Lithium ion battery defect recognition method based on multimodal feature consistency evolution
By using multimodal feature consistency evolution and deep neural network recognition models, the problems of misjudgment of single parameters and high cost of complex detection in lithium-ion battery defect identification are solved, and accurate identification and low-cost detection of defects are achieved.
Patent Information
- Application Number
- CN202611288722.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-06
AI Technical Summary
Existing lithium-ion battery defect identification methods are unable to accurately depict the complete development process of defects from latent to deterioration. Single electrical parameters are easily affected by fluctuations in operating conditions, resulting in a high risk of misjudgment and missed judgment. Complex physical detection methods are costly and difficult to promote on a large scale.
The method of consistent evolution of multimodal features is adopted. By synchronously collecting the time-domain signals of battery terminal voltage, terminal current and surface pressure response, incremental capacity curve data, online impedance spectrum data and pressure response data are constructed. Combined with the deep neural network recognition model, the nonlinear mapping relationship of multimodal features with the cyclic process is learned to achieve automatic recognition.
It achieves multi-modal signal collaborative characterization of the internal electrochemical characteristics and mechanical response of lithium-ion batteries, accurately depicts the defect evolution process, reduces the risk of misjudgment and missed judgment, and realizes low-cost and high-efficiency quasi-online detection.
Smart Images
Figure CN123049301A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lithium-ion battery testing technology, specifically to a lithium-ion battery defect identification method based on the consistent evolution of multimodal features. Background Technology
[0002] Lithium-ion batteries have been widely used in new energy vehicles, energy storage systems, and portable electronic devices due to their advantages such as high energy density, long cycle life, and low self-discharge rate. In the manufacturing process of pouch batteries, there are unavoidable process fluctuations in processes such as electrode preparation, stacking, welding, and packaging. These fluctuations can easily introduce production line defects such as metal foreign objects, separator damage, electrode misalignment, poor local contact, and microstructure stress concentration. In the early stages of operation, these defects usually do not show obvious electrical performance degradation, but they may gradually develop into local micro-short circuits, abnormal polarization, enhanced interface side reactions, or gas generation during charge-discharge cycles, resting, or aging. This can lead to serious safety problems such as cell swelling, accelerated capacity decay, and even thermal runaway.
[0003] Currently, the methods for identifying internal defects in lithium-ion batteries mainly fall into two categories: one is based on electrical parameter detection, which involves screening and evaluating cells using single or a few electrical indicators such as open-circuit voltage, DC internal resistance, and capacity; the other is based on physical characterization, which involves using electrochemical impedance spectroscopy, ultrasonic testing, or X-ray imaging to analyze the internal structure of the battery. In practical applications, detection methods based on a single electrical parameter are insufficient to characterize the dynamic evolution of defective batteries during cycling. Furthermore, a single parameter is susceptible to fluctuations in operating conditions and manufacturing variations, making it difficult to distinguish between normal performance differences and defect-induced abnormalities, resulting in a high risk of misjudgment and missed detection. While complex detection methods such as electrochemical impedance spectroscopy, ultrasound, and X-ray imaging can provide more internal information, they require significant equipment investment and have low detection efficiency, making them difficult to promote and apply in large-scale production lines. Analyzing each parameter independently or using simple threshold superposition for judgment fails to fully utilize the correlation and evolution information of multimodal signals over time, making it difficult to accurately depict the complete development process of defects from latent to deterioration, leading to insufficient ability to identify early latent defects. Therefore, a lithium-ion battery defect identification method based on the consistent evolution of multimodal features is proposed to address the above problems. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a lithium-ion battery defect identification method based on the consistent evolution of multimodal features. It has the advantages of multimodal signal collaborative characterization, full life cycle evolution feature tracking, and non-contact online detection. It solves the problems in existing technologies, such as the difficulty of capturing the dynamic evolution features of defects with a single electrical parameter, the high cost and long cycle of complex physical detection methods, and the insufficient ability to identify early latent defects due to the lack of multimodal collaborative analysis.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a lithium-ion battery defect identification method based on the consistency evolution of multimodal features, comprising the following steps; S1. Apply standard charge-discharge cycle conditions to the battery under test, collect and process multimodal signals during the cycle, and then construct a multimodal dataset; S2. Extract the original features of each loop node in the multimodal dataset, calculate the feature change of the current loop node based on the original features, perform vector fusion of the feature changes under the same loop node, construct a multimodal feature set, and normalize the multimodal feature set to obtain a normalized multimodal feature set. S3. Embed the Softmax function into the output layer of the deep neural network recognition model, train the deep neural network recognition model, input the normalized multimodal feature set into the deep neural network recognition model, learn the nonlinear mapping relationship of multimodal features as the cyclic process changes, and output the corresponding class probability; S4. Compare the probabilities of each category and determine the state category corresponding to the highest probability as the identification category of the battery to be tested.
