A high-efficiency yarn tension adjustment system

Abstract

The invention relates to the technical field of yarn tension regulation, and provides an efficient yarn tension regulation system; the system comprises a yarn tension regulation device and a central processing unit; in the central processing unit, the system uses a Pearson correlation coefficient and a Spearman correlation coefficient to perform linear and nonlinear analysis on comprehensive time series data, extracts high-dimensional time features through a TCN layer and a TM-BiDir-LSTM layer in a TCN-TM-BiDir-LSTM model, accurately predicts tension anomalies and outputs an abnormal warning probability value through TCN-TM-BiDir-LSTM model hyperparameters optimized by PSO-ISSA; through the above steps, the invention effectively improves the response speed and regulation accuracy of the yarn tension regulation system, enhances the adaptability to dynamic environments and tension anomalies, and provides an efficient and reliable tension control scheme for a textile production process.

Description

Technical Field

[0001] The invention relates to the technical field of tension regulation, in particular to a high-efficiency yarn tension regulation system. Background Art

[0002] Existing yarn tension adjustment systems are difficult to meet the production requirements for high precision and rapid response; existing tension adjustment methods usually fail to take into account the dynamic changes of tension fluctuations and environmental factors, and lack sensitivity to abnormal fluctuations during the adjustment process, resulting in low adjustment accuracy and unstable performance; traditional methods often use simple linear models when processing multidimensional time series data, ignoring the nonlinear change characteristics and multi-factor coupling relationship of yarn tension, and are unable to achieve fine adjustment and prediction of tension, affecting yarn quality and production efficiency; therefore, there is an urgent need for a yarn tension control system that can perform real-time analysis, multi-dimensional prediction, and efficient adjustment to improve the stability and product quality of the textile production process. Summary of the invention

[0003] The present invention provides an efficient yarn tension adjustment system, aiming to solve the problems in the prior art of low tension adjustment accuracy and poor response performance due to the insufficient processing capability of traditional algorithms for multi-dimensional time series data; the system adopts a PSO-ISSA algorithm to optimize the hyperparameter configuration of a TCN-TM-BiDir-LSTM model, thereby improving the model's adaptability and prediction accuracy to tension changes in complex environments; in addition, the system extracts time series features through a TCN layer, and further extracts long-term and short-term dependency features of yarn tension in combination with a TM-BiDir-LSTM layer, thereby ensuring the real-time and stability of tension adjustment; the scheme not only effectively improves the accuracy of yarn tension adjustment and the system response speed, but also enhances its sensitivity to abnormal tension fluctuations during the production process, thereby significantly improving production efficiency and yarn quality.

[0004] The present invention provides an efficient yarn tension adjustment system, which includes a yarn tension adjustment device and a central processing unit, wherein the central processing unit includes a data acquisition module, a data preprocessing module, a feature selection module, a prediction module and a hyperparameter optimization module;

[0005] The data acquisition module collects yarn tension time series data and environmental time series data to obtain comprehensive time series data;

[0006] The data preprocessing module uses an interpolation method to fill missing values ​​in the comprehensive time series data, uses an IQR method to remove outliers in the comprehensive time series data, and then performs normalization processing to obtain preprocessed comprehensive time series data;

[0007] The feature selection module performs linear analysis on the preprocessed comprehensive time series data through the Pearson correlation coefficient to obtain linear feature data, performs nonlinear analysis on the preprocessed comprehensive time series data through the Spearman correlation coefficient to obtain nonlinear feature data, and merges the linear feature data and the nonlinear feature data to obtain comprehensive time series feature data; the prediction module establishes a TCN-TM-BiDir-LSTM model, inputs the comprehensive time series feature data into the TCN-TM-BiDir-LSTM model, obtains an abnormal warning probability value, and performs abnormal alarm and shutdown protection according to the abnormal warning probability value. The TCN-TM-BiDir-LSTM model includes an input layer, a TCN layer, a TM-BiDir-LSTMc layer, a fully connected layer, and an output layer;

[0008] The hyperparameter optimization module optimizes the hyperparameters of the TCN-TM-BiDir-LSTM model through the PSO-ISSA algorithm to obtain an optimized hyperparameter configuration.

