Die casting island process parameter optimization method and system based on machine learning

Abstract

The invention provides a die-casting island process parameter optimization method and system based on machine learning, and relates to the technical field of die-casting processes, and the method comprises the steps: obtaining process parameters in a die-casting process, calculating an energy loss value and an entropy weight of each process stage, recognizing a target process stage, and carrying out the local optimization. Constructing a low-dimensional space for the historical process parameters, and clustering based on geodesic distances; selecting a historical parameter with the optimal quality as a search starting point, and executing parallel parameter search to obtain an optimal process parameter combination; and an adjusting instruction is issued to the die-casting island control system. According to the method, the quality of die-casting products can be improved, energy consumption is reduced, and self-adaptive optimization is achieved.

Description

Technical Field

[0001] This invention relates to the field of die casting technology, and in particular to a method and system for optimizing process parameters of die casting islands based on machine learning. Background Technology

[0002] Die casting is an important metal forming technology widely used in industries such as automotive, electronics, and aerospace. With the development of Industry 4.0, die casting islands, as a typical equipment combination in intelligent manufacturing, integrate multiple devices such as die casting machines, furnaces, spray systems, and part-retrieving robots. They are characterized by numerous process parameters, complex control, and high product quality requirements. Traditional die casting process parameter optimization mainly relies on experience accumulation and repeated experiments, which is time-consuming and has limited effectiveness. In recent years, the application of machine learning technology in the manufacturing field has provided new ideas for process parameter optimization.

[0003] Existing machine learning-based methods for optimizing die-casting process parameters primarily employ neural networks and genetic algorithms to construct a mapping relationship between process parameters and product quality, obtaining the optimal combination of process parameters through reverse derivation. Other research combines physical models with data-driven methods to establish multi-physics models of the die-casting process for parameter optimization. Furthermore, some researchers are exploring the use of artificial intelligence technologies such as reinforcement learning for online optimization and control of the die-casting process to adapt to dynamic changes in production.

[0004] Current technologies lack a systematic analysis of energy utilization efficiency at different stages of the die-casting process, making it difficult to identify critical stages with high energy losses. This results in a lack of clear optimization direction for parameter optimization, reducing optimization efficiency. Traditional methods fail to effectively handle the high-dimensional characteristics of die-casting process parameters. The complex coupling relationships between parameters lead to a large optimization search space, resulting in high computational complexity, low optimization efficiency, and difficulty in achieving rapid optimization. Therefore, designing a fast, efficient, and accurate method for optimizing die-casting island process parameters is of great significance. Summary of the Invention

[0005] This invention provides a method and system for optimizing process parameters of die-casting islands based on machine learning, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for optimizing die-casting island process parameters based on machine learning, comprising: The process parameters of the die casting process are obtained and divided into multiple process stages according to the die casting process. Based on the energy conversion relationship in the die casting process, the energy loss value of each process stage is calculated. Based on the product quality data, the entropy weight of each process stage is calculated, the target process stage affecting the quality is identified, and when the energy loss value of the target process stage is higher than the loss threshold, the process parameters of that process stage are optimized to obtain locally optimized process parameters. Historical process parameters are constructed into a historical parameter matrix, and eigenvalue decomposition is performed on its covariance matrix to construct a low-dimensional parameter space. The historical process parameters are mapped to the low-dimensional parameter space. Based on the locally optimized process parameters, the geodesic distance is calculated, and historical process parameter combinations with geodesic distances less than a distance threshold are grouped into the same category. The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination. Based on the optimal combination of process parameters, an adjustment command for the process parameters of the die-casting island is generated and sent to the die-casting island control system to adjust the process parameters of each piece of equipment in real time.

[0007] Based on product quality data, the entropy weight of each process stage is calculated to identify the target process stage affecting quality. When the energy loss value of the target process stage exceeds the loss threshold, the process parameters of that process stage are optimized to obtain locally optimized process parameters, including: Collect multidimensional quality inspection index data of die-cast products, and normalize the multidimensional quality inspection index data to obtain normalized inspection data. The weight value is obtained by dividing the normalized detection data by the sum of all normalized data of the corresponding detection indicator. The information entropy of each detection indicator is obtained by dividing the sum of the products of the weight value and its natural logarithm by the sample size. The correlation coefficient of each process parameter to each detection index is calculated using the Spearman correlation coefficient. The correlation coefficients are weighted and summed to obtain the correlation degree between the process stage and each detection index. The correlation degree is added to the product of the corresponding information entropy to obtain the entropy weight of each process stage. The process stage with the largest entropy weight is selected as the target process stage. When the energy loss value of the target process stage is higher than the preset loss threshold, calculate the first ratio of the energy loss value to the preset loss threshold and the second ratio of the mean of each detection index to the standard value, and then sum the first ratio and the second ratio by weight to obtain the local optimization target value. Calculate the gradient value of the local optimization target value with respect to each process parameter, and optimize the process parameters of the target process stage based on the gradient value; when the deviation value of the local optimization target value between two adjacent iterations is less than the convergence threshold, record the current process parameter as the local optimization process parameter.

[0008] Historical process parameters are constructed into a historical parameter matrix. Eigenvalue decomposition is performed on its covariance matrix, and a low-dimensional parameter space is constructed. The historical process parameters are mapped to this low-dimensional parameter space. Geodesic distances are calculated based on the locally optimized process parameters. Historical process parameters with geodesic distances less than a distance threshold are grouped into the same category, including: The historical process parameter combinations during the die casting process are obtained and constructed into a historical parameter matrix; the historical parameter matrix is ​​multiplied by its transpose and divided by the number of samples to obtain the covariance matrix; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and eigenvectors. The historical parameter matrix is ​​then multiplied by the eigenvectors whose eigenvalues ​​are greater than the eigenvalue threshold to obtain the projection values ​​of the historical process parameter combinations in each eigendirection. The projection values ​​are then multiplied by the contribution rates corresponding to each eigenvalue to obtain the mapping coordinates of the historical process parameters in the low-dimensional parameter space. The contribution rates are calculated by dividing the eigenvalues ​​by the sum of the eigenvalues. Calculate the partial derivatives of the local optimized process parameters with respect to the coordinates in the low-dimensional parameter space; multiply the partial derivatives by the randomly initialized weight parameters, sum them, and add them to the identity matrix to obtain the Riemann metric matrix; use the Riemann metric matrix to solve for the minimum path length between any two historical process parameters in the low-dimensional parameter space using the variational method, and use the minimum path length as the geodesic distance; Historical process parameters whose geodesic distance is less than a preset distance threshold are grouped into the same category to obtain the classification results of historical process parameters.

[0009] Using the Riemann metric matrix, the optimal geodesic path for any two historical process parameters in the low-dimensional parameter space is solved by variational method. The path length of the optimal geodesic path is taken as the geodesic distance, including: For any two historical process parameters, the geodesic path is calculated using linear interpolation in a low-dimensional parameter space. The gradient norm of the Riemann metric matrix on the geodesic path is calculated. The gradient norm is then mapped exponentially to obtain the path weight. The path weight is multiplied by the square root of the derivative of the geodesic path under the Riemann metric matrix to obtain the path length. Calculate the variational derivative of the path length with respect to the geodesic path, and add the variational derivative to an exponentially decaying regularization term to obtain the path update direction; calculate the first-order and second-order derivative norms of the geodesic path, and divide the second-order derivative norm by the square of the first-order derivative norm to obtain the local curvature of the path; multiply the baseline iteration step size by the negative exponential function of the local curvature of the path to obtain the iteration step size; In each iteration, the relative change rate of length is obtained by dividing the difference in path length between two adjacent iterations by the path length of the previous iteration, and the difference in local curvature between two adjacent iterations is obtained by calculating the curvature difference value. When the relative change rate of length is less than the length stability threshold and the curvature difference value is less than the curvature stability threshold, the geodesic path of the current iteration is determined as the optimal geodesic path, and the path length of the optimal geodesic path is used as the geodesic distance between two historical process parameter points.

