A cross-process quality real-time correction method for multi-process manufacturing process

Abstract

The application relates to a cross-process quality real-time correction method for a multi-process manufacturing process, which comprises collecting process parameter data, quality detection data and process meta-information corresponding to each manufacturing process of a part, constructing a manufacturing process sequence dataset according to the process sequence, constructing a cross-process quality propagation model to extract a whole-process state vector representing the quality state of the whole part manufacturing process and a final quality prediction result, constructing and training a cross-process correction decision model to obtain an optimal compensation vector to correct the process parameters, realizing dynamic control and real-time correction of the part manufacturing equipment, and designing a fine-tuning strategy based on incremental learning and reinforcement learning to dynamically optimize the control and correction process in each control cycle. The application solves the technical bottlenecks in traditional quality control, such as the difficulty in describing the quality transmission relationship between processes, the lag of dynamic correction and the lack of self-adaptive ability of the model, and improves the overall quality stability and consistency in the complex part manufacturing scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and process quality control technology, and in particular to a method for real-time quality correction across multiple manufacturing processes. Background Technology

[0002] In the manufacturing process of complex structural components and high-precision parts, multiple consecutive machining or assembly processes are often required. These processes differ significantly in terms of technology, equipment conditions, and control parameters. The process execution and quality results of each process are not independent but rather form a continuous transmission relationship in terms of materials, geometry, and stress states. Small errors in one process can be amplified in subsequent processes, ultimately leading to assembly difficulties, performance fluctuations, or even product failure. Therefore, the quality control problem in multi-process manufacturing is essentially a typical problem of cross-process error propagation and accumulation.

[0003] Existing quality control methods in manufacturing processes mainly focus on single or partial processes. They involve setting up inspection points at specific processes to assess the current processing quality, and then having on-site personnel take rework or localized adjustment measures upon detecting deviations. This approach generally relies on manual experience for parameter correction, lacking systematic modeling of error sources and propagation paths. On the one hand, it makes it difficult to intervene before errors become apparent; on the other hand, it easily leads to randomness and lag in correction actions, failing to meet the consistency and stability requirements of high-precision manufacturing.

[0004] In recent years, with the application of data-driven methods in the manufacturing field, some studies have begun to attempt to use machine learning models to predict processing quality or to use optimization algorithms to adjust process parameters offline. However, these methods are mostly aimed at single-process quality modeling, often simplifying or ignoring the correlation between processes, and failing to characterize the dynamic evolution of quality in the process chain. Even if there are a few quality analysis methods for multiple processes, they usually remain at the level of statistical correlation analysis or rule-driven approaches, lacking a unified modeling framework and making it difficult to support parameter-level corrective decisions.

[0005] In addition, traditional parameter optimization methods are mostly based on static process window settings or manual parameter adjustment strategies, which are difficult to cope with quality fluctuations caused by complex factors such as material batch changes, equipment status changes and environmental disturbances. The parameter update cycle is long and the response speed is slow. They are still mainly based on post-event correction and lack a mechanism to make real-time adjustments based on prediction results during the manufacturing process.

[0006] In summary, existing technologies still have significant shortcomings in cross-process quality propagation modeling, real-time parameter correction, and adaptive model updates. There is an urgent need for a method that can model multi-process manufacturing from a full-process perspective and intelligently correct process parameters driven by quality prediction, so as to realize the transformation of multi-process manufacturing systems from "experience control" to "model-driven and data-driven collaborative control". Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a real-time cross-process quality correction method for multi-process manufacturing. It solves problems such as information fragmentation between processes, difficulty in eliminating error accumulation, and reliance on experience for parameter adjustment in traditional multi-process manufacturing. Based on multi-process data fusion modeling, it combines cross-process quality propagation prediction with parameter compensation decision-making to overcome technical bottlenecks in traditional quality control, such as difficulty in depicting quality transfer relationships between processes, lag in dynamic correction, and lack of model adaptability. It constructs a closed-loop control system for multi-process quality prediction, parameter compensation, and strategy optimization, thereby achieving feedforward intervention and proactive correction of manufacturing process quality and improving the overall quality stability and consistency in complex manufacturing scenarios.

[0008] To achieve the above technical objectives, the present invention provides the following technical solution: a method for real-time quality correction across multiple manufacturing processes, comprising the following steps: Collect process parameter data, quality inspection data and process metadata corresponding to each manufacturing process of the parts and quantize them. Perform preprocessing and alignment to generate process parameter vectors, quality feature vectors and process metadata vectors. Construct process state triplets for each manufacturing process and form a manufacturing process sequence dataset according to the order of the processes. A cross-process quality propagation model is constructed and pre-trained. The manufacturing process sequence dataset is input into the pre-trained cross-process quality propagation model to model the quality dependency relationship between each manufacturing process, extract the full-process state vector representing the quality status of the entire part manufacturing process, and generate the final quality prediction result. Construct a cross-process correction decision model, and input the full-process state vector and process position code into the cross-process correction decision model to obtain the parameter compensation vector; The process location code includes the current process index and the identifier of the subsequent adjustable process; A multi-objective optimization function is constructed using the final quality prediction result, energy consumption level, and defect risk as joint optimization objectives. The cross-process correction decision model is optimized, and the parameter compensation vector generated by the optimized cross-process correction decision model is used as the optimal compensation vector. The process parameter vector is corrected by the optimal compensation vector, and parameter instructions are generated to update the process parameters of subsequent manufacturing processes, so as to realize dynamic control and real-time correction of the parts manufacturing equipment. After completing a multi-process manufacturing process in a control cycle, the cross-process quality propagation model is fine-tuned based on incremental learning based on the measured quality results of the currently manufactured parts, and the cross-process correction decision model is fine-tuned based on reinforcement learning, so as to realize dynamic optimization of multi-process manufacturing and cross-process quality real-time correction of parts.

