Extreme manufacturing process parameter optimization method and system fused with machine learning

Abstract

The application relates to the technical field of intelligent manufacturing, and discloses an extreme manufacturing process parameter optimization method and system fusing machine learning. The method comprises the following steps: obtaining multi-source data from a manufacturing equipment sensor, and fusing to generate a material state vector; inputting a pre-training model to obtain a material coefficient transition trend; judging whether the trend fluctuation amplitude exceeds a preset threshold value, and if yes, marking a key node and extracting a characteristic parameter; for the key node, a control algorithm is used to calculate a parameter adjustment amount according to the characteristic parameter and real-time data; based on the parameter adjustment amount, a control parameter is optimized, a control instruction sequence is generated and transmitted to an actuator; feedback data after adjustment is obtained, compared with the transition trend, and if the deviation exceeds the allowable range, the pre-training model is updated. Through the deep fusion of predictive monitoring and intelligent control, the present application realizes accurate optimization and self-adaptive control of process parameters, improves the stability of the extreme manufacturing process and the product quality, and reduces energy consumption and the defect rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing, and particularly relates to an extreme manufacturing process parameter optimization method and system fusing machine learning. BACKGROUND

[0002] In extreme manufacturing processes (such as material processing and forming under extreme conditions such as high temperature, high pressure, high stress, etc.), precise control of process parameters (such as heating rate, pressure gradient, etc.) plays a decisive role in product quality, production efficiency and cost control. With the development of manufacturing technology towards high precision and high reliability, the demand for process parameter optimization is increasingly urgent.

[0003] In the early days, the process parameter control of extreme manufacturing processes mainly relied on the experience of operators, and the stability of the process was maintained through manual observation and manual adjustment. However, this method is highly dependent on human experience and is difficult to cope with complex and variable working conditions, with low control precision and low efficiency.

[0004] With the development of sensor technology, data feedback-based control methods have gradually emerged. Through sensors deployed on manufacturing equipment, real-time data such as temperature and pressure are collected, and process parameters are adjusted based on pre-set rules or simple feedback control algorithms. This method has improved the automation level of control to some extent, but still has limitations in dealing with the nonlinear changes of material properties and the complex coupling relationship between multiple parameters in extreme manufacturing processes.

[0005] In recent years, the development of machine learning technology has brought new opportunities for process parameter optimization in extreme manufacturing processes. Through the learning and analysis of historical data by machine learning models, the trend of material state changes can be predicted to provide reference for the adjustment of process parameters. However, existing machine learning-based methods mostly focus on single prediction function, lack of precise identification of key nodes, effective fusion of multi-source data, and dynamic optimization ability of control parameters, making it difficult to achieve precise and adaptive control of process parameters in extreme manufacturing processes.

[0006] The present application aims to fuse machine learning technology to solve the problems of low control precision, poor adaptability, lack of dynamic optimization ability, etc. in existing process parameter control methods of extreme manufacturing processes, to realize precise optimization and adaptive control of process parameters, and to improve the quality and efficiency of extreme manufacturing processes. SUMMARY

[0007] The application provides an extreme manufacturing process parameter optimization method and system fused with machine learning, aiming to solve the problems of low control precision, poor adaptability and insufficient response to dynamic changes of material state in existing extreme manufacturing processes, which can respond to material state mutations in real time, realize forward-looking adaptive optimization of process parameters, and improve the efficiency and quality of extreme manufacturing processes.

[0008] In a first aspect, the application provides an extreme manufacturing process parameter optimization method fused with machine learning, which comprises:

[0009] Step 1: Obtain multi-source data from sensors deployed on manufacturing equipment, and fuse the multi-source data to generate a material state vector;

[0010] Step 2: Input the material state vector into a pre-trained model to obtain a material coefficient transition trend;

[0011] Step 3: Determine whether the fluctuation amplitude of the material coefficient transition trend exceeds a preset threshold, and if so, mark a key node and extract a feature parameter;

[0012] Step 4: For the marked key node, calculate a parameter adjustment amount according to the feature parameter and real-time data using a control algorithm;

[0013] Step 5: Optimize the control parameters according to the parameter adjustment amount, generate a control instruction sequence, and transmit it to an actuator;

[0014] Step 6: Obtain adjusted feedback data and compare it with the material coefficient transition trend, and update the pre-trained model based on the comparison result.

[0015] In a second aspect, the application provides an extreme manufacturing process parameter optimization system fused with machine learning, which comprises:

[0016] A data acquisition module for obtaining multi-source data from sensors deployed on manufacturing equipment, and fusing the multi-source data to generate a material state vector;

[0017] A trend analysis module for inputting the material state vector into a pre-trained model to obtain a material coefficient transition trend;

[0018] A fluctuation detection module for determining whether the fluctuation amplitude of the material coefficient transition trend exceeds a preset threshold, and if so, marking a key node and extracting a feature parameter;

[0019] A parameter calculation module for calculating a parameter adjustment amount according to the feature parameter and real-time data using a control algorithm for the marked key node;

[0020] An instruction generation module is configured to optimize the control parameter according to the parameter adjustment amount, generate a control instruction sequence, and transmit the control instruction sequence to the actuator.

[0021] A model updating module is configured to obtain the adjusted feedback data, compare the adjusted feedback data with the transition trend of the material coefficient, and update the pre-trained model based on a comparison result.