[0006] Furthermore, a method for constructing a multimodal dataset is introduced in S1; Among them, the multimodal signals include the battery's terminal voltage time-domain signal, terminal current time-domain signal, and surface pressure response time-domain signal; Discrete voltage-capacity sampling data are extracted from the terminal voltage time-domain signal and the terminal current time-domain signal. The sampling data includes the voltage value and capacity value of each sampling point. For the current sampling point, obtain the capacity value and voltage value of its adjacent sampling points, calculate the capacity difference and voltage difference between the current sampling point and its adjacent sampling points respectively, and divide the capacity difference by the voltage difference to obtain the incremental capacity value of the current sampling point. The incremental capacity value of each sampling point is correlated with its corresponding voltage value to construct an incremental capacity curve data composed of discrete points of voltage-incremental capacity. Continuous wavelet transform is performed on the time-domain signals of terminal voltage and terminal current to obtain the wavelet coefficients of the voltage signal and the wavelet coefficients of the current signal. Based on the ratio of the wavelet coefficients of the voltage signal to the wavelet coefficients of the current signal, the real part and the imaginary part of the online complex impedance are obtained. Calculate the magnitude and phase angle of the online complex impedance based on the real and imaginary parts of the online complex impedance. Online impedance spectrum data is constructed based on preset frequency, online complex impedance real part, online complex impedance imaginary part, online complex impedance magnitude, and online complex impedance phase angle; Extract the complete data sequence of pressure change over time from the surface pressure response time domain signal. Based on the start and end times of each cycle in the data sequence, divide the data sequence into cycle nodes to obtain the pressure change sequence corresponding to each cycle node. Pressure response data is constructed based on the pressure change sequence corresponding to each loop node; The incremental capacity curve data, online impedance spectrum data, and pressure response data are time-aligned according to the cycle node. The incremental capacity curve data, online impedance spectrum data, and pressure response data under the same cycle node are correlated and integrated to construct a multimodal dataset.
[0007] Furthermore, S2 further introduces the original feature extraction method and the calculation method for feature changes; Based on the incremental capacity curve data, the main peak position, main peak height, and main peak area of the incremental capacity curve are extracted as the original features of incremental capacity. Based on the original incremental capacity features of the initial loop node, the original incremental capacity features of the current loop node are subtracted from the original incremental capacity features of the initial loop node to obtain the change in the peak position, the change in the peak height, and the change in the area of the main peak, and an incremental capacity feature vector is constructed. Based on the online impedance spectrum data, the local online impedance spectrum is fitted with a circle to obtain the impedance arc radius. The average phase angle and impedance spectrum area within the local frequency range are calculated and used as the original features of the online impedance spectrum. Based on the original online impedance spectrum features of the initial loop node, the original online impedance spectrum features of the current loop node are subtracted from the original online impedance spectrum features of the initial loop node to obtain the change in impedance arc radius, the change in average phase angle, and the change in impedance spectrum area, and an online impedance spectrum feature vector is constructed. Extract the maximum pressure, residual pressure, and pressure change amplitude from the pressure response data and use them as the original features of the pressure response; Based on the original pressure response features of the initial loop node, the original pressure response features of the current loop node are subtracted from the original pressure response features of the initial loop node to obtain the maximum pressure change, residual pressure change, and pressure change amplitude change, and a pressure response feature vector is constructed.
[0008] Furthermore, a normalized multimodal feature set construction method is introduced in S2; The incremental capacity feature vector, online impedance spectrum feature vector, and pressure response feature vector are spliced and fused according to the same loop node to construct a multimodal fused feature vector; The multimodal fusion feature vectors are arranged in a cyclical order to construct a multimodal feature set. The multimodal feature set is then normalized to obtain a normalized multimodal feature set.
[0009] Furthermore, step S3 further introduces the construction and training of a deep neural network recognition model; Select battery samples with known state categories, extract the normalized multimodal feature set of each battery sample, and assign the corresponding state category label according to the actual state of the battery sample. The status categories include normal status, slightly abnormal status, moderately abnormal status, and severely abnormal status. The number of input layer nodes of the deep neural network recognition model is set according to the feature dimension of the normalized multimodal feature set of the battery samples, and the number of output layer nodes is set according to the preset number of state categories. A deep neural network recognition model containing an input layer, multiple hidden layers, and an output layer is constructed. The output layer uses the Softmax function. The normalized multimodal feature set is taken as input, and after being processed by each hidden layer, the predicted probability of each training sample belonging to each category is output by the Softmax function of the output layer. The state category labels are converted into one-hot codes. The cross-entropy classification loss is calculated based on the predicted probability and the one-hot codes. The backpropagation algorithm is used to calculate the gradient of the classification loss with respect to the parameters of each layer of the network, and the parameters of each layer of the network are updated according to the gradient. The above process is repeated iteratively until the convergence condition is met, resulting in a trained deep neural network recognition model.
[0010] Furthermore, in step S4, the normalized multimodal feature set obtained in step S2 is input into the trained deep neural network recognition model for forward computation. The normalized multimodal feature set obtained in step S2 is used as the output of the input layer and passed sequentially to the first hidden layer; For each hidden layer other than the first hidden layer, the output values of each neuron from the previous hidden layer are multiplied by their corresponding weights and summed, then the corresponding bias term is added, and after processing by a non-linear activation function, the result is used as the output of the current hidden layer. Then it is passed layer by layer to the last hidden layer; The output of the last hidden layer is passed to the output layer, which calculates the raw scores for each category and uses the Softmax function to map each raw score to the corresponding category probability.
