Ex-service power battery sorting method and device based on multi-modal information fusion

Abstract

The invention discloses an out-of-service power battery sorting method and device based on multi-modal information fusion, and relates to the technical field of battery recycling. Multi-modal input is formed based on one-dimensional structure features, a two-dimensional GASF image and a two-dimensional GADF image, and a multi-modal sorting model is constructed to realize modal deep feature extraction and interactive enhancement processing for classification prediction. Wherein the one-dimensional structure feature is extracted based on a partial discharge curve, full-coverage type charging and discharging are avoided, the time consumption of a charging and discharging experiment is shortened, and the data volume of model processing is reduced; the two-dimensional GASF image and the two-dimensional GADF image are obtained based on medium and high frequency band EIS data conversion, and the processing efficiency is considered under the condition that the integrity of key information is ensured. Compared with an existing sorting scheme which only depends on a single mode or is only based on capacity, the internal and external states of the batteries can be comprehensively reflected, and the consistency and accuracy of battery sorting are improved.

Description

Technical Field

[0001] This invention relates to the field of battery recycling technology, specifically to: 1. a method for sorting retired power batteries based on multimodal information fusion; 2. a sorting device for retired power batteries based on multimodal information fusion. Background Technology

[0002] Utilizing retired power batteries in a tiered manner can fully tap into their value throughout their entire life cycle and promote their sustainable development. To ensure the consistency and safety of battery utilization, accurate sorting of individual cells is essential during the initial recycling stage.

[0003] Currently, the widely used approach is to sort batteries based on their external characteristic parameters using threshold methods or simple classification models. For example, indicators such as voltage, capacity, constant current charge-discharge curves, and DC internal resistance are used to evaluate and classify batteries. This approach relies solely on external characteristics, and the raw data is one-dimensional time-series data with limited expressive power, resulting in inherent limitations. While subsequent attempts have focused on converting the raw data into two-dimensional image features, enhancing feature expression and amplifying differences between data points while preserving the original sequence relationships, these methods still perform image processing on external characteristic parameters, representing a deep dive into a single modal characteristic. Although the feature expression forms are becoming more diversified, the inherent limitations remain. Furthermore, both methods are overly dependent on the completeness of the raw data, generally requiring a full charge-discharge test, leading to a time-consuming and lengthy testing process.

[0004] Although a few cutting-edge studies have used classification models to fuse multi-source information for sorting, they have the following problems: 1. The selection of multimodal information is unreasonable, leading to redundant data calculations; 2. Existing multimodal fusion strategies only stop at simple data layer splicing or decision layer voting / averaging - this shallow fusion fails to design effective feature interaction and fusion mechanisms, lacks the synergistic utilization of different modal information, resulting in poor performance.

[0005] Furthermore, both the single-modal and multi-modal methods mentioned above generally assume that all batteries to be sorted belong to categories known in the training set. Therefore, when faced with abnormal batteries not covered by the training data, the model is prone to forcibly classifying them into existing groups, which poses a potential safety risk. Summary of the Invention

[0006] Therefore, it is necessary to provide a sorting method and device for retired power batteries based on multimodal information fusion, addressing the problems of redundant data calculation and failure to achieve effective feature interaction and fusion of different modal information in existing multi-source information fusion methods.

[0007] This invention is achieved using the following technical solution: In a first aspect, this invention discloses a method for sorting retired power batteries based on multimodal information fusion, comprising: Step 1: Perform a non-full-coverage discharge experiment on the target battery to obtain a partial discharge curve, and extract one-dimensional structural features based on the partial discharge curve. f s Among them, the voltage starting point of part of the discharge curve is a preset voltage threshold. V start The voltage endpoint is the discharge cutoff voltage. V end ; Step two: Apply an excitation current to the target battery that has been processed in step one and has SOC=0, and acquire high-frequency EIS data. Then, convert the data to obtain the corresponding two-dimensional GASF image. P GASF and 2D GADF images P GADF The frequency band of the mid-to-high frequency EIS data is from 1Hz to 10kHz. Step 3, f s , P GASF , P GADF The input is processed by a trained multimodal sorting model to obtain the fused classification result. Res The multimodal sorting model includes: a structure processing branch, an image processing branch, a bidirectional cross-attention layer, and a main classification head; the structure processing branch is used to sort data from... f s Extract structural feature vectors F s Image processing branch is used to combine P GASF , P GADF Extracting image structural features F i The bidirectional cross-attention layer is used to implement a bidirectional cross-attention mechanism based on mutual information attention modulation. F s , F i Enhance the interaction and merge it into a fusion feature. F fused The main classification header is used for classification based on... F fused Perform classification prediction to obtain Res .

