A method and system for predicting the weld area of an ultrasonic weld of a battery cell

Abstract

The application discloses a kind of electric core ultrasonic welding weld mark area prediction method and system, it is related to lithium ion electric core manufacturing technical field, the method includes: obtaining multi-dimensional time series data in welding process and weld mark area after welding;Multi-dimensional time series data is reconstructed using autoencoder processing;According to the reconstruction time series data of each dimension and the weld mark area after welding, a one-level sample set of each dimension is formed;A one-level prediction model is constructed for the one-level sample set of each dimension, and is trained by one-level sample set, generates the weld mark area prediction value of each dimension after training and is composed into multi-dimensional feature vector, according to multi-dimensional feature vector and the weld mark area after welding, a two-level sample set is formed;A two-level prediction model is constructed, and the two-level prediction model is trained by two-level sample set;Weld mark area prediction of electric core ultrasonic welding is realized by the two-level prediction model trained.The application can realize high accuracy and high timeliness of weld mark area prediction.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery cell manufacturing technology, and in particular to a method and system for predicting the area of ​​ultrasonic welding stamps in battery cells. Background Technology

[0002] In the lithium-ion battery cell manufacturing process, ultrasonic welding is the core process for achieving reliable tab connection. The welding quality directly determines the battery cell's electrical performance, safety stability, and cycle life. Ultrasonic welding quality is determined by the synergistic effect of multi-dimensional time-series waveforms, including energy, pressure, amplitude, and power, during the welding process. The temporal variation characteristics of each parameter (such as energy rise rate and power peak magnitude) lead to changes in the welding area. Currently, the industry commonly uses CCD visual inspection technology to determine whether the welding quality is acceptable (OK / NG) by measuring the size of the solder area. However, this method can only achieve post-process screening and cannot predict welding quality in advance. Furthermore, when a batch of NG cells appears, it is difficult to trace the root cause at the time-series parameter level. Existing related technologies mostly focus on static threshold control of single parameters or simple time-series characteristic statistics, and have not yet formed a deep fusion analysis system for multi-dimensional time-series data. Even more so, there is a lack of technical solutions for predicting solder area in advance through "virtual measurement."

[0003] As the lithium battery industry continues to demand higher standards for cell consistency (such as the requirement to control the deviation of the solder area of ​​power batteries within +5%), traditional detection methods based on static thresholds and single models can no longer meet the needs of large-scale, high-frequency production. There is an urgent need for a technical solution that can accurately process multi-dimensional time-series data and achieve high-precision prediction of solder area. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting the area of ​​ultrasonic welding marks on battery cells, solving the technical problems that existing detection methods either can only achieve post-screening or have low detection accuracy.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for predicting the area of ​​ultrasonic welding solder marks in battery cells, comprising:

[0007] Acquire multi-dimensional time-series data during the welding process and the weld area after welding;

[0008] An autoencoder is used to reconstruct the multi-dimensional time series data to generate multi-dimensional reconstructed time series data;

[0009] The reconstructed time series data and the weld area after welding are used to construct the first-level sample set for each dimension.

[0010] A first-level CNN prediction model is constructed for each dimension of the first-level sample set, and the corresponding first-level CNN prediction model is trained using the first-level sample set;

[0011] The reconstructed time series data of each dimension are input into the corresponding trained first-level CNN prediction model to generate the predicted solder area values ​​of each dimension.

[0012] The predicted solder area values ​​of each dimension are combined into a multi-dimensional feature vector, and a secondary sample set is formed based on the multi-dimensional feature vector and the solder area after welding.

[0013] Construct a two-level CNN prediction model and train the two-level CNN prediction model using the two-level sample set;

[0014] The trained secondary CNN prediction model is used to predict the weld area of ​​the ultrasonic welding of battery cells.

[0015] The present invention provides a method for predicting the weld area of ​​ultrasonic welding of battery cells. This method reconstructs multi-dimensional time-series data during the welding process using an autoencoder, ensuring the integrity of the temporal dynamic features and providing high-quality input for subsequent processing. It employs a two-layer architecture: a first-level CNN prediction model and a second-level CNN prediction model. The first-level CNN prediction model is independently pre-trained for single-dimensional time-series data, mining local temporal features and outputting preliminary prediction values. The second-level CNN prediction model uses the prediction results of the first-level CNN prediction model as input, learns multi-parameter global correlations, and outputs the final prediction value. This addresses the problem of insufficient feature mining in traditional single-model methods, effectively improving prediction accuracy. Furthermore, the prediction method provided by this invention performs synchronous prediction based on multi-dimensional time-series data during the welding process. Compared to post-screening methods, it offers better timeliness, enabling timely warnings when problems are detected and reducing losses.

