Chip production logistics optimization method and system based on multi-objective genetic algorithm

Abstract

The present invention provides a chip production logistics optimization method and system based on a multi-objective genetic algorithm, which relates to the technical field of chip production scheduling, including collecting real-time operation data of a production line, extracting task execution features and equipment layout features using a dual-channel deep neural network, and performing feature fusion through a recurrent neural network to generate a production state vector containing multidimensional scheduling index prediction values; constructing a multidimensional optimization target set based on the production state vector, and using a double-layer gene coding structure to generate an initial population; performing non-dominated sorting and crowding distance calculation based on fitness values, and selecting high-quality individuals to form a parent population; performing adaptive crossover, taboo search optimization and adaptive mutation operations on the parent population, iteratively selecting a new generation population until convergence, and outputting an optimal scheduling solution.

Description

Technical Field

[0001] The present invention relates to the technical field of chip production scheduling, and in particular to a chip production logistics optimization method and system based on a multi-objective genetic algorithm. Background Art

[0002] With the rapid development of the integrated circuit industry, production scheduling and logistics optimization problems in the chip manufacturing process are becoming increasingly complex. Chip production lines have the characteristics of long process routes, diverse processing equipment, and frequent material transportation. They need to take into account multiple optimization objectives such as task completion time, equipment utilization efficiency, and logistics path length. Traditional heuristic scheduling methods are difficult to effectively handle complex scheduling scenarios under multi-objective constraints, and intelligent optimization methods based on deep learning and evolutionary algorithms provide new ideas for solving such problems.

[0003] However, there are some shortcomings in the existing technology. Traditional scheduling algorithms often use a single data processing model, which is unable to capture the task timing characteristics and equipment space characteristics at the same time, resulting in inaccurate description of complex scenarios of chip production lines; existing genetic coding methods usually mix task sorting, equipment allocation and path planning, making the search space too large and affecting the convergence efficiency of the algorithm; conventional genetic algorithms use fixed evolutionary operator parameters, lack an adaptive adjustment mechanism, are prone to fall into local optimal solutions, and find it difficult to achieve a good balance between multiple optimization objectives.

[0004] In summary, a dual-channel deep neural network is constructed to realize multi-dimensional feature extraction and fusion of production data, a double-layer gene coding structure is used to reduce the search complexity of the problem, and an adaptive evolutionary operator is used to improve the effect of multi-objective optimization, thereby realizing comprehensive optimization of the chip production line in three dimensions: task completion time, equipment utilization efficiency, and logistics path length. This provides an intelligent scheduling solution that can adapt to complex production environments. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a chip production logistics optimization method and system based on a multi-objective genetic algorithm, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A chip production logistics optimization method based on a multi-objective genetic algorithm is provided, comprising:

[0008] Collect real-time operation data of the chip production line, and process the data through a dual-channel deep neural network, including a first channel and a second channel, wherein the first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer; input the task execution features and the equipment layout features into a recurrent neural network, perform feature fusion, determine multidimensional scheduling indicators including task completion time, equipment utilization efficiency, and logistics path length, and generate a production state vector including multidimensional scheduling indicator prediction values; based on the production state vector, construct a multidimensional optimization target set including task sorting priority, equipment resource allocation, and logistics path planning;

[0009] Based on the multi-dimensional optimization target set, a double-layer gene coding structure is constructed, the task sorting priority is encoded as an outer gene sequence, and the equipment resource allocation and the logistics path planning are encoded as an inner gene sequence; the double-layer gene coding structure is used to generate an initial population, and the fitness value of each individual in the initial population on the multi-dimensional scheduling index is calculated; based on the fitness value, a non-dominated sort is performed on the initial population to obtain a population stratification result, and the crowding distance between individuals is calculated; according to the population stratification result and the crowding distance, high-quality individuals are selected to form a parent population;

[0010] An adaptive crossover operation is performed on the individuals of the parent population, and the crossover position is dynamically adjusted based on the fitness value of the individuals to generate offspring individuals; taboo search optimization is performed on the offspring individuals to generate a local search solution; the mutation probability is adaptively adjusted according to the distribution density of the individuals in the solution space, and a mutation operation is performed on the local search solution to generate mutant individuals; the mutant individuals are merged with the parent individuals, and a new generation of population is iteratively selected; the execution is repeated until the population converges, and a scheduling plan that achieves optimal balance on multi-dimensional scheduling indicators is output.

[0011] In an optional embodiment,

[0012] Collect real-time operation data of the chip production line and process the data through a dual-channel deep neural network, including a first channel and a second channel. The first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer, including:

[0013] Collecting real-time operation data of the chip production line, the real-time operation data includes equipment operation status data, material transmission speed data and buffer area occupancy data;

[0014] According to a preset fixed length, a data detection window is constructed, a distribution skewness value of the real-time operation data is calculated within the data detection window, abnormal data points are identified according to a preset skewness threshold, and corrected data is obtained after the abnormal data points are removed;

[0015] Perform fast Fourier transform to process the correction data, extract the fluctuation frequency value of the logistics system, calculate and set the hole convolution rate of the first branch, the second branch and the third branch in the timing feature extraction network according to the fluctuation frequency value; input the correction data to the timing feature extraction network, extract local timing features through the dynamic deformable convolution layer in the first branch, the second branch and the third branch, and the dynamic deformable convolution layer adaptively adjusts the convolution sampling position and weight parameters according to the distribution characteristics of the local timing features; calculate the conditional entropy value and mutual information value of the local timing features, determine the weight coefficients of the features at different time scales based on the conditional entropy value and the mutual information value, and obtain the task execution features by weighted fusion;

[0016] Construct a heterogeneous graph network model of the production line, model the processing equipment, transmission equipment and cache area as processing nodes, transportation nodes and storage nodes respectively, and obtain the correlation strength between nodes through multi-head attention calculation based on the actual distance and historical logistics data between the processing nodes, the transportation nodes and the storage nodes, and generate an initial adjacency matrix;

[0017] Based on the heterogeneous graph network model of the production line, a community discovery algorithm is executed to identify the load intensity and transmission capacity saturation areas of the logistics channel from the initial adjacency matrix. Based on the corresponding spatial distribution law, the connection relationship between the processing nodes, transportation nodes and storage nodes is adjusted, and the initial adjacency matrix is ​​updated to obtain an optimized adjacency matrix. The optimized adjacency matrix is ​​input into a graph convolutional network based on Chebyshev polynomials to extract and generate equipment layout features.

[0018] In an optional embodiment,

[0019] The task execution features and the equipment layout features are input into a recurrent neural network for feature fusion, and a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length is determined. A production state vector including a predicted value of the multidimensional scheduling index is generated, including:

[0020] Inputting the task execution features and the equipment layout features into a double-layer recurrent neural network, wherein the double-layer recurrent neural network includes a feature decomposition layer and a feature combination layer;

[0021] A plurality of recursive processing units of time scales are arranged in the feature decomposition layer, the task execution features and the equipment layout features are divided into a plurality of groups of time series segments according to preset time intervals, and the segments are respectively input into the recursive processing units of the corresponding time scales to generate a plurality of groups of time series features; the recursive processing units of different time scales are connected by a feature cascade method to realize information transmission between the plurality of groups of time series features and obtain the fused time series features;

[0022] Calculate the information entropy value and mutual information value of the fused time series features to generate feature channel weights; calculate the position coding difference and time distance of adjacent time series segments in the fused time series features to generate time series position weights;

[0023] The feature channel weight and the time sequence position weight are input into an adaptive gating unit, and the adaptive gating unit dynamically adjusts the information flow ratio according to the weight to generate a fusion weight;

[0024] The fusion weight is multiplied by the fused time series feature to obtain a weighted feature vector; the weighted feature vector is input into a fully connected layer to generate a production state vector including a task completion time prediction value, an equipment utilization efficiency prediction value, and a logistics path length prediction value.

[0025] In an optional embodiment,

[0026] Based on the multi-dimensional optimization target set, a double-layer gene encoding structure is constructed, the task sorting priority is encoded as an outer gene sequence, and the equipment resource allocation and the logistics path planning are encoded as an inner gene sequence; the double-layer gene encoding structure is used to generate an initial population, and the fitness value of each individual in the initial population on the multi-dimensional scheduling index is calculated; based on the fitness value, a non-dominated sorting is performed on the initial population to obtain a population stratification result, and the crowding distance between individuals is calculated; and according to the population stratification result and the crowding distance, high-quality individuals are selected to form a parent population, which includes:

[0027] Construct a double-layer gene coding structure, encode the execution priority of the task to be processed as a real number between zero and one to form an outer gene sequence, encode the equipment allocation plan and the logistics path plan of the task to be processed as integers to form an inner gene sequence, the equipment allocation plan in the inner gene sequence corresponds to the equipment number, and the logistics path plan corresponds to the path number;

[0028] Through the double-layer gene coding structure, a preset number of individuals are randomly generated to form an initial population, and each individual in the initial population includes a group of the outer gene sequences and a group of the inner gene sequences;

[0029] Prioritize the tasks to be processed according to the numerical values ​​of the outer gene sequence, read the corresponding equipment numbers and path numbers in the inner gene sequence in sequence according to the priority ranking, generate a scheduling plan, count the maximum completion time of the scheduling plan, the inverse of the minimum equipment utilization rate and the total load of the logistics path, and calculate the multidimensional fitness value;

[0030] Performing non-dominated sorting on the initial population, dividing the initial population into multiple population levels based on dominance relations; calculating the crowding distance of each individual in the initial population, sorting the multidimensional fitness values ​​of individuals in the same population level according to each dimension, calculating the difference in multidimensional fitness values ​​of adjacent individuals, dividing the difference in multidimensional fitness values ​​by the extreme difference in fitness values ​​on the corresponding dimension to obtain a standardized distance, summing the standardized distances of each dimension to obtain the crowding distance of the individual;

[0031] Two individuals are randomly selected for tournament selection, and the individual with a smaller number corresponding to the population level is selected. When the population levels of the two individuals are the same, the individual with a larger crowding distance is selected, and the tournament selection is repeated until a preset number of individuals are selected to form the parent population.