[0010] The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination, including: The quality evaluation value is obtained by multiplying the data of multiple quality inspection indicators corresponding to the process parameters with the corresponding weight coefficients and summing them. In each category, the historical process parameter with the highest quality evaluation value is selected as the search starting point. The mutual information entropy between the search starting point and the quality evaluation value is calculated. For search starting points with mutual information entropy higher than the importance threshold, the range of their parameter values ​​is divided into grid cells of multiple times the baseline number to construct a local search space. For search starting points with mutual information entropy lower than the importance threshold, the range of their parameter values ​​is divided into grid cells of the baseline number to construct a global search space. Calculate the process parameter difference vector between the current iteration position and the previous iteration position of each search starting point, and divide the difference vector by its magnitude to obtain the search direction vector; calculate the quality evaluation value difference between adjacent iteration positions of each search starting point to obtain the quality improvement value; When the maximum quality improvement value of all search starting points is less than the quality improvement threshold, obtain the process parameter combination corresponding to the current iteration position of all search starting points, and select the process parameter combination with the highest quality evaluation value as the optimal process parameter combination.

[0011] After obtaining the quality improvement value, the method further includes: Based on the search direction vector and quality improvement value of each search starting point, a direction-reward mapping matrix for parallel search is established; The expected return of each candidate search direction is calculated based on the direction-return mapping matrix, and the candidate search direction with the highest expected return is selected as the search direction of each search starting point. When the expected return of a certain search starting point is lower than the expected value after a preset number of consecutive searches, the search direction with the highest expected return is obtained from the adjacent search starting point.

[0012] Based on the direction-reward mapping matrix, the expected reward of each candidate search direction is calculated, and the candidate search direction with the highest expected reward is selected as the search direction for each search starting point, including: Based on the direction-reward mapping matrix, the angle between adjacent search direction vectors is calculated and negative exponential operation is performed to obtain the direction similarity. The correlation coefficient between adjacent quality improvement values ​​is calculated. The direction similarity and the correlation coefficient are weighted and summed to obtain the correlation degree. For a candidate search direction, calculate the angle between it and each search direction vector in the direction-reward mapping matrix and perform a negative exponential operation to obtain the candidate direction similarity; multiply the candidate direction similarity by the sum of the corresponding quality improvement value and the correlation degree, and normalize and weight the calculation results of all search direction vectors to obtain the initial expected reward of the candidate search direction; Calculate the negative exponential product of the current time and the historical time difference as the time decay weight; adjust the initial expected return based on the time decay weight to obtain the expected return, and select the candidate search direction with the highest expected return as the search direction of each search starting point.

[0013] A second aspect of the present invention provides a machine learning-based die-casting island process parameter optimization system, comprising: The first unit is used to acquire process parameters in the die casting process, divide the process parameters into multiple process stages according to the die casting process, calculate the energy loss value of each process stage according to the energy conversion relationship in the die casting process, calculate the entropy weight of each process stage according to product quality data, identify the target process stage that affects quality, and when the energy loss value of the target process stage is higher than the loss threshold, optimize the process parameters of that process stage to obtain locally optimized process parameters. The second unit is used to construct a historical parameter matrix from historical process parameters, perform eigenvalue decomposition on its covariance matrix, and construct a low-dimensional parameter space. The historical process parameters are mapped to the low-dimensional parameter space. Based on the locally optimized process parameters, the geodesic distance is calculated, and historical process parameter combinations with geodesic distances less than a distance threshold are grouped into the same category. The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination. The third unit is used to generate adjustment instructions for the process parameters of the die-casting island based on the optimal combination of process parameters, and to send the adjustment instructions to the die-casting island control system to adjust the process parameters of each piece of equipment in real time.

[0014] A third aspect of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] Fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] The beneficial effects of this application are as follows: This invention utilizes a machine learning-based method for optimizing process parameters in die-casting islands. This method can accurately identify key process stages that affect product quality. By combining entropy weight calculation with energy loss analysis, it achieves precise local optimization of process parameters, reduces unnecessary parameter adjustments, and improves optimization efficiency.

[0017] This invention employs eigenvalue decomposition and low-dimensional parameter space construction techniques, combined with geodesic distance classification methods, to effectively reduce the search complexity of process parameters. It can quickly locate the optimal combination of process parameters in a complex multi-parameter space, achieving an effective transition from local optimization to global optimization and improving the accuracy and efficiency of parameter optimization.

[0018] This invention achieves dynamic optimization and automatic adjustment of process parameters by sending adjustment commands to the die-casting island control system in real time, reducing manual intervention, lowering the risk of operational errors, improving the stability and consistency of die-casting production, effectively improving product quality, and reducing energy consumption and production costs. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the machine learning-based method for optimizing process parameters of a die-casting island according to an embodiment of the present invention. Figure 2 A schematic diagram of the system architecture for optimizing die-casting process parameters based on energy loss and product quality data. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the machine learning-based die-casting island process parameter optimization method according to an embodiment of the present invention. Figure 1 As shown, the method includes: The process parameters of the die casting process are obtained and divided into multiple process stages according to the die casting process. Based on the energy conversion relationship in the die casting process, the energy loss value of each process stage is calculated. Based on the product quality data, the entropy weight of each process stage is calculated, the target process stage affecting the quality is identified, and when the energy loss value of the target process stage is higher than the loss threshold, the process parameters of that process stage are optimized to obtain locally optimized process parameters. Historical process parameters are constructed into a historical parameter matrix, and eigenvalue decomposition is performed on its covariance matrix to construct a low-dimensional parameter space. The historical process parameters are mapped to the low-dimensional parameter space. Based on the locally optimized process parameters, the geodesic distance is calculated, and historical process parameter combinations with geodesic distances less than a distance threshold are grouped into the same category. The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination. Based on the optimal combination of process parameters, an adjustment command for the process parameters of the die-casting island is generated and sent to the die-casting island control system to adjust the process parameters of each piece of equipment in real time.

[0023] Figure 2 This is a schematic diagram of a system architecture for optimizing die-casting process parameters based on energy loss and product quality data. In one optional implementation, the entropy weight of each process stage is calculated based on product quality data to identify target process stages affecting quality. When the energy loss value of a target process stage exceeds a loss threshold, the process parameters of that process stage are optimized to obtain locally optimized process parameters, including: Collect multidimensional quality inspection index data of die-cast products, and normalize the multidimensional quality inspection index data to obtain normalized inspection data. The weight value is obtained by dividing the normalized detection data by the sum of all normalized data of the corresponding detection indicator. The information entropy of each detection indicator is obtained by dividing the sum of the products of the weight value and its natural logarithm by the sample size. The correlation coefficient of each process parameter to each detection index is calculated using the Spearman correlation coefficient. The correlation coefficients are weighted and summed to obtain the correlation degree between the process stage and each detection index. The correlation degree is added to the product of the corresponding information entropy to obtain the entropy weight of each process stage. The process stage with the largest entropy weight is selected as the target process stage. When the energy loss value of the target process stage is higher than the preset loss threshold, calculate the first ratio of the energy loss value to the preset loss threshold and the second ratio of the mean of each detection index to the standard value, and then sum the first ratio and the second ratio by weight to obtain the local optimization target value. Calculate the gradient value of the local optimization target value with respect to each process parameter, and optimize the process parameters of the target process stage based on the gradient value; when the deviation value of the local optimization target value between two adjacent iterations is less than the convergence threshold, record the current process parameter as the local optimization process parameter.

[0024] This invention relates to a method for optimizing die-casting process parameters based on energy loss and product quality data. In practical implementation, by calculating the entropy weight of each process stage, the target process stage with the most significant impact on product quality is identified. When the energy loss in this process stage is too high, the process parameters are optimized to improve product quality and reduce energy consumption.

[0025] In implementing this invention, it is necessary to collect multi-dimensional quality inspection index data of die-cast products. Taking a die-casting production line as an example, data on five quality inspection indicators—surface roughness, dimensional accuracy, porosity, hardness, and tensile strength—were collected from 100 aluminum alloy die-cast parts. Since these indicators have different dimensions and numerical ranges, they need to be normalized to eliminate the influence of dimensions. For indicators where smaller values ​​are better (such as surface roughness, dimensional deviation, and porosity), the minimum-maximum normalization method is used, with the formula being: the normalized value equals the maximum value minus the current value, then divided by the maximum value minus the minimum value. For indicators where larger values ​​are better (such as hardness and tensile strength), the formula is: the normalized value equals the current value minus the minimum value, then divided by the maximum value minus the minimum value. After normalization, the values ​​of all inspection indicators are mapped to the range of 0 to 1.

