Lithium battery recycling dangerous gas concentration prediction method based on convolution attention and residual time convolution network

Abstract

This invention discloses a method for predicting hazardous gas concentrations in lithium battery recycling based on convolutional attention and residual temporal convolutional networks. The method includes: 1. Collecting and preprocessing multi-channel hazardous gas sensor response data during lithium battery recycling to construct a concentration prediction dataset; 2. Establishing a hazardous gas concentration prediction network, utilizing a convolutional attention feature extraction layer to enhance key time locations and important channel information, and extracting local temporal features and global correlation features through temporal attention residual temporal convolutional blocks, lightweight self-attention encoding blocks, and gated fusion units. The concentration prediction results are obtained through a temporal feature extraction layer and a regression output layer; 3. Constructing a weighted mean square error loss function between the predicted concentration and the actual concentration, and obtaining the optimal prediction model through training iterations. This invention can improve the accuracy and stability of predicting different hazardous gas concentrations, meeting the needs of gas monitoring and safety early warning at lithium battery recycling sites.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery safety monitoring technology, specifically a method for predicting the concentration of hazardous gases in lithium battery recovery based on convolutional attention and residual temporal convolutional networks. Background Technology

[0002] With the rapid development of new energy vehicles, energy storage power stations, and portable electronic devices, the number of retired lithium batteries continues to increase, making lithium battery recycling a crucial link in resource recycling and environmental safety management. During the dismantling, crushing, screening, and heat treatment of lithium batteries, residual electrolytes and organic components inside the batteries may volatilize or decompose, generating hazardous gases. These gases are toxic, flammable, or explosive, and if they accumulate at the recycling site, they can easily threaten personnel safety and equipment operation. Therefore, effective monitoring and prediction of their concentration is of great significance.

[0003] Currently, lithium battery recycling sites primarily rely on gas sensors for concentration detection and alarm triggering based on set thresholds. While this method is structurally simple, it typically relies on data from a single moment, making it difficult to reflect dynamic trends in gas concentration. In situations of rapid gas changes or mixing of multiple gases, it is prone to response lag or misjudgments. Furthermore, in practical applications, sensors are susceptible to changes in temperature and humidity, drift, and cross-sensitivity, and the complex coupling relationships between multi-channel responses introduce significant uncertainty into the detection results.

[0004] The key shortcomings of existing technologies are twofold: firstly, they lack effective utilization of multi-channel sensor response sequences, making it difficult to fully extract the local variation characteristics and dynamic evolution laws of sensor responses during hazardous gas release; secondly, in the presence of multiple hazardous gases, response coupling and cross-sensitivity occur between different sensor channels, making it difficult for traditional methods to accurately establish the nonlinear mapping relationship between multi-channel sensor responses and the concentrations of multiple gases. Furthermore, while some deep learning methods can model sensor time-series data, they are still susceptible to redundant responses, noise disturbances, and channel differences under complex operating conditions, making it difficult to stably capture response information that plays a crucial role in gas concentration prediction, resulting in insufficient prediction accuracy and robustness. Therefore, a hazardous gas concentration prediction method is needed that can fully extract local response features, enhance key response information, and strengthen time-series modeling capabilities from multi-sensor response data at lithium battery recycling sites, thereby improving the accuracy, stability, and engineering application value of hazardous gas concentration prediction. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a method for predicting hazardous gas concentrations in lithium battery recycling based on convolutional attention and residual temporal convolutional networks. The aim is to automatically predict hazardous gas concentration changes based on multi-sensor response data during the lithium battery recycling process, thereby improving the accuracy and stability of hazardous gas concentration prediction. This provides data support for gas concentration identification, abnormal state judgment, and safety early warning at the recycling site, enhancing the real-time performance and engineering practical value of hazardous gas monitoring.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting the concentration of hazardous gases in lithium battery recovery based on convolutional attention and residual temporal convolutional networks, characterized by the following steps: Step 1: Collect response data from multi-channel hazardous gas sensors and their corresponding real gas concentration values ​​during the lithium battery recycling process, and preprocess the data to construct a hazardous gas concentration prediction dataset. ;in, Indicates the first preprocessed... n The six-channel sensor response matrix of a sample. Indicates the first preprocessed... n The vector of real gas concentrations corresponding to each sample N Indicates the total number of samples; Step 2: Construct a hazardous gas concentration prediction network, including: a convolutional attention feature extraction layer (ConvAttLayer), a hybrid temporal feature enhancement block (Hybrid Block), a temporal feature extraction layer (Temporal), and a regression output layer (HeadRegression). Processing is performed to obtain Corresponding hazardous gas concentration prediction vector ; Step 3: Based on and The weighted mean square error loss function is constructed using equation (1). L : (1) In equation (1), Indicates the types and quantities of hazardous gases. Indicates the first g Loss weights corresponding to different hazardous gases; express The Middle g Predicted concentrations of several hazardous gases express The Middle g The actual concentration of the hazardous gas; Step 4: Train the hazardous gas concentration prediction network using the backpropagation algorithm, and based on... Iteratively update network parameters, when Furthermore, training stops when the validation set evaluation index reaches its optimum, thus obtaining the optimal hazardous gas concentration prediction model, which is used for online prediction of hazardous gas concentration under multi-sensor response at lithium battery recycling sites.