[0032] In an optional embodiment,

[0033] Performing non-dominated sorting on the initial population, and dividing the initial population into a plurality of population levels based on dominance relationships includes:

[0034] Constructing a fitness difference record matrix and a dominance relationship mark matrix, wherein the number of rows and columns of the fitness difference record matrix and the dominance relationship mark matrix is ​​equal to the number of individuals in the initial population;

[0035] For any two individuals in the initial population, compare the fitness values ​​in three dimensions: maximum completion time, the inverse of minimum equipment utilization, and total load of the logistics path, calculate the fitness difference between the two individuals in the three dimensions, fill the calculated fitness difference into the corresponding position of the fitness difference record matrix, calculate in sequence, and complete the filling action of the fitness difference record matrix;

[0036] Read the difference information stored in the fitness difference record matrix to determine whether the first individual dominates the second individual. When the fitness values ​​of the first individual in the three dimensions of maximum completion time, the inverse of the minimum equipment utilization rate, and the total load of the logistics path are all less than or equal to the second individual, and at least one dimension is less than the second individual, mark the first individual dominating the second individual at the corresponding position of the domination relationship marking matrix, and make judgments in turn to complete the marking action of the domination relationship marking matrix;

[0037] Scan each row and each column of the dominance relationship labeling matrix, count the number of times each individual is dominated by other individuals, record the dominance count of the corresponding individual, and record other individuals dominated by each individual to form a set of dominated individuals;

[0038] Select individuals with a dominance count of zero and place them in the first frontier layer. Traverse the dominated individual set of each individual in the first frontier layer, reduce the dominance count of the individual in the dominated individual set by one, and obtain the updated dominance count. When the updated dominance count becomes zero, place the corresponding individual in the second frontier layer.

[0039] Repeat the process until all individuals in the initial population are divided into the corresponding frontier layers, and the population level division result is obtained.

[0040] In an optional embodiment,

[0041] Performing an adaptive crossover operation on the individuals of the parent population, dynamically adjusting the crossover position based on the fitness value of the individual, and generating offspring individuals; performing taboo search optimization on the offspring individuals to generate a local search solution; adaptively adjusting the mutation probability according to the distribution density of the individual in the solution space, performing a mutation operation on the local search solution to generate a mutant individual; merging the mutant individual with the parent individual, iteratively selecting a new generation population; repeating the execution until the population converges, and outputting a scheduling solution that achieves optimal balance on multi-dimensional scheduling indicators, including:

[0042] Calculate the fitness value of each parent individual in the parent population on the multidimensional scheduling index, and divide the fitness value by the total fitness value of the parent individual on the corresponding dimension to obtain a normalized fitness value;

[0043] Calculate the product of the arithmetic mean of the normalized fitness value and the adaptive adjustment factor to determine the crossover position selection probability, determine the crossover position of the parent individuals according to the crossover position selection probability, and exchange the gene fragments of the two parent individuals between the crossover positions to generate offspring individuals;

[0044] Construct a taboo table, and perform neighborhood search for task order exchange, equipment allocation change, and logistics path adjustment under the constraints of the taboo table. Use the weighted fitness value as a criterion to select the optimal neighborhood action and update the taboo table. After multiple iterations, a local search solution is obtained.

[0045] Perform task sequence mutation and device path mutation on the local search solution according to the mutation probability to generate mutant individuals; merge the mutant individuals with the parent individuals to generate a temporary population;

[0046] Calculating the change of each temporary individual in the temporary population in the multidimensional scheduling index relative to the parent population, and the distribution uniformity of the individuals in the three-dimensional target space;

[0047] Based on the change amount and the distribution uniformity, when the difference in the change amounts of adjacent temporary individuals is less than a preset change amount, secondary sorting is performed in descending order of the distribution uniformity, and a corresponding number of temporary individuals are selected from the order after the secondary sorting according to a preset selection ratio to form a new generation population;

[0048] Repeat the process and output a scheduling solution that achieves the optimal balance on multi-dimensional scheduling indicators.

[0049] In an optional embodiment,

[0050] A taboo table is constructed, and neighborhood search for task order exchange, equipment allocation change, and logistics path adjustment is performed under the constraints of the taboo table. The weighted fitness value is used as a criterion to select the optimal neighborhood action and update the taboo table. After multiple iterations, the local search solutions include:

[0051] The maximum capacity of the taboo table is set to the weighted sum of the number of tasks, the number of devices, and the number of feasible paths;

[0052] The first type of neighborhood solution set is generated by exchanging the processing order of adjacent tasks, the second type of neighborhood solution set is generated by changing the processing equipment of the task, and the third type of neighborhood solution set is generated by adjusting the logistics path of the task. The first type of neighborhood solution set, the second type of neighborhood solution set and the third type of neighborhood solution set are combined to generate the candidate neighborhood solution of the offspring individual:

[0053] Calculate the fitness value of each candidate neighborhood solution on the multi-dimensional scheduling index, determine the evaluation index based on the fitness value through weighted summation, and record the current optimal evaluation index;

[0054] Traverse each neighborhood solution. If the action corresponding to the current neighborhood solution is not in the taboo table and the corresponding evaluation index is greater than the current optimal evaluation index, the current neighborhood solution is taken as a candidate optimal solution. If the action corresponding to the current neighborhood solution is in the taboo table and the fitness value of any dimension in the corresponding multi-dimensional scheduling index is greater than the historical optimal value, the current neighborhood solution is taken as a candidate optimal solution.

[0055] Execute the optimal action corresponding to the candidate optimal solution, update the current neighborhood solution, determine the taboo action based on the reverse action of the optimal action, and add the taboo action to the taboo table; when the taboo table exceeds the preset capacity, delete the earliest added taboo action;

[0056] Repeat until the preset number of iterations is reached, and the final solution obtained is used as the local search solution.

[0057] According to a second aspect of the embodiments of the present invention,

[0058] A chip production logistics optimization system based on a multi-objective genetic algorithm is provided, comprising:

[0059] The first unit is used to collect real-time operation data of the chip production line and perform data processing through a dual-channel deep neural network, including a first channel and a second channel, wherein the first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer; the task execution features and the equipment layout features are input into a recurrent neural network for feature fusion, and a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length is determined, and a production state vector including a predicted value of the multidimensional scheduling index is generated; based on the production state vector, a multidimensional optimization target set including task sorting priority, equipment resource allocation, and logistics path planning is constructed;

[0060] The second unit is used to construct a double-layer gene coding structure based on the multi-dimensional optimization target set, encode the task sorting priority as an outer gene sequence, and encode the equipment resource allocation and the logistics path planning as an inner gene sequence; generate an initial population using the double-layer gene coding structure, calculate the fitness value of each individual in the initial population on the multi-dimensional scheduling index; perform non-dominated sorting on the initial population based on the fitness value, obtain a population stratification result, and calculate the crowding distance between individuals; select high-quality individuals to form a parent population according to the population stratification result and the crowding distance;

[0061] The third unit is used to perform an adaptive crossover operation on the individuals of the parent population, dynamically adjust the crossover position based on the fitness value of the individuals, and generate offspring individuals; perform taboo search optimization on the offspring individuals to generate a local search solution; adaptively adjust the mutation probability according to the distribution density of the individuals in the solution space, perform a mutation operation on the local search solution to generate mutant individuals; merge the mutant individuals with the parent individuals, and iteratively select a new generation of population; repeat the execution until the population converges, and output a scheduling plan that achieves optimal balance on multi-dimensional scheduling indicators.

[0062] According to a third aspect of the embodiments of the present invention,

[0063] An electronic device is provided, comprising:

[0064] processor;

[0065] a memory for storing processor-executable instructions;

[0066] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0067] A fourth aspect of the embodiments of the present invention is:

[0068] 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.

[0069] In an embodiment of the present invention, task execution features and equipment layout features are extracted through a dual-channel deep neural network, and feature fusion is performed in combination with a recurrent neural network, so that the production status can be predicted more accurately, thereby optimizing task sorting priority, equipment resource allocation and logistics path planning, and ultimately shortening the task completion time and improving equipment utilization efficiency; through multi-objective genetic algorithm and taboo search optimization, the logistics path can be effectively planned, the logistics path length can be shortened, and the logistics cost and time can be reduced, thereby improving the overall production efficiency; a multi-dimensional optimization target set including task completion time, equipment utilization efficiency and logistics path length is constructed, and multi-objective balanced optimization is achieved through methods such as non-dominated sorting and congestion distance calculation, and finally a scheduling plan that achieves the optimal balance on multi-dimensional scheduling indicators is output. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the process of a chip production logistics optimization method based on a multi-objective genetic algorithm according to an embodiment of the present invention;

[0071] Figure 2 It is a structural schematic diagram of a chip production logistics optimization system based on a multi-objective genetic algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] 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.

[0073] 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.

[0074] Figure 1 FIG. 1 is a flow chart of a chip production logistics optimization method based on a multi-objective genetic algorithm according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0075] S101. Collect real-time operation data of the chip production line, and process the data through a dual-channel deep neural network, including a first channel and a second channel, wherein the first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer; the task execution features and the equipment layout features are input into a recurrent neural network for feature fusion, and a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length is determined, and a production state vector including a predicted value of the multidimensional scheduling index is generated; based on the production state vector, a multidimensional optimization target set including task sorting priority, equipment resource allocation, and logistics path planning is constructed;

[0076] Real-time operation data such as process flow data, equipment status data and logistics trajectory data are collected from the chip production line, and a dual-channel deep neural network is constructed for feature extraction: the first channel uses the temporal convolution layer to process the temporal dependency and processing time series of task execution to extract task execution features; the second channel uses the spatial convolution layer to process the positional relationship and connection topology of the equipment layout to extract the equipment layout features.

