Battery state evaluation method based on machine learning

Abstract

The invention discloses a battery state evaluation method based on machine learning, and the method comprises the steps: collecting battery data in a plurality of battery use environments, and marking a battery state; carrying out sample generation on the collected battery data by adopting an SMOTE algorithm based on low-rank matrix interpolation; forming an expanded data set by the generated data and the original data, training the data pair feature extraction model, and optimizing the training process of the feature extraction model by adopting a self-adaptive oscillation method in the training process; inputting the features corresponding to the battery state data extracted by the feature extraction model into a classifier model of a Riemann neural network classification algorithm based on fractional derivation to train the classifier model; and inputting the collected battery data of the to-be-evaluated battery into the trained feature extraction model for feature extraction, and inputting the extracted features into the trained classifier model for classification to obtain the state of the to-be-evaluated battery. According to the invention, the health state evaluation precision of the battery can be improved.

Description

Technical Field

[0001] The present invention relates to the field of battery management, and particularly to a method for evaluating battery state based on machine learning. Background Art

[0002] With the popularization of applications such as electronic devices and electric vehicles, as a core energy component, the accurate evaluation of the health state of the battery has become increasingly important. The health state of the battery directly affects the performance, safety and service life of the device. Especially in high-demand applications such as electric vehicles, the decline and failure of the battery may lead to extremely serious consequences. Therefore, how to evaluate the health state of the battery in real time and accurately has become a key problem to be solved urgently. At present, the evaluation methods of battery health state mainly rely on algorithms based on traditional physical models, methods based on empirical formulas, and machine learning methods based on data-driven. Traditional physical models need to consider multiple complex factors of the battery, such as temperature, load, voltage, etc. during the calculation process, usually relying on a large amount of experimental data and complex mathematical models, which makes the calculation process cumbersome and difficult to cope with the dynamic changes of the battery state. And the evaluation method based on empirical formula is often a simplified description of the battery health state, unable to effectively capture the complex non-linear relationship of battery performance changing with time. In contrast, the data-driven machine learning method can self-learn the relationship between the battery state and performance based on a large amount of data, and achieve accurate evaluation of battery health through model training and optimization.

[0003] The existing technologies have the following deficiencies:

[0004] 1. In the battery state evaluation task, the traditional SMOTE (Synthetic Minority Over-sampling Technique) algorithm relies on simple interpolation methods for sample augmentation. The generated samples are prone to losing the structural characteristics of the data, resulting in insufficient diversity of the training data and weak ability to capture battery state changes.

[0005] 2. When conventional neural networks process high-dimensional battery state data, they are prone to problems such as gradient disappearance or gradient explosion, which affect the training stability, resulting in insufficient extracted features to support the classification requirements of complex states. The fixed network parameter update method has poor adaptability to complex data distributions and is difficult to efficiently extract the key features of battery states.

[0006] 3. The classification model based on Euclidean space is difficult to fully capture the complex non-linear relationships between battery health state features, resulting in low classification accuracy for sub-healthy and unhealthy battery states. The static regularization method fails to dynamically adapt to the overfitting situation of the model and is prone to classification bias for small sample data or unbalanced data. Summary of the Invention

[0007] In view of the deficiencies of the prior art, the present invention proposes a battery state evaluation method based on machine learning, and the specific technical solution is as follows:

[0008] A battery state evaluation method based on machine learning, the method comprising the following steps:

[0009] Step 1: Collect battery data under various battery usage environments, store the battery data as a structured vector; and label the collected data with 3 categories: "battery healthy state", "battery sub-healthy", and "battery unhealthy state";

[0010] Step 2: Use the SMOTE algorithm based on low-rank matrix interpolation to perform low-rank matrix decomposition on the battery data collected in Step 1, extract the potential feature structure of the battery data, and generate samples; and together with the original data in Step 1, form an expanded data set with the newly generated data;

[0011] Step 3: Use the data in the expanded data set to train the feature extraction model, the feature extraction model is a fully connected neural network, and an adaptive oscillation method is used to optimize the training process of the feature extraction model during the training process;

[0012] Step 4: Input the features corresponding to the battery state data extracted by the feature extraction model into the classifier model for training of the classifier model; the classifier model is a Riemann neural network classification algorithm based on fractional-order derivative;

[0013] Step 5: Input the battery data of the battery to be evaluated collected into the trained feature extraction model for feature extraction, and then input the extracted features into the trained classifier model for classification to obtain the state of the battery to be evaluated.

[0014] Further, the second step includes the following sub-steps:

[0015] S2.1: Process the battery data collected in Step 1 through a low-rank matrix decomposition algorithm, convert the sparse and high-dimensional original data into a low-rank matrix representation, so as to extract the potential structural information in the data; during the low-rank matrix decomposition process, the structural regularization term constraint of the data needs to be followed;

[0016] S2.2: Select two rows of data from the feature matrix of the battery data and initialize the interpolation path;

[0017] S2.3: Based on the interpolation path, generate synthetic samples, and perform adaptive adjustment on each synthetic sample to make its distribution more consistent with the original data, obtaining optimized synthetic samples; and add the optimized generated samples to the original data set obtained in Step 1 to obtain an expanded data set.