[0022] Compared with the prior art, the technical scheme of the present application has at least the following advantages:

[0023] 1. The transition trend of the material coefficient is predicted by the neural network model, and the key node of fluctuation overrun is accurately identified, so that the action time of the control system is changed from "after-the-fact remediation" to "prevention", which fundamentally solves the problem of control instruction lag, realizes the advanced perception and active intervention of the material state mutation, and effectively avoids the generation of irreversible defects such as cracks and deformation.

[0024] 2. The hybrid control algorithm fusing fuzzy reasoning and data fusion is adopted, multi-dimensional information such as trend change rate (slope) and real-time state (pressure) is comprehensively considered, the calculated parameter adjustment amount is more scientific and more suitable for the nonlinear characteristics of the process, and high-precision, strong-robust synchronous optimization control of the heating and pressure system is realized.

[0025] 3. Through the closed-loop optimization mechanism, the fluctuation of different material batches and process conditions can be dynamically adapted, and the key process parameters can be always stabilized in the best window, so that the product performance dispersion and the waste rate are greatly reduced, and the invalid energy consumption is reduced through optimization and regulation, and the product quality consistency, the overall production efficiency and the economic benefit are improved.

[0026] 4. The deviation between the real-time feedback data and the predicted trend is used for online updating of the model, so that the system can continuously learn the new characteristics of the manufacturing process and constantly evolve and optimize the strategy, thereby maintaining a high-precision control level for a long time, meeting the harsh requirements of long-period and high-stability production in extreme manufacturing, and enhancing the self-learning and self-adaptive ability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical scheme of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0028] Figure 1 The flowchart of the extreme manufacturing process parameter optimization method fusing machine learning of the present application;

[0029] Figure 2A schematic diagram of a fusion machine learning extreme manufacturing process parameter optimization system of the present application. DETAILED DESCRIPTION

[0030] The terms "first", "second", "third", "fourth" and the like in the description and in the claims of the present application, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of these terms herein is merely for distinguishing between the similar objects and the same can be referenced in the same way by other terms, unless otherwise specified. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items. Expressions such as "at least one of," when preceding the syllables of a list of elements, modify the entire list of elements and do not modify the elements individually. As used herein, the term "if" can be construed to mean "when" or "upon" or "in response to the" or "determined by" or "in response to the determination of” or "in response to the" or "in response to the." Similarly, the term "response" can be construed to mean "determined by" or "in response to the determination of” or "in response to the" or "in response to the." As used herein, the term "exemplary" is used in the sense of serving as an example, instance, or illustration. Any implementation of the application that is described is not necessarily to be construed as preferred or advantageous over other implementations. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and the present disclosure, and will not be interpreted in an overly literal or overly formal sense unless expressly so defined herein.

[0031] For the purpose of facilitating understanding, the specific flow of the embodiments of the present application is described as follows, Figure 1 The flowchart of the fusion machine learning extreme manufacturing process parameter optimization method provided by the present application is shown, which specifically includes the following steps:

[0032] Step 1: Obtain multi-source data from sensors deployed on manufacturing equipment, and generate material state vectors by fusing the multi-source data.

[0033] In a specific embodiment, the multi-source data includes temperature, pressure and stress, and the process of step 1 can specifically include the following steps:

[0034] (1) Integrate the temperature data, pressure data and stress data by using data fusion technology, and generate an initial material state vector containing elastic modulus and yield strength;

[0035] (2) Preprocess the initial material state vector, remove noise data, and obtain the material state vector.

[0036] Specifically, the multi-source data comes from sensors deployed on key parts of manufacturing equipment, such as the surface of a hot press or an extruder die. The temperature, pressure and stress are measured by thermocouples, piezoresistive sensors and strain gauges, respectively, and are collected in real time at a certain frequency (such as 10 times per second) to ensure the capture of transient changes in the manufacturing process.

[0037] The multi-source data is fused via a Kalman filtering algorithm which iteratively updates the state estimation according to the covariance matrix of each sensor data to generate a fused signal, thereby suppressing single sensor error and improving data robustness. The fusion technology can reduce the error caused by temperature fluctuations and improve the accuracy of the vector, thereby more accurately predicting material deformation. The fusion algorithm highlights the dominant effect of thermal stress by weighted allocation, effectively reducing the characterization deviation caused by thermal fluctuations. For example, in a high temperature environment, the fusion algorithm can set the temperature data weight to 0.6, the pressure to 0.3, and the stress to 0.1, ensuring that the vector reflects the material behavior dominated by thermal stress, based on which the robustness of the state vector can be improved and subsequent prediction deviation can be reduced.

[0038] Based on the fused signal, the elastic modulus is calculated using a deformation form of Hooke's law, i.e., the elastic modulus is equal to the stress divided by the strain, where the strain is derived from the stress data and the temperature compensation factor; at the same time, based on the von Mises yield criterion, the pressure and stress data are converted into equivalent stress, and when the equivalent stress reaches a specific threshold value of the material, it is marked as a yield point, so as to calculate the yield strength index. The calculated elastic modulus and yield strength index are combined into a vector form to form an initial material state vector containing the elastic modulus and the yield strength, which is in the form of a two-dimensional numerical pair [elastic modulus value, yield strength value]. When preprocessing the initial material state vector, median filtering is applied first to remove abnormal noise, and then each index is converted to a distribution with a mean of zero and a standard deviation of one through z-score standardization, and finally a standardized material state vector is obtained.

