Multi-target collaborative optimization control method and system for igniter robot production line

Abstract

The invention provides a multi-target collaborative optimization control method and system for an igniter robot production line, and relates to the technical field of automatic control, and the method comprises the steps: collecting real-time operation parameters of the production line, carrying out the wavelet denoising, feature extraction and fusion, inputting a hierarchical asynchronous dominant actor commentator model, and in a strategy network layer, carrying out the wavelet denoising, feature extraction and fusion; a robot trajectory planning scheme is optimized and calculated through an improved near-end strategy, a production line speed control scheme is obtained through a layered dual-state action value function in combination with an adaptive Coubeck-Leibler divergence constraint, an energy consumption distribution scheme is obtained based on pseudo-parallel Monte Carlo tree search and an improved soft actor commentator algorithm, and an energy consumption control scheme is obtained. The value network layer evaluates an energy consumption distribution scheme and generates a combined control strategy, after the strategy is executed, a cycle converter model is improved to evaluate production indexes, model parameters are dynamically optimized, and multi-target collaborative optimization is achieved.

Description

Technical Field

[0001] The present invention relates to the field of automatic control technology, and in particular to a multi-objective collaborative optimization control method and system for an ignition tool robot production line. Background Art

[0002] Ignition tool robot production lines play an important role in modern manufacturing. The efficiency and stability of ignition tool robots directly affect the quality and output of products. In order to improve the comprehensive performance of the production line, it is necessary to coordinate and optimize the control of multiple goals in the production process, such as robot trajectory planning, production line speed control and energy consumption distribution. Traditional control methods usually optimize for a single goal and it is difficult to take into account the mutual influence between multiple goals, resulting in overall low efficiency. In addition, the production line operation process will be affected by various uncertain factors, such as equipment aging, environmental changes, etc., making it difficult for the control strategy to adapt to the dynamically changing production environment; Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0003] The embodiments of the present invention provide a multi-objective collaborative optimization control method and system for an ignition tool robot production line, which can at least solve some of the problems existing in the prior art.

[0004] A first aspect of an embodiment of the present invention provides a multi-objective collaborative optimization control method for an ignition tool robot production line, comprising: The real-time operating parameters of the ignition tool robot production line are collected, and based on the iterative wavelet threshold algorithm, the real-time operating parameters are denoised by an adaptive soft threshold function, the real-time operating parameters are decomposed into four layers of wavelet coefficients, and the threshold parameters are determined by calculating the Shannon entropy in each iteration to obtain denoised data, the maximum and minimum normalization calculations are performed on the denoised data, and feature extraction is performed in combination with an improved kernel principal component analysis algorithm, a mixed kernel matrix is ​​constructed and feature dimensions are calculated, a mixed feature matrix is ​​obtained and feature fusion is performed, multi-scale spatial feature weights are calculated in combination with dilated convolution and channel feature weights are calculated in combination with channel compression excitation structure, a fused feature vector is obtained and fuzzy grey correlation is calculated, the production line optimization target is converted into a production index value, the evaluation weight is calculated by an improved ant colony algorithm, and the target weight coefficient is obtained; The production index value, the target weight coefficient and the fusion feature vector are added to the preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantized orthogonal attention, and the optimal scheme combination is solved to generate a joint control strategy. The joint control strategy is implemented, a real-time evaluation model is constructed by improving the loop transformer model, a recursive state update mechanism is set in the improved loop transformer model and the relative position encoding is calculated, the control execution result is input into the real-time evaluation model, the timing characteristics are calculated by multi-head cross attention, the production evaluation index is obtained, and the network parameters of the hierarchical asynchronous advantage actor critic structure are calculated and updated by meta-reinforcement learning, the network node parameters are calculated and the optimal solution is searched by improving the whale optimization algorithm in combination with differential evolution and simulated annealing methods, the timing characteristics of historical data are determined by combining the exponential sliding average, and the learning parameters are adjusted by the fuzzy cognitive map to obtain a dynamic optimization solution.

[0005] In an optional embodiment, The real-time operating parameters of the ignition tool robot production line are collected, and based on the iterative wavelet threshold algorithm, the real-time operating parameters are denoised by an adaptive soft threshold function, the real-time operating parameters are decomposed into four layers of wavelet coefficients, and the threshold parameters are determined by calculating the Shannon entropy in each iteration to obtain denoised data, and the maximum and minimum value normalization calculations are performed on the denoised data, and feature extraction is performed in combination with the improved kernel principal component analysis algorithm, a mixed kernel matrix is ​​constructed, and the feature dimension is calculated, and the mixed feature matrix obtained includes: The real-time operating parameters of the ignition tool robot production line are collected by sensors, wherein the real-time operating parameters include current parameters, pressure parameters, temperature parameters and vibration frequency parameters of the equipment on the production line; Decomposing the real-time operation parameter into four layers of wavelet coefficients by wavelet transform, calculating the Shannon entropy value of each layer according to the wavelet coefficients, calculating the adaptive threshold value based on the Shannon entropy value, constructing a soft threshold function according to the adaptive threshold value, when the absolute value of the input wavelet coefficient is greater than the adaptive threshold value, the output of the soft threshold function is equal to the sign of the wavelet coefficient multiplied by the difference between the absolute value of the wavelet coefficient and the adaptive threshold value, when the absolute value of the input wavelet coefficient is less than or equal to the adaptive threshold value, the output of the soft threshold function is zero, and substituting the wavelet coefficient of each layer into the soft threshold function to obtain the wavelet coefficient after threshold processing; Constructing a reconstruction formula, wherein the reconstruction formula is equal to the sum of the product of the scale function and the approximate coefficient of the last layer, plus the sum of the product of the wavelet function and the detail coefficients of each layer after processing, substituting the wavelet coefficients after the threshold processing into the reconstruction formula, calculating the denoised parameter sequence, performing maximum and minimum value normalization processing on the denoised parameter sequence, and mapping the parameter sequence to the interval from zero to one; A hybrid kernel matrix is ​​obtained by linearly combining a linear kernel matrix, a polynomial kernel matrix, a Gaussian kernel matrix and a Laplace kernel matrix according to preset weight coefficients, and the hybrid kernel matrix is ​​eigendecomposed, and the decomposed features are sorted in descending order according to the eigenvalues, and the features in the top 10% are selected to construct a mapping feature matrix. The normalized parameter sequence is mapped to the feature space according to the mapping feature matrix to obtain the hybrid feature matrix.

[0006] In an optional embodiment, Perform feature fusion, combine dilated convolution to calculate multi-scale spatial feature weights and combine channel compression excitation structure to calculate channel feature weights, obtain fused feature vectors and calculate fuzzy grey correlation, convert production line optimization targets into production index values, calculate evaluation weights through improved ant colony algorithm, and obtain target weight coefficients including: For the mixed feature matrix, multi-scale spatial features are extracted by dilated convolution with different dilation rates, the multi-scale spatial features are weightedly fused to obtain a spatial feature matrix, channel features are extracted by using a channel compression excitation structure, a channel description vector is obtained by global average pooling, a channel weight vector is obtained by passing through two fully connected layers and a Sigmoid activation function, the channel weight vector is multiplied by the mixed feature matrix to obtain a channel feature matrix, and the spatial feature matrix and the channel feature matrix are spliced ​​in the channel dimension to obtain a fused feature matrix; Quantifying the production line optimization target into a production index, calculating the fuzzy grey correlation coefficient between each feature in the fusion feature matrix and each production index, and calculating the correlation between each feature and the production index based on the fuzzy grey correlation coefficient; Construct a complete graph of production indicators, take each indicator as a node in the graph, assign importance and historical evaluation score attribute values ​​to each node, calculate the similarity of attribute values ​​of adjacent nodes as the heuristic information of the current node, and initialize the pheromone concentration of the edges between nodes; Placing a plurality of ants with different search strategies in the complete graph, including global search ants and local search ants, wherein the global search ants give priority to unvisited paths, and control the local search ants to move near known paths, multiplying the pheromone concentration term, the heuristic information term, and the taboo term, and then normalizing them to obtain the adaptive transfer probability of each ant, dynamically adjusting the weight coefficients of the pheromone concentration term and the heuristic information term, and reducing the probability of repeated visits to similar nodes; Determine the path length, node attribute variance and path smoothness of each ant traversal path, and obtain a comprehensive evaluation value by weighted summation. Sort all paths according to the comprehensive evaluation value of the data, superimpose local pheromones on the currently traversed path, enhance the pheromone concentration on the optimal path every three iterations, apply random perturbations to the suboptimal path, randomly select nodes on the optimal path and exchange positions with neighboring nodes, compare the path evaluation values ​​before and after the mutation, retain the path with the highest path evaluation value and calculate the target weight coefficient corresponding to the production indicator in combination with the node attribute value.

