Lithium battery recovery process intelligent monitoring method and system combined with deep learning

Abstract

The invention discloses a lithium battery recovery process intelligent monitoring method and system combined with deep learning, and relates to the technical field of intelligent battery recovery monitoring, and the method comprises the steps: building a manifold Laplacian matrix based on a graph embedding feature matrix, defining an optimization objective function, obtaining a low-rank graph coefficient matrix through iterative optimization, and generating a final fusion matrix. Extracting a final fusion feature, predicting a health state and a recovery rate by using an SVR model, performing anomaly and fault probability detection by using RF, generating a monitoring report, and uploading the monitoring report to a cloud for storage; through combination of multi-dimensional data and Gibbs-Shannon entropy adaptive filtering, the adaptability and stability of data preprocessing are improved, the accuracy of state evaluation is improved, and through low-rank feature optimization and graph embedding fusion, accurate prediction of the battery health state and recovery rate is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent battery recycling monitoring technology, and in particular to an intelligent monitoring method and system for the lithium battery recycling process that combines deep learning. Background Technology

[0002] With the rapid development of the new energy industry, lithium batteries, as the core component of energy storage and power systems, have been widely used in electric vehicles, energy storage power stations, and consumer electronics. With the surge in battery usage, the recycling and reuse of retired batteries has become increasingly prominent. Traditional lithium battery recycling monitoring methods rely on single sensors or laboratory testing methods, which cannot fully reflect the state changes of batteries in complex environments. The integration of multimodal sensing and deep learning provides a new research direction for intelligent monitoring of the lithium battery recycling process. By deploying multiple types of sensors, it is possible to simultaneously acquire multidimensional features of the battery's internal and external components. Combined with deep learning structures such as convolutional neural networks, recurrent neural networks, and variational autoencoders, implicit spatiotemporal relationships can be extracted at the feature level, thereby achieving high-precision prediction of battery health status, material degradation mechanisms, and recycling rates.

[0003] However, existing technologies still have shortcomings. Existing methods rely on single or limited modal data, resulting in insufficient capture of complex dynamics such as battery damage areas, material particle distribution, and acoustic signals. This affects the comprehensiveness of health status assessment and the accuracy of recovery rate calculation. Existing preprocessing techniques use static filters, which cannot adapt to dynamic changes in data noise. They are prone to information loss or over-smoothing, resulting in non-adaptive denoising intensity and reduced robustness of subsequent segmentation and feature extraction. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent monitoring method and system for lithium battery recycling process that combines deep learning. This solves the problem that existing methods rely on single or limited modal data, resulting in insufficient capture of complex dynamics such as battery damage areas, material particle distribution, and acoustic signals. This affects the comprehensiveness of health status assessment and the accuracy of recycling rate calculation. Existing preprocessing techniques use static filters, which cannot adapt to dynamic changes in data noise, are prone to information loss or over-smoothing, and result in non-adaptive denoising intensity, reducing the robustness of subsequent segmentation and feature extraction.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent monitoring method for lithium battery recycling processes that incorporates deep learning, comprising the following steps:

[0008] Multimodal sensors are deployed to collect multimodal data. Non-image multimodal data is preprocessed, and equal-pixel segmentation is performed on battery surface images and material particle distribution images. Gibbs-Shannon entropy is calculated and combined with flexibility index to obtain denoising intensity. Filtering parameters are initialized, an initial optimization objective function is constructed, and meta-dynamics is used for optimization to obtain the optimal filtering parameters for denoising. Battery damage area and material particle images are segmented through U-Net model. Combined with the preprocessed non-image multimodal data, a multimodal dataset is generated.

[0009] Based on a multimodal dataset, select multimodal features are extracted, their Pearson correlation coefficients with the target parameters are calculated, features with correlation coefficients exceeding the threshold are selected, and the weights are optimized using the EPMA algorithm to construct a selected multimodal feature matrix.

[0010] The selected multimodal feature matrix is ​​set as the initial low-rank feature matrix. The optimized low-rank feature matrix is ​​obtained by optimizing it using the enhanced Lagrange multiplier method. The similarity between samples is calculated using the Gaussian kernel function to generate the weight matrix. The intra-class scatter matrix and inter-class scatter matrix are constructed to obtain the projection matrix. The optimized low-rank feature matrix is ​​projected to a low-dimensional space and normalized to obtain the graph embedding feature matrix.

[0011] Based on the graph embedding feature matrix, a manifold Laplacian matrix is ​​constructed, an optimization objective function is defined, and a low-rank graph coefficient matrix is ​​obtained through iterative optimization. The final fusion matrix is ​​generated, the final fusion features are extracted, and health status prediction and fault detection are performed. A monitoring report is generated and uploaded to the cloud for storage.

