A large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components

Abstract

The present invention relates to a large-model-driven adaptive closed-loop optimization method for machining process parameters of complex aviation thin-walled components, comprising the following steps: collecting multi-source heterogeneous machining data and quality inspection data of complex aviation components; constructing and training a machining quality prediction model and a multi-objective process parameter optimization model for complex aviation components; encapsulating the trained model as a microservice and deploying it to the DeepSeek large-model container platform; using the DeepSeek large model as a controller to implement closed-loop calls to the microservice; forming an online adaptive optimization process; automatically generating an analysis report based on the optimization results; and optimizing the machining quality prediction model using new data increments and designing an automatic rollback strategy. The present invention integrates deep learning and large-model technologies to achieve accurate prediction of machining quality and efficient regulation of process parameters, while effectively improving the manufacturing efficiency and quality of complex aviation components.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a large-model driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components. Background Art

[0002] As critical components like aerospace engine blades face higher demands for machining quality, traditional experience-driven or static optimization methods struggle to address these challenges in multi-objective coupled scenarios. During machining, various process parameters (such as feed rate and cutting speed) couple with each other under varying operating conditions and material properties, exhibiting complex, nonlinear dynamic changes. Existing methods lack intelligent perception and rapid response to these coupling relationships and real-time quality feedback, resulting in insufficient precision and stability in multi-objective optimization.

[0003] Furthermore, complex aviation components often face the challenge of producing a wide variety of components in small batches. Traditional methods are unable to effectively handle high-dimensional, heterogeneous data and the generalization issues inherent in sparse samples. This necessitates the introduction of large-scale, pre-trained deep learning models that learn latent features from massive historical and online real-time data, enabling multimodal information fusion and intelligent process prediction. This can overcome the bottlenecks in current process optimization, such as responsiveness and adaptability across multiple scenarios. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a large-model-driven adaptive closed-loop optimization method for the processing parameters of complex thin-walled aviation components, which solves the problems of insufficient multi-index optimization accuracy and stability of traditional methods and insufficient generalization ability to effectively handle high-dimensional heterogeneous data and sparse samples. This method significantly improves the multi-index optimization accuracy and stability by deeply integrating large-scale pre-training models with process control algorithms, and can respond quickly to disturbances, meeting the needs of intelligent processing of high-end parts.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components, comprising the following steps:

[0006] S1. Collect multi-source heterogeneous processing data and quality inspection data of complex aviation components;

[0007] S2. Based on the collected multi-source heterogeneous processing data and quality inspection data, a quality prediction model for complex aviation components is constructed and trained. A multi-objective process parameter optimization model is then constructed and trained to obtain a parameter optimization strategy.

[0008] S3. Encapsulate the two trained models as a quality prediction microservice and a process optimization microservice, respectively, publish them to a container platform based on the DeepSeek large model, and configure operation and maintenance policies.

[0009] S4, using the DeepSeek large model as a controller to implement closed-loop calls to microservices;

[0010] S5: Call the quality prediction microservice to calculate the error ratio of the target and generate an adaptive weight vector. Then call the process optimization microservice to generate action increments and apply them to the current environment state. This process iterates until all indicators converge, forming a multi-objective online closed-loop adaptive optimization process.

[0011] S6. Automatically write an analysis report based on the converged optimal process parameter feature vector and its corresponding predicted quality feature vector;

[0012] S7. Collect new processing data of a small batch of real workpieces, perform incremental fine-tuning and hyperparameter optimization on the previously trained processing quality prediction model, and design an automatic rollback strategy.

[0013] Furthermore, in step S1, the specific process includes the following steps:

[0014] S11. Deploy a unified data access module on the machine tool side, sensor side, and detection equipment side to collect process parameters, timing signals, and quality detection data, and push the data to the data aggregation layer and store it in the database in a standardized message format via industrial Ethernet or wireless edge gateway;

[0015] S12. Perform unified preprocessing on the stored data, accurately align the timing signals of different device sampling frequencies by timestamps and fill the sampling gaps, repair missing values, apply normalization or standardization to all numerical features and filter outliers to generate a data set with consistent distribution.

[0016] Furthermore, in step S2, the specific process includes the following steps:

[0017] S21. Based on the collected multi-source heterogeneous processing data and quality inspection data, a process parameter feature matrix and a corresponding quality inspection label matrix are constructed. A multi-layer perceptron and a multi-head attention mechanism are used to build a processing quality prediction model. The model training aims to minimize the mean square error loss and obtain the optimal weight of the training model.