[0007] In an optional embodiment, The production index value, the target weight coefficient and the fusion feature vector are added to the preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantizing orthogonal attention, and the optimal scheme combination is solved and the joint control strategy is generated, including: The production index value, the target weight coefficient and the fusion feature vector are input into a hierarchical asynchronous advantage actor-critic model, wherein the hierarchical asynchronous advantage actor-critic model includes a policy network layer and a value network layer. In the policy network layer, an improved proximal policy optimization algorithm is used to calculate the robot trajectory planning scheme, and the Kulbeck-Leibler divergence between the current policy distribution and the target policy distribution is calculated, and the upper and lower bounds of the policy clipping range are dynamically adjusted according to the Kulbeck-Leibler divergence. The policy parameters are updated within the adjusted clipping range and the optimization is iteratively optimized until the policy converges. Through the hierarchical dual-state action-value function, the decision task is mapped to the macro state space and the micro state space, the movement trend is determined by calculating the policy gradient in the macro state space, the input data is decomposed into macro strategies, and micro actions are generated by action sampling in the micro state space. Based on the macro strategies and the micro actions, multiple production line speed control schemes are obtained; For each production line speed control scheme, a search tree with the current state as the root node is constructed, and multiple child nodes are expanded in parallel to generate candidate schemes; a Monte Carlo simulation is performed on each candidate scheme, and the value estimation of the scheme is calculated in combination with an improved soft actor critic algorithm, wherein the soft actor critic algorithm uses a temperature parameter to adjust the balance between exploration and utilization, and selects the optimal energy consumption allocation scheme according to the value estimation; In the value network layer, the value function of the control scheme is calculated by quantizing the orthogonal attention mechanism, the attention matrix is ​​constructed to calculate the correlation between different states and the attention weights are orthogonally constrained to eliminate redundancy, the computational complexity is reduced by quantifying the attention output, and then the multi-head attention mechanism is integrated to extract multi-scale features, the state value function is calculated based on the attention output, the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are combined to form a candidate strategy set, and the strategy combination with the largest value function is selected as the optimal joint control strategy by calculating the comprehensive value function of each candidate strategy.

[0008] In an optional embodiment, The value estimation of the scheme combined with the improved soft actor critic algorithm is shown in the following formula: ; in, t 0 represents the initial temperature value, d 1 represents the temperature decay rate, t represents the time step, t min Indicates the minimum temperature value, d 2 represents the reward error decay rate, r avg represents the average reward value of the current strategy, r target Represents the target reward value.

[0009] In an optional embodiment, The joint control strategy is implemented, a real-time evaluation model is constructed by improving the cyclic transformer model, a recursive state update mechanism is set in the improved cyclic transformer model and the relative position encoding is calculated, the control execution result is input into the real-time evaluation model, the time series characteristics are calculated by multi-head cross attention, the production evaluation index is obtained, and the network parameters of the hierarchical asynchronous advantage actor critic structure are calculated and updated by meta-reinforcement learning, the network node parameters are calculated and the optimal solution is searched by improving the whale optimization algorithm by combining the differential evolution and simulated annealing methods, the time series characteristics of the historical data are determined by combining the exponential sliding average, and the learning parameters are adjusted by the fuzzy cognitive map, and the dynamic optimization scheme is obtained, including: Execute the joint control strategy, and at the same time, build a real-time evaluation model based on the improved recurrent transformer model, concatenate the control execution result and the historical state representation to form an input feature vector, and calculate the input feature vector through the multi-head self-attention mechanism of the recurrent transformer unit and the feedforward neural network to obtain an updated state representation, and introduce a relative position encoding matrix when calculating the attention weight to realize the representation of the timing information; The time series interaction features of state representation and production index are calculated through the multi-head cross attention mechanism. The state representation sequence is used as the query vector and the production index is used as the key value vector to be input into the multi-head cross attention module. The attention weights of the query vector and the key value vector are calculated to perform weighted aggregation on the value vector to obtain the output features of each attention head. The output features of each attention head are concatenated and linearly transformed to obtain the time series interaction feature vector. The temporal interaction feature vector is converted into a production evaluation index through a fully connected layer and an activation function, and the production evaluation index is used as an immediate reward for meta-reinforcement learning. The actor network parameters are updated through a policy gradient algorithm to optimize the control strategy. The critic network parameters are updated and the state action value function is estimated by combining a temporal difference algorithm. Differential evolution and simulated annealing mechanisms are introduced in the parameter updating process to improve the exploration ability of the model. Based on multiple production evaluation indicators, an improved whale optimization algorithm is used for multi-objective solution, and a whale population containing multiple candidate weight coefficients is initialized, wherein each individual in the whale population represents a candidate solution, and the comprehensive evaluation index of each candidate solution under the current weight coefficient is calculated and used as the fitness value, and a global search is performed through the bubble network mechanism, and the local development and update of the population position are combined with the spiral update mechanism. An adaptive weight factor is introduced to dynamically adjust the weights of exploration and development, and the update direction of the candidate solution is corrected based on the position information of the historical optimal solution to obtain the optimal solution; Based on the optimal solution, the exponential sliding average of the pre-acquired historical state representation sequence is performed to extract stable time series features, a fuzzy cognitive map is constructed to set fuzzy rules for the causal relationship between production indicators and the triggering conditions and impact intensity of the fuzzy rules, and the learning parameters of the real-time evaluation model are dynamically adjusted through the fuzzy reasoning mechanism according to the current production indicator value. When an abnormal deviation occurs in the production indicator, the early warning mechanism is triggered and the evaluation optimization process is started to generate an adjustment plan, thereby obtaining the dynamic optimization plan.

[0010] In an optional embodiment, The bubble network mechanism is used for global search, and the spiral update mechanism is used for local development and updating of population positions. An adaptive weight factor is introduced to dynamically adjust the weights of exploration and development. The update direction of the candidate solution is corrected based on the location information of the historical optimal solution. The optimal solutions include: Initialize a whale population containing multiple candidate solutions, each candidate solution represents a set of network node parameters, and the dimension of the candidate solution is the same as the number of network node parameters; A bubble network mechanism is used for global search, a reference point is randomly selected, a shrinking step size and a shrinking direction are calculated based on the distance between the reference point and the candidate solution, and the position of the candidate solution is updated for the first time according to the shrinking step size and the shrinking direction; Based on the spiral update mechanism, local development is performed. With the current optimal solution as the center, the candidate solutions are updated according to the spiral equation. The spiral equation includes a rotation angle parameter and a contraction distance parameter. The rotation angle parameter and the contraction distance parameter are used to control the rotation amplitude and radial movement distance of the candidate solutions in each dimension, and the position of the candidate solutions is updated for the second time. An adaptive weight factor is introduced to calculate the similarity between the candidate solution and the historical optimal network parameter configuration, and the weight ratio of the bubble network search and the spiral update is dynamically adjusted according to the similarity. When the similarity is lower than the preset similarity threshold, the weight of the bubble network search is increased, and when the similarity is higher than the similarity threshold, the weight of the spiral update is increased; Based on the location information of the historical optimal network parameter configuration, the update direction of the candidate solution is corrected, the direction vector of the candidate solution and the historical optimal solution is calculated, the direction vector is weightedly combined with the original update direction of the candidate solution to obtain a corrected update direction, and the location of the candidate solution is updated for the third time according to the corrected update direction; The update is repeated until the maximum number of iterations is reached or the preset convergence condition is met, and the optimal combination of network node parameters is output as the optimal solution.

[0011] A second aspect of the embodiments of the present invention provides a multi-objective collaborative optimization control system for an ignition tool robot production line, comprising: The first unit is used to collect the real-time operating parameters of the ignition tool robot production line, perform denoising on the real-time operating parameters through an adaptive soft threshold function based on an iterative wavelet threshold algorithm, decompose the real-time operating parameters into four layers of wavelet coefficients, and determine the threshold parameters by calculating the Shannon entropy in each iteration to obtain denoised data, perform maximum and minimum value normalization calculation on the denoised data and extract features in combination with an improved kernel principal component analysis algorithm, construct a mixed kernel matrix and calculate the feature dimension, obtain a mixed feature matrix and perform feature fusion, calculate multi-scale spatial feature weights in combination with dilated convolution and calculate channel feature weights in combination with a channel compression excitation structure, obtain a fused feature vector and calculate a fuzzy grey correlation degree, convert the production line optimization target into a production index value, calculate the evaluation weight through an improved ant colony algorithm, and obtain a target weight coefficient; The second unit is used to add the production index value, the target weight coefficient and the fusion feature vector to a preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantized orthogonal attention, and the optimal scheme combination is solved to generate a joint control strategy. The third unit is used to execute the joint control strategy, build a real-time evaluation model by improving the loop transformer model, set a recursive state update mechanism in the improved loop transformer model and calculate the relative position encoding, input the control execution result into the real-time evaluation model, calculate the timing characteristics through multi-head cross attention, obtain the production evaluation index, and calculate and update the network parameters of the hierarchical asynchronous advantage actor critic structure through meta-reinforcement learning, combine the differential evolution and simulated annealing methods, calculate the network node parameters and search for the optimal solution through the improved whale optimization algorithm, determine the timing characteristics of the historical data by combining the exponential sliding average, and adjust the learning parameters through the fuzzy cognitive map to obtain a dynamic optimization solution.