[0012] As a preferred embodiment of the intelligent monitoring method for lithium battery recycling process combining deep learning described in this invention, the step of generating a multimodal dataset by combining preprocessed non-image multimodal data includes:

[0013] The battery surface image and the material particle distribution image are segmented into non-overlapping image blocks using equal pixel segmentation. A gray-level histogram is calculated for each image block using a direct traversal statistical method. The counts in the gray-level histogram are normalized to obtain the gray-level probability. Gibbs-Shannon entropy is calculated, and the denoising intensity is obtained by combining it with a flexibility index. An initial optimization objective function is constructed, and the initial filtering parameters are optimized using meta-dynamics. The image blocks are then denoised, and the U-Net model is used for segmentation to obtain images of the battery damage area and material particles. Timestamp alignment is performed by combining normalized non-image multimodal data and frequency features, and a multimodal dataset is generated by horizontal arrangement.

[0014] As a preferred embodiment of the intelligent monitoring method for lithium battery recycling process combining deep learning described in this invention, the step of extracting selected multimodal features based on a multimodal dataset includes:

[0015] Based on the multimodal dataset, selected multimodal features are extracted, including the concatenation of normalized image features, normalized sensor features, normalized features, normalized derivative features, normalized SVD features, normalized capacity features, and normalized constant current features to obtain multimodal features.

[0016] As a preferred embodiment of the intelligent monitoring method for lithium battery recycling process combining deep learning described in this invention, the step of optimizing weights and constructing a selected multimodal feature matrix through the EPMA algorithm includes:

[0017] The correlation coefficient between multimodal features and target parameters is calculated using the Pearson correlation coefficient formula. Multimodal features with correlation coefficients greater than the correlation coefficient threshold are selected for retention. The weights of the retained multimodal features are then optimized using the EPMA algorithm to obtain a selected multimodal feature matrix.

[0018] As a preferred embodiment of the intelligent monitoring method for lithium battery recycling process combining deep learning described in this invention, wherein: the step of projecting the optimized low-rank feature matrix to a low-dimensional space and normalizing it to obtain a graph embedding feature matrix includes:

[0019] The selected multimodal feature matrix is ​​set as the initial low-rank feature matrix. It is optimized by the enhanced Lagrange multiplier method and stops when the maximum number of iterations is reached to obtain the optimized low-rank feature matrix. The similarity of the optimized low-rank feature matrix is ​​calculated using the Gaussian kernel function to construct the initial weight matrix, generate the intra-class scatter matrix and inter-class scatter matrix, and obtain the projection matrix. The optimized low-rank feature matrix is ​​projected to a low-dimensional space using the projection matrix to obtain the graph embedding feature matrix.

[0020] As a preferred embodiment of the intelligent monitoring method for lithium battery recycling process combining deep learning described in this invention, the step of extracting the final fusion features, performing health status prediction and fault detection, generating a monitoring report, and uploading it to the cloud for storage includes:

[0021] Based on the graph embedding feature matrix, a manifold Laplacian matrix is ​​constructed, an optimization objective function is defined, and a low-rank graph coefficient matrix is ​​obtained through iterative optimization. Feature scores are calculated to obtain a filter feature matrix. The final fusion matrix is ​​obtained by combining the projection matrix with matrix multiplication. The final fusion features are extracted, and the SVR model is used for prediction to obtain the predicted health status and recovery rate. RF is used for anomaly and fault detection to obtain the anomaly and fault probabilities. A monitoring report is generated and uploaded to the cloud for storage.

[0022] As a preferred embodiment of the intelligent monitoring method for lithium battery recycling process combining deep learning described in this invention, the deployment of multimodal sensors to collect multimodal data and the preprocessing of non-image multimodal data include:

[0023] Multimodal sensors are deployed to collect multimodal data. Non-image multimodal data is preprocessed, including denoising the non-image multimodal data using Savitzky-Golay filtering and normalizing it. Frequency domain features are extracted from the normalized acoustic signal using FFT.

[0024] Secondly, the present invention provides an intelligent monitoring system for lithium battery recycling processes that incorporates deep learning, comprising:

[0025] The data acquisition and processing module is used to deploy multimodal sensors to acquire battery images and multi-source physicochemical data, and to achieve data cleaning and enhancement through filtering, normalization and adaptive denoising optimization based on Gibbs-Shannon entropy.

[0026] The feature extraction and fusion module is used to extract multi-dimensional features from image and time-series sensor data based on deep models, and generate a selected multimodal feature matrix through feature stitching, correlation filtering and EPMA optimization.

[0027] The feature optimization and graph embedding module is used to optimize low-rank features using the enhanced Lagrange method and manifold regularization, construct graph embedding and manifold Laplacian matrix, and extract highly correlated low-dimensional embedding features.

[0028] The feature selection and fusion matrix generation module is used to calculate feature scores based on the sparse matrix of the low-rank graph, select high-contribution features, and combine them with the projection matrix to generate the final fusion feature matrix.