[0008] This method for sorting decommissioned power batteries based on multimodal information fusion implements the method or process according to embodiments of this disclosure.

[0009] Secondly, the present invention discloses a retired power battery sorting device based on multimodal information fusion, which uses the retired power battery sorting method based on multimodal information fusion of the first aspect.

[0010] The decommissioned power battery sorting device based on multimodal information fusion includes: a one-dimensional feature acquisition module, a two-dimensional feature acquisition module, and a multimodal sorting module.

[0011] The one-dimensional feature acquisition module is used to: perform a non-full-coverage discharge experiment on the target battery to obtain a partial discharge curve, and extract one-dimensional structural features based on the partial discharge curve. f s .

[0012] The two-dimensional feature acquisition module is used to: apply an excitation current to the target battery with SOC=0 after processing by the one-dimensional feature acquisition module, acquire high-frequency EIS data, and then obtain the corresponding two-dimensional GASF image through data conversion. P GASF and 2D GADF images P GADF .

[0013] The multimodal sorting module is used to: f s , P GASF , P GADF The input is processed by a trained multimodal sorting model to obtain the fused classification result. Res .

[0014] The decommissioned power battery sorting device based on multimodal information fusion implements the method or process according to embodiments of this disclosure.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses one-dimensional structural features, two-dimensional GASF images, and two-dimensional GADF images as multimodal inputs to construct a multimodal sorting model. This model performs deep feature extraction and interactive enhancement processing of the various modes for classification and prediction. Compared to existing sorting schemes that rely solely on a single mode or are based solely on capacity, this invention comprehensively reflects the internal and external states of the battery, improving the consistency and accuracy of battery sorting.

[0016] 2. The one-dimensional structural features of the present invention are extracted based on partial discharge curves, avoiding full-coverage charging and discharging, which not only shortens the charging and discharging experiment time, but also reduces the amount of data processed by the model; the two-dimensional GASF image and two-dimensional GADF image of the present invention are obtained based on mid-to-high frequency band EIS data conversion, which takes into account the processing efficiency while ensuring the integrity of key information.

[0017] 3. The multimodal sorting model of the present invention utilizes a dual-branch network to process one-dimensional structural features and two-dimensional images respectively. The model uses the image processing branch to convert the one-dimensional signal into a two-dimensional image, enhancing feature expression while preserving the frequency domain structure, and effectively amplifying the subtle differences in EIS data. The model also achieves intermodal interaction through a bidirectional cross-attention mechanism and integrates attention modulation based on mutual information, so that the parts with closer connections and stronger correlations between feature dimensions receive higher weights in the fusion, improving the fusion effect while avoiding information redundancy.

[0018] 4. This invention designs a joint decision-making system based on confidence level and modal consistency, which can effectively identify target batteries that may belong to new types or abnormal states, enhancing the practicality and safety of its application in real industrial scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method for sorting retired power batteries based on multimodal information fusion provided in Embodiment 1 of the present invention; Figure 2 for Figure 1 Structure diagram of the multimodal sorting model; Figure 3 for Figure 1 A structural diagram for processing branches in a mid-structure; Figure 4 for Figure 3 Two structural diagrams of the hidden layer; Figure 5 for Figure 1 Another structural diagram for processing branches in the middle structure; Figure 6 for Figure 1 Structure diagram of the image processing branch in Chinese; Figure 7 for Figure 1 Structure diagram of the bidirectional cross-attention layer and the main classification head. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] Example 1 Please see Figure 1 , Figure 1 This is a flowchart of the decommissioned power battery sorting method based on multimodal information fusion in this invention, which includes the following steps: Step 1: Perform a non-full-coverage discharge experiment on the target battery to obtain a partial discharge curve, and extract one-dimensional structural features based on the partial discharge curve. f s .

[0025] The target battery, as the battery to be sorted, has an uncertain initial remaining capacity. Therefore, to ensure data integrity and consistency, the following settings are made: If the initial residual voltage of the target battery is less than the preset voltage threshold V start First, the target battery is charged under constant current and constant voltage until its real-time remaining voltage reaches the specified value. V start Then, it is subjected to constant current discharge until its real-time residual voltage drops to the discharge cutoff voltage. V end and extract from V start arrive V end Partial discharge curve; Otherwise, directly discharge the target battery at a constant current until its real-time residual voltage drops to [a value missing]. V end and extract from V start arrive V end Partial discharge curve.