[0009] The present invention provides a high-efficiency yarn tension adjustment system, wherein the yarn tension adjustment device of the system comprises a spinning frame, a force sensor, an electric yarn guide, a tension adjustment roller, and a feedback sensor, wherein the force sensor is fixedly connected to the inlet end of the spinning frame, the electric yarn guide is fixed to the inner wall of the spinning frame, the tension adjustment roller is fixedly connected to the inner wall of the spinning frame, the feedback sensor is fixedly connected to the outlet end of the spinning frame, the inlet end of the spinning frame is provided with yarn, the yarn contacts the detection end of the force sensor, the yarn contacts the electric yarn guide and the upper wall of the tension adjustment roller, the yarn contacts the detection end of the feedback sensor, and the central processing unit is electrically connected to the force sensor, the electric yarn guide, the tension adjustment roller, and the feedback sensor.

[0010] Furthermore, the process of optimizing the hyperparameters of the TCN-TM-BiDir-LSTM model by the PSO-ISSA algorithm specifically includes the following steps:

[0011] Step S1: Initialize the TCN-TM-BiDir-LSTM model and set the initial hyperparameter configuration, train the TCN-TM-BiDir-LSTM model according to the initial hyperparameter configuration, obtain the training results, and evaluate the training results to obtain the initial fitness value;

[0012] Step S2: Initialize the PSO particle swarm, randomly generate particles, and initialize the particle speed and particle position; initialize the ISSA sparrow population, randomly generate explorers and followers, and set the initial explorer position and follower position;

[0013] Step S3: Calculate the fitness value of the particle according to the particle speed and particle position, and calculate the fitness value of the sparrow population according to the explorer position and follower position;

[0014] Step S4: By comparing the particle fitness values, find the optimal individual particle and obtain the global fitness value. Use the PSO particle update method to update the particle speed and particle position according to the particle fitness value and the global fitness value. Through iterative updates, generate the individual optimal position and the global optimal position, and gradually lock the particle optimization hyperparameter configuration. The formula used is as follows:

[0015] PSO particle velocity update formula:

[0016] ;

[0017] in, No. Particles in The speed of iterations, No. Particles in The speed of iterations, represents the inertia weight, represents the self-perception coefficient, represents the social cognition coefficient, and represents a random number in the interval [0,1], Indicates The individual best position of each particle, represents the global optimal position, Indicates The particle in The position of the iteration;

[0018] PSO particle position update formula:

[0019] ;

[0020] in, Indicates The particle in The position of the iteration;

[0021] Step S5: guiding the explorer position update according to the global optimal position, guiding the follower position update according to the explorer position update, and continuously optimizing the sparrow population fitness value by iteratively updating the explorer position and the follower position to obtain the sparrow optimization hyperparameter configuration;

[0022] Step S6: Collaborative information sharing is performed on the particle optimization hyperparameter configuration and the sparrow optimization hyperparameter configuration, and the optimized hyperparameter configuration is obtained through iterative training until the number of iterations is met.

[0023] Furthermore, the TCN layer in the prediction module specifically includes the following contents:

[0024] Step M1: receiving the comprehensive time series feature data of the input layer, performing channel transformation on the comprehensive time series feature data, and obtaining channel transformation feature data;

[0025] Step M2: Perform dilation convolution operation on the comprehensive time series feature data of the input layer to extract long-term and short-term dependency features;

[0026] Step M3: Standardize the long-term and short-term dependency features to obtain standard time feature data;

[0027] Step M4: nonlinearly activate the standard time feature data through the ReLU activation function to obtain nonlinear time feature data;

[0028] Step M5: Randomly discard some neurons of the TCN-TM-BiDir-LSTM model to obtain regularized time feature data;

[0029] Step M6: Add the channel transformation feature data and the regularized time feature data to obtain TCN comprehensive feature data.

[0030] Furthermore, the TM-BiDir-LSTM layer in the prediction module specifically includes the following contents:

[0031] Step B1: introducing a time mask matrix: reading TCN comprehensive feature data dimension information according to TCN comprehensive feature data, generating a time mask matrix according to the TCN comprehensive feature data dimension information, applying the time mask matrix, masking invalid time steps, and obtaining masked feature data;

[0032] Step B2: Forward LSTM traversal: Perform a forward LSTM traversal operation on the masked feature data from left to right to extract time-dependent information from the past to the future, and obtain forward traversal feature data;

[0033] Step B3: Backward LSTM traversal: Perform a backward LSTM traversal operation on the masked feature data from right to left to extract the time dependency information from the future to the past, and obtain the backward traversal feature data;

[0034] Step B4: forward and backward feature fusion: forward travel feature data and backward travel feature data are spliced ​​and fused to obtain comprehensive travel feature data. The formula used is as follows:

[0035] ;

[0036] in, Indicates Comprehensive travel feature data of time steps, Indicates The forward pass feature data for time steps, Indicates Backward travel feature data for time steps.

