Method and device for intelligently sorting positive electrode materials of retired lithium ion batteries based on spectrum method

Abstract

The invention discloses an intelligent sorting method and device for a positive electrode material of a retired lithium ion battery based on a spectrum method, and the device is connected in series with a pretreatment module, a cleaning module, an identification module and a sorting module through a conveying belt in a supporting module to form an automatic sorting assembly line. Pre-treating the retired battery to obtain a clean positive plate; an X-ray fluorescence spectrophotometer in the recognition module collects the spectrum of the positive plate and transmits the spectrum to an industrial personal computer, and the spectrum is preprocessed and then input into a classifier library; the classifier library comprises a main classifier and an increment classifier based on element and spectral shape dual channels, if the output confidence reaches a preset threshold value, the material category is judged, otherwise, the spectrum is temporarily stored as an unknown spectrum; after unknown spectrums are accumulated to a set number, new categories are identified through a clustering algorithm, and a lightweight dichotomy increment classifier is trained and added into a classifier library; and the sorting module drives the hydraulic pushing hand to sort the positive plates according to the identification result. Through cooperative evolution of the main classifier and the increment classifier, the sorting precision is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial waste treatment and resource utilization in the field of environmental protection technology, in particular to an intelligent sorting method and device for retired lithium ion battery positive electrode materials based on spectroscopy, which is an environmentally friendly and efficient method and device for comprehensively recovering heavy metals in retired lithium ion battery positive electrode materials. BACKGROUND

[0002] Energy and environmental problems have been the main concern worldwide, and with the gradual implementation of new energy industry plans, electric vehicles have developed to a certain extent. The vigorous development of the electric vehicle industry has also promoted the further development of the power battery industry. Electric vehicles provide a certain convenience for human travel, but the problems caused by the retirement of power batteries cannot be ignored.

[0003] With the increasing number of retired lithium ion batteries, from the perspective of environmental protection, lithium batteries contain heavy metals and harmful substances, and if they are not treated in time, they will cause significant harm to the environment; from the perspective of economic value, lithium ion batteries contain valuable precious metals with high market demand.

[0004] The current recycling process of retired lithium ion batteries mainly includes pretreatment, secondary treatment and deep treatment. After discharging, the battery can be obtained by crushing and physical sorting. Coarse positive and negative electrode powders (lithium-containing positive electrode materials and graphite), separators, aluminum foil, copper foil, metal shell, etc. However, the purity of the obtained positive electrode powder is low, containing 5-10% of impurity powder, which needs to be removed by chemical method. The pretreatment process of coarse crushing-sorting-impurity removal has the problems of long process, consumption of chemical reagents, sewage treatment, etc., and the recovery cost is relatively high. For different types of lithium battery positive electrode materials, such as lithium iron phosphate (LiFePO4, LFP), lithium cobaltate (LiCoO2, LCO), lithium manganate (LiMn2O4, LMO), and nickel-cobalt-manganese ternary material (LiNi x Co y Mn 1-x-y O2, NCM, including NCM523, NCM622 and NCM811 with Ni∶Co∶Mn ratio of 5∶2∶3, 6∶2∶2 and 8∶1∶1), the classification and sorting are only manually classified according to the battery nameplate and composition table. Therefore, there are prominent problems of backward classification and sorting technology at the pretreatment end of the retired battery.

[0005] Based on the above background, an integrated intelligent precise sorting and recycling equipment is designed to realize precise sorting of positive electrode materials and improve the purity and value of the recovered products. SUMMARY

[0006] The application provides a spectrum-based retired lithium-ion battery positive electrode material intelligent sorting method and device, which aims to improve the accurate classification of positive electrode materials in retired lithium batteries.

[0007] To achieve the above-mentioned purpose, the specific technical scheme of the application is:

[0008] A spectrum-based retired lithium-ion battery positive electrode material intelligent sorting method, the method comprising the following specific steps:

[0009] First, XRF spectrum collection is performed, and the spectra of various positive electrode materials are obtained in two ways of software simulation and experimental detection to construct an XRF spectrum database covering LiFePO4, LiCoO2, LiMn2O4 and LiNi x Co y Mn 1-x-y O2 four material types as the data basis for training the main classifier model; at the same time, a temporary database integrating a clustering algorithm and capable of caching multiple sample data is constructed as the data source for training the incremental classifier model; all spectra share the same energy axis, and the one-to-one corresponding intensity values are stored one by one, and the intensity values differ due to different samples;

[0010] Subsequently, a classifier library is constructed and stored on the industrial computer, which is composed of a main classifier based on a dynamic double-channel neural network model and an incremental classifier based on an incremental learning model trained from the XRF spectrum database; the construction of the main classifier is specifically as follows: the data in the XRF spectrum database is preprocessed, and then the sample size is expanded through data enhancement technology to alleviate the small sample problem; element features and spectral shape features are extracted in parallel for all enhanced spectrum data; the two types of features are input into the dynamic double-channel neural network model, and finally the probability distribution of the positive electrode material category is obtained through the classification output layer; the network parameters are optimized through the back propagation algorithm to train the model to maximize the classification accuracy of the training samples; the dynamic double-channel neural network model obtained after training is registered as the main classifier in the classifier library of the industrial computer;

[0011] Then when the to-be-tested positive electrode material passes through the industrial computer, the existing main classifier in the classifier library identifies the positive electrode material. In this process, the main classifier in the classifier library first performs prediction and result interpretation on the positive electrode material, and then determines two results of high confidence samples and low confidence samples according to the confidence threshold. If it is a low confidence sample, the sample is classified as an unknown type when output and the result is interpreted, and the XRF spectrum data of the sample is cached to a temporary database; the clustering algorithm integrated in the temporary database induces samples that may belong to the same unknown category into a new category cluster. When the number of sample data of the new category cluster accumulates to 100, the training incremental learning model process is automatically triggered, and the incremental learning model obtained after merging is registered as an incremental classifier in the classifier library; when the training incremental learning model process is triggered, the total amount of sample data cached in the temporary database will decrease, and when the number of sample data of the new category cluster in the cache reaches 100 again, the training incremental learning model process is automatically triggered. In this way, the classifier library is continuously expanded and updated.

[0012] Finally, the to-be-tested positive electrode material after material pretreatment and cleaning is identified by the classifier library in the industrial computer, and the material type is output to obtain the sorting result.

[0013] Further, the data in the XRF spectrum database is preprocessed, specifically:

[0014] Each original spectrum data is executed in turn:

[0015] Energy axis calibration, using standard sample characteristic peaks to perform nonlinear calibration on energy channels to control the calibration error within the set range;

[0016] Background subtraction, using an iterative background estimation algorithm to perform multiple iterations in an adaptive width iteration window;

[0017] Noise suppression, using wavelet transform to perform multi-scale decomposition and threshold reconstruction on the spectrum after background subtraction;

[0018] Normalization, using the Compton scattering peak integral intensity as an internal standard to normalize the overall amplitude of the spectrum to a unified dimension;

[0019] The spectrum data after executing data preprocessing is re-stored.