[0007] The method for predicting hazardous gas concentrations in lithium battery recovery based on convolutional attention and residual temporal convolutional networks, as described in this invention, is characterized in that step 2 includes the following steps: Step 2.1: The ConvAttLayer, a convolutional attention feature extraction layer, includes: convolutional units, temporal attention (TA) units, channel attention (SE) units, residual fusion units, and pooling sequence transformation units. Processing is performed to obtain the first... n Each self-convolutional temporal feature ; Step 2.2: The Hybrid Block for Temporal Feature Enhancement includes: Temporal Attention Residual Temporal Convolutional Block (TA-TCN), Lightweight Self-Attention Encoding Block (Mini Transformer), and Gate-Gated Fusion Unit, and performs... The input is processed to obtain the first... n A hybrid temporal enhancement feature ; Step 2.3: The Temporal feature extraction layer includes: one-dimensional convolutional units, residual TCN units, attention pooling units, and adaptive average pooling units, and performs... Processing is performed to obtain the first... n A fusion pooling feature ; Step 2.4: Head Regression of the Regression Output Layer uses Equation (2) to... Processing is performed to obtain Corresponding hazardous gas concentration prediction vector : (2) In equation (2), and This represents the weight matrix and bias terms of the first linear layer. and This represents the weight matrix and bias terms of the second linear layer. This represents the activation function. Dropout This indicates a random deactivation operation.

[0008] Furthermore, step 2.1 includes the following steps: Step 2.1.1: The convolutional unit pairs Perform convolution processing to extract the first... n The local response characteristics of the multi-channel hazardous gas sensor response data in the sample were obtained to obtain the first... n Convolutional features ; Step 2.1.2: The TA time attention unit is calculated. The importance weights of different time positions are used to determine the importance of different time positions. By weighting over the time dimension, we obtain the first... n Time attention enhancement features ; Step 2.1.3: Calculation of the SE channel attention unit The importance weights of different feature channels are used to determine the importance of different feature channels. Perform channel recalibration to obtain the first n Individual channel enhancement features ; Step 2.1.4: The residual fusion unit will... and Residual fusion is performed, and the pooling sequence transformation unit uses equation (3) to perform dimensionality reduction and feature rearrangement on the fused features to obtain the first... n Convolutional attention temporal features : (3) In equation (3), MaxPool This indicates a max pooling operation. Reshape This indicates a tensor reshaping operation.

[0009] Furthermore, step 2.2 includes the following steps: Step 2.2.1: The Temporal Attention Residual Temporal Convolutional Block TA-TCN utilizes equation (4) to... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics Perform residual temporal convolution processing to obtain the first... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics Thus, the first The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics : (4) In equation (4), Indicates the first The dilation rate of each residual temporal convolutional layer Indicates the expansion rate One-dimensional temporal convolution, BN Indicates batch normalization, Represents the activation function, when season ; Step 2.2.2: The lightweight self-attention encoding block Mini Transformer will... Mapped to query matrices respectively Key matrix Sum matrix Then, by utilizing the self-attention mechanism... Encoding is performed to obtain global temporal correlation features. ; Step 2.2.3: The Gate-gated fusion unit uses equation (5) to... and By merging, we obtain the first... n A hybrid temporal enhancement feature : (5) In equation (5), Indicates the first n Each gating weight.