[0026] After normalization, the information entropy of each test index is calculated. Taking surface roughness as an example, the proportion of the normalized surface roughness value of the first product to the sum of the normalized surface roughness values ​​of all 100 products is calculated, yielding 0.0103. This process is repeated to calculate the proportion of surface roughness values ​​for all 100 products. Then, the product of each proportion and its natural logarithm is calculated; for example, 0.0103 multiplied by ln(0.0103) yields -0.0472. This product is summed for all products, resulting in -4.5861. Dividing this by the sample size of 100, the information entropy of the surface roughness index is 0.4586. The information entropy of other test indexes is calculated using the same method. Assuming the information entropy of dimensional accuracy is 0.3921, porosity is 0.5217, hardness is 0.4102, and tensile strength is 0.4738, the information entropy is calculated accordingly.

[0027] In the die casting process, the process parameters can be divided into five stages: melting, conveying, injection, cooling, and demolding. Each stage contains multiple process parameters. In the melting stage, key process parameters include furnace power, heating curve, stirring speed, degassing time, and slag removal rate. In the conveying stage, key process parameters include conveying temperature, holding power, metal flow rate, conveying pressure, and filtration efficiency. In the injection stage, key process parameters include injection speed, injection pressure, speed switching point location, injection speed, and nozzle temperature. In the cooling stage, key process parameters include cooling time, die casting temperature, holding time, cooling channel distribution, and cooling efficiency. In the demolding stage, key process parameters include mold opening speed, mold opening distance, ejection speed, release agent spraying pressure, and spraying uniformity. Based on these process parameters, the energy loss values ​​for different process stages can be calculated for subsequent data analysis. On a certain die casting production line, a batch of aluminum alloy parts is produced, with each part weighing 2 kg. The melting stage processes 200 kg of aluminum alloy, producing 100 parts. Based on the energy conversion relationships at each stage, the energy loss value for each stage can be calculated. In the smelting stage, the rated power of the resistance furnace is 85 kW, the actual operating power is 78 kW, and the smelting time is 45 minutes. The input energy is calculated as: 78 kW × 0.75 hours = 58.5 kWh. The theoretical melting energy of the aluminum alloy is calculated as: 200 kg × [0.9 kJ / (kg·℃) × (690℃ - 25℃) + 397 kJ / kg] = 33.12 kWh. Where 0.9 kJ / (kg·℃) is the specific heat capacity of the aluminum alloy, 690℃ is the smelting temperature, 25℃ is the initial temperature, and 397 kJ / kg is the latent heat of fusion of the aluminum alloy. The total energy loss in the smelting stage is 58.5 - 33.12 = 25.38 kWh, and the energy loss per unit is 25.38 ÷ 100 = 0.2538 kWh. Energy loss is broken down as follows: furnace body heat dissipation loss of 12.69 kWh; heat loss during furnace opening of 7.61 kWh; and power conversion efficiency loss of 5.08 kWh.

[0028] During the conveying phase, the insulated conveying device has a power of 15 kW, with an actual operating power of 12 kW and a conveying time of 3.5 minutes. The input energy is calculated as: 12 kW × 0.058 hours = 0.7 kWh. During conveying, the metal temperature drops from 690℃ to 665℃, and the heat loss is calculated as: 200 kg × 0.9 kJ / (kg·℃) × (690℃ - 665℃) ÷ 3600 = 1.25 kWh. Considering the 0.7 kWh power supply from the insulation equipment, the actual energy loss during the conveying phase is 1.25 - 0.7 = 0.55 kWh, and the energy loss per unit is 0.55 ÷ 100 = 0.0055 kWh. The energy loss is further broken down as follows: heat dissipation from metal to air (0.275 kWh); conduction loss through the conveyor wall (0.165 kWh); and radiation loss (0.11 kWh).

[0029] During the injection stage, the hydraulic pump power of the injection system is 55 kW. The injection process is divided into two sub-stages: low-speed injection and high-speed injection. In the low-speed injection stage, the actual power of the hydraulic system is 25 kW, lasting for 2.5 seconds, with an energy input of 25 kW × (2.5 ÷ 3600) hours = 0.0174 kWh. In the high-speed injection stage, the actual power of the hydraulic system is 48 kW, lasting for 0.8 seconds, with an energy input of 48 kW × (0.8 ÷ 3600) hours = 0.0107 kWh. The total energy input for the injection stage is 0.0174 + 0.0107 = 0.0281 kWh. By measuring the pressure and displacement of the injection cylinder, the effective mechanical work is calculated as follows: 0.0063 kWh for the low-speed injection stage and 0.0046 kWh for the high-speed injection stage, for a total effective mechanical work of 0.0109 kWh. The total energy loss during the injection stage is 0.0281 - 0.0109 = 0.0172 kWh, and the energy loss per unit is 0.0172 kWh. This energy loss is broken down as follows: hydraulic system mechanical friction loss (0.0069 kWh); oil flow resistance loss (0.0052 kWh); internal leakage loss of hydraulic components (0.0034 kWh); and system vibration loss (0.0017 kWh).

[0030] During the cooling stage, the water pump power of the mold cooling system is 12 kW, the actual operating power is 10 kW, and the cooling time is 25 seconds. The energy input is calculated as: 10 kW × (25 ÷ 3600) hours = 0.0694 kWh. The cooling water flow rate is 85 liters / minute, the inlet water temperature is 25℃, and the outlet water temperature is 38℃. The heat absorbed by the cooling water is calculated as: 85 × (25 ÷ 60) × 4.2 × (38 - 25) ÷ 3600 = 0.2316 kWh. Here, 4.2 kJ / (kg·℃) is the specific heat capacity of water. The heat released when the metal cools from 665℃ to 180℃ is calculated as: 2 kg × [0.9 kJ / (kg·℃) × (665℃ - 180℃)] + 2 kg × 397 kJ / kg = 1.7154 kWh. The cooling system absorbs 0.2316 kWh of heat, plus an energy input of 0.0694 kWh, totaling 0.301 kWh. The energy loss during the cooling phase is 1.7154 - 0.301 = 1.4144 kWh, with a single product energy loss of 0.2015 kWh (considering the mold can produce 7 products simultaneously). The energy loss is broken down as follows: heat exchange loss between the mold and the environment (0.7072 kWh); heat loss through the cooling system piping (0.4243 kWh); and heat conduction loss between the cooling water and the piping (0.2829 kWh).

[0031] During the demolding stage, the hydraulic system has a power of 18 kW and consists of two actions: mold opening and ejection. During mold opening, the actual power of the hydraulic system is 15 kW, lasting for 4 seconds, with an energy input of 15 kW × (4 ÷ 3600) hours = 0.0167 kWh. During ejection, the actual power of the hydraulic system is 10 kW, lasting for 2 seconds, with an energy input of 10 kW × (2 ÷ 3600) hours = 0.0056 kWh. The total energy input during the demolding stage is 0.0167 + 0.0056 = 0.0223 kWh. The effective mechanical work is calculated using force and displacement sensors: 0.0085 kWh during mold opening and 0.0028 kWh during ejection, for a total effective mechanical work of 0.0113 kWh. The total energy loss during the demolding stage is 0.0223 - 0.0113 = 0.011 kWh, and the energy loss per product is 0.0016 kWh (considering that the mold demolds 7 products simultaneously). The energy loss is broken down as follows: hydraulic system mechanical friction loss of 0.0044 kWh (accounting for 40% of the total loss); oil flow resistance loss of 0.0033 kWh; mold guide post friction resistance loss of 0.0022 kWh; and system vibration loss of 0.0011 kWh.

[0032] Based on the above calculations, the energy losses at each stage of producing an aluminum alloy steering knuckle are as follows: melting stage 0.2538 kWh, conveying stage 0.0055 kWh, injection stage 0.0172 kWh, cooling stage 0.2015 kWh, and demolding stage 0.0016 kWh, for a total energy loss of 0.4796 kWh. The energy loss analysis shows that the melting and cooling stages have the highest energy losses, accounting for 52.9% and 42.0% of the total loss, respectively. The conveying, injection, and demolding stages combined account for only 5.1% of the total loss.

[0033] Determine the correlation between each process stage and quality indicators. Use Spearman's correlation coefficient to calculate the correlation between each process parameter and each quality inspection indicator. For example, for the injection speed parameter in the injection stage, the correlation coefficient with surface roughness is calculated to be 0.82, with dimensional accuracy to be 0.65, with porosity to be 0.79, with hardness to be 0.42, and with tensile strength to be 0.56. Similarly, calculate the correlation coefficients between all process parameters and each inspection indicator.