[0077] The extracted task execution features and equipment layout features are input into the recurrent neural network for multi-dimensional feature fusion. The network learns the importance weights of different features through the gating mechanism and establishes dynamic associations between features. Based on the fused feature representation, multi-dimensional scheduling indicators such as task completion time, equipment utilization efficiency, and logistics path length are predicted, and a production state vector containing these predicted values ​​is generated.

[0078] Based on the production state vector, combined with the chip production process requirements and resource constraints, a multi-dimensional optimization target set is constructed. This target set includes: minimizing the task sorting priority index based on task urgency and process dependency, maximizing the resource allocation efficiency index considering equipment capacity matching, and optimizing the logistics path planning index taking into account distance and congestion, providing evaluation criteria for the subsequent evolutionary optimization algorithm.

[0079] In this embodiment, by collecting the operation data of the chip production line in real time, the dynamic changes in the production process can be captured in time, and the latest input information can be provided for subsequent optimization; a dual-channel deep neural network is used to extract task execution features and equipment layout features respectively, which can fully consider different types of information and improve the accuracy and comprehensiveness of feature extraction; the temporal convolution layer of the first channel and the spatial convolution layer of the second channel focus on the timing changes and spatial layout respectively, ensuring that multi-dimensional features can be effectively captured, providing more accurate data support for subsequent scheduling optimization; feature fusion is performed through a recurrent neural network, and the multi-dimensional information of task execution and equipment layout is integrated to generate multi-dimensional scheduling indicators including task completion time, equipment utilization efficiency and logistics path length, providing comprehensive prediction data for production scheduling; based on the production state vector, a multi-dimensional optimization target set is constructed, including task sorting priority, equipment resource allocation and logistics path planning, to ensure the comprehensive optimization of the scheduling plan and meet the needs of production efficiency, equipment utilization and logistics path.

[0080] S102. Based on the multi-dimensional optimization target set, a double-layer gene encoding structure is constructed, the task sorting priority is encoded as an outer gene sequence, and the equipment resource allocation and the logistics path planning are encoded as an inner gene sequence; the initial population is generated by using the double-layer gene encoding structure, and the fitness value of each individual in the initial population on the multi-dimensional scheduling index is calculated; based on the fitness value, a non-dominated sort is performed on the initial population to obtain a population stratification result, and the crowding distance between individuals is calculated; according to the population stratification result and the crowding distance, high-quality individuals are selected to form a parent population;

[0081] In this embodiment, through multi-dimensional target optimization, multiple factors such as task sorting priority, equipment resource allocation and logistics path planning can be comprehensively considered to ensure that the scheduling scheme is balanced and optimized in multiple dimensions; the outer gene sequence is used to encode the task priority, and the inner gene sequence encodes the equipment allocation and logistics path planning respectively. The structure is clear and modular, which helps to effectively express complex scheduling problems and improve the solution space search ability of the algorithm; the generation of the initial population can effectively cover the solution space through a double-layer encoding structure, and the calculation of the fitness value can provide an effective evaluation basis for subsequent optimization, ensuring that excellent individuals are selected for evolution; the population is stratified by non-dominated sorting, which can identify and retain diverse high-quality individuals, avoid the population from converging to the local optimal solution in the early stage, and improve the diversity of solutions and global search capabilities; the calculation of the crowding distance helps to measure the distribution density between individuals, ensuring that when selecting the parent population, those individuals located in positions with high diversity and a wider search space are given priority, thereby promoting the diversity of the population and the quality of the solution; high-quality individuals are selected through stratification results and crowding distance, and individuals with high fitness and diversity are effectively screened out, ensuring that the parent population can guide the algorithm to converge to the global optimal solution or a near-optimal solution.

[0082] S103. Perform an adaptive crossover operation on the individuals of the parent population, dynamically adjust the crossover position based on the fitness value of the individuals, and generate offspring individuals; perform taboo search optimization on the offspring individuals to generate a local search solution; adaptively adjust the mutation probability according to the distribution density of the individuals in the solution space, perform a mutation operation on the local search solution, and generate mutant individuals; merge the mutant individuals with the parent individuals, and iteratively select a new generation of population; repeat the execution until the population converges, and output a scheduling plan that achieves optimal balance on multi-dimensional scheduling indicators.

[0083] In this embodiment, the crossover position is adjusted dynamically, and high-quality offspring individuals can be flexibly generated according to the changes in individual fitness, thereby improving the diversity and convergence of the search; local searches are performed on offspring individuals to avoid falling into local optimal solutions, further improving the quality of the solution, and expanding the exploration depth of the solution space; the mutation probability is adjusted according to the distribution density in the solution space, effectively balancing exploration and development, avoiding premature convergence, and enhancing the global search capability of the algorithm; by merging mutant individuals and parent individuals, the population structure is optimized so that the new generation of population contains better solutions, thereby enhancing the adaptability and stability of the algorithm; in multi-objective scheduling problems, by comprehensively considering multi-dimensional indicators such as maximum completion time, equipment utilization, and logistics path load, the optimal balanced scheduling solution can be found to improve production efficiency and resource utilization; by repeatedly iterating until the population converges, it is ensured that a global optimal or approximately optimal solution can be found, thereby avoiding the problem of local optimal solutions.

[0084] In an optional implementation, real-time operation data of a chip production line is collected and processed through a dual-channel deep neural network, including a first channel and a second channel, wherein the first channel passes through a temporal convolution layer to extract task execution features, and the second channel passes through a spatial convolution layer to extract device layout features, including:

[0085] Collecting real-time operation data of the chip production line, the real-time operation data includes equipment operation status data, material transmission speed data and buffer area occupancy data;

[0086] According to a preset fixed length, a data detection window is constructed, a distribution skewness value of the real-time operation data is calculated within the data detection window, abnormal data points are identified according to a preset skewness threshold, and corrected data is obtained after the abnormal data points are removed;

[0087] Perform fast Fourier transform to process the correction data, extract the fluctuation frequency value of the logistics system, calculate and set the hole convolution rate of the first branch, the second branch and the third branch in the timing feature extraction network according to the fluctuation frequency value; input the correction data to the timing feature extraction network, extract local timing features through the dynamic deformable convolution layer in the first branch, the second branch and the third branch, and the dynamic deformable convolution layer adaptively adjusts the convolution sampling position and weight parameters according to the distribution characteristics of the local timing features; calculate the conditional entropy value and mutual information value of the local timing features, determine the weight coefficients of the features at different time scales based on the conditional entropy value and the mutual information value, and obtain the task execution features by weighted fusion;

[0088] Construct a heterogeneous graph network model of the production line, model the processing equipment, transmission equipment and cache area as processing nodes, transportation nodes and storage nodes respectively, and obtain the correlation strength between nodes through multi-head attention calculation based on the actual distance and historical logistics data between the processing nodes, the transportation nodes and the storage nodes, and generate an initial adjacency matrix;

[0089] Based on the heterogeneous graph network model of the production line, a community discovery algorithm is executed to identify the load intensity and transmission capacity saturation areas of the logistics channel from the initial adjacency matrix. Based on the corresponding spatial distribution law, the connection relationship between the processing nodes, transportation nodes and storage nodes is adjusted, and the initial adjacency matrix is ​​updated to obtain an optimized adjacency matrix. The optimized adjacency matrix is ​​input into a graph convolutional network based on Chebyshev polynomials to extract and generate equipment layout features.

[0090] In an optional specific implementation, first, real-time operation data is collected from various sensors and databases in the chip production line. These data include equipment operation status data, such as the switch status, operation speed, temperature, pressure, etc. of the equipment; material transmission speed data, such as conveyor belt speed, robot arm movement speed, etc.; and buffer area occupancy data, such as the quantity and type of materials in the buffer area. Example data include: the operation speed of photolithography machine No. 1 is 100 wafers per hour, the speed of conveyor belt No. 2 is 1 meter per second, and buffer No. 3 currently stores 50 wafers.

[0091] Next, the collected real-time operation data is preprocessed to remove abnormal data points. The data is windowed according to a preset fixed length, such as 1 minute. In each data detection window, the distribution skewness value of the real-time operation data is calculated. If the skewness value exceeds a preset threshold, such as 2, the data point is considered abnormal. After removing the abnormal data points, the corrected data is obtained. For example, in a data detection window, the operating speed data of photolithography machine No. 1 has a peak value, and the skewness value is 2.5, which exceeds the preset threshold 2, and the peak data point is removed.

[0092] Then, the corrected data is analyzed in the frequency domain. Fast Fourier transform is performed on the corrected data to extract the fluctuation frequency values ​​of the logistics system. For example, the analysis found that the fluctuation frequency of the logistics system is mainly concentrated at 0.1Hz and 0.5Hz. According to these fluctuation frequency values, the hole convolution rates of the three branches in the time series feature extraction network are set. For example, the hole convolution rate of the first branch is set to 1, the second branch is set to 2, and the third branch is set to 4 to capture the characteristics of different time scales.

[0093] The corrected data is input into the temporal feature extraction network. The network contains three branches, each of which uses a dynamic deformable convolution layer to extract local temporal features. The dynamic deformable convolution layer can adaptively adjust the convolution sampling position and weight parameters according to the distribution characteristics of the local temporal features. For example, if the material transmission speed fluctuates greatly within a certain period of time, the dynamic deformable convolution layer will automatically adjust the sampling position and weight to better capture these fluctuation characteristics. The conditional entropy value and mutual information value of the extracted local temporal features are calculated. Based on these values, the weight coefficients of the features at different time scales are determined, and the features extracted by different branches are weighted and fused to obtain the task execution features. For example, if the feature conditional entropy value corresponding to the fluctuation frequency of 0.1Hz is higher, the feature is given a higher weight.