[0010] Furthermore, step 2.3 includes the following steps: Step 2.3.1: The one-dimensional convolutional unit is processed using equations (6) and (7) respectively. Temporal feature extraction is performed, and the corresponding result is obtained. n Intermediate convolutional features and the n Temporal convolutional features : (6) (7) In equations (6) and (7), Indicates the kernel size as a One-dimensional convolutional layer, Indicates the kernel size as b One-dimensional convolutional layer, Indicates batch normalization; Step 2.3.2: Extraction of Residual TCN Units Deep temporal features The data are then processed in the attention pooling unit and the adaptive average pooling unit, respectively, to obtain the corresponding results. n Attention pooling features and the n Average pooling features Thus and Concatenate them along the feature dimension to form the first n A fusion pooling feature .

[0011] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0012] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention proposes a multi-channel hazardous gas concentration prediction method for lithium battery recycling scenarios. By constructing a mapping relationship between multi-sensor response sequences and multi-gas concentrations, it achieves joint prediction of multiple hazardous gases such as CO, VOC, and NH3. This solves the problem in existing technologies that rely solely on single-point detection or threshold alarms and cannot predict gas concentration changes in advance. It enables not only real-time monitoring of hazardous gas concentrations but also advance prediction of their changing trends, thereby improving prediction accuracy and response timeliness.

[0014] 2. This invention constructs a unified feature modeling framework that integrates convolutional attention feature extraction, temporal attention residual temporal convolution, and lightweight self-attention encoding. It cleverly integrates the advantages of convolutional networks in extracting local response features, the ability of temporal attention mechanisms to enhance response information in key time segments, the ability of channel attention mechanisms to strengthen important feature channels, and the ability of self-attention mechanisms to model long-range temporal dependencies. This design strategy not only effectively compensates for the shortcomings of traditional single feature extraction methods in characterizing key response information in complex gas response sequences, but also overcomes the limitations of single temporal models in local feature extraction, channel difference expression, and global correlation modeling. Simultaneously, through residual temporal convolution, gated fusion, and fusion pooling strategies, it takes into account local response changes, key time segments, important channel information, and overall change trends, enabling the model to have stronger feature expression capabilities under complex mixed gas conditions and significantly improving prediction robustness in environments with noise interference, channel coupling, and sensor response fluctuations.

[0015] 3. This invention constructs a multi-gas joint regression output and weighted loss function mechanism, which enables the model to be optimized for the importance and range differences of different gases, thereby improving the overall accuracy and practicality of multi-gas concentration prediction. Attached Figure Description

[0016] Figure 1This is an overall flowchart of the method for predicting the concentration of hazardous gases in lithium battery recovery based on convolutional attention and residual temporal convolutional networks, as described in this invention. Figure 2 This is a schematic diagram of the temporal attention residual temporal convolutional block TA-TCN of the present invention; Figure 3 This is a schematic diagram of the structure of the lightweight self-attention coding block Mini Transformer of the present invention. Detailed Implementation

[0017] In this embodiment, a method for predicting hazardous gas concentrations in lithium battery recycling based on convolutional attention and residual temporal convolutional networks is presented. This method primarily establishes a gas concentration prediction model using multi-sensor response data during the lithium battery recycling process. It utilizes a convolutional attention feature extraction layer to extract local response features, key temporal response information, and important channel features from the sensor response sequences. Furthermore, it combines residual temporal convolution, self-attention encoding, and fusion pooling to model the gas response sequences, thereby improving the accuracy and stability of hazardous gas concentration predictions such as CO, VOCs, and NH3. Specifically, as... Figure 1 As shown, the method includes the following steps: Step 1: Collect response data from multi-channel hazardous gas sensors and their corresponding real gas concentration values ​​during the lithium battery recycling process, and preprocess the data to construct a hazardous gas concentration prediction dataset. ;in, Indicates the first preprocessed step n The six-channel sensor response matrix of a sample. Indicates the first preprocessed step n The vector of real gas concentrations corresponding to each sample N This represents the total number of samples.

[0018] Step 2: Construct a hazardous gas concentration prediction network, including: a convolutional attention feature extraction layer (ConvAttLayer), a hybrid temporal feature enhancement block (Hybrid Block), a temporal feature extraction layer (Temporal), and a regression output layer (HeadRegression). Processing is performed to obtain Corresponding hazardous gas concentration prediction vector .

[0019] Step 2.1: The ConvAttLayer, a convolutional attention feature extraction layer, includes: convolutional units, temporal attention (TA) units, channel attention (SE) units, residual fusion units, and pooling sequence transformation units. Processing is performed to obtain the first... n Each self-convolutional temporal feature ; Step 2.1.1: The convolutional unit pairs Perform convolution processing to extract the first... n The local response characteristics of the multi-channel hazardous gas sensor response data in the sample were obtained to obtain the first... n Convolutional features ; Step 2.1.2: The TA time attention unit is calculated. The importance weights of different time positions are used to calculate the importance weights using equation (1). By weighting over the time dimension, we obtain the first... n Time attention enhancement features .

