A method for sorting and clustering retired lithium-ion batteries based on semi-supervised learning

Abstract

The application discloses a kind of based on semi-supervised learning's retired lithium ion battery sorting and clustering method.The method includes: constructing retired lithium ion battery discharge data processing flow;Double image coding and pre-training feature extraction framework are constructed;Interval selection mechanism based on significant feature map reflection is constructed;Two-stage semi-supervised contrast learning sorting and clustering mechanism is constructed.The method is by constructing retired lithium ion battery discharge data processing flow, double image coding and pre-training feature extraction, significant interval selection, two-stage semi-supervised contrast learning sorting and clustering integrated technology chain link, effectively improves the sorting accuracy, consistency and generalization ability of retired battery under the condition of few labels, class imbalance;By introducing pre-training convolutional network, the cost of two-dimensional feature extraction is significantly reduced;By significant feature map reflection technology, the most discriminant discharge interval is accurately positioned;By semi-supervised contrast learning mechanism, the representation ability of minority class samples is strengthened.

Description

Technical Field

[0001] This application relates to the field of battery sorting technology, and in particular to a method for sorting and clustering retired lithium-ion batteries based on semi-supervised learning. Background Technology

[0002] Before retired lithium-ion batteries are reused, they need to be effectively sorted and grouped according to their capacity retention, discharge characteristics, and consistency differences to avoid performance mismatch, shortened lifespan, and increased safety risks after being directly grouped.

[0003] Existing technologies mainly fall into two categories: direct testing methods and data-driven methods. Direct testing methods typically obtain battery performance indicators through capacity testing, pulse testing, internal resistance testing, or open-circuit voltage testing; data-driven methods achieve the sorting and clustering of retired lithium-ion batteries by analyzing charge-discharge curves or statistical characteristics.

[0004] However, existing technologies still have significant shortcomings in the sorting of retired lithium-ion batteries. Specifically, these methods rely too heavily on human experience and features, making it difficult to comprehensively capture the global trend and local dynamic differences in battery degradation. Furthermore, even when using two-dimensional image encoding techniques to enhance feature representation, the high computational cost often arises from training complex networks from scratch. In addition, the lack of an effective mechanism for reverse-locating key discharge intervals from the two-dimensional feature space makes extracting the most discriminative voltage segments extremely difficult. Finally, in scenarios with few labels and class imbalance, existing methods have limited ability to identify a small number of retired lithium-ion batteries, easily leading to sorting biases. Summary of the Invention

[0005] The main objective of this application is to provide a semi-supervised learning-based method for sorting and clustering retired lithium-ion batteries to solve the technical problems in the prior art.

[0006] To achieve the above objectives, this application provides a semi-supervised learning-based method for sorting and clustering retired lithium-ion batteries, the method comprising: S10, Construct a discharge data processing flow for retired lithium-ion batteries; S20, Construct a dual image coding and pre-trained feature extraction framework; S30, Construct an interval selection mechanism based on salient feature map replay; S40, constructing a two-stage semi-supervised comparative learning sorting and clustering mechanism.

[0007] Optionally, step S10 includes the following steps: S101 performs capacity testing, pulse testing, and open-circuit voltage testing on retired lithium-ion batteries to obtain test results; S102, Extract the SOC-discharge voltage curve based on the test results; S103, the SOC-discharge voltage curve is denoised, interpolated and resampled, and its length is standardized to make different battery samples comparable under the same data scale.

[0008] Optionally, step S20 includes: S201, Perform one-dimensional curve preprocessing on the SOC-discharge voltage curve, and convert the SOC-discharge voltage curve into GASF image and RP image respectively; S202, perform image encoding on the GASF image and RP image to obtain two types of encoding results; S203, the two types of encoding results are combined to form a pseudo-color image adapted to the input of the convolutional neural network, and the pseudo-color image is input into a pre-trained VGG16 network for hierarchical feature extraction.

[0009] Optionally, step S201 includes the following steps: S2011, normalizes the SOC-discharge voltage curve after noise reduction, interpolation resampling and length unification. S2012, the normalized SOC-discharge voltage curve is mapped to polar coordinate space, and a GASF image is constructed in combination with the corresponding discharge voltage to express the correlation of the sequence over the global time span. S2013, perform linear interpolation and first-order difference processing on the mapped SOC-discharge voltage curve, and construct an RP image by the Manhattan distance between the difference values ​​at any two sampling times to characterize the local variation relationship and state recurrence characteristics during the discharge process.