[0094] At the same time, a heterogeneous graph network model of the production line is constructed. Processing equipment (such as lithography machines, etching machines), transmission equipment (such as conveyor belts, robotic arms) and cache areas are modeled as processing nodes, transportation nodes and storage nodes respectively. According to the actual distance between nodes and historical logistics data, such as average transmission time, number of transmissions, etc., the association strength between nodes is calculated through multi-head attention to generate the initial adjacency matrix. For example, if the average transmission time between lithography machine No. 1 and conveyor belt No. 2 is short and the number of transmissions is large, the association strength between them is high.

[0095] Next, based on the heterogeneous graph network model of the production line, a community discovery algorithm is executed. The algorithm can identify the load intensity of the logistics channel and the saturated area of ​​the transmission capacity from the initial adjacency matrix. For example, it is found that the load intensity of the logistics channel connecting the No. 1 lithography machine, the No. 2 conveyor belt, and the No. 3 buffer area is high. According to the results of the identification, the connection relationship between the processing nodes, the transportation nodes, and the storage nodes is adjusted, and the initial adjacency matrix is ​​updated to obtain the optimized adjacency matrix. For example, a path connecting the No. 1 lithography machine and the No. 4 conveyor belt can be added to share the load of the No. 2 conveyor belt. The optimized adjacency matrix is ​​input into the graph convolution network based on Chebyshev polynomials to extract and generate equipment layout features.

[0096] In this embodiment, by real-time analysis of production line data, identification of bottlenecks and abnormal situations, and dynamic adjustments, production scheduling can be optimized and production efficiency can be improved; by optimizing logistics routes and resource allocation, material waste and equipment idle time can be reduced, thereby reducing production costs; through real-time monitoring and analysis of equipment operating status, potential quality problems can be discovered in a timely manner, and measures can be taken to prevent and correct them, thereby improving product quality.

[0097] In an optional implementation, the task execution features and the equipment layout features are input into a recurrent neural network to perform feature fusion, determine a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length, and generate a production state vector including a predicted value of the multidimensional scheduling index, including:

[0098] Inputting the task execution features and the equipment layout features into a double-layer recurrent neural network, wherein the double-layer recurrent neural network includes a feature decomposition layer and a feature combination layer;

[0099] A plurality of recursive processing units of time scales are arranged in the feature decomposition layer, the task execution features and the equipment layout features are divided into a plurality of groups of time series segments according to preset time intervals, and the segments are respectively input into the recursive processing units of the corresponding time scales to generate a plurality of groups of time series features; the recursive processing units of different time scales are connected by a feature cascade method to realize information transmission between the plurality of groups of time series features and obtain the fused time series features;

[0100] Calculate the information entropy value and mutual information value of the fused time series features to generate feature channel weights; calculate the position coding difference and time distance of adjacent time series segments in the fused time series features to generate time series position weights;

[0101] The feature channel weight and the time sequence position weight are input into an adaptive gating unit, and the adaptive gating unit dynamically adjusts the information flow ratio according to the weight to generate a fusion weight;

[0102] The fusion weight is multiplied by the fused time series feature to obtain a weighted feature vector; the weighted feature vector is input into a fully connected layer to generate a production state vector including a task completion time prediction value, an equipment utilization efficiency prediction value, and a logistics path length prediction value.

[0103] The information entropy specifically refers to a measure of uncertainty or information volume, reflecting the distribution of possible states in a feature or system. Specifically, the higher the information entropy value, the more dispersed the state distribution of the feature or system and the greater the uncertainty; the lower the information entropy value, the more concentrated the state distribution and the smaller the uncertainty. In task feature analysis, information entropy can be used to evaluate the importance and distribution of features, thereby helping to identify and extract key information.

[0104] The mutual information specifically refers to the degree of association between two features or variables, reflecting the amount of information contained in one feature about another feature. If the mutual information value is high, it means that the correlation between the two features is strong, and one feature can predict the other feature to a large extent; if the mutual information value is low, it means that the dependence between the two is weak and may be relatively independent. In multi-feature analysis, the mutual information value can be used to judge the synergy between features, thereby optimizing the feature combination strategy.

[0105] In order to improve production scheduling efficiency and resource utilization, a production status prediction method based on a double-layer recurrent neural network is provided. The method integrates task execution characteristics and equipment layout characteristics, predicts multidimensional scheduling indicators such as task completion time, equipment utilization efficiency and logistics path length, and generates a production status vector to provide a basis for subsequent scheduling decisions.

[0106] First, collect and preprocess the task execution characteristics and equipment layout characteristics data in the production process. Task execution characteristics can include information such as task type, task priority, and task processing time. Equipment layout characteristics can include information such as equipment location, equipment type, and equipment status. Perform preprocessing operations such as cleaning and normalization on the collected data to ensure data quality and consistency. For example, quantify task priority as a value from 1 to 5, and the equipment status is represented as "idle" or "busy".

[0107] Then, the preprocessed task execution features and equipment layout features are input into a two-layer recurrent neural network. The network includes a feature decomposition layer and a feature combination layer. In the feature decomposition layer, recursive processing units of multiple time scales are set, such as hours, days and weeks. The task execution features and equipment layout features are divided into multiple groups of time series segments according to preset time intervals (such as 1 hour), and are respectively input into the recursive processing units of the corresponding time scales. For example, a week of data is divided into 168 hours of time series segments, which are respectively input into the recursive processing units of the hour scale. Each recursive processing unit processes the input time series segments to generate multiple groups of time series features. The recursive processing units of different time scales are connected by feature cascading to realize information transfer between multiple groups of time series features and obtain the fused time series features. For example, the time series features generated at the hour, day and week scales are connected together.

[0108] Next, the information entropy value and mutual information value of the fused time series features are calculated to generate feature channel weights. The information entropy value reflects the uncertainty of the feature, and the mutual information value reflects the correlation between the features. The higher the weight value, the more important the feature. For example, if the information entropy value of a feature is high and the mutual information value is also high, the feature is given a higher weight. At the same time, the position coding difference and time distance of adjacent time series segments in the fused time series features are calculated to generate the time series position weight. The position coding difference reflects the feature difference of adjacent time series segments, and the time distance reflects the time interval of adjacent time series segments. The higher the weight value, the stronger the correlation between adjacent time series segments. For example, if the position coding difference of adjacent time series segments is small and the time distance is short, the time series position is given a higher weight.

[0109] Subsequently, the feature channel weights and the time series position weights are input into the adaptive gating unit. The adaptive gating unit dynamically adjusts the information flow ratio according to the weight size to generate the fusion weight. For example, if the weight of a feature channel is high, a larger information flow ratio is given to the feature channel. The fusion weight is multiplied by the fused time series feature to obtain a weighted feature vector. Finally, the weighted feature vector is input into the fully connected layer to generate a production state vector containing the task completion time prediction value, the equipment utilization efficiency prediction value, and the logistics path length prediction value. For example, a vector containing three values ​​is output, indicating that the task completion time prediction value is 24 hours, the equipment utilization efficiency prediction value is 80%, and the logistics path length prediction value is 100 meters.

[0110] In this embodiment, by fusing the timing features of multiple time scales and considering the feature channel weights and time position weights, the dynamic changes in the production process can be captured more comprehensively, thereby improving the prediction accuracy; the adaptive gating unit can dynamically adjust the information flow ratio according to the weight size, thereby enhancing the robustness of the model to noise and abnormal data; by predicting multi-dimensional scheduling indicators, a more accurate decision-making basis can be provided for production scheduling, thereby improving decision-making efficiency.

[0111] In an optional implementation, based on the multidimensional optimization target set, a double-layer gene encoding structure is constructed, the task sorting priority is encoded as an outer gene sequence, and the equipment resource allocation and the logistics path planning are encoded as an inner gene sequence; the double-layer gene encoding structure is used to generate an initial population, and the fitness value of each individual in the initial population on the multidimensional scheduling index is calculated; based on the fitness value, a non-dominated sort is performed on the initial population to obtain a population stratification result, and the crowding distance between individuals is calculated; selecting high-quality individuals to form a parent population according to the population stratification result and the crowding distance includes:

[0112] Construct a double-layer gene coding structure, encode the execution priority of the task to be processed as a real number between zero and one to form an outer gene sequence, encode the equipment allocation plan and the logistics path plan of the task to be processed as integers to form an inner gene sequence, the equipment allocation plan in the inner gene sequence corresponds to the equipment number, and the logistics path plan corresponds to the path number;

[0113] Through the double-layer gene coding structure, a preset number of individuals are randomly generated to form an initial population, and each individual in the initial population includes a group of the outer gene sequences and a group of the inner gene sequences;

[0114] Prioritize the tasks to be processed according to the numerical values ​​of the outer gene sequence, read the corresponding equipment numbers and path numbers in the inner gene sequence in sequence according to the priority ranking, generate a scheduling plan, count the maximum completion time of the scheduling plan, the inverse of the minimum equipment utilization rate and the total load of the logistics path, and calculate the multidimensional fitness value;

[0115] Performing non-dominated sorting on the initial population, dividing the initial population into multiple population levels based on dominance relations; calculating the crowding distance of each individual in the initial population, sorting the multidimensional fitness values ​​of individuals in the same population level according to each dimension, calculating the difference in multidimensional fitness values ​​of adjacent individuals, dividing the difference in multidimensional fitness values ​​by the extreme difference in fitness values ​​on the corresponding dimension to obtain a standardized distance, summing the standardized distances of each dimension to obtain the crowding distance of the individual;

[0116] Two individuals are randomly selected for tournament selection, and the individual with a smaller number corresponding to the population level is selected. When the population levels of the two individuals are the same, the individual with a larger crowding distance is selected, and the tournament selection is repeated until a preset number of individuals are selected to form the parent population.