[0020] (1) In equation (1), , and This represents the learnable parameters in the TA temporal attention unit. T Indicates transpose. This indicates that normalization is performed along the time dimension. This represents the hyperbolic tangent activation function. This indicates the weight allocation.

[0021] Step 2.1.3: The SE channel attention unit first... Global average pooling is performed on each feature channel to obtain the global response value of each channel. Then, the importance weight corresponding to each feature channel is calculated through a fully connected layer and a sigmoid function. Subsequently, each importance weight is multiplied by... Finally, the weighted channels are recombined, and the channel enhancement features are obtained using equation (2). : (2) In equation (2), Indicates global average pooling. , This represents the weight matrix of the fully connected layer. , Indicates the bias term. This represents the activation function. This represents the Sigmoid function.

[0022] Step 2.1.4: The residual fusion unit uses equation (3) to combine the convolutional features With channel enhancement features The features are fused, and the fused features are then subjected to dimensionality reduction and feature rearrangement using a pooling sequence transformation unit to obtain the first... n Convolutional attention temporal features : (3) In equation (3), MaxPool This indicates a max pooling operation. Reshape This indicates a tensor reshaping operation.

[0023] Step 2.2: The Hybrid Block for Temporal Feature Enhancement includes: Temporal Attention Residual Temporal Convolutional Block (TA-TCN), Lightweight Self-Attention Encoding Block (Mini Transformer), and Gate-Gated Fusion Unit, and performs... The input is processed to obtain the first... n A hybrid temporal enhancement feature .

[0024] Step 2.2.1: Temporal Attention Residual Temporal Convolutional Block (TA-TCN), such as... Figure 2 As shown, it includes a Temporal Convolutional Network (TCN), a Temporal Attention mechanism, layer normalization, and a feedforward network. Equation (4) is used to adjust the... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics Residual temporal convolution processing is performed, where TCN uses one-dimensional temporal convolution with dilation rates of (1,2,4,8) to extract local temporal variation features under different receptive fields, resulting in the th... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics Furthermore, the response features at key time locations are weighted and enhanced using the TemporalAttention mechanism, thereby obtaining the first... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics : (4) In equation (4), Indicates the first The dilation rate of each residual temporal convolutional layer Indicates the expansion rate One-dimensional temporal convolution, BN Indicates batch normalization, This represents the activation function. Dropout Indicates a random deactivation operation, when season .

[0025] Step 2.2.2: Lightweight self-attention encoding block Mini Transformer such as Figure 3As shown, it includes a multi-head self-attention layer, residual connections, layer normalization, and a feedforward network to extract global temporal correlation features between different time positions. In this embodiment, the number of attention heads in the multi-head self-attention layer is set to 2, and the hidden layer dimension of the feedforward network is set to 64, to extract global temporal correlation features between different time positions in the gas response sequence with lower computational complexity. Mapped to query matrices respectively Key matrix Sum matrix Then, by utilizing the self-attention mechanism... Encode the data and use Equation (5) to obtain the global temporal correlation features. : (5) In equation (5), Represents the dimension of the key vector. Representation layer normalization, This represents a feedforward network.

[0026] Step 2.2.3: The Gate-gated fusion unit determines the input features based on the given characteristics. Calculate the first using equation (6) n individual gating weights This is used to adaptively adjust the fusion ratio of local temporal variation features and global temporal correlation features: (6) In equation (6), This indicates a global average pooling operation. Indicates the gating mapping weights, This indicates the bias term.

[0027] The Gate gate fusion unit utilizes formula (7) to... and By merging, we obtain the first... n A hybrid temporal enhancement feature : (7) Step 2.3: The Temporal feature extraction layer includes: one-dimensional convolutional units, residual TCN units, attention pooling units, and adaptive average pooling units, and performs... Processing is performed to further extract deep temporal dependencies, and the variable-length temporal features are aggregated into fixed-dimensional feature representations to obtain the first... n A fusion pooling feature .