[0010] Optionally, step S202 includes the following steps: S2021, normalize the SOC sequence to a preset range, and convert the normalized SOC value into an angle variable, with the discharge voltage as the polar coordinate radius component; S2022 constructs a GASF matrix based on the cosine of the sum of the angle components at any two sampling times, thereby transforming the global correlation between different time positions in a one-dimensional time series curve into a spatial texture distribution in a two-dimensional image.

[0011] Optionally, step S30 includes: S301, extract feature maps from the intermediate convolutional layers of the pre-trained VGG16 network, and aggregate the multi-channel activation responses to generate a two-dimensional saliency map reflecting the discrimination strength. S302 calculates the difference heatmap between salient feature maps for multiple battery samples, and locates the spatial region with the strongest class discrimination by cross-sample difference enhancement. S303, the spatial region with the strongest category differentiation is reverse-mapped to the original discharge curve to obtain the optimal discharge range with the strongest discriminative power.

[0012] Optionally, step S40 includes: S401, preheating training is performed based on a preset number of labeled samples to establish a basic inter-class structure; S402, construct an initial class prototype based on the labeled samples, construct a class prototype based on the representation center of each class of samples, and perform similarity evaluation on the unlabeled samples; S403 adds high-confidence samples to the training set after assigning pseudo-labels, and enhances intra-class compactness and inter-class separation by comparing the learning target. S404 introduces a classification head, combines labeled samples and high-confidence pseudo-labeled samples for supervised training, and outputs the final sorting and clustering results of retired lithium-ion batteries based on semi-supervised learning.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: The semi-supervised learning-based sorting and clustering method for retired lithium-ion batteries provided in this application effectively improves the sorting accuracy, consistency, and generalization ability of retired batteries under conditions of few labels and class imbalance by constructing an integrated technical link for retired lithium-ion battery discharge data processing, dual image encoding and pre-trained feature extraction, salient interval selection, and two-stage semi-supervised contrastive learning sorting and clustering. By introducing a pre-trained convolutional network, the cost of two-dimensional feature extraction is significantly reduced; the most discriminative discharge interval is accurately located through salient feature map reflection technology; and the representation ability of minority class samples is strengthened through a semi-supervised contrastive learning mechanism. This method achieves high efficiency, accuracy, and robustness in the sorting and clustering of retired lithium-ion batteries based on semi-supervised learning, providing an effective way to solve the pain points of existing methods. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a semi-supervised learning-based sorting and clustering method for retired lithium-ion batteries provided in this application embodiment.

[0015] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the application. Rather, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0017] To address the aforementioned technical problems, this application provides a semi-supervised learning-based method for sorting and clustering retired lithium-ion batteries. This method can be executed by a computer, such as... Figure 1 As shown, the method may include the following steps: S10, Construct a discharge data processing flow for retired lithium-ion batteries.

[0018] In an exemplary embodiment, step S10 may include the following steps: S101 performs capacity testing, pulse testing, and open-circuit voltage testing on retired lithium-ion batteries to obtain test results; S102, Extract the SOC-discharge voltage curve based on the test results; S103, the SOC-discharge voltage curve is denoised, interpolated and resampled, and its length is standardized to make different battery samples comparable under the same data scale, providing a consistent data foundation for subsequent image encoding and feature learning.

[0019] In this embodiment, in order to obtain capacity characteristics, pulse characteristics, and open-circuit voltage characteristics for sorting and clustering retired lithium-ion batteries based on semi-supervised learning, retired lithium-ion battery cells are tested under uniform operating conditions. The test content includes capacity testing, pulse testing, and open-circuit voltage testing.

[0020] Specifically, the test platform includes a battery test system, a constant temperature chamber, and a host computer. The battery test system is used to perform charge and discharge control and synchronously collect voltage and current data. The constant temperature chamber is used to maintain the test environment at a constant temperature. The host computer is used to issue test commands, record the test process, and store test results.