[0037] Furthermore, the fully connected layer in the prediction module maps the TCN comprehensive feature data and the comprehensive travel feature data to the output space of the TCN-TM-BiDir-LSTM model through the ReLU activation function to obtain the activated feature data.

[0038] Furthermore, in the output layer of the prediction module, the activated feature data is mapped to an abnormal warning probability value through a Sigmoid activation function, and the formula used is as follows:

[0039] ;

[0040] in, Indicates activation feature data, Indicates that the Sigmoid activation function is used to The output is mapped to the (0,1) interval.

[0041] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0042] The present invention optimizes the hyperparameter configuration of the TCN-TM-BiDir-LSTM model through the PSO-ISSA algorithm, thereby realizing accurate modeling and real-time response to complex time series data in the yarn tension regulation system; the PSO-ISSA algorithm effectively adjusts the model's learning rate, number of layers, number of nodes and other hyperparameters through a global and local dual optimization strategy, thereby significantly improving the sensitivity and adaptability of the TCN-TM-BiDir-LSTM model to tension changes; compared with the traditional static hyperparameter adjustment method, the present invention can dynamically adapt to environmental changes and tension fluctuations in the production process, thereby solving the problems of response lag and adjustment inaccuracy caused by poor hyperparameter configuration in the traditional adjustment method, thereby enhancing the stability and adjustment accuracy of the system.

[0043] In terms of feature extraction, the present invention adopts a combination of TCN layer and TM-BiDir-LSTM layer to perform multi-dimensional feature analysis and extraction, making full use of the TCN layer's ability to capture the dependencies of long time series and the TM-BiDir-LSTM layer's ability to respond to short-term fluctuations, effectively solving the problem of insufficient processing of nonlinear characteristics of yarn tension by traditional methods; the TCN layer first extracts the time series dependencies of the comprehensive time series feature data, and the TM-BiDir-LSTM layer further extracts the bidirectional time dependency information in the feature data, thereby achieving fine capture of long-term and short-term changes in yarn tension; the application of this technology improves the adaptability of the yarn tension adjustment system to dynamic environments, and ensures high precision and high response speed in the adjustment process.

[0044] By applying the above-mentioned optimization algorithm and feature extraction technology to the yarn tension adjustment system, the present invention achieves more efficient and accurate yarn tension control, effectively improving the tension stability and product quality in the textile production process; when facing a complex production environment and changeable tension conditions, the system can analyze and adjust the tension value in real time, solving the problem of adjustment lag and poor control accuracy caused by the traditional monitoring system's insensitivity to tension fluctuations; the system not only improves the work efficiency and product consistency of the production line, but also reduces the risk of yarn breakage and quality defects caused by abnormal tension, providing an advanced and reliable tension adjustment solution for the modern textile industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of the overall structure of an efficient yarn tension adjustment system proposed by the present invention;

[0046] Figure 2 This is a flowchart of the TM-BiDir-LSTMc layer of the TCN-TM-BiDir-LSTM model in the prediction module proposed in the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] Embodiment 1, the present invention provides an efficient yarn tension adjustment system, the system comprises a yarn tension adjustment device and a central processing unit, the central processing unit comprises a data acquisition module, a data preprocessing module, a feature selection module, a prediction module and a hyperparameter optimization module;

[0049] The data acquisition module collects yarn tension time series data and environmental time series data to obtain comprehensive time series data;

[0050] The data preprocessing module uses an interpolation method to fill missing values ​​in the comprehensive time series data, uses an IQR method to remove outliers in the comprehensive time series data, and then performs normalization processing to obtain preprocessed comprehensive time series data; the feature selection module performs a linear analysis on the preprocessed comprehensive time series data through a Pearson correlation coefficient to obtain linear feature data, performs a nonlinear analysis on the preprocessed comprehensive time series data through a Spearman correlation coefficient to obtain nonlinear feature data, and merges the linear feature data and the nonlinear feature data to obtain comprehensive time series feature data;

[0051] The prediction module establishes a TCN-TM-BiDir-LSTM model, inputs the comprehensive time series feature data into the TCN-TM-BiDir-LSTM model, obtains an abnormal warning probability value, and performs abnormal alarm and shutdown protection according to the abnormal warning probability value. The TCN-TM-BiDir-LSTM model includes an input layer, a TCN layer, a TM-BiDir-LSTMc layer, a fully connected layer and an output layer;

[0052] The hyperparameter optimization module optimizes the hyperparameters of the TCN-TM-BiDir-LSTM model through the PSO-ISSA algorithm to obtain an optimized hyperparameter configuration.