[0020] Further, the data enhancement technique is a random disturbance operation performed on the XRF spectrum after data preprocessing, specifically:

[0021] Introducing noise consistent with the photon counting statistics to simulate signal fluctuations; applying a small random shift to the energy axis and re-interpolating the samples; applying different degrees of Gaussian blur to change the spectral resolution; superimposing a random low-order polynomial background to simulate the difference in matrix effects; randomly adjusting the intensity of the spectral peak interval corresponding to the main characteristic element within a range of ±15%; then fine-tuning the normalization to ensure that each sample is finally consistent in scale; thereby generating spectral samples that are consistent in physical mechanism and significantly enhanced in diversity, for expanding the XRF spectral database.

[0022] Further, the element feature and the spectral shape feature of all the enhanced spectrum data are extracted in parallel, specifically:

[0023] The element feature extraction is completed once for all the energy spectrum:

[0024] After inputting the energy axis and intensity axis data, first, the main peak energy position is determined according to the maximum intensity, and it is detected whether there is a significant feature peak exceeding 20-40% of the main peak intensity in each pre-defined element feature energy window, according to which the spectrum is classified into one of the pre-defined positive material types; then, according to the determined material type, the corresponding element feature energy interval is called, the peak area in each interval is calculated by using the numerical integration method for the intensity-energy curve, and the sum of all element peak areas is taken as the total effective area; if the total effective area exceeds the trace threshold, the element peak area is divided by the total effective area to obtain the normalized ratio, otherwise the ratio is set to zero; finally, the fixed-dimension ratio vector is arranged in the order of the pre-defined standard elements as the element ratio feature output;

[0025] Spectral shape feature extraction: the complete intensity sequence after data preprocessing or data enhancement is resampled to a fixed dimension as the feature output.

[0026] Further, the dynamic dual-channel neural network model is constructed as follows:

[0027] The dynamic dual-channel deep learning model is used for processing, which includes a spectral shape data input channel and a dynamic ratio vector input channel; the spectral shape data input channel uses a one-dimensional convolutional neural network containing multiple residual block structures to extract features, and includes an attention mechanism realized by attention weight generation and feature weighting; the dynamic ratio vector input channel performs feature transformation through a fully connected layer; the features of the two channels are fused through a dynamic weighting mechanism, which combines the spectral shape feature and the material type information to generate an adaptive weight and weight the spectral shape feature; the weighted spectral shape feature and the processed ratio feature are fused, and the fused feature is processed through a fully connected layer and a random inactivation layer, and finally the result is output through a classification output layer.

[0028] Further, the dynamic dual-channel neural network model is trained as follows:

[0029] The extracted element features and spectral shape features are constructed into a data set and divided into a training set and a validation set according to a predetermined proportion; the connection weights of the dynamic double-channel neural network model are initialized; callback functions used in the training process are configured, including an early stopping strategy and a dynamic learning rate decay strategy; the model is trained through multiple rounds of iteration using the training set data, and the model performance is evaluated using the validation set data after each round of iteration; when the validation set performance no longer improves in continuous multiple rounds of iteration, the training process is automatically terminated and the model weight with the optimal validation performance is restored, and finally the trained model is obtained as the main classifier.

[0030] Further, the main classifier in the classifier library first predicts and interprets the positive electrode material, specifically:

[0031] When the to-be-tested positive electrode material passes through the industrial computer, the main classifier in the classifier library performs data preprocessing on the XRF spectrum data of the positive electrode material collected by the X-ray fluorescence spectrometer carried in the industrial computer. The data preprocessing is consistent with the process of preprocessing the spectrum data, and the standardized element feature vector and the spectral shape feature vector are extracted. The two kinds of feature vectors are simultaneously input into the trained dynamic double-channel neural network model for forward propagation calculation to obtain the probability distribution belonging to each material category. If there is a result reaching or exceeding the predetermined confidence threshold in the output of each material confidence, i.e., the four types of materials covered in the XRF spectrum database, it is determined that this type of material is a high-confidence sample, and the corresponding positive electrode material type and result interpretation are output. If the output positive electrode material confidence is lower than the predetermined confidence threshold, for example: LiNi x Co y Al 1-x-y O2 material, i.e., not within the range of the four types of materials covered in the XRF spectrum database, it is determined that this type of material is a low-confidence sample, which is classified as an unknown type and interpreted in the output, and the XRF spectrum data of the sample is cached to a temporary database. All output result interpretations have the following contents: extracting the output features of the intermediate layer in the model, including spectral feature compression representation and attention weight distribution, generating an interpretable result report containing predicted material type, element ratio information, probability of each category, and model decision basis.

[0032] Further, the training incremental learning model process is automatically triggered, and the obtained incremental learning model after merging is registered as an incremental classifier in the classifier library, specifically:

[0033] Whenever the number of sample data of a new category cluster in the temporary database accumulates to 100, 100 sample data of the new category cluster are automatically extracted from the temporary database, and the total amount of sample data in the temporary database is reduced; then for each new category cluster that has passed the validity verification, a lightweight binary classification convolutional neural network model is constructed; the model is initialized using the knowledge distillation technology, and the existing knowledge is maintained by reusing and freezing the weights of part of the feature extraction layers of the main model; then the model is trained using positive samples from the new category cluster and an equal amount of negative samples extracted from known categories, so that it learns to distinguish the new category from all existing categories; from the beginning of construction to the completion of training, it is a cycle, and the binary classification model generated each time is merged with the binary classification model generated last time, and the merged binary classification model is called an incremental learning model, which is used as an incremental classifier through iteration and registered in the classifier library to achieve the expansion and update of the classifier library; wherein the positive sample is a low-confidence sample that cannot be recognized by the main classifier, and the sample data of the new category cluster obtained by the clustering algorithm meets the sample quantity of 100; the negative sample is a high-confidence sample that has been successfully recognized by the main classifier, and the quantity is consistent with the positive sample.

[0034] An intelligent sorting device for retired lithium-ion battery positive electrode materials based on the above method, the device comprises a material pretreatment module, a cleaning module, an identification module, a sorting module and a support module.

[0035] The support module comprises a rack structure and a conveying belt assembly, which together constitute an integrated material conveying platform; the conveying belt passes through the material pretreatment module, the cleaning module, the identification module and the sorting module along the conveying path in turn, for realizing automatic transfer of waste lithium-ion batteries between the functional modules; the rack structure is composed of a platform support, a top support and a bottom support, the platform support is used for bearing the conveying belt and fixing the functional modules, and the top and bottom supports provide overall stability and structural support to ensure mechanical rigidity and reliability of the equipment during operation; the support module realizes orderly layout and continuous operation support of the equipment structure, and is an important foundation for efficient operation of the system;

[0036] The material pretreatment module is used for mechanical pretreatment of waste lithium-ion batteries, mainly comprising a cutter assembly and a clamp assembly; the cutter assembly is composed of a set of fixed knives and a set of movable knives, which can effectively cut the battery shell and internal structure through relative motion; the clamp assembly is provided with two rotatable clamping mechanisms, which can multi-axis position and clamp the battery sample, thereby realizing omnidirectional cutting in different directions; the module realizes efficient, stable and accurate disintegration of the battery sample by cooperatively controlling the relative motion of the cutter and the clamp, and provides convenient conditions for subsequent material sorting and recycling;

[0037] The cleaning module comprises a cleaning box and a spring vibration screen; the cleaning box is internally provided with a material channel, and adjustable movable baffles are arranged at the inlet and outlet of the material channel for controlling the flow rate of the material and adjusting the flow rate of the cleaning liquid; the spring vibration screen is installed below the cleaning box and is used for realizing water-solid separation and impurity screening while the material is subjected to spray cleaning; liquid spray nozzles are arranged above the cleaning box and are used for uniformly and highly pressurizing and spraying the passing material to improve the cleaning effect; the vibration screen realizes high-frequency vibration by means of a spiral spring support structure, and the cleaning liquid is used to carry away the surface-attached impurities and particles, so that the material is ensured to be clean; the cleaning module has compact structure and high cleaning efficiency, and is suitable for efficient pretreatment of waste lithium ion battery materials.