[0016] Optionally, the multi-dimensional time-series data includes time-series data of welding energy, welding pressure, welding amplitude, and welding power.

[0017] In this invention, time-series data of four dimensions—welding energy, welding pressure, welding amplitude, and welding power—are selected from among the many factors that affect welding quality during the welding process. This facilitates the subsequent integration of multi-dimensional time-series information, thereby improving prediction accuracy and generalization ability.

[0018] Optionally, the autoencoder includes an encoder and a decoder;

[0019] The encoder uses a three-layer LSTM network to downsample the input time-series data, extracts deep time-series features through the ReLU activation function, and outputs a feature encoding vector of the target dimension.

[0020] The decoder uses a three-layer LSTM network to upsample the feature encoding vector of the target dimension, and then outputs the reconstructed temporal data through a fully connected layer.

[0021] The mean squared error is used as the loss function, and the network parameters of the autoencoder are optimized with the goal of minimizing the error between the time series data and the reconstructed time series data.

[0022] This invention employs an autoencoder containing an LSTM network. The LSTM network serves as the backbone network for both the encoder and decoder, and is specifically designed to model and reconstruct long-term dependencies in time-series data. This ensures the generation of a reconstructed sequence that is temporally coherent, smooth, and conforms to historical patterns.

[0023] Optionally, the reconstruction process of the multi-dimensional time-series data using an autoencoder includes:

[0024] Normalize the time series data for each dimension;

[0025] The time series data after normalization is filled with missing values ​​using the autoencoder to complete the missing value repair, and the reconstructed time series data of the target length is generated using the autoencoder to achieve length standardization.

[0026] This invention utilizes an autoencoder to perform missing value repair and length standardization of time-series data. Missing value repair provides subsequent models with a complete and continuous input sequence. This avoids biases or complex processing introduced by missing values, enabling the model to learn true and coherent temporal dependencies rather than being interfered with by data gaps. Simultaneously, the autoencoder, based on the contextual information of the entire sequence, can generate more accurate missing values. Length standardization also solves the problem of inconsistent original time series lengths, ensuring the uniformity of the temporal dimensions of each parameter and meeting the input requirements of subsequent models.

[0027] Optionally, cells with solder area within the acceptable range are designated as OK cells, and cells outside the acceptable range are designated as NG cells.

[0028] The ratio of OK cells to NG cells corresponding to the solder area in the primary sample set for each dimension is 1:1.

[0029] This invention avoids data distribution deviations affecting the pre-training effect by controlling the ratio of OK cells to NG cells.

[0030] Optionally, the initial parameters of the first-level CNN prediction model and the second-level CNN prediction model are generated independently and randomly.

[0031] This invention ensures the independence of training by independently and randomly assigning initial parameters. After the first-level CNN prediction model is trained, the parameters are fixed and used only as a feature extractor to output preliminary prediction values. The second-level CNN prediction model is trained independently and only updates its own parameters, thus avoiding prediction bias caused by cross-stage parameter interference.

[0032] In a second aspect, the present invention provides a battery cell ultrasonic welding solder mark area prediction system, comprising:

[0033] The data acquisition module is configured to acquire multi-dimensional time-series data during the welding process and the weld area after welding.

[0034] The data reconstruction module is configured to use an autoencoder to reconstruct the multi-dimensional time-series data and generate multi-dimensional reconstructed time-series data.

[0035] The first-level sample set module is configured to construct first-level sample sets for each dimension based on the reconstructed time-series data and the weld area after welding.

[0036] The first-level model training module is configured to construct a first-level CNN prediction model for each dimension of the first-level sample set, and train the corresponding first-level CNN prediction model using the first-level sample set.

[0037] The first-level model prediction module is configured to input the reconstructed time series data of each dimension into the corresponding trained first-level CNN prediction model to generate the predicted values ​​of the solder area of ​​each dimension.