[0012] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0014] In the present invention, the real-time operation parameters are denoised by iterative wavelet threshold algorithm, and the feature extraction is combined with the improved kernel principal component analysis algorithm, which can effectively remove noise interference, extract key features, and improve the accuracy and robustness of the model. The multi-scale spatial features and channel feature weights are calculated by combining the dilated convolution and channel compression excitation structure, which can better capture the changes in the operation status of the production line and achieve more accurate control. The evaluation weights calculated by the improved ant colony algorithm can more accurately reflect the importance of the production target, thereby optimizing the operation efficiency of the production line. Through the hierarchical asynchronous dominant actor critic model, combined with the adaptive Kulbeck-Leibler divergence constraint and The hierarchical dual-state action-value function can achieve more sophisticated trajectory planning, speed control and energy consumption allocation. By expanding the action space based on the pseudo-parallel Monte Carlo tree and combining it with the improved soft actor critic algorithm, more potential control strategies can be explored and more energy-saving solutions can be found. By improving the cyclic transformer model to build a real-time evaluation model and combining it with multi-head cross-attention to calculate the timing characteristics, the control effect can be evaluated in real time and the control strategy can be dynamically adjusted according to the production evaluation indicators. By combining the exponential sliding average to determine the timing characteristics of historical data and adjusting the learning parameters through the fuzzy cognitive map, the production line can adaptively adjust the control strategy to cope with changes in the production environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a multi-objective collaborative optimization control method for an ignition tool robot production line according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of the multi-objective collaborative optimization control system of the ignition tool robot production line according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 FIG. 1 is a flow chart of a multi-objective collaborative optimization control method for an ignition tool robot production line according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Collect the real-time operating parameters of the ignition tool robot production line, perform denoising on the real-time operating parameters through an adaptive soft threshold function based on an iterative wavelet threshold algorithm, decompose the real-time operating parameters into four layers of wavelet coefficients, and determine the threshold parameters by calculating the Shannon entropy in each iteration to obtain denoised data, perform maximum and minimum value normalization calculations on the denoised data and extract features in combination with an improved kernel principal component analysis algorithm, construct a mixed kernel matrix and calculate the feature dimension, obtain a mixed feature matrix and perform feature fusion, calculate multi-scale spatial feature weights in combination with dilated convolution and calculate channel feature weights in combination with channel compression excitation structure, obtain a fused feature vector and calculate fuzzy grey correlation, convert the production line optimization target into a production index value, calculate the evaluation weight through an improved ant colony algorithm, and obtain the target weight coefficient; The adaptive soft threshold function is a nonlinear transformation function used in signal processing and compressed sensing. Its main function is to adaptively adjust the threshold according to the size or characteristics of the signal, and highlight the important signal characteristics by removing smaller signal components. The Shannon entropy is a basic concept in information theory, which is used to measure the randomness or uncertainty of information. The mixed kernel matrix is ​​a matrix that combines multiple kernel functions and is used to process complex data distribution. The fuzzy gray correlation is a measurement method based on fuzzy mathematics and gray system theory, which is used to evaluate the correlation between different objects. When processing incomplete, fuzzy or uncertain information, the fuzzy gray correlation can help analyze the similarity or dependence between various factors in the system. By calculating the fuzzy gray correlation, the potential relationship between variables in the system can be revealed.

[0019] In an optional embodiment, The real-time operating parameters of the ignition tool robot production line are collected, and based on the iterative wavelet threshold algorithm, the real-time operating parameters are denoised by an adaptive soft threshold function, the real-time operating parameters are decomposed into four layers of wavelet coefficients, and the threshold parameters are determined by calculating the Shannon entropy in each iteration to obtain denoised data, and the maximum and minimum value normalization calculations are performed on the denoised data, and feature extraction is performed in combination with the improved kernel principal component analysis algorithm, a mixed kernel matrix is ​​constructed, and the feature dimension is calculated, and the mixed feature matrix obtained includes: The real-time operating parameters of the ignition tool robot production line are collected by sensors, wherein the real-time operating parameters include current parameters, pressure parameters, temperature parameters and vibration frequency parameters of the equipment on the production line; Decomposing the real-time operation parameter into four layers of wavelet coefficients by wavelet transform, calculating the Shannon entropy value of each layer according to the wavelet coefficients, calculating the adaptive threshold value based on the Shannon entropy value, constructing a soft threshold function according to the adaptive threshold value, when the absolute value of the input wavelet coefficient is greater than the adaptive threshold value, the output of the soft threshold function is equal to the sign of the wavelet coefficient multiplied by the difference between the absolute value of the wavelet coefficient and the adaptive threshold value, when the absolute value of the input wavelet coefficient is less than or equal to the adaptive threshold value, the output of the soft threshold function is zero, and substituting the wavelet coefficient of each layer into the soft threshold function to obtain the wavelet coefficient after threshold processing; Constructing a reconstruction formula, wherein the reconstruction formula is equal to the sum of the product of the scale function and the approximate coefficient of the last layer, plus the sum of the product of the wavelet function and the detail coefficients of each layer after processing, substituting the wavelet coefficients after the threshold processing into the reconstruction formula, calculating the denoised parameter sequence, performing maximum and minimum value normalization processing on the denoised parameter sequence, and mapping the parameter sequence to the interval from zero to one; A hybrid kernel matrix is ​​obtained by linearly combining a linear kernel matrix, a polynomial kernel matrix, a Gaussian kernel matrix and a Laplace kernel matrix according to preset weight coefficients, and the hybrid kernel matrix is ​​eigendecomposed, and the decomposed features are sorted in descending order according to the eigenvalues, and the features in the top 10% are selected to construct a mapping feature matrix. The normalized parameter sequence is mapped to the feature space according to the mapping feature matrix to obtain the hybrid feature matrix.

[0020] The reconstruction formula refers to a mathematical formula for restoring a signal or data from its sparse representation or compressed state to its original form. The soft threshold function is a nonlinear function commonly used in signal denoising and sparse coding. It compresses the amplitude of the signal and compresses smaller amplitude values ​​to zero, thereby removing noise or unimportant signal components.

[0021] Collect the real-time operating parameters of the ignition robot production line, including the current, pressure, temperature and vibration frequency of the equipment on the production line. The sensors can be current sensors, pressure sensors, temperature sensors and vibration sensors, which collect the corresponding parameters respectively. For example, data can be collected once per second and stored in the database. Assume that the collected current value is 10A, the pressure value is 2MPa, the temperature value is 200℃, and the vibration frequency is 100Hz.

[0022] The collected real-time operating parameters are decomposed into four layers of wavelet coefficients using wavelet transform, and the four layers are decomposed using db4 wavelet. After decomposition, four layers of detail coefficients and one layer of approximate coefficients are obtained. For example, the first layer of detail coefficients may contain high-frequency noise information, while the fourth layer of approximate coefficients contains the main signal trend information. Assume that a certain detail coefficient after decomposition is [0.1, 0.2, 0.3, 0.4].

[0023] According to the decomposed wavelet coefficients, the Shannon entropy value of each layer is calculated. Shannon entropy is used to measure the uncertainty of the signal. Noise usually has a higher Shannon entropy. The calculation method is to first calculate the probability density of each wavelet coefficient, and then calculate the entropy value according to the definition formula of Shannon entropy. Assume that the Shannon entropy value of a certain layer is 0.8.

[0024] Based on the calculated Shannon entropy value, an adaptive threshold is calculated, and the threshold is used to distinguish between signals and noise. The calculation method of the adaptive threshold can be adjusted according to the Shannon entropy value. For example, the threshold can be set to a certain multiple of the Shannon entropy value or a function value related to the Shannon entropy value. Assume that the adaptive threshold calculated based on the Shannon entropy of 0.8 is 0.2.

[0025] Construct a soft threshold function based on the adaptive threshold. The function of the soft threshold function is to set the wavelet coefficients less than the threshold to zero, and the wavelet coefficients greater than the threshold to shrink to zero. For example, if a wavelet coefficient is 0.3 and the threshold is 0.2, the output of the soft threshold function is 0.1; if a wavelet coefficient is 0.1 and the threshold is 0.2, the output of the soft threshold function is 0. Substitute the wavelet coefficients of each layer into the soft threshold function for processing to obtain the wavelet coefficients after threshold processing. Assume that a certain detail coefficient after processing is [0, 0, 0.1, 0.2].

[0026] Construct a reconstruction formula, substitute the wavelet coefficients after threshold processing into the reconstruction formula, and calculate the denoised parameter sequence. The reconstruction formula uses the inverse wavelet transform to recombine the processed wavelet coefficients into the denoised signal. Assume that the denoised current value is 9.8A, the pressure value is 1.9MPa, the temperature value is 198℃, and the vibration frequency is 99Hz.

[0027] The denoised parameter sequence is normalized to its maximum and minimum values. The parameter sequence is mapped to the interval from zero to one. For example, assuming that the denoised current value sequence is [9.5, 9.8, 10.2, 10.5], the maximum value is 10.5, and the minimum value is 9.5, then the normalized current value sequence is [0, 0.429, 0.857, 1].

[0028] The linear kernel matrix, the polynomial kernel matrix, the Gaussian kernel matrix and the Laplace kernel matrix are linearly combined according to the preset weight coefficients to obtain a hybrid kernel matrix. The preset weight coefficients can be determined based on experience or experimental results. For example, the weight coefficients of the four kernel matrices can be set to 0.25 respectively. Assuming that the value of the linear kernel matrix is ​​1, the value of the polynomial kernel matrix is ​​2, the value of the Gaussian kernel matrix is ​​3, and the value of the Laplace kernel matrix is ​​4, the value of the hybrid kernel matrix is ​​2.5.

[0029] Perform eigendecomposition on the mixed kernel matrix, sort the decomposed features in descending order according to the eigenvalues, and select the features in the top 10% to construct the mapping feature matrix. Assuming the eigenvalues ​​are [10, 5, 2, 1], select the features corresponding to the eigenvalue 10 to construct the mapping feature matrix.

[0030] The normalized parameter sequence is mapped to the feature space according to the mapping feature matrix to obtain a mixed feature matrix. The mixed feature matrix contains the extracted feature information and can be used for subsequent fault diagnosis or other analysis.

[0031] In this embodiment, noise interference is effectively removed by iterative wavelet threshold algorithm and adaptive soft threshold function, which improves the accuracy and reliability of collected data and provides a more accurate data basis for subsequent analysis. Combined with the improved kernel principal component analysis algorithm and mixed kernel matrix, more effective feature information can be extracted, avoiding the limitations of a single kernel function, improving the comprehensiveness and accuracy of feature expression, and being able to quickly process real-time operating parameters to meet the needs of real-time monitoring and control, which helps to promptly discover and handle abnormal situations on the production line.