[0029] The intelligent prediction and upload module is used to predict the health status and recycling rate using SVR, and combine the anomaly and failure probability of RF detection to generate and upload cloud monitoring reports, thereby realizing intelligent monitoring of the lithium battery recycling process.

[0030] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent monitoring method for lithium battery recycling process incorporating deep learning as described in the first aspect of the present invention.

[0031] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent monitoring method for lithium battery recycling process incorporating deep learning as described in the first aspect of the present invention.

[0032] The beneficial effects of this invention are as follows: This invention constructs a manifold Laplacian matrix based on a graph embedding feature matrix, defines an optimization objective function, iteratively optimizes to obtain a low-rank graph coefficient matrix, generates a final fusion matrix, extracts the final fusion features, uses an SVR model to predict health status and recovery rate, uses RF for anomaly and fault probability detection, generates a monitoring report, and uploads it to the cloud for storage; it improves the adaptability and stability of data preprocessing, enhances the accuracy of state assessment, and achieves accurate prediction of battery health status and recovery rate. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart illustrating the operation of the intelligent monitoring method for lithium battery recycling process that incorporates deep learning, as described in Example 1.

[0035] Figure 2 This is a schematic diagram of the intelligent monitoring system for lithium battery recycling process that incorporates deep learning, as shown in Example 1. Detailed Implementation

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0038] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0039] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides an intelligent monitoring method for lithium battery recycling processes combined with deep learning, including the following steps:

[0040] S1. Deploy multimodal sensors to collect multimodal data, preprocess non-image multimodal data, perform equal-pixel segmentation on battery surface images and material particle distribution images, calculate Gibbs-Shannon entropy and combine it with flexibility index to obtain denoising intensity, initialize filtering parameters, construct initial optimization objective function, use meta-dynamics for optimization, obtain optimal filtering parameters for denoising, segment the battery damage area and material particle images through U-Net model, and combine the preprocessed non-image multimodal data to generate a multimodal dataset;

[0041] Specifically, deploying multimodal sensors to collect multimodal data and preprocessing non-image multimodal data includes:

[0042] Deploy multimodal sensors to collect multimodal data and preprocess non-image multimodal data;

[0043] The multimodal sensors include industrial cameras, temperature sensors, voltage sensors, current sensors, chemical sensors, pressure sensors, acoustic sensors, and humidity sensors.

[0044] The multimodal data includes battery surface images, material particle distribution images, battery temperature, voltage, current, chemical, pressure, acoustic signals, humidity, and air pressure data.

[0045] The chemistry includes the content of lithium, cobalt, and nickel;

[0046] The preprocessing of non-image multimodal data includes denoising the non-image multimodal data using Savitzky-Golay filtering, normalizing the data, and extracting frequency domain features from the normalized acoustic signal using FFT.

[0047] By deploying multimodal sensors and preprocessing non-image multimodal data, high-fidelity fusion of multi-source information is achieved. By introducing multi-dimensional sensing modules including industrial cameras, temperature, voltage, current, chemical, pressure, acoustic, and humidity sensors, a composite sensing system covering electrochemistry, thermodynamics, mechanics, and acoustics is formed. This system enables the synchronous acquisition of multi-dimensional features of the battery during charging and discharging, providing rich input information for subsequent feature mining. The use of Savitzky-Golay filters to denoise non-image signals not only effectively maintains the continuity of signal morphology and peak structure characteristics but also reduces random noise interference, making the spectral features of time-series signals more physically interpretable. By extracting frequency domain features of acoustic signals through normalization and FFT, information expansion from the time domain to the frequency domain is achieved, making the feature dimensions more discriminative and thus enhancing the sensitivity and robustness of internal battery anomaly detection.

[0048] Furthermore, by combining the preprocessed non-image multimodal data, a multimodal dataset is generated, including:

[0049] The battery surface image and the material particle distribution image are segmented into non-overlapping image blocks using equal pixel segmentation. A gray-level histogram is calculated for each image block using a direct traversal statistical method. The counts in the gray-level histogram are normalized to obtain the gray-level probability. The Gibbs-Shannon entropy is calculated, and the denoising intensity is obtained by combining it with a flexibility index. The formula is as follows:

[0050] ,

[0051] ,

[0052] in, As a flexible indicator, For the quantity of material types, For materials Flexible weights, For materials The proportion of chemical components represents the percentage of normalized chemical data. For noise reduction intensity, The initial denoising intensity, For adjustment coefficients, This represents the maximum value of the Gibbs–Shannon entropy. Gibbs–Shannon entropy;

[0053] Based on the Gibbs–Shannon entropy, the filter parameters are initialized, including the dynamic window size and the dynamic polynomial order, which are consistent with the above formula for calculating the denoising intensity.

