A liquid metal battery sorting index prediction method, system and device based on timing characteristics and deep learning and a storage medium

Abstract

This invention discloses a method, system, device, and storage medium for predicting sorting indicators of liquid metal batteries based on temporal features and deep learning. The method includes: collecting discharge curve data of liquid metal batteries during the activation phase to construct an original input sequence; dividing the discharge curve into several two-dimensional temporal segments according to a preset window length and step size, and using each segment as an independent input to a deep learning model; the model extracts local morphological features through one-dimensional convolution and pooling layers, and uses long short-term memory units to model temporal dependencies, outputting a final state vector; inputting the final state vector into a fully connected layer to obtain the capacity prediction value of the corresponding segment, and performing a cycle-based voting ensemble on the prediction results of all segments of the same battery to obtain a cycle-level capacity prediction value; and making classification and sorting decisions based on the comparison results of the cycle-level capacity prediction value and a preset threshold. This method enables automatic feature extraction and capacity prediction, improving battery sorting efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of liquid metal battery applications, and in particular to a method, system, device, and storage medium for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning. Background Technology

[0002] Currently, liquid metal batteries show promising application prospects in large-scale energy storage, with their high power density, long cycle life, and low manufacturing cost making them an important technological direction for replacing lithium-ion batteries. However, liquid metal batteries exhibit significant activation period fluctuations during production and operation, with parameters such as voltage, capacity, and internal resistance dynamically changing over time, making it difficult to accurately predict sorting indicators using traditional methods. Existing sorting technologies based on empirical formulas or single algorithms are mostly suitable for lithium-ion battery systems with stable data characteristics, but for liquid metal batteries, which have strong nonlinearity and time dependence, their prediction accuracy and stability are significantly insufficient.

[0003] In the early stages of liquid metal batteries, many key performance indicators are hidden in the subtle fluctuations of time-series data. Traditional manual feature extraction methods often fail to fully capture these subtle changes, resulting in insufficient generalization ability of the model and difficulty in achieving high-precision capacity prediction and sorting. Therefore, how to fully utilize the time-series information during the activation period of liquid metal batteries while improving the model's ability to identify and predict complex dynamic signals has become a key technical problem for the quality control and automated sorting of liquid metal batteries. Summary of the Invention

[0004] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a method, system, device, and storage medium for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning, addressing the problems of existing methods relying on manual features and insufficient generalization performance.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning, including: collecting discharge curve data of liquid metal batteries during the activation period and constructing an original input sequence; The discharge curve is divided into several two-dimensional time segments according to a preset window length and step size, and each segment is used as an independent input to a deep learning model. The model extracts local morphological features through one-dimensional convolution and pooling layers, and uses long short-term memory units to model temporal dependencies, outputting a final state vector. The final state vector is input into the fully connected layer to obtain the capacity prediction value of the corresponding segment, and the prediction results of all segments of the same battery are integrated based on a cycle-based voting to obtain the cycle-level capacity prediction value. Based on the comparison between the predicted cycle capacity and the preset threshold, the batteries are classified and sorted.

[0006] As a preferred embodiment of the liquid metal battery sorting index prediction method based on time-series features and deep learning described in this invention, the step of classifying and sorting the batteries based on the comparison result of the cycle-level capacity prediction value and the preset threshold includes: If the predicted cycle capacity is less than the preset threshold, the liquid metal battery is determined to be a defective product; if the predicted cycle capacity is not less than the preset threshold, the liquid metal battery is determined to be a qualified product.

[0007] As a preferred embodiment of the liquid metal battery sorting index prediction method based on temporal features and deep learning described in this invention, the deep learning model includes: an input layer, two levels of convolutional layers, two levels of max pooling layers, a long short-term memory network layer, and a fully connected output layer; the input of the deep learning model is a two-dimensional sequence of size L×2, and the output is the capacity prediction value of the corresponding segment. For the number of filters in the convolutional layer, the size of the convolutional window, the size of the pooling window, and the number of long short-term memory units, an automatic tuning strategy is adopted to optimize them layer by layer. With the other hyperparameters fixed, the hyperparameters to be optimized are randomly sampled several times within a preset search interval. After each sampling, the model is trained and the validation set error is calculated. The sampled value that minimizes the error of the validation set is selected as the optimal value of the hyperparameter, and this optimal value is fixed. The optimization of the next hyperparameter is then carried out until all hyperparameters are optimized, resulting in the best-performing deep learning model structure and parameter configuration.