[0032] In an optional embodiment, Perform feature fusion, combine dilated convolution to calculate multi-scale spatial feature weights and combine channel compression excitation structure to calculate channel feature weights, obtain fused feature vectors and calculate fuzzy grey correlation, convert production line optimization targets into production index values, calculate evaluation weights through improved ant colony algorithm, and obtain target weight coefficients including: For the mixed feature matrix, multi-scale spatial features are extracted by dilated convolution with different dilation rates, the multi-scale spatial features are weightedly fused to obtain a spatial feature matrix, channel features are extracted by using a channel compression excitation structure, a channel description vector is obtained by global average pooling, a channel weight vector is obtained by passing through two fully connected layers and a Sigmoid activation function, the channel weight vector is multiplied by the mixed feature matrix to obtain a channel feature matrix, and the spatial feature matrix and the channel feature matrix are spliced ​​in the channel dimension to obtain a fused feature matrix; Quantifying the production line optimization target into a production index, calculating the fuzzy grey correlation coefficient between each feature in the fusion feature matrix and each production index, and calculating the correlation between each feature and the production index based on the fuzzy grey correlation coefficient; Construct a complete graph of production indicators, take each indicator as a node in the graph, assign importance and historical evaluation score attribute values ​​to each node, calculate the similarity of attribute values ​​of adjacent nodes as the heuristic information of the current node, and initialize the pheromone concentration of the edges between nodes; Placing a plurality of ants with different search strategies in the complete graph, including global search ants and local search ants, wherein the global search ants give priority to unvisited paths, and control the local search ants to move near known paths, multiplying the pheromone concentration term, the heuristic information term, and the taboo term, and then normalizing them to obtain the adaptive transfer probability of each ant, dynamically adjusting the weight coefficients of the pheromone concentration term and the heuristic information term, and reducing the probability of repeated visits to similar nodes; Determine the path length, node attribute variance and path smoothness of each ant traversal path, and obtain a comprehensive evaluation value by weighted summation. Sort all paths according to the comprehensive evaluation value of the data, superimpose local pheromones on the currently traversed path, enhance the pheromone concentration on the optimal path every three iterations, apply random perturbations to the suboptimal path, randomly select nodes on the optimal path and exchange positions with neighboring nodes, compare the path evaluation values ​​before and after the mutation, retain the path with the highest path evaluation value and calculate the target weight coefficient corresponding to the production indicator in combination with the node attribute value.

[0033] The dilation rate refers to the degree of expansion of the receptive field in the convolution operation in the convolutional neural network. The dilated convolution is a convolution operation applied in the convolutional neural network. The dilation rate is introduced to expand the receptive field of the convolution kernel to capture a wider range of feature information. The fuzzy grey correlation coefficient is a metric used to measure the strength of the relationship between different factors in the fuzzy system. By combining fuzzy theory and grey system theory, the degree of association between different objects or variables is calculated. The pheromone concentration is an indicator used to indicate the quality of the search path in the ant colony algorithm. The local pheromone refers to the amount of pheromone released by ants in a local area in the ant colony algorithm, which is usually used to indicate the quality of a local search path.

[0034] The data obtained from the production line contains different types of production data collected by multiple sensors, such as temperature, pressure, speed, etc. These different types of production data form a mixed feature matrix. For example, if 5 different types of production data are collected at 100 time points, the size of the mixed feature matrix is ​​100x5.

[0035] Extract multi-scale spatial features and use dilated convolutions with different dilation rates to perform convolution operations on the mixed feature matrix. For example, use three dilated convolution kernels with dilation rates of 1, 2, and 3 to extract spatial features of three different scales. Perform weighted fusion on the three spatial features of different scales to obtain a spatial feature matrix. The weighted fusion method can be a simple average, or different weights can be set according to the actual situation. For example, multiply the spatial features of the three scales by weights of 0.2, 0.3, and 0.5 respectively, and then add them together to obtain the final spatial feature matrix.

[0036] Extract channel features. Use the channel compression excitation structure to extract channel features from the mixed feature matrix, perform global average pooling on the mixed feature matrix, and obtain a channel description vector. For example, for a 100x5 mixed feature matrix, a 1x5 channel description vector is obtained after global average pooling. The channel description vector is input into two fully connected layers, and the channel weight vector is obtained using the Sigmoid activation function. For example, if the output dimensions of the two fully connected layers are 3 and 5, respectively, a 1x5 channel weight vector is finally obtained. Multiply the channel weight vector with the mixed feature matrix to obtain the channel feature matrix. For example, multiply the 1x5 channel weight vector with the 100x5 mixed feature matrix to obtain a 100x5 channel feature matrix.

[0037] The spatial feature matrix and the channel feature matrix are concatenated in the channel dimension to obtain a fused feature matrix. For example, a 100x5 spatial feature matrix and a 100x5 channel feature matrix are concatenated to obtain a 100x10 fused feature matrix.

[0038] Quantify the production line optimization goal into production indicators. For example, if the optimization goal is set to improve production efficiency and reduce production costs, the corresponding production indicators can be unit time output and unit product cost. Assume that there are three production indicators, namely unit time output, unit product cost and product qualification rate.

[0039] Calculate the fuzzy grey correlation coefficient between each feature and each production index in the fusion feature matrix. For example, calculate the fuzzy grey correlation coefficient between 10 features and 3 production indicators in the fusion feature matrix to obtain a 10x3 matrix. Calculate the correlation between each feature and the production index based on the fuzzy grey correlation coefficient.

[0040] Construct a complete graph of production indicators, and use each indicator as a node in the graph. For example, the three indicators of unit time output, unit product cost, and product qualification rate are used as three nodes in the graph. Assign importance and historical evaluation score attribute values ​​to each node. For example, based on experience or expert knowledge, the importance of unit time output is set to 0.5, and the historical evaluation score is set to 0.8; the importance of unit product cost is set to 0.3, and the historical evaluation score is set to 0.7; the importance of product qualification rate is set to 0.2, and the historical evaluation score is set to 0.9. Calculate the similarity of attribute values ​​of adjacent nodes as heuristic information for the current node.

[0041] Place multiple ants with different search strategies in the complete graph, including global search ants and local search ants. Global search ants give priority to unvisited paths and control local search ants to move near known paths. Calculate the adaptive transition probability of each ant. The pheromone concentration term, heuristic information term, and taboo term are multiplied and normalized. Dynamically adjust the weight coefficients of the pheromone concentration term and the heuristic information term to reduce the probability of repeated visits to similar nodes.

[0042] Determine the path length, node attribute variance, and path smoothness of each ant traversal path. Obtain a comprehensive evaluation value by weighted summation. Sort all paths according to the comprehensive evaluation value. Superimpose local pheromones on the currently traversed path. Enhance the pheromone concentration on the optimal path every three iterations. Apply random perturbations to the suboptimal path. Randomly select nodes on the optimal path and exchange positions with neighboring nodes. Compare the path evaluation values ​​before and after the mutation, retain the path with the highest path evaluation value, and calculate the target weight coefficient corresponding to the production index in combination with the node attribute value. For example, the target weight coefficients corresponding to the three indicators of unit time output, unit product cost, and product qualification rate are 0.4, 0.35, and 0.25, respectively.

[0043] In this embodiment, by considering multi-scale spatial features and channel features, it is possible to capture information in production data more comprehensively, thereby more accurately evaluating the importance of production indicators and ultimately improving production line performance. The improved ant colony algorithm can adaptively adjust the search strategy to avoid falling into local optimal solutions, thereby enhancing the generalization ability of the model and enabling it to adapt to different production line environments. It can quickly and effectively determine the target weight coefficients of production indicators, thereby guiding the optimization direction of the production line and improving the optimization efficiency of the production line.

[0044] S2. Add the production index value, the target weight coefficient and the fusion feature vector to the preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. Through the hierarchical dual-state action value function, the decision task is decomposed into macro-strategy and micro-action to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantized orthogonal attention, and the optimal scheme combination is solved to generate a joint control strategy. The hierarchical asynchronous advantage actor-critic model is a reinforcement learning algorithm that combines the advantage actor-critic (A2C) model and the hierarchical learning framework. It improves learning efficiency by decomposing tasks into sub-tasks at multiple levels and making decisions and learning at the macro and micro levels respectively. The improved proximal policy optimization (PPO) is a commonly used policy optimization algorithm in reinforcement learning. Its main purpose is to improve the traditional policy gradient method and make the update process more stable. The hierarchical dual-state action value function is a reinforcement learning algorithm that combines dual Q learning and a hierarchical reinforcement learning framework. The pseudo-parallel Monte Carlo tree is an improved version of the Monte Carlo tree search algorithm. It accelerates the search process by simulating multiple independent random paths for parallel evaluation while avoiding excessive consumption of computing resources. It is suitable for decision-making tasks with high time requirements.