[0038] The identification module comprises an X-ray fluorescence spectrometer, an industrial computer and a gantry structure, the X-ray fluorescence spectrometer and the industrial computer are connected through a data communication interface and are jointly and fixedly installed on the gantry structure; the X-ray fluorescence analyzer is arranged at the inlet position of the sorting module and is used for analyzing the element composition of the cathode material entering the sample; the industrial computer is internally integrated with a pre-trained classifier library, can receive the spectral data collected by the XRF in real time, automatically identify different types of cathode materials, and transmit the identification results to the subsequent sorting module to guide the material diversion; the identification module has the characteristics of compact structure, accurate identification and high processing efficiency, and provides key technical support for realizing intelligent sorting of waste lithium ion battery materials with high precision.

[0039] The sorting module comprises a plurality of hydraulic drive push rods, each push rod is driven by a hydraulic cylinder and is used for sorting the identified different types of cathode materials from the conveying platform to the corresponding recycling path; the plurality of hydraulic drive push rods are arranged at equal intervals along one side of the conveying platform, can push the target material in a direction according to the sorting instruction output by the upper identification module, and realize rapid and accurate diversion operation; each hydraulic cylinder is controlled by a control valve group. The sorting module has the characteristics of simple structure, rapid response and accurate control, and is a key execution module for realizing automatic sorting operation.

[0040] Compared with the problems described in the background art, the present application first discharges and disassembles the retired lithium ion battery to obtain a lithium battery positive electrode material with a current collector as a carrier. As can be imagined, there are many kinds of positive electrode materials, and the common recycling device on the market is to crush the lithium ion battery for recycling, which has certain safety risks and cannot distinguish the types of recycled materials, so that all the different positive electrode materials obtained by crushing can only obtain rare metals in the same way, which greatly weakens the recycling effect of different types of rare metals, so an intelligent sorting device that can accurately identify different positive electrode materials is urgently needed. Therefore, further, the XRF spectrum data double-channel neural network model and the incremental learning model of the positive electrode material are developed and deployed to the computer, and the positive electrode material is accurately sorted by controlling the force of the mechanical arm. When constructing the model, first, obtain the accurate element composition and density of the common positive electrode material from ICSD, use DTSA-II software to obtain the simulated XRF spectrum, and collect the measured XRF in the laboratory or production line, covering different conditions, to obtain two kinds of original XRF spectrum, and form a feature lithium ion battery material database. The spectrum is pretreated, including energy calibration, background subtraction, noise suppression, intensity normalization, feature peak enhancement and other steps, to remove noise and interference signals, so that the spectrum data is standardized and neat, and then data enhancement is performed to generate spectrum samples that are consistent in physical mechanism and have significantly enhanced diversity, which are used to expand the training database. Then extract the features to obtain the element feature branch and the spectral feature branch. Create a double-channel neural network classification model, including the input layer of the element feature branch and the spectral feature branch, each convolutional layer, etc., and finally output the material category probability distribution through the feature fusion layer. High confidence directly outputs material type information, and low confidence is processed accordingly, such as marking "unknown", storing feature data, triggering incremental learning, etc. The application also provides a corresponding sorting device, which can cut the sample in all directions through the pretreatment module, clean the positive electrode sheet through the cleaning module, identify the positive electrode material through the identification module, and sort the positive electrode material through the sorting module according to the identification result. The support module realizes the connection and support of each module, and the intelligent sorting of the positive electrode material is completed by the cooperation of each part. Therefore, the recycling method and device of the retired lithium ion battery proposed in the present application mainly aims to improve the accurate classification of the positive electrode material of the retired lithium ion battery.

[0041] Advantages of the present application:

[0042] 1. Intelligent identification and sorting, accurate and reliable classification

[0043] Based on the spectrum method combined with intelligent identification technology, the present application can realize high-precision identification and automatic sorting of different types of retired lithium battery positive electrode materials, avoid errors caused by manual or inefficient mechanical methods, and still maintain stable performance in complex scenes, ensuring the accuracy and consistency of material classification.

[0044] 2. Non-contact detection, material integrity fidelity

[0045] Spectrometry is a non-contact, non-destructive, non-destructive testing technology that does not cause physical damage or chemical contamination to the positive electrode material during the sorting process, which is conducive to maintaining high value in subsequent smelting, step utilization or regeneration processes, avoiding material quality degradation caused by the detection process, and improving overall recycling efficiency.

[0046] 3. Strong general adaptability, large-scale application

[0047] The present application can adapt to different models and brands of lithium battery positive electrode materials at different retirement stages, and is not affected by factors such as surface contamination and oxidation degree, has strong general adaptability, and the device operating parameters can be flexibly adjusted to meet the demand of large-scale industrial application, and is suitable for deployment in retired lithium battery recycling plants to realize continuous and large-scale operation.

[0048] 4. Deep learning driven, autonomous iterative optimization

[0049] The fusion of deep learning algorithm model has the ability of autonomous learning and continuous optimization. In the initial state, based on large-scale sample training, it has super strong recognition accuracy, and in the processing process, it can also constantly learn and accumulate experience from the samples to be tested, identify positive electrode materials that have not been seen before or have complex composition, thereby greatly improving the adaptability and accuracy of the sorting system to unknown or mixed materials. This makes it more intelligent and forward-looking, and still maintains efficient and stable sorting performance when facing material updates and iterations.

[0050] 5. Module integration, stable and efficient device

[0051] The device is composed of five modules of material pretreatment, cleaning, identification, sorting and support, forming a complete flow processing chain, which can realize the whole process integration from raw material entering to finished product classification output, greatly improving the engineering practicability and popularization value of the equipment; and the modular integrated design scheme can ensure that the system structure is compact, easy to operate, and the device runs stably, is easy to maintain, reduces external dependence, and can realize long-term continuous operation.

[0052] 6. Realize green environmental protection, promote technological progress

[0053] Through accurate identification technology, accurate sorting of different types of positive electrode materials can be completed. Then it is convenient to introduce a directed regeneration process, which not only improves the metal recovery efficiency and reduces the secondary pollution of harmful substances, but also reduces the waste of resources caused by mixed treatment. This technological innovation effectively promotes the transformation of the lithium battery recycling industry to automation and intelligence, marking the lithium battery recycling technology towards green and low-carbon, intelligent and efficient, and meeting the global trend of sustainable development. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flow chart of the intelligent sorting method of the retired lithium ion battery positive electrode material is provided for embodiment 1 of the application.

[0055] Figure 2 The structural schematic diagram of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application.

[0056] Figure 3 The schematic diagram of the clamp part of the material pretreatment module of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application.

[0057] Figures 4-5 The schematic diagram of the cutting part of the material pretreatment module of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application.

[0058] Figure 6 The schematic diagram of the cleaning tank of the cleaning module of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application.