[0008] The beneficial effect of this preferred technical solution is that by introducing a greedy optimization strategy during the training process, the hyperparameters of the model are adaptively adjusted, so that the model can maintain high prediction accuracy and stability under different operating conditions and battery batches.

[0009] As a preferred embodiment of the liquid metal battery sorting index prediction method based on temporal features and deep learning described in this invention, the model extracts local morphological features through one-dimensional convolution and pooling layers, including: setting two levels of one-dimensional convolution and max pooling layers in the input layer to extract local features and achieve downsampling. The output of a convolutional layer is defined as: in, Indicates the first In a one-dimensional convolutional network, the first layer Each convolutional kernel in time Output feature value at, To represent the layer index of a convolutional network, To indicate the number of the feature channels in the current convolutional layer, Represents a non-linear activation function. Indicates the first The first in the layer Each convolutional kernel is located at... The weight parameters at that location, To indicate the first The receptive field size of a one-dimensional convolution kernel. Indicates the time of the upper layer network Input feature value at; when hour, Corresponding two-dimensional time segment of the input ; Indicates the first The first in the layer The bias parameters corresponding to each convolutional kernel; The max pooling operation is as follows: in, Indicates the first Layered pooling operations within a time interval The pooling output feature values ​​obtained within the inner pooling are Indicates the first Layered convolutional networks in time Output feature value at, This represents the set of time indices covered by the pooling window. Indicates time index It is within the current pooling window.

[0010] As a preferred embodiment of the liquid metal battery sorting index prediction method based on temporal features and deep learning described in this invention, the model extracts local morphological features through one-dimensional convolution and pooling layers, including: setting two levels of one-dimensional convolution and max pooling layers in the input layer to extract local features and achieve downsampling. The output of a convolutional layer is defined as: in, Indicates the first In a one-dimensional convolutional network, the first layer Each convolutional kernel in time Output feature value at, To represent the layer index of a convolutional network, To indicate the number of the feature channels in the current convolutional layer, Represents a non-linear activation function. Indicates the first The first in the layer Each convolutional kernel is located at... The weight parameters at that location, To indicate the first The receptive field size of a one-dimensional convolution kernel. Indicates the time of the upper layer network Input feature value at; when hour, Corresponding two-dimensional time segment of the input ; Indicates the first The first in the layer The bias parameters corresponding to each convolutional kernel; The max pooling operation is as follows: in, Indicates the first Layered pooling operations within a time interval The pooling output feature values ​​obtained within the inner pooling are Indicates the first Layered convolutional networks in time Output feature value at, This represents the set of time indices covered by the pooling window. Indicates time index It is within the current pooling window.

[0011] As a preferred embodiment of the liquid metal battery sorting index prediction method based on temporal features and deep learning described in this invention, the method utilizes long short-term memory units to model temporal dependencies, and the output final state vector includes: the pooled sequence is input into the long short-term memory unit to learn temporal correlation, and the calculation process is as follows: in, This represents the input gate output at time step t. This represents the output of the forget gate at time step t. The output of the gate at time step t is represented by... Represents the candidate long-term memory state at time step t. This represents the long-term memory state at time step t. This represents the short-term memory output at time step t. This represents the feature vector input to the Long Short-Term Memory unit at time step t. This represents the short-term memory output at time step t−1. This represents the long-term memory state at time step t−1. Representing the input features respectively The weight matrix for the input gate, forget gate, output gate, and candidate memory units. These represent the outputs of short-term memory at the previous time step. The weight matrix for the input gate, forget gate, output gate, and candidate memory units. These represent the bias terms corresponding to the input gate, forget gate, output gate, and candidate memory units, respectively. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

[0012] As a preferred embodiment of the liquid metal battery sorting index prediction method based on time-series features and deep learning described in this invention, the following steps are taken: Cyclic voting integration is performed on the prediction results of all segments of the same battery to obtain the cycle-level capacity prediction value, including: the values ​​obtained by dividing the same battery into segments during a complete discharge cycle. S There are 1 time series segments, each segment corresponding to a capacity prediction value. The cycle-level capacity prediction value of the battery is defined as the arithmetic mean of the prediction values ​​of all segments.