[0011] Furthermore, in step S4, the state determination is based on the category probability; By comparing the probabilities of each category, the state category corresponding to the highest probability is determined as the identification category of the battery under test; When the identification category is normal, the battery under test is determined to be a normal battery; When the identification category is a minor abnormal state, the battery under test is determined to be a minor abnormal battery; When the identification category is moderately abnormal state, the battery under test is determined to be a moderately abnormal battery; When the identification category is a severe abnormal state, the battery under test is determined to be a severely abnormal battery.
[0012] Compared with existing technologies, this invention provides a lithium-ion battery defect identification method based on the consistent evolution of multimodal features, which has the following advantages: 1. This lithium-ion battery defect identification method based on the consistent evolution of multimodal features simultaneously collects the time-domain signals of battery terminal voltage, terminal current, and surface pressure response, and constructs incremental capacity curve data, online impedance spectrum data, and pressure response data respectively. This achieves the coordinated characterization of multimodal signals of the battery's internal electrochemical characteristics, conduction state, and mechanical response, overcoming the limitation that a single electrical parameter cannot fully reflect the multidimensional evolution characteristics of defective batteries. 2. This lithium-ion battery defect identification method based on the consistent evolution of multimodal features calculates the feature changes of each current cycle node by taking the original features of the initial cycle node as a benchmark and tracking the degradation trajectory of multimodal features with the cycle process. It can accurately depict the complete evolution process of defects from latent to deterioration, solve the problem that existing technologies have difficulty distinguishing between normal performance differences and defect-induced abnormal performance, and effectively reduce the risk of misjudgment and missed judgment. 3. This lithium-ion battery defect identification method based on the consistent evolution of multimodal features integrates incremental capacity features, online impedance spectrum features, and pressure response features into a deep neural network identification model. The model automatically learns the nonlinear mapping relationship between the changes of multimodal features with the cycling process and the battery state. This overcomes the problems of strong subjectivity and poor adaptability caused by the independent analysis of each parameter or simple threshold superposition in existing methods, and realizes the automated and intelligent identification of battery state. 4. This lithium-ion battery defect identification method based on the consistent evolution of multimodal features can complete data acquisition by only collecting end electrical signals and surface pressure signals during the battery charging and discharging process. The entire detection process does not require disassembling the battery, adding an external excitation source, or a complex detection environment. It can be directly deployed on the basis of the existing production line charging and discharging steps. It solves the problems of high equipment cost, long detection cycle, and difficulty in large-scale production line promotion of complex detection methods such as electrochemical impedance spectroscopy, ultrasonic detection, and X-ray imaging. It achieves low-cost and high-efficiency quasi-online detection. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a detailed flowchart of step S2 of the present invention; Figure 3 This is a detailed flowchart of step S4 of the present invention. Detailed Implementation
[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0015] Example 1: Please refer to Figure 1-3 The lithium-ion battery defect identification method based on the consistency evolution of multimodal features in this embodiment includes the following steps; S1. Apply standard charge-discharge cycle conditions to the battery under test, collect and process multimodal signals during the cycle, and then construct a multimodal dataset; S2. Extract the original features of each loop node in the multimodal dataset, calculate the feature change of the current loop node based on the original features, perform vector fusion of the feature changes under the same loop node, construct a multimodal feature set, and normalize the multimodal feature set to obtain a normalized multimodal feature set. S3. Embed the Softmax function into the output layer of the deep neural network recognition model, train the deep neural network recognition model, input the normalized multimodal feature set into the deep neural network recognition model, learn the nonlinear mapping relationship of multimodal features as the cyclic process changes, and output the corresponding class probability; S4. Compare the probabilities of each category and determine the state category corresponding to the highest probability as the identification category of the battery to be tested.
[0016] It should be noted that a complete technical closed loop of "signal acquisition - multimodal data construction - feature change extraction - deep neural network probability output - state determination" has been constructed, providing a unified upper-level framework for subsequent weighting and realizing end-to-end automated processing from raw signals to defect identification. It solves the problem that existing technologies lack a systematic technical solution from multimodal signal acquisition to defect state determination, and that the links are isolated from each other, making it difficult to form an effective automated identification process.