[0054] Based on the denoising intensity, an initial optimization objective function is constructed, as follows:

[0055] ,

[0056] in, To initially optimize the objective function, For dynamic window size, For the order of the dynamic polynomial, , as well as These are the weighting coefficients. Mean square error, and For the denoised image and the reference image, Gibbs–Shannon entropy for the denoised image, The noise reduction intensity for practical applications;

[0057] The initial filter parameters are optimized using meta-dynamics, stopping when the maximum number of iterations is reached. The initial objective function value is calculated for each iteration, sorted in descending order, and the filter parameters corresponding to the minimum initial objective function value are selected. Noise is then denoised from the image patches, and Gaussian smoothing is applied to the patch boundaries to generate smooth battery surface and material particle distribution images. Contrast stretching is used to enhance these images, and the U-Net model is used for segmentation to obtain images of the damaged battery area and material particles. The formula is:

[0058] ,

[0059] in, These are the denoised pixel values ​​within the image patch. The size of the dynamic window obtained by filtering. window offset position The weighting factor, For the pixels of the image block, The horizontal coordinate is... Vertical coordinates;

[0060] Normalized non-image multimodal data, frequency features, battery damage areas, and material particle images are time-stamped and then horizontally arranged to generate a multimodal dataset.

[0061] By combining image patch segmentation with Gibbs-Shannon entropy calculation, a quantitative correlation between image statistical characteristics and information entropy was established. This not only enhances the expressive power of image structural information but also enables image noise intensity to be dynamically mapped as denoising parameters in the form of information entropy. This achieves adaptive filtering adjustment based on a flexible index, which integrates material composition ratios and chemical weights, reflecting the differentiated responses of different battery materials to denoising strategies. The constructed optimization objective function unifies denoising effect, entropy change, and practical application parameters within an adjustable constraint system, achieving multi-objective optimization of filtering parameters. Combined with the meta-dynamic optimization algorithm, it can search for the global optimum in complex non-convex spaces, avoiding the trap of traditional gradient descent methods in local optima. After optimization, Gaussian smoothing and contrast stretching are applied to the image patches, followed by region segmentation using a U-Net network. This accurately identifies battery damage areas and material particle boundaries, providing a refined input foundation for subsequent structure-performance correlation analysis.

[0062] S2. Based on the multimodal dataset, select multimodal features are extracted, their Pearson correlation coefficient with the target parameter is calculated, features with correlation coefficients exceeding the threshold are selected, and the weights are optimized using the EPMA algorithm to construct a selected multimodal feature matrix.

[0063] Specifically, based on the multimodal dataset, selected multimodal features are extracted, including:

[0064] Based on the multimodal dataset, selected multimodal features were extracted, including images of battery damage areas and material particles. The ResNet-50 model was used to extract image features and normalize them to obtain normalized image features, including battery surface texture and material particle morphology features.

[0065] For normalized temperature, voltage, current, humidity and air pressure data, missing values ​​are filled by linear interpolation. The time-dependent features are output by the LSTM model, and the features are spliced ​​and normalized to obtain normalized sensor features.

[0066] For the normalized pressure and frequency domain features, SPCA is used for feature extraction and normalization to obtain normalized features;

[0067] For normalized temperature and voltage data, the derivative of temperature with respect to voltage is calculated by differential thermal voltammetry. Spline interpolation is performed to obtain the derivative curve. Peak value, valley value, peak position, valley position, peak width, and area are extracted. The area is calculated using the trapezoidal rule. The extracted results are normalized to obtain normalized derivative features.

[0068] For normalized voltage and temperature data, a matrix is ​​constructed using horizontal splicing. The matrix is ​​then decomposed using SVD to obtain a left singular vector matrix. The first N columns of the left singular vectors are extracted and normalized to obtain normalized SVD features.

[0069] Accumulate and integrate the normalized current data to obtain the cumulative capacity. Calculate the derivative of the cumulative capacity with respect to the normalized voltage data. Using the same method as described above, generate the capacity derivative curve, extract the peak value and peak position, and normalize it to obtain the normalized capacity characteristics.

[0070] For normalized current data, the medium constant current segment is extracted. This involves calculating the ratio of the absolute value of the difference between time j+1 and time j to the time interval to obtain the current change rate. Based on empirical rules, a current change rate threshold is set, and continuous current segments with current change rates less than the threshold are selected as the medium constant current segment. Correspondingly, the normalized voltage data is used to obtain the constant current discharge time, which is then normalized to obtain the normalized constant current characteristic. The formula is as follows:

[0071] ,

[0072] in, The constant current discharge time, The time it takes for the voltage to drop to 2V. The time when the voltage first reaches 4.2V, where 2V and 4.2V are the classic operating range of lithium batteries;

[0073] The normalized image features, normalized sensor features, normalized features, normalized derivative features, normalized SVD features, normalized capacity features, and normalized constant current features are concatenated to obtain multimodal features.