[0009] Optionally, the process parameter data, quality inspection data, and process metadata corresponding to each manufacturing process of the component are collected and quantized, preprocessed and aligned to generate process parameter vectors, quality feature vectors, and process metadata vectors, and a process state triplet is constructed for each manufacturing process, including: A data acquisition interface is set in the equipment control system corresponding to each manufacturing process of the parts to collect process parameter data in real time or near real time, and the collected process parameter data is vectorized to obtain the original process parameter vector. The original process parameter vector includes at least the fields of cutting speed, feed rate, depth of cut, tool compensation amount, path strategy code, machining cycle time, and a uniformly introduced timestamp field, and is recorded with a uniform sampling period and timestamp format. After each manufacturing process is completed, the generated parts are inspected according to the preset quality inspection plan, the quality inspection data is obtained and quantified to form an original quality feature vector; the original quality feature vector is associated with the corresponding timestamp and the unique identifier of the part; the quality inspection data includes at least dimensional deviation, form and position deviation, critical wall thickness, hole position deviation, surface quality grade, and defect type; For each manufacturing process, collect the corresponding process metadata and quantize it to obtain the original process metadata vector; the process metadata includes at least the process type, equipment number and machine tool model, clamping method, clamping posture, workpiece batch number, and ambient temperature and humidity, as well as the process context and equipment and environmental conditions used to describe the manufacturing process when it occurs. The obtained original process parameter vector, original quality feature vector, and original process element information vector are uniformly aligned in time, outlier removal is performed, missing data is imputed, and numerical standardization is performed to obtain process parameter vector, quality feature vector, and process element information vector. Then, the process parameter vector, quality feature vector, and process element information vector of each manufacturing process are concatenated to construct process state triplet.

[0010] Optionally, the cross-process quality propagation model includes a process embedding layer, a cross-process encoding layer, a convergence layer, and a quality prediction head; the process embedding layer is a multilayer perceptron or other nonlinear mapping structure; the cross-process encoding layer is a Transformer encoder with a self-attention mechanism or other equivalent cross-sequence modeling structure. The process involves inputting the manufacturing process sequence dataset into a pre-trained cross-process quality propagation model to model the quality dependencies between various manufacturing processes, extracting a full-process state vector representing the quality status of the entire component manufacturing process, and generating the final quality prediction result, including: The process state triplet of each manufacturing process in the manufacturing process sequence dataset is input into the process embedding layer. The process parameter vector, quality feature vector, and process meta-information vector in the process state triplet are jointly encoded through a nonlinear mapping structure to obtain the process vector of each manufacturing process. The process vectors of all manufacturing processes are arranged into a process sequence according to the order of the processes. This sequence is then input into the cross-process coding layer. The mutual influence weights between different processes are calculated through multi-head self-attention, and the process vectors of each manufacturing process in the process sequence are weighted to obtain the hidden state sequence. The hidden state sequence is input into the aggregation layer, and the hidden state vectors of each manufacturing process are aggregated to obtain the full process state vector. Input the entire process state vector into the quality prediction head to obtain the final quality prediction result.

[0011] Optionally, the step of inputting the entire process state vector and the process position code into the cross-process correction decision model to obtain the parameter compensation vector includes: The entire process state vector and the process position code are concatenated at the vector level to form a joint state vector, which is then input into the cross-process correction decision model to output parameter compensation vectors for subsequent manufacturing processes. The cross-process correction decision model employs a multilayer perceptron, attention network, or other nonlinear regression structure.

[0012] Optionally, the mathematical representation of the constructed multi-objective optimization function is as follows: ; ; in, For the constructed multi-objective optimization function; This represents the parameter compensation vector generated by the cross-process correction decision model. Represents the optimal compensation vector; This indicates the final quality prediction result; Indicates energy consumption level, Indicates defect risk; , , These are the weighting parameters for the final quality prediction result, energy consumption level, and defect risk, respectively. Indicates the need to make smallest That is, the optimization objective of this multi-objective optimization function.

[0013] Optionally, the step of correcting the process parameter vector with the optimal compensation vector includes: The optimal compensation vector is superimposed onto the process parameter vector of the corresponding manufacturing process to generate the corrected process parameter vector. Set lower and upper limits for the modified process parameter vector. If the modified process parameter vector exceeds the range of these limits, it will be trimmed or rolled back to the state before overlay.

[0014] Optionally, the step of correcting the process parameter vector with the optimal compensation vector further includes: By setting a gradual control curve or interpolation trajectory, the process parameter vector is gradually transitioned to the corrected process parameter vector, avoiding dynamic shocks to the machine tool, increased vibration, or instability in the machining process caused by sudden parameter changes.

