An adaptive sorting method and system for waste lithium battery recycling pretreatment

Abstract

This invention belongs to the field of lithium battery recycling technology and relates to an adaptive sorting method and system for pre-processing waste lithium battery recycling. The method involves: acquiring images of the batch of crushed lithium battery material to be processed; inputting these images into a lithium battery full-component segmentation model to output a material composition vector containing component categories and area percentages; inputting the material composition vector into a pre-trained process parameter prediction model to output a combination of sorting process parameters; and dynamically adjusting the operating parameters of the sorting equipment based on this parameter combination to complete the sorting. The system includes an image acquisition module, a vision processing module, a parameter prediction module, and a control adjustment module, with each module working together to achieve adaptive and precise sorting. This invention can acquire detailed compositional information of the crushed waste lithium battery material in real time, achieving precise quantitative characterization of all components and dynamic control of sorting parameters, thus improving the separation purity and recycling efficiency of each component.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery recycling technology, specifically relating to an adaptive sorting method and system for the pretreatment of waste lithium batteries. Background Technology

[0002] With the explosive growth of the new energy vehicle industry, the number of retired lithium batteries in the market has been increasing year by year. Waste lithium batteries contain a large amount of valuable metals and chemical components. Randomly discarding them not only wastes mineral resources but also easily causes environmental pollution and safety hazards. Therefore, the standardized recycling and utilization of waste lithium batteries has become an industry necessity.

[0003] The pretreatment of waste lithium batteries mainly includes key processes such as discharge, dismantling, crushing, and sorting. Among them, the sorting process plays a core separation role, and its goal is to efficiently separate copper foil, aluminum foil, separator, shell plastic and black powder from the crushed material, so as to provide high-purity raw materials for subsequent hydrometallurgical or pyrometallurgical processes.

[0004] However, in actual production, the diverse sources of waste batteries result in significant fluctuations in the composition of different batches of crushed materials. Traditional sorting equipment mostly operates with fixed parameters, using preset, uniform process parameters for airflow separation, magnetic separation, and eddy current separation for different batches of crushed materials. This fails to adapt to the dynamic fluctuations in the composition characteristics of the crushed materials, easily leading to unstable sorting efficiency and large fluctuations in the purity of each component. Simultaneously, existing pretreatment process control methods are relatively lagging, unable to dynamically adapt to changes in the characteristics of the crushed materials in real time, and unable to provide timely warnings of abnormal component conditions, making it difficult to meet the current development needs of the lithium-ion battery recycling industry.

[0005] For example, patent application CN112718800A discloses a method for recycling and processing waste lithium batteries. This method uses a dry crushing and sorting process, employing a graded crushing process involving primary, secondary, and tertiary crushing to separate and recover the components in the waste lithium batteries, thereby improving the component recovery rate to a certain extent. Another example is patent application CN118367253A, which discloses a method for separating and recovering copper and aluminum from waste lithium batteries. This method uses a low-energy mechanical physical processing approach, wet ball milling the crushed copper and aluminum particles, followed by fine screening and wet gravity separation. This method features a short process flow, low energy consumption, and high purity and recovery rate of copper and aluminum. While the aforementioned existing technologies can achieve the separation and recovery of components after crushing waste lithium batteries and each has its own technological advantages, none of them consider the dynamic fluctuation characteristics of the component properties of different batches of crushed materials. They still cannot adapt to real-time changes in material components, inevitably resulting in unstable sorting efficiency and large fluctuations in the purity of each component. Furthermore, they lack the ability to dynamically adapt process parameters and provide early warning of component anomalies.

[0006] Therefore, it is necessary to propose an adaptive sorting method and system for the pretreatment of waste lithium batteries to solve the above-mentioned technical problems existing in the prior art. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies and provide an adaptive sorting method and system for the pretreatment of recycled waste lithium batteries.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: An adaptive sorting method for pretreatment of waste lithium batteries includes the following steps: Acquire images of the current batch of crushed lithium battery material awaiting processing; The image is input into the lithium battery full component segmentation model, which outputs a material composition vector representing the batch of lithium battery crushed material to be processed. The material composition vector includes the category and area ratio of each component in the material. The lithium battery full component segmentation model is used to accurately identify and quantify the components of lithium battery crushed material in real time. The material composition vector is input into a pre-trained process parameter prediction model, which outputs a combination of sorting process parameters that match the characteristics of the broken lithium battery material in the batch to be processed; thus enabling the process parameters to be flexibly adapted and adjusted according to the material characteristics. The control system dynamically adjusts the operating parameters of the sorting equipment based on the combination of sorting process parameters to sort the batches of crushed lithium battery materials to be processed.

[0009] As a preferred technical solution: The adaptive sorting method for waste lithium battery recycling pretreatment described above uses an improved UNet model for the full-component segmentation of lithium batteries, which includes an encoder, a decoder, and a detail-enhancing convolutional module. The encoder is used to extract multi-level features from the input image and output multi-scale feature maps; The decoder is used to upsample multi-scale feature maps step by step to restore image resolution, and during the sampling process, it fuses shallow features from the corresponding level of the encoder through feature splicing to output a pixel-level image segmentation mask. The detail enhancement convolution module is located on the feature stitching path between the encoder and decoder. It includes at least one convolution operation to enhance the edge features at the boundaries of different regions (i.e., different components in the image of broken lithium battery materials) in the multi-scale feature map before feature fusion.

[0010] As described above, the adaptive sorting method for preprocessing waste lithium battery recycling involves an encoder built on a backbone network and an attention mechanism module integrated after at least one feature extraction level. The attention mechanism module is a convolutional attention module, which includes a channel attention submodule and a spatial attention submodule set sequentially. The channel attention submodule uses global pooling and a multilayer perceptron to weight the channel dimensions of the multi-scale feature map, focusing on the feature channels used for material component identification; the spatial attention submodule performs pooling and convolution operations along the channel dimensions of the multi-scale feature map, weights the spatial location of the multi-scale feature map, and focuses on the region where the material component edges are located in the image.