[0017] In particular, S1 introduces a method for constructing multimodal datasets; Among them, the multimodal signals include the battery's terminal voltage time-domain signal, terminal current time-domain signal, and surface pressure response time-domain signal; Discrete voltage-capacity sampling data are extracted from the terminal voltage time-domain signal and the terminal current time-domain signal. The sampling data includes the voltage value and capacity value of each sampling point. For the current sampling point, obtain the capacity value and voltage value of its adjacent sampling points, calculate the capacity difference and voltage difference between the current sampling point and its adjacent sampling points respectively, and divide the capacity difference by the voltage difference to obtain the incremental capacity value of the current sampling point. The incremental capacity value of each sampling point is correlated with its corresponding voltage value to construct an incremental capacity curve data composed of discrete points of voltage-incremental capacity. It should be noted that after obtaining the voltage-capacity data, the incremental capacity curve represents the differential relationship between capacity and voltage, obtained by differentiating capacity Q with respect to voltage V. ;in, Indicates voltage The incremental capacity value at the location; For capacity; For voltage, This represents the rate of change of capacitance Q with voltage V; For the actual collected discrete voltage-capacity data: ;in, Indicates the first The voltage value corresponding to each sampling point; This represents the capacity value corresponding to the i-th sampling point; Indicates the discrete sampling point number. ; This represents the total number of sampling points for discrete voltage-capacity data; The incremental capacity value is calculated using a differential method: Since the experimental data consists of discrete sampling points, the incremental capacity value is calculated using a differential method. ;in, This represents the incremental capacity value corresponding to the nth sampling point; and These represent the capacities of adjacent sampling points; and These represent the voltages of adjacent sampling points; This yields the corresponding IC curve sampling sequence: ;in, This represents the set of discrete sampling points for the IC curve; therefore, the first and last two sampling points do not participate in the central difference calculation. Continuous wavelet transform is performed on the time-domain signals of terminal voltage and terminal current to obtain the wavelet coefficients of the voltage signal and the wavelet coefficients of the current signal. Based on the ratio of the wavelet coefficients of the voltage signal to the wavelet coefficients of the current signal, the real part and the imaginary part of the online complex impedance are obtained. Calculate the magnitude and phase angle of the online complex impedance based on the real and imaginary parts of the online complex impedance. Online impedance spectrum data is constructed based on preset frequency, online complex impedance real part, online complex impedance imaginary part, online complex impedance magnitude, and online complex impedance phase angle; It should be noted that during the cyclic aging process, the final step synchronously acquires the battery terminal voltage signal at a frequency of 100Hz. and current signal And time-frequency analysis of both was performed using an online impedance spectroscopy program: ; ;in, and The wavelet coefficients for voltage and current signals are respectively. The time-domain signal representing the change of battery terminal voltage over time. A time-domain signal representing the change of battery terminal current over time; Indicates the sampling time; Represents the wavelet transform scaling parameter; Indicates the time shift parameter; Represents the mother wavelet function; The conjugate function of the mother wavelet function; Represents the time variable Integrate points; Within the local frequency range, the real part and imaginary part of the online complex impedance are obtained by the ratio of the corresponding voltage wavelet coefficient to the current wavelet coefficient: ;in, For the m-th frequency point, For the corresponding scale.
[0018] The online complex impedance is expressed as: ;in, Represents frequency Online complex impedance at the location; Represents the real part of the online complex impedance; This represents the imaginary part of the online complex impedance; j is the imaginary unit. The impedance magnitude and phase angle are as follows: ; ;in, Represents frequency The magnitude of the complex impedance at the line; Represents the real part of the online complex impedance; Represents the imaginary part of the online complex impedance; Represents frequency Phase angle at the online complex impedance; Represents the arctangent function; Therefore, the local online impedance spectrum data obtained based on wavelet transform can be expressed as: ;in, This represents the set of online impedance spectrum features; m represents the frequency point number. M represents the total number of frequency points selected in the local impedance spectrum; Extract the complete data sequence of pressure change over time from the surface pressure response time domain signal. Based on the start and end times of each cycle in the data sequence, divide the data sequence into cycle nodes to obtain the pressure change sequence corresponding to each cycle node. Pressure response data is constructed based on the pressure change sequence corresponding to each loop node; It should be noted that the pressure response on the battery surface is collected in real time by a pressure sensor, resulting in a data sequence of pressure changes over time. ; in, This represents the pressure value collected at the nth time point, where N is the number of pressure sampling points.
[0019] For the k-th cycle, the pressure data can be represented as: ; in, This represents the sequence of pressure changes during the k-th cycle.
[0020] The incremental capacity curve data, online impedance spectrum data, and pressure response data are time-aligned according to the cycle node. The incremental capacity curve data, online impedance spectrum data, and pressure response data under the same cycle node are correlated and integrated to construct a multimodal dataset.
[0021] It should be noted that by synchronously acquiring the time-domain signals of terminal voltage, terminal current, and surface pressure response, and processing them through differential operation, continuous wavelet transform, and cyclic segmentation, incremental capacity curve data, online impedance spectrum data, and pressure response data were constructed, realizing the multi-modal signal collaborative characterization of the battery's internal electrochemical characteristics, conduction state, and mechanical response.
[0022] This method solves the problem that existing single electrical parameter detection methods can only reflect the performance indicators of a battery in one dimension, and cannot comprehensively characterize the synergistic evolution of defective batteries in terms of electrochemical activity, conduction state and mechanical response, and have limited ability to identify early latent defects.
[0023] In S2, the original feature extraction method and the calculation method of feature change are further introduced; Based on the incremental capacity curve data, the main peak position, main peak height, and main peak area of the incremental capacity curve are extracted as the original features of incremental capacity. Based on the original incremental capacity features of the initial loop node, the original incremental capacity features of the current loop node are subtracted from the original incremental capacity features of the initial loop node to obtain the change in the peak position, the change in the peak height, and the change in the area of the main peak, and an incremental capacity feature vector is constructed. It should be noted that the IC curve of the i-th battery at the k-th cycle node is represented by the following method: ;in, This represents the nth sampling voltage point. This represents the voltage of the i-th battery at the k-th cycle node. The incremental capacity value at the location, where N represents the total number of sampling points.