[0015] Optionally, the fine-tuning of the cross-process quality propagation model based on incremental learning includes: The quality prediction error is calculated based on the final quality prediction result and the measured quality result. ; With the aforementioned quality prediction error Constructing the loss function of the prediction model The mathematical representation is as follows: ; in Represents the calculation of the L2 norm; based on the loss function of the prediction model. Incremental learning is used to fine-tune and update the parameters of the cross-process quality propagation model.

[0016] Optionally, the fine-tuning of the cross-process correction decision model based on reinforcement learning includes: The full-process state vector corresponding to each control cycle is used as the reinforcement learning state, the parameter compensation vector executed in that control cycle is used as the action, and the final quality prediction result in the corresponding cycle is used as the immediate reward, thereby constructing a reinforcement learning sample triplet. Estimating cumulative returns using a discounted return method: ; in, Indicates the current control cycle The cumulative income, For subsequent control cycles Discount factor, Indicates in Instant rewards for controlling the cycle, Indicates infinity; Based on this discount return, reinforcement learning is applied to the cross-process correction decision model to optimize parameter adjustment patterns while continuously absorbing production data and quality feedback from new control cycles.

[0017] The present invention also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the aforementioned cross-process quality real-time correction method for multi-process manufacturing.

[0018] By employing the above technical solution, the present invention provides a method for real-time quality correction across multiple manufacturing processes, which has at least the following beneficial effects: (1) This invention vectorizes the collected process parameter data, quality inspection data, and process element information, providing an input basis for the subsequent modeling of the cross-process quality propagation model. Specifically, a timestamp field is uniformly introduced into the process parameter data to achieve time alignment between different processes, enabling the cross-process quality propagation model to restore the actual execution trajectory of the manufacturing process according to the time sequence during modeling, avoiding data misalignment caused by differences in sampling frequency; the quality inspection data is doubly bound to the corresponding timestamp or unique component identifier, which can accurately establish the mapping relationship between "parameter-quality" in the subsequent modeling stage, thereby supporting causal inference in the cross-process error propagation analysis and correction decision process; the vectorization of process element information provides auxiliary information and enhances the model's adaptability to equipment heterogeneity and batch differences.

[0019] (2) This invention performs unified preprocessing and alignment on the acquired original process parameter vector, original quality feature vector, and original process meta-information vector, so that the process state triplet of each manufacturing process has a unified scale in the numerical space and can be directly used as the input of the subsequent cross-process quality propagation model, so that the data layer and the model layer are consistent in structure and semantics; at the same time, by forming a manufacturing process sequence dataset according to the order of the processes, it provides a structured standard data form that can be directly input into the model for cross-process quality propagation modeling and real-time correction.

[0020] (3) This invention transforms the original heterogeneous vector data into process vectors with unified semantic expression by jointly encoding the process parameter vector, quality feature vector, and process meta-information vector in the process state triplet, thereby enabling the subsequent cross-process quality propagation model and cross-process correction decision model to learn the quality correlation patterns between different processes in the same representation space.

[0021] (4) This invention organizes the process vectors of all manufacturing processes into a process sequence and encodes the hidden state sequence, so that the small deviations generated by the early manufacturing processes can be propagated in the feature space to the representation process of the subsequent manufacturing processes, thereby reflecting the transmission law of "preceding errors affecting subsequent quality". Compared with the traditional modeling method based on fixed windows, this invention automatically learns the correlation weights between key processes in the modeling stage, and gives more attention to manufacturing processes with high quality contribution, thereby improving the accuracy and expressive ability of cross-process quality propagation modeling.

[0022] (5) This invention compresses the quality influence factors that were originally scattered in multiple manufacturing processes into a single full-process hidden state vector by converging the hidden state vectors of each manufacturing process, so that the hidden state vector can represent the state of the entire manufacturing chain in subsequent decision-making and prediction.

[0023] (6) This invention extracts the state vector of the entire process by constructing a cross-process quality propagation model and generates the final quality prediction result. The final quality prediction result is used as the evaluation basis of the quality status of the currently processed parts and as the constraint and optimization target of subsequent parameter compensation, correction and optimization control. This allows the prediction result of the cross-process quality propagation model to directly participate in the decision-making process in the closed-loop correction system, thereby realizing a cross-process parameter adjustment mechanism driven by quality prediction.

[0024] (7) This invention integrates process position coding with state vectors throughout the entire process, making the correction decision forward-looking and executable, thereby supporting the subsequent parameter adjustment scheme to be optimized in a coordinated manner with the overall process quality target as the guide; at the same time, the constructed cross-process correction decision model learns the mapping relationship between "quality deviation - parameter adjustment - result improvement" in historical process data, so that the output result can reflect the parameter adjustment direction that is most conducive to correcting the quality trend under the current working conditions.

[0025] (8) This invention constructs a multi-objective optimization function by taking the final quality prediction result, energy consumption level and defect risk as joint optimization objectives, and maps the originally scattered quality indicators, energy consumption indicators and risk indicators into the same optimization space, so that the parameter correction process can be comprehensively weighed under the framework of a single objective function. At the same time, the optimization objective can reflect the quality results, resource consumption and potential risks simultaneously, thereby avoiding the bias problem that the correction process only pursues quality and ignores safety and economy.