[0018] S22. Use the trained processing quality prediction model as the reward function in the reinforcement learning environment. Define the environment state as the current process parameter feature vector and its predicted quality feature vector. Define the action increment as the adjustment increment of the process parameters. Use the multi-agent SAC reinforcement learning algorithm improved by the attention mechanism to maximize the expected reward. Generate and train a multi-objective process parameter optimization model to obtain the parameter optimization strategy.

[0019] S23. Export and serialize the trained processing quality prediction model and multi-objective process parameter optimization model for subsequent deployment and calling.

[0020] Furthermore, in step S3, the specific process includes the following steps:

[0021] S31. Define input and output interface call libraries for the quality prediction microservice and process optimization microservice;

[0022] S32. Package the above two microservices into a lightweight image. The image construction adopts a multi-stage construction strategy, and the model files, normalization parameters, and DeepSeek interface metadata are uniformly placed in the image read-only layer. The health check and readiness probes are pre-installed in the image.

[0023] S33. Publish the image to a container platform based on the DeepSeek large model, create deployment units based on resource quotas, allocate routes and domain names, configure circuit breakers, rate limiters, and retry policies at the call level, and set real-time warnings for key indicators at the monitoring level to ensure that exceptions are captured in a timely manner during closed-loop calls.

[0024] Furthermore, in step S4, the specific process includes the following steps:

[0025] S41. Configure the operating environment of the DeepSeek large model, deploy the model weights and vocabulary files, and lock the version;

[0026] S42. Use DeepSeek's input and output interface definition specifications to describe the call contracts of each microservice, and map the quality prediction microservice and process optimization microservice into formal functions respectively:

[0027] ;

[0028] in, Input process parameter feature vector , output prediction quality feature vector , Input environment status , output action increment In the input and output interface definition documents, declare the type, unit, and value range of each field using JSON Schema or IDL, and introduce version numbers, interface signature hashes, and timeout thresholds to ensure the stability and traceability of interface calls during the inference process;

[0029] S43, design prompt words and dialogue context, the context is determined by system instructions , Historical Dialogue and current user intent The total length does not exceed the model context window ,Right now:

[0030] ;

[0031] In the prompt template, the interface name and parameters are dynamically filled in through the placeholders {SERVICE} and {INPUT}, and the model is explicitly required to be output in the format of "calling DeepSeek microservice";

[0032] S44. Construct a reasoning process based on chain thinking, and let the context state be , the tool set is , then in Step 1 requires calculation for each tool The calling probability is:

[0033] ;

[0034] in, Calling for the model The original score, is the temperature coefficient, then select Generates RPC format call instructions based on the prompt words, and returns the results after execution Append to context: , loop until convergence is met or the predefined step limit is reached, forming a closed-loop collaborative call to the microservices by the DeepSeek large model as the controller.

[0035] Furthermore, in step S5, the specific process includes the following steps:

[0036] S51. Call the quality prediction microservice and input the process parameter feature vector Returns the predicted quality feature vector , and according to the preset target mass vector Compute the multi-objective error vector:

[0037] ;

[0038] in Indicates the The prediction error of the target, represents the total number of targets, is the number of iterations, and then calculate the proportion of each error in all errors:

[0039] ;

[0040] Will As the adaptive weight component of the target, the adaptive weight vector ;

[0041] S52, the current environment state and adaptive weight vector Also passed to the process optimization microservice, its internal parameter optimization strategy Generate action increments based on a multi-objective weighted reward function , the multi-objective weighted reward function is defined as:

[0042] ;

[0043] S53, increment the obtained action Applies to the current environment state , and obtain the new process parameter characteristic vector , and call the quality prediction microservice again to calculate the new prediction quality feature vector , providing accurate feedback for the next round of weight update;

[0044] S54, according to the pre-set convergence threshold vector , when all new prediction errors All satisfied When the multi-objective optimization is considered to have converged, the system outputs the final process parameter feature vector and the final prediction quality feature vector If any indicator exceeds the threshold, the environment state of the next cycle is set to , and and Re-enter the closed-loop iteration of steps S51 to S54 until all indicators meet the convergence criteria.

[0045] Furthermore, in step S6, the specific process includes the following steps:

[0046] S61, performing structured aggregation on the optimal process parameter feature vector obtained by final convergence and its corresponding predicted quality feature vector to obtain a structured horizontal table;

[0047] S62. Automatically generate visual charts around key indicators in the iteration process, including error convergence curves, weight evolution curves, and parameter performance comparison tables;

[0048] S63. Automatically write an analysis report based on the preset report template. The report is divided into four parts: "Optimization Parameter Summary", "Convergence Analysis", "Weight Evolution", and "Performance Comparison". The table of S61 and all the charts of S62 are embedded in the corresponding chapters in sequence, and the experimental configuration and timestamp are attached at the beginning of the report.