[0059] Figure 7 The schematic diagram of the air nozzle of the cleaning module of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application.

[0060] Figure 8 The schematic diagram of the sorting module of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application.

[0061] Figure 9 The schematic diagram of the collection tank of the sorting module of the intelligent sorting device of the retired lithium ion battery positive electrode material is provided for embodiment 2 of the application. DETAILED DESCRIPTION

[0062] The application will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0063] The application includes:

[0064] An intelligent sorting method of a retired lithium ion battery positive electrode material based on a spectral method, a main classifier based on a dynamic double-channel neural network model and an incremental classifier based on an incremental learning model for developing and identifying XRF spectral data of the positive electrode material, which are deployed to an industrial computer to realize accurate sorting of the positive electrode material by controlling the strength of a mechanical arm. The construction of the classifier in the industrial computer includes a series of steps.

[0065] The construction of the main classifier is firstly the acquisition of the XRF spectrum of common positive electrode materials. In this step, the simulation data: the accurate element composition (chemical formula) and density of common positive electrode materials are obtained from ICSD, and the XRF simulation software (DTSA-II) is used to input these information, so that the simulated XRF spectrum can be obtained conveniently. The experimental data: the XRF measured in the laboratory or production line is collected, covering different states of charge (SOC), cycle times and synthesis batches. Two kinds of original XRF spectra are obtained, and the horizontal and vertical coordinates are energy and relative intensity. All the spectra share the same energy axis, and the corresponding intensity values are stored one by one. The intensity values are different due to different samples, and the two kinds of data are combined into a feature lithium ion battery material database. The data includes but is not limited to lithium iron phosphate (LiFePO4, LFP), lithium cobaltate (LiCoO2, LCO), lithium manganate (LiMn2O4, LMO), nickel cobalt manganese ternary material (LiNi x Co y Mn 1-x-yXRF spectrum data of four material types of O2, NCM, including NCM523, NCM622 and NCM811 with Ni:Co:Mn ratio of 5:2:3, 6:2:2 and 8:1:1, and LiMnO2. The obtained spectrum is preprocessed, and the spectrum data preprocessing includes: 1) energy calibration, the energy axis is calibrated by a standard sample characteristic peak (such as Fe Kα = 6.40 keV), the actual peak position is searched near the theoretical peak position, the energy offset is calculated, the offset is applied to correct the entire energy axis, the influence of device drift is eliminated, the comparability of data of different devices and different batches is ensured, and the influence of device drift is eliminated; 2) background deduction, SNIP (Smoothing by Nonlinear Iterative Peak-clipping) algorithm is used, Compton continuous spectrum is estimated by multi-stage iteration (typically 8 times), and the window is gradually optimized from wide window (15 points) to narrow window (3 points), so that the characteristic peak and background noise are effectively separated; 3) noise suppression, wavelet transform is used for noise reduction (sym8 wavelet basis), 3-level wavelet decomposition and reconstruction, adaptive threshold processing based on noise statistical characteristics, retaining characteristic peaks while suppressing high-frequency noise; 4) intensity normalization, taking the Compton scattering peak (9-10 keV region) as the reference, calculating the maximum intensity value of the Compton peak, dividing the entire spectrum intensity by the reference value, and eliminating the influence of sample quantity difference; in the present application, in order to improve the robustness of the model to instrument drift, statistical fluctuations and matrix differences, integrated data enhancement is performed on the preprocessed XRF spectrum data: first, under the premise of keeping the overall distribution unchanged, random count noise is applied to each energy channel according to Poisson statistics to simulate low count scenarios; then the energy axis is translated by a random offset not exceeding ±0.05 keV and is remapped to the calibration grid by interpolation to train the energy drift tolerance; then the spectrum is convolved with a Gaussian kernel function with a variable width, and the Gaussian standard deviation is randomly changed within the range of 0.8-1.2 pixels to simulate the peak broadening caused by different spectrometer resolutions; then the background generated by a random second-order polynomial is superimposed, and the coefficients are selected from the interval [-0.001, 0.001] to reproduce the matrix and scattering background differences; then, for Co-Kα (6.93 keV), Ni-Kα (7.47 keV) and Mn-Kα (5.89 keV) characteristic peaks, the peak intensity is fine-tuned within ±15% to make the model learn the small fluctuations of element content; finally, the enhanced spectrum data is fine-tuned and normalized (min-max or z-score) to unify the input feature scale and prevent network gradient explosion. Next, the data enhancement data is extracted to obtain element feature branches and spectrum shape feature branches, and the feature acquisition process of the element feature branch is as follows: first, the energy axis array and the intensity axis array of the energy spectrum are input. By finding the maximum value in the intensity array, the energy value corresponding to the maximum value is determined as the main peak energy position of the energy spectrum.Subsequently, a series of predefined elemental characteristic energy windows (e.g. nickel elemental characteristic peak window is 7.35 keV to 7.60 keV) are detected to see if there is a significant characteristic peak with intensity exceeding 30% of the main peak intensity. Based on the main peak position and the presence of characteristic peaks, according to the preset decision logic (e.g. the main peak is located at 7.3 - 7.6 keV and there is an aluminum characteristic peak, it is judged as NCA material), the spectrum is automatically classified into a predefined positive electrode material type (such as NCM, LFP, LCO, etc.). After determining the material type, the corresponding elemental characteristic energy interval of the type is called. Preferably, the trapezoidal rule in numerical integration is used to calculate the peak area of each element in its specific energy interval. The peak areas of all elements are summed to obtain the total effective area. In a preferred embodiment, a trace threshold (such as 10. -6), if the total effective area is greater than the threshold, the element peak area is divided by the total effective area to obtain the normalized content ratio of the element; if the total effective area is too small, all the ratios are set to zero to avoid unstable values. Finally, the obtained element ratio is arranged in a standard order, for example: nickel (Ni), cobalt (Co), manganese (Mn), iron (Fe), aluminum (Al), phosphorus (P), oxygen (O), to generate a fixed dimension (for example, 7 dimensions) ratio vector as the element ratio feature output of the spectrum, which is used for subsequent model training or classification identification. The feature acquisition process of the spectral shape feature branch: the energy axis and intensity axis data are resampled, and the data sequence is converted into a standard sequence of fixed length (such as 1000 dimensions) by an interpolation method, so as to ensure that all input spectrum data have a unified dimension and meet the requirements of subsequent model processing. Next, a double-channel neural network model is constructed, and the method comprises the following steps: a deep learning model architecture with double input channels is constructed. The first input channel receives the standard length spectral shape intensity data sequence obtained by the spectral shape feature branch; the second input channel receives the standardized element ratio feature vector obtained by the element feature branch. The spectral shape data input channel adopts a deep neural network structure based on one-dimensional convolution for feature extraction. The network contains an initial feature extraction layer, followed by a plurality of residual block structures cascaded in turn. The use of residual blocks effectively alleviates the gradient vanishing problem in deep network training and improves the feature extraction capability. The network backend also integrates an attention mechanism, which generates attention weights corresponding to the spatial dimension of the feature map, adaptively weights the extracted spectral shape features, and thus strengthens the key feature area and suppresses irrelevant information. The element ratio feature vector input channel first performs nonlinear transformation and embedding representation on the input element ratio feature through a fully connected layer to generate a low-dimensional embedding vector containing material type information. The double-channel feature fusion adopts a dynamic weighting mechanism. This mechanism splices the high-level features extracted by the spectral shape channel and the material type embedding vector generated by the ratio channel, and calculates a set of adaptive weight coefficients through a fully connected layer and an activation function; the set of weight coefficients is multiplied with the spectral shape features element by element to realize the dynamic modulation and enhancement of the spectral shape features according to the material type. The modulated spectral shape features and the separately processed ratio features are spliced again to form the fused joint feature representation. Finally, the fused joint feature is nonlinearly transformed through one or more fully connected layers and processed by a random inactivation layer, and the final material category probability distribution is generated by the Softmax classification output layer to complete the model architecture design. After that, the dynamic double-channel neural network model is trained through the following steps: first, divide the entire dataset after completing the feature extraction. Preferably, a random division strategy is adopted, and about 20% of the total number of samples are reserved as a validation set, and the remaining samples are used as a training set to ensure the objectivity of model performance evaluation.Before model training, all weight parameters of the network are initialized, and a callback monitoring strategy for the training process is configured. Specifically, the callback strategy mainly includes: early stopping strategy: monitor the performance indicator on the validation set, if the indicator does not improve for a plurality of consecutive iterations (for example, 15 rounds), the training is automatically terminated to prevent overfitting; dynamic learning rate decay strategy: monitor the performance of the validation set, if the performance improvement is stuck in a plateau period (for example, 5 consecutive rounds without improvement), the current learning rate is decayed by a certain percentage (for example, halved) to promote the model to converge to a better performance extremum. During training, use the mini-batch gradient descent algorithm for multiple iterations. In an embodiment, the batch size is set to 32, and the maximum number of iterations is set to 100 rounds. In each iteration, the loss function gradient is calculated using the training set data and the network weights are updated. After the update is completed, the current performance of the model is immediately evaluated and monitored using the validation set data. The training process is eventually automatically terminated by the early stopping strategy, and automatically rolls back to the model weight snapshot with the best validation set performance, thereby obtaining a model with the best generalization ability for subsequent positive electrode material identification and sorting tasks. The main classifier based on the dynamic dual-channel neural network model is registered in the industrial computer.