[0034] To obtain the correlation between each process stage and the test index, the correlation coefficients of all process parameters and a certain test index under each process stage are weighted and summed. The weights can be determined according to the importance of each parameter. Assuming that the injection stage includes three parameters—injection speed, injection pressure, and clamping force—the weights are 0.5, 0.3, and 0.2, respectively. The calculated correlation between the injection stage and surface roughness is 0.74, with dimensional accuracy 0.61, with porosity 0.73, with hardness 0.48, and with tensile strength 0.58.

[0035] To calculate the entropy weight of each process stage, the correlation between each process stage and each detection index is multiplied by the information entropy of the corresponding detection index and then summed. Taking the injection stage as an example, its entropy weight is calculated as 0.74×0.4586 + 0.61×0.3921 + 0.73×0.5217 + 0.48×0.4102 + 0.58×0.4738 = 1.6436. Similarly, the entropy weights of other process stages are calculated: 1.2341 for the melting stage, 0.9872 for the conveying stage, 1.4510 for the cooling stage, and 0.8765 for the demolding stage. Comparing the entropy weights of each process stage, it is found that the injection stage has the largest entropy weight, at 1.6436. Therefore, the injection stage is selected as the target process stage.

[0036] Monitoring the energy loss during the injection stage, assuming the current energy loss is 85 kWh and the preset loss threshold is 75 kWh, the energy loss exceeds the preset threshold, requiring process parameter optimization. The first ratio of energy loss to the loss threshold is calculated as 85 / 75 = 1.133, indicating an energy exceedance of 13.3%. The second ratio of the average value of each test indicator to the standard value is calculated. For example, if the average surface roughness is 4.2 μm and the standard value is 3.5 μm, the second ratio is 4.2 / 3.5 = 1.2, indicating a quality exceedance of 20%. The second ratios for other test indicators are calculated similarly.

[0037] We assigned weights of 0.4 and 0.6 to energy loss and quality indicators, respectively, and calculated the local optimization target values. For surface roughness, the local optimization target value is 0.4 × 1.133 + 0.6 × 1.2 = 1.1732. Similarly, we calculated the local optimization target values ​​for other detection indicators.

[0038] Based on the local optimization objective value, the gradient descent method is used to optimize the process parameters. The gradient value of the local optimization objective value with respect to each process parameter is calculated, i.e., the sensitivity of the objective value to parameter changes. For example, for injection speed, the gradient value is 0.25, meaning that increasing the injection speed will cause the objective value to increase by 0.25; for injection pressure, the gradient value is -0.18, meaning that increasing the injection pressure will cause the objective value to decrease by 0.18. The process parameters are adjusted according to the gradient values, i.e., the parameters corresponding to positive gradients are decreased, and the parameters corresponding to negative gradients are increased.

[0039] After multiple iterations of optimization, when the deviation between the local optimization target values ​​of two adjacent iterations is less than the preset convergence threshold of 0.001, the iteration stops and the current process parameters are recorded as the local optimization process parameters. After optimization, the injection speed decreased from 2.8 m / s to 2.5 m / s, the injection pressure increased from 120 MPa to 135 MPa, and the clamping force was adjusted from 4500 tons to 4200 tons. With this set of optimized parameters, energy loss decreased to 72 kWh, lower than the preset threshold of 75 kWh, and product quality inspection indicators also improved, with surface roughness reduced to 3.7 micrometers, closer to the standard value of 3.5 micrometers.

[0040] In one optional implementation, historical process parameters are constructed into a historical parameter matrix, their covariance matrix is ​​decomposed using eigenvalue decomposition, and a low-dimensional parameter space is constructed. The historical process parameters are mapped to this low-dimensional parameter space. Geodesic distances are calculated based on the locally optimized process parameters. Historical process parameters with geodesic distances less than a distance threshold are grouped into the same category, including: The historical process parameter combinations during the die casting process are obtained and constructed into a historical parameter matrix; the historical parameter matrix is ​​multiplied by its transpose and divided by the number of samples to obtain the covariance matrix; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and eigenvectors. The historical parameter matrix is ​​then multiplied by the eigenvectors whose eigenvalues ​​are greater than the eigenvalue threshold to obtain the projection values ​​of the historical process parameter combinations in each eigendirection. The projection values ​​are then multiplied by the contribution rates corresponding to each eigenvalue to obtain the mapping coordinates of the historical process parameters in the low-dimensional parameter space. The contribution rates are calculated by dividing the eigenvalues ​​by the sum of the eigenvalues. Calculate the partial derivatives of the local optimized process parameters with respect to the coordinates in the low-dimensional parameter space; multiply the partial derivatives by the randomly initialized weight parameters, sum them, and add them to the identity matrix to obtain the Riemann metric matrix; use the Riemann metric matrix to solve for the minimum path length between any two historical process parameters in the low-dimensional parameter space using the variational method, and use the minimum path length as the geodesic distance; Historical process parameters whose geodesic distance is less than a preset distance threshold are grouped into the same category to obtain the classification results of historical process parameters.

[0041] This embodiment provides a method for classifying historical process parameters based on geodesic distance. The method constructs a historical parameter matrix, performs dimensionality reduction through eigenvalue decomposition, calculates geodesic distances in the low-dimensional space, and ultimately classifies the historical process parameters.

[0042] During the die-casting production process, a large amount of historical process parameter combination data is collected. These process parameters include, but are not limited to: mold temperature, alloy temperature, injection speed, boosting pressure, low-speed range, and holding time. For example, 100 sets of historical process parameter combinations are collected, each containing 10 different process parameters. These parameters are arranged in a row-column manner to construct a 100-row, 10-column historical parameter matrix P. In this matrix, each row represents a set of process parameter combinations, and each column represents a type of process parameter.

[0043] To calculate the covariance matrix, the history parameter matrix P and its transpose P T Multiply the results and then divide by the sample size of 100 to obtain a 10×10 covariance matrix C. The covariance matrix C reflects the correlation between different process parameters.

[0044] Eigenvalue decomposition of the covariance matrix C yields 10 eigenvalues ​​λ1, λ2, ..., λ3. 10 and their corresponding feature vectors v1, v2, ..., v 10 The eigenvalues ​​are arranged in descending order, for example, λ1=5.8, λ2=2.3, λ3=1.2, λ4=0.7, λ5=0.4, etc. A feature threshold of 0.8 is set, and the eigenvectors with eigenvalues ​​greater than 0.8, namely v1, v2, and v3, are selected as the primary feature directions.

[0045] Calculate the projection values ​​of historical process parameters onto each major characteristic direction. For each row of data p in the historical parameter matrix P... i The projection value proj is obtained by performing inner product operations with the feature vectors v1, v2, and v3 respectively. i1 ,proj i2 and proj i3 For example, for the first set of process parameters p1, the calculated proj 11 =25.6, proj 12 =10.2, proj 13 =-5.8.

[0046] Calculate the contribution rate corresponding to each eigenvalue. The contribution rate equals the eigenvalue divided by the sum of the eigenvalues. The sum of the eigenvalues ​​is λ1 + λ2 + ... + λ. 10=11.3. Therefore, the contribution rate of λ1 is 5.8 / 11.3=0.513, the contribution rate of λ2 is 2.3 / 11.3=0.204, and the contribution rate of λ3 is 1.2 / 11.3=0.106.

[0047] Multiplying the projected value by the contribution rate of the corresponding eigenvalue yields the mapped coordinates of the historical process parameters in the low-dimensional parameter space. Taking the first set of process parameters p1 as an example, its low-dimensional mapped coordinates are (25.6×0.513, 10.2×0.204, -5.8×0.106)=(13.13, 2.08, -0.61). Through similar calculations, the mapped coordinates of all historical process parameters in the three-dimensional low-dimensional parameter space can be obtained.

[0048] Obtain the locally optimized process parameters and map them to a low-dimensional parameter space, resulting in coordinates (15.5, 3.2, -0.8). Calculate the partial derivatives of these locally optimized process parameters with respect to the low-dimensional parameter space coordinates. Specifically, calculate the partial derivatives of the optimization objective function f (e.g., a product quality indicator) with respect to the low-dimensional coordinates x, y, and z, obtaining the partial derivative vector grad. f =(∂f / ∂x, ∂f / ∂y, ∂f / ∂z)=(0.23, 0.15, -0.08).

