SOH is the battery state of health. Time series data processing ensures that the data input has time correlation, making it suitable for subsequent deep learning model training.

[0057] Normalize and standardize the time series data so that the data range is scaled to [0, 1] to improve the training effect of the model.

[0058] The normalization formula is:

[0059] Normalization maps the data to the range [0,1].

[0060] The normalization formula is:

[0061] in, and Characteristics The mean and standard deviation of .

[0062] S5 divides the processed time series data into multiple time windows and generates a target label for each time window, indicating whether a short circuit has occurred. Specifically, each window contains a certain number of time steps of data, and a label is generated for each time window, indicating whether a battery short circuit has occurred within that time period.

[0063] Time window division includes: assuming that each time window contains T time step data, the original data is divided into N time windows, and the data in each window is:

[0064] in, is the input feature data at time t, and T is the window size.

[0065] Target label generation includes: target label is a binary classification label , indicating whether a short circuit occurs within the time window, Expressed as:

[0066] S6, build a temporal convolutional network model, which includes an input layer, multiple TCN layers, a global pooling layer, a fully connected layer, and an output layer; The temporal convolutional network model includes causal convolution and dilated convolution. Causal convolution is convenient for ensuring that future information is not used in prediction, that is, the output Only depends on and previous input.

[0067] Dilated convolution introduces a dilation factor during the convolution process. , expand the receptive field of view and enable the model to capture long-term dependencies. The expansion convolution operation includes:

[0068] in, is the convolution kernel size, is the weight of the convolution kernel, is the expansion factor.

[0069] In the temporal convolutional network model, the input layer receives time series data (shape is (T, d)); each TCN layer in multiple TCN layers captures local features and long-term dependencies in the time series through dilated convolution and causal convolution, and adopts residual connection training stability in each TCN layer; the global pooling layer performs global average pooling on the data output by the TCN layer to extract global feature representation; the fully connected layer maps the global features to the output space and outputs the final classification result; the output layer uses the Softmax activation function to directly output the binary classification result.

[0070] S7, using the cross entropy loss function to train the temporal convolutional network model. The cross entropy can effectively measure the difference between the predicted probability and the actual label, and efficiently adjust the learning rate to accelerate the convergence of the model; the cross entropy loss function is:

[0071] in, is the true label, is the short circuit risk probability predicted by the model.

[0072] S8 evaluates the trained temporal convolutional network model using accuracy, precision, and recall to ensure its predictive effectiveness on the test set. Accuracy measures the proportion of all correct predictions made by the model; precision measures the proportion of short circuits that were actually detected among the samples predicted by the model; and recall measures the proportion of correct predictions made by the model among the samples that actually were detected.

[0073] S9 uses the trained temporal convolutional network model to receive multi-dimensional data of the battery in real time and predict the short-circuit risk probability of the battery. When the risk probability exceeds the set threshold, an alarm is generated to indicate that the battery may have a short-circuit risk.

[0074] The real-time prediction formula is:

[0075] when (set threshold), the system generates an alert.

[0076] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A quantitative assessment method for power battery short-circuit risk that integrates physical mechanisms and deep learning is characterized by: include: S1, selects real vehicle operation data from the data cloud platform and performs data cleaning; S2, build an equivalent circuit model and identify key parameters based on actual vehicle operation data; S3, using recursive least squares method to fit the key parameters of the equivalent circuit model and obtain the optimal parameters; S4, combining the cleaned actual vehicle operation data with the optimal parameters and converting them into a time series format, and performing normalization and standardization on the time series data; S5, divides the processed time series data into multiple time windows, and generates a target label for each time window: whether a short circuit occurs; S6, build a temporal convolutional network model, which includes an input layer, multiple TCN layers, a global pooling layer, a fully connected layer, and an output layer; S7, training the temporal convolutional network model using a cross entropy loss function, adjusting the learning rate and accelerating convergence; S8, uses accuracy, precision and recall to evaluate the trained temporal convolutional network model; S9 uses the trained temporal convolutional network model to receive multi-dimensional data of the battery in real time, predicts the short circuit risk probability of the battery, and generates an alarm when the risk probability exceeds the set threshold.

2. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanism and deep learning according to claim 1 is characterized in that: The data cleaning includes removing outliers, filling missing data, and deleting irrelevant variables. The method of filling missing data is mean filling or interpolation. The formula of mean filling is: in, is the value after filling, is the existing data of the same feature column, is the number of data points.

3. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanism and deep learning according to claim 2 is characterized in that: The equivalent circuit model is: Where R is the internal resistance, C is the capacitance, is the open circuit voltage of the battery, and I is the current.

4. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanism and deep learning according to claim 1 is characterized in that: In S3, fitting the key parameters of the equivalent circuit model using the recursive least squares method includes: S31, linearize the current parameter estimates of the output equation; S32, initializing parameters and covariance matrix, calculating prediction error and gain vector, updating covariance matrix based on prediction error and gain vector, and fitting optimal parameters; S33, identifying the battery state of health (SOH) through the battery's capacity decay and internal resistance increase.

5. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanisms and deep learning according to claim 4 is characterized in that: The prediction error is: in, is the current parameter prediction, is the battery voltage at the current moment.

6. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanisms and deep learning according to claim 4 is characterized in that: The gain vector is: in, is the gradient vector, For the forgetting factor, is the covariance matrix.

7. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanism and deep learning according to claim 1 is characterized in that: The time series format is: Among them, V is voltage, I is current, T is temperature, S is charge and discharge time, R is internal resistance, C is capacitance, is the battery open circuit voltage, and SOH is the battery health state.

8. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanism and deep learning according to claim 1 is characterized in that: In S6, the temporal convolutional network model includes causal convolution and dilated convolution, and the dilated convolution operation includes: in, is the convolution kernel size, is the weight of the convolution kernel, is the expansion factor.

9. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanisms and deep learning according to claim 1 is characterized in that: In the temporal convolutional network model, the input layer receives time series data; each TCN layer in the multiple TCN layers captures local features and long-term dependencies in the time series through dilated convolution and causal convolution, and adopts residual connection training for stability; The global pooling layer performs global average pooling on the data output by the TCN layer; the fully connected layer maps the global features to the output space and outputs the final classification results; the output layer uses the Softmax activation function to directly output the binary classification results.

10. The method for quantitatively assessing power battery short-circuit risk by integrating physical mechanism and deep learning according to claim 1 is characterized in that: The cross entropy loss function is: in, is the true label, is the short circuit risk probability predicted by the model.

Images