[0038] The secondary sample set module is configured to form a multi-dimensional feature vector from the predicted solder area values ​​of each dimension, and to construct a secondary sample set based on the multi-dimensional feature vector and the solder area after welding.

[0039] The secondary model training module is configured to construct a secondary CNN prediction model and train the secondary CNN prediction model using the secondary sample set;

[0040] The secondary model prediction module is configured to predict the weld area of ​​the ultrasonic welding of the battery cell using the trained secondary CNN prediction model.

[0041] Thirdly, the present invention provides an electronic device, including a processor and a storage medium;

[0042] The storage medium is used to store instructions;

[0043] The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0044] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0045] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0046] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0047] This invention provides a method and system for predicting the weld area of ​​ultrasonic welding of battery cells. It reconstructs multi-dimensional time-series data during the welding process using an autoencoder, ensuring the integrity of the temporal dynamic features and providing high-quality input for subsequent processing. A two-layer architecture is employed: a first-level CNN prediction model and a second-level CNN prediction model. The first-level CNN prediction model is independently pre-trained for single-dimensional time-series data, mining local temporal features and outputting preliminary prediction values. The second-level CNN prediction model uses the prediction results of the first-level CNN prediction model as input, learns multi-parameter global correlations, and outputs the final prediction value. This addresses the problem of insufficient feature mining in traditional single-model approaches, effectively improving prediction accuracy. Furthermore, the prediction method provided by this invention performs synchronous prediction based on multi-dimensional time-series data during the welding process. Compared to post-screening methods, it offers better timeliness, enabling timely warnings when problems are detected and reducing losses. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method for predicting the area of ​​ultrasonic welding solder marks in battery cells provided in an embodiment of the present invention;

[0049] Figure 2 This is a flowchart of autoencoder reconstruction and two-level CNN prediction provided in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0051] Example 1

[0052] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for predicting the area of ​​ultrasonic welding marks in battery cells, including the following steps:

[0053] Step S1: Obtain multi-dimensional time-series data during the welding process and the weld area after welding.

[0054] Specifically, in this embodiment, the multi-dimensional time-series data includes time-series data of welding energy, welding pressure, welding amplitude, and welding power.

[0055] For a single lithium battery cell welding production line, multi-dimensional time-series data are collected during the welding process of each cell. Each data point contains 4 input time-series parameters and 1 output label (used for model training and validation). The specific specifications are shown in Table 1.

[0056] Table 1: Acquisition Specifications for Multi-Dimensional Time Series Data

[0057]

[0058] A high-speed data acquisition card is used to synchronously acquire four time-series parameters, ensuring that the sampling timestamps of each parameter are consistent and the time synchronization accuracy is <1µs. The raw data is stored in a time-series database (InfluxDB), which supports fast retrieval and batch export of single time-series data. The data storage latency is <50ms, which meets the data requirements for virtual measurement model training and real-time prediction.

[0059] In this invention, time-series data of four dimensions—welding energy, welding pressure, welding amplitude, and welding power—are selected from among the many factors that affect welding quality during the welding process. This facilitates the subsequent integration of multi-dimensional time-series information, thereby improving prediction accuracy and generalization ability.

[0060] Step S2: Use an autoencoder to reconstruct the multi-dimensional time series data to generate multi-dimensional reconstructed time series data.

[0061] Step S2.1 Specifically, in this embodiment, the autoencoder includes an encoder and a decoder;

[0062] The encoder uses a three-layer LSTM network to downsample the input temporal data, extracts deep temporal features through the ReLU activation function, and outputs a feature encoding vector of the target dimension.

[0063] According to the requirements of this embodiment, the time series data is padded with zeros to a fixed length of 512. The number of LSTM neurons in the three-layer LSTM network is 256, 128 and 64 respectively, and finally outputs a 64-dimensional feature encoding vector.

[0064] The decoder uses a three-layer LSTM network to upsample the feature encoding vector of the target dimension, and then outputs the reconstructed temporal data through a fully connected layer;

[0065] According to the requirements of this embodiment, the number of LSTM neurons in the three-layer LSTM network of the decoder is 64, 128, and 256 respectively.

[0066] The mean squared error is used as the loss function, and the network parameters of the autoencoder are optimized with the goal of minimizing the error between the time series data and the reconstructed time series data.