[0074] Image features extracted using ResNet-50 provide a deep semantic representation of battery surface texture and particle distribution, enabling high-precision capture of material degradation patterns. LSTM models are used for time-series modeling of temperature, voltage, current, humidity, and air pressure data, allowing sensor features to fully reflect dynamic thermo-electric behavior. The SPCA method extracts sparse principal components from pressure and frequency features, effectively compressing data dimensionality while preserving key waveform changes. Differential thermal voltammetry combined with spline interpolation generates derivative curves, quantifying the differential changes in the temperature-voltage relationship and significantly improving the sensitivity to changes in electrochemical characteristics. SVD decomposition extracts the left singular vector matrix, enabling subspace projection of multi-source data and revealing potential coupling modes between voltage and temperature. Cumulative capacity and constant current section analysis further reflect the consistency of energy output and lifetime degradation patterns of the battery during the discharge phase. Multi-channel features, after normalization and splicing, form high-dimensional fusion features, providing structured input for multimodal learning models.

[0075] Furthermore, the weights are optimized using the EPMA algorithm to construct a carefully selected multimodal feature matrix, including:

[0076] The correlation coefficient between multimodal features and target parameters, such as health status and recovery rate, is calculated using the Pearson correlation coefficient formula. Multimodal features with correlation coefficients greater than a threshold are selected and retained. The weights of the retained multimodal features are optimized using the EPMA-enhanced marine predator algorithm to obtain selected multimodal features. The selected multimodal features are defined as columns and the batteries are defined as rows to construct a selected multimodal feature matrix.

[0077] By optimizing the weights of selected multimodal features using the EPMA algorithm, a global search model based on an evolutionary mechanism was established. Combining Levy flight and dynamic predation strategies, EPMA exhibits high convergence speed and the ability to escape local optima in a complex multimodal optimization space. After calculating the Pearson correlation coefficient, features related to the target parameters are selected, and their weight distribution is dynamically adjusted to achieve adaptive optimization of the feature subset. EPMA can simultaneously consider global correlation and suppress local redundancy, ensuring the optimal balance of information content, robustness, and generalization ability in the feature matrix. The final selected multimodal feature matrix not only improves the accuracy of battery state prediction and degradation modeling, but also provides an scalable feature expression framework for subsequent data-driven electrochemical mechanism research.

[0078] S3. Set the selected multimodal feature matrix as the initial low-rank feature matrix, optimize it by enhancing the Lagrange multiplier method to obtain the optimized low-rank feature matrix, use the Gaussian kernel function to calculate the similarity between samples to generate the weight matrix, construct the intra-class scatter matrix and the inter-class scatter matrix to obtain the projection matrix, project the optimized low-rank feature matrix to the low-dimensional space and normalize it to obtain the graph embedding feature matrix.

[0079] Specifically, the optimized low-rank feature matrix is ​​projected to a low-dimensional space and normalized to obtain the graph embedding feature matrix, including:

[0080] The selected multimodal feature matrix is ​​set as the initial low-rank feature matrix, and optimized using the enhanced Lagrange multiplier method. The optimization stops when the maximum number of iterations is reached, yielding the optimized low-rank feature matrix, as shown in the formula:

[0081] ,

[0082] ,

[0083] ,

[0084] in, For iteration The low-rank eigenma matrix, For proximal operators, and The penalty coefficient is... For nuclear norm, For iteration The sparse noise matrix, For iteration The Lagrange multiplier matrix, To select the best multimodal feature matrix, For iteration The sparse noise matrix, For iteration The Lagrange multiplier matrix;

[0085] The similarity between rows I and J of the optimized low-rank feature matrix is ​​calculated using a Gaussian kernel function to construct an initial weight matrix, where rows and columns represent different batteries, and matrix elements represent similarity. This generates within-class and between-class scatter matrices, resulting in the projection matrix, as shown in the formula:

[0086] ,

[0087] ,

[0088]

[0089] in, and These are the within-class scatter matrix and the between-class scatter matrix. The weight matrix is ​​the first row and number Column elements, For transpose, and For the first row and number eigenvectors of the column and For batteries of the same category and different categories, such as normal batteries and normal batteries versus abnormal batteries and normal batteries, For the projection matrix, Let be the initial projection matrix. The trace of a matrix represents the sum of the diagonal elements of the matrix.

[0090] The optimized low-rank feature matrix is ​​projected to a low-dimensional space using matrix multiplication using a projection matrix, and then normalized to obtain the graph embedding feature matrix.