[0053] Embodiment 2, according to Figure 1 The present invention provides a high-efficiency yarn tension adjustment system, wherein the yarn tension adjustment device of the system includes a spinning frame, a force sensor, an electric yarn guide, a tension adjustment roller, and a feedback sensor. The force sensor is fixedly connected to the inlet end of the spinning frame, the electric yarn guide is fixed to the inner wall of the spinning frame, the tension adjustment roller is fixedly connected to the inner wall of the spinning frame, and the feedback sensor is fixedly connected to the outlet end of the spinning frame. The inlet end of the spinning frame is provided with yarn, the yarn contacts the detection end of the force sensor, the yarn contacts the electric yarn guide and the upper wall of the tension adjustment roller, the yarn contacts the detection end of the feedback sensor, and the central processing unit is electrically connected to the force sensor, the electric yarn guide, the tension adjustment roller and the feedback sensor.

[0054] Embodiment 3: This embodiment is based on the above embodiment, and the data acquisition module acquires yarn tension time series data and environmental time series data to obtain comprehensive time series data.

[0055] Yarn tension time series data include: real-time tension time series data, tension fluctuation time series data, historical tension time series data and tension abnormality event time series data;

[0056] The environmental time series data includes: temperature time series data, humidity time series data, air flow rate time series data, machine vibration time series data and tension regulator state time series data.

[0057] Embodiment 4, based on the above embodiment, the process of optimizing the hyperparameters of the TCN-TM-BiDir-LSTM model by the PSO-ISSA algorithm specifically includes the following steps:

[0058] Step S1: Initialize the TCN-TM-BiDir-LSTM model and set the initial hyperparameter configuration, train the TCN-TM-BiDir-LSTM model according to the initial hyperparameter configuration, obtain the training results, and evaluate the training results to obtain the initial fitness value;

[0059] Step S2: Initialize the PSO particle swarm, randomly generate particles, and initialize the particle speed and particle position; initialize the ISSA sparrow population, randomly generate explorers and followers, and set the initial explorer position and follower position

[0060] Step S3: Calculate the fitness value of the particle according to the particle speed and particle position, and calculate the fitness value of the sparrow population according to the explorer position and follower position;

[0061] Step S4: By comparing the particle fitness values, find the optimal individual particle and obtain the global fitness value. Use the PSO particle update method to update the particle speed and particle position according to the particle fitness value and the global fitness value. Generate the individual optimal position and the global optimal position through iterative update, and gradually lock the particle optimization hyperparameter configuration. The formula used is as follows:

[0062] PSO particle velocity update formula:

[0063] ;

[0064] in, No. Particles in The speed of iterations, No. Particles in The speed of iterations, represents the inertia weight, represents the self-perception coefficient, represents the social cognition coefficient, and represents a random number in the interval [0, 1], Indicates The individual best position of each particle, represents the global optimal position, Indicates The particle in The position of the iteration;

[0065] PSO particle position update formula:

[0066] ;

[0067] in, Indicates The particle in The position of the iteration;

[0068] Step S5: guiding the explorer position update according to the global optimal position, guiding the follower position update according to the explorer position update, and continuously optimizing the sparrow population fitness value by iteratively updating the explorer position and the follower position to obtain the sparrow optimization hyperparameter configuration;

[0069] Step S6: Collaborative information sharing is performed on the particle optimization hyperparameter configuration and the sparrow optimization hyperparameter configuration, and the optimized hyperparameter configuration is obtained through iterative training until the number of iterations is met.

[0070] Embodiment 5, based on the above embodiment, the TCN layer in the prediction module specifically includes the following contents:

[0071] Step M1: receiving the comprehensive time series feature data of the input layer, performing channel transformation on the comprehensive time series feature data, and obtaining channel transformation feature data;

[0072] Step M2: Perform dilation convolution operation on the comprehensive time series feature data of the input layer to extract long-term and short-term dependency features;

[0073] Step M3: Standardize the long-term and short-term dependency features to obtain standard time feature data;

[0074] Step M4: nonlinearly activate the standard time feature data through the ReLU activation function to obtain nonlinear time feature data;

[0075] Step M5: Randomly discard some neurons of the TCN-TM-BiDir-LSTM model to obtain regularized time feature data;

[0076] Step M6: Add the channel transformation feature data and the regularized time feature data to obtain TCN comprehensive feature data.