[0021] Furthermore, before testing, the batteries to be tested can be numbered, visually screened, and have their initial voltage consistent to eliminate samples with visual damage, abnormal terminal voltage, or unstable connection. For samples that pass the initial screening, capacity testing, pulse testing, and open-circuit voltage testing are then performed sequentially under a constant temperature environment to ensure that raw data for different battery samples are obtained under uniform operating conditions, a uniform testing sequence, and uniform data acquisition rules.

[0022] When extracting the SOC-discharge voltage curve from test results, the SOC-time relationship can be calculated based on the discharge capacity integral result or the ampere-hour measurement result recorded by the test system. The SOC sequence during the discharge process is then synchronized with the corresponding voltage sequence in time to form the SOC-discharge voltage curve for subsequent encoding. For noise suppression, moving average, local smoothing, or outlier removal methods can be used. For uniform length, different samples can be mapped to a uniform number of sampling points through interpolation, and then the normalized sampling sequence can be used as the input for subsequent two-dimensional encoding.

[0023] In capacity testing, the battery under test is first charged using a constant current and constant voltage charging protocol until it reaches full charge. Then, the battery is discharged at a rate of 0.5C and three charge-discharge cycles are performed continuously. The average discharge capacity of the three cycles is taken as the capacity characterization index of the battery, thereby obtaining basic data reflecting the remaining energy storage capacity and capacity retention level of retired lithium-ion batteries.

[0024] In the pulse test, the fully charged battery is first discharged to 80% SOC at a rate of 0.5C; then, charge and discharge pulses of 3C, 2C, 1C and 0.5C are applied sequentially to obtain the dynamic response characteristics of the battery under different load intensities; further, the battery internal resistance or polarization response is evaluated at 80%, 50% and 20% SOC nodes to obtain pulse response data for characterizing the dynamic power characteristics and consistency of retired lithium-ion batteries.

[0025] In the open-circuit voltage test, the battery was charged and discharged at a low rate of 0.05C. The voltage curves obtained during the low-rate charge and discharge process were averaged to obtain the open-circuit voltage curve characterizing the quasi-static voltage characteristics of the battery. After completing the above test, the SOC-discharge voltage curve was extracted from the obtained test data as the core input signal for subsequent sorting and clustering. At the same time, the capacity, internal resistance, and open-circuit voltage information were used as references for the characterization of the state of retired lithium-ion batteries and the verification of results.

[0026] S20, construct a dual image coding and pre-trained feature extraction framework.

[0027] In an exemplary embodiment, step S20 may include the following steps: S201, Perform one-dimensional curve preprocessing on the SOC-discharge voltage curve, and convert the SOC-discharge voltage curve into GASF image and RP image respectively; S202, perform image encoding on the GASF image and RP image to obtain two types of encoding results; S203, the two types of encoding results are combined to form a pseudo-color image adapted to the input of the convolutional neural network, and the pseudo-color image is input into a pre-trained VGG16 network for hierarchical feature extraction.

[0028] Among them, the GASF image is used to map the global correlation of the time series curve to the polar coordinate angle space to enhance the expression of the overall degradation trend; the RP image is used to characterize the recursive relationship of adjacent state changes and local dynamic differences during discharge to enhance the expression of local structural changes.

[0029] In an exemplary embodiment, step S201 may include the following steps: S2011, normalizes the SOC-discharge voltage curve after noise reduction, interpolation resampling and length unification. S2012, the normalized SOC-discharge voltage curve is mapped to polar coordinate space, and a GASF image is constructed in combination with the corresponding discharge voltage to express the correlation of the sequence over the global time span. S2013, perform linear interpolation and first-order difference processing on the mapped SOC-discharge voltage curve, and construct an RP image by the Manhattan distance between the difference values ​​at any two sampling times to characterize the local variation relationship and state recurrence characteristics during the discharge process.

[0030] In an exemplary embodiment, step S202 may include the following steps: S2021, normalize the SOC sequence to a preset range, and convert the normalized SOC value into an angle variable, with the discharge voltage as the polar coordinate radius component; S2022 constructs a GASF matrix based on the cosine of the sum of the angle components at any two sampling times, thereby transforming the global correlation between different time positions in a one-dimensional time series curve into a spatial texture distribution in a two-dimensional image.