[0091] By optimizing the low-rank feature matrix using the enhanced Lagrange multiplier method, the low-rank structure and sparse noise components in multimodal input data can be effectively separated. The synergistic constraints of the kernel norm and L1 norm ensure that the optimization process maintains the low-rank properties of the matrix while robustly resisting abnormal noise interference, thus preserving global consistency and local separability of features during multi-source data fusion. Stepwise updates implemented through proximal operators not only guarantee the convergence and stability of the algorithm but also reduce information redundancy in the high-dimensional feature space, improving computational efficiency. A Gaussian kernel function is used to calculate and optimize the similarity between rows of the low-rank feature matrix, constructing an initial weight matrix and further generating intra-class and inter-class scatter matrices to model the nonlinear correlation between battery features. The Gaussian kernel function can capture complex manifold structures in implicit high-dimensional space, allowing the geometric relationships between features to be accurately represented. By preserving the original feature structure information, the discriminative power of the projection matrix is ​​enhanced. Through the joint optimization criterion of maximizing inter-class divergence and minimizing intra-class divergence, the obtained projection matrix can achieve optimal separability of features in low-dimensional space while preserving the original feature structure information. This not only strengthens the boundary difference between normal and abnormal batteries in the feature space, but also improves the accuracy of subsequent classification, clustering and health status identification. After projecting the optimized low-rank feature matrix into the low-dimensional space and performing normalization processing, a graph embedding feature matrix is ​​formed, which further enhances the structural consistency and scale comparability among features. The normalization operation can avoid the bias caused by the inconsistency of the dimensions of features of different modalities, and ensure that the weights of each feature contribute in the graph embedding space are balanced. It not only has strong structural robustness and nonlinear expression ability, but also provides a high-quality input foundation for subsequent graph manifold modeling and health status prediction.

[0092] S4. Construct a manifold Laplacian matrix based on the graph embedding feature matrix, define the optimization objective function, iteratively optimize to obtain the low-rank graph coefficient matrix, generate the final fusion matrix, extract the final fusion features, perform health status prediction and fault detection, generate a monitoring report and upload it to the cloud for storage.

[0093] Specifically, the final fusion features are extracted to predict health status and detect faults, generating monitoring reports and uploading them to cloud storage, including:

[0094] Based on the graph embedding feature matrix, the manifold Laplacian matrix is ​​constructed as follows:

[0095] ,

[0096] ,

[0097] ,

[0098] in, This is the weight matrix. For the first The graph embedding features of the rows are extracted from the graph embedding feature matrix. For the first Column graph embedding features, For physical attribute weights, For Gaussian kernel weights, For the number of columns, For degree matrix, It is the Laplace matrix of the manifold;

[0099] Based on the flow-direction Laplacian matrix, an optimization objective function is defined, and the coefficient matrix of the low-rank graph is obtained through iterative optimization, as shown in the formula:

[0100] ,

[0101] ,

[0102] ,

[0103] ,

[0104] in, To optimize the objective function, This is the coefficient matrix of a low-rank graph. and For regularization parameters, For graph embedding features, Denotes the Frobenius norm. For manifold regularization, For iteration The coefficient matrix of the low-rank graph. For iteration The coefficient matrix of the low-rank graph. To optimize the objective function for the coefficient matrix of the low-rank graph gradient, For iteration The penalty coefficient, The maximum penalty coefficient, The rate of increase of the penalty coefficient. For iteration The penalty coefficient;

[0105] Based on the low-rank graph sparse matrix, feature scores are calculated, sorted in descending order, and the top L features corresponding to the graph embedding feature matrix are selected to obtain the selected feature matrix, as shown in the formula:

[0106] ,

[0107] in, For the first The feature score of the row, A sparse matrix of a low-rank graph The row and number Column elements;

[0108] The final fusion matrix is ​​obtained by multiplying the selected feature matrix and the projection matrix.

[0109] Construct an SVR model, including an input layer, a kernel function layer, a regression layer, and an output layer;

[0110] The SVR model was trained using the CALCE Battery Dataset dataset.

[0111] Construct a RF model, including an input layer, a decision tree layer, a voting layer, and an output layer;

[0112] The RF model was trained using the NASA Battery Fault Dataset.

[0113] Based on the final fusion matrix, the final fusion features are extracted, and the SVR model is used for prediction to obtain the predicted health status and recovery rate. RF is used for anomaly and fault detection to obtain the anomaly and fault probabilities.

[0114] The predicted health status, predicted recovery rate, anomaly probability, and failure probability are arranged vertically to generate a monitoring report, which is then uploaded to the cloud for storage.