[0039] In a preferred embodiment, principal component analysis is performed on the material state vector, the first n principal components are extracted, and a feature vector is generated by combining them. The Euclidean norm of the feature vector is calculated and defined as the comprehensive strength index. The feature vector and the comprehensive strength index are combined to generate a feature expression vector.

[0040] Specifically, this technical solution is applied in continuous rolling manufacturing, for example, when processing steel, the feature expression vector can highlight the pattern of uneven stress distribution, which is beneficial for real-time adjustment of rolling force; in the ceramic sintering process, the extracted feature expression vector helps to identify the modulus change caused by temperature gradient, improving product consistency.

[0041] For example, assuming that the standardized vector is [1.2, 0.8] and the feature expression vector after principal component analysis is [0.9, 0.5], the norm of which is 1.0, this feature expression vector is used for model input, which can accurately reflect the transition of materials from elasticity to plasticity, bringing more accurate manufacturing control effect.

[0042] For high-precision scenarios such as aerospace material processing, feature representation can additionally incorporate time series analysis, calculating the rate of change of the vector as a supplementary feature, which expands the sensitivity to dynamic changes and helps to avoid material failure.

[0043] Step 2: Input the material state vector into the pre-trained model to obtain the transition trend of the material coefficient.

[0044] In a specific embodiment, the process of performing step 2 can specifically include the following steps:

[0045] Input the material state vector into the pre-trained neural network model, and the neural network model processes the nonlinear change characteristics of the material coefficient through an activation function, outputs the predicted value of the material coefficient at future time series, and generates the transition trend of the material coefficient.

[0046] Specifically, the transition trend of the material coefficient refers to the dynamic change law of the key coefficient (i.e., the material coefficient) of the material in a specific process (such as extreme manufacturing process, covering material processing and forming under extreme conditions such as high temperature, high pressure, and high stress) with the change of time, temperature, pressure, etc. Process parameters, this trend is expressed in the form of time series data or curve, which describes the evolution law of the coefficient with time. Common material coefficients include but are not limited to elastic modulus, yield strength, thermal expansion coefficient, thermal conductivity, Poisson's ratio, and creep rate. In the metal forming or composite material curing scene, the elastic modulus and yield strength are the core coefficients, as they directly determine the processing deformation and the strength of the final component.

[0047] The neural network model is a multilayer perceptron trained based on historical manufacturing data, which can capture the nonlinear response of the material during heating. The model uses activation functions such as ReLU to process the nonlinear change characteristics of the material coefficient, calculates the input vector through the hidden layer, and outputs the predicted value of the material coefficient at future time series, thereby forming the transition trend curve of the material coefficient, which represents the dynamic change of material performance in the manufacturing process. For example, in high-temperature alloy forging, the input vector reflects the current material state, and the model predicts the change curve of the elastic modulus in the next 10 seconds, such as the elastic modulus nonlinearly decreasing from 200 GPa to 180 GPa in the next 10 seconds, which can capture the performance degradation trend caused by pressure mutation. This prediction capability helps to identify material fatigue risks in advance and improve the stability of the manufacturing process.

[0048] Preferably, the neural network model is trained through a backpropagation algorithm, with mean square error as the loss function to optimize the weight parameters, ensuring the mapping ability for nonlinear relationships.

[0049] The neural network processes complex nonlinear patterns in the material state vector, making up for the lack of accuracy of traditional physical models under extreme conditions. This processing and prediction of nonlinear characteristics enable the model to accurately predict the dynamic changes of the material under complex manufacturing conditions, providing data support for subsequent process parameter optimization. In this way, the problem of accurately predicting the dynamic changes of the material state is solved, and the material state under extreme manufacturing conditions is accurately controlled, improving the processing stability and product quality, and significantly reducing the defect rate and energy consumption in the production process.

[0050] Step 3, determine whether the fluctuation amplitude of the material coefficient transition trend exceeds the preset threshold, if yes, mark the key node and extract the feature parameter.

[0051] In a specific embodiment, the process of performing step 3 can specifically include the following steps:

[0052] (1) Based on the preset time length, the material coefficient transition trend is segmented, the fluctuation amplitude of each segment is calculated, the points with fluctuation amplitude exceeding the preset threshold are identified, and are marked as key nodes;

[0053] (2) Linear fitting is performed on a segment of data centered on the key node in the material coefficient transition trend, and the transition slope of the segment of data is calculated, which is taken as the feature parameter corresponding to the key node;

[0054] (3) Generate feature description data of the key node based on the feature parameter and the fluctuation amplitude.

[0055] Specifically, the material coefficient transition trend curve in the future short period is divided into multiple data segments according to the preset time length (such as 5 seconds), and the fluctuation amplitude of each segment is calculated based on this. The fluctuation amplitude is obtained by calculating the standard deviation of the trend curve within the specified time window (i.e. the preset time length) or the difference between the peak value and the valley value. In this way, the degree of change of the material coefficient in each segment is quantified.

[0056] Identify points with fluctuation amplitude exceeding the preset threshold. These points exceeding the threshold mean that the material coefficient has undergone a more significant change in that period. Mark them as key nodes, record the key nodes in the data log based on the timestamp and amplitude information, ensure that the system can respond to potential material mutation points in time, and thus avoid manufacturing defects such as crack formation.

[0057] After confirming the key nodes, a segment of data centered on the key node is extracted and linearly fitted. Preferably, the least squares method or other methods are used to fit the trend data points, calculate the transition slope of the segment of data, and the transition slope reflects the dynamic change rate of the material coefficient near the key node, and reflects the speed of the material state change.