[0026] It should be noted that, V start The voltage is below the upper limit of the charging limit, allowing for an effective discharge segment to be obtained without performing a full charging process. This not only shortens the discharge experiment time but also reduces the amount of data processed by the subsequent model. V start , V stop 、V end The value is set according to the actual type and specifications of the target battery.

[0027] In other words, the voltage starting point of the partial discharge curve must be at a preset voltage threshold. V start The voltage endpoint is the discharge cutoff voltage. V end This ensures that effective one-dimensional structural features can be extracted. f s .

[0028] The partial discharge curve is actually the target battery from V start Down to V end Corresponding voltage-time series data Data V→T To start from Data V→T Extract f s The specific process is as follows: First of all Data V→T Preprocessing (e.g., linear interpolation) is performed to standardize the data length, resulting in corrected sequence data. N_Data V→T ; Then from N_Data V→T Extensive extraction f s Specifically, including but not limited to: ① Time-domain statistical characteristics, for example: N_Data V→T The voltage average value, voltage variance, voltage skewness, voltage kurtosis, etc.

[0029] ② Local differential characteristics, such as dV / dt, dQ / dV, etc.; V represents the real-time remaining voltage, t represents time, and Q represents the real-time remaining charge.

[0030] ③ Time characteristics, such as the discharge duration of a partial discharge curve. T part ; T part = t end -t start ; t end This indicates that the real-time residual voltage has dropped to the discharge cutoff voltage. V end The corresponding moment; t start Indicates from V start The moment when discharge begins.

[0031] ④ Capacity characteristics, such as the amount of charge released in the partial discharge curve. Q part ; Q part = I × T part ; I This represents the discharge current.

[0032] ⑤ Voltage variation characteristics at fixed time intervals—that is, at a fixed time interval Δ in the partial discharge curve t The corresponding voltage change Δ V It characterizes the rate of voltage drop and stability of the battery.

[0033] ⑥. Time variation characteristics of fixed voltage interval – i.e., the fixed voltage interval Δ decreases in the partial discharge curve. u Required duration Δ τ This reflects the kinetic differences of the battery at different voltage ranges.

[0034] Step 2: Apply an excitation current (sinusoidal excitation current is recommended) to the target battery that has been processed in Step 1 and has a SOC (State of Charge) of 0, and acquire high-frequency EIS data. Then, convert the data to obtain the corresponding two-dimensional GASF image. P GASF and 2D GADF images P GADF .

[0035] Step two did not select full-band EIS data. Instead, taking into account both testing efficiency and the integrity of key information, an optimized frequency band acquisition strategy was adopted—collecting mid-to-high frequency EIS data from 1Hz to 10kHz. This significantly reduced the acquisition time from nearly ten minutes for the full frequency band to less than one minute.

[0036] After obtaining the mid-to-high frequency band EIS data, the data conversion process treats it as a one-dimensional time series and encodes it into a two-dimensional image using Gram angle field encoding. P GASF , P GADF .

[0037] To facilitate further understanding, the specific process of data transformation is as follows: First, the mid-to-high frequency EIS data is represented as a one-dimensional time series. Z ={ z 1, z 2,…, z K}, z k Indicates the first k EIS data at each frequency point k ∈[1, K ].

[0038] Then, to Z Normalization is performed to obtain the normalized time series. , Represents the normalized i-th k EIS data at several frequency points; among them min( Z )express Z The minimum value; max( Z )express Z The maximum value.

[0039] Next, Mapped to polar coordinates; where, Corresponding angle It can be represented as: .

[0040] Subsequently, the Gramian Angular Summation Field (GASF) matrix is ​​generated by performing cosine operations on the sum of angles and sine operations on the difference of angles, respectively. Gramian Angular Difference Field (GADF) matrix .in, , They are represented as follows: ; .

[0041] Finally, The matrix elements are mapped one-to-one with the image pixels to obtain P GASF ;Will The matrix elements are mapped one-to-one with the image pixels to obtain P GADF In other words, P GASF , P GADF All resolutions are K × K Two-dimensional image.