[0011] As described above, in the adaptive sorting method for pretreatment of waste lithium batteries, the output of the attention mechanism module is equipped with an edge enhancement module. The edge enhancement module consists of a 3×3 learnable convolutional layer and residual connections. The edge enhancement module is used to learn autonomously and enhance the edge features at the junction of different regions.

[0012] The adaptive sorting method for waste lithium battery recycling pretreatment described above, the improved UNet model also includes a feature fusion layer; The feature fusion layer is placed after the encoder outputs the multi-scale feature map and before the decoder performs upsampling. The feature fusion layer can be either a feature pyramid network structure or a high-resolution network fusion strategy. When the feature fusion layer adopts a feature pyramid network structure, it performs top-down feature fusion, fusing deep high-semantic features with shallow high-resolution features step by step to obtain multi-scale fused features for use by the decoder. When the feature fusion layer adopts a high-resolution network fusion strategy, the multi-scale feature maps of each layer are upsampled to the highest resolution, and then feature fusion is completed by element-wise addition to obtain a fused multi-scale feature map that maintains high resolution, so as to optimize the segmentation of small sub-segments.

[0013] The adaptive sorting method for pretreatment of waste lithium batteries described above includes a process parameter prediction model comprising a feature extraction layer, a correlation analysis layer, and a parameter output layer connected in sequence. The feature extraction layer is used to perform nonlinear transformations on the input material composition vector to extract deep features; The correlation analysis layer has a built-in physical characteristic mapping rule library, which is used to map material composition characteristics to the adjustable physical parameter range of the sorting equipment, and to dynamically weight and correct the fixed parameters of the sorting process in combination with material metadata. The parameter output layer is used to output recommended values ​​for the combination of sorting process parameters.

[0014] The adaptive sorting method for pretreatment of waste lithium batteries described above has a material metadata association correction logic in the process parameter prediction model. The correction logic can dynamically weight and correct the material composition vector or initial process parameters based on the material metadata, such as key information affecting the material composition characteristics, such as battery type, cell capacity, cycle number, and disassembly method.

[0015] The adaptive sorting method for pretreatment of recycled waste lithium batteries described above uses a process parameter prediction model trained using machine learning methods. The training process includes: Collect historical production data, which should include at least the material composition vector of the crushed lithium battery materials from historical batches, the actual combination of sorting process parameters used, and the corresponding sorting effect data. Using material composition vectors as input features and combinations of sorting process parameters as training labels, the model is validated by combining sorting effect data, and a process parameter prediction model is constructed and trained using historical production data.

[0016] The adaptive sorting method for pretreatment of waste lithium batteries as described above further includes the following steps: Collect actual sorting effect data of the batch of lithium battery crushed materials to be processed after sorting, including at least one of the following: copper recovery rate, aluminum recovery rate, black powder purity, separator removal rate, magnetic material residue, and energy consumption of sorting process. The actual sorting effect data, material composition vector, and sorting process parameters are combined to construct training sample data; The process parameter prediction model is retrained or its parameters are adjusted using training sample data to complete the continuous iterative optimization of the model.

[0017] The adaptive sorting method for pretreatment of waste lithium batteries as described above further includes the following steps: The area percentage of each component in the material composition vector is compared with the corresponding preset threshold. If the area percentage of any component exceeds the corresponding preset threshold, an abnormal warning signal is triggered, and a prompt message containing the category of the abnormal component and the value of the excess is generated.

[0018] An adaptive sorting system for pretreatment of waste lithium batteries, used to implement the method described in any of the preceding claims, includes: an image acquisition module, a vision processing module, a parameter prediction module, and a control adjustment module; The image acquisition module is used to acquire images of the batch of crushed lithium battery materials to be processed; The vision processing module is used to input images into the lithium battery full-component segmentation model for processing, identify the component categories of the crushed material and obtain the identification results, calculate the area ratio of each component based on the identification results, and construct a material composition vector that characterizes the characteristics of the crushed lithium battery material in the batch to be processed. The parameter prediction module is used to input the material composition vector into the pre-trained process parameter prediction model. The process parameter prediction model analyzes the material composition vector as input and outputs a combination of sorting process parameters that matches the characteristics of the lithium battery crushed material in the batch to be processed and is used to control the operation of the sorting equipment. The control adjustment module is used to send the combination of sorting process parameters to the control system of the sorting equipment, so that the control system can dynamically adjust the operating parameters of the sorting equipment according to the combination of sorting process parameters to perform sorting of the batch of crushed lithium battery materials to be processed.

[0019] Beneficial effects The adaptive sorting method and system for pretreatment of waste lithium battery recycling proposed in this invention can accurately identify and quantitatively characterize the components of crushed waste lithium battery materials, and can obtain the component composition details of each batch of crushed materials in real time, thereby improving the accuracy and timeliness of crushed material characteristic analysis.

[0020] The process parameter prediction model in this invention relies on material composition vectors combined with historical production data to output a combination of sorting process parameters that highly matches the composition characteristics of the current batch of crushed materials. At the same time, it combines material metadata to construct dynamic correction logic, enabling process parameters to be flexibly adapted and adjusted according to battery type, cell capacity, cycle number, and disassembly method, so that the sorting process is accurately matched with the characteristics of the crushed materials, effectively improving the purity of material sorting and resource recovery rate.