[0049] Furthermore, in step S7, the specific process includes the following steps:

[0050] S71, online collection of new processing data of small batches of real workpieces and execution of the same preprocessing process as S12;

[0051] S72. Incrementally fine-tune the previously trained processing quality prediction model by minimizing the synthetic loss function:

[0052] ;

[0053] in, is the mean square error loss of the original training set, is the loss of the new sample, and is the weight coefficient controlling the retention of old knowledge and the learning of new knowledge, is the total number of samples collected, and the final updated parameter to ;

[0054] S73. Calculate the average error on the newly added samples and compare it with the historical error threshold to dynamically adjust the model's hyperparameters.

[0055] S74. Generate a unique version identifier for each incremental fine-tuning and hyperparameter adjustment , and the corresponding model weights and strategy parameters Register to the model management system while retaining the previous version In preparation for rollback, when the maximum error of the online version in actual small-batch testing exceeds the set threshold, the system will trigger an automatic rollback to the previous stable version.

[0056] By means of the above technical solution, the present invention provides a large-model driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components, which has at least the following beneficial effects:

[0057] (1) This invention achieves intelligent perception, accurate prediction, and dynamic optimization of complex process parameters and multi-dimensional quality indicators by deeply integrating a large-scale pre-trained deep learning model with an online closed-loop adaptive microservice framework, thus overcoming the limitations of traditional experience-driven or static optimization methods in multi-objective coupling scenarios;

[0058] (2) By constructing a large model controller and generating parameter strategies for complex component processing processes, the present invention significantly improves the accuracy and robustness of multi-index optimization and enhances the generalization ability in high-dimensional heterogeneous data and small sample scenarios;

[0059] (3) The present invention constructs microservices to evaluate quality indicators such as roughness, residual stress and deformation error in real time, and adaptively adjusts the multi-objective optimization weights based on feedback, thereby solving the problem of insufficient multi-indicator optimization accuracy and stability in existing methods when various process parameters in the machining process are coupled. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0061] Figure 1 This is a flow chart of a large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to the present invention;

[0062] Figure 2 This is a framework diagram of a large-model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to the present invention. DETAILED DESCRIPTION

[0063] To make the above-mentioned objectives, features, and advantages of the present invention more clearly understood, the present invention is further described below in detail with reference to the accompanying drawings and specific embodiments. This will enable a full understanding of how this application uses technical means to solve technical problems and achieve technical effects, and to implement the invention accordingly.

[0064] Those skilled in the art will appreciate that all or part of the steps in the above-described embodiment methods can be accomplished by programming related hardware. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] Please refer to Figure 1-Figure 2 , shows a specific implementation method of this embodiment. This embodiment constructs and trains a complex aviation component processing quality prediction model and a multi-objective process parameter optimization model, which are then encapsulated as microservices and deployed to the DeepSeek container platform and a control environment is built. The DeepSeek large model is used as the controller to implement microservice closed-loop calls, form an online adaptive optimization process, and automatically generate analysis reports based on the optimization results. The model is optimized using new data increments, and an automatic rollback strategy is designed to achieve accurate prediction of processing quality and efficient control of process parameters, while effectively improving the manufacturing efficiency and quality of complex aviation components.

[0066] Please refer to Figure 1 This embodiment proposes a bearing fault diagnosis method based on a dual domain adaptive neural network, which includes the following steps:

[0067] S1. Collect multi-source heterogeneous processing data and quality inspection data of complex aviation components;

[0068] As a preferred implementation of step S1, the specific process includes the following steps:

[0069] S11. First, deploy a unified data access module on the machine tool, sensor, and inspection equipment sides. This module pushes process parameters from the CNC control system (such as feed per tooth, cutting speed, cutting depth, and cutting width), real-time collected tool load, vibration, temperature, and other timing signals, as well as quality inspection data from coordinate measuring machines (CMMs), roughness meters, and stress testing devices, to the data aggregation layer via industrial Ethernet or wireless edge gateways in a standardized message format (such as MQTT and OPC UA) and stores them in the database. During this process, interface adaptation is required for various data sources to ensure millisecond-level clock synchronization accuracy. Metadata tags are also introduced into the data pipeline to facilitate subsequent traceability and alignment.