[0058] The transition slope is taken as a feature parameter corresponding to the key node, and the fluctuation amplitude calculated in the foregoing is combined to generate feature description data containing information such as the fluctuation amplitude and the transition slope. The feature description data is encapsulated in a structured data such as a JSON format, and is used as an input of a subsequent fuzzy logic controller.

[0059] In the metal forging scenario, the key node of the elastic modulus decrease caused by the sudden increase in pressure can be found in time through this technical means, and the transition slope and the fluctuation amplitude extracted can be used to calculate the heating rate adjustment amount, so as to avoid defects caused by over-stress of the material, and solve the problem that the timing and the change rate of the material state mutation are difficult to accurately grasp in the existing manufacturing process, thereby improving the stability of the manufacturing process and the product quality.

[0060] For example, in the semiconductor wafer manufacturing scenario, the transition trend involves thermal stress change. If the fluctuation amplitude is 10%, which exceeds the threshold of 8%, the key node is marked, and the further calculated slope is 2% / s. The generated feature description data is [2, 10], which is used to adjust the heating rate. This means can prevent wafer warping and improve yield. In the plastic injection molding scenario, it can be used to handle stress trends. If the fluctuation amplitude is 15 kPa, which exceeds the threshold of 12 kPa, the key node is marked, and the further calculated slope is 1.2 kPa / s. The generated feature description data is [1.2, 15], which is used to support subsequent fuzzy control to adjust the pressure, thereby ensuring product uniformity. It optimizes the production process through real-time feedback and reduces the defect rate.

[0061] By combining trend analysis, fluctuation amplitude detection, and slope extraction, the mutation point of the material state can be effectively captured, and timely and accurate adjustment basis can be provided for the control system to avoid material defects and production efficiency decline caused by control lag. Through this efficient feedback mechanism, the stability of the manufacturing process and the product quality can be significantly improved, and energy waste and material loss in production can be reduced. In the field of high-precision manufacturing, especially in aerospace or precision alloy processing, rapid response can be achieved and the adverse effects caused by material mutation can be effectively prevented, which has a significant technical advantage.

[0062] Step 4, for the marked key node, a control algorithm is used to calculate the parameter adjustment amount according to the feature parameters and real-time data.

[0063] In a specific embodiment, the real-time data includes pressure data, and the process of performing step 4 can specifically include the following steps:

[0064] (1) obtaining an input variable combination corresponding to the key node;

[0065] (2) fuzzifying the input variable combination and mapping it to a fuzzy set, and performing fuzzy reasoning on the fuzzy set based on a pre-set fuzzy rule base;

[0066] (3) Calculate the control output value based on the defuzzification method based on the fuzzy inference result, and calculate the heating rate adjustment amount based on the control output value;

[0067] (4) Weighted fusion of transition slope and pressure data to obtain the optimized rate correction value.

[0068] Specifically, the input variable combination refers to a series of data selected and combined for the calculation of the parameter adjustment amount for the marked key node. For each key node, relevant input variable combinations are extracted from real-time data, which usually include pressure data and material characteristic parameters (such as transition slope, elastic modulus and yield strength, etc.) extracted through trend analysis.

[0069] The input variable combination is processed by a fuzzy logic controller, which is an example of a control method based on fuzzy set theory. It can handle uncertainties and nonlinearities in the manufacturing process by fuzzifying input variables, applying fuzzy rules for inference, and finally defuzzifying the output. First, the characteristic parameters such as transition slope and current pressure data are converted into fuzzy variables, for example, the transition slope is divided into low, medium and high fuzzy sets, and the pressure data is divided into small, medium and large fuzzy sets. Fuzzified data can effectively handle the uncertainty and nonlinearity of material state changes in the manufacturing process. Then, inference is made based on the pre-set fuzzy rule base, which contains rules trained based on expert experience or historical data. These rules guide the fuzzy inference process, for example, if the transition slope is high and the pressure is large, the control algorithm will output a larger adjustment amount. The design of these rules ensures that the algorithm can make appropriate adjustments based on the current material state and real-time pressure data, ensuring that it can respond promptly when the material undergoes a sudden change. Finally, the result of fuzzy inference is converted into a specific numerical value (i.e. control output value) through defuzzification methods such as center average method.

[0070] Exemplarily, assume that the feature parameters extracted at this time are: a transition slope of 3 (reflecting the rate of change of the material coefficient), an elastic modulus index of 180 GPa, a yield strength index of 400 MPa, and a real-time collected pressure data of 12 MPa. The transition slope is divided into three fuzzy sets: low (0-2), medium (2-4), and high (4-6), and the current transition slope of 3 is mapped to the medium fuzzy set; the elastic modulus index is divided into three fuzzy sets: low (< 150 GPa), medium (150-200 GPa), and high (> 200 GPa), and 180 GPa is mapped to the medium fuzzy set; the yield strength index is divided into three fuzzy sets: low (< 300 MPa), medium (300-450 MPa), and high (> 450 MPa), and 400 MPa is mapped to the medium fuzzy set; and the pressure data is divided into three fuzzy sets: small (< 10 MPa), medium (10-15 MPa), and large (> 15 MPa), and 12 MPa is mapped to the medium fuzzy set.