[0117] In genetic algorithms, gene sequences are a way of encoding solutions to problems. By simulating biological genetic mechanisms, the decision variables of the problem are represented as a sequence structure. In double-layer gene encoding, the outer and inner gene sequences represent different levels of information. The outer gene sequence usually reflects global decision information such as the execution priority of the task, while the inner gene sequence encodes more specific solutions, such as equipment allocation or path selection. The design of the gene sequence directly affects the algorithm's search efficiency and optimization capabilities for the solution space.

[0118] The range specifically refers to a simple indicator used in statistics to measure the range of data distribution, which represents the difference between the maximum and minimum values ​​in a data set. It reflects the degree of data dispersion. The larger the value, the wider the data distribution range; the smaller the value, the narrower the data distribution range. In the calculation of multidimensional fitness values, the range is used to standardize the data so that comparison and optimization can be performed between dimensions of different dimensions.

[0119] First, a double-layer gene encoding structure is constructed. The outer gene sequence consists of real numbers between 0 and 1, which represent the execution priority of the task to be processed. For example, there are three tasks, whose priorities are encoded as 0.8, 0.5 and 0.2, respectively, indicating that task 1 has the highest priority and task 3 has the lowest priority. The inner gene sequence consists of integers, which represent the equipment allocation plan and the logistics path plan respectively. Assuming there are 5 equipment and 10 paths, the equipment allocation plan is encoded with integers from 1 to 5, and the logistics path plan is encoded with integers from 1 to 10. For example, an inner gene sequence can be represented as [2,5;3,8], which means that task 1 is assigned to equipment 2 and path 5 is selected; task 2 is assigned to equipment 3 and path 8 is selected.

[0120] Next, generate the initial population. The initial population is formed by randomly generating a preset number of individuals. Each individual contains a set of outer gene sequences and a set of inner gene sequences. For example, if the population size is set to 10, 10 individuals need to be generated, and each individual contains the priority, equipment allocation, and path planning information of the corresponding number of tasks. Assume that an individual with three tasks has an outer gene sequence of [0.8, 0.5, 0.2] and an inner gene sequence of [2, 5; 3, 8; 1, 2].

[0121] Then, calculate the fitness value. Prioritize the tasks according to the numerical value of the outer gene sequence. Read the corresponding equipment number and path number in the inner gene sequence in order according to the priority order to generate a scheduling plan. For example, according to the gene sequence of the above individuals, the order of task execution is task 1, task 2, and task 3, and the corresponding equipment allocation and path planning are device 2 path 5, device 3 path 8, and device 1 path 2, respectively. Then, count the maximum completion time, the inverse of the equipment utilization rate, and the total load of the logistics path of the scheduling plan. Assuming that the maximum completion time is 10 hours, the equipment utilization rate is 80%, and the total load of the logistics path is 50 tons, these three indicators can be combined into a multidimensional fitness value vector, such as [10,-0.8,50].

[0122] Next, perform non-dominated sorting and crowding distance calculation. Perform non-dominated sorting on the initial population and divide the initial population into multiple population levels based on the dominance relationship. The definition of dominance relationship is: if an individual is better than another individual in all objectives, the former is said to dominate the latter. For example, if the fitness value of one individual is [8,-0.9,40] and the fitness value of another individual is [10,-0.8,50], the former dominates the latter. Sort all individuals in the population according to the dominance relationship and divide them into different levels. Then, calculate the crowding distance of each individual. Sort the multidimensional fitness values ​​of individuals in the same population level according to each dimension, and calculate the difference in multidimensional fitness values ​​of adjacent individuals. Divide the difference in multidimensional fitness values ​​by the extreme difference of fitness values ​​on the corresponding dimension to obtain the standardized distance. Sum the standardized distances of each dimension to obtain the crowding distance of the individual. The crowding distance is used to measure the density of solutions around an individual.

[0123] Finally, the parent population is selected. Two individuals are randomly selected for tournament selection, and the individual with the smaller number corresponding to the population level is selected. When the population levels of two individuals are the same, the individual with the larger crowding distance is selected. The tournament selection is repeated until a preset number of individuals are selected to form the parent population. For example, if the parent population size is set to 5, the tournament selection needs to be repeated 5 times to select 5 individuals to form the parent population.

[0124] In this embodiment, through priority coding and multi-dimensional optimization objectives, a better scheduling plan can be quickly found, the scheduling time can be shortened, and the scheduling efficiency can be improved; multiple optimization objectives can be comprehensively considered, such as the maximum completion time, equipment utilization, and the total load of the logistics path, so as to better balance resource utilization and improve resource utilization; the double-layer gene coding structure and non-dominated sorting strategy enable it to adapt to different types of scheduling problems, and it has strong robustness and adaptability.

[0125] In an optional implementation, performing non-dominated sorting on the initial population, and dividing the initial population into a plurality of population levels based on dominance relationships includes:

[0126] Constructing a fitness difference record matrix and a dominance relationship mark matrix, wherein the number of rows and columns of the fitness difference record matrix and the dominance relationship mark matrix is ​​equal to the number of individuals in the initial population;

[0127] For any two individuals in the initial population, compare the fitness values ​​in three dimensions: maximum completion time, the inverse of minimum equipment utilization, and total load of the logistics path, calculate the fitness difference between the two individuals in the three dimensions, fill the calculated fitness difference into the corresponding position of the fitness difference record matrix, calculate in sequence, and complete the filling action of the fitness difference record matrix;

[0128] Read the difference information stored in the fitness difference record matrix to determine whether the first individual dominates the second individual. When the fitness values ​​of the first individual in the three dimensions of maximum completion time, the inverse of the minimum equipment utilization rate, and the total load of the logistics path are all less than or equal to the second individual, and at least one dimension is less than the second individual, mark the first individual dominating the second individual at the corresponding position of the domination relationship marking matrix, and make judgments in turn to complete the marking action of the domination relationship marking matrix;

[0129] Scan each row and each column of the dominance relationship labeling matrix, count the number of times each individual is dominated by other individuals, record the dominance count of the corresponding individual, and record other individuals dominated by each individual to form a set of dominated individuals;

[0130] Select individuals with a dominance count of zero and place them in the first frontier layer. Traverse the dominated individual set of each individual in the first frontier layer, reduce the dominance count of the individual in the dominated individual set by one, and obtain the updated dominance count. When the updated dominance count becomes zero, place the corresponding individual in the second frontier layer.

[0131] Repeat the process until all individuals in the initial population are divided into the corresponding frontier layers, and the population level division result is obtained.

[0132] In an optional specific implementation, the population is first initialized. Assume that the population size is 5 individuals, each of which represents a possible production scheduling scheme. Each individual has three fitness values, corresponding to the maximum completion time, the opposite of the minimum equipment utilization rate (in order to unify the optimization direction, the minimization problem is converted into a maximization problem) and the total load of the logistics path.

[0133] For example: Individual 1: (10, 0.8, 150); Individual 2: (12, 0.7, 130); Individual 3: (8, 0.9, 160); Individual 4: (11, 0.6, 140); Individual 5: (9, 0.85, 170);

[0134] Next, a 5×5 fitness difference record matrix and a dominance relationship marking matrix of the same size are constructed, where the rows and columns of the matrix represent the individuals in the population.

[0135] Then, compare the fitness values ​​of any two individuals in the population on the three targets one by one. For example, compare individual 1 and individual 2, and calculate the difference between them on the three targets: the difference between the maximum completion time of individual 1 and the maximum completion time of individual 2 (10-12=-2), the difference between the inverse of the minimum equipment utilization of individual 1 and the inverse of the minimum equipment utilization of individual 2 (0.8-0.7=0.1), and the difference between the total load of the logistics path of individual 1 and the total load of the logistics path of individual 2 (150-130=20). Fill these differences into the corresponding positions of the fitness difference record matrix. Repeat this process until the entire fitness difference record matrix is ​​filled.

[0136] According to the information in the fitness difference record matrix, the dominance relationship between individuals is determined. For example, if the fitness value of an individual A on all three targets is less than or equal to the fitness value of another individual B, and is less than B on at least one target, then A is said to dominate B. This dominance relationship is recorded in the dominance relationship marking matrix. For example, if individual 1 dominates individual 2, then the first row and second column of the dominance relationship marking matrix is ​​marked as "dominance".

[0137] Scan the dominance relationship labeling matrix, count the number of times each individual is dominated, i.e., the dominance count, and record the other individuals dominated by each individual to form a dominated individual set. For example, if individual 1 is dominated by individual 3, individuals 2 and 4 do not dominate individual 1, and individual 5 dominates individual 1, then the dominance count of individual 1 is 2, and the dominated individual set is empty.

[0138] Individuals with a dominance count of zero are classified into the first frontier layer. For example, assuming that the dominance counts of individuals 1 and 3 are zero, they are classified into the first frontier layer.

[0139] Traverse the set of dominated individuals of each individual in the first frontier layer. For each individual in the set, reduce its dominance count by 1. If the dominance count of an individual becomes zero after being reduced by 1, it is placed in the second frontier layer.

[0140] Repeat the above steps until all individuals are divided into corresponding levels. Finally, the hierarchical division result of the population is obtained, for example: the first frontier layer: individual 1, individual 3; the second frontier layer: individual 2, individual 5; the third frontier layer: individual 4.

[0141] In this embodiment, by stratifying the population, excellent individuals can be screened out more effectively and the convergence speed can be accelerated; the non-dominated sorting retains individuals in different frontier layers, avoiding premature convergence to the local optimal solution, thereby maintaining the diversity of the population; the stratified population structure is clear, which is convenient for subsequent selection operations, for example, individuals with higher frontier levels can be given priority.