[0024] The peak position, peak height, and peak area of the main peak in the IC curve are extracted as incremental capacity features based on the IC sampling sequence. Let the sampling point of the IC curve at the main peak of the i-th battery at the k-th cycle node be . ,but: Main peak position: Main peak height: In the voltage range where the main peak is located Inside, the peak area is: ; in, This indicates the position of the main peak of the IC curve at the k-th cycle node. Indicates the height of the main peak. Indicates the area of the main peak. This represents the voltage interval between adjacent sampling points; The incremental capacity feature vector of the i-th battery at the k-th cycle node is represented as: ; in, This represents the change in the position of the main peak of the IC curve for the i-th battery at the k-th cycle node. This indicates the change in the peak height of the main peak of the IC curve. This indicates the change in the area of the main peak of the IC curve.
[0025] Based on the online impedance spectrum data, the local online impedance spectrum is fitted with a circle to obtain the impedance arc radius. The average phase angle and impedance spectrum area within the local frequency range are calculated and used as the original features of the online impedance spectrum. Based on the original online impedance spectrum features of the initial loop node, the original online impedance spectrum features of the current loop node are subtracted from the original online impedance spectrum features of the initial loop node to obtain the change in impedance arc radius, the change in average phase angle, and the change in impedance spectrum area, and an online impedance spectrum feature vector is constructed. It should be noted that, for the i-th battery at the k-th cyclic node, the local online impedance spectrum obtained by wavelet transform is expressed as: ;in, This represents the m-th frequency point. and Let represent the real and imaginary parts of the impedance of the i-th battery at the k-th cycle node and the m-th frequency point, respectively. Indicates the impedance magnitude. Indicates the impedance phase angle; Circular fitting of the local impedance spectrum: ;in, and These represent the x and y axes of the Nyquist plot corresponding to the m-th frequency point, respectively. The fitted center coordinates of the local impedance arc of the i-th battery at the k-th cycle node are represented. Indicates the radius of the corresponding local impedance arc; Average phase angle within the local frequency range: ; in, This represents the average phase angle of the i-th battery within the local frequency range at the k-th cycle node; This represents the impedance phase angle of the i-th battery at the k-th cycle node and the m-th frequency point; Local impedance spectrum area: ;in, This represents the local impedance spectrum area of the i-th cell at the k-th cycle node; This represents the magnitude of the online complex impedance corresponding to the m-th frequency point; This represents the frequency interval between the m-th and (m+1)-th frequency points; As the cyclic charge-discharge experiment proceeds, the impedance radius, phase angle, and spectral area inside the battery will change. The online impedance spectrum eigenvector of the i-th battery at the k-th cycle node is represented as: ; Extract the maximum pressure, residual pressure, and pressure change amplitude from the pressure response data and use them as the original features of the pressure response; Based on the original pressure response features of the initial loop node, the original pressure response features of the current loop node are subtracted from the original pressure response features of the initial loop node to obtain the maximum pressure change, residual pressure change, and pressure change amplitude change, and a pressure response feature vector is constructed.
[0026] It should be noted that the pressure response data of the i-th battery at the k-th cycle node is represented as: ;in, This represents the pressure value collected by the i-th battery at the k-th cycle node and at the n-th time, where N represents the total number of pressure sampling points; The pressure response feature vector of the i-th battery at the k-th cycle node is represented as: ;in, This represents the maximum pressure change of the i-th battery at the k-th cycle node. This represents the change in residual pressure. This indicates the amount of change in the magnitude of the pressure change; It should be noted that by using the original characteristics of the initial cycle node as a benchmark, the changes in the main peak position, peak height, peak area, impedance arc radius, average phase angle, impedance spectrum area, maximum pressure, residual pressure, and pressure change amplitude of each current cycle node are calculated. This quantifies the evolution process of multimodal characteristics into calculable numerical indicators, thus achieving an accurate characterization of the complete process of defects from latency to deterioration.
[0027] This solves the problem that existing technologies cannot distinguish between normal battery performance differences and abnormal performance induced by defects, and that independent analysis of each parameter cannot reflect the dynamic evolution process of defects, leading to a high risk of misjudgment and missed judgment.
[0028] In S2, a normalized multimodal feature set construction method is further introduced; The incremental capacity feature vector, online impedance spectrum feature vector, and pressure response feature vector are spliced and fused according to the same loop node to construct a multimodal fused feature vector; The multimodal fusion feature vectors are arranged in a cyclical order to construct a multimodal feature set. The multimodal feature set is then normalized to obtain a normalized multimodal feature set.
[0029] It should be noted that the multimodal fusion feature vector of the i-th battery at the k-th cycle node is constructed by splicing and fusing according to the same cycle node: ; The multimodal fusion feature vectors obtained from the i-th battery at multiple cyclic nodes are arranged in cyclic order to construct a multimodal feature set: ; in, Let K represent the set of multimodal features of the i-th battery, and K represent the number of selected cyclic nodes. This represents the j-th fusion feature of the i-th battery at the k-th cycle node.