[0066] Next, a batch of various types of retired batteries are sorted, and an incremental classifier based on an incremental learning model is constructed. First, each retired battery enters the pretreatment and cleaning module to obtain clean positive electrode materials, and then the positive electrode materials pass through the X-ray fluorescence spectrometer in the sorting module device to collect positive electrode material spectral data and transmit them to the industrial computer.

[0067] The industrial computer receives the XRF spectral data of the positive electrode materials of the retired batteries, and uses the same pretreatment, element ratio feature extraction, and spectral shape feature extraction process as in the main classifier stage to generate fixed-length spectral shape intensity feature vectors and parallelly calculate element composition content ratio feature vectors. Subsequently, the processed spectral shape features and ratio features are input into the dynamic dual-channel neural network model of the trained main classifier as dual inputs. The model calculates the confidence distribution of each pre-defined positive electrode material type by forward propagation. If there is a result that reaches or exceeds a predetermined confidence threshold (such as ≥0.85) in the output of each material confidence, the corresponding positive electrode material type is output, for example: LiFePO4 (LFP), LiCoO2 (LCO), LiMn2O4 (LMO), and LiNi x Co y Mn 1-x-yO2 (NCM, including NCM523, NCM622 and NCM811 with Ni:Co:Mn ratio of 5:2:3, 6:2:2 and 8:1:1) four material types, explainable result report; if the maximum confidence is lower than the threshold (<0.85), the sample is determined as a low confidence sample, the sample is classified as other type in output type and the explainable result report is output, at the same time, the XRF spectrum data of the sample is cached to the temporary database, the clustering algorithm in the temporary database induces the samples that may belong to the same unknown category into a new category cluster, when the number of sample data in the new category cluster accumulates to 100, the incremental classifier based on the incremental learning model is started to be constructed and is registered to the classifier library; in order to further enhance the transparency and credibility of model decision, the explainable result report method is obtained: by constructing an auxiliary explanation model, the activation values of the key layers in the model are extracted and output, such as the high-level spectral features compressed by the global average pooling layer and the attention weight distribution reflecting the focus of the model. These features together reveal the internal basis of the model for classification decision. Finally, a structured prediction and explanation report is generated, which comprehensively contains the predicted material type, the automatically detected element ratio information, the prediction probability of all categories, and the attention weight information reflecting the focus of the model decision. The report not only provides the final classification result, but also provides in-depth insight into the model decision process for users, meeting the dual needs of algorithm reliability and explainability in industrial sorting applications.

[0068] The basic steps of implementing incremental classifier construction based on incremental learning: cache the XRF spectrum data of the sample to a temporary database, the integrated HDBSCAN clustering algorithm in the temporary database will induce samples that may belong to the same unknown category into a new category cluster, and trigger training when the number of sample data in the new category cluster accumulates to 100. Identify potential unknown category samples, where the key parameters are the minimum number of clustering samples (e.g. 100), clustering selection epsilon (e.g. 0.5), then exclude outliers (noise samples), and finally assign a unique material ID to each valid cluster. Then train an independent incremental classifier for each valid cluster material, the key steps include data preparation: extract positive samples from the cluster, perform the same data augmentation as in the main classifier, then randomly extract the same number of negative samples from the known category samples used to build the main classifier, feature extraction layer (reuse main classifier knowledge), classification decision layer: lightweight CNN binary classifier fast training: after global feature compression, directly perform Sigmoid activation, directly output binary classification probability value. From the beginning of construction to the completion of training as a loop, the binary classification model produced each time will be merged with the binary classification model produced last time, the merged binary classification model is called an incremental learning model, which is registered to the classifier library as an incremental classifier through iterative method and disaster forgetting prevention. With the main classifier and incremental classifier in the classifier library, when a new batch of battery positive materials comes, the material XRF spectrum data can be identified through the models in the classifier library, and at the same time if there are low confidence samples, the same method as above is used to expand and update the incremental classifier.