[0013] The beneficial effect of this preferred technical solution is that by averaging the prediction results of multiple local segments, the prediction fluctuations caused by data noise or changes in local features are effectively smoothed, thereby improving the robustness and reliability of the model in practical applications.

[0014] Secondly, the present invention provides a liquid metal battery sorting index prediction system based on time-series features and deep learning, comprising: The data acquisition module is used to collect discharge curve data of liquid metal batteries during the activation phase and construct the original input sequence; The time-series segmentation module is used to divide the discharge curve into several two-dimensional time-series segments according to a preset window length and step size, and use each segment as an independent input to a deep learning model; The feature extraction module is used by the model to extract local morphological features through one-dimensional convolution and pooling layers, and to model temporal dependencies using long short-term memory units to output the final state vector. The cyclic voting integration module is used to input the final state vector into the fully connected layer to obtain the capacity prediction value of the corresponding segment, and to perform cyclic voting integration on the prediction results of all segments of the same battery to obtain the cyclic capacity prediction value. The classification and sorting decision module is used to classify and sort batteries based on the comparison results between the predicted cycle capacity value and the preset threshold.

[0015] Thirdly, the present invention provides an electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the liquid metal battery sorting index prediction method based on time-series features and deep learning.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the liquid metal battery sorting index prediction method based on time-series features and deep learning.

[0017] Compared with existing technologies, the advantages of this invention are as follows: This invention employs a hybrid deep learning model combining long short-term memory units and one-dimensional convolutional layers to automatically extract multi-dimensional dynamic features from the original activation period data, capturing short-term fluctuations and long-term trend information in the time series. Furthermore, it integrates multiple rounds of prediction results through a cyclic prediction mechanism and a voting ensemble strategy, reducing the uncertainty and noise interference of single predictions. In addition, a greedy optimization strategy is introduced during training to adaptively adjust the model's hyperparameters, ensuring high prediction accuracy and stability under different operating conditions and battery batches. This invention enables end-to-end prediction of sorting indicators for liquid metal batteries, avoiding the limitations of traditional manual feature selection and complex parameter modeling. It can be widely applied to the production testing, performance screening, and consistency evaluation of liquid metal batteries, providing a new technical path for their large-scale engineering application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments 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. Wherein: Figure 1 This is a schematic diagram of the process flow for a liquid metal battery sorting index prediction method based on time-series features and deep learning, according to an embodiment of the present invention. Figure 2 This is a battery sorting framework diagram based on a deep learning model, which is a method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning, according to an embodiment of the present invention. Figure 3 This is a diagram of the LSTM-CONV1D model structure of a liquid metal battery sorting index prediction method based on time-series features and deep learning, as described in one embodiment of the present invention. Figure 4This is a schematic diagram of the hyperparameter optimization process using a greedy strategy for a liquid metal battery sorting index prediction method based on time-series features and deep learning, as described in one embodiment of the present invention. Figure 5 This is a diagram illustrating the process of a greedy one-dimensional convolution in a liquid metal battery sorting index prediction method based on temporal features and deep learning, as described in an embodiment of the present invention. Figure 6 This is a structural diagram of a long short-term memory unit in a liquid metal battery sorting index prediction method based on temporal features and deep learning, as described in an embodiment of the present invention. Figure 7 This is a schematic diagram of the error distribution of a liquid metal battery sorting index prediction method based on time-series features and deep learning according to an embodiment of the present invention on a validation set. Figure 8 This is a schematic diagram showing the absolute value of the prediction error of a liquid metal battery sorting index prediction method based on time-series features and deep learning in one embodiment of the present invention for 174 cycles of training set and validation set batteries. Figure 9 This is a schematic diagram illustrating the absolute value of the prediction error of a liquid metal battery sorting index prediction method based on time-series features and deep learning for 21 cycles of a test set of batteries, as described in one embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Example 1, referring to Figures 1-7 This is one embodiment of the present invention, which provides a method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning, such as... Figure 1 As shown, it includes: S100: Collect discharge curve data of liquid metal batteries during the activation phase to construct the original input sequence; S200: Divide the discharge curve into several two-dimensional time segments according to a preset window length and step size, and use each segment as an independent input to the deep learning model; S300: The model extracts local morphological features through one-dimensional convolution and pooling layers, and uses long short-term memory units to model temporal dependencies, outputting a final state vector; S400: Input the final state vector into the fully connected layer to obtain the capacity prediction value of the corresponding segment, and perform a cycle-based voting integration of the prediction results of all segments of the same cell to obtain the cycle-level capacity prediction value. S500: Based on the comparison between the predicted cycle capacity and the preset threshold, the battery is classified and sorted.

