Lithium metal battery fault diagnosis method based on battery constant-current constant-voltage charging process

Abstract

The invention discloses a lithium metal battery fault diagnosis method based on a battery constant-current and constant-voltage charging process, and the method comprises the steps: taking the voltage and current evolution of the constant-current and constant-voltage process as a core, and obtaining the voltage change value of a constant-current stage and the current change value of a constant-voltage stage in different time periods through the charging process; and constructing a training set, a verification set and a test set for training a machine learning model by taking the current characteristic sequence as a characteristic for evaluating a battery fault / normal state in a battery aging process, thereby establishing an association relationship between the current characteristic sequence and the battery fault state so as to diagnose the safety state of the battery. Through voltage evolution and current evolution measurement at different time stages in the charging process, electrochemical mechanism analysis is combined, the characteristic that the charging stage is strongly correlated with key fault mechanisms such as short circuit and side reaction in the battery is proposed to evaluate the safety state of the battery, the evaluation method has higher precision, and the evaluation accuracy is improved. And the method is suitable for lithium (sodium) metal, ion battery monomers and modules of various specifications.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery diagnosis, and relates to a lithium metal battery fault diagnosis method, and specifically to a battery fault diagnosis method using current and voltage characteristics during a battery charging process. Background Art

[0002] As the most widely used energy storage and conversion device, lithium-ion batteries have great application prospects in the fields of energy storage and power. However, the energy density of conventional graphite-based power storage lithium-ion batteries has reached its upper limit. Replacing the current graphite negative electrode with a lithium alloy negative electrode has become an important way to develop high energy density batteries.

[0003] As a representative of high-energy-density battery systems, lithium metal batteries have been widely used in a variety of important fields such as aerospace and national defense, and unmanned aerial vehicles with their rapid development in recent years. However, lithium metal batteries are prone to internal short circuits / micro short circuits / electrolyte decomposition during use, which can lead to accelerated capacity decay, capacity diving, fire and explosion. However, the research on the failure mechanism of lithium metal batteries is incomplete and there is a lack of effective fault detection methods. In recent years, the rapid development of machine learning and its application in experiments and data science have promoted its development in the field of battery diagnosis. Combining machine learning methods to extract features and establish diagnostic models to achieve rapid and accurate diagnosis of battery safety is a current research hotspot. However, it is difficult to achieve effective key feature extraction and fault diagnosis by directly establishing diagnostic models based on data. The lack of understanding of the mechanism and the inability to obtain complete charging data during actual use also make it difficult to apply lithium metal battery fault diagnosis models.

[0004] Machine learning-based methods mainly include traditional linear regression, support vector machine, decision neural tree and other machine learning algorithms and neural network algorithms. Neural network algorithms rely on a relatively smaller amount of data and have a wide range of applicability, making them more widely used in the field of battery diagnosis. However, they use the back-propagation mechanism to establish a nonlinear mapping of input and output to achieve accurate predictions that rely on highly correlated feature outputs. Currently, high-accuracy feature extraction cannot be achieved directly based on input of current and voltage data, and fault diagnosis based on sensor methods is difficult to be implanted non-destructively on a large scale. Therefore, the voltage and current data collected by BMS in real time are the most ideal data source for real-time fault diagnosis, but there is a lack of effective feature mining methods and model methods for partial charge and discharge data. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process. The method uses differential capacity and differential voltage to mine the key fault feature interval and correlation verification as the core method. Through current and voltage acquisition under different cycles, combined with differential capacity ICA and differential voltage DVA analysis, the key fault feature interval is identified, and the corresponding features are extracted to construct a diagnostic model to diagnose battery safety. The diagnostic method has higher accuracy, lower data cost and wider applicability, and the implementation method is simple, low cost and fast diagnostic speed.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process comprises the following steps:

[0008] Step 1: Perform charge-discharge cycle tests on lithium metal batteries under different charge-discharge regimes to obtain battery data under different charge-discharge cycles until the coulomb efficiency of the battery drops to a threshold value;

[0009] Step 2: Mark the battery data obtained in step 1, mark the current and voltage data of the corresponding cycle above the threshold as normal (NOR), and the corresponding label is 0, and mark the current and voltage data of the corresponding cycle below the threshold as failure (FAIL), and the corresponding label is 1;

[0010] Step 3: Perform differential capacity ICA and differential voltage DVA analysis on the battery data obtained after marking in step 2 to identify the critical voltage / current ranges corresponding to fault / normal cycle conditions;

[0011] Step 4: Extract the data features of voltage, current, and time corresponding to the key voltage / current range under fault / normal cycle conditions, set random factors, and randomly divide the data set into training set, validation set, and test set;

[0012] Step 5: Use the training set obtained in step 4 to build a machine learning model for training. Compare the predicted labels of the training output with the true labels in the training set. Update the neuron weights in the neural network through the back-propagation mechanism to obtain and save the parameter distribution with the minimum training error.