[0018] Furthermore, the expression of low-rank matrix factorization in S2.1 is as follows:

[0019]

[0020] In the formula, is the left singular matrix; represents the spatial dimension; is a diagonal matrix, is the right singular matrix; k is the rank of the decomposition, which determines the low-dimensional space of the data;

[0021] The objective function constrained by the structural regularization term is:

[0022]

[0023] In the formula, represents the feature matrix X of the data c and the error of the low-rank approximation, ∥Σ c ∥ F is the Frobenius norm of Σ c , which is used to constrain the complexity of low-rank matrix factorization, prevent overfitting, and prompt the matrix factorization to maintain simplicity and structure; ∥∥ F represents the Frobenius norm; λ c is the parameter of the structural regularization term.

[0024] Furthermore, the interpolation path in S2.2 is expressed as:

[0025] γ c (X c,i ,X c,j )=β c X c,i +δ c α c β yce (X c,j -X c,i )

[0026] In the formula, X c,i and X c,j are the i-th row and j-th row data in the feature matrix X c of the data respectively, representing the features of different training samples. γ c (X c,i ,X c,j ) is the interpolation path between the two data points X c,i and X c,j , and this path describes the linear interpolation process from X c,i to X c,j ; α c ∈[0,1] is the interpolation factor, representing from X c,i to Xc,j Interpolation ratio; when α c = 0, it represents the data point X c,i ; when α c = 1, it represents the data point X c,j ; δ c is a correction term based on the error metric; β c is a dynamic adjustment factor; β yce is a data distribution adjustment factor;

[0027]

[0028] In the formula, Var(X c ) represents the variance of the current data set; is the mean of the current data set; η ckg is a data distribution weight factor;

[0029]

[0030] In the formula, λ vces is a correction term adjustment coefficient, ∈ c is a small constant to prevent the denominator from being zero, ∥ΔX c,i,j ∥ 2 is the Euclidean distance between the data points X c,i and X c,j ; ∥∥ is the L2 norm, and the calculation method is the same as the Euclidean distance;

[0031]

[0032] In the formula, t c is the time interval; γ ces is an expansion coefficient related to the data difference, θ c is a constant related to the change rate of the battery state.

[0033] Furthermore, in the S2.3, the generated synthetic sample X c,k is expressed as:

[0034] X c,k = γ c (X c,i , X c,j )

[0035] The adaptively adjusted sample is expressed as:

[0036]

[0037] Among them, is the mean of the current data set; is an adaptive adjustment factor; is the optimized generated sample; is the coefficient for adaptive adjustment, is the Euclidean distance between the generated sample and the mean of the current dataset.

[0038] Further, the third step includes the following sub-steps:

[0039] S3.1: Initialize the parameters of the fully connected neural network, and at the same time initialize the hyperparameters of the dynamic adaptive oscillation, including the initial phase and the initial amplitude of the oscillation;

[0040] S3.2: Calculate the oscillation frequency of each parameter based on the current loss function; calculate the phase of the current iteration oscillation through the success rate of the previous update and the interaction strength between parameters; adaptively adjust the amplitude size according to the effect of parameter updates in the past few iterations. If the update of a certain parameter continuously leads to a reduction in loss, increase its amplitude; otherwise, decrease the amplitude;

[0041] S3.3: Calculate the update value of each parameter of the fully connected neural network;

[0042] S3.4: Repeat steps S3.2 and S3.3 until the preset stop iteration condition is met, and complete the training of the feature extraction model.

[0043] Further, the calculation formula of the oscillation frequency is as follows:

[0044]

[0045] In the formula, β p is the historical gradient weight factor, which controls the influence of the historical gradient in the current frequency adjustment; ω p is the oscillation frequency, is the oscillation frequency of the t-th iteration; is the weight of the neural network in the (t - 1)-th iteration; is the weight of the neural network in the k-th iteration; Lp is the loss function of the neural network; is the partial derivative symbol; ω 0 is the basic oscillation frequency; α p is the hyperparameter for adjusting the oscillation responsiveness;

[0046] The phase of the oscillation is expressed as:

[0047]

[0048] In the formula, is the phase of the oscillation in the t-th iteration; is the phase of the oscillation in the (t - 1)-th iteration; δ p is the phase adjustment factor; tanh() is the hyperbolic tangent function; λ p is the hyperparameter for adjusting the phase sensitivity;

[0049] The amplitude is expressed as:

[0050]

[0051] wherein, is the amplitude of the oscillation at the t-th iteration; is the amplitude of the oscillation at the (t - 1)-th iteration; γ p is the amplitude adjustment coefficient;

[0052] The parameter update of the fully connected neural network is expressed as:

[0053]

[0054] wherein, is the weight of the neural network at the t-th iteration; is the bias of the neural network at the t-th iteration; is the weight of the neural network at the (t - 1)-th iteration; is the bias of the neural network at the (t - 1)-th iteration; t * is the current iteration number of the neural network.