[0042] It should be noted that this invention introduces two types of two-dimensional images based on mid-to-high frequency EIS data conversion, which can ensure the integrity of key information while taking into account processing efficiency.

[0043] Step 3, f s , P GASF , P GADF The input is processed by a trained multimodal sorting model to obtain the fused classification result. Res .

[0044] It should be noted that the sorting method for retired power batteries in this paper uses a pre-trained multimodal sorting model. The model is trained on a sample dataset with real labels, and a cross-entropy loss function is constructed. The Adam optimizer is used for end-to-end training until the training is completed.

[0045] It should be noted that the sample dataset can be a known sample set or a self-constructed labeled sample set.

[0046] I. The following section will introduce the construction of the labeled sample set: S1. Randomly select M retired batteries of the same model as the target battery as M samples to be labeled. The value of M should not be too small, generally not less than 500.

[0047] S2, conduct a full-coverage charge-discharge experiment on each sample to be labeled to obtain its true battery capacity value.

[0048] The full-coverage charge-discharge experiment includes: first, the sample to be labeled is charged with constant current and constant voltage until its real-time remaining voltage reaches the upper limit of the charging voltage. V stop The real-time residual current reaches the charging cutoff current.I stop Then perform constant current discharge until the real-time residual voltage drops to V end .

[0049] S3. Under the condition of SOC=0, apply an excitation current (sinusoidal excitation current is recommended) to each sample to be labeled to collect its full-band EIS data (i.e., frequency range: from 0.01Hz to 10kHz), and fit the full-band EIS data through the equivalent circuit model to obtain its ohmic internal resistance and charge transfer internal resistance.

[0050] Specifically, a typical circuit model (e.g., R(QR)(QR) model; where R represents resistance and Q represents a constant phase angle element) containing elements such as ohmic internal resistance, charge transfer process, and diffusion process can be used for fitting. The error between the measured value and the fitted value can be minimized by optimization algorithms such as the least squares method, thereby identifying the ohmic internal resistance characterizing the ohmic losses of electrolyte and current collector, and the charge transfer internal resistance characterizing the reaction kinetics of the electrode / electrolyte interface.

[0051] S4. The actual battery capacity, ohmic internal resistance, and charge transfer internal resistance of the M samples to be labeled are clustered using a clustering algorithm (such as K-means clustering, DBSCAN clustering, or hierarchical clustering) to obtain the category label of each sample to be labeled, thereby forming a labeled sample set.

[0052] It should be noted that S4 will produce several clustering results, which represent the corresponding sorting categories. These category labels can then be used as supervision signals for model training.

[0053] In addition, common internal metrics such as the Silhouette Coefficient (SC), Davies-Bouldin Index (DBI), Calinski-Harabasz Index (CHI), and Dunn Index (DI) can be used to evaluate the clustering results and verify their rationality.

[0054] II. See Figure 2 The structure of the multimodal sorting model is explained below. The multimodal sorting model can be divided into the following functional branches: structure processing branch, image processing branch, bidirectional cross-attention layer, and main classification head.

[0055] The following is a description of each part of the model: ① The structure processing branch is used to process branches from fs Extract structural feature vectors F s .

[0056] See Figure 3 The structural processing branch can employ a lightweight design based on a multilayer perceptron, specifically including: a multilayer perceptron. For example... Figure 3 As shown, the multilayer perceptron includes: N hidden layers and 1 fully connected layer. The output of the i-th hidden layer serves as the input of the (i+1)-th hidden layer; i∈[1,N-1]; the output of the N-th hidden layer serves as the input of the fully connected layer; the fully connected layer is used to perform feature mapping on the output of the N-th hidden layer to obtain... F s —That is, we obtain a high-dimensional abstract feature vector.

[0057] It should be noted that hidden layers can be used Figure 4 The design of region (a) includes: a fully connected layer, a batch normalization layer, and a ReLU activation layer connected in sequence. That is, it corresponds to: performing a linear transformation, batch normalization, and introducing nonlinearity through the ReLU activation function on the input IN of the hidden layer in sequence to obtain the output OUT of the hidden layer.

[0058] Of course, hidden layers can also be used to promote gradient flow and increase network depth. Figure 4 The design of the middle (b) area, in Figure 4 Based on region (a), a skip connection is added—IN is first superimposed with the output of the batch normalization layer, and then processed by the ReLU activation layer.