[0026] (9) This invention corrects the process parameter vector with the optimal compensation vector, transforming the originally irreversible preceding quality error into a controllable parameter offset in the subsequent process, thereby realizing the transfer of error from the "result space" to the "parameter space" and changing the correction behavior from post-process rework to process absorption. At the same time, safety boundary constraints are set on the corrected process parameter vector according to the equipment capacity, tool safety range and process specifications, ensuring that the parts processing equipment always operates within a controllable and safe process window during the real-time correction process, thereby improving the feasibility and reliability of the parameter adjustment scheme in the actual production environment. In addition, during the process parameter vector update process, the process parameter vector is gradually transitioned to the corrected process parameter vector to avoid the dynamic impact of parameter mutations on the machine tool, increased vibration or instability of the processing process.

[0027] (10) This invention fine-tunes the cross-process quality propagation model based on incremental learning by calculating the quality prediction error. While ensuring that the cross-process quality propagation model has the ability to quickly adapt to new working conditions, it avoids excessive destruction of the learned knowledge, enabling the cross-process quality propagation model to have continuous learning capabilities. At the same time, it constructs reinforcement learning sample triples to fine-tune the cross-process correction decision model based on reinforcement learning. In the process of continuously absorbing production data and quality feedback from new control cycles, it optimizes the parameter adjustment rules, thereby gradually improving the effect of real-time cross-process quality correction and the adaptability of the equipment control system to different working conditions, different structural components, and environmental disturbances. Through this update mechanism, the correction strategy can form a memory of different process combinations and quality response characteristics in long-term operation, realizing the transformation from "relying on experience" to "data-driven decision-making", enabling the equipment control system to continuously improve the correction accuracy and robustness in the continuous production process. Attached Figure Description

[0028] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a cross-process quality real-time correction method for multi-process manufacturing, according to the present invention. Detailed Implementation

[0029] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0030] Those skilled in the art will understand that all or part of the steps in the implementation of the methods of the embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can 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.

[0031] Please refer to Figure 1 This illustration shows a specific implementation of the present embodiment. This embodiment extracts the full-process state vector and the final quality prediction result that characterize the quality status of the entire component manufacturing process by constructing a cross-process quality propagation model; it constructs and trains a cross-process correction decision model to obtain the optimal compensation vector to correct process parameters, thereby realizing dynamic control and real-time correction of component manufacturing equipment; and it designs a fine-tuning strategy based on incremental learning and reinforcement learning to dynamically optimize the control and correction process in each control cycle. This solves the technical bottlenecks in traditional quality control, such as the difficulty in characterizing the quality transfer relationship between processes, the lag in dynamic correction, and the lack of adaptive capability of the model. It constructs a closed-loop control system for quality prediction, parameter compensation, and strategy optimization for multiple processes, thereby realizing feedforward intervention and proactive correction of manufacturing process quality and improving the overall quality stability and consistency in complex manufacturing scenarios.

[0032] This embodiment proposes a real-time cross-process quality correction method for multi-process manufacturing, which includes the following steps: S1. Collect process parameter data, quality inspection data and process element information corresponding to each manufacturing process of the parts and quantize them. Perform preprocessing and alignment to generate process parameter vectors, quality feature vectors and process element information vectors. Construct process state triplets for each manufacturing process and form a manufacturing process sequence dataset according to the order of the processes.

[0033] As a preferred embodiment of step S1, the specific process includes: S11. Set up a data acquisition interface in the equipment control system corresponding to each manufacturing process of the parts to collect process parameter data in real time or near real time, and vectorize the collected process parameter data to obtain the original process parameter vector; the original process parameter vector includes at least the fields of cutting speed, feed rate, cutting depth, tool compensation amount, path strategy code, machining cycle time, and a uniformly introduced timestamp field, and records them with a uniform sampling period and timestamp format, so that the process parameter data of each manufacturing process can be sequentially aligned and compared when modeling across processes.

[0034] The data acquisition interface is preferably integrated into a CNC system, programmable logic controller (PLC), or host computer monitoring system. It samples and records process parameters for each manufacturing process at the start, middle, and end stages. Based on the actual conditions of the parts manufacturing equipment, a fixed time interval sampling method or an event-triggered sampling method is set, ensuring the comparability of process parameter data collected under different equipment and process types. In specific implementation, the process parameter data is organized into vector form, mapping parameters such as rotational speed, feed rate, depth of cut, tool compensation speed, path segment index, and machining cycle time to continuous numerical variables in the same dimensional space, thus forming a multi-dimensional feature vector structure. This provides a foundation for subsequent cross-process quality propagation modeling and weight allocation of different parameters. Simultaneously, by uniformly introducing a timestamp field into the collected process parameter data, time alignment between different processes is achieved. This allows the cross-process quality propagation model to reconstruct the actual execution trajectory of the manufacturing process based on chronological order during modeling, avoiding data misalignment caused by differences in sampling frequency.

[0035] S12. After each manufacturing process is completed, the generated parts are inspected according to the preset quality inspection plan. Quality inspection data including dimensional deviation, form and position deviation, critical wall thickness, hole position deviation, surface quality grade, defect type, etc. are obtained and quantified to form an original quality feature vector. The original quality feature vector is associated with the corresponding timestamp and the unique identifier of the part to ensure that each quality feature can be traced back to the process parameter configuration of the specific process.