[0070] S12. Perform unified preprocessing on the collected multi-source heterogeneous data, including: first, accurately align the time series signals of different device sampling frequencies according to the timestamp, and fill the sampling gaps through linear or high-order interpolation; then, for different complex components, such as integral blades, thin-walled casings, etc., according to the process stage or tool path, slice the original time series data into feature segments of fixed length, and extract the statistics (such as mean, variance, and frequency domain energy) within each segment; then, for missing values ​​caused by occasional transmission packet loss or sensor failure, use interpolation or model prediction methods based on adjacent time windows to repair them; finally, apply normalization or standardization to all numerical features, and filter outliers based on upper and lower limits and box plots to ensure that the data input used for subsequent model training is consistent in distribution, dimensionless, clean and reliable.

[0071] S2. Based on the collected multi-source heterogeneous processing data and quality inspection data, a quality prediction model for complex aviation components is constructed and trained. A multi-objective process parameter optimization model is then constructed and trained to obtain a parameter optimization strategy.

[0072] As a preferred implementation of step S2, the specific process includes the following steps:

[0073] S21, based on the multi-source heterogeneous processing data and quality inspection data pre-processed in S1, establish the process parameter feature matrix , and the corresponding quality detection label matrix , and then build a processing quality prediction model based on deep learning ,in represents the sample size of training data, and Represent the dimensions of the input process parameter feature matrix and the quality inspection label matrix respectively, Represents the learnable parameters of the model; the processing quality prediction model first further normalizes the input features, then uses a multi-layer perceptron (MLP) combined with a multi-head attention mechanism as the backbone structure, and finally connects to a fully connected layer to output the predicted value. The model training goal is to minimize the mean square error loss:

[0074] ;

[0075] in represents the loss function, Indicates the sample number index. The Adam algorithm with weight decay is used in the optimization process. The learning rate and decay coefficient are determined by the grid search method, and indicators such as RMSE are monitored on the validation set. When the indicators no longer improve for several consecutive rounds, early stopping is triggered to finally obtain the optimal weight. , ensuring that the model can accurately fit the historical data and has good generalization ability, and obtain a well-trained processing quality prediction model ;

[0076] S22, the trained processing quality prediction model As a reward function in the reinforcement learning environment, it is used to build a multi-objective process parameter optimization model and define Environmental status at any moment , represents the current process parameter feature vector and its predicted quality feature vector, action increment Get instant returns for each step of incremental adjustment of process parameters :

[0077] ;

[0078] in, The target quality vector for quality optimization, the negative sign means the smaller the prediction error, the greater the reward, and then the discount factor calculate Cumulative returns at each moment , measure from the moment All future moments The weighted sum of the returns of the multi-agent SAC reinforcement learning algorithm using the improved attention mechanism maximizes the expected return :

[0079] ;

[0080] in Represents the cumulative return at time 0, the expectation Indicates that in the parameter optimization strategy The expectation of all possible trajectories under , using the policy gradient formula to iteratively update, so that the parameter adjustment sequence output under the same initial working conditions can continuously reduce the quality error and achieve adaptive optimization of multi-objective process parameters. The policy gradient formula is defined as:

[0081] ;

[0082] Among them, the policy gradient uses the advantage function To update, Represents the state value estimation, and finally obtains the trained multi-objective process parameter optimization model and its parameter optimization strategy ;

[0083] S23. Export and serialize the optimal parameters of the above-mentioned processing quality prediction model and multi-objective process parameter optimization model. For the processing quality prediction model, export the weight file pred_weight.pt, along with the model input and output dimensions, network structure description, and normalization parameters during training; for the multi-objective process parameter optimization model, export the parameter optimization strategy. and value network weights, and save them as binary files for easy loading. At the same time, write the corresponding configuration files (including hyperparameters, environment state action space definitions and version information). Finally, use standardized tools to package all serialized files and generate model cards (Model Cards) to record model accuracy, training data statistics and potential usage scenarios, ensuring that they can be loaded and called with one click during subsequent microservice deployment.

[0084] S3. Encapsulate the two trained models as a quality prediction microservice and a process optimization microservice, respectively, publish them to a container platform based on the DeepSeek large model, and configure operation and maintenance policies.

[0085] As a preferred implementation of step S3, the specific process includes the following steps:

[0086] S31, the above-mentioned saved processing quality prediction model and parameter optimization strategies for multi-objective process parameter optimization models Encapsulated into microservices identifiable by the DeepSeek platform, namely the quality prediction microservice and the process optimization microservice, their mapping functions are defined as:

[0087] ;

[0088] Among them, input is the feature vector of the process parameters to be predicted, output To predict the quality feature vector, input For the current environment state, output To increase the recommended actions, we then used the microservice interface description language provided by DeepSeek to clearly define the name, data type, and value range of each field in the interface definition. We also generated a client call library through the DeepSeek SDK to ensure that the DeepSeek large model controller could seamlessly discover, call, and parse the microservice return results. At the same time, we automatically generated use cases and documentation for all interfaces.