[0055] Furthermore, in the fourth step, the training process of the Riemann neural network classification algorithm based on fractional-order derivative is as follows:

[0056] S4.1: Initialize the parameters of the Riemann neural network:

[0057]

[0058] wherein, represents the initial value of the network weight, which is initialized with a standard normal distribution, n qin is the number of nodes in the input layer;

[0059] S4.2: In the forward propagation stage, the feature vector after feature extraction of the input is converted into a point on the Riemann manifold through the network layer, and corresponding manifold operations are performed on each layer; the forward propagation is specifically expressed as:

[0060]

[0061] wherein, is the output of the l-th layer of the Riemann neural network, is the output of the (l + 1)-th layer of the Riemann neural network, and are the weight and bias of the l-th layer of the Riemann neural network respectively, and Sig() is the Sigmoid non-linear activation function;

[0062] S4.3: Based on the data points on the Riemannian manifold, the cross-entropy loss function with a regularization term is used to calculate the network loss, expressed as:

[0063]

[0064] In the formula, L q is the loss function, y qc is the one-hot encoding of the true label of the sample, is the probability predicted by the Riemannian neural network model, C is the total number of classes, and R q (W) is the regularization term.

[0065] The calculation method of is expressed as:

[0066]

[0067] In the formula, z qc is the raw output score of the model for class c, which is transformed into the predicted probability through the Softmax function z qk is the output score of the classifier model for class k;

[0068]

[0069] In the formula, λ q (t) is the regularization coefficient that is dynamically adjusted according to the training progress, is the weight of the l-th layer, represents the sum of the squares of the weights;

[0070] The regularization coefficient λ q (t) is adjusted according to the current overfitting situation of the model, and the adjustment method is expressed as:

[0071]

[0072] In the formula, β, k s and m 0 are hyperparameters that control the height, steepness, and midpoint position of the curve respectively

[0073] S4.4: Use fractional-order differentiation to calculate the gradient of the network parameters to adapt to the distribution characteristics of the data on the manifold. The differentiation method is expressed as:

[0074]

[0075] In the formula, ΔW q is the update amount of the weight, η lm is the learning rate, represents the 0.5-order partial derivative of the loss function L q with respect to;

[0076] The fractional derivative is expressed as:

[0077]

[0078] where Γ is the Gamma function.

[0079] Update the network parameters according to the gradient calculated by backpropagation. The update method is expressed as:

[0080]

[0081] where is the updated weight.

[0082] Furthermore, the loss function of the fully connected neural network adopts cross-entropy loss, which is calculated by the preset Softmax function for the output of the last layer of the fully connected neural network.

[0083] Furthermore, the battery data at least includes battery capacity, charging time, discharge rate, number of cycles, temperature fluctuation, voltage fluctuation, internal resistance change, capacity attenuation rate, battery leakage current, and charging efficiency.

[0084] The beneficial effects of the present invention are as follows:

[0085] 1. Based on the traditional SMOTE algorithm, the present invention uses low-rank matrix decomposition to extract the potential feature structure of the data, avoiding the unnaturalness of the data caused by simple interpolation, and being able to better maintain the structural consistency of the data. Especially in the case of high-dimensional sparse data such as battery state assessment, it avoids the unnatural data generated by simple interpolation and improves the quality of sample generation.

[0086] 2. The traditional SMOTE algorithm uses a fixed interpolation path. The present invention dynamically adjusts the interpolation path, considering the rate of change of the battery state, to ensure that the generated new samples can more truly reflect the state change of the battery, enhancing the diversity and accuracy of data augmentation.

[0087] 3. The present invention uses an oscillation mechanism to optimize the neural network training process, which can smoothly explore local extrema in high-dimensional space, avoiding the problems of gradient disappearance and explosion in traditional neural networks, thereby improving the stability and training effect of the model.

[0088] 4. The present invention processes the data on the Riemannian manifold through fractional calculus, enabling the neural network to more finely adjust the weight update in the learning process, improving the adaptability and classification accuracy for the complex battery health state data structure. Description of the Drawings

[0089] Figure 1It is a flowchart of the battery state evaluation method based on machine learning of the present invention.

[0090] Figure 2 It is a flowchart for training the feature extraction model. Specific embodiments

[0091] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become more apparent. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0092] As Figure 1 shown, the battery state evaluation method based on machine learning of the present invention includes the following steps:

[0093] Step 1: Collect battery data under various battery usage environments, store the battery data as a structured vector; and label the collected data with 3 categories of "battery healthy state", "battery sub-healthy", and "battery unhealthy state".