[0028] Step 2.3.1: One-dimensional convolutional units are processed using equations (8) and (9) respectively. Temporal feature extraction is performed, and the corresponding result is obtained. n Intermediate convolutional features and the n Temporal convolutional features In this embodiment, the kernel size... a Set to 5, kernel size b Setting it to 3 allows for the initial extraction of local response changes within a wider time window, followed by further extraction of short-term fluctuation features. (8) (9) In equations (8) and (9), Indicates the kernel size as a One-dimensional convolutional layer, Indicates the kernel size as b One-dimensional convolutional layer, This represents the activation function. This indicates batch normalization.

[0029] Step 2.3.2: Extraction of Residual TCN Units Deep temporal features This is used to further expand the temporal receptive field based on the local response features extracted by one-dimensional convolution, in order to obtain a more stable deep temporal representation. The results are then processed by the attention pooling unit and the adaptive average pooling unit, respectively, to obtain the corresponding... n Attention pooling features and the n Average pooling features Attention pooling features are used to characterize key time segments that contribute significantly to gas concentration prediction, while average pooling features are used to characterize the overall change level of the entire response sequence; the two complement each other. Thus, equation (10) is used to... and Concatenate them along the feature dimension to form the first n A fusion pooling feature : (10) In equation (10), C express and Feature dimensions, express The Middle Each feature component express The Middle Each feature component express The Middle Each feature component.

[0030] Step 2.4: Head Regression of the Regression Output Layer uses Equation (11) to... Processing is performed to obtain Corresponding hazardous gas concentration prediction vector : (11) In equation (11), and This represents the weight matrix and bias terms of the first linear layer. and This represents the weight matrix and bias terms of the second linear layer.

[0031] Step 3: Reduce the impact of differences in concentration ranges of different hazardous gases on model training, and enhance the constraint effect of prediction errors of high-risk gas concentrations on the training process. First, based on the... g The concentration range and safety risk coefficient of each hazardous gas are used to determine its corresponding loss weight using equation (12). : (12) In equation (12), Indicates the types and quantities of hazardous gases. Indicates the first g Safety risk coefficient of the hazardous gas Indicates the first g Concentration range of the hazardous gas This represents a constant used to prevent the denominator from being zero.

[0032] based on and The weighted mean square error loss function is constructed using equation (13). L : (13) In equation (13), express The Middle g Predicted concentrations of several hazardous gases express The Middle g The actual concentration of a hazardous gas; in this embodiment, when the concentration range of a certain type of hazardous gas is small or the safety risk coefficient is high, its corresponding loss weight is... The increased size allows the model training process to take into account the differences in range and safety risks of different gases, thereby improving the stability of multi-gas joint prediction.

[0033] Step 4: Train the hazardous gas concentration prediction network using the backpropagation algorithm, and based on... Iteratively update network parameters, when Furthermore, training stops when the validation set evaluation index reaches its optimum, thus obtaining the optimal hazardous gas concentration prediction model, which is used for online prediction of hazardous gas concentration under multi-sensor response at lithium battery recycling sites.

[0034] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0035] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A method for predicting hazardous gas concentrations in lithium battery recovery based on convolutional attention and residual temporal convolutional networks, characterized in that, Includes the following steps: Step 1: Collect response data from multi-channel hazardous gas sensors and their corresponding real gas concentration values ​​during the lithium battery recycling process, and preprocess the data to construct a hazardous gas concentration prediction dataset. ;in, Indicates the first preprocessed... n The six-channel sensor response matrix of a sample. Indicates the first preprocessed... n The vector of real gas concentrations corresponding to each sample N Indicates the total number of samples; Step 2: Construct a hazardous gas concentration prediction network, including: a convolutional attention feature extraction layer (ConvAttLayer), a hybrid temporal feature enhancement block (Hybrid Block), a temporal feature extraction layer (Temporal), and a regression output layer (Head Regression). Processing is performed to obtain Corresponding hazardous gas concentration prediction vector ; Step 3: Based on and The weighted mean square error loss function is constructed using equation (1). L : (1) In equation (1), Indicates the types and quantities of hazardous gases. Indicates the first g Loss weights corresponding to different hazardous gases; express The Middle g Predicted concentrations of several hazardous gases express The Middle g The actual concentration of the hazardous gas; Step 4: Train the hazardous gas concentration prediction network using the backpropagation algorithm, and based on... Iteratively update network parameters, when Furthermore, training stops when the validation set evaluation index reaches its optimum, thus obtaining the optimal hazardous gas concentration prediction model, which is used for online prediction of hazardous gas concentration under multi-sensor response at lithium battery recycling sites.