[0045] In an optional embodiment, The production index value, the target weight coefficient and the fusion feature vector are added to the preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantizing orthogonal attention, and the optimal scheme combination is solved and the joint control strategy is generated, including: The production index value, the target weight coefficient and the fusion feature vector are input into a hierarchical asynchronous advantage actor-critic model, wherein the hierarchical asynchronous advantage actor-critic model includes a policy network layer and a value network layer. In the policy network layer, an improved proximal policy optimization algorithm is used to calculate the robot trajectory planning scheme, and the Kulbeck-Leibler divergence between the current policy distribution and the target policy distribution is calculated, and the upper and lower bounds of the policy clipping range are dynamically adjusted according to the Kulbeck-Leibler divergence. The policy parameters are updated within the adjusted clipping range and the optimization is iteratively optimized until the policy converges. Through the hierarchical dual-state action-value function, the decision task is mapped to the macro state space and the micro state space, the movement trend is determined by calculating the policy gradient in the macro state space, the input data is decomposed into macro strategies, and micro actions are generated by action sampling in the micro state space. Based on the macro strategies and the micro actions, multiple production line speed control schemes are obtained; For each production line speed control scheme, a search tree with the current state as the root node is constructed, and multiple child nodes are expanded in parallel to generate candidate schemes; a Monte Carlo simulation is performed on each candidate scheme, and the value estimation of the scheme is calculated in combination with an improved soft actor critic algorithm, wherein the soft actor critic algorithm uses a temperature parameter to adjust the balance between exploration and utilization, and selects the optimal energy consumption allocation scheme according to the value estimation; In the value network layer, the value function of the control scheme is calculated by quantizing the orthogonal attention mechanism, the attention matrix is ​​constructed to calculate the correlation between different states and the attention weights are orthogonally constrained to eliminate redundancy, the computational complexity is reduced by quantifying the attention output, and then the multi-head attention mechanism is integrated to extract multi-scale features, the state value function is calculated based on the attention output, the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are combined to form a candidate strategy set, and the strategy combination with the largest value function is selected as the optimal joint control strategy by calculating the comprehensive value function of each candidate strategy.

[0046] The macro state space refers to decomposing the state space of the entire task into multiple large-scale high-level states in the reinforcement learning task. The micro state space refers to the detailed level state in the task, which usually has a more direct impact on the operation or decision-making process of the system. The orthogonal constraint refers to restricting the relationship between different variables in the optimization problem to keep them independent or orthogonal.

[0047] Collect various indicator data during the operation of the production line, such as output, qualified rate, energy consumption, etc., and set the weight coefficient of each indicator according to the production target. At the same time, extract the fusion feature vector of the production line status, such as equipment status, material inventory, etc. Input these data into the pre-trained hierarchical asynchronous advantage actor critic model.

[0048] The core of the hierarchical asynchronous advantage actor-critic model is the policy network layer and the value network layer. In the policy network layer, an improved proximal policy optimization algorithm is used to calculate the robot trajectory planning scheme. Specifically, the robot trajectory generated by the current policy is compared with the target trajectory, and the difference between the two distributions is calculated. According to the size of the difference, the amplitude of the policy update is dynamically adjusted, and iterative optimization is performed within a limited range until the policy converges to obtain the optimal robot trajectory planning scheme. For example, assuming that the target trajectory is that the robot moves from point A to point B, the trajectory generated by the current policy has a slight deviation. By continuously adjusting the policy parameters, the generated trajectory is finally made consistent with the target trajectory.

[0049] Through the hierarchical dual state-action value function, the complex decision-making task is decomposed into macro-strategies and micro-actions. At the macro level, the overall operation trend of the production line is determined according to the production goals and current status, such as increasing output or reducing energy consumption. At the micro level, the production line speed is fine-tuned according to the macro strategy. For example, if the macro strategy is to increase output, the micro action can be to increase the speed of some production lines while reducing the speed of other production lines to balance overall efficiency and energy consumption. Through this hierarchical decision-making method, multiple candidate production line speed control solutions are generated.

[0050] For each speed control scheme, a pseudo-parallel Monte Carlo tree search method is used to expand the action space and evaluate the value of the scheme. Taking the current state as the root node, multiple child nodes are expanded in parallel, each of which represents a possible energy allocation scheme. Multiple simulations are performed for each scheme, and the value estimate is calculated in combination with the improved soft actor-critic algorithm. By introducing the temperature parameter, the relationship between exploring new schemes and using existing experience is balanced. For example, when the temperature parameter is high, it is more inclined to explore new energy allocation schemes; when the temperature parameter is low, it is more inclined to choose schemes with higher known values. Finally, the energy allocation scheme with the highest value estimate is selected.

[0051] At the value network layer, a quantized orthogonal attention mechanism is used to calculate the value function of each solution. By constructing an attention matrix, the correlation between different states is calculated, and the attention weights are orthogonally constrained to eliminate redundant information. For example, if a state is highly correlated with multiple other states, only the most important correlations are retained through orthogonal constraints to reduce computational complexity. At the same time, quantization technology is used to compress the attention output to further improve computational efficiency. Finally, a multi-head attention mechanism is integrated to extract multi-scale features and calculate the final value function of each solution.

[0052] In this embodiment, by optimizing robot trajectory planning and production line speed control, production time can be reduced and output can be increased. For example, by optimizing the robot motion path, collisions and waiting can be avoided, and the production cycle can be shortened. By optimizing the energy consumption allocation plan, energy consumption in the production process can be reduced and costs can be saved. For example, the power of the equipment can be dynamically adjusted according to production needs to avoid energy waste. Through hierarchical decision-making and multi-source information fusion, the production line can be more finely controlled to improve product quality and stability. For example, adjusting the production line speed according to the real-time status can avoid fluctuations in the production process and ensure product consistency.

[0053] In an optional embodiment, The value estimation of the scheme combined with the improved soft actor critic algorithm is shown in the following formula: ; in, t0 represents the initial temperature value, d 1 represents the temperature decay rate, t represents the time step, t min Indicates the minimum temperature value, d 2 represents the reward error decay rate, r avg represents the average reward value of the current strategy, r target Represents the target reward value.

[0054] In this embodiment, by dynamically adjusting the temperature parameter, exploration and utilization can be balanced, thereby improving learning efficiency. By introducing the reward error decay rate, the impact of reward value fluctuations can be reduced and the stability of the algorithm can be enhanced. By comprehensively considering the temperature parameter, reward error and reward error decay rate, the value can be estimated more accurately, thereby improving the performance of the strategy.

[0055] S3. Execute the joint control strategy, build a real-time evaluation model by improving the loop transformer model, set a recursive state update mechanism in the improved loop transformer model and calculate the relative position encoding, input the control execution result into the real-time evaluation model, calculate the timing characteristics through multi-head cross attention, obtain the production evaluation index, and calculate and update the network parameters of the hierarchical asynchronous advantage actor critic structure through meta-reinforcement learning, combine differential evolution and simulated annealing methods, calculate the network node parameters and search for the optimal solution through the improved whale optimization algorithm, determine the timing characteristics of historical data by combining the exponential sliding average, and adjust the learning parameters through the fuzzy cognitive map to obtain a dynamic optimization solution.

[0056] The improved recurrent transformer model is an improvement on the traditional recurrent transformer (such as Transformer), which aims to improve the performance of the model when processing sequence data. The relative position encoding is an improved position encoding method, which is usually used in models based on attention mechanism. The meta-reinforcement learning is a method that combines meta-learning (learning how to learn) with reinforcement learning. The goal is to enable the agent to quickly adjust its strategy when facing new tasks through multiple task training. The improved whale optimization algorithm is an optimization algorithm based on whale predation behavior, which imitates the behavior pattern of whale groups when preying. The exponential sliding average is a weighted average method, in which newer data points have higher weights, while the weights of older data points gradually decay over time. The fuzzy cognitive map is a graphical tool for representing and analyzing uncertainty and ambiguity in complex systems.

[0057] In an optional embodiment, The joint control strategy is implemented, a real-time evaluation model is constructed by improving the cyclic transformer model, a recursive state update mechanism is set in the improved cyclic transformer model and the relative position encoding is calculated, the control execution result is input into the real-time evaluation model, the time series characteristics are calculated by multi-head cross attention, the production evaluation index is obtained, and the network parameters of the hierarchical asynchronous advantage actor critic structure are calculated and updated by meta-reinforcement learning, the network node parameters are calculated and the optimal solution is searched by improving the whale optimization algorithm by combining the differential evolution and simulated annealing methods, the time series characteristics of the historical data are determined by combining the exponential sliding average, and the learning parameters are adjusted by the fuzzy cognitive map, and the dynamic optimization scheme is obtained, including: Execute the joint control strategy, and at the same time, build a real-time evaluation model based on the improved recurrent transformer model, concatenate the control execution result and the historical state representation to form an input feature vector, and calculate the input feature vector through the multi-head self-attention mechanism of the recurrent transformer unit and the feedforward neural network to obtain an updated state representation, and introduce a relative position encoding matrix when calculating the attention weight to realize the representation of the timing information; The time series interaction features of state representation and production index are calculated through the multi-head cross attention mechanism. The state representation sequence is used as the query vector and the production index is used as the key value vector to be input into the multi-head cross attention module. The attention weights of the query vector and the key value vector are calculated to perform weighted aggregation on the value vector to obtain the output features of each attention head. The output features of each attention head are concatenated and linearly transformed to obtain the time series interaction feature vector. The temporal interaction feature vector is converted into a production evaluation index through a fully connected layer and an activation function, and the production evaluation index is used as an immediate reward for meta-reinforcement learning. The actor network parameters are updated through a policy gradient algorithm to optimize the control strategy. The critic network parameters are updated and the state action value function is estimated by combining a temporal difference algorithm. Differential evolution and simulated annealing mechanisms are introduced in the parameter updating process to improve the exploration ability of the model. Based on multiple production evaluation indicators, an improved whale optimization algorithm is used for multi-objective solution, and a whale population containing multiple candidate weight coefficients is initialized, wherein each individual in the whale population represents a candidate solution, and the comprehensive evaluation index of each candidate solution under the current weight coefficient is calculated and used as the fitness value, and a global search is performed through the bubble network mechanism, and the local development and update of the population position are combined with the spiral update mechanism. An adaptive weight factor is introduced to dynamically adjust the weights of exploration and development, and the update direction of the candidate solution is corrected based on the position information of the historical optimal solution to obtain the optimal solution; Based on the optimal solution, the exponential sliding average of the pre-acquired historical state representation sequence is performed to extract stable time series features, a fuzzy cognitive map is constructed to set fuzzy rules for the causal relationship between production indicators and the triggering conditions and impact intensity of the fuzzy rules, and the learning parameters of the real-time evaluation model are dynamically adjusted through the fuzzy reasoning mechanism according to the current production indicator value. When an abnormal deviation occurs in the production indicator, the early warning mechanism is triggered and the evaluation optimization process is started to generate an adjustment plan, thereby obtaining the dynamic optimization plan.