[0031] Specifically, for the RP image, the SOC sequence can first be linearly interpolated to ensure that each battery sample has the same sampling length; then, the first-order difference of the interpolated SOC sequence can be calculated to obtain the state change between adjacent sampling points; finally, the Manhattan distance between the difference values ​​of any two sampling times can be calculated, and the RP matrix can be constructed using this distance. Thus, low-value regions in the RP image can reflect similar states or stable stages, while high-value regions can reflect local mutations and state transitions.

[0032] In the sub-process of hierarchical feature extraction using a pre-trained VGG16 network, the pre-trained weights of the VGG16 network can be derived from training results on a general image dataset. Specifically, the parameters of the convolutional backbone layers can be kept frozen or only fine-tuned to avoid overfitting when there are few retired lithium-ion battery samples. Simultaneously, saliency maps are output from the intermediate convolutional layers to analyze the required spatial features, and high-dimensional semantic features are output from the fully connected layers to represent the subsequent backbone network. In other words, the "conversion process" in this step addresses how a one-dimensional curve is transformed into a two-dimensional image, while the "pre-trained feature extraction" addresses how a two-dimensional image is transformed into a hierarchical feature representation usable for subsequent classification.

[0033] After obtaining the GASF and RP images, they are combined according to a preset channel order to form a pseudo-color image adapted to the input format of the convolutional neural network. The "input pre-trained VGG16 network for hierarchical feature extraction" specifically refers to: feeding the pseudo-color image into a VGG16 network that has been pre-trained based on general image data, and preferentially freezing the parameters of the convolutional feature extraction layer, adjusting them only in subsequent task-related layers, to utilize its existing visual representation capabilities to stably output intermediate convolutional feature maps and high-dimensional features from fully connected layers. The intermediate convolutional feature maps are mainly used for subsequent salient region selection, while the high-dimensional features output by the fully connected layers are mainly used for subsequent backbone representation network construction and two-stage semi-supervised comparative learning.

[0034] S30, Construct an interval selection mechanism based on salient feature map reflection.

[0035] In an exemplary embodiment, step S30 may include the following steps: S301, extract feature maps from the intermediate convolutional layers of the pre-trained VGG16 network, and aggregate the multi-channel activation responses to generate a two-dimensional saliency map reflecting the discrimination strength. S302 calculates the difference heatmap between salient feature maps for multiple battery samples, and locates the spatial region with the strongest class discrimination by cross-sample difference enhancement. S303, the spatial region with the strongest category differentiation is reverse-mapped to the original discharge curve to obtain the optimal discharge range with the strongest discriminative power.

[0036] The pre-trained VGG16 network in this step not only provides image features, but also outputs intermediate convolutional feature maps with hierarchical semantic information.

[0037] Specifically, the implementation of step S301, "extracting feature maps from intermediate convolutional layers of the pre-trained VGG16 network," is as follows: The aforementioned pseudo-color image is input into the pre-trained VGG16 network, and a feature tensor A is extracted from the preset intermediate convolutional layers. This feature tensor can be represented as A∈R^(C×H×W), where C represents the number of channels, and H and W represent the spatial dimensions. Since different convolutional layers correspond to different receptive fields and semantic abstraction levels, this embodiment preferentially selects deep convolutional layers as saliency region analysis layers to improve the stability of the discrimination of key discharge regions. This section explains how the intermediate convolutional feature maps are obtained and used for subsequent saliency map construction.

[0038] In practice, a pre-trained VGG16 algorithm is first used to perform forward propagation on all pseudo-color images, and the output feature maps of different convolutional blocks are obtained. Then, the performance of shallow and deep convolutional layers in distinguishing differences between regions is compared. Since shallow convolutional layers tend to preserve local textures, while deep convolutional layers tend to fuse long-range context and high-level semantic information, deep convolutional layers are preferred as salient region analysis layers. For example, the output of the final convolutional block is used as the basis for subsequent salient map construction.

[0039] After the feature tensors are output from the intermediate convolutional layers, a three-dimensional tensor can be obtained independently for each battery sample, and then saved according to the sample number. This tensor is used for subsequent construction of saliency maps sample by sample and difference heatmaps across samples. In other words, this embodiment does not directly obtain the optimal interval from the entire batch of images at once, but first calculates the saliency map for a single sample, then compares the differences in saliency maps between samples, and finally comprehensively determines the optimal discharge interval.