[0094] Here, the various battery usage environments include charging speed, temperature, battery cycle count, etc. In one embodiment, the battery data includes:

[0095] Battery capacity ax1, representing the maximum energy storage capacity of the battery;

[0096] Charging time ax2, representing the time required for the battery to charge from fully discharged to full;

[0097] Discharge rate ax3, representing the discharge speed of the battery;

[0098] Cycle count ax4, representing the charge and discharge cycle count of the battery;

[0099] Temperature fluctuation ax5, representing the average temperature change during battery use;

[0100] Voltage fluctuation ax6, representing the voltage fluctuation range during battery use;

[0101] Internal resistance change ax7, representing the change in the internal resistance of the battery;

[0102] Capacity attenuation rate ax8, representing the attenuation rate of the battery capacity over time;

[0103] Battery leakage current ax9, representing the leakage current of the battery in the non-working state;

[0104] Charging efficiency ax10, representing the efficiency of converting electrical energy into chemical energy during charging.

[0105] This is just one data format and type of battery data. In actual applications, the attributes of the data usually exceed 10, and the number of data attributes may reach dozens or even hundreds.

[0106] Step 2: Use the SMOTE algorithm based on low-rank matrix interpolation to perform low-rank matrix decomposition on the battery data collected in Step 1, effectively extract the potential feature structure of the battery data, and generate samples; then combine the newly generated data with the original data in Step 1 to form an expanded dataset.

[0107] Step 2 specifically includes the following sub-steps:

[0108] S2.1: Process the battery data collected in Step 1 through the low-rank matrix decomposition algorithm, convert the sparse and high-dimensional original data into a low-rank matrix representation, so as to extract the potential structural information in the data; during the low-rank matrix decomposition process, the constraint of the data structure regularization term needs to be followed.

[0109] Specifically, let the training dataset be {X c}, which is an m c ×n c matrix, X c represents the feature matrix of the data, m c represents the number of samples, n c represents the feature dimension of each sample. Through low-rank matrix decomposition, the data can be expressed as the product of two matrices, which is expressed as:

[0110]

[0111] In the formula, is the left singular matrix; represents the spatial dimension; is the diagonal matrix, is the right singular matrix; k is the rank of the decomposition, which determines the low-dimensional space of the data.

[0112] During the low-rank matrix decomposition process, the constraint of the data structure regularization term needs to be followed to better control the sparsity of the data and ensure that the low-rank approximation after matrix decomposition is more in line with the actual structure of the data. The objective function of the constraint is expressed as:

[0113]

[0114] In the formula, represents the error between the feature matrix X c of the data and the low-rank approximation, ∥Σ c ∥ F is Σ cThe Frobenius norm is used to constrain the complexity of low-rank matrix factorization, prevent overfitting, and encourage the matrix factorization to remain concise and structured; ∥∥ F represents the Frobenius norm; λ c is the parameter of the structural regularization term. Preferably, λ c is set to 0.3.

[0115] The principal components in the data are extracted through matrix factorization, representing the low-rank structure of the training data. The singular values in the diagonal matrix Σ c reveal the dominant features of the data, and the left singular matrix U c and the right singular matrix capture the spatial and feature representations of the data respectively.

[0116] S2.2: Select two rows of data from the feature matrix of the battery data and initialize the interpolation path, expressed as:

[0117] γ c (X c,i , X c,j ) = β c X c,i + δ c α c β yce (X c,j - X c,i )

[0118] In the formula, X c,i and X c,j are the i-th row and the j-th row of data in the feature matrix X c of the data respectively, representing the features of different training samples. γ c (X c,i , X c,j ) is the interpolation path between the two data points X c,i and X c,j , and this path describes the linear interpolation process from X c,i to X c,j ; α c ∈[0, 1] is the interpolation factor, representing the interpolation ratio from X c,i to X c,j ; when α c = 0, it represents the data point X c,i ; when α c = 1, it represents the data point X c,j ; δ c is the correction term based on the error metric; β c is the dynamic adjustment factor; β yce is the data distribution adjustment factor, calculated based on the new data distribution, expressed as:

[0119]

[0120] Wherein, Var(X c ) represents the variance of the current data set; is the mean of the current data set; η ckg is the data distribution weight factor. Preferably, η ckg is set to 0.1.

[0121] The correction term δ based on the error metric c dynamically corrects the generation of the interpolation path according to the distance between data points, ensuring that the generated synthetic samples can more accurately reflect the state of the battery. The calculation method is expressed as:

[0122]

[0123] Wherein, λ vces is the correction term adjustment coefficient, ∈ c is a small constant to prevent the denominator from being zero, ∥ΔX c,i,j ∥ 2 is the Euclidean distance between the data points X c,i to X c,j ; ∥∥ is the L2 norm, and the calculation method is the same as the Euclidean distance. Preferably, λ vces is set to 2, ∈ c is set to 0.001.

[0124] Based on the data representation obtained by low-rank matrix factorization, the algorithm initializes the interpolation path by selecting appropriate data points. The distance between each pair of data points is calculated based on the distance metric in the low-rank matrix, rather than the traditional Euclidean distance. When initializing the path, the algorithm preferentially selects those points located at the edge of the data distribution to enhance the diversity of data augmentation.