[0077] Embodiment 6, according to Figure 2 This embodiment is based on the above embodiment, and the TM-BiDir-LSTM layer in the prediction module specifically includes the following contents:

[0078] Step B1: introducing a time mask matrix: reading TCN comprehensive feature data dimension information according to TCN comprehensive feature data, generating a time mask matrix according to the TCN comprehensive feature data dimension information, applying the time mask matrix, masking invalid time steps, and obtaining masked feature data;

[0079] Step B2: Forward LSTM traversal: Perform a forward LSTM traversal operation on the masked feature data from left to right to extract time-dependent information from the past to the future, and obtain forward traversal feature data;

[0080] Step B3: Backward LSTM traversal: Perform a backward LSTM traversal operation on the masked feature data from right to left to extract the time dependency information from the future to the past, and obtain the backward traversal feature data;

[0081] Step B4: forward and backward feature fusion: forward travel feature data and backward travel feature data are spliced ​​and fused to obtain comprehensive travel feature data. The formula used is as follows:

[0082] ;

[0083] in, Indicates Comprehensive travel feature data of time steps, Indicates The forward pass feature data for time steps, Indicates Backward travel feature data for time steps.

[0084] Embodiment 7, this embodiment is based on the above embodiment, and the fully connected layer in the prediction module maps the TCN comprehensive feature data and the comprehensive travel feature data to the output space of the TCN-TM-BiDir-LSTM model through the ReLU activation function to obtain the activated feature data.

[0085] Embodiment 8, based on the above embodiment, the output layer in the prediction module maps the activated feature data to an abnormal warning probability value through a Sigmoid activation function, and the formula used is as follows:

[0086] ;

[0087] in, Indicates activation feature data, Indicates that the Sigmoid activation function is used to The output is mapped to the (0,1) interval.

[0088] Embodiment 9. This embodiment is based on the above embodiment. The prediction module establishes a TCN-TM-BiDir-LSTM model, inputs the comprehensive time series feature data into the TCN-TM-BiDir-LSTM model, obtains the abnormal warning probability value, and performs abnormal alarm and shutdown protection according to the abnormal warning probability value. The TCN-TM-BiDir-LSTM model includes an input layer, a TCN layer, a TM-BiDir-LSTMc layer, a fully connected layer and an output layer.

[0089] In this embodiment:

[0090] Abnormal warning probability value: 0.85, which is close to 1, and there is a high risk of abnormality;

[0091] Abnormal alarm: When the probability value is 0.85 and exceeds the set alarm threshold (0.7), the system triggers an abnormal alarm and the alarm light turns red and flashes;

[0092] Shutdown protection: If the warning probability value continues to exceed the safety threshold (0.8), the system will automatically shut down for protection.