[0067] This invention employs an autoencoder containing an LSTM network. The LSTM network serves as the backbone network for both the encoder and decoder, and is specifically designed to model and reconstruct long-term dependencies in time-series data. This ensures the generation of a reconstructed sequence that is temporally coherent, smooth, and conforms to historical patterns.

[0068] Step S2.1, specifically in this embodiment, the reconstruction processing of multi-dimensional time-series data using an autoencoder includes:

[0069] The time series data of each dimension are normalized to map the values ​​to the [0,1] interval, thus eliminating the influence of dimensions;

[0070] For the normalized time series data, an autoencoder is used to adaptively fill in missing values ​​to complete the missing value repair, and an autoencoder is used to generate reconstructed time series data of the target length to achieve length standardization.

[0071] This invention utilizes an autoencoder to perform missing value repair and length standardization of time-series data. Missing value repair provides subsequent models with a complete and continuous input sequence. This avoids biases or complex processing introduced by missing values, enabling the model to learn true and coherent temporal dependencies rather than being interfered with by data gaps. Simultaneously, the autoencoder, based on the contextual information of the entire sequence, can generate more accurate missing values. Length standardization also solves the problem of inconsistent original time series lengths, ensuring the uniformity of the temporal dimensions of each parameter and meeting the input requirements of subsequent models.

[0072] Step S3: Construct a first-level sample set for each dimension based on the reconstructed time series data and the weld area after welding.

[0073] Cells with solder area within the acceptable range are designated as OK cells, while cells outside the acceptable range are designated as NG cells.

[0074] The ratio of OK cells to NG cells corresponding to the solder area of ​​the primary sample set in each dimension is 1:1 to avoid the data distribution bias affecting the pre-training effect.

[0075] Step S4: Construct a first-level CNN prediction model for each dimension of the first-level sample set, and train the corresponding first-level CNN prediction model using the first-level sample set.

[0076] Specifically, in this embodiment, the model structure of each first-level CNN prediction model is completely identical, and all include:

[0077] ConV convolutional layer 1: The kernel size is 3x1, the number of kernels is 32, the stride is 1, the padding method is "SAME", the activation function is ReLU, and the output feature map dimension is 32×256;

[0078] Pooling layer 1: Max pooling is used, with a pooling kernel size of 2x1, a stride of 2, and an output feature map dimension of 32×128;

[0079] ConV convolutional layer 2: The kernel size is 3x1, the number of kernels is 64, the stride is 1, the padding method is "SAME", the activation function is ReLU, and the output feature map dimension is 64x128.

[0080] Pooling layer 2: Max pooling is used, with a pooling kernel size of 2x1, a stride of 2, and an output feature map dimension of 64x64;

[0081] Fully connected layer 1: Input dimension is 64x64=4096, number of neurons is 128, activation function is ReLU;

[0082] Fully connected layer 2: The number of neurons is 1, and the output is the preliminary prediction value of the solder surface area corresponding to a single time sequence;

[0083] The initial parameters of each first-level CNN prediction model are randomly assigned to avoid cross-model parameter interference.

[0084] Independent training: Train the corresponding first-level CNN prediction model separately for each first-level sample set of each dimension. The optimization objective is to minimize the MSE between the predicted and actual solder area on the pre-training set. The number of training iterations is set to 150, the learning rate is 0.001, and the optimizer is Adam. After each iteration, the model performance is verified through the pre-validation set to avoid overfitting.

[0085] Pre-training validation: Evaluate model performance using a pre-validation set. This requires a pre-validation set for each Level 1 CNN prediction model. Models that fail to meet the standards need to be retrained by adjusting the number of convolutional kernels (±8) or the number of iterations (±50) to ensure that the pre-trained model has stable single-parameter prediction capabilities and provides reliable features for secondary training.

[0086] Pre-trained model solidification: After validation, all parameters (convolutional layer weights, fully connected layer biases, etc.) of the four first-level CNN prediction models are fixed, and only their "input-preliminary prediction value" mapping function is retained as feature extractors for second-level training, and they no longer participate in subsequent parameter updates.

[0087] Step S5: Input the reconstructed time series data of each dimension into the corresponding trained first-level CNN prediction model to generate the predicted solder area values ​​of each dimension.

[0088] This invention performs independent pre-training on the correlation between a single-dimensional time-series parameter and the solder area, achieving a preliminary mapping of "single parameter - solder area", providing high-quality basic features for secondary overall training, and reducing the difficulty of virtual measurement caused by multi-parameter coupling.