[0069] The application also provides a retired lithium ion battery positive electrode material intelligent sorting device based on spectroscopy, which comprises: the material pretreatment module, which is used for mechanical pretreatment of waste lithium ion batteries, mainly including a cutter assembly and a clamp assembly. The cutter assembly is composed of a set of fixed knives and a set of movable moving knives, and effective cutting of the battery shell and internal structure is realized through relative movement. The clamp assembly is provided with two groups of rotatable clamping mechanisms, which can multi-axis position and clamp the battery sample, so that omnidirectional cutting in different directions is realized. The cleaning module includes a cleaning tank and a spring vibration screen. The inside of the cleaning tank is provided with a material channel, and the inlet and outlet thereof are provided with adjustable movable baffles for controlling the material flow rate and adjusting the flow of the cleaning liquid. The spring vibration screen is installed below the cleaning tank and is used for water-solid separation and impurity screening while the material is subjected to spray cleaning. A plurality of liquid spray heads are arranged above the cleaning tank for uniformly high-pressure spray cleaning of the passing material to improve the cleaning effect. The vibration screen realizes high-frequency vibration by relying on the spring support structure to assist the cleaning liquid to carry away the surface attached impurities and particles to ensure the cleanliness of the material. The identification module includes an X-ray fluorescence spectrometer (XRF) and an industrial personal computer (IPC), which are connected through a data communication interface and are jointly fixed and installed on the gantry structure. The XRF instrument uses primary X-rays (or particles) to excite atoms in the sample, so that inner layer electrons are knocked out and outer layer electrons jump to fill the vacancies, releasing X-ray fluorescence (characteristic X-rays) with characteristic energy of the element. By detecting the energy or wavelength of these characteristic X-rays, the element type in the sample can be determined (qualitative analysis), and quantitative analysis can be performed by measuring the intensity, so the recorded curve has the horizontal coordinate as the energy of X-ray photons (unit: keV) and the vertical coordinate as the photon intensity (count rate); the industrial personal computer integrates a pre-trained classifier library, which can receive the spectrum data collected by the XRF in real time, automatically identify different types of positive electrode materials, and transmit the identification results to the subsequent sorting module to guide the material diversion. The sorting module includes four hydraulic drive push rod devices, each push rod is driven by a hydraulic cylinder, and is used for sorting the identified different types of positive electrode materials from the conveying platform to the corresponding recycling path. The hydraulic push rod is arranged at equal intervals along one side of the conveying platform, can push the target material in a directional manner according to the sorting instruction output by the upper identification module, and realizes rapid and accurate diversion operation. Each hydraulic cylinder is controlled through a control valve group. The support module includes a rack structure and a conveying belt assembly, which jointly constitute an integrated material conveying platform. The conveying belt sequentially penetrates the material pretreatment module, the cleaning module, the identification module and the sorting module along the conveying path, and is used for realizing automatic transfer of the waste lithium ion battery between the functional modules. The rack structure is composed of a platform support, a top support and a bottom support, the platform support is used for bearing the conveying belt and fixing the functional modules, and the top and bottom supports provide overall stability and structural support to ensure the mechanical rigidity and reliability of the equipment during operation.

[0070] Example 1

[0071] Referring to Figure 1 The flowchart of the intelligent sorting method of the retired lithium-ion battery positive electrode material provided in this embodiment is shown. A batch of retired lithium-ion battery positive electrode material samples are obtained from a new energy vehicle retired battery recycling enterprise. These samples contain multiple different types of positive electrode materials mixed together. These mixed samples are placed in a pretreatment module, cleaned by a metal grid dispersion screen and an adjustable air flow nozzle, and then transferred to an identification module. A gantry XRF spectrometer performs spectral scanning on the cleaned samples to obtain their original XRF spectral data, which are transmitted to an industrial computer. The industrial computer automatically performs energy calibration, multi-window iterative SNIP background subtraction, wavelet transform denoising, and intensity normalization preprocessing operations on the spectral data in sequence. The preprocessed spectral data are subjected to element feature and spectral shape feature extraction to obtain corresponding element feature vectors and spectral shape feature vectors. The collected XRF spectral data are preprocessed, data enhanced, and the corresponding element feature vectors and spectral shape feature vectors are obtained, and a dual-channel neural network model is constructed and trained to obtain a pre-trained main classifier based on the dual-channel neural network. When the main classifier receives these feature information, it outputs the probability distribution of the positive electrode material type to which each sample belongs, the corresponding confidence and the result explanation through forward propagation operation and analysis inside the model. For samples with a confidence of ≥0.85, their corresponding positive electrode materials are directly pushed into the collection area of the corresponding type according to the classification result; for samples with a confidence of <0.85, they are marked as “unknown type” and pushed into the collection area of other types, and the XRF spectral data of the sample are cached to a temporary database. The HDBSCAN clustering algorithm integrated in the temporary database induces samples that may belong to the same unknown category into a new category cluster. When the number of sample data in the new category cluster in the temporary database accumulates to 100, an incremental learning process is started to develop an incremental classifier. According to the steps described in the invention content, such as the training of a lightweight CNN binary classifier, the incremental classifier obtained after merging is registered to the classifier library, realizing the continuous expansion and update of the classifier library, so that it can identify new positive electrode material categories. Result verification: through the actual operation of this example, the method and device of the present application successfully classified the mixed retired lithium-ion battery positive electrode materials with a classification accuracy of more than 97%, and for unknown positive electrode materials, they can also be quickly and accurately identified after incremental learning, effectively improving the classification accuracy and recycling efficiency of retired lithium-ion battery positive electrode materials, verifying the feasibility and practicality of the invention content.

[0072] Example 2

[0073] Referring to Figure 2This is a schematic diagram of the intelligent sorting device for cathode materials of retired lithium-ion batteries provided in this embodiment. It includes a pretreatment module 1, a cleaning module 2, an identification module 3, a sorting module 4, and a support module 5. Modules 1-4 are connected via module 5. In this embodiment, the intelligent sorting device for cathode materials of retired lithium-ion batteries includes a frame structure and a conveyor belt assembly, forming an integrated material conveying platform. The conveyor belt sequentially passes through the material pretreatment module, cleaning module, identification module, and sorting module along the conveying path. The frame structure consists of a platform support, a top support frame, and a bottom support frame. First, the retired lithium-ion battery is cut into blocks by the pretreatment module to remove unknown cathode materials. The clamping part in the pretreatment module consists of two sets of clamps. Figure 3 The fixtures 6, 7, 8, and 9 are all capable of lifting, lowering, and horizontal movement. Fixtures 6 and 7 form one group, and fixtures 8 and 9 form another. Fixtures 6 and 7, along with fixtures 8 and 9, are always at the same height. The fixture assembly has two sets of rotatable clamping mechanisms, enabling multi-axis positioning and clamping of the battery sample, thus achieving omnidirectional cutting in different directions. The cutting process is as follows: Fixtures 6 and 7 clamp the battery, fixtures 8 and 9 descend to a height that does not affect the cutting process, and then the cutter makes a longitudinal cut on the battery. After cutting, fixtures 8 and 9 rise to the same height as fixtures 6 and 7, move horizontally, clamp the retired lithium-ion battery, fixtures 6 and 7 move horizontally, release the battery, and descend to a height that does not affect the cutting process. The entire fixture assembly rotates 90°, so that fixtures 8 and 9 replace the positions of fixtures 6 and 7, and then the cutter makes a transverse cut on the battery. Cutting section ( Figure 4 , 5 The system mainly consists of a fixed cutting unit 10, a movable cutting unit 11, and a motion mechanism 12 that controls the vertical movement of the entire cutting unit. The horizontal movement of the cutting unit 10 is achieved by a lead screw mechanism 13 controlled by a motor 14, while the vertical movement mechanism 12 relies on a built-in fixed belt drive 15 to move the cutting tool up and down. The material obtained after cutting contains both positive and negative electrode materials, as well as electrolyte components. The main components of lithium battery electrolyte are lithium-containing salts and some corrosive chemical components. The lithium-containing salts may affect the subsequent identification process, and the corrosive liquid may affect the entire mechanical device. To prevent the electrolyte from affecting the entire process, a cleaning module is needed to clean the pre-treated material. The purpose of the cleaning module is to completely wash away the electrolyte components in the material by rinsing with chemical reagents. In addition, the pre-treated material needs to be separated into positive and negative electrodes for subsequent identification and sorting processes. The cleaning tank (…) Figure 6The system mainly consists of three parts: a spray nozzle 16 for spraying liquid, a spring vibrating screen 17, and a baffle 18. The spray nozzle 16 sprays cleaning liquid to wash the material. The spring vibrating screen 17 consists of four helical springs, a screen plate with a certain inclination, and a vibrating motor. The spring vibrating screen has a certain inclination, with a height difference between the inlet and outlet, allowing the cleaned material to enter the next module for further processing. After the electrode material undergoes cutting in the pretreatment module, electrolyte cleaning in the cleaning module, and separation of the positive and negative electrodes, its main components are positive electrode material, negative electrode material, cleaning liquid, dust, and other impurities. Before identifying small blocky materials, a purification process is required. This purification process mainly uses six air nozzles 19 (… Figure 7 The process involves air nozzles using airflow to clean contaminants, removing residual liquid, drying materials, and removing debris. After this cleaning process, the electrode material enters the identification module for spectral analysis. The identification module contains an X-ray fluorescence spectrometer (XRF). By capturing X-ray fluorescence spectra and combining this with a deep learning model of the battery cathode material classification system built into the host computer using a unified classifier and incremental learning scheme, the module captures and analyzes the eigenvalues ​​and characteristic peaks of the spectrum to identify the cathode material. Finally, the identified data is transmitted to the sorting module. The sorting module (…) Figure 8 Two hydraulic actuators (30, 31) are used to concentrate the positive electrode material after passing through the identification module onto the same track, facilitating the subsequent sorting process. The identification module first determines the material type, and the control valve adjusts the thrust of the two hydraulic actuators (28, 29) according to the material type, pushing the material to different conveying tracks. Lithium iron phosphate (LiFePO4) material is pushed to track 20, and ternary lithium (LiNi) material is pushed to track 20. x Co y Mn 1-x-y O2 material is propelled to track 21, lithium cobalt oxide (LiCoO2) material is propelled to track 22, and other unknown types of cathode materials are propelled to track 23. Different transport tracks lead to different collection devices, ultimately achieving precise sorting of cathode material types. The collection area is used to collect different types of cathode materials. The collection area is divided into four collection boxes (…). Figure 9 Collection box 24 collects lithium iron phosphate (LiFePO4) material, and collection box 25 collects ternary lithium (LiNi) material. x Co y Mn 1-x-y The cathode material is divided into two categories: one for collecting lithium cobalt oxide (LiCoO2) materials, and the other for collecting other unknown types of cathode materials. Therefore, this embodiment details the device process, improving the accuracy of cathode material sorting within retired batteries.