[0071] The preset fuzzy rule base contains a series of "if-then" rules, such as "if the transition slope is medium and the pressure is medium, then the heating rate adjustment direction is appropriately reduced, and the adjustment amplitude is small". The input fuzzy sets are matched with the premise conditions of each rule in the rule base, and through matching rules, the conclusions corresponding to the rules that meet the premise conditions are extracted to form a comprehensive fuzzy conclusion set. For each element in the fuzzy conclusion set, such as "the heating rate is appropriately reduced and the adjustment amplitude is small", a membership function curve needs to be constructed. The membership function is used to describe the possibility (membership) of each heating rate adjustment degree value belonging to the fuzzy description, and the type (such as triangle, trapezoid, etc.) and parameters of the membership function are generally determined according to actual production experience, expert knowledge or historical data. Assuming that a triangular membership function is used to describe "the heating rate is appropriately reduced and the adjustment amplitude is small", and the heating rate adjustment ratio is taken as the abscissa and the membership is taken as the ordinate. According to experience, it is determined that the vertex (point of maximum membership) of the triangular membership function corresponds to a heating rate adjustment ratio of 0.9 times the original rate, and the left and right bottom points correspond to adjustment ratios of 0.8 times and 1.0 times the original rate, respectively. In this way, the membership variation of the fuzzy description under different heating rate adjustment degrees is determined, and the membership function curves of other fuzzy description elements are also determined in a similar manner. Subsequently, the centroids of each membership function curve in the fuzzy conclusion set are calculated, and the centroids are then weighted and averaged to obtain the control output value.

[0072] The control output value is multiplied by a preset scaling factor to calculate the heating rate adjustment amount. Exemplarily, when the result of fuzzy reasoning is 0.8, the control system can adjust the heating rate to 80% of the original rate, i.e., reduce the heating rate, to adapt to the change of the material state.

[0073] When fusing the transition slope and pressure data, different weights are assigned to the transition slope and pressure data, typically with a higher weight for the transition slope as it is more significant in reflecting the material state change. For example, the weight of the transition slope can be set to 0.6, while the weight of the pressure data is 0.4. Through weighted averaging, the fused rate correction value can comprehensively reflect the influence of both, ensuring more accurate adjustment. The rate correction value after weighted fusion will be used for subsequent parameter optimization, such as updating the pressure gradient, to ensure the matching of material coefficient change rules.

[0074] In practical applications, the fuzzy rule base can also be adapted according to material characteristics for different material types in manufacturing scenarios. For example, for aluminum alloys, the fuzzy rules may emphasize rapid adjustment of the heating rate under low pressure conditions to avoid material overheating and deformation; for steel, the rules may emphasize reducing the heating rate under high pressure conditions to prevent stress concentration in the material. This diversified control method improves the adaptability of the algorithm and can provide higher adjustment accuracy in the manufacturing process of different materials, reducing energy consumption and improving product quality.

[0075] By fusing the fuzzy logic control algorithm with material state characteristic data, not only the real-time response capability of the manufacturing process is enhanced, but also effective adjustment of manufacturing parameters in complex changing environments is ensured, greatly improving manufacturing precision and efficiency. Through meticulous adjustment and optimization, material defects can be effectively avoided, product quality can be improved, and energy can be saved.

[0076] In a preferred embodiment, a feature selection mechanism is introduced to dynamically adjust the input variables for fuzzy reasoning, including the following steps:

[0077] (1) Based on the importance of input variables analyzed by the neural network, the contribution of each input variable under the current material state is evaluated, and the weight of the input variable is adjusted according to the contribution;

[0078] (2) Principal component analysis is used to reduce the dimensionality of multi-dimensional input data, extracting the first K main components affecting material performance change, and generating input variable combinations based on the first K main components;

[0079] (3) Based on real-time sensor feedback signals, the real-time importance of each input variable under different process stages and material conditions is evaluated, and the input variable combination is dynamically adjusted based on the real-time importance to ensure the best match between the input variable and the material state.

[0080] Specifically, by introducing a feature selection mechanism, the input variables for fuzzy reasoning are dynamically adjusted, thereby optimizing the process parameters in the manufacturing process. The feature selection mechanism is to dynamically analyze the input variables through various data processing methods to ensure the best match between the input variables and the material state under different process stages and material conditions, thereby improving the accuracy and adaptability of the control process.

[0081] The neural network model learns the nonlinear relationship between material properties and process parameters through training data, and can evaluate the influence of each input variable on the current material state. Through the learning of the weights and biases of the neural network, the contribution of different input variables (such as temperature, pressure, elastic modulus, yield strength, etc.) in a specific state can be accurately reflected. Based on these contribution degrees, the system can dynamically adjust the weight of each input variable, so that important variables are given higher weights in the fuzzy reasoning process, so that the control system can more accurately reflect the current state of the material. Exemplarily, in a metal processing process, there are multiple input variables to describe the state of the material, such as temperature (T), pressure (P), stress (S), elastic modulus (E) and yield strength (YS), and the contribution of each input variable in the current material state is analyzed through a neural network, and the result is: temperature (T) 40%, pressure (P) 20%, stress (S) 15%, elastic modulus (E) 10%, and yield strength (YS) 15%. These contribution values indicate that temperature has the greatest impact on the material state, so temperature will be given a higher weight in model reasoning, while the elastic modulus may have less impact on the current material state, and the weight is lower.