[0125] Through this interpolation path, new data points can be generated at any position between two points, and then new synthetic samples can be generated for data augmentation.

[0126] When generating new samples, the traditional SMOTE algorithm usually adopts a fixed interpolation path. However, in actual data, the state data of the battery may have continuity and time series. Therefore, the present invention adopts a dynamic interpolation path optimization mechanism, which dynamically adjusts the path during the interpolation process by real-time evaluating the state change of the battery, ensuring that the generated new samples can accurately reflect the state change of the battery. Specifically, the dynamic adjustment factor β c is calculated related to time and depends on the state change rate of the battery. For example, when the state of the battery changes rapidly, β c will increase, indicating that the generated new data points should reflect a greater change; when the state of the battery is relatively stable, β c is smaller, and the calculation method is expressed as:

[0127]

[0128] wherein, t c is the time interval, preset artificially; γ ces is an expansion coefficient related to the data difference, and θ c is a constant related to the change rate of the battery state, preset artificially.

[0129] S2.3: Based on the interpolation path, generate synthetic samples, and perform adaptive adjustment on each synthetic sample to make its distribution more consistent with the original data, obtaining optimized synthetic samples; and add the optimized generated samples to the original dataset obtained in Step 1 to obtain an augmented dataset.

[0130] After dynamic path optimization, generate synthetic samples. By performing interpolation in the low-rank matrix space and combining the adjustment of the interpolation path, new data points are generated. Each new sample is obtained by performing interpolation calculations between the original samples. Through dynamic path optimization, ensure its consistency with the original data while having sufficient diversity. Let the newly generated synthetic sample be X c,k , which is generated from the interpolation results of X c,i and X c,j after dynamic path optimization, and is expressed as:

[0131] X c,k = γ c (X c,i , X c,j )

[0132] After the new samples are generated, to ensure the quality of the generated samples, perform adaptive adjustment on each synthetic sample to make its distribution more consistent with the original data. Specifically, optimize the new samples through an adaptive adjustment factor, and the optimization method is expressed as:

[0133]

[0134] wherein, is the mean of the current dataset; is the adaptive adjustment factor; is the optimized generated sample.

[0135] The adaptive adjustment factor is adjusted according to the difference between the generated samples and the overall data distribution, so as to ensure that the generated synthetic samples are more consistent with the original data and avoid excessive deviation from the data distribution. The calculation method is expressed as:

[0136]

[0137] wherein, is the coefficient for adaptive adjustment, is the Euclidean distance between the generated sample and the mean of the current dataset.

[0138] The newly generated sample will be added to the training dataset. Let the augmented dataset be X c,aug , which is the synthesis result of the original dataset and the newly generated sample, expressed as:

[0139] X c,aug ={X c} ∪ {X c,k}

[0140] In the formula, {X c,k} represents the newly generated synthetic sample set.

[0141] Step 3: Train the feature extraction model using the data in the augmented dataset. The feature extraction model is a fully connected neural network, and an adaptive oscillation method is adopted to optimize the training process of the feature extraction model during training.

[0142] In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention adopts an algorithm based on dynamic adaptive oscillation to optimize the training process of the feature extraction model. By simulating the nonlinear oscillation behavior in physical phenomena, the algorithm can effectively explore and utilize local extreme values in the high-dimensional parameter space, thereby achieving the purpose of optimizing the neural network.

[0143] Step 3 specifically includes the following sub-steps:

[0144] S3.1: Initialize the parameters of the fully connected neural network and simultaneously initialize the hyperparameters of the dynamic adaptive oscillation, including the initial phase and the initial amplitude of the oscillation; the fully connected layer has 6 layers, and the number of nodes in each layer is 128, that is, this layer contains 128 neurons;

[0145] In one embodiment, the initialization method is expressed as:

[0146]

[0147] φ 0p = 0

[0148] A 0p = 1

[0149] In the formula, ~ means following a specific distribution; represents a normal distribution with a mean of 0 and a variance of ; represents a normal distribution; Represents the initial variance of the fully connected neural network; W 0p Is the initial weight of the fully connected neural network, b 0p Is the initial bias of the fully connected neural network; φ 0p Is the initial phase of the oscillation; A 0p Is the initial amplitude of the oscillation. Preferably, Is set to 0.01.

[0150] S3.2: Calculate the oscillation frequency of each parameter based on the current loss function; Calculate the phase of the current iteration oscillation through the success rate of the previous update and the interaction strength between parameters; According to the effect of parameter updates in the past few iterations, adaptively adjust the amplitude size. If the update of a certain parameter continuously leads to a reduction in loss, increase its amplitude; otherwise, decrease the amplitude.