[0021] It should be noted that traditional manual feature extraction methods often fail to fully capture these subtle changes, resulting in insufficient generalization ability of the model and difficulty in achieving high-precision capacity prediction and sorting. This invention constructs a hybrid deep neural network using Long Short-Term Memory (LSTM) units and a One-Dimensional Convolutional Neural Network (CONV1D) to achieve automatic feature extraction and capacity prediction of the discharge curve of liquid metal batteries. Furthermore, it combines a recurrence-based voting ensemble algorithm to improve overall prediction stability and reliability, optimize the feature extraction and selection problem, and improve battery sorting efficiency, thereby meeting the application requirements of liquid metal battery models.

[0022] In the embodiments of the present invention Figure 2 A battery sorting framework based on a deep learning model is shown; further, discharge curve data of liquid metal batteries during the activation phase are collected, and the original input sequence is constructed, including: Let the original input be: in, Indicates time The original input feature vectors collected at the site are used to characterize the instantaneous state of the liquid metal battery during the discharge process in the activation period; It represents a time variable or sampling time, corresponding to discrete or continuous time points in the discharge process; Indicates the time of liquid metal battery The terminal voltage value at the point is used to reflect the voltage change characteristics of the battery during the discharge process; Indicates time The corresponding discharge quantity or differential discharge quantity is used to describe the capacity release characteristics of the battery during the discharge process. Furthermore, the entire discharge curve is divided according to the window length. and step length Divided into multiple two-dimensional time segments: in, Indicates the first A two-dimensional time-series segment sample obtained by segmenting the original discharge curve. Indicates the index number of the time segment. This represents the total number of time segments obtained from the entire discharge curve using the sliding window segmentation method. Let represent a two-dimensional matrix space over the real number field, where: the first dimension The first dimension represents the length of the time step contained in each time segment; the second dimension 2 represents the number of feature dimensions contained in each time step, corresponding to voltage features and discharge quantity features. This indicates the length of the sliding window, used to control the time coverage of a single time segment.

[0023] Reference Figure 3 In this embodiment of the invention, the deep learning model in step S200 includes: an input layer, two levels of convolutional layers, two levels of max pooling layers, a long short-term memory network layer, and a fully connected output layer; the input of the deep learning model is a two-dimensional sequence of size L×2, and the output is the capacity prediction value of the corresponding segment; For the number of filters in the convolutional layer, the size of the convolutional window, the size of the pooling window, and the number of long short-term memory units, an automatic tuning strategy is adopted to optimize them layer by layer. With the other hyperparameters fixed, the hyperparameters to be optimized are randomly sampled several times within a preset search interval. After each sampling, the model is trained and the validation set error is calculated. The sampled value that minimizes the error of the validation set is selected as the optimal value of the hyperparameter, and this optimal value is fixed. The optimization of the next hyperparameter is then carried out until all hyperparameters are optimized, resulting in the best-performing deep learning model structure and parameter configuration.

[0024] It should be noted that the model structure of this embodiment can simultaneously extract the local trend and time dependency of voltage changes, realizing end-to-end mapping from the original timing sequence to capacity.

[0025] Specifically, the model training uses the Keras framework, with RMSprop as the optimizer, mean squared error (MSE) as the loss function, a learning rate of 0.005, 300 training epochs, and a batch size of 32. The relevant parameter configurations are shown in Table 1.

[0026] Table 1. Configuration of the training process

[0027] Reference Figure 4 In one possible implementation, this embodiment of the invention employs a layer-by-layer greedy search strategy to optimize the model hyperparameters. With the remaining parameters fixed, the key parameters of each layer (number of convolutional kernels, window size, pooling size, number of LSTM units) are randomly sampled 10 times and the model is trained. The parameter corresponding to the minimum MSE value on the validation set is taken as the optimal value.

[0028] The parameter search range for the optimization process is shown in Table 2, and the final optimization results and corresponding validation set MSE are shown in Table 3.

[0029] Table 2 Search Space of Model Parameters

[0030] Table 3 shows the optimization results for each parameter and the corresponding MSE on the validation set.