[0013] Step 6: Substitute the validation set into the machine learning model saved in step 5, monitor the loss on the validation set, and if the model error reduction value is less than the set value after the number of continuous iterations reaches the predetermined number of rounds, stop training in advance to avoid overfitting of the model, and save the model as the best model for diagnosing the battery safety status;

[0014] Step 7: Bring the test set into the best model saved in step 6 for diagnosis to obtain the accuracy results of the predicted faulty / normal battery cycle.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] 1. The present invention mines the high correlation intervals and features corresponding to the fault diagnosis of lithium metal based on the differential capacity ICA and differential voltage DVA of the charging process, which reduces the cumbersome and inaccurate feature identification, has high feature extraction efficiency and requires less data.

[0017] 2. The present invention only relies on the local voltage, current, capacity and other information collected by the BMS, and does not require additional equipment, so it is relatively easy to obtain.

[0018] 3. The present invention is based on the characteristics of the local current and voltage data of the charging and discharging process, such as the mean, time length, steepness, slope, and duration of constant voltage and low current. The characteristics are calculated only numerically, which is simple and efficient, and the input model iterative mapping is accurate.

[0019] 4. The present invention measures the voltage evolution and current evolution at different time stages of the charging process, combines it with electrochemical mechanism analysis, and proposes to use the characteristics of the charging stage that are strongly correlated with key failure mechanisms such as short circuits and side reactions in the battery to evaluate the battery safety status. The evaluation method has higher accuracy and is applicable to lithium (sodium) metal, ion battery cells, and modules of various specifications. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of a lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process.

[0021] Figure 2 It is the high correlation interval diagram of faults under ICA and DVA analysis in CCCV charging process.

[0022] Figure 3 It is the confusion matrix diagram of the fault diagnosis results of the validation set. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0024] The present invention provides a lithium metal battery fault diagnosis method based on the constant current and constant voltage charging process of the battery. The method is based on the voltage and current evolution of the constant current and constant voltage process. The charging process is used to obtain the voltage change in the constant current stage and the current change in the constant voltage stage in different time periods as the characteristics of evaluating the battery failure / normal state during the battery aging process. A training set, a validation set, and a test set are constructed for training machine learning models such as neural networks and Gaussian process regression, thereby establishing a correlation between the current feature sequence and the battery failure state to diagnose the battery safety state. Specifically, the following steps are included:

[0025] Step 1: Perform charge and discharge cycle tests on lithium metal batteries under different charge and discharge regimes to obtain battery data under different charge and discharge cycles until the coulombic efficiency of the battery drops to a threshold value (the threshold value is set to 98.5%, 98%, 97.5%, 97%, 97.5%, 96% or below, depending on the battery system, model and size, etc.).

[0026] In this step, the charge and discharge cycle test of the homemade soft-pack lithium metal battery includes: constant current and constant voltage test or multi-stage constant current and constant voltage test at different temperatures and different charge rates (same discharge rate); the collected battery data includes: current, voltage, time, capacity and other data during the charge and discharge process.

[0027] Step 2: Mark the battery data obtained in step 1, mark the current and voltage data of the corresponding cycle above the threshold as normal (NOR), the corresponding label is 0, and mark the current and voltage data of the corresponding cycle below the threshold as failure (FAIL), the corresponding label is 1.

[0028] Step 3: Perform differential capacity ICA and differential voltage DVA analysis on the battery data obtained after marking in step 2 to identify the critical voltage / current ranges corresponding to fault / normal cycle conditions.

[0029] In this step, the differential capacity ICA and differential voltage DVA analysis are based on the curves of current, voltage, capacity, and time during the battery charging process, and are achieved by constructing derivative curves. When the differential capacity voltage takes the value Vi, it is determined according to the differential capacity dQ / dV value corresponding to the point. When the dQ / dV value is small, the value of Vi is sparse, and when the dQ / dV value is large, the value of Vi is dense. In the constant voltage stage, the value rule of dQ / dI is the same. Then, the Savitzky-Golay filter is used to reduce noise and smooth the curve.

[0030] Step 3: Perform differential capacity ICA and differential voltage DVA analysis on the battery data obtained after marking in step 2 to identify the critical voltage / current ranges corresponding to fault / normal cycle conditions;

[0031] Step 4: Extract the data characteristics of voltage, current, and time corresponding to the key voltage / current range under fault / normal cycle conditions (such as mean, time length, steepness, slope, constant voltage and low current duration, etc.), set random factors, and randomly divide the data set into training set, test set, and validation set.