[0142] In an optional implementation, an adaptive crossover operation is performed on individuals of the parent population, and the crossover position is dynamically adjusted based on the fitness value of the individual to generate a child individual; taboo search optimization is performed on the child individual to generate a local search solution; the mutation probability is adaptively adjusted according to the distribution density of the individual in the solution space, and a mutation operation is performed on the local search solution to generate a mutant individual; the mutant individual is merged with the parent individual, and a new generation population is iteratively selected; the execution is repeated until the population converges, and the scheduling scheme that achieves the optimal balance on the multi-dimensional scheduling index is outputted, including:

[0143] Calculate the fitness value of each parent individual in the parent population on the multidimensional scheduling index, and divide the fitness value by the total fitness value of the parent individual on the corresponding dimension to obtain a normalized fitness value;

[0144] Calculate the product of the arithmetic mean of the normalized fitness value and the adaptive adjustment factor to determine the crossover position selection probability, determine the crossover position of the parent individuals according to the crossover position selection probability, and exchange the gene fragments of the two parent individuals between the crossover positions to generate offspring individuals;

[0145] Construct a taboo table, and perform neighborhood search for task order exchange, equipment allocation change, and logistics path adjustment under the constraints of the taboo table. Use the weighted fitness value as a criterion to select the optimal neighborhood action and update the taboo table. After multiple iterations, a local search solution is obtained.

[0146] Perform task sequence mutation and device path mutation on the local search solution according to the mutation probability to generate mutant individuals; merge the mutant individuals with the parent individuals to generate a temporary population;

[0147] Calculating the change of each temporary individual in the temporary population in the multidimensional scheduling index relative to the parent population, and the distribution uniformity of the individuals in the three-dimensional target space;

[0148] Based on the change amount and the distribution uniformity, when the difference in the change amounts of adjacent temporary individuals is less than a preset change amount, secondary sorting is performed in descending order of the distribution uniformity, and a corresponding number of temporary individuals are selected from the order after the secondary sorting according to a preset selection ratio to form a new generation population;

[0149] Repeat the process and output a scheduling solution that achieves the optimal balance on multi-dimensional scheduling indicators.

[0150] First, calculate the fitness value of each individual in the parent population on each scheduling indicator. Assume that there are three scheduling indicators: time, cost, and quality. For each parent individual, calculate its fitness value on the three indicators of time, cost, and quality. For example, the fitness values ​​of parent individual 1 on time, cost, and quality are 10, 8, and 9, respectively. Then, divide the fitness value of each indicator by the total fitness value of the indicator in all parent individuals to obtain the normalized fitness value. Assuming that the total fitness values ​​of time, cost, and quality are 100, 80, and 90, respectively, the normalized fitness values ​​of parent individual 1 on time, cost, and quality are 0.1, 0.1, and 0.1, respectively.

[0151] Next, the product of the arithmetic mean of the normalized fitness value of each indicator and the adaptive adjustment factor is calculated to obtain the probability of crossover position selection. Assuming that the adaptive adjustment factor is 0.5, the probability of crossover position selection of parent individual 1 is (0.1+0.1+0.1) / 3*0.5=0.05. According to the crossover position selection probability, the crossover position of the parent individual is determined. For example, for two parent individuals, if their crossover position selection probabilities are 0.05 and 0.08 respectively, the two parent individuals can be crossovered between the 5th and 8th positions of the gene fragment to generate offspring individuals. Assuming that the gene fragment of parent individual 1 is [A, B, C, D, E, F, G] and the gene fragment of parent individual 2 is [H, I, J, K, L, M, N], the offspring individuals generated after crossover may be [A, B, C, D, L, M, N] and [H, I, J, K, E, F, G].

[0152] Then, taboo search optimization is performed on the generated offspring individuals. A taboo table is constructed, and neighborhood search operations such as task order exchange, equipment allocation change, and logistics path adjustment are performed under the constraints of the taboo table. Using the weighted fitness value as the evaluation criterion, the optimal neighborhood action is selected and the taboo table is updated. After multiple iterations, a local search solution is obtained. For example, assuming that the current solution is [A, B, C], the neighborhood solutions are [B, A, C], [A, C, B], and [C, B, A], and their weighted fitness values ​​are 12, 15, and 10, respectively, and [B, A, C] is not in the taboo table, then [B, A, C] is selected as the new current solution and added to the taboo table.

[0153] Next, the mutation probability is adaptively adjusted according to the distribution density of individuals in the solution space. For example, if individuals are densely distributed in the solution space, the mutation probability is reduced; otherwise, the mutation probability is increased. Assuming the mutation probability is 0.1, the local search solution is subjected to tasks order mutation and equipment path mutation to generate mutant individuals. For example, for the local search solution [B, A, C], after mutation, it may generate [A, B, C] or [B, C, A].

[0154] Merge the mutant individuals with the parent individuals to generate a temporary population. Calculate the change of each individual in the temporary population relative to the parent population in terms of multidimensional scheduling indicators, as well as the distribution uniformity of the individuals in the three-dimensional target space. Based on the change and distribution uniformity, when the difference in the change of adjacent temporary individuals is less than the preset change, perform secondary sorting in descending order of distribution uniformity, and select a certain number of temporary individuals from the secondary sorting order according to the preset selection ratio to form a new generation of population.

[0155] Repeat the above steps until the population converges, and finally output a scheduling solution that achieves the optimal balance on multi-dimensional scheduling indicators.

[0156] In this embodiment, through adaptive crossover, taboo search and density variation-based evolutionary algorithms, a better scheduling scheme can be quickly searched, thereby improving scheduling efficiency; multiple scheduling indicators can be considered at the same time, and the best balance can be achieved between these indicators, avoiding the negative impact of single indicator optimization; by considering the distribution of individuals in the solution space, more diverse scheduling schemes can be generated, thereby enhancing the robustness of the scheme and better coping with uncertainties in the actual environment.

[0157] In an optional implementation, a taboo table is constructed, and a neighborhood search of task order exchange, equipment allocation change, and logistics path adjustment is performed under the constraints of the taboo table. The weighted fitness value is used as a criterion to select the optimal neighborhood action and update the taboo table. After multiple iterations, the local search solution includes:

[0158] The maximum capacity of the taboo table is set to the weighted sum of the number of tasks, the number of devices, and the number of feasible paths;

[0159] The first type of neighborhood solution set is generated by exchanging the processing order of adjacent tasks, the second type of neighborhood solution set is generated by changing the processing equipment of the task, and the third type of neighborhood solution set is generated by adjusting the logistics path of the task. The first type of neighborhood solution set, the second type of neighborhood solution set and the third type of neighborhood solution set are combined to generate the candidate neighborhood solution of the offspring individual:

[0160] Calculate the fitness value of each candidate neighborhood solution on the multi-dimensional scheduling index, determine the evaluation index based on the fitness value through weighted summation, and record the current optimal evaluation index;

[0161] Traverse each neighborhood solution. If the action corresponding to the current neighborhood solution is not in the taboo table and the corresponding evaluation index is greater than the current optimal evaluation index, the current neighborhood solution is taken as a candidate optimal solution. If the action corresponding to the current neighborhood solution is in the taboo table and the fitness value of any dimension in the corresponding multi-dimensional scheduling index is greater than the historical optimal value, the current neighborhood solution is taken as a candidate optimal solution.

[0162] Execute the optimal action corresponding to the candidate optimal solution, update the current neighborhood solution, determine the taboo action based on the reverse action of the optimal action, and add the taboo action to the taboo table; when the taboo table exceeds the preset capacity, delete the earliest added taboo action;

[0163] Repeat until the preset number of iterations is reached, and the final solution obtained is used as the local search solution.

[0164] First, determine the size of the taboo table. The taboo table is used to record the operations that have been executed recently to prevent the search process from falling into the local optimum. The size of the taboo table is set to an appropriate value, such as twice the sum of the number of tasks, the number of devices, and the number of feasible paths. This value can be adjusted according to the actual problem. For example, if there are 10 tasks, 5 devices, and 20 feasible paths, the taboo table size can be set to 70.

[0165] Next, generate an initial scheduling plan. This plan can be generated randomly or based on experience or heuristic rules. The initial plan includes the processing order, processing equipment, and logistics path of each task. For example, task 1 is processed on equipment A and then transported to equipment B via path 1 for the next process; task 2 is processed on equipment C and then transported to equipment D via path 5.

[0166] Then, the iterative optimization process begins. In each iteration, a set of candidate neighborhood solutions is generated. Neighborhood solutions are generated by making minor changes to the current solution. Here, three methods are used to generate neighborhood solutions: swapping the processing order of adjacent tasks, such as swapping the order of task 1 and task 2; changing the processing equipment of the task, such as changing the processing equipment of task 1 from A to B; adjusting the logistics path of the task, such as changing the transportation path of task 1 from path 1 to path 3. The neighborhood solutions generated by these three methods are combined to form a candidate neighborhood solution set. Assuming that 2 neighborhood solutions are generated by swapping the order, 3 neighborhood solutions are generated by changing the equipment, and 4 neighborhood solutions are generated by adjusting the path, the candidate neighborhood solution set has a total of 9 solutions.

[0167] After that, evaluate the fitness value of each candidate neighborhood solution. The fitness value is calculated based on multi-dimensional scheduling indicators, such as total processing time, total transportation time, equipment utilization, etc. Each indicator is assigned a weight, and the weighted sum of each indicator is used as the final evaluation indicator. For example, the weight of the total processing time is 0.5, the weight of the total transportation time is 0.3, and the weight of the equipment utilization is 0.2. Then the evaluation indicator of a candidate solution is: 0.5×total processing time+0.3×total transportation time+0.2×equipment utilization. Record the optimal evaluation indicator of the current iteration.

[0168] Next, select the optimal neighborhood action. Traverse each candidate neighborhood solution. If the action corresponding to the solution is not in the taboo table and its evaluation index is better than the current optimal evaluation index, then the solution is regarded as a candidate optimal solution. If the action corresponding to the solution is in the taboo table, but its fitness value on a certain scheduling index is better than the historical optimal value, it is also regarded as a candidate optimal solution. This allows some excellent solutions to break through the limitations of the taboo table.