2. The method for predicting hazardous gas concentrations in lithium battery recovery based on convolutional attention and residual temporal convolutional networks according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: The ConvAttLayer, a convolutional attention feature extraction layer, includes convolutional units, temporal attention (TA) units, channel attention (SE) units, residual fusion units, and pooling sequence transformation units. Processing is performed to obtain the first... n Each self-convolutional temporal feature ; Step 2.2: The Hybrid Block for Temporal Feature Enhancement includes: Temporal Attention Residual Temporal Convolutional Block (TA-TCN), Lightweight Self-Attention Encoding Block (Mini Transformer), and Gate-Gated Fusion Unit, and performs... The input is processed to obtain the first... n A hybrid temporal enhancement feature ; Step 2.3: The Temporal feature extraction layer includes: one-dimensional convolutional units, residual TCN units, attention pooling units, and adaptive average pooling units, and performs... Processing is performed to obtain the first... n A fusion pooling feature ; Step 2.4: Head Regression of the Regression Output Layer uses Equation (2) to... Processing is performed to obtain Corresponding hazardous gas concentration prediction vector : (2) In equation (2), and This represents the weight matrix and bias terms of the first linear layer. and This represents the weight matrix and bias terms of the second linear layer. This represents the activation function. Dropout This indicates a random deactivation operation.

3. The method for predicting hazardous gas concentrations in lithium battery recovery based on convolutional attention and residual temporal convolutional networks according to claim 2, characterized in that, Step 2.1 includes the following steps: Step 2.1.1: The convolutional unit pairs Perform convolution processing to extract the first... n The local response characteristics of the multi-channel hazardous gas sensor response data in the sample were obtained to obtain the first... n Convolutional features ; Step 2.1.2: The TA time attention unit is calculated. The importance weights of different time positions are used to determine the importance of different time positions. By weighting over the time dimension, we obtain the first... n Time attention enhancement features ; Step 2.1.3: Calculation of the SE channel attention unit The importance weights of different feature channels are used to determine the importance of different feature channels. Perform channel recalibration to obtain the first n Individual channel enhancement features ; Step 2.1.4: The residual fusion unit will... and Residual fusion is performed, and the pooling sequence transformation unit uses equation (3) to perform dimensionality reduction and feature rearrangement on the fused features to obtain the first... n Convolutional attention temporal features : (3) In equation (3), MaxPool This indicates a max pooling operation. Reshape This indicates a tensor reshaping operation.

4. The method for predicting hazardous gas concentrations in lithium battery recovery based on convolutional attention and residual temporal convolutional networks according to claim 3, characterized in that, Step 2.2 includes the following steps: Step 2.2.1: The temporal attention residual temporal convolutional block TA-TCN uses equation (4) to process the first... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics Perform residual temporal convolution processing to obtain the first... The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics Thus, the first The output of the residual temporal convolutional layer is the first n Local temporal variation characteristics : (4) In equation (4), Indicates the first The dilation rate of each residual temporal convolutional layer Indicates the expansion rate One-dimensional temporal convolution, BN Indicates batch normalization, Represents the activation function, when season ; Step 2.2.2: The lightweight self-attention encoding block Mini Transformer will... Mapped to query matrices respectively Key matrix Sum matrix Then, by utilizing the self-attention mechanism... Encoding is performed to obtain global temporal correlation features. ; Step 2.2.3: The Gate-gated fusion unit uses equation (5) to... and By merging, we obtain the first... n A hybrid temporal enhancement feature : (5) In equation (5), Indicates the first n Each gating weight.

5. The method for predicting hazardous gas concentrations in lithium battery recovery based on convolutional attention and residual temporal convolutional networks according to claim 4, characterized in that, Step 2.3 includes the following steps: Step 2.3.1: The one-dimensional convolutional unit uses equations (6) and (7) respectively to... Temporal feature extraction is performed, and the corresponding result is obtained. n Intermediate convolutional features and the n Temporal convolutional features : (6) (7) In equations (6) and (7), Indicates the kernel size as a One-dimensional convolutional layer, Indicates the kernel size as b One-dimensional convolutional layer, Indicates batch normalization; Step 2.3.2: Extraction of Residual TCN Units Deep temporal features The data are then processed in the attention pooling unit and the adaptive average pooling unit, respectively, to obtain the corresponding results. n Attention pooling features and the n Average pooling features Thus and Concatenate them along the feature dimension to form the first n A fusion pooling feature .

6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-5, the processor being configured to execute the program stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-5.

Images