[0021] This invention can dynamically adjust the operating parameters of the sorting equipment according to the combination of sorting process parameters, realize the automated and intelligent control of the sorting equipment, avoid subjective deviations and operational errors caused by manual intervention, and greatly improve the automation level and overall operating efficiency of the recycling production process. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the steps of an adaptive sorting method for pretreatment of waste lithium batteries according to the present invention; Figure 2 This is a schematic diagram of the module structure of an adaptive sorting system for pretreatment of waste lithium batteries according to the present invention; Figure 3 This is a schematic diagram of the overall process of an adaptive sorting method for pretreatment of waste lithium batteries according to the present invention; Figure 4 This is a schematic diagram of the processing flow of the lithium battery full-component segmentation model in the adaptive sorting method for waste lithium battery recycling pretreatment of the present invention; Figure 5 This is a schematic diagram of the process parameter prediction model in the adaptive sorting method for pretreatment of waste lithium batteries according to the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0024] The following reference Figure 1-5 The present invention will be further described below.

[0025] An adaptive sorting method for pretreatment of waste lithium batteries includes the following steps: Step S1: First, perform the image acquisition operation of the current batch of crushed lithium battery materials to be processed.

[0026] A linear array CMOS camera is installed above the conveyor belt for crushed materials. It uses a combination of visible light and near-infrared dual light sources for illumination, with an image resolution of no less than 0.1 mm / pixel to enhance the contrast between metal, plastic, and black powder, and to acquire images of the current batch of crushed lithium battery materials to be processed. Step S2: Input the image into the pre-trained lithium battery full-component segmentation model for processing.

[0027] The pre-trained lithium battery full-component segmentation model is an improved UNet model, including an encoder, decoder, and detail-enhancing convolutional module; The encoder is used to perform multi-level feature extraction on the input image of crushed lithium battery materials, and output multi-scale feature maps. The multi-scale feature maps are image feature maps of different resolutions. Specifically, the encoder is used to perform multi-level feature extraction operations on the input image of crushed lithium battery materials, and gradually compresses the spatial size of the image through multiple rounds of convolution and downsampling processing, while continuously improving the semantic abstraction of the features, and finally outputs multi-scale feature maps. The multi-scale feature maps are image feature maps of different resolutions, including shallow texture edge details and semantic category abstraction information.

[0028] The decoder is used to progressively upsample multi-scale feature maps to restore image resolution. During the upsampling process, it fuses shallow features from the corresponding level of the encoder through feature concatenation to output a pixel-level image segmentation mask. Specifically, the decoder progressively upsamples the multi-scale feature maps output by the encoder, gradually restoring the spatial resolution of the image through deconvolution or interpolation, so that the size of the feature maps gradually matches the size of the original input image. In each upsampling level, the decoder fuses the deep features of the current level with the shallow features output by the corresponding level of the encoder through feature concatenation. The shallow features can supplement the deep features with rich details, effectively avoiding the loss of details during the upsampling process. Finally, the decoder outputs a pixel-level image segmentation mask to achieve pixel-level localization and classification of various components in the broken lithium battery materials.

[0029] The detail enhancement convolution module is integrated into the feature concatenation path between the encoder and decoder. This module contains at least one convolution operation to enhance edge features at the boundaries of different regions in the multi-scale feature map before feature fusion. Specifically, before fusing the shallow features output by the encoder with the deep features by the decoder, targeted convolution processing is performed on the shallow features. Learnable convolution kernels are used to enhance the detailed feature representation of component edges and textures, highlighting the boundary features between different components, weakening interference from irrelevant backgrounds, and ensuring that the fused features retain the semantic information of the deep features while possessing clearer detailed features. This improves the model's segmentation accuracy for small component edges and ensures the accuracy and completeness of component recognition results.

[0030] The encoder is built upon a backbone network. Based on its feature extraction capabilities, the encoder performs multi-level feature extraction operations on the input image of broken lithium battery materials. An attention mechanism module, specifically a convolutional attention module, is integrated after at least one feature extraction level to enhance the model's ability to capture component features and improve the accuracy of component identification. The convolutional attention module comprises sequentially arranged channel attention and spatial attention sub-modules, performing weighted optimization of multi-scale feature maps from both channel and spatial dimensions. The channel attention sub-module performs global pooling on the input multi-scale feature map, compressing spatial features into channel-dimensional global features. A multilayer perceptron is then used to perform a non-linear transformation on the global features, learning the importance weights of different feature channels. These importance weights are then applied to the channel dimension of the original multi-scale feature map, weighting each feature channel to strengthen the feature channels crucial for component identification and suppress redundant feature channels irrelevant to component identification, thereby enhancing the expressive power of effective features. The spatial attention submodule is activated after the channel attention submodule has finished processing. First, it performs pooling operations along the channel dimension on the multi-scale feature map to extract global features in the spatial dimension. Then, it learns the importance weights of spatial locations through convolution operations and applies these weights to the spatial location dimension of the multi-scale feature map. This weighted processing of features at different spatial locations focuses on the regions where components are located in the image, weakens the feature interference between the background region and the effective component region, and enables the model to more accurately locate component regions, thereby improving the accuracy and completeness of component segmentation.

[0031] The attention mechanism module's output includes an edge enhancement module, which consists of a 3×3 learnable convolutional layer and residual connections. This edge enhancement module autonomously learns and enhances edge features at the boundaries of different regions, i.e., the edge features of the boundary regions between different components. Specifically, after the attention mechanism module completes the feature weighting for channel and spatial dimensions, it inputs a multi-scale feature map into the edge enhancement module. The learnable convolutional layer, through continuously optimized kernel parameters during model training, autonomously learns the edge feature patterns of the boundary regions between different components, extracting and enhancing the edge information of the multi-scale feature map. This highlights the boundary contours between different components, effectively solving the segmentation bias problem caused by blurred component edges and adhesion between small components and the background. The residual connections fuse the original multi-scale feature map input to the edge enhancement module with the enhanced feature map output by the convolutional layer. This preserves the effective information in the original features, avoiding feature loss due to convolution operations, while also enhancing the expressive power of edge features. This allows the model to capture the complete contours of various components, significantly improving the accuracy of pixel-level segmentation.