[0151] Among them, the adjustment of the oscillation frequency depends on the local curvature estimation of the loss surface, and is adjusted using historical gradient information to make the parameter update smoother and enable more effective search in the parameter space. The adjustment method is expressed as:

[0152]

[0153] In the formula, β p Is the historical gradient weight factor, which controls the influence of the historical gradient in the current frequency adjustment; ω p Is the oscillation frequency, Is the oscillation frequency of the t-th iteration; Is the weight of the neural network in the (t - 1)-th iteration; Is the weight of the neural network in the k-th iteration; Lp is the loss function of the neural network; Is the partial derivative symbol; ω 0 Is the basic oscillation frequency; α p Is the hyperparameter that adjusts the oscillation responsiveness. Preferably, α p Is set to 5, β p Is set to 0.3; The loss function of the neural network adopts cross-entropy loss, which is calculated through the preset Softmax function for the output of the last layer of the neural network.

[0154] Among them, the update of the phase depends on the effect of the previous parameter update, and is used to simulate the delay effect of causal relationships. The update method is expressed as:

[0155]

[0156] In the formula, φ p Is the phase of the oscillation, Is the phase of the oscillation in the t-th iteration; Is the phase of the oscillation in the (t - 1)-th iteration; δ pis the phase adjustment factor; tanh() is the hyperbolic tangent function, which can limit the adjustment range of the phase and avoid instability of the algorithm caused by excessive adjustment; λ p is the hyperparameter for adjusting the phase sensitivity. Preferably, δ p is set to 0.1, and λ p is set to 0.3.

[0157] Among them, the adaptive adjustment method of the amplitude is expressed as:

[0158]

[0159] In the formula, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the t-th iteration; is the amplitude of the oscillation at the (t - 1)-th iteration; γ p is the amplitude adjustment coefficient. Preferably, γ p is set to 0.95.

[0160] S3.3: Calculate the updated values of the parameters of each fully connected neural network.

[0161] Specifically, combined with the oscillation behavior for parameter update, the sine function and cosine function are used to simulate the oscillation behavior, allowing the parameters to perform periodic exploration in the gradient direction of the loss function. The update method is expressed as:

[0162]

[0163] In the formula, is the weight of the neural network at the t-th iteration; is the bias of the neural network at the t-th iteration; is the weight of the neural network at the (t - 1)-th iteration; is the bias of the neural network at the (t - 1)-th iteration; t * is the current iteration number of the neural network.

[0164] S3.4: Repeat steps S3.2 and S3.3 until the preset stop iteration condition is satisfied, and complete the training of the feature extraction model.

[0165] In one embodiment, the preset stop iteration condition is to reach the preset maximum iteration number. Preferably, the preset maximum iteration number is set to 1000 times.

[0166] Step Four: Input the features corresponding to the battery state data extracted by the feature extraction model into the classifier model for training the classifier model; the classifier model is the Riemann neural network classification algorithm based on fractional order derivative.

[0167] The present invention adopts a Riemann neural network classification algorithm based on fractional-order derivatives, projects the battery state data after feature extraction onto points in a non-Euclidean space, and further processes the data embedded on the Riemann manifold. Based on the traditional Riemann neural network algorithm, the present invention uses fractional-order calculus to process the data on the Riemann manifold, enabling the network to more finely adjust the weight update in the learning process and improving the adaptability and classification accuracy for complex data structures.

[0168] The training process of the Riemann neural network classification algorithm based on fractional-order derivatives is as follows:

[0169] (1) Initialize the parameters of the Riemann neural network, and the initialization method is expressed as:

[0170]

[0171] In the formula, represents the initial value of the network weights, which is initialized using a standard normal distribution, and n qin is the number of nodes in the input layer.

[0172] (2) In the forward propagation stage, the input feature vector after feature extraction is converted into points on the Riemann manifold through the network layer, and corresponding manifold operations are performed on each layer. Then the forward propagation method is expressed as:

[0173]

[0174] In the formula, is the output of the l-th layer of the Riemann neural network, is the output of the (l + 1)-th layer of the Riemann neural network, and are the weights and biases of the l-th layer of the Riemann neural network respectively, and Sig() is the Sigmoid non-linear activation function.

[0175] (3) According to the data points on the Riemann manifold, calculate the network loss using a cross-entropy loss function based on a regularization term, which is expressed as:

[0176]

[0177] In the formula, L q is the loss function, y qc is the one-hot encoding of the true label of the sample, is the probability predicted by the Riemann neural network model, C is the total number of categories, and R q (W) is the regularization term.

[0178] The calculation method of

[0179]

[0180] where z qc is the original output score of the model for class c, which is transformed into a predicted probability through the Softmax function z qk is the output score of the classifier model for class k.

[0181] The regularization term R q (W) is calculated as follows:

[0182]

[0183] where λ q (t) is a regularization coefficient that is dynamically adjusted according to the training progress, is the weight of the l-th layer, represents the sum of the squares of the weights.

[0184] The regularization coefficient λ q (t) is adjusted based on the current overfitting situation of the model, and the adjustment method is as follows:

[0185]

[0186] where β, k s and m 0 are hyperparameters that control the height, steepness, and midpoint position of the curve respectively. Preferably, β, k s and m 0 are set to 1, 3, and 0.5 respectively.