[0089] Step S6: Combine the predicted solder area values ​​of each dimension into a multi-dimensional feature vector, and construct a secondary sample set based on the multi-dimensional feature vector and the solder area after welding.

[0090] Specifically, all original samples (including training and test sets) are input into four solidified first-level CNN prediction models to obtain four preliminary prediction values ​​for each sample (y1-predicted value corresponding to energy time series, y2-predicted value corresponding to pressure time series, y3-predicted value corresponding to amplitude time series, and y4-predicted value corresponding to power time series), forming a 4-dimensional feature vector [y-y1, y-y2, y-y3, y-y4] (y is the actual solder area value). This vector integrates the influence information of each single parameter on the solder area.

[0091] The secondary sample set containing "4-dimensional feature vector + actual solder area" is divided into a secondary training set (80%) and a secondary test set (20%) in an 8:2 ratio to ensure that there is no overlap with the primary pre-training dataset (i.e., the secondary training set samples are not included in the primary pre-training / pre-validation set) to avoid data leakage affecting the model's generalization ability.

[0092] Step S7: Construct a two-level CNN prediction model and train the two-level CNN prediction model using a two-level sample set.

[0093] Specifically, in this embodiment, the model structure of the two-level CNN prediction model is as follows:

[0094] ConV convolutional layer: kernel size is 2x1, number is 16, stride is 1, padding method is "SAME", activation function is ReLU, output feature map dimension is 16×4;

[0095] Pooling layer: Average pooling is used, with a kernel size of 2x1, a stride of 2, and an output feature map dimension of 16×2;

[0096] Fully connected layer 1: Input dimension is 16×2=32, number of neurons is 32, and activation function is ReLU;

[0097] Fully connected layer 2 (FC): It has 1 neuron and outputs the final virtual measurement value y of the solder area.

[0098] The initial parameters of the model are assigned independently and randomly, without inheriting the parameters of the first-level CNN, to ensure training independence.

[0099] Using the second-level training set as input, the optimization objective is to minimize the MSE between the final predicted value output by the second-level CNN prediction model and the actual solder area. The number of training iterations is set to 150, the learning rate is 0.0005, and the optimizer is Adam. During training, the parameters of the first-level CNN prediction model are kept fixed, and only the parameters of the second-level CNN prediction model are updated to avoid cross-stage parameter interference.

[0100] The virtual satellite measurement performance is evaluated using a secondary test set, and the final prediction results are required. If the prediction accuracy ACC is greater than 98%, and the target is not met, the size of the convolution kernel of the second-level CNN prediction model (e.g., from 2x1 to 3x1) or the number of neurons in the fully connected layer (±8) should be adjusted and retrained until the performance target is met.

[0101] This invention takes the initial prediction value output by the first-level pre-trained model (first-level CNN prediction model) as input, learns the global correlation between multiple parameters (such as the synergistic effect of energy and power), optimizes the final virtual measurement accuracy of solder area, and achieves accurate mapping of "multi-parameter synergy - solder area".

[0102] Step S8: Predict the area of ​​ultrasonic welding stamps in battery cells using a trained secondary CNN prediction model.

[0103] The present invention provides a method for predicting the weld area of ​​ultrasonic welding of battery cells. This method reconstructs multi-dimensional time-series data during the welding process using an autoencoder, ensuring the integrity of the temporal dynamic features and providing high-quality input for subsequent processing. It employs a two-layer architecture: a first-level CNN prediction model and a second-level CNN prediction model. The first-level CNN prediction model is independently pre-trained for single-dimensional time-series data, mining local temporal features and outputting preliminary prediction values. The second-level CNN prediction model uses the prediction results of the first-level CNN prediction model as input, learns multi-parameter global correlations, and outputs the final prediction value. This addresses the problem of insufficient feature mining in traditional single-model methods, effectively improving prediction accuracy. Furthermore, the prediction method provided by this invention performs synchronous prediction based on multi-dimensional time-series data during the welding process. Compared to post-screening methods, it offers better timeliness, enabling timely warnings when problems are detected and reducing losses.