[0032] The improved UNet model includes a feature fusion layer after the encoder outputs multi-scale feature maps and before the decoder begins upsampling. This feature fusion layer can be either a feature pyramid network structure or a high-resolution network fusion strategy, with each strategy adapting to different component segmentation requirements and providing fused features to the decoder. When the feature fusion layer adopts a feature pyramid network structure, it performs top-down feature fusion, fusing deep high-semantic features with shallow high-resolution features step by step to generate multi-scale fused features that combine semantic and detail information for the decoder. Specifically, when the feature fusion layer adopts a feature pyramid network structure, it performs a top-down feature fusion process. In the multi-scale features output by the encoder, the deep features have rich high-semantic information and can identify the category attributes of the components. The shallow features retain complete high-resolution detail information and clearly present the edges and textures of the components. The feature fusion layer fuses the deep high-semantic features with the shallow high-resolution features step by step. Each level of fusion combines the upsampled deep features with the corresponding level of shallow features to generate multi-scale fused features that combine semantic and detail information. The multi-scale fused features are sent to the decoder, enabling the decoder to distinguish different components based on deep semantic features and to outline the components with shallow detail features during the process of restoring image resolution. For example, for the conventional components of broken materials such as copper foil, aluminum foil, and diaphragm, this structure can completely segment the overall shape of the components while ensuring the accuracy of category recognition, thus improving the segmentation effect of conventional components. When the feature fusion layer adopts a high-resolution network fusion strategy, the multi-scale feature maps of each level are upsampled to the highest resolution, and then feature fusion is completed through element-wise addition to generate and maintain a high-resolution fused multi-scale feature map, thereby optimizing the segmentation of small particles. Specifically, when the feature fusion layer adopts a high-resolution network fusion strategy, the multi-scale feature maps of all levels output by the encoder are uniformly upsampled to the highest resolution, so that the features of different levels have the same high resolution as the original input image. All high-resolution feature maps are fused through element-wise addition to generate and maintain a high-resolution fused multi-scale feature map. The fused multi-scale feature map retains rich detail information throughout the process, which can effectively capture the weak features of small particles and optimize the segmentation effect of small particles. For example, for black powder in crushed materials, the problem of missing segmentation of small particles due to insufficient feature resolution is avoided through high-resolution feature fusion throughout the process, thereby improving the recognition and segmentation accuracy of small particles.

[0033] Based on the recognition results of the improved UNet model, the area proportion of each component is calculated, thereby constructing a material composition vector characterizing the properties of the batch of crushed lithium battery materials to be processed. The material composition vector includes the category and area proportion of each component in the material. Specifically, the area proportion of each component is calculated based on the recognition results of pixel-level image segmentation mask. The pixel-level image segmentation mask of the recognition results labels the category attribute corresponding to each pixel of the image, including the category of each component and the background category. First, the total number of pixels corresponding to each category of component in the image is counted, and then the total number of pixels in the entire acquired image is counted. The total number of pixels corresponding to each category of component is divided by the total number of pixels in the image to obtain the area proportion of that category of component in the current batch of crushed lithium battery materials to be processed. This area proportion can truly reflect the actual proportion of each component in the current batch of crushed lithium battery materials to be processed. The area proportion of each category of component is arranged in an orderly manner according to the preset component category order to form a structured numerical vector, which is the material composition vector that can completely characterize the composition of the current batch of crushed lithium battery materials to be processed. The material composition vector transforms the visual component distribution information into quantitative features that can be used for subsequent process analysis, realizing the digital characterization of the component properties of the batch of crushed lithium battery materials to be processed.

[0034] For example, after a batch of crushed lithium battery materials is identified by the improved UNet model, copper foil, aluminum foil, separator, outer plastic casing, and black powder components are distinguished. By statistically analyzing the number of pixels corresponding to each component and the total number of pixels in the image, the area percentage of each component is calculated in the format: [copper, aluminum, separator, outer plastic casing, black powder]. These percentage values ​​are arranged in order according to the preset component category order to form the material composition vector of the batch of crushed lithium battery materials to be processed. For example, the material composition vector of a certain batch is [12%, 8%, 15%, 10%, 55%]. The visual component distribution information is transformed into structured data that can be used for process analysis.

[0035] Step S3: After receiving the material composition vector, the process parameter prediction model matches the corresponding sorting process parameters according to the proportion of each component.

[0036] For example, given the high proportion of copper in this batch, the frequency of the eddy current separator is increased to enhance the repulsion force on copper particles, and the airflow velocity is appropriately reduced to prevent fine copper powder from being accidentally blown into the lighter components. This adapts to the actual working conditions of the current material and ultimately outputs a combination of sorting process parameters that are highly matched to the characteristics of the batch, thereby achieving dynamic adaptive control of the lithium battery recycling pretreatment process.

[0037] The pre-trained process parameter prediction model is obtained through machine learning methods. The training process of the process parameter prediction model includes: Collect historical production data, which should include at least the material composition vector of the crushed materials in historical batches, the actual combination of sorting process parameters used, and the corresponding sorting effect data. Using material composition vectors as input features and combining sorting effect data for verification, sorting process parameter combinations are used as training labels, and historical production data is used to construct and train a process parameter prediction model.

[0038] Specifically, the collected historical production data includes material composition vectors of historical batches of crushed materials and the actual combinations of sorting process parameters used. The material composition vectors are a quantitative representation of the area proportion of various components in the historical batches of materials, while the combinations of sorting process parameters include parameters such as air classifier wind speed, magnetic separator magnetic field strength, eddy current separator frequency, and vibrating screen frequency used in the actual production of the corresponding batches. Based on the historical production data, using the material composition vectors as input features, and combining them with the sorting effect data of the corresponding batches for validation, the optimal combination of sorting process parameters validated in actual production is used as training labels to construct a standardized training dataset. Based on this dataset, a process parameter prediction model is constructed and iteratively trained multiple times, enabling the model to learn the intrinsic relationship between component composition and process parameters, providing support for parameter prediction in actual production.