[0059] See Figure 5 The structural processing branch can also be designed as a network structure for multi-scale extraction, which specifically includes: a preliminary processing unit, a low-scale processing unit, a medium-scale processing unit, a high-scale processing unit, and a splicing layer.

[0060] like Figure 5 As shown, the preliminary processing unit includes: a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer connected in sequence, corresponding to: for f s The process involves sequentially performing one-dimensional convolution, batch normalization, ReLU activation, and max pooling to obtain preliminary structural features. f F .

[0061] The low-scale processing unit includes: a low-scale convolutional layer, a batch normalization layer, a ReLU activation layer, another low-scale convolutional layer, a layer normalization layer, a residual connection, a ReLU activation layer, and an average pooling layer connected in sequence. This corresponds to: first processing... f FThe process involves sequentially performing low-scale convolution, batch normalization, ReLU activation, low-scale convolution, and layer normalization, followed by... f F The layers are then stacked, followed by ReLU activation and average pooling to obtain low-scale structural features. f L Specifically, the kernel size of the low-scale convolutional layer is set to 3×1.

[0062] The mesoscale processing unit comprises: a mesoscale convolutional layer, a batch normalization layer, a ReLU activation layer, another mesoscale convolutional layer, a layer normalization layer, a residual connection, a ReLU activation layer, and an average pooling layer, connected sequentially. This corresponds to: first processing... f F The process involves sequentially performing mesoscale convolution, batch normalization, ReLU activation, mesoscale convolution, and layer normalization, followed by... f F The layers are then stacked, followed by ReLU activation and average pooling to obtain mesoscale structural features. f M Specifically, the kernel size of the mesoscale convolutional layer is set to 5×1.

[0063] The high-scale processing unit includes: a high-scale convolutional layer, a batch normalization layer, a ReLU activation layer, another high-scale convolutional layer, a layer normalization layer, a residual connection, a ReLU activation layer, and an average pooling layer connected in sequence. This corresponds to: first processing... f F The process involves sequentially performing high-scale convolution, batch normalization, ReLU activation, high-scale convolution, and layer normalization, followed by... f F The layers are then stacked, followed by ReLU activation and average pooling to obtain high-scale structural features. f H Specifically, the kernel size of the high-scale convolutional layer is set to 7×1.

[0064] The splicing layer is used to: f L , f M , f H To obtain by splicing F s —That is, we obtain a high-dimensional abstract feature vector.

[0065] ② The image processing branch is used to combine P GASF , P GADF Extracting image structural featuresF i .

[0066] In other words, the image processing branch is built upon convolutional neural networks to extract deep features from images. See also Figure 6 The image processing branch is designed to include: one stitching layer and one ResNet50 network layer. The stitching layer is used to: stitch... P GASF , P GADF The concatenation process is performed; the ResNet50 network layer is a ResNet50 model with the last fully connected layer removed, which is used to: process the output of the concatenation layer to obtain... F i —That is, we obtain another high-dimensional abstract feature vector.

[0067] It should be noted that ResNet50 network layers are recommended to be truncated from ResNet-50 models pre-trained on large image datasets, as this can improve the overall training performance of the model.

[0068] ③ A bidirectional cross-attention layer is used to implement a bidirectional cross-attention mechanism based on mutual information attention modulation. F s , F i Enhance the interaction and merge it into a fusion feature. F fused .

[0069] In other words, the bidirectional cross-attention layer realizes the interaction between modalities through the bidirectional cross-attention mechanism and integrates attention modulation based on mutual information, so that the parts with closer connections and stronger correlations between feature dimensions get higher weights in the fusion, improving the fusion effect while avoiding information redundancy.

[0070] See Figure 7 The bidirectional cross-attention layer consists of: one linear mapping layer, two cross-attention layers, and one interaction processing layer.

[0071] Linear mapping layers are used to: F s Linear mapping to query vector Q s Key vector K s Value vector V s ,Will F i Linear mapping to query vector Q i Key vector K i Value vectorV i .

[0072] Of course, the processing of the linear mapping layer can also be expressed by the formula: ; ; In the formula, This represents six learnable transformation matrices.

[0073] The first cross-attention layer is used for: based on Q s , K i Calculate the initial attention weights λ s The second cross-attention layer is used for: based on Q i , K s Calculate the initial attention weights λ i .

[0074] Specifically, λ s , λ i The calculation formula is as follows: ; ; In the formula, d s express Q s The dimension; d i express Q i Dimensions.