[0031] After optimization, the validation set MSE decreased from 0.5210 to 0.4040, and the test set RMSE decreased from 0.5296 to 0.4414, as shown in Table 4, proving that the optimization strategy effectively improves the model performance.

[0032] Table 4 shows the model's performance on the validation and test sets (fragments) before and after optimization.

[0033] Reference Figure 5 In this embodiment of the invention, step S300 involves the model extracting local morphological features through one-dimensional convolution and pooling layers, including: setting two levels of one-dimensional convolution and max pooling layers in the input layer to extract local features and achieve downsampling. The output of a convolutional layer is defined as: in, Indicates the first In a one-dimensional convolutional network, the first layer Each convolutional kernel in time Output feature value at, To represent the layer index of a convolutional network, To indicate the number of the feature channels in the current convolutional layer, Represents a non-linear activation function. Indicates the first The first in the layer Each convolutional kernel is located at... The weight parameters at that location, To indicate the first The receptive field size of a one-dimensional convolution kernel. Indicates the time of the upper layer network Input feature value at; when hour, Corresponding two-dimensional time segment of the input ; Indicates the first The first in the layer The bias parameters corresponding to each convolutional kernel; The max pooling operation is as follows: in, Indicates the first Layered pooling operations within a time interval The pooling output feature values ​​obtained within the inner pooling are Indicates the first Layered convolutional networks in time Output feature value at, This represents the set of time indices covered by the pooling window. Indicates time index It is within the current pooling window.

[0034] It should be noted that the combination of multiplication layers and pooling layers can effectively compress the length of the input sequence while retaining key information features.

[0035] Reference Figure 6 In this embodiment of the invention, the Long Short-Term Memory (LSTM) unit achieves the fusion of long-term and short-term memory through a gating structure. In this embodiment, step S300 utilizes the LSTM unit to model temporal dependencies, and the output final state vector includes: the pooled sequence is input into the LSTM unit to learn temporal correlations; the calculation process is as follows: in, Indicates time step Input gate output at any time, Indicates time step The output of the forget gate at any moment, Indicates time step The output gate outputs at any time, using Indicates time step Candidate long-term memory states at time points Indicates time step Long-term memory state at any given moment Indicates time step Short-term memory output at any given moment Indicates time step The feature vectors input into the long short-term memory unit at each time step, Indicates time step Short-term memory output at any given moment Indicates time step Long-term memory state at any given moment Representing the input features respectively The weight matrix for the input gate, forget gate, output gate, and candidate memory units. These represent the outputs of short-term memory at the previous time step. The weight matrix for the input gate, forget gate, output gate, and candidate memory units. These represent the bias terms corresponding to the input gate, forget gate, output gate, and candidate memory units, respectively. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

[0036] In this embodiment of the invention, step S400, which inputs the final state vector into the fully connected layer to obtain the capacity prediction value of the corresponding segment, includes: inputting the final state vector output by the Long Short-Term Memory network layer into the fully connected layer, and outputting the predicted capacity of the corresponding segment: in, This represents the prediction capacity value corresponding to the s-th time segment. Let w represent the hidden state vector of the Long Short-Term Memory network at the last time step T of the current time segment, where T represents the total number of time steps in the current time segment, and w represents the weight vector of the fully connected layer. represents the transpose of the weight vector w; b represents the bias term of the fully connected layer.

[0037] Mean squared error is used as the loss function: Where L represents the loss function value used during model training. This indicates the number of samples participating in training or evaluation. Indicates the sample index number. Indicates the first The true capacity value corresponding to each sample The model represents the first The capacity value predicted for each sample The term represents the squared error between the predicted capacity and the actual capacity, and λ represents the regularization coefficient. This represents the complete set of trainable parameters in the model. This represents the squared L2 norm of the model parameters.

[0038] In this embodiment of the invention, step S400, which involves performing cycle-based voting integration on the prediction results of all segments of the same battery to obtain the cycle-level capacity prediction value, includes: for the same battery, the predicted values ​​obtained from the full discharge cycle... S There are 1 time series segments, each segment corresponding to a capacity prediction value. The cycle-level capacity prediction value of the battery is defined as the arithmetic mean of the prediction values ​​of all segments.

[0039] Reference Figure 7 Specifically, for the same battery Segment prediction results Define the cycle-level capacity prediction value: in, This represents the cycle-stage predicted capacity value of the same battery within a complete discharge cycle, where S represents the total number of time segments obtained from the discharge curve of the same battery. Indicates the first The prediction capacity value corresponding to each time segment, 1 / S represents the normalization coefficient of the arithmetic mean of the prediction results of all segments.