[0030] To reduce the impact of multimodal features on deep neural network training, the multimodal feature set is normalized to obtain a normalized multimodal feature set: in, This represents the normalization processing function. This represents the normalized set of multimodal features.
[0031] It should be noted that by concatenating and fusing the incremental capacity feature vector, the online impedance spectrum feature vector, and the pressure response feature vector according to the same cyclic node, and then normalizing them, the differences in dimensions and orders of magnitude between different modal features are eliminated, providing standardized input data for deep neural networks and improving the training efficiency and recognition accuracy of the model.
[0032] This addresses the problem that when different modal features have significantly different dimensions and numerical ranges (e.g., voltage is measured in V, impedance in Ω, and pressure in N), directly inputting them into a neural network can lead to unstable training and slow convergence. Existing technologies lack effective methods for fusing and standardizing multimodal features.
[0033] Example 2: Please refer to Figure 1-3 Based on Example 1, the lithium-ion battery defect identification method based on the consistency evolution of multimodal features further introduces the construction and training of a deep neural network identification model in step S3. Battery samples with known state categories are selected, and the normalized multimodal feature set of each battery sample is extracted. The corresponding state category label is assigned according to the actual state of the battery sample. The state categories include normal state, slightly abnormal state, moderately abnormal state, and severely abnormal state. The number of input layer nodes of the deep neural network recognition model is set according to the feature dimensions of the normalized multimodal feature set of the battery samples, and the number of output layer nodes is set according to the preset number of state categories. A deep neural network recognition model containing an input layer, multiple hidden layers, and an output layer is constructed. The output layer uses the Softmax function. The normalized multimodal feature set is used as input, processed by each hidden layer, and the predicted probability of each training sample belonging to each category is output by the Softmax function of the output layer. The state category label is converted into one-hot encoding, and the cross-entropy classification loss is calculated based on the predicted probability and the one-hot encoding. The gradient of the classification loss with respect to the network parameters of each layer is calculated using the backpropagation algorithm, and the network parameters of each layer are updated according to the gradient. The above process is iteratively executed until the convergence condition is met, and the trained deep neural network recognition model is obtained.
[0034] It should be noted that the predicted probability of the i-th training battery belonging to the c-th state is: ;in, This represents the output value obtained by the output layer for the c-th state; c represents the state category number; r represents the category traversal number when summing the values of each state category; C represents the preset total number of state categories; Represents an exponential function; Based on the pre-labeled real-state categories of the training batteries, the real-state labels are represented using one-hot encoding as follows: ; in, This represents the true label value of the c-th class state corresponding to the i-th training battery; The classification loss is calculated based on the predicted probability corresponding to the true state category of each training battery, and is expressed by the following classification loss function: Where L represents the loss function value during model training; N represents the total number of training samples; and i represents the training battery number. Represent the natural logarithm function; It should be noted that by using the normalized multimodal feature set of battery samples with known state categories as training samples, and employing the cross-entropy loss function and backpropagation algorithm to iteratively optimize network parameters, the deep neural network automatically learns the nonlinear mapping relationship between the changes of multimodal features with the cyclic process and the battery state, thus avoiding the subjectivity and limitations of manually setting thresholds.
[0035] This solves the problem that existing technologies often rely on independent analysis of each parameter or simple threshold superposition for judgment, making it difficult to uncover the complex nonlinear correlation between multimodal features, resulting in poor adaptability and difficulty in meeting the battery status recognition needs of different batches and operating conditions.
[0036] In step S4, the normalized multimodal feature set obtained in step S2 is further introduced into a forward computation process, which is then input into the trained deep neural network recognition model. The normalized multimodal feature set obtained in step S2 is used as the output of the input layer and passed sequentially to the first hidden layer; For each hidden layer other than the first hidden layer, the output values of each neuron from the previous hidden layer are multiplied by their corresponding weights and summed, then the corresponding bias term is added, and after being processed by a non-linear activation function, it is used as the output of the current hidden layer; then it is passed layer by layer to the last hidden layer. The output of the last hidden layer is passed to the output layer, which calculates the raw scores for each category and uses the Softmax function to map each raw score to the corresponding category probability.
[0037] It should be noted that the normalized multimodal feature set of the battery under test is as follows: ;in, Let K represent the normalized multimodal fusion feature vector of the j-th battery under test at the k-th cyclic node, where K represents the number of cyclic nodes selected. The normalized multimodal feature set of the j-th battery to be tested The trained deep neural network recognition model is input and used as the output of the network input layer, i.e.: ; For the Each hidden layer performs weighted calculations and nonlinear mappings on the output of the previous layer based on the weight matrix and bias terms determined during the training phase. ;in, Indicates the first The output of the layer, Indicates the first The weight matrix of each hidden layer Indicates the first The bias term of the hidden layer Represents a non-linear activation function. Indicates the first The output of the hidden layer.
[0038] No. The computation process of the r-th neuron in each hidden layer is represented as follows: ;in, Indicates the first The number of layer output features Indicates the first Layer The first feature and the second Layer Connection weights between neurons Indicates the first Layer Bias terms corresponding to each neuron; After calculations through L hidden layers, the deep feature representation of the j-th cell under test is obtained: in, This indicates the number of neurons contained in the last hidden layer.