[0036] In practical implementation, various quality inspection data can be acquired through inspection equipment such as coordinate measuring machines, online probes, laser scanners, vision systems, and surface roughness testers, and uniformly encoded into a vector format. Dimensional deviations and hole position deviations are encoded as continuous features, while surface quality grades and defect types are encoded as discrete features. The original quality feature vector not only reflects the processing results of a single manufacturing process but also implicitly contains the cumulative quality effects of preceding processes. By doubly binding it with the corresponding timestamp or unique component identifier, a precise mapping relationship between "parameters" and "quality" can be established in the subsequent modeling stage, thereby supporting causal inference in the cross-process error propagation analysis and correction decision-making process.

[0037] S13. Collect the corresponding process metadata for each manufacturing process and quantize it to obtain the original process metadata vector. The process metadata includes at least the process type (such as turning, milling, drilling, boring, five-axis machining), equipment number and machine tool model, clamping method, clamping posture, workpiece batch number, and ambient temperature and humidity. It also includes the process context and equipment and environmental conditions used to describe the manufacturing process when it occurs, so that subsequent modeling can explicitly consider the impact of equipment differences and environmental disturbances on quality.

[0038] Among them, the process type is used to characterize the structural differences introduced by different processing methods; the equipment number and machine tool model are used to describe the differences in CNC system, spindle stiffness and accuracy level; the clamping method and clamping posture reflect the support state and stress of the parts during processing; the workpiece batch number is used to track the systematic differences introduced by material batches; and the ambient temperature and humidity characterize the influence of the external environment on thermal deformation and equipment stability. The above process meta-information can be combined with numerical features and discrete labels to form auxiliary variables (i.e., original process meta-information vector) describing the process environment state, so as to enhance the model's adaptability to equipment heterogeneity and batch differences.

[0039] S14. Perform unified time alignment, outlier removal, missing data imputation, and numerical standardization on the obtained original process parameter vector, original quality feature vector, and original process meta-information vector to obtain the process parameter vector, quality feature vector, and process meta-information vector. Then, concatenate the process parameter vector, quality feature vector, and process meta-information vector for each manufacturing process to construct a process state triplet, mathematically represented as follows: ; in Indicates manufacturing process The process status triplet, , , They represent the first The process parameter vector, quality feature vector, and process metadata vector are used for each manufacturing step. Next, a manufacturing process sequence dataset is formed according to the order of the steps: This provides a structured, standard data format that can be directly input into the model for cross-process quality propagation modeling and real-time correction; In the above formula This represents a dataset representing the manufacturing process sequence. This represents the total number of manufacturing processes.

[0040] During preprocessing and alignment, time alignment is based on the unique identifier of the component and the order of process execution. Outlier removal is determined based on process experience thresholds or statistical distributions. Missing values ​​required for data imputation can be filled by historical sample interpolation or model prediction. Numerical standardization is used to eliminate the influence of different units on model training. The following standardization or normalization methods are preferred: ; or ; in, This represents the original feature values ​​(i.e., the feature values ​​of the features that need to be standardized or normalized). The feature values ​​of the features after standardization or normalization. , They are respectively The statistical mean and standard deviation, , They are respectively The upper and lower bounds are defined. After processing, the process state triples of each manufacturing process have a unified scale in the numerical space, which can be directly used as input for the subsequent cross-process quality propagation model, so that the data layer and the model layer are consistent in structure and semantics.

[0041] S2. Construct a cross-process quality propagation model and pre-train it. Input the manufacturing process sequence dataset into the pre-trained cross-process quality propagation model to model the quality dependency relationship between each manufacturing process, extract the full-process state vector representing the quality status of the entire part manufacturing process, and generate the final quality prediction result.

[0042] As a preferred embodiment of step S2, the specific process includes: S21. Before inputting the real-time manufacturing process sequence dataset into the cross-process quality propagation model for online prediction, the cross-process quality propagation model needs to be pre-trained offline using historical batch manufacturing data. The specific process includes: The manufacturing process sequence dataset of historical batches of parts is obtained as the training input, and the corresponding historical measured quality results are obtained as the true labels. A pre-trained loss function is constructed based on the error between historical quality prediction results and the true labels. The mathematical representation is as follows: ; in, The number of historical training samples, For the first Model prediction results for each sample For the first The true label of each sample; Based on the pre-training loss function The backpropagation algorithm is adopted, and the network parameters of the process embedding layer, cross-process encoding layer, convergence layer and quality prediction head are updated through the gradient descent optimizer until the pre-training loss function converges, thus obtaining the pre-trained cross-process quality propagation model.

[0043] S22. Input the process state triplet of each manufacturing process in the manufacturing process sequence dataset into the process embedding layer. Use a multilayer perceptron or other nonlinear mapping structure to jointly encode the process parameter vector, quality feature vector, and process meta-information vector in the process state triplet to obtain the process vector of each manufacturing process, mathematically represented as: ; in Indicates the process embedding layer. Indicates the first The process vector of the manufacturing process.

[0044] The process vectors uniformly represent the state characteristics of different types of processes in the numerical space, facilitating subsequent sequence modeling within the same network. The process embedding layer fuses process parameter vectors, quality feature vectors, and process metadata vectors at the feature level, preventing single-modal features from dominating prediction results during modeling. In specific implementations, the multilayer perceptron can consist of several fully connected layers, nonlinear activation functions, and normalization layers. It performs scale unification processing on continuous-valued features and uses embedded vectors to encode discrete feature values, ensuring comparability of data from different sources and with different scales within the embedding space. Through embedding mapping via the process embedding layer, the original heterogeneous vector data is transformed into process vectors with unified semantic expression capabilities. This enables subsequent cross-process quality propagation models and cross-process correction decision models to learn quality correlation patterns between different processes within the same representation space.