[0187] (4) Use fractional-order differentiation to calculate the gradient of the network parameters to adapt to the distribution characteristics of the data on the manifold. The differentiation method is as follows:

[0188]

[0189] where ΔW q is the update amount of the weight, η lm is the learning rate, represents the 0.5-order partial derivative of the loss function L q Preferably, η lm is set to 0.01.

[0190] The method of fractional-order differentiation is as follows:

[0191]

[0192] where Γ is the Gamma function.

[0193] Update the network parameters according to the gradient calculated by backpropagation. The update method is as follows:

[0194]

[0195] Wherein, is the updated weight.

[0196] When the iteration meets the stop iteration condition, it means that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0197] Step Five: Input the battery data of the battery to be evaluated collected into the trained feature extraction model for feature processing, and then input the processed features into the trained classifier model for classification to obtain the state of the battery to be evaluated.

[0198] Those of ordinary skill in the art can understand that the above are only preferred examples of the invention and are not used to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing examples, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, etc. made within the spirit and principle of the invention shall be included within the protection scope of the invention.

Claims

1. A battery status assessment method based on machine learning, characterized in that: The method comprises the following steps: Step 1: Collect battery data under various battery usage environments and store the battery data as a structured vector; and label the collected data into three categories: "battery health status", "battery sub-health" and "battery unhealth status"; Step 2: Use the SMOTE algorithm based on low-rank matrix interpolation to perform low-rank matrix decomposition on the battery data collected in step 1, extract the potential feature structure of the battery data, and generate samples; and combine the newly generated data with the original data in step 1 to form an expanded data set; Step 3: Use the data in the expanded data set to train the feature extraction model, which is a fully connected neural network. During the training process, an adaptive oscillation method is used to optimize the training process of the feature extraction model. Step 4: Input the features corresponding to the battery status data extracted by the feature extraction model into the classifier model to train the classifier model; the classifier model is a Riemann neural network classification algorithm based on fractional order derivation; Step 5: Input the collected battery data of the battery to be evaluated into the trained feature extraction model for feature extraction, and then input the extracted features into the trained classifier model for classification to obtain the state of the battery to be evaluated.

2. The battery status assessment method based on machine learning according to claim 1, characterized in that: The step 2 includes the following sub-steps: S2.1: The battery data collected in step 1 is processed by a low-rank matrix decomposition algorithm, and the sparse and high-dimensional original data is converted into a low-rank matrix representation, so as to extract the potential structural information in the data; during the low-rank matrix decomposition process, the structural regularization term constraints of the data must be followed; S2.2: Select two rows of data from the feature matrix of the battery data and initialize the interpolation path; S2.3: Based on the interpolation path, synthetic samples are generated, and each synthetic sample is adaptively adjusted to make it more consistent with the distribution of the original data, so as to obtain an optimized synthetic sample; and the optimized generated sample is added to the original data set obtained in step 1 to obtain an expanded data set.

3. The battery status assessment method based on machine learning according to claim 2, characterized in that: The expression of the low-rank matrix decomposition in S2.1 is: In the formula, is a left singular matrix; Represents spatial dimension; is a diagonal matrix, is a right singular matrix; k is the rank of decomposition, which determines the low-dimensional space of the data; The objective function of the structural regularization term constraint is: In the formula, The feature matrix X representing the data c Error with low-rank approximation, ∥Σ c ∥ F is Σ c The Frobenius norm is used to constrain the complexity of low-rank matrix decomposition, prevent overfitting, and keep the matrix decomposition simple and structured; ∥∥ F represents the Frobenius norm; λ c is the structural regularization parameter.

4. The battery status assessment method based on machine learning according to claim 3, characterized in that: The interpolation path in S2.2 is expressed as: c c (X c,i ,X c,j )=β c X c,i +d c a c b yce (X c,j -X c,i ) Where, X c,i and X c,j are the feature matrices X of the data respectively. c The i-th and j-th rows of data in represent the characteristics of different training samples, γ c (X c,i ,X c,j ) is the data point X c,i and X c,j The interpolation path between two data points describes the path from X c,i To X c,j The linear interpolation process between α c ∈[0,1] is the interpolation factor, which means from X c,i To X c,j The interpolation ratio; when α c =0, indicating that the data point X c,i ; When α c =1, indicating that the data point X c,j ; δ c is a correction term based on error measurement; β c is the dynamic adjustment factor; β yce is the data distribution adjustment factor; In the formula, Var(X c ) represents the variance of the current data set; is the mean of the current data set; η ckg is the data distribution weight factor; In the formula, λ vces is the correction term adjustment coefficient, ∈ c To prevent a small constant with a denominator of zero, ∥ΔX c,i,j ∥2 is the data point X c,i To X c,j The Euclidean distance between them; ∥∥ is the L2 norm, which is calculated in the same way as the Euclidean distance; In the formula, t c is the time interval; γ ces is the expansion coefficient related to the data difference, θ c is a constant related to the rate of change of the battery state.