[0039] Input the output of the last hidden layer into the output layer, and calculate the original scores of the j-th battery under test for each of the four state categories: ; in, This represents the original score of the j-th battery under test, indicating that it belongs to the c-th state. This represents the connection weight between the r-th neuron in the last hidden layer and the c-th class in the output layer. This represents the bias term for the c-th category of the output layer; The original score vector obtained from the output layer is represented as follows: ; The Softmax function is used to convert the raw scores of each state category into corresponding class probabilities:
[0040] in, Let represent the probability that the j-th battery to be tested belongs to the c-th state. This indicates the traversal number of the state category.
[0041] It should be noted that by inputting the normalized multimodal feature set into the trained deep neural network recognition model, and performing weighted calculations and nonlinear mapping layer by layer through each hidden layer, the Softmax function outputs the probability of the battery under test belonging to each category. This achieves an end-to-end mapping from multimodal features to defect state probabilities, and detection can be completed with only conventional production line signals, without the need for external excitation sources or complex detection environments.
[0042] It solves the problems of complex detection methods such as electrochemical impedance spectroscopy, ultrasonic testing, and X-ray imaging, which require external excitation sources or special equipment, have high equipment costs, long detection cycles, and strict requirements on the environment and operators, making them difficult to promote and apply in large-scale production lines.
[0043] Example 3: Please refer to Figure 1-3 Based on Example 2, the lithium-ion battery defect identification method based on the consistency evolution of multimodal features determines the state based on category probability in step S4. By comparing the probabilities of each category, the state category corresponding to the highest probability is determined as the identification category of the battery under test; When the identification category is normal, the battery under test is determined to be a normal battery; when the identification category is slightly abnormal, the battery under test is determined to be slightly abnormal; when the identification category is moderately abnormal, the battery under test is determined to be moderately abnormal; when the identification category is severely abnormal, the battery under test is determined to be severely abnormal.
[0044] It should be noted that the state category probability vector of the j-th battery under test, transformed by the Softmax function in S3, is: ; in, Let represent the probability that the j-th battery under test belongs to a normal state, a slightly abnormal state, a moderately abnormal state, and an abnormal state, and satisfy the following: ; By comparing the probabilities corresponding to the four state categories, the state category with the highest probability is determined as the identification category of the j-th battery to be tested. ; in, This indicates the initial identification category of the j-th battery to be tested. This indicates the state category number corresponding to the highest probability.
[0045] The specific determination method is as follows: .
[0046] It should be noted that by comparing the probabilities of each category, the state category corresponding to the highest probability is determined as the identification category of the battery under test. The judgment results are output according to four categories: normal state, slightly abnormal state, moderately abnormal state, and severely abnormal state, thus realizing the graded early warning of latent defects.
[0047] This solves the problem that existing technologies lack the ability to classify and identify the degree of battery defects, and can only judge "qualified / unqualified", which cannot provide a refined condition assessment basis for production line screening and tiered utilization.
[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for identifying defects in lithium-ion batteries based on the consistent evolution of multimodal features, characterized in that, Includes the following steps; S1. Apply standard charge-discharge cycle conditions to the battery under test, collect and process multimodal signals during the cycle, and then construct a multimodal dataset; S2. Extract the original features of each loop node in the multimodal dataset, calculate the feature change of the current loop node based on the original features, perform vector fusion of the feature changes under the same loop node, construct a multimodal feature set, and normalize the multimodal feature set to obtain a normalized multimodal feature set. S3. Embed the Softmax function into the output layer of the deep neural network recognition model, train the deep neural network recognition model, input the normalized multimodal feature set into the deep neural network recognition model, learn the nonlinear mapping relationship of multimodal features as the cyclic process changes, and output the corresponding class probability; S4. Compare the probabilities of each category and determine the state category corresponding to the highest probability as the identification category of the battery to be tested.
2. The lithium-ion battery defect identification method based on multimodal feature consistency evolution according to claim 1, characterized in that: The method for constructing a multimodal dataset is further introduced in S1; Among them, the multimodal signals include the battery's terminal voltage time-domain signal, terminal current time-domain signal, and surface pressure response time-domain signal; Discrete voltage-capacity sampling data are extracted from the terminal voltage time-domain signal and the terminal current time-domain signal. The sampling data includes the voltage value and capacity value of each sampling point. For the current sampling point, obtain the capacity value and voltage value of its adjacent sampling points, calculate the capacity difference and voltage difference between the current sampling point and its adjacent sampling points respectively, and divide the capacity difference by the voltage difference to obtain the incremental capacity value of the current sampling point. The incremental capacity value of each sampling point is correlated with its corresponding voltage value to construct an incremental capacity curve data composed of discrete points of voltage-incremental capacity. Continuous wavelet transform is performed on the time-domain signals of terminal voltage and terminal current to obtain the wavelet coefficients of the voltage signal and the wavelet coefficients of the current signal. Based on the ratio of the wavelet coefficients of the voltage signal to the wavelet coefficients of the current signal, the real part and the imaginary part of the online complex impedance are obtained. Calculate the magnitude and phase angle of the online complex impedance based on the real and imaginary parts of the online complex impedance. Online impedance spectrum data is constructed based on preset frequency, online complex impedance real part, online complex impedance imaginary part, online complex impedance magnitude, and online complex impedance phase angle; Extract the complete data sequence of pressure change over time from the surface pressure response time domain signal. Based on the start and end times of each cycle in the data sequence, divide the data sequence into cycle nodes to obtain the pressure change sequence corresponding to each cycle node. Pressure response data is constructed based on the pressure change sequence corresponding to each loop node; The incremental capacity curve data, online impedance spectrum data, and pressure response data are time-aligned according to the cycle node. The incremental capacity curve data, online impedance spectrum data, and pressure response data under the same cycle node are correlated and integrated to construct a multimodal dataset.