[0058] The fuzzy rule is a rule expression method based on fuzzy logic, usually composed of an "if-then" structure, and is used to describe the relationship between fuzzy variables in the system. The fuzzy reasoning mechanism is a process of reasoning the input fuzzy information using fuzzy logic.

[0059] Execute the preset joint control strategy, and input the control execution results into the real-time evaluation model. At the same time, build a real-time evaluation model based on the improved cyclic transformer model. The model concatenates the control execution results with the historical state representation to form an input feature vector. For example, assuming that the control execution results are output and quality, and the historical state is represented by temperature, pressure, and raw material ratio, these five indicators are concatenated to form a five-dimensional vector as the input feature vector. The vector is input to the cyclic transformer unit, and after calculation by the multi-head self-attention mechanism and the feedforward neural network, an updated state representation is obtained. When calculating the attention weight, a relative position encoding matrix is ​​introduced, for example, an encoding matrix generated using sine and cosine functions, to characterize the timing information, so that the model can better capture the dependencies in the time series data.

[0060] The multi-head cross-attention mechanism is used to calculate the temporal interaction features of state representation and production indicators. The state representation sequence is used as the query vector and the production indicator is used as the key-value vector, which are input into the multi-head cross-attention module. By calculating the attention weights of the query vector and the key-value vector, the value vector is weightedly aggregated to obtain the output features of each attention head. For example, if there are two attention heads, two output feature vectors will be obtained. These output feature vectors are concatenated and linearly transformed to obtain the final temporal interaction feature vector.

[0061] The temporal interaction feature vector is converted into production evaluation indicators such as production efficiency, product quality, and energy consumption through a fully connected layer and an activation function (such as the ReLU activation function). These evaluation indicators are used as immediate rewards for meta-reinforcement learning to update the actor network parameters to optimize the control strategy. At the same time, the critic network parameters are updated in combination with the temporal difference algorithm, and the state action value function is estimated. In the parameter update process, differential evolution and simulated annealing mechanisms are introduced to improve the exploration ability of the model and avoid falling into local optimal solutions. For example, in differential evolution, new candidate solutions are generated through mutation and crossover operations, and in simulated annealing, poorer solutions are accepted with a certain probability to jump out of the local optimum.

[0062] Based on multiple production evaluation indicators, an improved whale optimization algorithm is used for multi-objective solution. Initialize a whale population containing multiple candidate weight coefficients, where each individual represents a candidate solution. For example, assuming there are three evaluation indicators, each individual contains three weight coefficients, corresponding to the weights of the three indicators. Calculate the comprehensive evaluation index of each candidate solution under the current weight coefficient and use it as the fitness value. Perform global search through the bubble network mechanism, and combine the spiral update mechanism for local development to update the population position. Introduce adaptive weight factors to dynamically adjust the weights of exploration and development. For example, in the early stage of the search, give exploration a greater weight, while in the later stage of the search, give development a greater weight. Based on the position information of the historical optimal solution, the update direction of the candidate solution is corrected, and finally the optimal solution, that is, the optimal combination of weight coefficients, is obtained.

[0063] Based on the optimal solution, perform exponential sliding average on the pre-acquired historical state representation sequence to extract stable time series features. Construct a fuzzy cognitive map, set fuzzy rules for the causal relationship between production indicators, as well as the triggering conditions and impact intensity of these rules. For example, set the fuzzy rule "If the temperature is too high, the product quality will decline", and set the temperature threshold and the degree of quality decline. According to the current production indicator value, dynamically adjust the learning parameters of the real-time evaluation model, such as learning rate and batch size, through the fuzzy reasoning mechanism. When there is an abnormal deviation in the production indicator, such as a sudden drop in output or an increase in the rate of quality failure, the early warning mechanism is triggered, and the evaluation and optimization process is started to generate an adjustment plan, and finally a dynamic optimization plan is obtained.

[0064] In this embodiment, through real-time evaluation and dynamic optimization, problems in the production process can be discovered and solved in a timely manner, thereby improving production efficiency. By real-time monitoring and adjustment of production parameters, product quality can be effectively controlled, the defective rate can be reduced, and the product qualification rate can be improved. By optimizing the production process, energy consumption can be reduced, and waste of raw materials can be reduced, thereby reducing production costs.

[0065] In an optional embodiment, The bubble network mechanism is used for global search, and the spiral update mechanism is used for local development and updating of population positions. An adaptive weight factor is introduced to dynamically adjust the weights of exploration and development. The update direction of the candidate solution is corrected based on the location information of the historical optimal solution. The optimal solutions include: Initialize a whale population containing multiple candidate solutions, each candidate solution represents a set of network node parameters, and the dimension of the candidate solution is the same as the number of network node parameters; A bubble network mechanism is used for global search, a reference point is randomly selected, a shrinking step size and a shrinking direction are calculated based on the distance between the reference point and the candidate solution, and the position of the candidate solution is updated for the first time according to the shrinking step size and the shrinking direction; Based on the spiral update mechanism, local development is performed. With the current optimal solution as the center, the candidate solutions are updated according to the spiral equation. The spiral equation includes a rotation angle parameter and a contraction distance parameter. The rotation angle parameter and the contraction distance parameter are used to control the rotation amplitude and radial movement distance of the candidate solutions in each dimension, and the position of the candidate solutions is updated for the second time. An adaptive weight factor is introduced to calculate the similarity between the candidate solution and the historical optimal network parameter configuration, and the weight ratio of the bubble network search and the spiral update is dynamically adjusted according to the similarity. When the similarity is lower than the preset similarity threshold, the weight of the bubble network search is increased, and when the similarity is higher than the similarity threshold, the weight of the spiral update is increased; Based on the location information of the historical optimal network parameter configuration, the update direction of the candidate solution is corrected, the direction vector of the candidate solution and the historical optimal solution is calculated, the direction vector is weightedly combined with the original update direction of the candidate solution to obtain a corrected update direction, and the location of the candidate solution is updated for the third time according to the corrected update direction; The update is repeated until the maximum number of iterations is reached or the preset convergence condition is met, and the optimal combination of network node parameters is output as the optimal solution.

[0066] The spiral update mechanism is a strategy in the improved whale optimization algorithm, which is used to optimize the update method of individual solutions. The bubble grid mechanism simulates the behavior of whales releasing bubbles to attract fish schools when hunting. When whales hunt through bubble grids, the shape and distribution of bubble grids will affect the hunting effect of whales.

[0067] Initialize the whale population. Assuming the network contains 10 nodes and each node needs to configure 3 parameters, each candidate solution is a 30-dimensional vector. The initial population contains 50 candidate solutions, and the 30 parameter values ​​of each candidate solution are randomly generated between -1 and 1.

[0068] Perform a global search. Use a bubble network mechanism to simulate the process of bubbles rising in water. Randomly select a candidate solution as a reference point and calculate the distance between other candidate solutions and the reference point. The farther the distance between the candidate solutions is, the larger the shrinking step size is, and the shrinking direction points to the reference point. For example, if the distance between candidate solution A and the reference point is 0.8, then its shrinking step size is set to 0.08, and the shrinking direction is the unit vector of A pointing to the reference point. According to the calculated shrinking step size and shrinking direction, the position of each candidate solution is updated for the first time.

[0069] Perform local development. With the current optimal solution as the center, use the spiral update mechanism to update the position of the candidate solution. The spiral update mechanism simulates the trajectory of a whale circling and rising around its prey. The rotation angle parameter and the contraction distance parameter control the rotation amplitude and radial movement distance of the candidate solution in each dimension. For example, if the rotation angle parameter is set to 0.5 and the contraction distance parameter is set to 0.1, the candidate solution rotates 0.5 radians in each dimension and moves 0.1 distance toward the optimal solution. Perform a second update on the position of each candidate solution.

[0070] An adaptive weight factor is introduced to dynamically adjust the weight ratio of global search and local development. The similarity between each candidate solution and the historical optimal network parameter configuration is calculated, and the similarity is calculated using cosine similarity. The preset similarity threshold is 0.8. If the similarity between the candidate solution and the historical optimal solution is lower than 0.8, the weight of the bubble network search is increased, for example, the weight of the bubble network search is set to 0.7 and the weight of the spiral update is set to 0.3; if the similarity is higher than 0.8, the weight of the spiral update is increased, for example, the weight of the bubble network search is set to 0.3 and the weight of the spiral update is set to 0.7.

[0071] The update direction of the candidate solution is corrected using the location information of the historical optimal solution. The direction vector of each candidate solution and the historical optimal solution is calculated, and the direction vector is weighted and combined with the original update direction of the candidate solution to obtain the corrected update direction. For example, the weight of the direction vector is set to 0.2, and the weight of the original update direction is set to 0.8. According to the corrected update direction, the position of each candidate solution is updated for the third time.