[0082] Multi-dimensional data often contains redundant information, and PCA, as an effective dimensionality reduction method, transforms the original data into a few principal components through linear transformation, which can retain as much variance of the original data as possible. By extracting the top K principal components that affect the change of material performance, the system can eliminate the interference of irrelevant variables and focus on the most relevant features of the material state, thereby generating a more simplified and information-rich input feature combination, which helps to remove unnecessary noise, optimize the selection of input variables, and further improve the efficiency and accuracy of the model. Exemplarily, in a plastic injection molding process, the original input variables may include temperature, pressure, injection speed, etc. Through principal component analysis, temperature and pressure are extracted as the input variable combination for subsequent processing, thereby reducing the data dimension and improving the calculation efficiency.

[0083] On this basis, the system also combines real-time sensor feedback signals to evaluate the real-time importance of each input variable under different process stages and material conditions. During the manufacturing process, the properties of the material and the processing conditions may change dramatically, and the importance of the input variables fluctuates accordingly. Therefore, the system needs to monitor real-time feedback signals such as temperature, pressure, etc., and adjust the weights of the input features based on these signals. Through the real-time importance feedback mechanism, the system can dynamically adjust the input feature combination to ensure that the input variable combination at each moment can best match the current material state and processing conditions. In this way, the control system can better adapt to the dynamically changing environment and improve the real-time and accuracy of decision-making. For example, in the ceramic sintering process, as the sintering temperature rises, the influence of stress on material performance gradually increases. At this time, according to the real-time sensor feedback signal, the input variable combination is dynamically adjusted, the weight of stress is increased, and the weight of temperature is reduced, so as to ensure that the input variable combination can accurately reflect the current material state at all times.

[0084] For example, in the processing of high-temperature alloys, as the temperature rises, the yield strength and elastic modulus of the material may change significantly. At this time, the system can adjust the process parameters such as pressure and heating rate based on real-time feedback signals and the contribution of neural network analysis. In the low-temperature stage, more reliance may be placed on the stress and temperature data of the material, and less reliance may be placed on the changes in the elastic modulus. By evaluating the importance of variables in real time and dynamically adjusting the input feature combination, the system can ensure that the process parameters always meet the optimal processing requirements of the material, thereby reducing the defect rate and improving product quality.

[0085] By introducing the feature selection mechanism and adaptive input variable generation logic, it is ensured that under different materials and process conditions, the control algorithm can dynamically adjust the input variables according to real-time data and historical feedback, thereby optimizing the fuzzy reasoning process, solving the problem of insufficient adaptability of fixed input feature combinations in complex and dynamic manufacturing environments, and improving the precision control ability and stability of the manufacturing process, thereby effectively improving the production efficiency and product quality.

[0086] Step 5, optimizing the control parameters according to the parameter adjustment amount, generating a control instruction sequence, and transmitting it to the actuator.

[0087] In a specific embodiment, in step 5, optimizing the control parameters according to the parameter adjustment amount, generating a control instruction sequence, includes:

[0088] (1) updating the current pressure gradient parameter based on the rate correction value, and simultaneously, obtaining the optimized heating rate based on the heating rate adjustment amount;

[0089] (2) The updated pressure gradient parameter is iteratively optimized using a gradient descent algorithm to match the transition trend of the material coefficient, and an optimized pressure gradient parameter is obtained, wherein the gradient descent algorithm calculates the parameter deviation based on the loss function and updates the parameter value;

[0090] (3) According to the optimized heating rate and the optimized pressure gradient parameter, a control instruction sequence is generated to adjust the heating and pressure process parameters in the material manufacturing process.

[0091] Specifically, the rate correction value is multiplied by a preset weight factor and superimposed on the existing pressure gradient parameter to form a preliminary updated value. The product of the heating rate adjustment amount and the current heating rate is taken as the adjusted heating rate.

[0092] The gradient descent algorithm evaluates the parameter deviation by calculating the loss function, and updates the parameter value according to the gradient direction in each iteration. The loss function is calculated based on the deviation between the pressure gradient parameter and the change rule of the material coefficient, which comes from the pre-set material coefficient transition trend model. Through step-by-step iteration optimization, the gradient descent algorithm makes the pressure gradient parameter consistent with the actual change trend of the material, so as to ensure accurate control of the heating and applied pressure of the material during the manufacturing process.

[0093] In this optimization process, by calculating the loss function and updated parameter value after each iteration, an optimized pressure gradient parameter is finally obtained, which can more accurately reflect the dynamic changes of the material coefficient and adapt to the changing manufacturing environment. For example, in the process of metal heat treatment, the pressure gradient parameter optimized by the gradient descent algorithm can better match the changes of the yield strength and elastic modulus of the material, reducing material defects caused by improper pressure control. In plastic extrusion manufacturing, the optimized pressure gradient parameter can effectively cope with the fluctuation of the yield strength of the material caused by temperature changes, ensuring the consistency and quality of the product.

[0094] Based on the optimized heating rate and pressure gradient parameter, the control system generates a corresponding control instruction sequence, which is used to adjust the heating and pressure parameters in the subsequent manufacturing process. The control instruction sequence is transmitted to the actuator through the communication interface, and the corresponding adjustment action is performed in the actuator, such as adjusting the heating rate and applying the pressure gradient, to ensure that the state of the material is always maintained within the optimal range during the production process.

[0095] Through the combination of gradient descent algorithm and fuzzy control system, dynamic optimization and precise control of process parameters can be achieved, effectively avoiding material defects, improving product quality, and reducing energy consumption and production cost. It can respond and adjust the process parameters in real time during the material processing process, ensuring the accuracy and consistency of each link, and significantly improving the stability and reliability of production.