[0040] It should be noted that the integration strategy in this embodiment transforms the originally scattered segment-level predictions into a global prediction value representing the entire discharge process, achieving information aggregation from the segment level to the cycle level. Since each segment only covers a part of the discharge curve, its prediction may have biases; however, by averaging multiple segments, the influence of individual abnormal segments can be suppressed, enhancing the consistency and reliability of the overall prediction.

[0041] In a preferred implementation, to reduce fragment-level fluctuations and improve prediction stability, thereby minimizing the impact of anomalous fragments, a robust weighted approach can also be adopted, specifically as follows: in, This represents the predicted capacity value obtained based on the weighted voting mechanism at the cyclic level. This represents the weighting coefficient corresponding to the prediction result of the s-th time segment, which is used to adjust the contribution of this segment in the cyclic prediction. This represents the result of a weighted sum of the predicted capacities of each segment according to their respective weights; This represents the normalization factor for all segment weight coefficients; Represents the set of predicted capacities for all segments. The median is used to reflect the central tendency of the segment prediction results; This represents the absolute deviation between the predicted capacity of the s-th segment and the median; This indicates the definition of the weight coefficients, where the segment whose prediction deviates more from the median has a smaller corresponding weight.

[0042] It should be noted that the average error of the weighted average method on the validation set fragments is 0.020356 Ah, and the error of the vast majority of fragments is within ±0.75 Ah, which proves the consistency and robustness of the prediction results.

[0043] In this embodiment of the invention, step S500, which involves classifying and sorting batteries based on the comparison between the predicted cycle capacity and a preset threshold, includes: If the predicted cycle capacity is less than the preset threshold, the liquid metal battery is determined to be a defective product; if the predicted cycle capacity is not less than the preset threshold, the liquid metal battery is determined to be a qualified product.

[0044] It should be noted that, in this embodiment of the invention, the preset threshold can be set to... .

[0045] The above is an illustrative scheme of a liquid metal battery sorting index prediction method based on time-series features and deep learning according to this embodiment. It should be noted that the technical solution of this liquid metal battery sorting index prediction system based on time-series features and deep learning belongs to the same concept as the technical solution of the liquid metal battery sorting index prediction method based on time-series features and deep learning described above. Details not described in detail in the technical solution of the liquid metal battery sorting index prediction system based on time-series features and deep learning in this embodiment can be found in the description of the technical solution of the liquid metal battery sorting index prediction method based on time-series features and deep learning described above.

[0046] This embodiment presents a liquid metal battery sorting index prediction system based on time-series features and deep learning, comprising: The data acquisition module is used to collect discharge curve data of liquid metal batteries during the activation phase and construct the original input sequence; The time-series segmentation module is used to divide the discharge curve into several two-dimensional time-series segments according to a preset window length and step size, and use each segment as an independent input to the deep learning model; The feature extraction module is used by the model to extract local morphological features through one-dimensional convolution and pooling layers, and to model temporal dependencies using long short-term memory units to output the final state vector. The cyclic voting integration module is used to input the final state vector into the fully connected layer to obtain the capacity prediction value of the corresponding segment, and to perform cyclic voting integration on the prediction results of all segments of the same cell to obtain the cyclic capacity prediction value. The classification and sorting decision module is used to classify and sort batteries based on the comparison results between the cycle-level capacity prediction value and the preset threshold.

[0047] This embodiment also provides an electronic device applicable to the prediction method of sorting indicators for liquid metal batteries based on time-series features and deep learning, including: The memory and processor are used to store computer-executable instructions and execute computer-executable instructions to implement the liquid metal battery sorting index prediction method based on time-series features and deep learning proposed in the above embodiments.

[0048] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the liquid metal battery sorting index prediction method based on time-series features and deep learning as proposed in the above embodiments.

[0049] The storage medium proposed in this embodiment and the liquid metal battery sorting index prediction method based on time-series features and deep learning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0050] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0051] Example 2, refer to Figures 8-9 This is one embodiment of the present invention, and this embodiment verifies the beneficial effects of the present invention through scientific experiments.