[0115] By utilizing the manifold Laplacian matrix based on graph embedding features, the local neighborhood relationships of battery samples in a high-dimensional feature space can be effectively characterized. By introducing physical attribute weights and Gaussian kernel weights, the algorithm not only considers the geometric manifold distribution of data at the graph structure level but also integrates the intrinsic coupling relationship between electrochemical and physical features. This strengthens the physical neighborhood correlation while maintaining topological consistency. Through a multi-objective optimization strategy that introduces kernel norm regularization, Frobenius norm constraints, and manifold regularization terms, a unified modeling of feature sparsity, global similarity, and local geometric structure is achieved. The gradient update mechanism, combined with the dynamic adjustment of the adaptive penalty parameter ζ, not only accelerates the convergence process but also adaptively adjusts the low-rank constraint strength according to different feature distributions. Ensuring the global optimality of the optimization results in balancing sparsity and structural integrity allows for more accurate capture of potential patterns in battery performance evolution and reduces the interference of noise features on health assessment. By calculating and sorting the feature scores of the low-rank graph sparse matrix Z, automatic feature selection and compression can be achieved, effectively eliminating invalid or weakly correlated features and retaining the most discriminative feature subset. The fusion features combined with the projection matrix further enhance the complementarity between cross-modal features. SVR and RF models are used to perform health status prediction and anomaly detection, respectively. Through the collaborative modeling of linear and nonlinear feature relationships by different models, a high-precision multi-task learning framework is realized. After the prediction results are generated and uploaded to the cloud, intelligent monitoring and remote diagnosis of the battery system can be achieved.

[0116] Example 2, refer to Figure 2 As a second embodiment of the present invention, an intelligent monitoring system for lithium battery recycling process incorporating deep learning includes:

[0117] The data acquisition and processing module is used to deploy multimodal sensors to acquire battery images and multi-source physicochemical data, and to achieve data cleaning and enhancement through filtering, normalization and adaptive denoising optimization based on Gibbs-Shannon entropy.

[0118] The feature extraction and fusion module is used to extract multi-dimensional features from image and time-series sensor data based on deep models, and generate a selected multimodal feature matrix through feature stitching, correlation filtering and EPMA optimization.

[0119] The feature optimization and graph embedding module is used to optimize low-rank features using the enhanced Lagrange method and manifold regularization, construct graph embedding and manifold Laplacian matrix, and extract highly correlated low-dimensional embedding features.

[0120] The feature selection and fusion matrix generation module is used to calculate feature scores based on the sparse matrix of the low-rank graph, select high-contribution features, and combine them with the projection matrix to generate the final fusion feature matrix.

[0121] The intelligent prediction and upload module is used to predict the health status and recycling rate using SVR, and combine the anomaly and failure probability of RF detection to generate and upload cloud monitoring reports, thereby realizing intelligent monitoring of the lithium battery recycling process.

[0122] This embodiment also provides a computer device suitable for intelligent monitoring methods of lithium battery recycling processes combined with deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent monitoring method of lithium battery recycling processes combined with deep learning as proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium on which a computer program is stored. When executed by a processor, the program implements the intelligent monitoring method for lithium battery recycling process combined with deep learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent monitoring of lithium battery recycling processes combining deep learning, characterized in that: Includes the following steps: Multimodal sensors are deployed to collect multimodal data. Non-image multimodal data is preprocessed, and equal-pixel segmentation is performed on battery surface images and material particle distribution images. Gibbs-Shannon entropy is calculated and combined with flexibility index to obtain denoising intensity. Filtering parameters are initialized, an initial optimization objective function is constructed, and meta-dynamics is used for optimization to obtain the optimal filtering parameters for denoising. Battery damage area and material particle images are segmented through U-Net model. Combined with the preprocessed non-image multimodal data, a multimodal dataset is generated. Based on a multimodal dataset, select multimodal features are extracted, their Pearson correlation coefficients with the target parameters are calculated, features with correlation coefficients exceeding the threshold are selected, and the weights are optimized using the EPMA algorithm to construct a selected multimodal feature matrix. The selected multimodal feature matrix is ​​set as the initial low-rank feature matrix. The optimized low-rank feature matrix is ​​obtained by optimizing it using the enhanced Lagrange multiplier method. The similarity between samples is calculated using the Gaussian kernel function to generate the weight matrix. The intra-class scatter matrix and inter-class scatter matrix are constructed to obtain the projection matrix. The optimized low-rank feature matrix is ​​projected to a low-dimensional space and normalized to obtain the graph embedding feature matrix. Based on the graph embedding feature matrix, a manifold Laplacian matrix is ​​constructed, an optimization objective function is defined, and a low-rank graph coefficient matrix is ​​obtained through iterative optimization. The final fusion matrix is ​​generated, the final fusion features are extracted, and health status prediction and fault detection are performed. A monitoring report is generated and uploaded to the cloud for storage.

2. The intelligent monitoring method for lithium battery recycling process combined with deep learning as described in claim 1, characterized in that: The process of combining preprocessed non-image multimodal data to generate a multimodal dataset includes: The battery surface image and the material particle distribution image are segmented into non-overlapping image blocks using equal pixel segmentation. A gray-level histogram is calculated for each image block using a direct traversal statistical method. The counts in the gray-level histogram are normalized to obtain the gray-level probability. Gibbs-Shannon entropy is calculated, and the denoising intensity is obtained by combining it with a flexibility index. An initial optimization objective function is constructed, and the initial filtering parameters are optimized using meta-dynamics. The image blocks are then denoised, and the U-Net model is used for segmentation to obtain images of the battery damage area and material particles. Timestamp alignment is performed by combining normalized non-image multimodal data and frequency features, and a multimodal dataset is generated by horizontal arrangement.