[0096] In a specific embodiment, in step 5, the control instruction sequence is sent to the actuator to adjust the manufacturing process, including:

[0097] (1) The control instruction sequence is formatted into instruction data recognizable by the actuator and transmitted to the actuator through a communication interface;

[0098] (2) The actuator applies the instruction data in real time to adjust the heating rate and pressure gradient;

[0099] (3) Monitor the running state of the actuator to generate an adjustment log.

[0100] Specifically, the control instruction sequence contains numerical instructions such as the heating rate per second temperature value and the pressure gradient per unit time change value, which are converted into a standard protocol format, such as binary encoding, to ensure that the actuator can directly parse it, avoid transmission errors, improve data compatibility, and ensure accurate execution of instructions. Use wired or wireless interfaces such as RS-485 or Ethernet to send the formatted data packet to the actuator to ensure real-time performance with a transmission delay of less than 50 milliseconds to maintain the continuity of the manufacturing process.

[0101] After the actuator system receives the instruction data, it parses the heating rate and pressure gradient parts, activates the heating element to gradually raise the temperature to the target value, while monitoring the current temperature to match the predicted material coefficient transition trend, adjusts the pressure application device to ensure synchronization with the yield strength index of the material state vector, avoids material mutations, and verifies the adjustment effect in real time. Through the built-in sensor feedback of the current heating and pressure value, compare it with the instruction, if the deviation exceeds 2%, then fine-tune the execution parameters.

[0102] Through formatting, transmission, and real-time adjustment, the actuator can accurately execute the control instruction, optimize the adjustment of the heating rate and pressure gradient, while monitoring the running state to generate a log, forming a closed-loop optimization, and improving the overall prediction accuracy and stability of the manufacturing process. product quality.

[0103] Step 6, obtain the feedback data after adjustment and compare it with the material coefficient transition trend, and update the pre-trained model based on the comparison result.

[0104] In a specific embodiment, the process of performing step 6 can specifically include the following steps:

[0105] (1) Obtain feedback data including temperature, pressure and stress from the sensor;

[0106] (2) Compare the feedback data with the material coefficient transition trend and calculate the deviation value;

[0107] (3) Determine whether the deviation value exceeds the preset allowable range, if yes, generate a model update signal;

[0108] (4) Adjust the weight parameters of the pre-trained model according to the model update signal to generate an updated pre-trained model.

[0109] Specifically, the feedback data includes but is not limited to temperature, pressure, stress, crystal structure information, part size accuracy, shape error, and surface roughness. The adjusted feedback data is obtained in real time from the sensor array deployed on the manufacturing equipment, which reflects the actual material state after executing the control instruction sequence, ensuring the timeliness of the feedback data. The obtained feedback data is compared with the predicted transition trend of the material coefficient of the neural network model, and the prediction error is quantified by calculating the deviation at the corresponding time point. The deviation value can be calculated using the root mean square error method to ensure the accuracy of the quantitative comparison. If the calculated deviation value exceeds the preset allowable range, a model update signal is automatically generated, which is used to trigger subsequent adjustments to avoid unnecessary computational overhead.

[0110] According to the model update signal, the weight parameters of the pre-trained model are adjusted using the backpropagation algorithm. The feature vectors are extracted from the feedback data and input into the backpropagation algorithm. The gradient is calculated and the weight parameters of the neural network are adjusted. Iterative adjustment is performed until the deviation value is within the allowable range, generating an updated pre-trained model.

[0111] Through dynamic updating of the pre-trained model, the prediction accuracy of the model for material state changes is improved. The model parameters can be automatically adjusted according to different manufacturing scenarios and material states, thereby achieving precise control of the material state in extreme manufacturing processes. The problem of inaccurate material state prediction by the pre-trained model and the inability to dynamically adjust according to real-time feedback is solved, precise control of the material state in extreme manufacturing processes is achieved, processing stability and product quality are improved, and defect rate and energy consumption in the production process are significantly reduced.

[0112] The extreme manufacturing process parameter optimization method incorporating machine learning in the embodiments of the present application is described above, and the extreme manufacturing process parameter optimization system incorporating machine learning in the embodiments of the present application is described below. Please refer to Figure 2 The extreme manufacturing process parameter optimization system incorporating machine learning provided by the present application has a structure diagram as shown in the figure, and the system comprises:

[0113] The data acquisition module 10 is used to obtain multi-source data from sensors deployed on the manufacturing equipment, and to generate a material state vector by fusing the multi-source data.

[0114] The trend analysis module 20 is used to input the material state vector into the pre-trained model to obtain the transition trend of the material coefficient.

[0115] The fluctuation detection module 30 is configured to determine whether the fluctuation amplitude of the material coefficient transition trend exceeds a preset threshold value, and if yes, mark a key node and extract a characteristic parameter.

[0116] The parameter calculation module 40 is configured to calculate a parameter adjustment amount by using a control algorithm according to the characteristic parameter and real-time data for the marked key node.

[0117] The instruction generation module 50 is configured to optimize the control parameter according to the parameter adjustment amount, generate a control instruction sequence, and transmit the control instruction sequence to an actuator.

[0118] The model updating module 60 is configured to obtain adjusted feedback data, compare the adjusted feedback data with the material coefficient transition trend, and update the pre-trained model based on a comparison result.