[0040] The specific process for generating a two-dimensional saliency map reflecting the discrimination strength is as follows: The activation responses of each channel in the selected convolutional layer are aggregated by channel mean averaging at the same spatial location. That is, the average of all channel responses Ac(x,y) at each spatial coordinate (x,y) is calculated to obtain the two-dimensional saliency map F(x,y). The larger the obtained F(x,y) value, the stronger the contribution of that location to class discrimination, thus forming an interpretable two-dimensional discrimination strength distribution.

[0041] The process of locating the spatial region with the strongest class discrimination through cross-sample difference enhancement is described in detail below: For each battery sample, a two-dimensional saliency map Fk is obtained. Then, the difference map D(i,j) = |Fi-Fj| between any two battery sample saliency maps is calculated. Finally, the average of all difference maps is calculated to obtain a comprehensive difference heatmap Davg. The regions with higher responses in the comprehensive difference heatmap correspond to the locations where the differences between different battery categories are most significant, and therefore can be used as the basis for locating candidate key intervals.

[0042] Next, the comprehensive difference heatmap is upsampled to the same spatial resolution as the encoded image, and then the significant response rectangular region Ω=[i1,i2]×[j1,j2] is extracted from this heatmap. Since there is a one-to-one correspondence between the two-dimensional encoded image and the original SOC-discharge voltage curve in terms of sampling index, this significant region can be mapped to a one-dimensional interval [a,b]=[min(i1,j1),max(i2,j2)]. This interval is the optimal discharge interval, which is used to represent the discharge segment with the strongest discriminative information that is preferentially retained during learning and classification decisions, reducing interference from irrelevant data.

[0043] Specifically, two-dimensional saliency map aggregation can be achieved by averaging the activation responses of each channel in the selected convolutional layer to obtain an interpretable two-dimensional discriminative intensity distribution. The optimal interval mapping then determines the corresponding one-dimensional SOC-voltage index interval based on the coordinate range of the salient region in the encoding plane. Through this design, the high-response region in the two-dimensional feature space can be effectively mapped back to the key discharge segment in the original curve.

[0044] Next, the activation values ​​of all channels in the output feature tensor of the convolutional layer at the same spatial location are averaged. This process yields a two-dimensional discriminative intensity distribution. This involves fixing the output feature tensor of a convolutional layer, then summing the activation values ​​of all channels at that spatial location point by point using spatial coordinates, and dividing by the total number of channels. The result is a two-dimensional matrix with the same spatial size as the convolutional layer. Each element in this matrix represents the average response intensity of the classification task at the corresponding spatial location, and therefore can be directly used as a saliency map.

[0045] To determine the one-dimensional index interval, an interval validity check can be added after mapping. This check examines whether the interval length is too short, exceeds the original sampling range, and covers the main discharge stage. If the preset conditions are not met, the mapping can be reverted to the next candidate significant region. With these additions, the process of "determining the one-dimensional SOC-voltage index interval based on the coordinate range" can be interpreted as a clear process of "checking the sampling point location based on the index mapping table and performing validity screening."

[0046] When determining the corresponding one-dimensional SOC-voltage index interval based on the coordinate range, the index mapping relationship retained during the two-dimensional encoding construction stage is used to map the starting and ending coordinates of the salient region back to the original curve sampling point positions. For the rectangular region Ω=[i1,i2]×[j1,j2], min(i1,j1) is selected as the starting point of the one-dimensional interval, and max(i2,j2) is selected as the ending point of the one-dimensional interval, thus obtaining the one-dimensional SOC-voltage index interval corresponding to the salient region. Based on this method, the high-response discrimination region in the two-dimensional image can be accurately mapped back to the key discharge segment in the original discharge curve.

[0047] S40, constructing a two-stage semi-supervised comparative learning sorting and clustering mechanism.

[0048] The two-stage semi-supervised contrastive learning sorting and clustering mechanism includes a first training stage and a second training stage.

[0049] In an exemplary embodiment, step S40 may include the following steps: S401, preheating training is performed based on a preset number of labeled samples to establish a basic inter-class structure; S402, construct an initial class prototype based on the labeled samples, construct a class prototype based on the representation center of each class sample (determined by selectively incorporating high-confidence pseudo-label samples after updating based on the initial class prototype), and evaluate the similarity of unlabeled samples; S403, add high-confidence samples to the training set after assigning pseudo-labels, and enhance intra-class compactness and inter-class separation by comparing the learning objectives. S404 introduces a classification head, combines labeled samples and high-confidence pseudo-labeled samples for supervised training, and outputs the final sorting and clustering results of retired lithium-ion batteries based on semi-supervised learning.