Claims

1. A smart sorting method for cathode materials of retired lithium-ion batteries based on spectroscopic methods, characterized in that, The method includes the following specific steps: First, XRF spectra were collected, and the spectra of various cathode materials were obtained through software simulation and experimental detection. A spectrum covering LiFePO4, LiCoO2, LiMn2O4, and LiNi was constructed. x Co y Mn 1-x-y An XRF spectral database of O2 material type is used as the data foundation for training the main classifier; at the same time, a temporary database integrating clustering algorithm and capable of caching multiple sample data is constructed as the data source for training the incremental classifier; all spectra share the same set of energy axes, and the corresponding intensity values ​​are stored one by one, with the intensity values ​​varying from sample to sample. Subsequently, a classifier library was built and stored on the industrial control computer. This classifier library consists of a main classifier based on a dynamic dual-channel neural network model trained from the XRF spectral database and an incremental classifier based on an incremental learning model. The construction of the main classifier is as follows: data preprocessing is performed on the data in the XRF spectral database, followed by data augmentation to expand the sample size; elemental features and spectral shape features are extracted in parallel from all augmented spectral data; the two types of features are input into the dynamic dual-channel neural network model, and finally, the probability distribution of the cathode material category is obtained through the classification output layer; the network parameters are optimized using the backpropagation algorithm to train the model to maximize the classification accuracy of the training samples; the dynamic dual-channel neural network model obtained after training is registered as the main classifier in the classifier library of the industrial control computer. Then, when the cathode material to be tested passes through the industrial control computer, the main classifier in its classifier library identifies the cathode material. During this process, the main classifier in the classifier library first predicts and interprets the cathode material, and then determines high-confidence samples and low-confidence samples based on its confidence threshold. If it is a low-confidence sample, it is classified as an unknown type in the output and the result is interpreted. At the same time, the XRF spectral data of the sample is cached in a temporary database. The clustering algorithm integrated in the temporary database will group samples that may belong to the same unknown category into a new category cluster. When the number of sample data in the new category cluster reaches 100, the incremental learning model training process is automatically triggered. The incremental learning model obtained after merging is used as the incremental classifier and registered in the classifier library. When the incremental learning model training process is triggered, the total amount of sample data cached in the temporary database will decrease accordingly until the number of sample data in the new category cluster in the cache reaches 100 again, at which point the incremental learning model training process is automatically triggered. This cycle continues, thereby realizing the continuous expansion and updating of the classifier library. Finally, after material pretreatment and cleaning, the cathode material to be tested is identified by the classifier library in the industrial control computer, the material type is output, and the sorting result is obtained.

2. The intelligent sorting method for cathode materials of retired lithium-ion batteries as described in claim 1, characterized in that, The data preprocessing of the XRF spectral database specifically includes: Perform the following steps sequentially for each piece of raw spectral data: Energy axis calibration uses the characteristic peaks of standard samples to perform nonlinear calibration of the energy channel in order to control the calibration error within a set range; Background subtraction employs an iterative background estimation algorithm, performing multiple iterations within an adaptive-width iteration window; Noise suppression is achieved by using wavelet transform to perform multi-scale decomposition and threshold reconstruction on the spectrum after background subtraction. Normalization uses the integrated intensity of the Compton scattering peak as an internal standard to normalize the overall spectral amplitude to a uniform dimension. The spectral data after data preprocessing will be re-entered into the database.

3. The intelligent sorting method for cathode materials of retired lithium-ion batteries as described in claim 1, characterized in that, The data augmentation technique involves sequentially performing random perturbation operations on the preprocessed XRF spectra, specifically: Noise consistent with photon counting statistics is introduced to simulate signal fluctuations; a small random offset is applied to the energy axis and re-interpolated and sampled; different degrees of Gaussian blur are applied to change the spectral resolution; a random low-order polynomial background is superimposed to simulate matrix effect differences; the intensity of the spectral peaks corresponding to the feature elements is randomly adjusted within a range of ±15%; then fine-tuning and normalization are performed to ensure that the final scale of each sample is consistent; thus generating spectral samples that are physically consistent and significantly more diverse to expand the XRF spectral database.

4. The intelligent sorting method for cathode materials of retired lithium-ion batteries as described in claim 1, characterized in that, The parallel extraction of elemental and spectral features from all enhanced spectral data specifically involves: Element feature extraction is completed in one operation for all energy spectra: After inputting the energy axis and intensity axis data, the main peak energy position is first determined based on the maximum intensity value, and it is then checked whether there are significant characteristic peaks exceeding the main peak intensity by 20-40% within the characteristic energy window of each predefined element. Based on this, the spectrum is classified into one of the predefined cathode material types. Subsequently, based on the determined material type, the corresponding elemental characteristic energy range is called, and the peak area of ​​the intensity-energy curve in each range is calculated using the numerical integration method. The sum of the peak areas of all elements is then taken as the total effective area. If the total effective area exceeds a small threshold, the peak area of ​​each element is divided by the total effective area to obtain the normalized ratio; otherwise, the ratio is set to zero. Finally, the ratios are arranged into a fixed-dimensional ratio vector according to a predetermined standard element order, which is used as the element ratio feature output. Spectral feature extraction: The complete intensity sequence after data preprocessing or data augmentation is resampled to a fixed dimension and used as the feature output.