[0039] The process parameter prediction model consists of a feature extraction layer, a correlation analysis layer, and a parameter output layer connected in sequence. The feature extraction layer is used to perform nonlinear transformation on the input material composition vector to extract deep features. The correlation analysis layer has a built-in physical property mapping rule library, which is used to map the material composition features to the adjustable physical parameter range of the sorting equipment, and to dynamically weight and correct the fixed parameters of the sorting process in combination with material metadata. The parameter output layer is used to output recommended values ​​for the combination of sorting process parameters.

[0040] Specifically, the feature extraction layer receives the input material composition vector and performs a deep nonlinear transformation on it. Through a multi-layer neural network structure, it mines the potential correlations between the proportions of various components, extracting deep features that reflect the characteristics of the material components and the requirements of the pretreatment process. This overcomes the linear limitations of the original vector and significantly improves the representational ability of the features. For example, for a material composition vector with a high copper content, the feature extraction layer extracts the characteristic that the copper component has a high demand for eddy current separator frequency and air classifier wind speed parameters. The correlation analysis layer receives the deep features output by the feature extraction layer and combines them with material metadata for weighted correction. The process parameter prediction model includes material metadata correlation correction logic, which can dynamically weight and correct the material composition vector or initial process parameters based on the material metadata.

[0041] Material metadata includes key information that affects the material composition characteristics, such as battery type, cell capacity, cycle count, and disassembly method. Different battery types correspond to different component ratios, and the physical properties of components vary significantly under different cell capacities, cycle counts, and disassembly methods. The correlation analysis layer integrates this metadata into feature correction based on a preset physical property mapping rule library, making the deep features more consistent with actual working conditions.

[0042] Specifically, the correlation analysis layer first reads the metadata information of the current batch of crushed material, including battery type, cell capacity, cycle count, and dismantling method. Different dismantling methods directly affect the composition and physical state of the material. Materials from mechanical dismantling: The metal fragments are small and heavily mixed with black powder; the diaphragm is fibrous; there is a lot of residual black powder on the surface of the copper and aluminum foils; and the overall particle size distribution of the material is wide. Materials disassembled manually: metal sheets are relatively large and intact, diaphragms are large sheets, copper foil and aluminum foil surfaces are relatively clean, and the particle size distribution of the materials is concentrated.

[0043] Different battery types affect the metal composition and impurity content of the components: Materials for power batteries: The black powder contains high-value nickel, cobalt, and manganese, so it is necessary to prioritize the purity of the black powder and reduce metal inclusions; Materials for consumer batteries: The black powder contains cobalt, has a high proportion of plastic impurities, and contains small metal fragments. It is necessary to focus on removing plastic impurities and recycling small metal particles. Materials for energy storage batteries: The black powder is mainly lithium iron phosphate, which does not contain precious metals. The copper and aluminum foils are intact, and the focus of recycling is on the efficient separation of copper and aluminum metals.

[0044] Based on the aforementioned differential characteristics, the association analysis layer performs dynamic weighting on the features: For materials that are mechanically disassembled: the reasoning process is to increase the weight of fine metal components so that the model focuses on the recovery needs of difficult-to-separate fine components; increase the weight of fibrous membrane components to enhance the removal effect of air separation on light fibers; improve the detection sensitivity of fine metal inclusions in black powder and optimize the subsequent magnetic separation and eddy current separation parameters.

[0045] For materials that are manually disassembled: the reasoning process is to increase the weight of large metal components to enhance the recovery efficiency of eddy current separation for large metals; reduce the weight of fine particle separation parameters and appropriately reduce the separation intensity to achieve energy-saving operation; increase the weight of large membrane components and optimize the coordination between screening and air separation.

[0046] For materials used in power batteries: increase the weighting of parameters related to the purity of black powder, prioritizing the quality of nickel-cobalt-manganese black powder; increase the weighting of magnetic separation parameters, and enhance the removal of magnetic impurities.

[0047] For materials used in consumer batteries: increase the weighting of parameters related to plastic impurity removal and enhance the light component separation capability of air classifiers; increase the weighting of parameters related to fine metal recovery and optimize the frequency of eddy current separators; increase the weighting of parameters related to screening accuracy and use finer screen mesh sizes to separate fine components.

[0048] For materials used in energy storage batteries: increase the proportion of copper and aluminum recovery parameters, increase the frequency of eddy current separators, and reduce the need for magnetic separation.

[0049] After the above weighted adjustment, the weighted logic is transformed into specific executable parameter adjustment instructions, as shown in the following example: For materials that are mechanically dismantled: increase the frequency of the eddy current separator to enhance the recovery of fine metal particles, increase the wind speed of the air classifier to enhance the separation effect of the fibrous membrane, increase the frequency of the vibrating screen to optimize the fine particle separation accuracy, and reduce the screen aperture to reduce metal entrainment.

[0050] For materials sourced from manual dismantling: reduce the frequency of the eddy current separator, reduce the wind speed of the air classifier to reduce the loss of black powder due to accidental blowing, increase the screen aperture to prioritize the separation of large metal pieces, and reduce the load on subsequent sorting.

[0051] For materials used in power batteries: increase the magnetic field strength of the magnetic separator to enhance the removal of magnetic impurities such as nickel, optimize the frequency of the eddy current separator to protect the integrity of large copper foil sheets, and adjust the position of the distribution baffle to ensure the collection of high-purity black powder.

[0052] For materials used in consumer batteries: increase the wind speed of the air separator to enhance the separation of lightweight components such as plastics; increase the frequency of the eddy current separator to enhance the recovery of fine metal particles; and increase the frequency of the vibrating screen and reduce the screen aperture to improve the fine particle screening accuracy.

[0053] For materials in energy storage batteries: increase the frequency of the eddy current separator to the medium-high range to prioritize copper and aluminum recovery rates, reduce the magnetic field strength of the magnetic separator or shut it down to save energy, and increase the conveyor belt speed to improve processing efficiency.