[0045] S23. Arrange the process vectors of all manufacturing processes into a process sequence according to the process order, as shown mathematically below: ; in The process sequence is represented; the process sequence is input into the cross-process encoding layer, and a Transformer encoder with self-attention mechanism or an equivalent cross-sequence modeling structure is used. The mutual influence weights between different processes are calculated and weighted through multi-head self-attention, and the relationship between quality propagation and error accumulation between each manufacturing process is explicitly modeled to obtain the encoded hidden state sequence, which is mathematically represented as follows: ; in Indicates a cross-process coding layer. The hidden state sequence is represented by a hidden state vector, where each hidden state vector corresponds to the context representation of a manufacturing process. The cross-process encoding layer establishes global dependencies between different manufacturing processes through a self-attention mechanism, enabling small deviations generated in earlier manufacturing processes to propagate in the feature space to the representation of subsequent manufacturing processes, thus reflecting the propagation law that "preceding errors affect subsequent quality." Compared to traditional fixed-window-based modeling methods, the aforementioned self-attention structure can automatically learn the correlation weights between key processes during the modeling stage, giving greater attention to manufacturing processes with high quality contributions, thereby improving the accuracy and expressive power of cross-process quality propagation modeling.

[0046] S24. Input the hidden state sequence into the aggregation layer, and use methods such as weighted averaging, max pooling or selecting the hidden state of the last manufacturing process to aggregate the hidden state vectors of each manufacturing process to obtain the full process state vector that represents the quality status of the entire part manufacturing process. This full process state vector comprehensively reflects the combined effect of the process parameter configuration and quality characteristics of each manufacturing process on the final quality level.

[0047] In practical implementation, an appropriate convergence strategy can be selected based on the complexity of the component manufacturing process and the quality objectives. When more attention is paid to the overall process consistency, a weighted average method can be used; when more attention is paid to extreme value risks, a max pooling method can be used; and when emphasizing the decisive influence of the final manufacturing process on the final quality, the hidden state of the final manufacturing process can be selected as the global representation. Through the above convergence process, the quality influence factors that were originally scattered in multiple manufacturing processes are compressed into a single full-process hidden state vector, so that the hidden state vector has the ability to represent the state of the entire manufacturing chain in subsequent decision-making and prediction.

[0048] S25. Input the entire process state vector into the quality prediction head, and obtain the final quality prediction result through a regression network or multi-task output structure, which is mathematically represented as follows: ; in This indicates the final quality prediction result. Indicates the quality prediction head, This represents the state vector of the entire process. The final quality prediction result can include one (single-dimensional) or multiple (multi-dimensional) quality indicators such as comprehensive quality score, key assembly dimension deviation, hole position accuracy, and minimum wall thickness, which are used to predict and evaluate the final quality level of the current process sequence under default subsequent parameters.

[0049] The quality prediction head can employ a single-output regression structure to predict a single-dimensional comprehensive quality index, or it can use a multi-output structure to predict various key quality characteristics separately and form multi-dimensional final key quality characteristics based on set weights. The resulting final quality prediction results not only serve as the evaluation basis for the quality status of currently processed parts, but also as constraints and optimization objectives for subsequent parameter compensation, correction, and optimization control. This allows the prediction results of the cross-process quality propagation model to directly participate in the decision-making process within the closed-loop correction system, thereby realizing a cross-process parameter adjustment mechanism driven by quality prediction.

[0050] S3. Construct a cross-process correction decision model, inputting the full-process state vector and process position code into the cross-process correction decision model to obtain the parameter compensation vector.

[0051] As a preferred embodiment of step S3, the specific process includes: S31. The entire process state vector and the process position code are concatenated at the vector level to form a joint state vector, which is then input into the cross-process correction decision model.

[0052] The process location code includes a current process index and subsequent adjustable process identifiers, used to depict the combined state of "processed processes + unprocessed processes" at the current moment, providing a complete quality and process context for generating targeted parameter compensation schemes. The current process index indicates the position of the manufacturing process, and the subsequent adjustable process identifier clarifies the set of processes with adjustable degrees of freedom, ensuring that corrective decisions are made only within the scope of processes with execution capabilities, thereby avoiding the generation of invalid instructions for non-adjustable processes.

[0053] In specific implementation, the process position code and the full-process state vector are concatenated at the vector level to form a joint state vector, which is then input into the cross-process correction decision model. The joint state vector not only contains the quality information that has occurred, but also explicitly distinguishes between the process boundaries of "deviations that have occurred" and "compensable processes", making the correction decision forward-looking and executable, thereby supporting the subsequent parameter adjustment scheme to be optimized in a coordinated manner with the overall process quality objective as the guide.

[0054] S32. The cross-process correction decision model outputs parameter compensation vectors for subsequent manufacturing processes. This cross-process correction decision model employs a multilayer perceptron, attention network, or other nonlinear regression structure to comprehensively evaluate quality status and process margin, outputting parameter compensation vectors for one or more subsequent adjustable processes. The mathematical representation is as follows: ; in, Indicates the first The parameter compensation vector for the first and subsequent manufacturing processes can include subdivided parameters such as feed rate correction, depth of cut correction, and coordinate offset correction. This represents a cross-process corrective decision-making model. This represents the joint state vector formed by vector-level concatenation of the overall process state vector and the process position code.