[0052] In this embodiment, the baseline capacity of the liquid metal battery is 50 Ah. For example... Figure 7 As shown, the absolute values ​​of prediction errors of the models on the 174 cycles of the training and validation sets are generally low, and the LSTM-CONV1D model has the smallest error fluctuation, showing good stability and robustness. Its overall RMSE is 0.2945 and RMSPE is 0.589%, which is significantly better than the control group RNN (1.0552%) and CNN (1.101%).

[0053] like Figure 8 As shown, the model's maximum error on the 21 cycles of the test set was only 0.4976 Ah, corresponding to an overall RMSPE of 0.7524%, which is 27.7% lower than the control model on average, demonstrating strong generalization ability.

[0054] During the classification and determination phase, a capacity threshold is set. When predicting capacity If the result is within the specified range, it is considered qualified; otherwise, it is considered unqualified. A binary classification experiment was conducted using the prediction results of the above model, and the resulting confusion matrix is ​​shown in Table 5: the proportion of samples on the main diagonal is as high as 90.77%, which is significantly better than the control models RNN and CNN (79.49% and 76.41%, respectively).

[0055] Table 5. Classification results (confusion matrix) of the three models (a) LSTM-CONV1D model

[0056] (b) RNN model

[0057] (c) CNN model

[0058] Table 5 shows the effectiveness and accuracy of the LSTM-CONV1D model and the recurrence-based prediction method in liquid metal battery sorting applications. This invention employs a hybrid deep learning model combining long short-term memory units and one-dimensional convolutional layers to automatically extract multi-dimensional dynamic features from the original activation period data, capturing short-term fluctuations and long-term trend information in the time series. Furthermore, it integrates multiple rounds of prediction results through a recurrence prediction mechanism and a voting ensemble strategy, reducing the uncertainty and noise interference of single predictions. In addition, a greedy optimization strategy is introduced during training to adaptively adjust the model's hyperparameters, ensuring high prediction accuracy and stability under different operating conditions and battery batches. This invention enables end-to-end prediction of liquid metal battery sorting indicators, avoiding the limitations of traditional manual feature selection and complex parameter modeling. It can be widely applied to the production testing, performance screening, and consistency evaluation of liquid metal batteries, providing a new technical path for their large-scale engineering applications.

[0059] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning, characterized in that, include: Discharge curve data of liquid metal batteries during the activation phase were collected to construct the original input sequence; The discharge curve is divided into several two-dimensional time segments according to a preset window length and step size, and each segment is used as an independent input to a deep learning model. The model extracts local morphological features through one-dimensional convolution and pooling layers, and uses long short-term memory units to model temporal dependencies, outputting a final state vector. The final state vector is input into the fully connected layer to obtain the capacity prediction value of the corresponding segment, and the prediction results of all segments of the same battery are integrated based on a cycle-based voting to obtain the cycle-level capacity prediction value. Based on the comparison between the predicted cycle capacity and the preset threshold, the batteries are classified and sorted.

2. The method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning as described in claim 1, characterized in that, The battery classification and sorting decision based on the comparison result between the cycle-level capacity prediction value and the preset threshold includes: If the predicted cycle capacity is less than the preset threshold, the liquid metal battery is determined to be a defective product; if the predicted cycle capacity is not less than the preset threshold, the liquid metal battery is determined to be a qualified product.

3. The method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning as described in claim 2, characterized in that, The deep learning model includes: an input layer, two levels of convolutional layers, two levels of max pooling layers, a long short-term memory network layer, and a fully connected output layer; the input of the deep learning model is a two-dimensional sequence of size L×2, and the output is the capacity prediction value of the corresponding segment; For the number of filters in the convolutional layer, the size of the convolutional window, the size of the pooling window, and the number of long short-term memory units, an automatic tuning strategy is adopted to optimize them layer by layer. With the other hyperparameters fixed, the hyperparameters to be optimized are randomly sampled several times within a preset search interval. After each sampling, the model is trained and the validation set error is calculated. The sampled value that minimizes the error of the validation set is selected as the optimal value of the hyperparameter, and this optimal value is fixed. The optimization of the next hyperparameter is then carried out until all hyperparameters are optimized, resulting in the best-performing deep learning model structure and parameter configuration.