3. The lithium-ion battery defect identification method based on multimodal feature consistency evolution according to claim 2, characterized in that: S2 further introduces the original feature extraction method and the calculation method for feature changes; Based on the incremental capacity curve data, the main peak position, main peak height, and main peak area of the incremental capacity curve are extracted as the original features of incremental capacity. Based on the original incremental capacity features of the initial loop node, the original incremental capacity features of the current loop node are subtracted from the original incremental capacity features of the initial loop node to obtain the change in the peak position, the change in the peak height, and the change in the area of the main peak, and an incremental capacity feature vector is constructed. Based on the online impedance spectrum data, the local online impedance spectrum is circle-fitted to obtain the impedance arc radius, and the average phase angle and impedance spectrum area within the local frequency range are calculated and used as the original features of the online impedance spectrum. Based on the original online impedance spectrum features of the initial loop node, the original online impedance spectrum features of the current loop node are subtracted from the original online impedance spectrum features of the initial loop node to obtain the change in impedance arc radius, the change in average phase angle, and the change in impedance spectrum area, and an online impedance spectrum feature vector is constructed. Extract the maximum pressure, residual pressure, and pressure change amplitude from the pressure response data and use them as the original features of the pressure response; Based on the original pressure response features of the initial loop node, the original pressure response features of the current loop node are subtracted from the original pressure response features of the initial loop node to obtain the maximum pressure change, residual pressure change, and pressure change amplitude change, and a pressure response feature vector is constructed.
4. The lithium-ion battery defect identification method based on multimodal feature consistency evolution according to claim 3, characterized in that: The S2 further introduces a normalized multimodal feature set construction method; The incremental capacity feature vector, online impedance spectrum feature vector, and pressure response feature vector are spliced and fused according to the same loop node to construct a multimodal fused feature vector; The multimodal fusion feature vectors are arranged in a cyclical order to construct a multimodal feature set. The multimodal feature set is then normalized to obtain a normalized multimodal feature set.
5. The lithium-ion battery defect identification method based on multimodal feature consistency evolution according to claim 4, characterized in that: The S3 step further introduces the construction and training of a deep neural network recognition model; Select battery samples with known state categories, extract the normalized multimodal feature set of each battery sample, and assign the corresponding state category label according to the actual state of the battery sample. The status categories include normal status, slightly abnormal status, moderately abnormal status, and severely abnormal status. The number of input layer nodes of the deep neural network recognition model is set according to the feature dimension of the normalized multimodal feature set of the battery samples, and the number of output layer nodes is set according to the preset number of state categories. A deep neural network recognition model containing an input layer, multiple hidden layers, and an output layer is constructed. The output layer uses the Softmax function. The normalized multimodal feature set is taken as input, and after being processed by each hidden layer, the predicted probability of each training sample belonging to each category is output by the Softmax function of the output layer. The state category labels are converted into one-hot codes. The cross-entropy classification loss is calculated based on the predicted probability and the one-hot codes. The backpropagation algorithm is used to calculate the gradient of the classification loss with respect to the parameters of each layer of the network, and the parameters of each layer of the network are updated according to the gradient. The above process is repeated iteratively until the convergence condition is met, resulting in a trained deep neural network recognition model.
6. The lithium-ion battery defect identification method based on multimodal feature consistency evolution according to claim 5, characterized in that: In step S4, the normalized multimodal feature set obtained in step S2 is further introduced into a forward computation process, which is then input into the trained deep neural network recognition model. The normalized multimodal feature set obtained in step S2 is used as the output of the input layer and passed sequentially to the first hidden layer; For each hidden layer other than the first hidden layer, the output values of each neuron from the previous hidden layer are multiplied by their corresponding weights and summed, then the corresponding bias term is added, and after processing by a non-linear activation function, the result is used as the output of the current hidden layer. Then it is passed layer by layer to the last hidden layer; The output of the last hidden layer is passed to the output layer, which calculates the raw scores for each category and uses the Softmax function to map each raw score to the corresponding category probability.
7. The lithium-ion battery defect identification method based on multimodal feature consistency evolution according to claim 6, characterized in that: In step S4, the state determination is based on the category probability; By comparing the probabilities of each category, the state category corresponding to the highest probability is determined as the identification category of the battery under test; When the identification category is normal, the battery under test is determined to be a normal battery; When the identification category is a minor abnormal state, the battery under test is determined to be a minor abnormal battery; When the identification category is moderately abnormal state, the battery under test is determined to be a moderately abnormal battery; When the identification category is a severe abnormal state, the battery under test is determined to be a severely abnormal battery.