[0089] S32. Based on the best practices of containerization, the above microservices and their runtime dependencies are packaged into lightweight images for rapid deployment on the DeepSeek container platform. The image construction adopts a multi-stage construction strategy, retaining only the components required for production environment operation, and placing the model files, normalization parameters and DeepSeek interface metadata in the image read-only layer, so as to control the final image size within the threshold. The image tag uses "deepseek.registry / project:v time " format, where v is the model version number and time is the build timestamp, so that the operation and maintenance system can perform rolling updates or rollbacks based on the version. To support high concurrency and low latency calls, health checks and readiness probes are pre-installed in the image, which automatically register with the DeepSeek service discovery module when started and are based on the peak request rate. Maximum processing capacity of a single instance The required number of copies is calculated based on the ratio ;

[0090] S33. Publish the image to the container platform based on the DeepSeek large model and complete the operation and maintenance configuration so that the microservices can be dynamically scheduled in the chain call process of the DeepSeek large model controller. First, pull the image to the computing node through the DeepSeek CI / CD pipeline, create a deployment unit based on predefined resource quotas (including CPU, GPU, and memory), and assign a unified route and domain name to each service; at the call level, configure circuit breaking, current limiting, and retry strategies to cope with instantaneous traffic peaks; at the monitoring level, set real-time alarms for key indicators such as latency, error rate, and throughput, and use log aggregation and link tracing functions to ensure that when using the DeepSeek large model as a controller for closed-loop calls to microservices, anomalies can be captured in a timely manner and trigger downgrades or manual intervention;

[0091] S4, using the DeepSeek large model as a controller to implement closed-loop calls to microservices;

[0092] As a preferred implementation of step S4, the specific process includes the following steps:

[0093] S41. Prepare the operating environment for the DeepSeek large model, including configuring high-performance computing nodes and necessary hardware resources, installing runtime libraries and dependency management tools compatible with DeepSeek LLM at the software level, and configuring environment variables. Deploy model weights and vocabulary files to distributed storage, and lock versions through container images or virtual environments to ensure a consistent and reproducible environment on different nodes.

[0094] S42. Use DeepSeek's interface definition specification to describe the call contracts of each microservice, and map the quality prediction microservice and process optimization microservice into formal functions respectively:

[0095] ;

[0096] in, Input process parameter feature vector , output prediction quality feature vector , Input environment status , output action increment In the interface definition document, declare the type, unit, and value range of each field using JSON Schema or IDL, and introduce the version number, interface signature hash, and timeout threshold to ensure the stability and traceability of interface calls during the reasoning process;

[0097] S43, design prompt words and dialogue context to guide the DeepSeek large model to gradually generate tool call instructions and explanatory output. The context is determined by the system instructions. , Historical Dialogue and current user intent The total length does not exceed the model context window ,Right now:

[0098] ;

[0099] In the prompt template, the interface name and parameters are dynamically filled in through the placeholders {SERVICE} and {INPUT}, and the model is explicitly required to be output in the format of "calling DeepSeek microservice";

[0100] S44. Build a reasoning process based on chain thinking, breaking down the entire optimization process into a series of tool call decisions, and let the context state be , the tool set is , then in Step 1 requires calculation for each tool The calling probability is:

[0101] ;

[0102] in, Calling for the model The original score, is the temperature coefficient, then select Generates RPC format call instructions based on the prompt words, and returns the results after execution Append to context: , loop until convergence is met or the predefined step limit is reached, forming a closed-loop collaborative call to the microservices by the DeepSeek large model as the controller.

[0103] In this embodiment, by constructing a DeepSeek large model controller, such as Figure 2 As shown in the figure, the large-scale pre-trained deep learning model is deeply integrated with the online closed-loop adaptive microservice framework to generate parameter strategies for the processing state of complex components, realizing intelligent perception, accurate prediction and dynamic optimization of complex process parameters and multi-dimensional quality indicators, solving the limitations of traditional experience-driven or static optimization methods in multi-objective coupling scenarios, significantly improving the accuracy and robustness of multi-indicator optimization, and enhancing the generalization ability in high-dimensional heterogeneous data and few-sample scenarios.