[0169] Then, the optimal action is executed and the taboo table is updated. The action corresponding to the candidate optimal solution is applied to the current solution to generate a new current solution. The reverse action of the action is added to the taboo table. For example, if the optimal action is to swap the order of task 1 and task 2, then "swap the order of task 2 and task 1" is added to the taboo table. If the taboo table is full, the earliest added taboo action is deleted.

[0170] Finally, repeat the above iterative process until the preset number of iterations is reached. The final solution is output as the local optimal solution. For example, if the number of iterations is set to 100, the above steps are repeated 100 times, and finally a scheduling plan is obtained, which includes the processing order, processing equipment and logistics path of each task.

[0171] In this embodiment, the taboo table mechanism is used to avoid the search process from falling into the local optimum, thereby improving the search efficiency and finding a better solution more quickly; the neighborhood search strategy can explore more solution spaces and find a better solution. The introduction of multi-dimensional evaluation indicators can better balance the relationship between different goals and obtain a scheduling solution that better meets actual needs; the size of the taboo table, neighborhood structure, evaluation indicators and other parameters can be adjusted according to actual problems, so that the algorithm can adapt to different production environments and scheduling needs.

[0172] The present invention provides another optional embodiment, including:

[0173] First, a complete system data collection mechanism needs to be established. At the rail network level, it is necessary to obtain the occupancy status, passable direction, and connection relationship with other rails of each track segment in real time to build a complete track topology map. For the overhead crane OHT in operation, it is necessary to continuously collect dynamic information such as its current position, task status, and load conditions. At the same time, information such as the source location, target station, and expected completion time of the materials to be transported is collected and updated. These data will serve as the basic input for subsequent optimization.

[0174] Next, a dual-channel neural network is used to extract and fuse the collected data. The temporal channel uses a convolutional layer to process the temporal characteristics of the overhead crane task sequence and material flow to capture the dynamic changes of the system; the spatial channel focuses on the layout characteristics of the rail network to identify potential congestion points and critical paths. The features of the two channels are fused through an attention mechanism to generate a comprehensive feature vector that reflects the current state of the system.

[0175] Based on feature processing, a mathematical model for scheduling optimization is constructed. The four objective functions of shortest delivery time, least deadline violation, shortest delivery distance, and maximum total throughput are standardized and weighted combined. Constraints include track carrying capacity constraints, overhead crane kinematic constraints, intersection mutual exclusion constraints, etc. In this way, the complex logistics optimization problem is transformed into a solvable mathematical model.

[0176] The optimization solution adopts an improved genetic algorithm framework. First, the material handling sequence and overhead crane allocation scheme are encoded to generate the initial population. During the evolution process, taboo search is combined to optimize the local path planning to avoid falling into the local optimum. The crossover operation is mainly performed at the level of handling sequence and overhead crane allocation, while the mutation operation focuses on the fine-tuning of the local path. During the optimization process, an elite solution set is dynamically maintained to save high-quality solutions with different characteristics.

[0177] Finally, there is the scheme execution and dynamic adjustment mechanism. The system converts the optimized scheduling scheme into a specific overhead crane scheduling instruction sequence. The system status is continuously monitored during the execution process. When a large deviation occurs, the local re-planning mechanism is triggered. During re-planning, the adjustment scheme with the least impact on the original scheme is given priority to ensure the stable operation of the system. At the same time, the data during the execution process is fed back to the feature extraction module to form a closed-loop optimization.

[0178] Figure 2 FIG. 1 is a schematic diagram of the structure of a chip production logistics optimization system based on a multi-objective genetic algorithm according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0179] The first unit is used to collect real-time operation data of the chip production line and perform data processing through a dual-channel deep neural network, including a first channel and a second channel, wherein the first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer; the task execution features and the equipment layout features are input into a recurrent neural network for feature fusion, and a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length is determined, and a production state vector including a predicted value of the multidimensional scheduling index is generated; based on the production state vector, a multidimensional optimization target set including task sorting priority, equipment resource allocation, and logistics path planning is constructed;

[0180] The second unit is used to construct a double-layer gene coding structure based on the multi-dimensional optimization target set, encode the task sorting priority as an outer gene sequence, and encode the equipment resource allocation and the logistics path planning as an inner gene sequence; generate an initial population using the double-layer gene coding structure, calculate the fitness value of each individual in the initial population on the multi-dimensional scheduling index; perform non-dominated sorting on the initial population based on the fitness value, obtain a population stratification result, and calculate the crowding distance between individuals; select high-quality individuals to form a parent population according to the population stratification result and the crowding distance;

[0181] The third unit is used to perform an adaptive crossover operation on the individuals of the parent population, dynamically adjust the crossover position based on the fitness value of the individuals, and generate offspring individuals; perform taboo search optimization on the offspring individuals to generate a local search solution; adaptively adjust the mutation probability according to the distribution density of the individuals in the solution space, perform a mutation operation on the local search solution to generate mutant individuals; merge the mutant individuals with the parent individuals, and iteratively select a new generation of population; repeat the execution until the population converges, and output a scheduling plan that achieves optimal balance on multi-dimensional scheduling indicators.

[0182] According to a third aspect of the embodiments of the present invention,

[0183] An electronic device is provided, comprising:

[0184] processor;

[0185] a memory for storing processor-executable instructions;

[0186] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0187] A fourth aspect of the embodiments of the present invention is:

[0188] 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.

[0189] 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.

[0190] 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 chip production logistics optimization method based on a multi-objective genetic algorithm, characterized in that: include: Collect real-time operation data of the chip production line and process the data through a dual-channel deep neural network, including a first channel and a second channel. The first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer. Input the task execution features and the equipment layout features into a recurrent neural network, perform feature fusion, determine a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length, and generate a production state vector including a predicted value of the multidimensional scheduling index; based on the production state vector, construct a multidimensional optimization target set including task sorting priority, equipment resource allocation, and logistics path planning; Based on the multi-dimensional optimization target set, a double-layer gene coding structure is constructed, the task sorting priority is encoded as an outer gene sequence, and the equipment resource allocation and the logistics path planning are encoded as an inner gene sequence; the double-layer gene coding structure is used to generate an initial population, and the fitness value of each individual in the initial population on the multi-dimensional scheduling index is calculated; based on the fitness value, a non-dominated sort is performed on the initial population to obtain a population stratification result, and the crowding distance between individuals is calculated; Selecting high-quality individuals to form a parent population according to the population stratification result and the crowding distance; An adaptive crossover operation is performed on the individuals of the parent population, and the crossover position is dynamically adjusted based on the fitness value of the individuals to generate offspring individuals; taboo search optimization is performed on the offspring individuals to generate a local search solution; the mutation probability is adaptively adjusted according to the distribution density of the individuals in the solution space, and a mutation operation is performed on the local search solution to generate mutant individuals; the mutant individuals are merged with the parent individuals, and a new generation of population is iteratively selected; the execution is repeated until the population converges, and a scheduling plan that achieves optimal balance on multi-dimensional scheduling indicators is output.

2. The method according to claim 1, characterized in that: Collect real-time operation data of the chip production line and process the data through a dual-channel deep neural network, including a first channel and a second channel. The first channel extracts task execution features through a temporal convolution layer, and the second channel extracts equipment layout features through a spatial convolution layer, including: Collecting real-time operation data of the chip production line, the real-time operation data includes equipment operation status data, material transmission speed data and buffer area occupancy data; According to a preset fixed length, a data detection window is constructed, a distribution skewness value of the real-time operation data is calculated within the data detection window, abnormal data points are identified according to a preset skewness threshold, and corrected data is obtained after the abnormal data points are removed; Perform fast Fourier transform to process the correction data, extract the fluctuation frequency value of the logistics system, calculate and set the hole convolution rate of the first branch, the second branch and the third branch in the timing feature extraction network according to the fluctuation frequency value; input the correction data to the timing feature extraction network, extract local timing features through the dynamic deformable convolution layer in the first branch, the second branch and the third branch, and the dynamic deformable convolution layer adaptively adjusts the convolution sampling position and weight parameters according to the distribution characteristics of the local timing features; calculate the conditional entropy value and mutual information value of the local timing features, determine the weight coefficients of the features at different time scales based on the conditional entropy value and the mutual information value, and obtain the task execution features by weighted fusion; Construct a heterogeneous graph network model of the production line, model the processing equipment, transmission equipment and cache area as processing nodes, transportation nodes and storage nodes respectively, and obtain the correlation strength between nodes through multi-head attention calculation based on the actual distance and historical logistics data between the processing nodes, the transportation nodes and the storage nodes, and generate an initial adjacency matrix; Based on the heterogeneous graph network model of the production line, a community discovery algorithm is executed to identify the load intensity and transmission capacity saturation areas of the logistics channel from the initial adjacency matrix. Based on the corresponding spatial distribution law, the connection relationship between the processing nodes, transportation nodes and storage nodes is adjusted, and the initial adjacency matrix is ​​updated to obtain an optimized adjacency matrix. The optimized adjacency matrix is ​​input into a graph convolutional network based on Chebyshev polynomials to extract and generate equipment layout features.

3. The method according to claim 2, characterized in that The task execution features and the equipment layout features are input into a recurrent neural network for feature fusion, and a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length is determined. A production state vector including a predicted value of the multidimensional scheduling index is generated, including: Inputting the task execution features and the equipment layout features into a double-layer recurrent neural network, wherein the double-layer recurrent neural network includes a feature decomposition layer and a feature combination layer; A plurality of recursive processing units of time scales are arranged in the feature decomposition layer, the task execution features and the equipment layout features are divided into a plurality of groups of time series segments according to preset time intervals, and the segments are respectively input into the recursive processing units of the corresponding time scales to generate a plurality of groups of time series features; the recursive processing units of different time scales are connected by a feature cascade method to realize information transmission between the plurality of groups of time series features and obtain the fused time series features; Calculate the information entropy value and mutual information value of the fused time series features to generate feature channel weights; calculate the position coding difference and time distance of adjacent time series segments in the fused time series features to generate time series position weights; The feature channel weight and the time sequence position weight are input into an adaptive gating unit, and the adaptive gating unit dynamically adjusts the information flow ratio according to the weight to generate a fusion weight; The fusion weight is multiplied by the fused time series feature to obtain a weighted feature vector; the weighted feature vector is input into a fully connected layer to generate a production state vector including a task completion time prediction value, an equipment utilization efficiency prediction value, and a logistics path length prediction value.