[0049] Randomly initialize the weight parameters w1=0.5, w2=0.3, w3=0.2. Multiply the partial derivatives by the weight parameters and sum them to obtain w1×(∂f / ∂x). 2 + w2×(∂f / ∂y) 2 + w3×(∂f / ∂z) 2 =0.5×0.23 2 + 0.3×0.15 2 + 0.2×(-0.08) 2 =0.03. Adding this value to the 3×3 identity matrix yields the Riemann metric matrix G.

[0050] Using the Riemannian metric matrix G, the minimum path length, i.e., the geodesic distance, between any two historical process parameters in the low-dimensional parameter space is calculated using the variational method. For example, the geodesic distance between the locally optimized process parameter and the 25th set of historical process parameters is 2.6, the geodesic distance between it and the 37th set of historical process parameters is 1.8, and the geodesic distance between it and the 42nd set of historical process parameters is 3.5.

[0051] A preset distance threshold of 3.0 was set, and historical process parameters with geodesic distances less than 3.0 were grouped into the same category. Therefore, historical process parameters in groups 25 and 37 were grouped into the same category as locally optimized process parameters, while historical process parameters in group 42 were grouped into a different category. In this way, all historical process parameters can be divided into several categories, and the combinations of process parameters within each category have similar performance characteristics.

[0052] This geodesic distance-based classification method fully considers the geometric structure of process parameters in the low-dimensional parameter space and the gradient information of locally optimized process parameters. It can more accurately reflect the intrinsic relationships between parameter combinations, providing effective support for optimizing die-casting process parameters. By analyzing the common characteristics of similar process parameters, key factors for process improvement and adjustment directions can be further extracted, guiding the optimization and adjustment of the die-casting production process.

[0053] In one optional implementation, the optimal geodesic path for any two historical process parameters in the low-dimensional parameter space is solved using the Riemann metric matrix via variational method, and the path length of the optimal geodesic path is taken as the geodesic distance, including: For any two historical process parameters, the geodesic path is calculated using linear interpolation in a low-dimensional parameter space. The gradient norm of the Riemann metric matrix on the geodesic path is calculated. The gradient norm is then mapped exponentially to obtain the path weight. The path weight is multiplied by the square root of the derivative of the geodesic path under the Riemann metric matrix to obtain the path length. Calculate the variational derivative of the path length with respect to the geodesic path, and add the variational derivative to an exponentially decaying regularization term to obtain the path update direction; calculate the first-order and second-order derivative norms of the geodesic path, and divide the second-order derivative norm by the square of the first-order derivative norm to obtain the local curvature of the path; multiply the baseline iteration step size by the negative exponential function of the local curvature of the path to obtain the iteration step size; In each iteration, the relative change rate of length is obtained by dividing the difference in path length between two adjacent iterations by the path length of the previous iteration, and the difference in local curvature between two adjacent iterations is obtained by calculating the curvature difference value. When the relative change rate of length is less than the length stability threshold and the curvature difference value is less than the curvature stability threshold, the geodesic path of the current iteration is determined as the optimal geodesic path, and the path length of the optimal geodesic path is used as the geodesic distance between two historical process parameter points.

[0054] The technical solution described in this paper provides a method for calculating the optimal geodesic path between any two historical process parameters in a low-dimensional parameter space based on the Riemannian metric matrix and variational method, and uses the length of this path as the geodesic distance. To achieve this goal, this technical solution adopts the following specific implementation steps.

[0055] In the low-dimensional parameter space, for any two historical process parameter points p1 and p2, an initial geodesic path γ(t) is constructed using linear interpolation, where t∈[0,1]. Assuming p1 corresponds to parameter values ​​(2.5, 3.7) and p2 corresponds to parameter values ​​(8.3, 6.2), the initial path can be expressed as γ(t) = (2.5 + 5.8t, 3.7 + 2.5t). The gradient norm of the established Riemann metric matrix G on this path is calculated. For example, at the path point γ(0.5)=(5.4, 4.95), four points are selected near this point for sampling: f(5.4+h, 4.95) = 8.24, f(5.4-h, 4.95) = 8.06, f(5.4, 4.95+h) = 8.32, f(5.4, 4.95-h) = 8.07, where h=0.2 is the sampling step size. The gradient is calculated using the central difference formula: ∂f / ∂x1 = [f(5.4+h, 4.95) - f(5.4-h, 4.95)] / (2h) = (8.24 - 8.06) / (2×0.2) = 0.18 / 0.4 = 0.45, ∂f / ∂x2 = [f(5.4, 4.95+h) - f(5.4, 4.95-h)] / (2h) = (8.32 - 8.07) / (2×0.2) = 0.25 / 0.4 = 0.625 ≈ 0.62. Therefore, the gradient vector at the path point γ(0.5) is: ▽f|(5.4, 4.95) = [0.45, 0.62]. Under the Riemann metric matrix G, the γ(0.5) gradient norm of the path is ||▽f|| G = √(▽f T ·G -1 ·▽f), the specific calculation steps are as follows: G -1 ·▽f = [0.8724, -0.1758; -0.1758, 0.7577]·[0.45; 0.62]= [0.8724×0.45 + (-0.1758)×0.62, (-0.1758)×0.45 + 0.7577×0.62]= [0.3926 - 0.1090, -0.0791 + 0.4698]= [0.2836, 0.3907], then calculate ▽f. T ·(G -1 ·▽f):▽f T ·(G -1·▽f) = [0.45, 0.62]·[0.2836; 0.3907]= 0.45×0.2836 + 0.62×0.3907= 0.1276 + 0.2422= 0.3698. Finally, the gradient norm is obtained by calculating the square root: ||▽f|| G = √0.3698 = 0.608. This gradient norm is converted into a path weight using the exponential mapping function exp(gradient norm), where the weight value is exp(0.608)≈1.84. This path weight is multiplied by the square root of the geodesic path derivative under the Riemannian metric matrix to obtain the path element length. By integrating over the entire path, the initial total path length is calculated to be 12.36.

[0056] To find the optimal geodesic path, it is necessary to calculate the variational derivative of the path length with respect to the geodesic path. For each path point γ(t) i ), calculate its variational derivative value. For example, the variational derivative value at γ(0.5) is ∂E / ∂γ(0.5) = (∂w / ∂γ) · ds + w · (∂ds / ∂γ)= [0.552, 0.92] · 7.56 + 1.84 · [0.536,0.573]= [4.173, 6.955] + [0.986, 1.054]= [5.159, 8.009], then the variational derivative unit vector is: [5.159, 8.009] / 9.514 = [0.542, 0.842], after scaling and normalization, the variational derivative value is approximately [0.43, 0.67].

[0057] Add this variational derivative to the exponentially decaying regularization term to obtain the path update direction. If an exponential decay factor of 0.95 is used, the regularization coefficient for the first iteration is 0.05, and the regularized path update direction is (1-0.05)·[0.43, 0.67] = 0.95·[0.43, 0.67]= [0.41, 0.64].

[0058] The step size needs to be adaptively adjusted during the iteration process. The first and second derivative norms of the geodesic path γ(t) are calculated. Assuming that in a certain iteration, the first derivative norm is 6.32 and the second derivative norm is 2.15, the local curvature of the path is calculated as follows. If the baseline iteration step size is set to 0.1, then the actual iteration step size is 0.1 × exp(-0.054) ≈ 0.095.

[0059] The geodesic path is updated using the calculated update direction and iteration step size. For example, if the current path point γ(0.5) = (5.4, 4.95), the update direction is (0.41, 0.64), and the iteration step size is 0.095, then the updated path point is (5.4, 4.95) + 0.095 × (0.41, 0.64) = (5.439, 5.011). All points on the path are updated in this way to obtain the new path γ'(t).

[0060] After each iteration, it's necessary to evaluate whether the iteration has converged. The relative rate of change of length is obtained by dividing the difference in path length between two adjacent iterations by the path length of the previous iteration. Assuming the path length in the previous iteration was 12.36 and the current iteration's path length is 12.31, the relative rate of change of length is |12.31-12.36| / 12.36 ≈ 0.004. Simultaneously, the difference in local curvature between two adjacent iterations is calculated. Assuming the average local curvature in the previous iteration was 0.054 and the average local curvature in the current iteration is 0.052, the curvature difference is |0.052-0.054| = 0.002.