[0104] After achieving the prediction of the weld area during ultrasonic welding of battery cells, further steps can be taken:

[0105] 1) Virtual measurement result determination: Based on the cell model, the default solder area range is preset (e.g., the standard solder area for a certain cell model is 104). The acceptable range is set at [72.8]. 104 The virtual measurement output by the second-level CNN prediction model will be used to... The system automatically outputs the quality judgment result after comparing the result with the acceptable range. (If within the range, it is OK; if outside, it is NG).

[0106] 2) Anomaly Warning and Traceability: When an anomaly is detected as NG, an audible and visual warning is automatically triggered, and an "Anomaly Traceability Report" is generated. This report includes the reconstruction curves of the cell's four timing parameters, a deviation analysis of the preliminary prediction values ​​of each first-level CNN and the final prediction values ​​of the second-level CNN, and identifies key parameters that may lead to the anomaly (such as...). and If the deviation exceeds 10%, the energy time series is marked as a suspected anomaly.

[0107] 3) Data push and storage: Virtual measurement values, judgment results, and anomaly reports are pushed to the production line MES system in real time and stored in the time series database to support historical data backtracking (such as querying the virtual measurement error distribution of NG cells in the past 7 days) and process optimization analysis.

[0108] 4) Individual update of the first-level CNN prediction model: When the acquisition device of a certain time series parameter is replaced (such as power sensor upgrade) or the parameter distribution changes significantly (such as energy time series peak drift exceeding 10%), it is only necessary to extract the sample corresponding to the parameter from the historical data and retrain the corresponding first-level CNN prediction model. There is no need to adjust other first-level models, and the update cycle is less than 1 hour.

[0109] 5) Collaborative fine-tuning of the second-level CNN model: After the first-level model is updated, or every 1000 new production samples are accumulated, the second-level CNN prediction model is fine-tuned using only the 4-dimensional feature vector generated by the new samples (50 iterations, 0.0001 learning rate). After fine-tuning, the model performance is restored to its previous state, and the adaptation cycle is shortened by 80% compared to full retraining.

[0110] 6) Rapid adaptation to multiple production lines: To address the differences in welding machine models across different production lines, transfer learning is used to take the parameters of the already trained first- and second-level CNN prediction models as initial values. Only 200 samples from the target production line need to be input for fine-tuning to complete the adaptation. The deployment cycle of the new production line model is shortened from the original 4 hours to 1 hour.

[0111] By establishing a dynamic mechanism of updating the first-level model independently and fine-tuning the second-level model collaboratively, when the acquisition device for a certain time-series parameter is replaced or its distribution changes, only the corresponding first-level CNN needs to be retrained. The second-level CNN can be adapted with a small number of new samples and fine-tuning, which greatly shortens the working condition adaptation cycle of the virtual measurement model.

[0112] Example 2

[0113] This invention provides a battery cell ultrasonic welding solder mark area prediction system, comprising:

[0114] The data acquisition module is configured to acquire multi-dimensional time-series data during the welding process and the weld area after welding.

[0115] The data reconstruction module is configured to use an autoencoder to reconstruct multi-dimensional time-series data and generate multi-dimensional reconstructed time-series data.

[0116] The first-level sample set module is configured to construct first-level sample sets for each dimension based on the reconstructed time-series data and the weld area after welding.

[0117] The first-level model training module is configured to build a first-level CNN prediction model for each dimension of the first-level sample set, and train the corresponding first-level CNN prediction model using the first-level sample set.

[0118] The first-level model prediction module is configured to input the reconstructed time series data of each dimension into the corresponding trained first-level CNN prediction model to generate the predicted values ​​of the solder area of ​​each dimension.

[0119] The secondary sample set module is configured to combine the predicted solder area values ​​of each dimension into a multi-dimensional feature vector, and to construct a secondary sample set based on the multi-dimensional feature vector and the solder area after welding.

[0120] The secondary model training module is configured to build a secondary CNN prediction model and train the secondary CNN prediction model using a secondary sample set;

[0121] The secondary model prediction module is configured to predict the weld area of ​​the ultrasonic welding of the battery cell using a trained secondary CNN prediction model.

[0122] Example 3

[0123] Based on the method for predicting the solder area of ​​ultrasonic welding of battery cells provided in Embodiment 1, this embodiment of the invention provides an electronic device, including a processor and a storage medium;

[0124] Storage media are used to store instructions;

[0125] The processor is used to perform operations according to instructions to execute the steps according to the method described above.