Claims

1. An efficient yarn tension adjustment system, characterized in that: It includes a yarn tension adjustment device and a central processing unit, wherein the central processing unit includes a data acquisition module, a data preprocessing module, a feature selection module, a prediction module and a hyperparameter optimization module; The data acquisition module collects yarn tension time series data and environmental time series data to obtain comprehensive time series data; The data preprocessing module uses an interpolation method to fill missing values ​​in the comprehensive time series data, uses an IQR method to remove outliers in the comprehensive time series data, and then performs normalization processing to obtain preprocessed comprehensive time series data; The feature selection module performs linear analysis on the preprocessed comprehensive time series data through the Pearson correlation coefficient to obtain linear feature data, performs nonlinear analysis on the preprocessed comprehensive time series data through the Spearman correlation coefficient to obtain nonlinear feature data, and combines the linear feature data and the nonlinear feature data to obtain comprehensive time series feature data; The prediction module establishes a TCN-TM-BiDir-LSTM model, inputs the comprehensive time series feature data into the TCN-TM-BiDir-LSTM model, obtains an abnormal warning probability value, and performs abnormal alarm and shutdown protection according to the abnormal warning probability value. The TCN-TM-BiDir-LSTM model includes an input layer, a TCN layer, a TM-BiDir-LSTMc layer, a fully connected layer and an output layer; The hyperparameter optimization module optimizes the hyperparameters of the TCN-TM-BiDir-LSTM model through the PSO-ISSA algorithm to obtain an optimized hyperparameter configuration; The TCN layer in the prediction module specifically includes the following contents: Step M1: receiving the comprehensive time series feature data of the input layer, performing channel transformation on the comprehensive time series feature data, and obtaining channel transformation feature data; Step M2: Perform dilation convolution operation on the comprehensive time series feature data of the input layer to extract long-term and short-term dependency features; Step M3: Standardize the long-term and short-term dependency features to obtain standard time feature data; Step M4: nonlinearly activate the standard time feature data through the ReLU activation function to obtain nonlinear time feature data; Step M5: Randomly discard some neurons of the TCN-TM-BiDir-LSTM model to obtain regularized time feature data; Step M6: Add the channel transformation feature data and the regularized time feature data to obtain TCN comprehensive feature data; The TM-BiDir-LSTM layer in the prediction module specifically includes the following contents: Step B1: introducing a time mask matrix: reading TCN comprehensive feature data dimension information according to TCN comprehensive feature data, generating a time mask matrix according to the TCN comprehensive feature data dimension information, applying the time mask matrix, masking invalid time steps, and obtaining masked feature data; Step B2: Forward LSTM traversal: Perform a forward LSTM traversal operation on the masked feature data from left to right to extract time-dependent information from the past to the future, and obtain forward traversal feature data; Step B3: Backward LSTM traversal: Perform a backward LSTM traversal operation on the masked feature data from right to left to extract the time dependency information from the future to the past, and obtain the backward traversal feature data; Step B4: forward and backward feature fusion: forward travel feature data and backward travel feature data are spliced ​​and fused to obtain comprehensive travel feature data; The fully connected layer in the prediction module maps the TCN comprehensive feature data and the comprehensive travel feature data to the output space of the TCN-TM-BiDir-LSTM model through the ReLU activation function to obtain the activated feature data; The output layer in the prediction module maps the activated feature data to an abnormal warning probability value through a Sigmoid activation function.

2. The high-efficiency yarn tension adjustment system according to claim 1, characterized in that: The yarn tension adjustment device includes a spinning frame, a force sensor, an electric yarn guide, a tension adjustment roller, and a feedback sensor. The force sensor is fixedly connected to the inlet end of the spinning frame, the electric yarn guide is fixed to the inner wall of the spinning frame, the tension adjustment roller is fixedly connected to the inner wall of the spinning frame, and the feedback sensor is fixedly connected to the outlet end of the spinning frame. The inlet end of the spinning frame is provided with yarn, the yarn contacts the detection end of the force sensor, the yarn contacts the electric yarn guide and the upper wall of the tension adjustment roller, the yarn contacts the detection end of the feedback sensor, and the central processing unit is electrically connected to the force sensor, the electric yarn guide, the tension adjustment roller, and the feedback sensor.

3. The high-efficiency yarn tension adjustment system according to claim 1, characterized in that: The process of optimizing the hyperparameters of the TCN-TM-BiDir-LSTM model by the PSO-ISSA algorithm specifically includes the following steps: Step S1: Initialize the TCN-TM-BiDir-LSTM model and set the initial hyperparameter configuration, train the TCN-TM-BiDir-LSTM model according to the initial hyperparameter configuration, obtain the training results, and evaluate the training results to obtain the initial fitness value; Step S2: Initialize the PSO particle swarm, randomly generate particles, and initialize the particle speed and particle position; initialize the ISSA sparrow population, randomly generate explorers and followers, and set the initial explorer position and follower position Step S3: Calculate the fitness value of the particle according to the particle speed and particle position, and calculate the fitness value of the sparrow population according to the explorer position and follower position; Step S4: By comparing the particle fitness values, find the optimal individual particle, obtain the global fitness value, use the PSO particle update method, update the particle speed and particle position according to the particle fitness value and the global fitness value, generate the individual optimal position and the global optimal position through iterative update, and gradually lock the particle optimization hyperparameter configuration; Step S5: guiding the explorer position update according to the global optimal position, guiding the follower position update according to the explorer position update, and continuously optimizing the sparrow population fitness value by iteratively updating the explorer position and the follower position to obtain the sparrow optimization hyperparameter configuration; Step S6: Collaborative information sharing is performed on the particle optimization hyperparameter configuration and the sparrow optimization hyperparameter configuration, and the optimized hyperparameter configuration is obtained through iterative training until the number of iterations is met.

Images