[0061] The length stability threshold is set to 0.005, and the curvature stability threshold is set to 0.003. When the relative rate of change of length is less than the length stability threshold and the curvature difference is less than the curvature stability threshold, the iteration is considered converged, and the geodesic path of the current iteration is determined as the optimal geodesic path. In this example, the relative rate of change of length (0.004) is less than the threshold (0.005), and the curvature difference (0.002) is less than the threshold (0.003), therefore the iteration converges, and the current path is the optimal geodesic path. Its path length of 12.31 is used as the geodesic distance between the two historical process parameter points.

[0062] In practical applications, multiple iterations are required to reach convergence. For example, when dealing with more complex parameter spaces, 20-30 iterations may be needed. To prevent infinite iterations, a maximum number of iterations can be set to 50. If convergence is not achieved after reaching the maximum number of iterations, the path from the last iteration is taken as the approximate optimal solution.

[0063] Once the geodesic distances are calculated, they can be used for subsequent similarity analysis of process parameters. For example, for a set of historical process parameters {p1, p2, p3, p4, p5}, the calculated geodesic distance matrix shows that the distance between p1 and p3 is the smallest, at 8.76, indicating that these two sets of process parameters are most similar; while the distance between p2 and p5 is the largest, at 25.42, indicating that their process characteristics are significantly different. Based on these geodesic distances, a reference basis is provided for the design of new process parameters.

[0064] Through the above implementation methods, this technical solution can accurately calculate the optimal geodesic path and geodesic distance between any two historical process parameters in a low-dimensional parameter space, providing a theoretical basis and practical tool for measuring the similarity of process parameters.

[0065] In one optional implementation, the historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination, including: The quality evaluation value is obtained by multiplying the data of multiple quality inspection indicators corresponding to the process parameters with the corresponding weight coefficients and summing them. In each category, the historical process parameter with the highest quality evaluation value is selected as the search starting point. The mutual information entropy between the search starting point and the quality evaluation value is calculated. For search starting points with mutual information entropy higher than the importance threshold, the range of their parameter values ​​is divided into grid cells of multiple times the baseline number to construct a local search space. For search starting points with mutual information entropy lower than the importance threshold, the range of their parameter values ​​is divided into grid cells of the baseline number to construct a global search space. Calculate the process parameter difference vector between the current iteration position and the previous iteration position for each search starting point, and divide this difference vector by its magnitude to obtain the search direction vector; calculate the quality evaluation value difference between adjacent iteration positions for each search starting point to obtain the quality improvement value. When the maximum quality improvement value of all search starting points is less than the quality improvement threshold, obtain the process parameter combination corresponding to the current iteration position of all search starting points, and select the process parameter combination with the highest quality evaluation value as the optimal process parameter combination.

[0066] After obtaining the quality improvement value, the method further includes: Based on the search direction vector and quality improvement value of each search starting point, a direction-reward mapping matrix for parallel search is established; The expected return of each candidate search direction is calculated based on the direction-return mapping matrix, and the candidate search direction with the highest expected return is selected as the search direction of each search starting point. When the expected return of a certain search starting point is lower than the expected value after a preset number of consecutive searches, the search direction with the highest expected return is obtained from the adjacent search starting point.

[0067] In a specific implementation, a process parameter optimization method based on multi-starting point parallel search can be achieved through the following technical means.

[0068] This method collects a large amount of historical production data, including combinations of process parameters and corresponding quality inspection indicators. For the historical data, an unsupervised clustering algorithm is used to classify the combinations of process parameters. For example, the K-means clustering method can be used to divide the parameter space into multiple categories. In practice, the parameters can be divided into 5 to 8 categories, each representing a different parameter configuration pattern.

[0069] For each combination of process parameters, the system calculates its quality evaluation value through a weighted summation. Assuming the quality inspection index corresponding to a certain combination of process parameters is [98.2, 0.05, 142.3], and the preset weight coefficients are [0.5, -0.3, 0.2], then the quality evaluation value for this parameter combination is calculated as follows: 98.2 × 0.5 + 0.05 × (-0.3) + 142.3 × 0.2 = 49.1 + (-0.015) + 28.46 = 77.545. The weight coefficients can be set according to the actual production needs of the enterprise; a positive weight indicates that the higher the index, the better, and a negative weight indicates that the lower the index, the better.

[0070] Within each category, the system selects the historical process parameter combination with the highest quality evaluation value as the search starting point. For example, in the five categories, the process parameter combinations with quality evaluation values ​​of [82.3, 79.5, 84.1, 78.6, 81.2] are selected as the search starting points.

[0071] For each search starting point, the system calculates the mutual information entropy between it and the quality evaluation value. The mutual information entropy calculation is based on the degree of influence of parameter changes on the quality evaluation value, obtained through parameter sensitivity analysis. A higher mutual information entropy value indicates a significant impact of parameter changes on quality. The system sets an importance threshold of 0.65. For search starting points with a mutual information entropy higher than 0.65, the parameter value range is divided into grid cells of multiples of the baseline number. For example, if the baseline number is 10, it can be set to 30 grid cells to construct a local search space. For search starting points with a mutual information entropy lower than 0.65, the parameter value range is divided into 10 grid cells to construct a global search space.

[0072] The search process employs an iterative optimization approach. In each iteration, the system calculates the parameter difference vector between the current position of the search starting point and the position of the previous iteration. For example, if the parameters of a search starting point in the previous iteration were [210, 55, 0.8], and the parameters at the current position are [215, 58, 0.75], then the difference vector is [5, 3, -0.05]. Dividing this difference vector by its magnitude yields the search direction vector, calculated as √(5^2 + 3^2 + (-0.05)^2) = 5.83. Therefore, the search direction vector is [0.857, 0.514, -0.0086].

[0073] Simultaneously, the difference in quality evaluation values ​​between adjacent iteration positions is calculated. For example, if the quality evaluation value at the previous position is 79.5 and the current position is 82.3, then the quality improvement value is 2.8. Based on the search direction vector and quality improvement value at each search starting point, the system establishes a direction-reward mapping matrix for parallel search.

[0074] The direction-benefit mapping matrix records each search direction and its corresponding quality improvement value. The system uses this matrix to calculate the expected benefit of different candidate search directions. For example, if there are 3 candidate search directions for a certain search starting point, with corresponding expected benefits of [2.1, 1.8, 3.2], the system will select the candidate direction with an expected benefit of 3.2 as the search direction for that search starting point.

[0075] During the search process, if the expected return of a search starting point is lower than the expected value of 0.5 for three consecutive iterations, the system will select the search direction with the highest expected return from the adjacent search starting points. For example, if the expected return of search starting point 1 is consistently lower than 0.5, while the expected returns of adjacent search starting points 2 and 3 are 1.2 and 0.8 respectively, the system will instruct search starting point 1 to adopt the search direction of search starting point 2 to improve search efficiency.

[0076] The search process continues until the maximum quality improvement value of all search starting points is less than the quality improvement threshold of 0.1, indicating that the search paths have converged. At this point, the system obtains the process parameter combinations corresponding to the current iteration position of all search starting points, compares their quality evaluation values, and selects the highest value as the optimal process parameter combination.

[0077] In a real-world application case, a manufacturing company initially had five typical combinations of process parameters with quality evaluation values ​​of [82.3, 79.5, 84.1, 78.6, 81.2]. After 20 rounds of iterative optimization, the quality evaluation values ​​of the optimal parameter combinations evolved from these search starting points improved to [86.7, 84.2, 89.5, 83.1, 85.8]. The system selected the process parameter combination [225, 62, 0.72] corresponding to a quality evaluation value of 89.5 as the final optimal process parameter combination. Compared to the original parameter combination [215, 58, 0.75], the quality evaluation value improved by 5.4, achieving a significant quality improvement.

[0078] This method makes full use of historical data and overcomes the problem of single-point search easily getting trapped in local optima through a multi-starting-point parallel search strategy, thereby improving the efficiency and effectiveness of process parameter optimization and making it suitable for process parameter optimization in complex industrial production environments.

[0079] In one optional implementation, the expected return of each candidate search direction is calculated based on the direction-return mapping matrix, and the candidate search direction with the highest expected return is selected as the search direction for each search starting point, including: Based on the direction-reward mapping matrix, the angle between adjacent search direction vectors is calculated and negative exponential operation is performed to obtain the direction similarity. The correlation coefficient between adjacent quality improvement values ​​is calculated. The direction similarity and the correlation coefficient are weighted and summed to obtain the correlation degree. For a candidate search direction, calculate the angle between it and each search direction vector in the direction-reward mapping matrix and perform a negative exponential operation to obtain the candidate direction similarity; multiply the candidate direction similarity by the sum of the corresponding quality improvement value and the correlation degree, and normalize and weight the calculation results of all search direction vectors to obtain the initial expected reward of the candidate search direction; Calculate the negative exponential product of the current time and the historical time difference as the time decay weight; adjust the initial expected return based on the time decay weight to obtain the expected return, and select the candidate search direction with the highest expected return as the search direction of each search starting point.