[0032] In this step, the data features include mean, time length, steepness, slope, and duration of constant voltage and low current. The specific extraction method is as follows: by removing outliers and applying cubic spline interpolation or linear interpolation methods, the voltage, current, and time data corresponding to the key voltage / current range under fault / normal cycle conditions are processed into equal-length sequences, and then the mean, time length, steepness, slope, duration of constant voltage and low current and other features are extracted.

[0033] Step 5: Use the training set samples obtained in step 4 to build a machine learning model for training. Compare the predicted labels of the training output with the true labels in the training set. Update the neuron weights in the neural network through the back-propagation mechanism to obtain and save the parameter distribution with the minimum training error.

[0034] In this step, the machine learning model adopts a convolutional neural network (CNN) structure, which includes 7 convolutional layers. The input layer receives the predetermined battery feature data. The convolution layer extracts features through multiple convolution kernels. Each convolution layer is followed by a ReLU activation function to increase nonlinearity. The number of channels in the convolution layer gradually increases from 8 to 64 to abstract features layer by layer. Features are extracted from the input data through convolution operations. The output layer contains a neuron for predicting the safety status of the battery (NOR corresponds to 0 / FAIL corresponds to 1). ReLU is selected as the activation function of the convolution layer, and StandardScaler is selected as the data normalization method. The optimization algorithm uses the Adam optimizer, the learning rate is set to 0.001, the number of training cycles is selected to 100 times, and the error function selects the cross entropy error.

[0035] Step 6: Substitute the validation set into the machine learning model saved in step 5, monitor the loss on the validation set, and if the model error reduction value is less than the set value after the number of continuous iterations reaches the predetermined number of rounds, stop training in advance to avoid model overfitting, and save the model as the best model for diagnosing the battery safety status.

[0036] In this step, the number of continuous iteration rounds reaches a predetermined number, which is generally set to 5 to 10, and the initial value of the reduction value of the current loss minus the loss of the previous round is set to zero.

[0037] Step 7: Bring the test set into the best model saved in step 6 for diagnosis to obtain the accuracy results of the predicted faulty / normal battery cycle.

[0038] Example:

[0039] In this embodiment, lithium metal battery data with a capacity greater than 10 Ah is used for constant current and constant voltage tests or multi-stage constant current and constant voltage tests at different temperatures and different charging rates (the discharge rate is the same). The battery data collected includes: current, voltage, time, capacity, temperature, etc. during the charge and discharge process. After the cycle reaches the set Coulomb efficiency threshold, it still cycles more than 5 times to ensure sufficient fault and safety battery cycle data. Differentiate the voltage-capacity data during the constant current process to obtain the differential capacity curve, and differentiate the current-capacity data during the constant voltage process to obtain the differential current curve, and then select the interval with obvious fault / normal characteristics. As Figure 2 shown, extract the voltage, current, and time data characteristics (such as voltage / current mean value, interval duration, voltage / current steepness, voltage / current slope, and the duration of low current during constant voltage Δt _lowDI / Dt etc.) corresponding to the interval (in the constant current stage (3.6V - 3.8V, 4.0 - 4.1V, 4.1V - 4.25V) and the constant voltage stage (Imin < I < 1 / 4Imax)) of the battery in this system. Mark the current and voltage data corresponding to the cycles higher than the threshold (98.5%) as normal (NOR), with the corresponding label being 0, and mark the current and voltage data corresponding to the cycles lower than the threshold as faulty (FAIL), with the corresponding label being 1. Randomly shuffle the dataset and divide it into training set: test set: validation set = 7:2:1 for training, validation, and testing.

[0040] Design a convolutional neural network (CNN) with 7 convolutional layers. The input layer receives the predetermined battery feature data. The convolutional layers extract features through multiple convolutional kernels. After each convolution layer, a ReLU activation function is connected to increase non-linearity. The number of channels in the convolutional layers gradually increases from 8 to 64 to abstract features layer by layer. Extract features from the input data through convolutional operations. The output layer contains a neuron for predicting the safety state of the battery (NOR corresponds to 0 / FAIL corresponds to 1). The activation function of the convolutional layer is selected as ReLU, and the data normalization method is selected as StandardScaler. The optimization algorithm uses the Adam optimizer, the learning rate is set to 0.001, the number of training loops is selected as 100 times, and the error function is selected as the cross-entropy error.