5. The intelligent sorting method for cathode materials of retired lithium-ion batteries according to claim 1, characterized in that, The dynamic dual-channel neural network model is constructed as follows: A dynamic dual-channel deep learning model is used for processing. The model includes a spectral data input channel and a dynamic ratio vector input channel. The spectral data input channel uses a one-dimensional convolutional neural network with multiple residual block structures to extract features and includes an attention mechanism implemented through attention weight generation and feature weighting. The dynamic ratio vector input channel performs feature transformation through a fully connected layer. The features of the two channels are fused through a dynamic weighting mechanism, which combines spectral features with material type information to generate adaptive weights and weight the spectral features. The weighted spectral features are fused with the processed ratio features. The fused features are then processed through a fully connected layer and a random deactivation layer, and finally the results are output through a classification output layer.

6. The intelligent sorting method for cathode materials of retired lithium-ion batteries according to claim 1, characterized in that, The training of the dynamic dual-channel neural network model is specifically as follows: The extracted elemental features and spectral features are used to construct a dataset, which is then divided into a training set and a validation set according to a predetermined ratio. The connection weights of the dynamic dual-channel neural network model are initialized. The callback functions used in the training process are configured, including an early stopping strategy and a dynamic learning rate decay strategy. The model is trained iteratively using the training set data, and the model performance is evaluated using the validation set data after each iteration. When the validation set performance no longer improves in consecutive iterations, the training process is automatically terminated and the model weights with the best validation performance are restored. Finally, the trained model is obtained as the main classifier.

7. The intelligent sorting method for cathode materials of retired lithium-ion batteries according to claim 1, characterized in that, The main classifier in the classifier library will first predict and interpret the results for the cathode material, specifically as follows: When the cathode material under test passes through the industrial control computer, the main classifier in the classifier library preprocesses the XRF spectral data of the cathode material collected by the X-ray fluorescence spectrometer mounted on the industrial control computer, and extracts its standardized elemental feature vector and spectral shape feature vector. These two feature vectors are simultaneously input into the trained dynamic dual-channel neural network model for forward propagation calculation to obtain the probability distribution of each material category. If any of the output confidence scores for each material category reach or exceed a predetermined confidence threshold, the test is performed. If the XRF spectral database contains four types of materials, then the material is classified as a high-confidence sample, and the corresponding cathode material type and result interpretation are output. If any cathode material in the output has a confidence level below a predetermined confidence threshold, meaning it is not within the range of materials covered by the XRF spectral database, then the material is classified as a low-confidence sample. This sample is then classified as an unknown type in the output and the result interpretation is performed. At the same time, the XRF spectral data of this sample is cached in a temporary database. All output result interpretations include the following: extracting the output features of the intermediate layers inside the model, including spectral feature compression representation and attention weight distribution, and generating an interpretable result report containing the predicted material type, element ratio information, probability of each category, and the basis for model decision-making.

8. The intelligent sorting method for cathode materials of retired lithium-ion batteries as described in claim 1, characterized in that, The process of automatically triggering the training of the incremental learning model is as follows: The incremental learning model obtained after merging is used as an incremental classifier and registered in the classifier library. Whenever the number of sample data for a new category cluster in the temporary database reaches 100, 100 sample data for the new category cluster are automatically extracted from the temporary database, thus reducing the total amount of sample data cached in the temporary database. Then, for each validated new category cluster, a lightweight binary classification convolutional neural network model is constructed. This model is initialized using knowledge distillation, which preserves existing knowledge by reusing and freezing the weights of some feature extraction layers in the main model. Subsequently, the model is trained using positive samples from the new category cluster and an equal number of negative samples extracted from known categories, enabling it to learn to distinguish the new category from all other categories. There are existing categories; the process from initial construction to completion of training is considered a loop. Each loop generates a binary classification model that is merged with the previously generated model. The merged binary classification model is called an incremental learning model, which is used iteratively as an incremental classifier and registered in a classifier library to expand and update the library. Positive samples are low-confidence samples that the main classifier could not identify, and these samples are clustered to form new clusters with a sample size of 100. Negative samples are high-confidence samples that the main classifier has already successfully identified, and their number is the same as the positive samples.

9. A smart sorting device for cathode materials of retired lithium-ion batteries based on the method of claim 1, characterized in that, The device includes: a material pretreatment module, a cleaning module, an identification module, a sorting module, and a support module; The support module, comprising a frame structure and a conveyor belt assembly, forms an integrated material conveying platform. The conveyor belt sequentially passes through the material pretreatment module, cleaning module, identification module, and sorting module along the conveying path, enabling automatic transfer of waste lithium-ion batteries between these modules. The frame structure consists of a platform support, a top support frame, and a bottom support frame. The platform support supports the conveyor belt and secures each module, while the top and bottom support frames provide overall stability and structural support, ensuring mechanical rigidity and reliability during operation. This support module achieves an orderly layout and continuous operation support, forming the foundation for efficient operation. The material pretreatment module is used for mechanical pretreatment of waste lithium-ion batteries, including a tool assembly and a fixture assembly. The tool assembly consists of a set of fixed blades and a set of movable blades, which achieve effective cutting of the battery shell and internal structure through relative motion. The fixture assembly is equipped with two sets of rotatable clamping mechanisms, which can perform multi-axis positioning and clamping of the battery sample, thereby achieving omnidirectional cutting in different directions. This module achieves efficient, stable and precise disassembly of the battery sample by coordinating the relative motion of the tool and the fixture. The cleaning module includes a cleaning tank and a spring vibrating screen. The cleaning tank has a material channel with adjustable baffles at its inlet and outlet to control the material flow rate and the flow rate of the cleaning solution. The spring vibrating screen is installed below the cleaning tank to achieve water-solid separation and impurity removal while the material is being sprayed clean. Spray nozzles are arranged above the cleaning tank to uniformly spray the passing material with high pressure. The vibrating screen relies on a helical spring support structure to achieve high-frequency vibration, assisting the cleaning solution in carrying away surface-adhered impurities and particles, ensuring the cleanliness of the material. The identification module includes an X-ray fluorescence spectrometer, an industrial control computer, and a gantry structure. The X-ray fluorescence spectrometer and the industrial control computer are connected via a data communication interface and are fixedly installed on the gantry structure. The X-ray fluorescence analyzer is located at the entrance of the sorting module and is used to perform elemental composition analysis on the cathode material entering the sample. The industrial control computer integrates a pre-trained classifier library, which can receive the spectral data acquired by XRF in real time, automatically identify different types of cathode materials, and transmit the identification results to the subsequent sorting module to guide material separation. The sorting module includes multiple hydraulically driven push rods, each driven by a hydraulic cylinder, used to sort different types of identified cathode materials from the conveyor platform to their corresponding recycling paths. The multiple hydraulically driven push rods are arranged at equal intervals along one side of the conveyor platform, and can directionally push the target material according to the sorting instructions output by the upper identification module to achieve rapid and accurate diversion operation. Each hydraulic cylinder is controlled by a control valve group. This sorting module is the key execution module for realizing automated sorting operations.

Images