4. The method according to claim 1, characterized in that: Based on the multi-dimensional optimization target set, a double-layer gene coding structure is constructed, the task sorting priority is encoded as an outer gene sequence, and the equipment resource allocation and the logistics path planning are encoded as an inner gene sequence; the double-layer gene coding structure is used to generate an initial population, and the fitness value of each individual in the initial population on the multi-dimensional scheduling index is calculated; based on the fitness value, a non-dominated sort is performed on the initial population to obtain a population stratification result, and the crowding distance between individuals is calculated; Selecting high-quality individuals to form a parent population according to the population stratification result and the crowding distance includes: Construct a double-layer gene coding structure, encode the execution priority of the task to be processed as a real number between zero and one to form an outer gene sequence, encode the equipment allocation plan and the logistics path plan of the task to be processed as integers to form an inner gene sequence, the equipment allocation plan in the inner gene sequence corresponds to the equipment number, and the logistics path plan corresponds to the path number; Through the double-layer gene coding structure, a preset number of individuals are randomly generated to form an initial population, and each individual in the initial population includes a group of the outer gene sequences and a group of the inner gene sequences; Prioritize the tasks to be processed according to the numerical values ​​of the outer gene sequence, read the corresponding equipment numbers and path numbers in the inner gene sequence in sequence according to the priority ranking, generate a scheduling plan, count the maximum completion time of the scheduling plan, the inverse of the minimum equipment utilization rate and the total load of the logistics path, and calculate the multidimensional fitness value; Performing non-dominated sorting on the initial population, dividing the initial population into multiple population levels based on dominance relations; calculating the crowding distance of each individual in the initial population, sorting the multidimensional fitness values ​​of individuals in the same population level according to each dimension, calculating the difference in multidimensional fitness values ​​of adjacent individuals, dividing the difference in multidimensional fitness values ​​by the extreme difference in fitness values ​​on the corresponding dimension to obtain a standardized distance, summing the standardized distances of each dimension to obtain the crowding distance of the individual; Two individuals are randomly selected for tournament selection, and the individual with a smaller number corresponding to the population level is selected. When the population levels of the two individuals are the same, the individual with a larger crowding distance is selected, and the tournament selection is repeated until a preset number of individuals are selected to form the parent population.

5. The method according to claim 4, characterized in that Performing non-dominated sorting on the initial population, and dividing the initial population into a plurality of population levels based on dominance relationships includes: Constructing a fitness difference record matrix and a dominance relationship mark matrix, wherein the number of rows and columns of the fitness difference record matrix and the dominance relationship mark matrix is ​​equal to the number of individuals in the initial population; For any two individuals in the initial population, compare the fitness values ​​in three dimensions: maximum completion time, the inverse of minimum equipment utilization, and total load of the logistics path, calculate the fitness difference between the two individuals in the three dimensions, fill the calculated fitness difference into the corresponding position of the fitness difference record matrix, calculate in sequence, and complete the filling action of the fitness difference record matrix; Read the difference information stored in the fitness difference record matrix to determine whether the first individual dominates the second individual. When the fitness values ​​of the first individual in the three dimensions of maximum completion time, the inverse of the minimum equipment utilization rate, and the total load of the logistics path are all less than or equal to the second individual, and at least one dimension is less than the second individual, mark the first individual dominating the second individual at the corresponding position of the domination relationship marking matrix, and make judgments in turn to complete the marking action of the domination relationship marking matrix; Scan each row and each column of the dominance relationship labeling matrix, count the number of times each individual is dominated by other individuals, record the dominance count of the corresponding individual, and record other individuals dominated by each individual to form a set of dominated individuals; Select individuals with a dominance count of zero and place them in the first frontier layer. Traverse the dominated individual set of each individual in the first frontier layer, reduce the dominance count of the individual in the dominated individual set by one, and obtain the updated dominance count. When the updated dominance count becomes zero, place the corresponding individual in the second frontier layer. Repeat the process until all individuals in the initial population are divided into the corresponding frontier layers, and the population level division result is obtained.

6. The method according to claim 1, characterized in that Performing an adaptive crossover operation on the individuals of the parent population, dynamically adjusting the crossover position based on the fitness value of the individual, and generating offspring individuals; performing taboo search optimization on the offspring individuals to generate a local search solution; adaptively adjusting the mutation probability according to the distribution density of the individual in the solution space, performing a mutation operation on the local search solution to generate a mutant individual; merging the mutant individual with the parent individual, iteratively selecting a new generation population; repeating the execution until the population converges, and outputting a scheduling solution that achieves optimal balance on multi-dimensional scheduling indicators, including: Calculate the fitness value of each parent individual in the parent population on the multidimensional scheduling index, and divide the fitness value by the total fitness value of the parent individual on the corresponding dimension to obtain a normalized fitness value; Calculate the product of the arithmetic mean of the normalized fitness value and the adaptive adjustment factor to determine the crossover position selection probability, determine the crossover position of the parent individuals according to the crossover position selection probability, and exchange the gene fragments of the two parent individuals between the crossover positions to generate offspring individuals; Construct a taboo table, and perform neighborhood search for task order exchange, equipment allocation change, and logistics path adjustment under the constraints of the taboo table. Use the weighted fitness value as a criterion to select the optimal neighborhood action and update the taboo table. After multiple iterations, a local search solution is obtained. Perform task sequence mutation and device path mutation on the local search solution according to the mutation probability to generate mutant individuals; merge the mutant individuals with the parent individuals to generate a temporary population; Calculating the change of each temporary individual in the temporary population in the multidimensional scheduling index relative to the parent population, and the distribution uniformity of the individuals in the three-dimensional target space; Based on the change amount and the distribution uniformity, when the difference in the change amounts of adjacent temporary individuals is less than a preset change amount, secondary sorting is performed in descending order of the distribution uniformity, and a corresponding number of temporary individuals are selected from the order after the secondary sorting according to a preset selection ratio to form a new generation population; Repeat the process and output a scheduling solution that achieves the optimal balance on multi-dimensional scheduling indicators.

7. The method according to claim 6, characterized in that A taboo table is constructed, and neighborhood search for task order exchange, equipment allocation change, and logistics path adjustment is performed under the constraints of the taboo table. The weighted fitness value is used as a criterion to select the optimal neighborhood action and update the taboo table. After multiple iterations, the local search solutions include: The maximum capacity of the taboo table is set to the weighted sum of the number of tasks, the number of devices, and the number of feasible paths; The first type of neighborhood solution set is generated by exchanging the processing order of adjacent tasks, the second type of neighborhood solution set is generated by changing the processing equipment of the task, and the third type of neighborhood solution set is generated by adjusting the logistics path of the task. The first type of neighborhood solution set, the second type of neighborhood solution set and the third type of neighborhood solution set are combined to generate the candidate neighborhood solution of the offspring individual: Calculate the fitness value of each candidate neighborhood solution on the multi-dimensional scheduling index, determine the evaluation index based on the fitness value through weighted summation, and record the current optimal evaluation index; Traverse each neighborhood solution. If the action corresponding to the current neighborhood solution is not in the taboo table and the corresponding evaluation index is greater than the current optimal evaluation index, the current neighborhood solution is taken as a candidate optimal solution. If the action corresponding to the current neighborhood solution is in the taboo table and the fitness value of any dimension in the corresponding multi-dimensional scheduling index is greater than the historical optimal value, the current neighborhood solution is taken as a candidate optimal solution. Execute the optimal action corresponding to the candidate optimal solution, update the current neighborhood solution, determine the taboo action based on the reverse action of the optimal action, and add the taboo action to the taboo table; when the taboo table exceeds the preset capacity, delete the earliest added taboo action; Repeat until the preset number of iterations is reached, and the final solution obtained is used as the local search solution.

8. A chip production logistics optimization system based on a multi-objective genetic algorithm, 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 real-time operation data of the chip production line and process the data through a dual-channel deep neural network, including a first channel and a second channel, wherein the first channel extracts task execution features through a temporal convolution layer, and the second channel extracts device layout features through a spatial convolution layer; Input the task execution features and the equipment layout features into a recurrent neural network, perform feature fusion, determine a multidimensional scheduling index including task completion time, equipment utilization efficiency, and logistics path length, and generate a production state vector including a predicted value of the multidimensional scheduling index; based on the production state vector, construct a multidimensional optimization target set including task sorting priority, equipment resource allocation, and logistics path planning; The second unit is used to construct a double-layer gene coding structure based on the multi-dimensional optimization target set, encode the task sorting priority as an outer gene sequence, and encode the equipment resource allocation and the logistics path planning as an inner gene sequence; generate an initial population using the double-layer gene coding structure, calculate the fitness value of each individual in the initial population on the multi-dimensional scheduling index; perform non-dominated sorting on the initial population based on the fitness value, obtain a population stratification result, and calculate the crowding distance between individuals; Selecting high-quality individuals to form a parent population according to the population stratification result and the crowding distance; The third unit is used to perform an adaptive crossover operation on the individuals of the parent population, dynamically adjust the crossover position based on the fitness value of the individuals, and generate offspring individuals; perform taboo search optimization on the offspring individuals to generate a local search solution; adaptively adjust the mutation probability according to the distribution density of the individuals in the solution space, perform a mutation operation on the local search solution to generate mutant individuals; merge the mutant individuals with the parent individuals, and iteratively select a new generation of population; repeat the execution until the population converges, and output a scheduling plan that achieves optimal balance on multi-dimensional scheduling indicators.

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