[0050] Specifically, a high-confidence sample is a sample whose maximum class probability is higher than a preset confidence threshold, which is on the order of 0.8.

[0051] Introducing a classification head means that after obtaining the shared backbone representation in the first stage, the projection head used only for contrastive learning is removed or frozen, and the shared representation is directly connected to the linear classification layer, thereby forming a classification head structure for supervised training.

[0052] The two-stage semi-supervised contrastive learning process is built upon the unified feature representation of the backbone representation network. Specifically, firstly, a pre-trained VGG16 network extracts features from the input battery image representation and outputs a high-dimensional feature vector from the fully connected layers. Subsequently, this high-dimensional feature vector is input into the backbone representation network. The backbone representation network consists of linear layers, batch normalization layers, and ReLU activation layers, which compress and reconstruct the original high-dimensional features to obtain a low-dimensional compact representation that can be shared for subsequent tasks. This low-dimensional compact representation is then connected to the projection head and the classification head, forming the network structure corresponding to the two training stages. The projection head maps the shared representation to a contrastive learning space suitable for measuring sample similarity, while the classification head maps the shared representation to specific class probability outputs. Therefore, the two-stage semi-supervised contrastive learning process is not an additional module independent of the backbone representation network, but rather a two-step training process centered around the same shared representation.

[0053] The first stage is the semi-supervised contrastive representation learning stage, whose core objective is to learn a representation space with good discriminative ability. In this stage, a "backbone representation network + projector head" structure is used for training. Initially, a small number of labeled samples are used for warm-up training, allowing similar samples to gradually approach each other in the projection space and dissimilar samples to gradually separate, thus establishing an initial inter-class structure. Based on this, the augmentation training stage begins: first, based on the feature representations of labeled samples in the projection space, the average representation of each class is calculated and used as the class prototype for the corresponding class; then, unlabeled samples are input into the same network to obtain their feature representations in the projection space, and their similarity to each class prototype is calculated; then, the class probability distribution is obtained based on the similarity results. When the maximum class probability is higher than a preset confidence threshold, a pseudo-label is assigned to the unlabeled sample, and it is added to the training set for further training. According to the structure settings in the paper, this confidence threshold can be set to the order of 0.8. Through the above process, the first stage can gradually incorporate high-confidence unlabeled samples into the training, thereby enhancing intra-class compactness and inter-class separability, ultimately obtaining a more stable shared representation.

[0054] The second stage is the discriminative refinement training stage, whose core objective is to further establish clear and stable classification boundaries based on the representation space already formed in the first stage. In this stage, a "backbone representation network + classification head" structure is used for training. Specifically, the shared backbone representation obtained in the first stage is used as the input basis. Instead of focusing on calculating similarity between samples, the projection head, which only serves contrastive learning, is removed or frozen. A classification head is then introduced, and supervised training is conducted using labeled samples and high-confidence pseudo-labeled samples. At this point, the focus of the classification head's learning is no longer on the sample representation itself, but on how to further transform the representation space, which has been initially separated in the first stage, into clear category discrimination boundaries. Thus, the first stage is mainly responsible for "learning representations," while the second stage is mainly responsible for "learning classification," with both stages connected and co-optimized through the same backbone representation network.

[0055] The key to the above two-stage mechanism is that, on the one hand, a shared representation space with discriminative capabilities is first established by using a preset number of labeled samples, reducing the reliance on large-scale manually labeled data; on the other hand, through class prototype guidance and pseudo-label screening mechanisms, high-confidence unlabeled samples are gradually incorporated into the training, thereby improving the identification ability of minority class samples in the representation space and further improving the sorting and clustering effects under class imbalance conditions.