3. The intelligent monitoring method for lithium battery recycling process combined with deep learning as described in claim 2, characterized in that: The extraction of selected multimodal features based on the multimodal dataset includes: Based on the multimodal dataset, selected multimodal features are extracted, including the concatenation of normalized image features, normalized sensor features, normalized features, normalized derivative features, normalized SVD features, normalized capacity features, and normalized constant current features to obtain multimodal features.

4. The intelligent monitoring method for lithium battery recycling process combined with deep learning as described in claim 3, characterized in that: The step of optimizing weights using the EPMA algorithm to construct a carefully selected multimodal feature matrix includes: The correlation coefficient between multimodal features and target parameters is calculated using the Pearson correlation coefficient formula. Multimodal features with correlation coefficients greater than the correlation coefficient threshold are selected for retention. The weights of the retained multimodal features are then optimized using the EPMA algorithm to obtain a selected multimodal feature matrix.

5. The intelligent monitoring method for lithium battery recycling process combined with deep learning as described in claim 4, characterized in that: The step of projecting the optimized low-rank feature matrix into a low-dimensional space and normalizing it to obtain the graph embedding feature matrix includes: The selected multimodal feature matrix is ​​set as the initial low-rank feature matrix. It is optimized by the enhanced Lagrange multiplier method and stops when the maximum number of iterations is reached to obtain the optimized low-rank feature matrix. The similarity of the optimized low-rank feature matrix is ​​calculated using the Gaussian kernel function to construct the initial weight matrix, generate the intra-class scatter matrix and inter-class scatter matrix, and obtain the projection matrix. The optimized low-rank feature matrix is ​​projected to a low-dimensional space using the projection matrix to obtain the graph embedding feature matrix.

6. The intelligent monitoring method for lithium battery recycling process combined with deep learning as described in claim 5, characterized in that: The process of extracting the final fusion features, performing health status prediction and fault detection, generating a monitoring report, and uploading it to the cloud for storage includes: Based on the graph embedding feature matrix, a manifold Laplacian matrix is ​​constructed, an optimization objective function is defined, and a low-rank graph coefficient matrix is ​​obtained through iterative optimization. Feature scores are calculated to obtain a filter feature matrix. The final fusion matrix is ​​obtained by combining the projection matrix with matrix multiplication. The final fusion features are extracted, and the SVR model is used for prediction to obtain the predicted health status and recovery rate. RF is used for anomaly and fault detection to obtain the anomaly and fault probabilities. A monitoring report is generated and uploaded to the cloud for storage.

7. The intelligent monitoring method for lithium battery recycling process combined with deep learning as described in claim 1, characterized in that: The deployment of multimodal sensors collects multimodal data, and preprocesses the non-image multimodal data, including: Multimodal sensors are deployed to collect multimodal data. Non-image multimodal data is preprocessed, including denoising the non-image multimodal data using Savitzky-Golay filtering and normalizing it. Frequency domain features are extracted from the normalized acoustic signal using FFT.

8. A deep learning-based intelligent monitoring system for lithium battery recycling processes, used to implement the deep learning-based intelligent monitoring method for lithium battery recycling processes as described in any one of claims 1 to 7, characterized in that: include: The data acquisition and processing module is used to deploy multimodal sensors to acquire battery images and multi-source physicochemical data, and to achieve data cleaning and enhancement through filtering, normalization and adaptive denoising optimization based on Gibbs-Shannon entropy. The feature extraction and fusion module is used to extract multi-dimensional features from image and time-series sensor data based on deep models, and generate a selected multimodal feature matrix through feature stitching, correlation filtering and EPMA optimization. The feature optimization and graph embedding module is used to optimize low-rank features using the enhanced Lagrange method and manifold regularization, construct graph embedding and manifold Laplacian matrix, and extract highly correlated low-dimensional embedding features. The feature selection and fusion matrix generation module is used to calculate feature scores based on the sparse matrix of the low-rank graph, select high-contribution features, and combine them with the projection matrix to generate the final fusion feature matrix. The intelligent prediction and upload module is used to predict the health status and recycling rate using SVR, and combine the anomaly and failure probability of RF detection to generate and upload cloud monitoring reports, thereby realizing intelligent monitoring of the lithium battery recycling process.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent monitoring method for lithium battery recycling process combined with deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent monitoring method for lithium battery recycling process combined with deep learning as described in any one of claims 1 to 7.

Images