[0126] Example 4

[0127] Based on the method for predicting the solder area of ​​ultrasonic welding of battery cells provided in Embodiment 1, this embodiment of the invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above method.

[0128] Example 5

[0129] Based on the method for predicting the solder area of ​​ultrasonic welding of battery cells provided in Embodiment 1, this embodiment of the invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above method.

[0130] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0134] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the area of ​​ultrasonic welding marks in battery cells, characterized in that, include: Acquire multi-dimensional time-series data during the welding process and the weld area after welding; An autoencoder is used to reconstruct the multi-dimensional time series data to generate multi-dimensional reconstructed time series data; The reconstructed time series data and the weld area after welding are used to construct the first-level sample set for each dimension. A first-level CNN prediction model is constructed for each dimension of the first-level sample set, and the corresponding first-level CNN prediction model is trained using the first-level sample set; The reconstructed time series data of each dimension are input into the corresponding trained first-level CNN prediction model to generate the predicted solder area values ​​of each dimension. The predicted solder area values ​​of each dimension are combined into a multi-dimensional feature vector, and a secondary sample set is formed based on the multi-dimensional feature vector and the solder area after welding. Construct a two-level CNN prediction model and train the two-level CNN prediction model using the two-level sample set; The trained secondary CNN prediction model is used to predict the weld area of ​​the ultrasonic welding of battery cells.

2. The method for predicting the weld area of ​​ultrasonic welding of battery cells according to claim 1, characterized in that, The multi-dimensional time-series data includes time-series data of welding energy, welding pressure, welding amplitude, and welding power.

3. The method for predicting the weld area of ​​ultrasonic welding of battery cells according to claim 1, characterized in that, The autoencoder includes an encoder and a decoder; The encoder uses a three-layer LSTM network to downsample the input time-series data, extracts deep time-series features through the ReLU activation function, and outputs a feature encoding vector of the target dimension. The decoder uses a three-layer LSTM network to upsample the feature encoding vector of the target dimension, and then outputs the reconstructed temporal data through a fully connected layer. The mean squared error is used as the loss function, and the network parameters of the autoencoder are optimized with the goal of minimizing the error between the time series data and the reconstructed time series data.

4. The method for predicting the weld area of ​​ultrasonic welding of battery cells according to claim 1, characterized in that, The reconstruction process of the multi-dimensional time-series data using an autoencoder includes: Normalize the time series data for each dimension; For the normalized time series data, the autoencoder is used to adaptively fill in missing values ​​to complete the missing value repair, and the autoencoder is used to generate reconstructed time series data of target length to achieve length standardization.

5. The method for predicting the weld area of ​​ultrasonic welding of battery cells according to claim 1, characterized in that, Cells with solder area within the acceptable range are designated as OK cells, while cells outside the acceptable range are designated as NG cells. The ratio of OK cells to NG cells corresponding to the solder area in the primary sample set for each dimension is 1:

1.

6. The method for predicting the weld area of ​​ultrasonic welding of battery cells according to claim 1, characterized in that, The initial parameters of both the first-level CNN prediction model and the second-level CNN prediction model are generated independently and randomly.

7. A system for predicting the area of ​​ultrasonic welding marks on battery cells, characterized in that, include: The data acquisition module is configured to acquire multi-dimensional time-series data during the welding process and the weld area after welding. The data reconstruction module is configured to use an autoencoder to reconstruct the multi-dimensional time-series data and generate multi-dimensional reconstructed time-series data. The first-level sample set module is configured to construct first-level sample sets for each dimension based on the reconstructed time-series data and the weld area after welding. The first-level model training module is configured to construct a first-level CNN prediction model for each dimension of the first-level sample set, and train the corresponding first-level CNN prediction model using the first-level sample set. The first-level model prediction module is configured to input the reconstructed time series data of each dimension into the corresponding trained first-level CNN prediction model to generate the predicted values ​​of the solder area of ​​each dimension. The secondary sample set module is configured to form a multi-dimensional feature vector from the predicted solder area values ​​of each dimension, and to construct a secondary sample set based on the multi-dimensional feature vector and the solder area after welding. The secondary model training module is configured to construct a secondary CNN prediction model and train the secondary CNN prediction model using the secondary sample set; The secondary model prediction module is configured to predict the weld area of ​​the ultrasonic welding of the battery cell using the trained secondary CNN prediction model.

8. An electronic device, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.

Images