[0056] The semi-supervised learning-based sorting and clustering method for retired lithium-ion batteries provided in the above embodiments of this application effectively improves the sorting accuracy, consistency, and generalization ability of retired batteries under conditions of few labels and class imbalance by constructing an integrated technical link for retired lithium-ion battery discharge data processing, dual image encoding and pre-trained feature extraction, salient interval selection, and two-stage semi-supervised contrastive learning sorting and clustering. By introducing a pre-trained convolutional network, the cost of two-dimensional feature extraction is significantly reduced; the most discriminative discharge interval is accurately located through salient feature map reflection technology; and the representation ability of minority class samples is strengthened through a semi-supervised contrastive learning mechanism. This method achieves high efficiency, accuracy, and robustness in the sorting and clustering of retired lithium-ion batteries based on semi-supervised learning, providing an effective way to solve the pain points of existing methods.

[0057] In the description of this application, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0058] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

[0059] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A method for sorting and clustering retired lithium-ion batteries based on semi-supervised learning, characterized in that, The method includes the following steps: S10, Construct a discharge data processing flow for retired lithium-ion batteries; S20, Construct a dual image coding and pre-trained feature extraction framework; S30, Construct an interval selection mechanism based on salient feature map replay; S40, constructing a two-stage semi-supervised comparative learning sorting and clustering mechanism.

2. The method for sorting and clustering retired lithium-ion batteries based on semi-supervised learning according to claim 1, characterized in that, Step S10 includes the following steps: S101 performs capacity testing, pulse testing, and open-circuit voltage testing on retired lithium-ion batteries to obtain test results; S102, Extract the SOC-discharge voltage curve based on the test results; S103, the SOC-discharge voltage curve is denoised, interpolated and resampled, and its length is standardized to make different battery samples comparable under the same data scale.

3. The method according to claim 1, characterized in that, Step S20 includes: S201, Perform one-dimensional curve preprocessing on the SOC-discharge voltage curve, and convert the SOC-discharge voltage curve into GASF image and RP image respectively; S202, perform image encoding on the GASF image and RP image to obtain two types of encoding results; S203, the two types of encoding results are combined to form a pseudo-color image adapted to the input of the convolutional neural network, and the pseudo-color image is input into a pre-trained VGG16 network for hierarchical feature extraction.

4. The method for sorting and clustering retired lithium-ion batteries based on semi-supervised learning according to claim 3, characterized in that, Step S201 includes the following steps: S2011, normalizes the SOC-discharge voltage curve after noise reduction, interpolation resampling and length unification. S2012, the normalized SOC-discharge voltage curve is mapped to polar coordinate space, and a GASF image is constructed in combination with the corresponding discharge voltage to express the correlation of the sequence over the global time span. S2013, perform linear interpolation and first-order difference processing on the mapped SOC-discharge voltage curve, and construct an RP image by the Manhattan distance between the difference values ​​at any two sampling times to characterize the local variation relationship and state recurrence characteristics during the discharge process.

5. The method for sorting and clustering retired lithium-ion batteries based on semi-supervised learning according to claim 3, characterized in that, Step S202 includes the following steps: S2021, normalize the SOC sequence to a preset range, and convert the normalized SOC value into an angle variable, with the discharge voltage as the polar coordinate radius component; S2022 constructs a GASF matrix based on the cosine of the sum of the angle components at any two sampling times, thereby transforming the global correlation between different time positions in a one-dimensional time series curve into a spatial texture distribution in a two-dimensional image.

6. The method according to claim 1, characterized in that, Step S30 includes: S301, extract feature maps from the intermediate convolutional layers of the pre-trained VGG16 network, and aggregate the multi-channel activation responses to generate a two-dimensional saliency map reflecting the discrimination strength. S302 calculates the difference heatmap between salient feature maps for multiple battery samples, and locates the spatial region with the strongest class discrimination by cross-sample difference enhancement. S303, the spatial region with the strongest category differentiation is reverse-mapped to the original discharge curve to obtain the optimal discharge range with the strongest discriminative power.

7. The method according to claim 1, characterized in that, Step S40 includes: S401, preheating training is performed based on a preset number of labeled samples to establish a basic inter-class structure; S402, construct an initial class prototype based on the labeled samples, construct class prototypes based on the representation centers of each class of samples, and evaluate the similarity of unlabeled samples; S403, add high-confidence samples to the training set after assigning pseudo-labels, and enhance intra-class compactness and inter-class separation by comparing the learning objectives. S404 introduces a classification head, combines labeled samples and high-confidence pseudo-labeled samples for supervised training, and outputs the final sorting and clustering results of retired lithium-ion batteries based on semi-supervised learning.

Images