[0072] Repeat the update until the maximum number of iterations (e.g. 100) is reached or the pre-set convergence condition is met (e.g. the improvement of the optimal solution for 10 consecutive iterations is less than 0.001). Finally, the optimal combination of network node parameters is output as the optimal solution.

[0073] In this embodiment, by combining global search and local development strategies, the area where the optimal solution is located can be quickly located, and a detailed search can be performed, thereby improving the search efficiency. The bubble network mechanism can effectively explore the search space and avoid falling into the local optimal solution. The spiral update mechanism and the historical optimal solution guidance mechanism can help the algorithm find a better solution.

[0074] Figure 2 FIG. 1 is a schematic diagram of the structure of a multi-objective collaborative optimization control system for an ignition tool robot production line according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect the real-time operating parameters of the ignition tool robot production line, perform denoising on the real-time operating parameters through an adaptive soft threshold function based on an iterative wavelet threshold algorithm, decompose the real-time operating parameters into four layers of wavelet coefficients, and determine the threshold parameters by calculating the Shannon entropy in each iteration to obtain denoised data, perform maximum and minimum value normalization calculation on the denoised data and extract features in combination with an improved kernel principal component analysis algorithm, construct a mixed kernel matrix and calculate the feature dimension, obtain a mixed feature matrix and perform feature fusion, calculate multi-scale spatial feature weights in combination with dilated convolution and calculate channel feature weights in combination with a channel compression excitation structure, obtain a fused feature vector and calculate a fuzzy grey correlation degree, convert the production line optimization target into a production index value, calculate the evaluation weight through an improved ant colony algorithm, and obtain a target weight coefficient; The second unit is used to add the production index value, the target weight coefficient and the fusion feature vector to a preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantized orthogonal attention, and the optimal scheme combination is solved to generate a joint control strategy. The third unit is used to execute the joint control strategy, build a real-time evaluation model by improving the loop transformer model, set a recursive state update mechanism in the improved loop transformer model and calculate the relative position encoding, input the control execution result into the real-time evaluation model, calculate the timing characteristics through multi-head cross attention, obtain the production evaluation index, and calculate and update the network parameters of the hierarchical asynchronous advantage actor critic structure through meta-reinforcement learning, combine the differential evolution and simulated annealing methods, calculate the network node parameters and search for the optimal solution through the improved whale optimization algorithm, determine the timing characteristics of the historical data by combining the exponential sliding average, and adjust the learning parameters through the fuzzy cognitive map to obtain a dynamic optimization solution.

[0075] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0076] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0077] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective collaborative optimization control method for an ignition tool robot production line, characterized in that: include: The real-time operating parameters of the ignition tool robot production line are collected, and based on the iterative wavelet threshold algorithm, the real-time operating parameters are denoised by an adaptive soft threshold function, the real-time operating parameters are decomposed into four layers of wavelet coefficients, and the threshold parameters are determined by calculating the Shannon entropy in each iteration to obtain denoised data, the maximum and minimum normalization calculations are performed on the denoised data, and feature extraction is performed in combination with an improved kernel principal component analysis algorithm, a mixed kernel matrix is ​​constructed and feature dimensions are calculated, a mixed feature matrix is ​​obtained and feature fusion is performed, multi-scale spatial feature weights are calculated in combination with dilated convolution and channel feature weights are calculated in combination with channel compression excitation structure, a fused feature vector is obtained and fuzzy grey correlation is calculated, the production line optimization target is converted into a production index value, the evaluation weight is calculated by an improved ant colony algorithm, and the target weight coefficient is obtained; The production index value, the target weight coefficient and the fusion feature vector are added to the preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantized orthogonal attention, and the optimal scheme combination is solved to generate a joint control strategy. The joint control strategy is implemented, a real-time evaluation model is constructed by improving the loop transformer model, a recursive state update mechanism is set in the improved loop transformer model and the relative position encoding is calculated, the control execution result is input into the real-time evaluation model, the timing characteristics are calculated by multi-head cross attention, the production evaluation index is obtained, and the network parameters of the hierarchical asynchronous advantage actor critic structure are calculated and updated by meta-reinforcement learning, the network node parameters are calculated and the optimal solution is searched by improving the whale optimization algorithm in combination with differential evolution and simulated annealing methods, the timing characteristics of historical data are determined by combining the exponential sliding average, and the learning parameters are adjusted by the fuzzy cognitive map to obtain a dynamic optimization solution.

2. The method according to claim 1, characterized in that The real-time operating parameters of the ignition tool robot production line are collected, and based on the iterative wavelet threshold algorithm, the real-time operating parameters are denoised by an adaptive soft threshold function, the real-time operating parameters are decomposed into four layers of wavelet coefficients, and the threshold parameters are determined by calculating the Shannon entropy in each iteration to obtain denoised data, and the maximum and minimum value normalization calculations are performed on the denoised data, and feature extraction is performed in combination with the improved kernel principal component analysis algorithm, a mixed kernel matrix is ​​constructed, and the feature dimension is calculated, and the mixed feature matrix obtained includes: The real-time operating parameters of the ignition tool robot production line are collected by sensors, wherein the real-time operating parameters include current parameters, pressure parameters, temperature parameters and vibration frequency parameters of the equipment on the production line; Decomposing the real-time operation parameter into four layers of wavelet coefficients by wavelet transform, calculating the Shannon entropy value of each layer according to the wavelet coefficients, calculating the adaptive threshold value based on the Shannon entropy value, constructing a soft threshold function according to the adaptive threshold value, when the absolute value of the input wavelet coefficient is greater than the adaptive threshold value, the output of the soft threshold function is equal to the sign of the wavelet coefficient multiplied by the difference between the absolute value of the wavelet coefficient and the adaptive threshold value, when the absolute value of the input wavelet coefficient is less than or equal to the adaptive threshold value, the output of the soft threshold function is zero, and substituting the wavelet coefficient of each layer into the soft threshold function to obtain the wavelet coefficient after threshold processing; Constructing a reconstruction formula, wherein the reconstruction formula is equal to the sum of the product of the scale function and the approximate coefficient of the last layer, plus the sum of the product of the wavelet function and the detail coefficients of each layer after processing, substituting the wavelet coefficients after the threshold processing into the reconstruction formula, calculating the denoised parameter sequence, performing maximum and minimum value normalization processing on the denoised parameter sequence, and mapping the parameter sequence to the interval from zero to one; A hybrid kernel matrix is ​​obtained by linearly combining a linear kernel matrix, a polynomial kernel matrix, a Gaussian kernel matrix and a Laplace kernel matrix according to preset weight coefficients, and the hybrid kernel matrix is ​​eigendecomposed, and the decomposed features are sorted in descending order according to the eigenvalues, and the features in the top 10% are selected to construct a mapping feature matrix. The normalized parameter sequence is mapped to the feature space according to the mapping feature matrix to obtain the hybrid feature matrix.

3. The method according to claim 1, characterized in that: Perform feature fusion, combine dilated convolution to calculate multi-scale spatial feature weights and combine channel compression excitation structure to calculate channel feature weights, obtain fused feature vectors and calculate fuzzy grey correlation, convert production line optimization targets into production index values, calculate evaluation weights through improved ant colony algorithm, and obtain target weight coefficients including: For the mixed feature matrix, multi-scale spatial features are extracted by dilated convolution with different dilation rates, the multi-scale spatial features are weightedly fused to obtain a spatial feature matrix, channel features are extracted by using a channel compression excitation structure, a channel description vector is obtained by global average pooling, a channel weight vector is obtained by passing through two fully connected layers and a Sigmoid activation function, the channel weight vector is multiplied by the mixed feature matrix to obtain a channel feature matrix, and the spatial feature matrix and the channel feature matrix are spliced ​​in the channel dimension to obtain a fused feature matrix; Quantifying the production line optimization target into a production index, calculating the fuzzy grey correlation coefficient between each feature in the fusion feature matrix and each production index, and calculating the correlation between each feature and the production index based on the fuzzy grey correlation coefficient; Construct a complete graph of production indicators, take each indicator as a node in the graph, assign importance and historical evaluation score attribute values ​​to each node, calculate the similarity of attribute values ​​of adjacent nodes as the heuristic information of the current node, and initialize the pheromone concentration of the edges between nodes; Placing a plurality of ants with different search strategies in the complete graph, including global search ants and local search ants, wherein the global search ants give priority to unvisited paths, and control the local search ants to move near known paths, multiplying the pheromone concentration term, the heuristic information term, and the taboo term, and then normalizing them to obtain the adaptive transfer probability of each ant, dynamically adjusting the weight coefficients of the pheromone concentration term and the heuristic information term, and reducing the probability of repeated visits to similar nodes; Determine the path length, node attribute variance and path smoothness of each ant traversal path, and obtain a comprehensive evaluation value by weighted summation. Sort all paths according to the comprehensive evaluation value of the data, superimpose local pheromones on the currently traversed path, enhance the pheromone concentration on the optimal path every three iterations, apply random perturbations to the suboptimal path, randomly select nodes on the optimal path and exchange positions with neighboring nodes, compare the path evaluation values ​​before and after the mutation, retain the path with the highest path evaluation value and calculate the target weight coefficient corresponding to the production indicator in combination with the node attribute value.