[0104] S5: Call the quality prediction microservice to calculate the error ratio of the target and generate an adaptive weight vector. Then call the process optimization microservice to generate action increments and apply them to the current environment state. This process iterates until all indicators converge, forming a multi-objective online closed-loop adaptive optimization process.

[0105] As a preferred implementation of step S5, the specific process includes the following steps:

[0106] S51. Call the quality prediction microservice and input the process parameter feature vector Returns the predicted quality feature vector , and according to the preset target mass vector Compute the multi-objective error vector:

[0107] ;

[0108] in Indicates the The prediction error of the target, represents the total number of targets, is the number of iterations, and then calculate the proportion of each error in all errors:

[0109] ;

[0110] Will As the adaptive weight component of the target, the adaptive weight vector , to reflect which indicator deviates more seriously from the target, so as to give it a higher priority in the next optimization step;

[0111] S52, the current environment state and adaptive weight vector Also passed to the process optimization microservice, its internal parameter optimization strategy Generate action increments based on a multi-objective weighted reward function , the multi-objective weighted reward function is defined as:

[0112] ;

[0113] Action Increment Maximize the comprehensive return in a weighted sense so that the error of the adjusted parameters on each quality indicator is reduced simultaneously;

[0114] S53, increment the obtained action Applies to the current environment state , and obtain the new process parameter characteristic vector , and call the quality prediction microservice again to calculate the new prediction quality feature vector , providing accurate feedback for the next round of weight update;

[0115] S54, pre-set convergence threshold vector ,according to Determine whether the convergence condition is met, that is, when all new prediction errors All satisfied When the multi-objective optimization is considered to have converged, the system outputs the final process parameter feature vector and the final prediction quality feature vector If any indicator exceeds the threshold, the environment state of the next cycle is set to , and and Re-enter the closed-loop iteration of steps S51 to S54 until all indicators meet the convergence criteria.

[0116] In this embodiment, by building microservices, quality indicators such as roughness, residual stress and deformation error are evaluated in real time, and the multi-objective optimization weights are adaptively adjusted according to feedback, which solves the problem of insufficient multi-indicator optimization accuracy and stability when the existing methods are faced with various process parameters in the processing process presenting a coupled relationship.

[0117] S6. Automatically write an analysis report based on the converged optimal process parameter feature vector and its corresponding predicted quality feature vector;

[0118] As a preferred implementation of step S6, the specific process includes the following steps:

[0119] S61, the optimal process parameter feature vector obtained by the final convergence and the corresponding prediction quality feature vector Perform structured summary to obtain a structured horizontal table , including the parameter names of cutting speed and cutting depth, as well as the names and unit descriptions of surface roughness and residual stress quality indicators, which can clearly show the final values ​​of various parameters and indicators before and after optimization;

[0120] S62, automatically generate visual charts around the key indicators in the iterative process, first draw the error convergence curve, and convert the multi-objective error norm of each iteration , with the number of iterations Draw a line graph on the horizontal axis to visually show how the error decreases with iteration; then draw a weight evolution curve for each quality target separately, and convert the adaptive weight Follow The changing curves are displayed side by side to observe the priority dynamics of each indicator during the optimization process; finally, a parameter performance comparison table is generated to compare each process parameter. In the initial state and the final state And the corresponding change in quality indicators List by row to clearly compare the performance differences before and after optimization;

[0121] S63, based on the preset report template, automatically writes the analysis report, which includes four parts: "Summary of optimization parameters", "Convergence analysis", "Weight evolution", and "Performance comparison". The structured horizontal table of S61 and all the visual charts of S62 are embedded in the corresponding chapters in order, and the report is preceded by a report including the learning rate, discount factor, and the performance comparison. , convergence threshold Experiment configuration and timestamp.

[0122] S7. Collect new machining data of a small batch of real workpieces, perform incremental fine-tuning and hyperparameter optimization on the previously trained machining quality prediction model, and design an automatic rollback strategy;

[0123] As a preferred implementation of step S7, the specific process includes the following steps:

[0124] S71. Collect new processing data of small batches of real workpieces online, and merge the process parameter feature matrix and the corresponding quality inspection label matrix of each workpiece into the incremental data pool. The newly added sample process parameter feature matrix is ​​recorded as , the quality inspection label matrix is , and perform the same preprocessing process as S12 on these real-time data to ensure that the incremental data and historical data are compared and integrated on the same distribution basis;

[0125] S72, the above-trained processing quality prediction model Perform incremental fine-tuning by minimizing the synthetic loss function:

[0126] ;