5. The battery status assessment method based on machine learning according to claim 4, characterized in that: In S2.3, the generated synthetic sample X c,k It is expressed as: X c,k =c c (X c,i ,X c,j ) The adaptively adjusted sample is expressed as: in, is the mean of the current data set; is the adaptive adjustment factor; is the generated sample after optimization; is the coefficient of adaptive adjustment, is the Euclidean distance between the generated sample and the mean of the current dataset.

6. The battery status assessment method based on machine learning according to claim 5, characterized in that: The step three includes the following sub-steps: S3.1: Initialize the parameters of the fully connected neural network and the hyperparameters of the dynamic adaptive oscillation, including the initial phase and initial amplitude of the oscillation; S3.2: Calculate the oscillation frequency of each parameter based on the current loss function; calculate the phase of the current iteration oscillation based on the success rate of the previous update and the interaction strength between the parameters; adaptively adjust the amplitude based on the effect of the parameter update in the past few iterations, and increase its amplitude if the update of a certain parameter continues to reduce the loss; On the contrary, the amplitude is reduced; S3.3: Calculate the updated values ​​of the parameters of each fully connected neural network; S3.4: Repeat steps S3.2 and S3.3 until the preset stop iteration condition is met, completing the training of the feature extraction model.

7. The battery status assessment method based on machine learning according to claim 6, characterized in that: The calculation formula of the oscillation frequency is as follows: In the formula, β p is the historical gradient weight factor, which controls the influence of the historical gradient on the current frequency adjustment; ω p is the oscillation frequency, is the oscillation frequency of the tth iteration; is the weight of the neural network at the t-1th iteration; is the weight of the neural network at the kth iteration; Lp is the loss function of the neural network; is the sign of the partial derivative; ω0 is the fundamental oscillation frequency; α p is a hyperparameter that adjusts the oscillation responsiveness; The phase of the oscillation is expressed as: In the formula, is the phase of the oscillation at the tth iteration; is the phase of the oscillation at the t-1th iteration; δ p is the phase adjustment factor; tanh() is the hyperbolic tangent function; λ p is a hyperparameter that adjusts phase sensitivity; The amplitude is expressed as: In the formula, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment factor; The parameter update of the fully connected neural network is expressed as: In the formula, is the weight of the neural network at the tth iteration; is the bias of the neural network at the tth iteration; is the weight of the neural network at the t-1th iteration; is the bias of the neural network at the t-1th iteration; t * is the current iteration number of the neural network.

8. The battery status assessment method based on machine learning according to claim 7, characterized in that: In step 4, the training process of the Riemann neural network classification algorithm based on fractional order derivation is as follows: S4.1: Initialize the parameters of the Riemann neural network: In the formula, Represents the initial value of the network weight, which is initialized using a standard normal distribution, n qin is the number of nodes in the input layer; S4.2: In the forward propagation stage, the feature vector after feature extraction from the input is converted into a point on the Riemann manifold through the network layer, and each layer performs the corresponding manifold operation; the forward propagation is specifically expressed as: In the formula, is the output of the lth layer of the Riemann neural network, is the output of the l+1th layer of the Riemann neural network, and are the weight and bias of the lth layer of the Riemann neural network, and Sig() is the Sigmoid nonlinear activation function; S4.3: According to the data points on the Riemann manifold, the network loss is calculated using the cross entropy loss function based on the regularization term, expressed as: Where, L q is the loss function, y qc is the one-hot encoding of the true label of the sample, is the probability predicted by the Riemann neural network model, C is the total number of categories, R q (W) is the regularization term. The calculation method is expressed as: In the formula, z qc It is the original output score of the model for category c, which is converted into predicted probability through the Softmax function z qk is the output score of the classifier model for category k; In the formula, λ q (t) is the regularization coefficient that is dynamically adjusted as the training progresses. is the weight of the lth layer, represents the sum of squares of weights; Regularization coefficient λ q The adjustment of (t) is based on the current overfitting condition of the model, and the adjustment method is expressed as: In the formula, β, k s and m0 are hyperparameters that control the height, steepness, and midpoint of the curve, respectively. S4.4: Use fractional order derivatives to calculate the gradient of network parameters to adapt to the distribution characteristics of data on the manifold. The derivative method is expressed as: In the formula, ΔW q is the weight update amount, η lm is the learning rate, Represents the loss function L q Take partial derivatives of order 0.5; The fractional derivative is expressed as: Where Γ is the Gamma function. The network parameters are updated according to the gradient calculated by back propagation. The update method is expressed as: In the formula, is the updated weight.

9. The battery status assessment method based on machine learning according to claim 7, characterized in that: The loss function of the fully connected neural network adopts the cross entropy loss, which is calculated by the preset Softmax function on the output of the last layer of the fully connected neural network.

10. The battery status assessment method based on machine learning according to claim 1, characterized in that: The battery data includes at least battery capacity, charging time, discharge rate, number of cycles, temperature fluctuation, voltage fluctuation, internal resistance change, capacity decay rate, battery leakage current and charging efficiency.

Images