4. The method according to claim 1, characterized in that The production index value, the target weight coefficient and the fusion feature vector are added to the preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantizing orthogonal attention, and the optimal scheme combination is solved and the joint control strategy is generated, including: The production index value, the target weight coefficient and the fusion feature vector are input into a hierarchical asynchronous advantage actor-critic model, wherein the hierarchical asynchronous advantage actor-critic model includes a policy network layer and a value network layer. In the policy network layer, an improved proximal policy optimization algorithm is used to calculate the robot trajectory planning scheme, and the Kulbeck-Leibler divergence between the current policy distribution and the target policy distribution is calculated, and the upper and lower bounds of the policy clipping range are dynamically adjusted according to the Kulbeck-Leibler divergence. The policy parameters are updated within the adjusted clipping range and the optimization is iteratively optimized until the policy converges. Through the hierarchical dual-state action-value function, the decision task is mapped to the macro state space and the micro state space, the movement trend is determined by calculating the policy gradient in the macro state space, the input data is decomposed into macro strategies, and micro actions are generated by action sampling in the micro state space. Based on the macro strategies and the micro actions, multiple production line speed control schemes are obtained; For each production line speed control scheme, a search tree with the current state as the root node is constructed, and multiple child nodes are expanded in parallel to generate candidate schemes; a Monte Carlo simulation is performed on each candidate scheme, and the value estimation of the scheme is calculated in combination with an improved soft actor critic algorithm, wherein the soft actor critic algorithm uses a temperature parameter to adjust the balance between exploration and utilization, and selects the optimal energy consumption allocation scheme according to the value estimation; In the value network layer, the value function of the control scheme is calculated by quantizing the orthogonal attention mechanism, the attention matrix is ​​constructed to calculate the correlation between different states and the attention weights are orthogonally constrained to eliminate redundancy, the computational complexity is reduced by quantifying the attention output, and then the multi-head attention mechanism is integrated to extract multi-scale features, the state value function is calculated based on the attention output, the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are combined to form a candidate strategy set, and the strategy combination with the largest value function is selected as the optimal joint control strategy by calculating the comprehensive value function of each candidate strategy.

5. The method according to claim 4, characterized in that The value estimation of the scheme combined with the improved soft actor critic algorithm is shown in the following formula: ； in, τ 0 represents the initial temperature value, δ 1 represents the temperature decay rate, t represents the time step, τ min Indicates the minimum temperature value, δ 2 represents the reward error decay rate, r avg represents the average reward value of the current strategy, r target Represents the target reward value.

6. The method according to claim 1, characterized in that The joint control strategy is implemented, a real-time evaluation model is constructed by improving the cyclic transformer model, a recursive state update mechanism is set in the improved cyclic transformer model and the relative position encoding is calculated, the control execution result is input into the real-time evaluation model, the time series characteristics are calculated by multi-head cross attention, the production evaluation index is obtained, and the network parameters of the hierarchical asynchronous advantage actor critic structure are calculated and updated by meta-reinforcement learning, the network node parameters are calculated and the optimal solution is searched by improving the whale optimization algorithm by combining the differential evolution and simulated annealing methods, the time series characteristics of the historical data are determined by combining the exponential sliding average, and the learning parameters are adjusted by the fuzzy cognitive map, and the dynamic optimization scheme is obtained, including: Execute the joint control strategy, and at the same time, build a real-time evaluation model based on the improved recurrent transformer model, concatenate the control execution result and the historical state representation to form an input feature vector, and calculate the input feature vector through the multi-head self-attention mechanism of the recurrent transformer unit and the feedforward neural network to obtain an updated state representation, and introduce a relative position encoding matrix when calculating the attention weight to realize the representation of the timing information; The time series interaction features of state representation and production index are calculated through the multi-head cross attention mechanism. The state representation sequence is used as the query vector and the production index is used as the key value vector to be input into the multi-head cross attention module. The attention weights of the query vector and the key value vector are calculated to perform weighted aggregation on the value vector to obtain the output features of each attention head. The output features of each attention head are concatenated and linearly transformed to obtain the time series interaction feature vector. The temporal interaction feature vector is converted into a production evaluation index through a fully connected layer and an activation function, and the production evaluation index is used as an immediate reward for meta-reinforcement learning. The actor network parameters are updated through a policy gradient algorithm to optimize the control strategy. The critic network parameters are updated and the state action value function is estimated by combining a temporal difference algorithm. Differential evolution and simulated annealing mechanisms are introduced in the parameter updating process to improve the exploration ability of the model. Based on multiple production evaluation indicators, an improved whale optimization algorithm is used for multi-objective solution, and a whale population containing multiple candidate weight coefficients is initialized, wherein each individual in the whale population represents a candidate solution, and the comprehensive evaluation index of each candidate solution under the current weight coefficient is calculated and used as the fitness value, and a global search is performed through the bubble network mechanism, and the local development and update of the population position are combined with the spiral update mechanism. An adaptive weight factor is introduced to dynamically adjust the weights of exploration and development, and the update direction of the candidate solution is corrected based on the position information of the historical optimal solution to obtain the optimal solution; Based on the optimal solution, the exponential sliding average of the pre-acquired historical state representation sequence is performed to extract stable time series features, a fuzzy cognitive map is constructed to set fuzzy rules for the causal relationship between production indicators and the triggering conditions and impact intensity of the fuzzy rules, and the learning parameters of the real-time evaluation model are dynamically adjusted through the fuzzy reasoning mechanism according to the current production indicator value. When an abnormal deviation occurs in the production indicator, the early warning mechanism is triggered and the evaluation optimization process is started to generate an adjustment plan, thereby obtaining the dynamic optimization plan.

7. The method according to claim 6, characterized in that The bubble network mechanism is used for global search, and the spiral update mechanism is used for local development and updating of population positions. An adaptive weight factor is introduced to dynamically adjust the weights of exploration and development. The update direction of the candidate solution is corrected based on the location information of the historical optimal solution. The optimal solutions include: Initialize a whale population containing multiple candidate solutions, each candidate solution represents a set of network node parameters, and the dimension of the candidate solution is the same as the number of network node parameters; A bubble network mechanism is used for global search, a reference point is randomly selected, a shrinking step size and a shrinking direction are calculated based on the distance between the reference point and the candidate solution, and the position of the candidate solution is updated for the first time according to the shrinking step size and the shrinking direction; Based on the spiral update mechanism, local development is performed. With the current optimal solution as the center, the candidate solutions are updated according to the spiral equation. The spiral equation includes a rotation angle parameter and a contraction distance parameter. The rotation angle parameter and the contraction distance parameter are used to control the rotation amplitude and radial movement distance of the candidate solutions in each dimension, and the position of the candidate solutions is updated for the second time. An adaptive weight factor is introduced to calculate the similarity between the candidate solution and the historical optimal network parameter configuration, and the weight ratio of the bubble network search and the spiral update is dynamically adjusted according to the similarity. When the similarity is lower than the preset similarity threshold, the weight of the bubble network search is increased, and when the similarity is higher than the similarity threshold, the weight of the spiral update is increased; Based on the location information of the historical optimal network parameter configuration, the update direction of the candidate solution is corrected, the direction vector of the candidate solution and the historical optimal solution is calculated, the direction vector is weightedly combined with the original update direction of the candidate solution to obtain a corrected update direction, and the location of the candidate solution is updated for the third time according to the corrected update direction; The update is repeated until the maximum number of iterations is reached or the preset convergence condition is met, and the optimal combination of network node parameters is output as the optimal solution.

8. A multi-objective collaborative optimization control system for an ignition tool robot production line, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the real-time operating parameters of the ignition tool robot production line, perform denoising on the real-time operating parameters through an adaptive soft threshold function based on an iterative wavelet threshold algorithm, decompose the real-time operating parameters into four layers of wavelet coefficients, and determine the threshold parameters by calculating the Shannon entropy in each iteration to obtain denoised data, perform maximum and minimum value normalization calculation on the denoised data and extract features in combination with an improved kernel principal component analysis algorithm, construct a mixed kernel matrix and calculate the feature dimension, obtain a mixed feature matrix and perform feature fusion, calculate multi-scale spatial feature weights in combination with dilated convolution and calculate channel feature weights in combination with a channel compression excitation structure, obtain a fused feature vector and calculate a fuzzy grey correlation degree, convert the production line optimization target into a production index value, calculate the evaluation weight through an improved ant colony algorithm, and obtain a target weight coefficient; The second unit is used to add the production index value, the target weight coefficient and the fusion feature vector to a preset hierarchical asynchronous advantage actor-critic model. In the strategy network layer, the robot trajectory planning scheme is calculated by optimizing the improved proximal strategy, combined with the adaptive Kulbeck-Leibler divergence constraint and dynamically adjusting the clipping range. The decision task is decomposed into macro-strategies and micro-actions through the hierarchical dual-state action value function to obtain the production line speed control scheme. The action space is expanded based on the pseudo-parallel Monte Carlo tree and combined with the improved soft actor-critic algorithm to calculate the energy consumption allocation scheme. In the value network layer, the value functions corresponding to the robot trajectory planning scheme, the production line speed control scheme and the energy consumption allocation scheme are calculated by quantized orthogonal attention, and the optimal scheme combination is solved to generate a joint control strategy. The third unit is used to execute the joint control strategy, build a real-time evaluation model by improving the loop transformer model, set a recursive state update mechanism in the improved loop transformer model and calculate the relative position encoding, input the control execution result into the real-time evaluation model, calculate the timing characteristics through multi-head cross attention, obtain the production evaluation index, and calculate and update the network parameters of the hierarchical asynchronous advantage actor critic structure through meta-reinforcement learning, combine the differential evolution and simulated annealing methods, calculate the network node parameters and search for the optimal solution through the improved whale optimization algorithm, determine the timing characteristics of the historical data by combining the exponential sliding average, and adjust the learning parameters through the fuzzy cognitive map to obtain a dynamic optimization solution.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images