[0054] After the above weighted correction, the correlation analysis layer outputs the corrected deep features to the parameter output layer.

[0055] The parameter output layer outputs recommended values ​​for the combination of sorting process parameters. These recommended values ​​include various key parameters required for the operation of the sorting equipment, ensuring a high degree of matching with the compositional characteristics of the current crushed material. These parameters include: air classifier wind speed (m / s), eddy current separator frequency (Hz), magnetic separator magnetic field strength (mT), vibrating screen frequency (Hz), screen aperture (mm), material distribution baffle position (%), and conveyor belt speed (m / s). These parameters can be directly used for the dynamic control of the sorting equipment to achieve intelligent adaptation to the pretreatment process of waste lithium battery recycling.

[0056] Step S4: Send the combination of sorting process parameters to the control system of the sorting equipment. The control system dynamically adjusts the operating parameters of the sorting equipment according to the combination of sorting process parameters to perform adaptive sorting of the current batch of crushed materials.

[0057] Specifically, the sorting process parameter combinations are sent to the control system of the sorting equipment via an industrial communication link. The control system directly interfaces with the actuators of various sorting equipment. After receiving the sorting process parameter combinations, the control system analyzes and verifies the parameter combinations to confirm their completeness and rationality. Based on the various indicators of the parameter combinations, it dynamically adjusts the various operating parameters of the sorting equipment. The parameter adjustment process covers key equipment throughout the pretreatment process, including airflow separators, eddy current separators, magnetic separators, and vibrating screens. Specifically, the airflow setpoint and the position of the material distribution baffle of the airflow separator are adjusted according to the parameter combinations; the operating frequency and the gap of the material distribution baffle of the eddy current separator are adjusted; the magnetic field strength, magnetic roller gap, and discharge scraper position of the magnetic separator are adjusted; and the screen aperture, vibration frequency, and screen inclination angle of the vibrating screen are adjusted to ensure that all operating parameters are highly compatible with the component characteristics of the current batch of crushed lithium battery material.

[0058] The sorting equipment operates according to updated parameters, performing adaptive sorting of the current batch of crushed lithium battery materials. During the sorting process, the control system monitors the equipment's operating status and material processing progress in real time, and can make adaptive fine-tuning adjustments to the operating parameters based on actual working conditions, ensuring the stability and efficiency of the sorting process. Through this dynamic control mechanism, while ensuring the separation effect of each component, energy consumption such as electricity and compressed air is optimized, effectively improving the sorting efficiency and the purity of each component recovered, and realizing intelligent management and control of the pre-treatment process for waste lithium battery recycling.

[0059] The adaptive sorting method for pretreatment of recycled lithium batteries also includes the following steps: The actual sorting effect data of the current batch of crushed lithium battery materials after sorting is collected. The actual sorting effect data includes at least one of the following: copper recovery rate, aluminum recovery rate, black powder purity, separator removal rate, magnetic material residue, and energy consumption of the sorting process. Among them, copper recovery rate directly reflects the separation and recovery effect of copper foil fragments and is the core indicator for evaluating sorting efficiency; aluminum recovery rate reflects the separation efficiency of aluminum foil components and directly affects the economic value of metal recovery; black powder purity characterizes the purity of positive and negative electrode active materials, and high-purity black powder is beneficial for subsequent hydrometallurgical and pyrometallurgical processes; separator removal rate reflects the separation effect between light organic components (separator, plastic) and heavy components (metal, black powder); magnetic material residue, such as iron and nickel content, is used to measure the impurity removal depth of the magnetic separation process and directly affects the quality of recycled materials; energy consumption of the sorting process is used to measure the energy saving and operational economy of the combination of sorting process parameters.

[0060] The actual sorting performance data is matched and correlated with the material composition vector and sorting process parameter combination corresponding to the batch to construct a complete training sample data. Among them, the material composition vector is a quantitative representation of the material composition characteristics of the batch, the sorting process parameter combination is the equipment control command output by the process parameter prediction model, and the actual sorting performance data is the real feedback of the set of parameters in actual production. The three are combined to construct a complete mapping relationship between material composition, process parameters and sorting performance, providing a real and reliable data basis for model iterative optimization.

[0061] The generated training sample data is used to retrain or adjust the parameters of the process parameter prediction model. Through multiple rounds of iteration, the internal weight parameters of the model are continuously optimized, enabling the model to continuously learn the correlation between the optimal process parameters and the actual sorting effect under different component characteristics. This corrects the initial prediction bias of the model, gradually improves the accuracy and adaptability of process parameter prediction, and realizes the long-term iterative upgrade of the process parameter prediction model.

[0062] The adaptive sorting method for pretreatment of recycled lithium batteries also includes the following steps: The area percentage of each component in the material composition vector is compared with the corresponding preset threshold. If the area percentage of any component exceeds the corresponding preset threshold, an abnormal warning signal is triggered, and a prompt message containing the category of the abnormal component and the value of the excess is generated.

[0063] Obtain the material composition vector corresponding to the current batch of crushed lithium battery materials. The material composition vector records the area percentage of each component in the material. Compare the area percentage of each component in the material composition vector with its corresponding preset threshold. The preset threshold is a safety upper limit standard set separately for each component category based on historical production data and process control requirements. Different component categories correspond to different threshold limits. Verify whether the area percentage of each component meets the process threshold requirements one by one.

[0064] When the area percentage of any component exceeds its corresponding preset threshold, an abnormal warning signal is immediately triggered, and a prompt message containing the abnormal component category and the exceeding value is generated. The prompt message clearly indicates the exceeding component category, the actual area percentage of the component, and the difference between the actual value and the preset threshold, which makes it easy for staff to quickly locate the source of the abnormality.