4. The method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning as described in claim 3, characterized in that, The model extracts local morphological features through one-dimensional convolution and pooling layers, including: setting two levels of one-dimensional convolution and max pooling layers in the input layer to extract local features and achieve downsampling. The output of a convolutional layer is defined as: in, Indicates the first In a one-dimensional convolutional network, the first layer Each convolutional kernel in time Output feature value at, To represent the layer index of a convolutional network, To indicate the number of the feature channels in the current convolutional layer, Represents a non-linear activation function. Indicates the first The first in the layer Each convolutional kernel is located at... The weight parameters at that location, To indicate the first The receptive field size of a one-dimensional convolution kernel. Indicates the time of the upper layer network Input feature value at; when hour, Corresponding two-dimensional time segment of the input ; Indicates the first The first in the layer The bias parameters corresponding to each convolutional kernel; The max pooling operation is as follows: in, Indicates the first Layered pooling operations within a time interval The pooling output feature values ​​obtained within the inner pooling are Indicates the first Layered convolutional networks in time Output feature value at, This represents the set of time indices covered by the pooling window. Indicates time index It is within the current pooling window.

5. The method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning as described in claim 4, characterized in that, Using Long Short-Term Memory (LSTM) units to model temporal dependencies, the output final state vector includes: the pooled sequence input into the LTM unit to learn temporal correlations, the calculation process is as follows: in, This represents the input gate output at time step t. This represents the output of the forget gate at time step t. The output of the gate at time step t is represented by... Represents the candidate long-term memory state at time step t. This represents the long-term memory state at time step t. This represents the short-term memory output at time step t. This represents the feature vector input to the Long Short-Term Memory unit at time step t. This represents the short-term memory output at time step t−1. This represents the long-term memory state at time step t−1. Representing the input features respectively The weight matrix for the input gate, forget gate, output gate, and candidate memory units. These represent the outputs of short-term memory at the previous time step. The weight matrix for the input gate, forget gate, output gate, and candidate memory units. These represent the bias terms corresponding to the input gate, forget gate, output gate, and candidate memory units, respectively. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

6. The method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning as described in claim 5, characterized in that, The process of inputting the final state vector into the fully connected layer to obtain the capacity prediction value of the corresponding segment includes: inputting the final state vector output from the Long Short-Term Memory (LSTM) network layer into the fully connected layer, and outputting the predicted capacity of the corresponding segment. in, This represents the prediction capacity value corresponding to the s-th time segment. Let w represent the hidden state vector of the Long Short-Term Memory network at the last time step T of the current time segment, where T represents the total number of time steps in the current time segment, and w represents the weight vector of the fully connected layer. represents the transpose of the weight vector w; b represents the bias term of the fully connected layer. Mean squared error is used as the loss function: Where L represents the loss function value used during model training. This indicates the number of samples participating in training or evaluation. Indicates the sample index number. Indicates the first The true capacity value corresponding to each sample The model represents the first The capacity value predicted for each sample The term represents the squared error between the predicted capacity and the actual capacity, and λ represents the regularization coefficient. This represents the complete set of trainable parameters in the model. This represents the squared L2 norm of the model parameters.

7. The method for predicting sorting indicators of liquid metal batteries based on time-series features and deep learning as described in claim 6, characterized in that, Cyclic voting-based ensemble is performed on all segment prediction results for the same battery to obtain cycle-level capacity predictions, including: predictions obtained by dividing the same battery into segments during a complete discharge cycle. S There are 1 time series segments, each segment corresponding to a capacity prediction value. The cycle-level capacity prediction value of the battery is defined as the arithmetic mean of the prediction values ​​of all segments.

8. A liquid metal battery sorting index prediction system based on time-series features and deep learning, applied to the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect discharge curve data of liquid metal batteries during the activation phase and construct the original input sequence; The time-series segmentation module is used to divide the discharge curve into several two-dimensional time-series segments according to a preset window length and step size, and use each segment as an independent input to a deep learning model; The feature extraction module is used by the model to extract local morphological features through one-dimensional convolution and pooling layers, and to model temporal dependencies using long short-term memory units to output the final state vector. The cyclic voting integration module is used to input the final state vector into the fully connected layer to obtain the capacity prediction value of the corresponding segment, and to perform cyclic voting integration on the prediction results of all segments of the same battery to obtain the cyclic capacity prediction value. The classification and sorting decision module is used to classify and sort batteries based on the comparison results between the predicted cycle capacity value and the preset threshold.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the liquid metal battery sorting index prediction method based on time-series features and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the liquid metal battery sorting index prediction method based on time-series features and deep learning as described in any one of claims 1 to 7.

Images