[0055] The cross-process correction decision model learns the mapping relationship between "quality deviation - parameter adjustment - result improvement" in historical process data, enabling the output to reflect the parameter adjustment direction most conducive to correcting quality trends under the current operating conditions. In its implementation, different dimensional parameter compensation vectors can be configured for different manufacturing process types, allowing the output structure of the cross-process correction decision model to adapt to diverse equipment control interfaces and process specification requirements, thereby ensuring the universality and scalability of the correction strategy in multi-equipment and multi-process scenarios.

[0056] S4. Construct a multi-objective optimization function using the final quality prediction result, energy consumption level, and defect risk as joint optimization objectives, optimize the cross-process correction decision model, and use the parameter compensation vector generated by the optimized cross-process correction decision model as the optimal compensation vector.

[0057] As a preferred embodiment of step S4, the mathematical representation of the constructed multi-objective optimization function is as follows: ; ; in, For the constructed multi-objective optimization function; This represents the parameter compensation vector generated by the cross-process correction decision model. Represents the optimal compensation vector; This indicates the final quality prediction result; Indicates energy consumption level, Indicates defect risk; , , These are the weighting parameters for the final quality prediction result, energy consumption level, and defect risk, used to balance the three objectives of quality, energy consumption, and risk. Indicates the need to make smallest That is, the optimization objective of this multi-objective optimization function.

[0058] By constructing the aforementioned multi-objective loss function, the originally dispersed quality, energy consumption, and risk indicators are mapped to the same optimization space, enabling the parameter correction process to comprehensively weigh these factors within a single objective function framework. In practical implementation, energy consumption levels can be constructed based on spindle power integrals, feed load models, or cutting force models. Defect risk can be calculated by constructing a defect classification model to obtain confidence probabilities or non-conforming rate estimates. This allows the optimization objective to simultaneously reflect quality results, resource consumption, and potential risks, thereby avoiding the bias problem of the correction process pursuing only quality while neglecting safety and economy.

[0059] The multi-objective optimization function is optimized by using gradient descent, policy gradient, genetic algorithm, Bayesian optimization, or a combination thereof to search for the parameter compensation vector, thereby obtaining the optimal compensation vector that minimizes the multi-objective optimization function. This yields a parameter adjustment scheme that satisfies quality requirements and achieves the best overall balance between energy consumption and defect risk. In the specific solution process, when using gradient-based methods, end-to-end differentiability modeling of the cross-process correction decision model can be performed to make the objective function differentiable with respect to the parameter compensation vector, thus supporting real-time gradient backpropagation. When using gradient-independent methods such as evolutionary algorithms or Bayesian optimization, the parameter compensation space search and surrogate model can be completed offline, and an approximate optimal solution can be quickly output based on an empirical model during online operation. Simultaneously, a multi-policy parallel optimization mechanism is introduced to flexibly select the search method for the optimal compensation vector according to different manufacturing task scales and real-time requirements, thereby achieving a balance between efficiency and accuracy.

[0060] S5. Correct the process parameter vector with the optimal compensation vector, generate parameter commands and send them to the control channel of the component manufacturing equipment in real time to update the process parameters of subsequent manufacturing processes, so as to realize the dynamic control and real-time correction of the component manufacturing equipment.

[0061] As a preferred embodiment of step S5, the specific process includes: S51. Superimpose the optimal compensation vector onto the process parameter vector of the corresponding manufacturing process to generate the corrected process parameter vector, mathematically represented as follows: ; in, , , They represent the first The corrected process parameter vector, process parameter vector, and optimal compensation vector for each manufacturing process.

[0062] The corrected process parameter vector is used in subsequent manufacturing processes to actively absorb and compensate for errors generated in previous processes by finely adjusting feed strategies, depth of cut allocation, or coordinate offsets. The parameter correction process can fine-tune a single parameter or simultaneously apply to multiple parameters to form a combined compensation strategy, distributing the error across multiple machining degrees of freedom. In this way, the originally irreversible quality error from the previous process is transformed into a controllable parameter offset in subsequent processes, realizing the transfer of error from the "result space" to the "parameter space," and transforming the corrective action from post-process rework to in-process absorption.

[0063] S52. Apply safety boundary constraints to the modified process parameter vector, setting lower and upper limits, mathematically represented as follows: ; in, , They represent respectively to The lower and upper limits of the parameters are determined by equipment capacity, tool safety range, and process specifications. This ensures that the compensated parameter combination will not lead to equipment overload, tool damage, or process instability, thus balancing machining safety and process feasibility while ensuring the correction effect. The safety boundary can be configured based on material properties, tool life model, and equipment load curve. When the corrected process parameter vector exceeds the safety boundary range, the corrected process parameter vector is automatically trimmed or rolled back to ensure that the part processing equipment always operates within a controllable and safe process window during real-time correction, thereby improving the executability and reliability of the parameter adjustment scheme in the actual production environment.