[0075] The interaction processing layer is used to: λ s With prior mutual information matrix M Multiply to obtain enhanced attention weights γ s and further with V i Multiply to obtain enhanced features F i→s ,Will λ i and M Multiply to obtain enhanced attention weights γ i and further with V s Multiply to obtain enhanced features Fs→i Then F i→s , F s→i Add to obtain F fused .

[0076] It is important to note that M It is obtained from model training calculations and satisfies: .in, D s and D i These represent the dimensions of structural features and image features, respectively. M Element M ij The information dependence between the i-th structural feature and the j-th image feature in the entire training set was quantified.

[0077] Therefore, first by utilizing M The initial attention weights are element-wise weighted and modulated so that the cross-modal feature dimension combinations that are statistically more closely related receive more attention during the attention allocation process. The value vector is then weighted and summed using the modulated enhanced attention weights to generate enhanced features that have been enhanced by information from another modality. This process is carried out bidirectionally between the two modalities, thus effectively achieving interactive enhancement.

[0078] ④ The main category header is used for... F fused Perform classification prediction to obtain Res .

[0079] See Figure 7 The main classification head consists of one linear layer and one softmax layer. The linear layer is used for: F fused Perform linear transformations; the Softmax layer is used to: process the output of the linear layer through the Softmax activation function to obtain... Res .

[0080] It should be noted that, Res To predict the probability distribution vector, which represents the probability distribution of the target battery belonging to each sorting category—generally, the category with the highest probability value can be taken as the category to which the target battery belongs. However, considering that training with a training set in the background technology would lead to the problem of forced classification during actual deployment, this retired power battery sorting method can also add the following step: Step 4, based on Res Calculate classification confidence based on F s , F i Calculate similarity and determine the classification of the target battery based on preset decisions.

[0081] Specifically, it can be based on Res The classification confidence score can be calculated using the maximum probability value, and cosine similarity can be used for quantification. F s , F i The similarity.

[0082] In other words, the pre-defined decision is a joint judgment decision based on confidence level and modality consistency, specifically: like Res The confidence level reaches the confidence threshold T1, and F s and F i If the similarity reaches the consistency threshold T2, then... Res The class of the target battery is determined by the highest predicted probability value; otherwise, the target battery is classified into the unknown class—because if... Res If the confidence level is lower than T1, it indicates that the target battery lacks sufficient confidence; if F s and F i If the similarity between the two is lower than that between T2, it indicates that the predicted categories may be inconsistent.

[0083] The value of T1 is limited to 0.6~0.95, and the value of T2 is limited to 0.5~0.9. The specific values ​​may be adjusted according to the actual situation.

[0084] Of course, for target batteries classified as unknown, a manual review or secondary confirmation process can be introduced to determine their classification.

[0085] Simulation verification First, the above-mentioned method for sorting retired power batteries was used as a baseline method, and two single-mode processing methods were compared (1, input only). f s 2. Input only P GASF and P GADF The results are shown in Table 1 below.

[0086] Table 1 Performance Comparison of the Three Methods method accuracy Accuracy Recall rate F1 score Baseline method 0.94 0.95 0.95 0.95 <![CDATA[Only input P GASF and P GADF > 0.81 0.80 0.80 0.79 <![CDATA[Only input f s > 0.86 0.88 0.88 0.88 As shown in Table 1, compared with the two single-mode processing methods, the baseline method has significant improvements in all four indicators, verifying its rationality and superiority.

[0087] Example 2 This embodiment 2 provides a retired power battery sorting device based on multimodal information fusion, which uses the retired power battery sorting method based on multimodal information fusion from the first aspect.

[0088] The decommissioned power battery sorting device based on multimodal information fusion includes: a one-dimensional feature acquisition module, a two-dimensional feature acquisition module, and a multimodal sorting module.

[0089] The one-dimensional feature acquisition module is used to: perform a non-full-coverage discharge experiment on the target battery to obtain a partial discharge curve, and extract one-dimensional structural features based on the partial discharge curve. f s .

[0090] The two-dimensional feature acquisition module is used to: apply an excitation current to the target battery with SOC=0 after processing by the one-dimensional feature acquisition module, acquire high-frequency EIS data, and then obtain the corresponding two-dimensional GASF image through data conversion. P GASF and 2D GADF images P GADF .

[0091] The multimodal sorting module is used to: f s , P GASF , P GADF The input is processed by a trained multimodal sorting model to obtain the fused classification result. Res .