[0127] in, is the mean square error loss of the original training set, is the loss of the new sample, and is the weight coefficient controlling the retention of old knowledge and the learning of new knowledge, is the total number of samples collected, and the final updated parameter to , in order to keep the original generalization ability while taking into account the fitting of new samples;

[0128] S73. Dynamically adjust the hyperparameter set of the optimization model based on online performance feedback , first calculate the average error on the newly added samples and compare it with the historical error threshold Compare and if there is a significant increase, adjust the strategy learning rate proportionally to , adjust the discount factor to , ;

[0129] in is the adjustment rate factor. This adaptive mechanism enables the optimization model to converge to the new optimal strategy in real time when data drifts or working conditions change;

[0130] S74. Generate a unique version identifier for each incremental fine-tuning and hyperparameter adjustment , and the corresponding model weights and strategy parameters Register to the model management system while retaining the previous version In preparation for rollback, when the maximum error of the online version exceeds the set threshold in the actual small batch test situation, that is:

[0131] ;

[0132] in, It is a testing process The real The system will trigger an automatic rollback by calling the previous stable version. , and record the rollback reasons and performance comparisons in the management panel to ensure high availability and continuous adaptive optimization capabilities of the entire service.

[0133] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless they are mutually inconsistent.

[0134] The logic and / or steps represented in the flowchart or otherwise described herein may be considered, for example, as an ordered list of executable instructions for implementing logical functions, and may be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device).

[0135] The above embodiments provide a detailed introduction to the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A large model driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components, characterized by: The following steps are involved: S1. Collect multi-source heterogeneous processing data and quality inspection data of complex aviation components; S2. Based on the collected multi-source heterogeneous processing data and quality inspection data, a quality prediction model for complex aviation components is constructed and trained. A multi-objective process parameter optimization model is then constructed and trained to obtain a parameter optimization strategy. S3. Encapsulate the two trained models as a quality prediction microservice and a process optimization microservice, respectively, publish them to a container platform based on the DeepSeek large model, and configure operation and maintenance policies. S4, using the DeepSeek large model as a controller to implement closed-loop calls to microservices; S5: Call the quality prediction microservice to calculate the error ratio of the target and generate an adaptive weight vector. Then call the process optimization microservice to generate action increments and apply them to the current environment state. This process iterates until all indicators converge, forming a multi-objective online closed-loop adaptive optimization process. S6. Automatically write an analysis report based on the converged optimal process parameter feature vector and its corresponding predicted quality feature vector; S7. Collect new processing data of a small batch of real workpieces, perform incremental fine-tuning and hyperparameter optimization on the previously trained processing quality prediction model, and design an automatic rollback strategy.

2. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 1 is characterized by: The specific process of step S1 includes the following steps: S11. Deploy a unified data access module on the machine tool side, sensor side, and detection equipment side to collect process parameters, timing signals, and quality detection data, and push the data to the data aggregation layer and store it in the database in a standardized message format via industrial Ethernet or wireless edge gateway; S12. Perform unified preprocessing on the stored data, accurately align the timing signals of different device sampling frequencies by timestamps and fill the sampling gaps, repair missing values, apply normalization or standardization to all numerical features and filter outliers to generate a data set with consistent distribution.

3. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 1 is characterized by: The specific process of step S2 includes the following steps: S21. Based on the collected multi-source heterogeneous processing data and quality inspection data, a process parameter feature matrix and a corresponding quality inspection label matrix are constructed. A multi-layer perceptron and a multi-head attention mechanism are used to build a processing quality prediction model. The model training aims to minimize the mean square error loss and obtain the optimal weight of the training model. S22. Use the trained processing quality prediction model as the reward function in the reinforcement learning environment. Define the environment state as the current process parameter feature vector and its predicted quality feature vector. Define the action increment as the adjustment increment of the process parameters. Use the multi-agent SAC reinforcement learning algorithm improved by the attention mechanism to maximize the expected reward. Generate and train a multi-objective process parameter optimization model to obtain the parameter optimization strategy. S23. Export and serialize the trained processing quality prediction model and multi-objective process parameter optimization model for subsequent deployment and calling.

4. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 1 is characterized by: The specific process of step S3 includes the following steps: S31. Define input and output interface call libraries for the quality prediction microservice and process optimization microservice; S32. Package the above two microservices into a lightweight image. The image construction adopts a multi-stage construction strategy, and the model files, normalization parameters, and DeepSeek interface metadata are uniformly placed in the image read-only layer. The health check and readiness probes are pre-installed in the image. S33. Publish the image to a container platform based on the DeepSeek large model, create deployment units based on resource quotas, allocate routes and domain names, configure circuit breakers, rate limiters, and retry policies at the call level, and set real-time warnings for key indicators at the monitoring level to ensure that exceptions are captured in a timely manner during closed-loop calls.

5. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 4 is characterized by: The specific process of step S4 includes the following steps: S41. Configure the operating environment of the DeepSeek large model, deploy the model weights and vocabulary files, and lock the version; S42. Use DeepSeek's input and output interface definition specifications to describe the call contracts of each microservice, and map the quality prediction microservice and process optimization microservice into formal functions respectively: ； in, Input process parameter feature vector , output prediction quality feature vector , Input environment status , output action increment In the input and output interface definition documents, declare the type, unit, and value range of each field using JSON Schema or IDL, and introduce version numbers, interface signature hashes, and timeout thresholds to ensure the stability and traceability of interface calls during the inference process; S43, design prompt words and dialogue context, the context is determined by system instructions , Historical Dialogue and current user intent The total length does not exceed the model context window ,Right now: ； In the prompt template, the interface name and parameters are dynamically filled in through the placeholders {SERVICE} and {INPUT}, and the model is explicitly required to output in the format of calling the DeepSeek microservice; S44. Construct a reasoning process based on chain thinking, and let the context state be , the tool set is , then in Step 1 requires calculation for each tool The calling probability is: ； in, Calling for the model The original score, is the temperature coefficient, then select Generates RPC format call instructions based on the prompt words, and returns the results after execution Append to context: , loop until convergence is met or the predefined step limit is reached, forming a closed-loop collaborative call to the microservices by the DeepSeek large model as the controller.

6. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 1 is characterized by: The specific process of step S5 includes the following steps: S51. Call the quality prediction microservice and input the process parameter feature vector Returns the predicted quality feature vector , and according to the preset target mass vector Compute the multi-objective error vector: ； in Indicates the The prediction error of the target, represents the total number of targets, is the number of iterations, and then calculate the proportion of each error in all errors: ； Will As the adaptive weight component of the target, the adaptive weight vector ; S52, the current environment state and adaptive weight vector Also passed to the process optimization microservice, its internal parameter optimization strategy Generate action increments based on a multi-objective weighted reward function , the multi-objective weighted reward function is defined as: ； S53, increment the obtained action Applies to the current state of the environment , and obtain the new process parameter feature vector , and call the quality prediction microservice again to calculate the new prediction quality feature vector , providing accurate feedback for the next round of weight update; S54, according to the pre-set convergence threshold vector , when all new prediction errors All satisfied When the multi-objective optimization is considered to have converged, the system outputs the final process parameter feature vector and the final prediction quality feature vector If any indicator exceeds the threshold, the environment state of the next cycle is set to , and and Re-enter the closed-loop iteration of steps S51 to S54 until all indicators meet the convergence criteria.

7. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 1 is characterized by: The specific process of step S6 includes the following steps: S61, performing structured aggregation on the optimal process parameter feature vector obtained by final convergence and its corresponding predicted quality feature vector to obtain a structured horizontal table; S62. Automatically generate visual charts around key indicators in the iteration process, including error convergence curves, weight evolution curves, and parameter performance comparison tables; S63. Automatically write an analysis report based on the preset report template. The report is divided into four parts: optimization parameter summary, convergence analysis, weight evolution, and performance comparison. The table of S61 and all the charts of S62 are embedded in the corresponding chapters in sequence, and the experimental configuration and timestamp are attached at the beginning of the report.

8. The large model-driven adaptive closed-loop optimization method for machining process parameters of complex thin-walled aviation components according to claim 1 is characterized by: The specific process of step S7 includes the following steps: S71, online collection of new processing data of small batches of real workpieces and execution of the same preprocessing process as S12; S72. Incrementally fine-tune the previously trained processing quality prediction model by minimizing the synthetic loss function: ； in, is the mean square error loss of the original training set, is the loss of the new sample, and is the weight coefficient controlling the retention of old knowledge and the learning of new knowledge, is the total number of samples collected, and the final updated parameter to ; S73. Calculate the average error on the newly added samples and compare it with the historical error threshold to dynamically adjust the model's hyperparameters. S74. Generate a unique version identifier for each incremental fine-tuning and hyperparameter adjustment , and the corresponding model weights and strategy parameters Register to the model management system while retaining the previous version In preparation for rollback, when the maximum error of the online version in actual small-batch testing exceeds the set threshold, the system will trigger an automatic rollback to the previous stable version.

Images