[0080] This embodiment provides a method for calculating the expected return of each candidate search direction based on a direction-return mapping matrix, and selecting the candidate search direction with the highest expected return as the search starting point for each search. The specific implementation steps of this method will be described in detail below.

[0081] In this embodiment, a direction-benefit mapping matrix is ​​obtained, which contains the search direction vectors recorded during the historical search process and their corresponding quality improvement values. For example, suppose that five search direction vectors D1(0.5, 0.5, 0.7), D2(0.6, 0.2, 0.8), D3(0.3, 0.7, 0.6), D4(0.8, 0.1, 0.6), and D5(0.4, 0.5, 0.7) were recorded during the historical search process, with corresponding quality improvement values ​​of Q1=0.25, Q2=0.18, Q3=0.22, Q4=0.30, and Q5=0.20, respectively.

[0082] Based on the aforementioned direction-reward mapping matrix, the angle between adjacent search direction vectors is calculated and a negative exponential operation is performed to obtain the direction similarity. Specifically, the angle between D1 and D2 is calculated to be approximately 25 degrees, and a negative exponential operation is performed to obtain the direction similarity S. 12 =0.78; the angle between D2 and D3 is calculated to be approximately 39 degrees, yielding the directional similarity S. 23 =0.65; the angle between D3 and D4 is calculated to be approximately 44 degrees, yielding the directional similarity S. 34=0.60; the angle between D4 and D5 is calculated to be approximately 32 degrees, yielding the directional similarity S. 45 =0.70.

[0083] Next, calculate the correlation coefficient between adjacent quality improvement values. For example, the correlation coefficient C between Q1 and Q2. 12 =0.85, the correlation coefficient C between Q2 and Q3 23 =0.78, the correlation coefficient C between Q3 and Q4 34 =0.82, the correlation coefficient C between Q4 and Q5 45 =0.79. The correlation coefficient is obtained by weighting and summing the directional similarity and correlation coefficient. The weight of directional similarity is set to 0.6, and the weight of the correlation coefficient is set to 0.4. Therefore, the correlation coefficient R between D1 and D2 is... 12 =0.6×0.78+0.4×0.85=0.81; The correlation R between D2 and D3 23 =0.6×0.65+0.4×0.78=0.70; The correlation R between D3 and D4 34 =0.6×0.60+0.4×0.82=0.69; The correlation R between D4 and D5 45 =0.6×0.70+0.4×0.79=0.74.

[0084] For a new candidate search direction D c (0.55, 0.35, 0.75), calculate the angle between this value and the historical search direction vector, and perform a negative exponential operation to obtain the candidate direction similarity. Calculate D c The angle between D1 and D1 is approximately 15 degrees. A negative exponential calculation is performed to obtain the candidate direction similarity S. c1 =0.87; Similarly, S is calculated to be... c2 =0.90, S c3 =0.72, S c4 =0.65, S c5 =0.80.

[0085] The similarity of the candidate directions is multiplied by the sum of the corresponding quality improvement value and the relevance. For D1, the relevance calculation result is 0.87×(0.25+0.81)=0.92; for D2, the relevance calculation result is 0.90×(0.18+0.70+0.81)=1.52; for D3, the relevance calculation result is 0.72×(0.22+0.70+0.69)=1.16; for D4, the relevance calculation result is 0.65×(0.30+0.69+0.74)=1.13; and for D5, the relevance calculation result is 0.80×(0.20+0.74)=0.75. The calculation results of all search direction vectors are normalized and weighted to obtain the candidate search direction D.c Initial expected return E c =(0.92+1.52+1.16+1.13+0.75) / 5=1.10.

[0086] Next, the negative exponential product of the current time and the historical time difference is calculated as the time decay weight. Assume the current time is T, and the recorded times for the historical search directions are T1, T2, T3, T4, and T5. The time difference between T and T1 is calculated to be 5 hours, and the negative exponential calculation yields a time decay weight W1 = 0.78; the time difference between T and T2 is calculated to be 4 hours, yielding a time decay weight W2 = 0.82; the time difference between T and T3 is calculated to be 2 hours, yielding a time decay weight W3 = 0.90; the time difference between T and T4 is calculated to be 8 hours, yielding a time decay weight W4 = 0.65; and the time difference between T and T5 is calculated to be 3 hours, yielding a time decay weight W5 = 0.86.

[0087] The initial expected return is adjusted based on the time decay weight to obtain the expected return. The adjusted expected return E' is then calculated. c =E c ×(W1+W2+W3+W4+W5) / 5=1.10×(0.78+0.82+0.90+0.65+0.86) / 5=1.10×0.80=0.88.

[0088] For other candidate search directions, such as D a (0.62, 0.45, 0.64) and D b (0.48, 0.52, 0.71), calculate its expected return following the steps above. Assume D a The expected return is 0.76, D b The expected return is 0.92. Because D b The expected return is the highest, so choose D. b The search direction that serves as the starting point for the search.

[0089] Considering the case of multiple search starting points, for each starting point, calculate the expected return of each candidate search direction according to the steps described above, and select the candidate search direction with the highest expected return as the search direction for that starting point. For example, for search starting point P1, candidate search direction D... b The highest expected return is 0.92, so choose D. b The search direction for P1; for the search starting point P2, the candidate search direction D c The highest expected return is 0.88, so choose D. c As the search direction for P2.

[0090] The above embodiments describe in detail a method for calculating the expected return of each candidate search direction based on a direction-return mapping matrix, and selecting the candidate search direction with the highest expected return as the starting point for each search. This method evaluates and selects candidate search directions by calculating parameters such as direction similarity, correlation coefficient, association degree, candidate direction similarity, initial expected return, time decay weight, and expected return, making the search process more efficient and accurate, and improving the overall search quality. This method can be applied to various scenarios requiring search optimization, such as path planning, resource scheduling, and parameter optimization.

[0091] The die-casting island process parameter optimization system based on machine learning in this embodiment of the invention includes: The first unit is used to acquire process parameters in the die casting process, divide the process parameters into multiple process stages according to the die casting process, calculate the energy loss value of each process stage according to the energy conversion relationship in the die casting process, calculate the entropy weight of each process stage according to product quality data, identify the target process stage that affects quality, and when the energy loss value of the target process stage is higher than the loss threshold, optimize the process parameters of that process stage to obtain locally optimized process parameters. The second unit is used to construct a historical parameter matrix from historical process parameters, perform eigenvalue decomposition on its covariance matrix, and construct a low-dimensional parameter space. The historical process parameters are mapped to the low-dimensional parameter space. Based on the locally optimized process parameters, the geodesic distance is calculated, and historical process parameter combinations with geodesic distances less than a distance threshold are grouped into the same category. The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination. The third unit is used to generate adjustment instructions for the process parameters of the die-casting island based on the optimal combination of process parameters, and to send the adjustment instructions to the die-casting island control system to adjust the process parameters of each piece of equipment in real time.

[0092] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0093] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0094] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 method for optimizing process parameters of die-casting islands based on machine learning, characterized in that, include: The process parameters in the die casting process are obtained, and the process parameters are divided into multiple process stages according to the die casting process. Based on the energy conversion relationship in the die casting process, the energy loss value of each process stage is calculated. Calculate the entropy weight of each process stage based on product quality data, identify the target process stage that affects quality, and optimize the process parameters of the target process stage when the energy loss value of the target process stage is higher than the loss threshold to obtain locally optimized process parameters. Historical process parameters are constructed into a historical parameter matrix, and eigenvalue decomposition is performed on its covariance matrix. A low-dimensional parameter space is constructed, and historical process parameters are mapped to the low-dimensional parameter space. Based on the locally optimized process parameters, geodesic distances are calculated, and historical process parameters with geodesic distances less than a distance threshold are grouped into the same category. The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination. Based on the optimal combination of process parameters, an adjustment command for the process parameters of the die-casting island is generated and sent to the die-casting island control system to adjust the process parameters of each piece of equipment in real time.