[0065] For example, if the preset threshold for the proportion of plastic components is 12%, when the actual proportion of plastic components in a batch of broken consumer battery materials reaches 18%, exceeding the corresponding preset threshold, an abnormal warning signal will be triggered immediately, and a prompt message will be generated simultaneously, clearly indicating that the plastic component exceeds the standard, the actual proportion is 18%, and the excess value is 6%.

[0066] Based on the prompts, the cause of the anomaly is usually determined to be: incomplete separation of the battery casing from the cell, or an inherently high proportion of plastic in the casing of the consumer batteries in this batch. Recommended optimization measures include: increasing the airflow separation machine's operating speed to 1.2 times the original set value, while simultaneously conducting incoming material sampling inspections; staff can intervene promptly based on the prompts to adjust the sorting process parameters or upstream dismantling process parameters to avoid problems such as decreased sorting efficiency, equipment blockage, or substandard recycled product quality caused by excessive component proportions, thus ensuring the overall stable operation of the waste lithium battery recycling pretreatment process.

[0067] An adaptive sorting system for pretreatment of waste lithium batteries includes: an image acquisition module, a vision processing module, a parameter prediction module, and a control adjustment module.

[0068] The image acquisition module is positioned above the conveyor belt for crushed materials and is used to acquire images of the batch of crushed lithium battery materials to be processed.

[0069] Preferably, the image acquisition module includes a linear CMOS camera, a dual-light source composite illumination device, a sealed protective cover, and a trigger synchronization unit; wherein, the linear CMOS camera has a resolution of not less than 0.1mm / pixel and a line frequency of not less than 10kHz; the dual-light source composite illumination device includes a visible light source and a near-infrared light source to enhance the image contrast between metals, plastics, and black powder in the material; the photoelectric sensor monitors the average brightness of the material layer in real time, dynamically adjusts the LED power, compensates for uneven illumination caused by changes in material accumulation thickness and color, and ensures the grayscale consistency of the acquired image; the sealed protective cover is a sealed protective cover with an air curtain blowing function; the trigger synchronization unit establishes a connection with the conveyor belt encoder to realize the synchronous triggering of image acquisition operation and material conveying process.

[0070] The image acquisition process employs active anti-interference measures, enabling it to adapt to changes in industrial lighting and dust pollution conditions.

[0071] The vision processing module is connected to the image acquisition module to receive and process images of the batch of broken lithium battery materials to be processed acquired by the image acquisition module. Specifically, the vision processing module inputs the received images into a pre-trained lithium battery full-component segmentation model for processing, identifies the component categories of the broken materials and obtains the identification results, calculates the area ratio of each component based on the identification results, and then constructs a material component vector characterizing the properties of the batch of broken lithium battery materials to be processed.

[0072] Preferably, the visual processing module includes a GPU-accelerated computing unit, an image preprocessing unit, a semantic segmentation engine, and a feature quantization unit. The GPU-accelerated computing unit runs a full-component segmentation model for lithium batteries, with a single-frame inference time of less than 15ms. The image preprocessing unit performs Retinex illumination correction, median filtering for noise reduction, and histogram equalization image processing operations. The semantic segmentation engine runs an improved UNet network, outputting pixel-level segmentation masks for five components: copper foil, aluminum foil, separator, outer plastic casing, and black powder. The feature quantization unit calculates the pixel area ratio of each component based on the segmentation masks and generates a material composition vector by arranging them in a preset order.

[0073] The parameter prediction module is connected to the vision processing module and is used to receive the material composition vector output by the vision processing module. Specifically, the parameter prediction module inputs the material composition vector into the pre-trained process parameter prediction model. After analyzing the material composition vector as input, the process parameter prediction model outputs a combination of sorting process parameters that matches the composition characteristics of the lithium battery crushed material in the batch to be processed and is used to control the operation of the sorting equipment.

[0074] The parameter prediction module includes a feature extraction layer, a correlation analysis layer, a parameter output layer, and a model storage and update unit. The feature extraction layer performs a deep nonlinear transformation on the material composition vector to extract deep features that reflect the characteristics of the material components. The correlation analysis layer receives material metadata, including battery type, cell capacity, cycle count, and disassembly method. The correlation analysis layer combines the material metadata to perform weighted correction on the deep features, making the feature representation more consistent with actual production conditions. The parameter output layer outputs recommended values ​​for the combination of sorting process parameters, including at least one of the following: wind speed of the air separator, frequency of the eddy current separator, magnetic field strength of the magnetic separator, frequency of the vibrating screen, screen aperture, position of the material distribution baffle, and conveyor belt speed. The model storage and update unit is used to store pre-trained model parameters and supports incremental model learning and online updates.

[0075] The control adjustment module is connected to the parameter prediction module and the control system of the sorting equipment, respectively, and is used to receive the sorting process parameter combination output by the parameter prediction module. Specifically, the control adjustment module sends the sorting process parameter combination to the control system of the sorting equipment. The control system dynamically adjusts the operating parameters of the sorting equipment according to the sorting process parameter combination to perform adaptive sorting of the batch of crushed lithium battery materials to be processed.

[0076] Preferably, the control and adjustment module includes a communication interface unit, a parameter parsing and verification unit, an instruction issuing unit, and a status monitoring unit; wherein, the communication interface unit supports at least one industrial communication protocol among OPC UA, Modbus TCP, and PROFINET; the parameter parsing and verification unit performs integrity checks and rationality verification on the received combination of sorting process parameters; the instruction issuing unit converts the verified parameters into register values ​​recognizable by the PLC / DCS and issues them for execution; the status monitoring unit monitors the operating status and parameter execution of the sorting equipment in real time and supports equipment abnormality alarm functions.