[0041] Substitute the organized training set samples into the above model for training. Through model training and the backpropagation algorithm, the hyperparameters are iteratively updated, and the hyperparameters with the minimum training error are saved. After obtaining the above optimal neural network, substitute the validation set into the saved hyperparameters and the neural network model, monitor its loss on the validation set. If the reduction value of the model error is less than the set value (1 / 5 of the previous round) under the condition that the continuous iteration rounds reach the predetermined number of rounds (set to 5), stop training in advance to avoid model overfitting, and save this model as the best model best_model for diagnosing the safety state of the battery.

[0042] The test set is brought into the model for diagnosis, and the accuracy results of predicting faulty / normal battery cycles are obtained as follows: Figure 3 As shown, Figure 3 The numbers on the diagonal line represent the number of normal cycles predicted as normal and faulty cycles predicted as faulty in the validation set. The overall diagnosis accuracy is (42+4) / 50=92%, which has good prediction accuracy.

Claims

1. A lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process, characterized in that The method comprises the following steps: Step 1: Perform charge-discharge cycle tests on lithium metal batteries under different charge-discharge regimes to obtain battery data under different charge-discharge cycles until the coulomb efficiency of the battery drops to a threshold value; Step 2: Mark the battery data obtained in step 1, mark the current and voltage data of the corresponding cycle above the threshold as normal NOR, with the corresponding label 0, and mark the current and voltage data of the corresponding cycle below the threshold as fault FAIL, with the corresponding label 1; Step 3: Perform differential capacity ICA and differential voltage DVA analysis on the battery data obtained after marking in step 2 to identify the critical voltage / current ranges corresponding to fault / normal cycle conditions; Step 4: Extract the data features of voltage, current, and time corresponding to the key voltage / current range under fault / normal cycle conditions, set random factors, and randomly divide the data set into training set, validation set, and test set; Step 5: Use the training set obtained in step 4 to build a machine learning model for training. Compare the predicted labels of the training output with the true labels in the training set. Update the neuron weights in the neural network through the back-propagation mechanism to obtain and save the parameter distribution with the minimum training error. Step 6: Substitute the validation set into the machine learning model saved in step 5, monitor the loss on the validation set, and if the model error reduction value is less than the set value after the number of continuous iterations reaches the predetermined number of rounds, stop training in advance to avoid overfitting of the model, and save the model as the best model for diagnosing the battery safety status; Step 7: Bring the test set into the best model saved in step 6 for diagnosis to obtain the accuracy results of the predicted faulty / normal battery cycle.

2. The lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process according to claim 1 is characterized in that In step 1, the charge-discharge cycle test includes: a constant current and constant voltage test at different temperatures and different charge rates or a multi-stage constant current and constant voltage test.

3. The lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process according to claim 1 is characterized in that In step 1, the battery data includes: current, voltage, time, and capacity data of the charging and discharging process.

4. The lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process according to claim 1 is characterized in that In step 2, the differential capacity ICA and differential voltage DVA analysis is based on the curves of current, voltage, capacity and time during the battery charging process and is achieved by constructing derivative curves.

5. The lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process according to claim 1 is characterized in that In step 4, the data features include mean, time length, steepness, slope, and duration of constant voltage and low current. The specific extraction method is as follows: by removing outliers and applying cubic spline interpolation or linear interpolation methods, the voltage, current, and time data corresponding to the key voltage / current range under fault / normal cycle conditions are processed into equal-length sequences, and then the mean, time length, steepness, slope, and duration of constant voltage and low current are extracted.

6. The lithium metal battery fault diagnosis method based on the battery constant current and constant voltage charging process according to claim 1 is characterized in that In step 5, the machine learning model adopts a convolutional neural network structure, including 7 convolution layers, the input layer receives the predetermined battery feature data, the convolution layer extracts features through multiple convolution kernels, and each convolution layer is followed by a ReLU activation function to increase nonlinearity; the number of channels of the convolution layer gradually increases from 8 to 64 to abstract features layer by layer; Extract features from input data through convolution operations; The output layer contains a neuron for predicting the safety status of the battery, NOR corresponds to 0 / FAIL corresponds to 1; the activation function of the convolutional layer selects ReLU, and the data normalization method selects StandardScaler; the optimization algorithm uses the Adam optimizer, the learning rate is set to 0.001, the number of training cycles is selected as 100, and the error function selects the cross entropy error.

7. The lithium metal battery fault diagnosis method based on the constant current and constant voltage charging process of the battery according to claim 1 A method for cutting, characterized in that In step 6, the number of iterations is set to 5 to 10. The initial value of the reduction of the current loss minus the previous round's loss is set to zero.

Images