[0119] The above-described embodiments are merely used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for optimizing extreme manufacturing process parameters by integrating machine learning, characterized in that, The process includes: S1, acquiring multi-source data from sensors deployed on manufacturing equipment and fusing the multi-source data to generate a material state vector; S2, inputting the material state vector into a pre-trained model to obtain the material coefficient transition trend, which refers to the dynamic change pattern of key material coefficients during the process as time, temperature, and pressure process parameters change; and S3, determining whether the fluctuation amplitude of the material coefficient transition trend exceeds a preset threshold. If so, key nodes are marked and feature parameters are extracted. S4. For the marked key nodes, calculate the parameter adjustment amount using the control algorithm based on the feature parameters and real-time data; S5. Optimize the control parameters based on the parameter adjustment amount, generate a control command sequence, and transmit it to the actuator; S6. Obtain the adjusted feedback data and compare it with the transition trend of the material coefficient, and update the pre-trained model based on the comparison results. S3 includes segmenting the transition trend of the material coefficient based on a preset duration, calculating the fluctuation amplitude of each segment, which is obtained by calculating the standard deviation of the trend curve within a specified time window, identifying points where the fluctuation amplitude exceeds a preset threshold and marking them as key nodes; performing linear fitting on a segment of data centered on the key nodes in the transition trend of the material coefficient, calculating the transition slope of the segment of data, and using the transition slope as the feature parameter corresponding to the key node. Generate feature description data for key nodes based on feature parameters and fluctuation amplitude; S4 includes obtaining the input variable combination corresponding to the key node; fuzzifying the input variable combination and mapping it to a fuzzy set; and performing fuzzy inference on the fuzzy set based on a preset fuzzy rule base. The control output value is calculated based on the fuzzy inference results using a defuzzification method, and the heating rate adjustment is calculated based on the control output value. The transition slope and pressure data are weighted and fused to obtain an optimized rate correction value; Furthermore, a feature selection mechanism is introduced to dynamically adjust the input variables used for fuzzy inference. This includes analyzing the importance of input variables based on neural networks, evaluating the contribution of each input variable in the current material state, and adjusting the weights of the input variables according to their contribution; using principal component analysis to reduce the dimensionality of multidimensional input data, extracting the top K principal components affecting changes in material properties, and generating input variable combinations based on the top K principal components; and evaluating the real-time importance of each input variable under different process stages and material conditions based on real-time sensor feedback signals, and dynamically adjusting the input variable combinations based on real-time importance to ensure the best match between input variables and material state. The multi-source data includes temperature, pressure, and stress. S1 includes integrating temperature, pressure, and stress data using data fusion technology to generate an initial material state vector containing elastic modulus and yield strength; the initial material state vector is preprocessed to remove noisy data and obtain the material state vector. S2 includes inputting the material state vector into a pre-trained neural network model. The neural network model processes the nonlinear variation characteristics of the material coefficients through activation functions, outputs predicted material coefficient values ​​in future time series, and generates material coefficient transition trends.

2. The method as described in claim 1, characterized in that, Real-time data includes stress data.

3. The method as described in claim 2, characterized in that, In S5, the control parameters are optimized based on the parameter adjustment amount, generating a control command sequence, including: updating the current pressure gradient parameters based on the rate correction value, and simultaneously obtaining the optimized heating rate based on the heating rate adjustment amount; iteratively optimizing the updated pressure gradient parameters using a gradient descent algorithm to match them with the transition trend of the material coefficients, obtaining the optimized pressure gradient parameters, wherein the gradient descent algorithm calculates the parameter deviation and updates the parameter values ​​based on the loss function; and generating a control command sequence based on the optimized heating rate and optimized pressure gradient parameters to adjust the heating and pressure process parameters in the material manufacturing process.

4. The method as described in claim 3, characterized in that, In S5, the control command sequence is sent to the actuator to regulate the manufacturing process, including: formatting the control command sequence into instruction data that the actuator can recognize, and transmitting it to the actuator through the communication interface; the actuator applies the instruction data in real time to adjust the heating rate and pressure gradient; monitoring the operating status of the actuator and generating an adjustment execution log.

5. The method as described in claim 1, characterized in that, S6 include: Obtain feedback data, including temperature, pressure, and stress, from sensors; The feedback data is compared with the transition trend of the material coefficients to calculate the deviation value; it is determined whether the deviation value exceeds the preset allowable range. If so, a model update signal is generated; the weight parameters of the pre-trained model are adjusted according to the model update signal to generate the updated pre-trained model.

6. An extreme manufacturing process parameter optimization system integrating machine learning, used to implement the method as described in any one of claims 1 to 5, characterized in that, include: The data acquisition module is used to acquire multi-source data from sensors deployed on manufacturing equipment and fuse the multi-source data to generate a material state vector; The trend analysis module is used to obtain the transition trend of material coefficients based on the material state vector input to the pre-trained model. The fluctuation detection module is used to determine whether the fluctuation amplitude of the transition trend of the material coefficient exceeds the preset threshold. If so, the key nodes are marked and the feature parameters are extracted. The parameter calculation module is used to calculate the parameter adjustment amount for key nodes marked, based on feature parameters and real-time data, using a control algorithm. The instruction generation module is used to optimize the control parameters based on the parameter adjustment amount, generate a sequence of control instructions, and transmit them to the actuator. The model update module is used to obtain the adjusted feedback data and compare it with the transition trend of the material coefficients, and update the pre-trained model based on the comparison results.

Images