[0092] Since this device uses the multimodal information fusion-based sorting method for retired power batteries in Example 1, it has the same effect, so it will not be repeated here.

[0093] Example 3 This embodiment 3 discloses a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the decommissioned power battery sorting method based on multimodal information fusion disclosed in embodiment 1.

[0094] The computer equipment can be either a mobile terminal or a fixed terminal. Examples of the former include mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (such as in-vehicle navigation terminals); examples of the latter include digital TVs and desktop computers.

[0095] This embodiment 3 also discloses a readable storage medium that stores computer program instructions. When the computer program instructions are read and run by a processor, the steps of the decommissioned power battery sorting method based on multimodal information fusion disclosed in embodiment 1 are executed.

[0096] The readable storage medium may include, but is not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

[0097] This embodiment 3 also discloses a computer program product, including a computer program. When executed by a processor, the computer program implements the steps of the decommissioned power battery sorting method based on multimodal information fusion disclosed in embodiment 1.

[0098] It should be noted that the computer program used to execute the above can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as C or similar languages. The computer program can be executed entirely on the user's computer, partially on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer through any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN).

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for sorting decommissioned power batteries based on multimodal information fusion, characterized in that, It includes: Step 1: Perform a non-full-coverage discharge experiment on the target battery to obtain a partial discharge curve, and extract one-dimensional structural features based on the partial discharge curve. f s Among them, the voltage starting point of part of the discharge curve is a preset voltage threshold. V start The voltage endpoint is the discharge cutoff voltage. V end ; Step two: Apply an excitation current to the target battery that has been processed in step one and has SOC=0, and acquire high-frequency EIS data. Then, convert the data to obtain the corresponding two-dimensional GASF image. P GASF and 2D GADF images P GADF The frequency band of the mid-to-high frequency EIS data is from 1Hz to 10kHz. Step 3, f s , P GASF , P GADF The input is processed by a trained multimodal sorting model to obtain the fused classification result. Res The multimodal sorting model includes: a structure processing branch, an image processing branch, a bidirectional cross-attention layer, and a main classification head; the structure processing branch is used to sort data from... f s Extract structural feature vectors F s Image processing branch is used to combine P GASF , P GADF Extracting image structural features F i The bidirectional cross-attention layer is used to implement a bidirectional cross-attention mechanism based on mutual information attention modulation. F s , F i Enhance the interaction and merge it into a fusion feature. F fused The main classification header is used for classification based on... F fused Perform classification prediction to obtain Res .

2. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, In step one, if the initial remaining voltage of the target battery is less than... V start First, the target battery is charged under constant current and constant voltage until its real-time remaining voltage reaches the specified value. V start Then, it is subjected to constant current discharge until its real-time residual voltage drops to V end and extract from V start arrive V end Partial discharge curve; Otherwise, directly discharge the target battery at a constant current until its real-time residual voltage drops to [a value missing]. V end and extract from V start arrive V end Partial discharge curve; Or / and, in step one, f s This includes: time-domain statistical characteristics, local differential characteristics, time characteristics, capacity characteristics, voltage variation characteristics at fixed time intervals, and time variation characteristics at fixed voltage intervals.

3. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, In step two, the data transformation methods include: Mid-to-high frequency EIS data is treated as a one-dimensional time series and encoded into a two-dimensional image using Gram angle field encoding to obtain... P GASF , P GADF .

4. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, The structural processing branch includes: multilayer perceptron; A multilayer perceptron consists of N hidden layers and one fully connected layer; the output of the i-th hidden layer serves as the input of the (i+1)-th hidden layer; i ∈ [1, N-1]; the output of the N-th hidden layer serves as the input of the fully connected layer; the fully connected layer is used to perform feature mapping on the output of the N-th hidden layer to obtain... F s ; Alternatively, the structural processing branches include: preliminary processing section, low-scale processing section, mid-scale processing section, high-scale processing section, and splicing layer; the preliminary processing section is used for: […]. f s The process involves sequentially performing one-dimensional convolution, batch normalization, ReLU activation, and max pooling to obtain preliminary structural features. f F The low-scale processing unit is used for: first processing... f F The process involves sequentially performing low-scale convolution, batch normalization, ReLU activation, low-scale convolution, and layer normalization, followed by... f F The layers are then stacked, followed by ReLU activation and average pooling to obtain low-scale structural features. f L The mesoscale processing unit is used to: first process... f F The process involves sequentially performing mesoscale convolution, batch normalization, ReLU activation, mesoscale convolution, and layer normalization, followed by... f F The layers are then stacked, followed by ReLU activation and average pooling to obtain mesoscale structural features. f M The high-scale processing unit is used for: first processing... f F The process involves sequentially performing high-scale convolution, batch normalization, ReLU activation, high-scale convolution, and layer normalization, followed by... f F The layers are then stacked, followed by ReLU activation and average pooling to obtain high-scale structural features. f H The splicing layer is used to: f L , f M , f H To obtain by splicing F s .

5. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, The image processing branch includes: one stitching layer and one ResNet50 network layer; The splicing layer is used to: P GASF , P GADF The splicing process is performed; the ResNet50 network layer is a ResNet50 model with the last fully connected layer removed, which is used to: process the output of the splicing layer to obtain... F i .

6. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, The bidirectional cross-attention layer consists of: one linear mapping layer, two cross-attention layers, and one interaction processing layer; Linear mapping layers are used to: F s Linear mapping to query vector Q s Key vector K s Value vector V s ,Will F i Linear mapping to query vector Q i Key vector K i Value vector V i The first cross-attention layer is used for: based on Q s , K i Calculate the initial attention weights λ s The second cross-attention layer is used for: based on Q i , K s Calculate the initial attention weights λ i ; The interaction processing layer is used to: λ s With prior mutual information matrix M Multiply to obtain enhanced attention weights γ s and further with V i Multiply to obtain enhanced features F i→s ,Will λ i and M Multiply to obtain enhanced attention weights γ i and further with V s Multiply to obtain enhanced features F s→i Then F i→s , F s→i Add to obtain F fused .

7. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, The main classification head consists of: one linear layer and one softmax layer; the linear layer is used for: F fused Perform linear transformations; the Softmax layer is used to: process the output of the linear layer through the Softmax activation function to obtain... Res .

8. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, The method for sorting retired power batteries also includes: Step four, based on... Res Calculate classification confidence based on F s , F i Calculate similarity and determine the classification of the target battery based on preset decisions; Among them, the pre-set decisions include: like Res The confidence level reaches the confidence threshold T1, and F s and F i If the similarity reaches the consistency threshold T2, then... Res The class with the highest predicted probability value is assigned to the target battery; otherwise, the target battery is assigned to the unknown class.

9. The method for sorting decommissioned power batteries based on multimodal information fusion according to claim 1, characterized in that, The multimodal sorting model is trained based on a labeled sample set; The methods for constructing the labeled sample set include: M retired batteries of the same model as the target battery were randomly selected as M samples to be labeled; A full-coverage charge-discharge experiment was conducted on each sample to be labeled to obtain its true battery capacity. The full-coverage charge-discharge experiment included: first, the sample to be labeled was charged with constant current and constant voltage until its real-time remaining voltage reached the upper limit of the charging voltage. V stop The real-time residual current reaches the charging cutoff current. I stop Then perform constant current discharge until the real-time residual voltage drops to V end ; Under the condition of SOC=0, an excitation current is applied to each sample to be labeled to collect its full-band EIS data, and the full-band EIS data is fitted by an equivalent circuit model to obtain its ohmic internal resistance and charge transfer internal resistance. The actual battery capacity, ohmic internal resistance, and charge transfer internal resistance of M unlabeled samples are clustered using a clustering algorithm to obtain the category label for each unlabeled sample, thus forming a labeled sample set.

10. A sorting device for decommissioned power batteries based on multimodal information fusion, characterized in that, It uses the multimodal information fusion-based sorting method for retired power batteries as described in any one of claims 1-9; The decommissioned power battery sorting device based on multimodal information fusion includes: The one-dimensional feature acquisition module is used to: perform a non-full-coverage discharge experiment on the target battery to obtain a partial discharge curve, and extract one-dimensional structural features based on the partial discharge curve. f s ; The two-dimensional feature acquisition module is used to: apply an excitation current to the target battery with SOC=0 after processing by the one-dimensional feature acquisition module, acquire high-frequency EIS data, and then obtain the corresponding two-dimensional GASF image through data conversion. P GASF and 2D GADF images P GADF ;as well as The multimodal sorting module is used to: f s , P GASF , P GADF The input is processed by a trained multimodal sorting model to obtain the fused classification result. Res .

Images