2. The method according to claim 1, characterized in that, Based on product quality data, the entropy weight of each process stage is calculated to identify the target process stage affecting quality. When the energy loss value of the target process stage exceeds the loss threshold, the process parameters of that process stage are optimized to obtain locally optimized process parameters, including: Collect multidimensional quality inspection index data of die-cast products, and normalize the multidimensional quality inspection index data to obtain normalized inspection data. The weight value is obtained by dividing the normalized detection data by the sum of all normalized data of the corresponding detection indicator. The information entropy of each detection indicator is obtained by dividing the sum of the products of the weight value and its natural logarithm by the sample size. The correlation coefficient of each process parameter to each detection index is calculated using the Spearman correlation coefficient. The correlation coefficients are weighted and summed to obtain the correlation degree between the process stage and each detection index. The correlation degree is added to the product of the corresponding information entropy to obtain the entropy weight of each process stage. The process stage with the largest entropy weight is selected as the target process stage. When the energy loss value of the target process stage is higher than the preset loss threshold, calculate the first ratio of the energy loss value to the preset loss threshold and the second ratio of the mean of each detection index to the standard value, and then sum the first ratio and the second ratio by weight to obtain the local optimization target value. Calculate the gradient value of the local optimization target value with respect to each process parameter, and optimize the process parameters of the target process stage based on the gradient value; when the deviation value of the local optimization target value between two adjacent iterations is less than the convergence threshold, record the current process parameter as the local optimization process parameter.

3. The method according to claim 1, characterized in that, Historical process parameters are constructed into a historical parameter matrix. Eigenvalue decomposition is performed on its covariance matrix, and a low-dimensional parameter space is constructed. The historical process parameters are mapped to this low-dimensional parameter space. Geodesic distances are calculated based on the locally optimized process parameters. Historical process parameters with geodesic distances less than a distance threshold are grouped into the same category, including: The historical process parameter combinations during the die casting process are obtained and constructed into a historical parameter matrix; the historical parameter matrix is ​​multiplied by its transpose and divided by the number of samples to obtain the covariance matrix; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and eigenvectors. The historical parameter matrix is ​​then multiplied by the eigenvectors whose eigenvalues ​​are greater than the eigenvalue threshold to obtain the projection values ​​of the historical process parameter combinations in each eigendirection. The projection values ​​are then multiplied by the contribution rates corresponding to each eigenvalue to obtain the mapping coordinates of the historical process parameters in the low-dimensional parameter space. The contribution rates are calculated by dividing the eigenvalues ​​by the sum of the eigenvalues. Calculate the partial derivatives of the local optimized process parameters with respect to the coordinates in the low-dimensional parameter space; multiply the partial derivatives by the randomly initialized weight parameters, sum them, and add them to the identity matrix to obtain the Riemann metric matrix; use the Riemann metric matrix to solve for the minimum path length between any two historical process parameters in the low-dimensional parameter space using the variational method, and use the minimum path length as the geodesic distance; Historical process parameters whose distance from the geodesic line is less than a preset distance threshold are grouped into the same category to obtain the historical process parameter classification results.

4. The method according to claim 3, characterized in that, Using the Riemann metric matrix, the optimal geodesic path for any two historical process parameters in the low-dimensional parameter space is solved by variational method. The path length of the optimal geodesic path is taken as the geodesic distance, including: For any two historical process parameters, the geodesic path is calculated using linear interpolation in a low-dimensional parameter space. The gradient norm of the Riemann metric matrix on the geodesic path is calculated. The gradient norm is then mapped exponentially to obtain the path weight. The path weight is multiplied by the square root of the derivative of the geodesic path under the Riemann metric matrix to obtain the path length. Calculate the variational derivative of the path length with respect to the geodesic path, and add the variational derivative to an exponentially decaying regularization term to obtain the path update direction; calculate the first-order and second-order derivative norms of the geodesic path, and divide the second-order derivative norm by the square of the first-order derivative norm to obtain the local curvature of the path; multiply the baseline iteration step size by the negative exponential function of the local curvature of the path to obtain the iteration step size; In each iteration, the relative change rate of length is obtained by dividing the difference in path length between two adjacent iterations by the path length of the previous iteration, and the difference in local curvature between two adjacent iterations is obtained by calculating the curvature difference value. When the relative change rate of length is less than the length stability threshold and the curvature difference value is less than the curvature stability threshold, the geodesic path of the current iteration is determined as the optimal geodesic path, and the path length of the optimal geodesic path is used as the geodesic distance between two historical process parameter points.

5. The method according to claim 1, characterized in that, The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination, including: The quality evaluation value is obtained by multiplying the data of multiple quality inspection indicators corresponding to the process parameters with the corresponding weight coefficients and summing them. In each category, the historical process parameter with the highest quality evaluation value is selected as the search starting point. The mutual information entropy between the search starting point and the quality evaluation value is calculated. For search starting points with mutual information entropy higher than the importance threshold, the range of their parameter values ​​is divided into grid cells of multiple times the baseline number to construct a local search space. For search starting points with mutual information entropy lower than the importance threshold, the range of their parameter values ​​is divided into grid cells of the baseline number to construct a global search space. Calculate the process parameter difference vector between the current iteration position and the previous iteration position of each search starting point, and divide the difference vector by its magnitude to obtain the search direction vector; calculate the quality evaluation value difference between adjacent iteration positions of each search starting point to obtain the quality improvement value; When the maximum quality improvement value of all search starting points is less than the quality improvement threshold, obtain the process parameter combination corresponding to the current iteration position of all search starting points, and select the process parameter combination with the highest quality evaluation value as the optimal process parameter combination.

6. The method according to claim 5, characterized in that, After obtaining the quality improvement value, the method further includes: Based on the search direction vector and quality improvement value of each search starting point, a direction-reward mapping matrix for parallel search is established; The expected return of each candidate search direction is calculated based on the direction-return mapping matrix, and the candidate search direction with the highest expected return is selected as the search direction of each search starting point. When the expected return of a certain search starting point is lower than the expected value after a preset number of consecutive searches, the search direction with the highest expected return is obtained from the adjacent search starting point.

7. The method according to claim 6, characterized in that, Based on the direction-reward mapping matrix, the expected reward of each candidate search direction is calculated, and the candidate search direction with the highest expected reward is selected as the search direction for each search starting point, including: Based on the direction-reward mapping matrix, the angle between adjacent search direction vectors is calculated and negative exponential operation is performed to obtain the direction similarity. The correlation coefficient between adjacent quality improvement values ​​is calculated. The direction similarity and the correlation coefficient are weighted and summed to obtain the correlation degree. For a candidate search direction, calculate the angle between it and each search direction vector in the direction-reward mapping matrix and perform a negative exponential operation to obtain the candidate direction similarity; multiply the candidate direction similarity by the sum of the corresponding quality improvement value and the correlation degree, and normalize and weight the calculation results of all search direction vectors to obtain the initial expected reward of the candidate search direction; Calculate the negative exponential product of the current time and the historical time difference as the time decay weight; adjust the initial expected return based on the time decay weight to obtain the expected return, and select the candidate search direction with the highest expected return as the search direction of each search starting point.

8. A machine learning-based die-casting island process parameter optimization system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to acquire process parameters during the die casting process, divide the process parameters into multiple process stages according to the die casting process, and calculate the energy loss value of each process stage based on the energy conversion relationship during the die casting process. Calculate the entropy weight of each process stage based on product quality data, identify the target process stage that affects quality, and optimize the process parameters of the target process stage when the energy loss value of the target process stage is higher than the loss threshold to obtain locally optimized process parameters. The second unit is used to construct a historical parameter matrix from historical process parameters, perform eigenvalue decomposition on its covariance matrix, construct a low-dimensional parameter space, map the historical process parameters to the low-dimensional parameter space, calculate the geodesic distance based on the locally optimized process parameters, and classify historical process parameters whose geodesic distance is less than the distance threshold into the same category. The historical process parameter combination with the highest product quality evaluation value in each category is marked as the search starting point. Parallel parameter search is performed on the search starting point to obtain candidate process parameter combinations. The candidate process parameter combination with the highest quality evaluation value is selected as the optimal process parameter combination. The third unit is used to generate adjustment instructions for the process parameters of the die-casting island based on the optimal combination of process parameters, and to send the adjustment instructions to the die-casting island control system to adjust the process parameters of each piece of equipment in real time.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to 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 the processor, they implement the method described in any one of claims 1 to 7.

Images