[0077] The modules mentioned above in this system can be independent physical devices or integrated into the same computing unit. One approach is centralized deployment: the vision processing module, parameter prediction module, and control adjustment module are deployed uniformly in the same industrial computer or edge computing gateway, suitable for small to medium-sized waste lithium battery recycling production lines. The other approach is distributed deployment: the vision processing module is deployed at an edge computing node close to the camera, the parameter prediction module is deployed in the cloud or on a central server, and the control adjustment module is deployed at the field PLC, suitable for large-scale or multi-plant recycling production lines.

Claims

1. An adaptive sorting method for pretreatment of waste lithium batteries, characterized in that, Includes the following steps: Acquire images of the current batch of crushed lithium battery material awaiting processing; The image is input into the lithium battery full-component segmentation model, which outputs a material composition vector representing the crushed lithium battery material of the batch to be processed; the material composition vector includes the category and area percentage of each component in the material. The material composition vector is input into a pre-trained process parameter prediction model, which outputs a combination of sorting process parameters. The control system dynamically adjusts the operating parameters of the sorting equipment according to the combination of sorting process parameters to sort the batch of crushed lithium battery materials to be processed.

2. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 1, characterized in that, The lithium battery full-component segmentation model is an improved UNet model, including an encoder, decoder, and detail-enhancing convolutional module; The encoder is used to perform multi-level feature extraction on the input image and output a multi-scale feature map; The decoder is used to upsample the multi-scale feature map step by step, and during the sampling process, it fuses the shallow features from the corresponding level of the encoder by feature splicing to output a pixel-level image segmentation mask. The detail enhancement convolution module is located on the feature concatenation path between the encoder and the decoder. It includes at least one convolution operation to enhance edge features at the junction of different regions in the multi-scale feature map before feature fusion.

3. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 2, characterized in that, The encoder is built on a backbone network and integrates an attention mechanism module after at least one feature extraction level. The attention mechanism module is a convolutional attention module, which includes a channel attention submodule and a spatial attention submodule arranged sequentially. The channel attention submodule weights the channel dimensions of the multi-scale feature map through global pooling and multilayer perceptron; the spatial attention submodule performs pooling and convolution operations along the channel dimensions of the multi-scale feature map to weight the spatial position of the multi-scale feature map.

4. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 3, characterized in that, An edge enhancement module is provided at the output of the attention mechanism module; The edge enhancement module consists of a 3×3 learnable convolutional layer and residual connections. The edge enhancement module is used to learn autonomously and enhance the edge features at the junction of different regions.

5. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 2, characterized in that, The improved UNet model also includes a feature fusion layer; The feature fusion layer is positioned after the encoder outputs the multi-scale feature map and before the decoder performs upsampling. The feature fusion layer can be either a feature pyramid network structure or a high-resolution network fusion strategy. When the feature fusion layer adopts a feature pyramid network structure, it performs top-down feature fusion, which merges deep high-semantic features with shallow high-resolution features step by step to obtain multi-scale fused features. When the feature fusion layer adopts a high-resolution network fusion strategy, the multi-scale feature maps of each layer are upsampled to the highest resolution, and then feature fusion is completed by element-wise addition to obtain a fused multi-scale feature map that maintains high resolution.

6. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 1, characterized in that, The process parameter prediction model includes a feature extraction layer, a correlation analysis layer, and a parameter output layer connected in sequence. The feature extraction layer is used to perform nonlinear transformation on the input material composition vector to extract deep features; The correlation analysis layer has a built-in physical characteristic mapping rule library, which is used to map material composition characteristics to the adjustable physical parameter range of the sorting equipment, and to dynamically weight and correct the fixed parameters of the sorting process in combination with material metadata. The parameter output layer is used to output the combination of sorting process parameters.

7. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 1, characterized in that, The process parameter prediction model includes a material metadata association correction logic, which can dynamically weight and correct the material composition vector or initial process parameters based on the material metadata.

8. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 1, characterized in that, The training process of the process parameter prediction model includes: Collect historical production data, which includes at least the material composition vector of historical batches of crushed lithium battery materials, the actual combination of sorting process parameters used, and the corresponding sorting effect data. Using material composition vectors as input features and sorting process parameter combinations as training labels, the model is validated by combining sorting effect data, and a process parameter prediction model is constructed and trained using the historical production data.

9. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 1, characterized in that, It also includes the following steps: Collect actual sorting effect data of the batch of crushed lithium battery materials to be processed after sorting; The actual sorting effect data, material composition vector, and sorting process parameters are combined to construct training sample data. The process parameter prediction model is retrained or its parameters are adjusted using the training sample data to complete the continuous iterative optimization of the model.

10. The adaptive sorting method for pretreatment of waste lithium batteries according to claim 1, characterized in that, It also includes the following steps: The area percentage of each component in the material composition vector is compared with the corresponding preset threshold. If the area ratio of any component exceeds the corresponding preset threshold, an abnormal warning signal will be triggered, and a prompt message containing the category of the abnormal component and the value exceeding the standard will be generated.

11. An adaptive sorting system for pretreatment of waste lithium batteries, used to implement the method described in any one of claims 1-10, characterized in that, include: Image acquisition module, vision processing module, parameter prediction module, and control adjustment module; The image acquisition module is used to acquire images of the batch of broken lithium battery materials to be processed. The visual processing module is used to input the image into the lithium battery full-component segmentation model for processing, identify the component categories of the crushed material and obtain the identification results, calculate the area ratio of each component based on the identification results, and construct a material component vector characterizing the characteristics of the crushed lithium battery material in the batch to be processed. The parameter prediction module is used to input the material composition vector into the pre-trained process parameter prediction model. The process parameter prediction model analyzes the material composition vector as input and outputs a combination of sorting process parameters that matches the characteristics of the lithium battery crushed material in the batch to be processed and is used to control the operation of the sorting equipment. The control adjustment module is used to send the sorting process parameter combination to the control system of the sorting equipment, so that the control system can dynamically adjust the operating parameters of the sorting equipment according to the sorting process parameter combination and perform sorting on the batch of lithium battery crushed materials to be processed.

Images