[0064] S53. Generate parameter instructions based on the corrected process parameter vector and send them in real time to the control channel of the parts manufacturing equipment through the CNC system, PLC, or host computer interface to update process parameters such as feed rate, depth of cut, tool compensation, and coordinate offset for subsequent manufacturing processes. In specific implementation, the process parameter vector can be updated in batches or in real time. According to the control protocol of the parts manufacturing equipment, the parameters are written into the corresponding registers or control buffers, so that the optimal compensation vector takes effect before the start of the next process or smoothly takes effect during the processing. At the same time, a communication interface can be set up between the equipment control system and the optimization module to detect the issuance status and execution result feedback in real time, ensuring the integrability and executability of the parameter correction strategy in the industrial field.

[0065] S54. During the update of the process parameter vector, a gradual control curve or interpolation trajectory is set to progressively transition the process parameter vector to the corrected vector, avoiding dynamic shocks to the machine tool, increased vibration, or instability in the machining process caused by sudden parameter changes. Key monitoring quantities (such as spindle current, vibration acceleration, and cutting force) are monitored in real time throughout the update process. If abnormal fluctuations or out-of-limit trends are detected, the current parameter update process can be stopped or reverted to a safe parameter configuration to ensure that the correction process operates within stable and controllable boundaries.

[0066] S6. After completing a multi-process manufacturing process in one control cycle, the final quality prediction result generated by the cross-process quality propagation model is compared with the measured quality result of the currently manufactured parts. The cross-process quality propagation model is fine-tuned based on incremental learning. Based on the full-process state vector generated by the cross-process quality propagation model, the final quality prediction result, and the parameter compensation vector generated by the cross-process correction decision model, the cross-process correction decision model is fine-tuned based on reinforcement learning to achieve dynamic optimization of multi-process manufacturing and real-time cross-process quality correction of parts.

[0067] As a preferred embodiment of step S6, the process of correcting the process parameter vector with the optimal compensation vector specifically includes: S61. Calculate the quality prediction error based on the final quality prediction result and the measured quality result. Quality prediction error The error can be a scalar or a vector, used to measure the prediction bias of the current cross-process quality propagation model. When the final quality prediction result is a comprehensive quality index, the quality prediction error can be expressed as a single numerical bias (single dimension). When the final quality prediction result includes multiple dimensions (such as critical dimensions, hole position accuracy, minimum wall thickness, etc.), the quality prediction error can form a multi-dimensional vector to characterize the degree of prediction inaccuracy of the model on each quality sub-index. Through time-series statistics of this error, the stability of the cross-process quality propagation model under different process ranges, different process combinations, and different parameter configurations can be further analyzed, thus providing directional guidance for subsequent updates to the cross-process quality propagation model.

[0068] S62, with the aforementioned quality prediction error Constructing the loss function of the prediction model The mathematical representation is as follows: ; in This indicates the calculation of the L2 norm.

[0069] S63, Based on the loss function of the prediction model Incremental learning is used to fine-tune and update the parameters of the cross-process quality propagation model. The preferred approach is to fine-tune some weights or use efficient parameter fine-tuning strategies to maintain the model's predictive performance under existing conditions while introducing new working condition knowledge, thereby mitigating catastrophic forgetting.

[0070] In practical implementation, the basic structures such as the process embedding layer and cross-process coding layer can be frozen, and only the parameters of the convergence layer or quality prediction head can be updated. Alternatively, regularization terms can be introduced to constrain the deviation between new and historical parameters. This ensures that the cross-process quality propagation model can quickly adapt to new operating conditions while avoiding excessive destruction of learned knowledge. The parameter update process in this step can be performed periodically offline or in batches when online conditions permit, enabling the cross-process quality propagation model to have continuous learning capabilities.

[0071] S64, Each control cycle Corresponding full-process state vector As a reinforcement learning state : ; During this control cycle Execution parameter compensation vector As an action : ; The final quality prediction result for the corresponding period will be used as an immediate reward: ; This constructs reinforcement learning sample triplets. The data is written to the experience cache for fine-tuning of the cross-process correction decision model.

[0072] The experience buffer can be managed using a circular queue or a priority experience replay structure, so that samples with a large amount of error information or a high contribution to quality improvement can be sampled with a higher probability. This accelerates the convergence speed of the strategy for key scenarios. The setting of the instant reward can further introduce quality threshold constraints or risk penalty terms to guide the strategy to perform stable optimization while ensuring that the quality meets the standards.

[0073] S65. Estimating cumulative returns using a discounted return method: ; in, Indicates the current control cycle The cumulative income, For subsequent control cycles The discount factor is used to balance the weights of short-term quality improvement and long-term quality stability; Indicates in Instant rewards for controlling the cycle, It represents infinity.

[0074] S66. Based on this discounted return, the cross-process correction decision model is reinforced through policy gradient, near-end policy optimization, or other reinforcement learning algorithms. By continuously absorbing production data and quality feedback from new control cycles, the parameter adjustment rules are optimized, thereby gradually improving the effectiveness of real-time cross-process quality correction and the adaptability of the equipment control system to different operating conditions, structural components, and environmental disturbances. Through this update mechanism, the correction strategy can form a memory of different process combinations and quality response characteristics during long-term operation, realizing a shift from "experience-dependent" to "data-driven decision-making," enabling the equipment control system to continuously improve correction accuracy and robustness during continuous production.

[0075] This application also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the cross-process quality real-time correction method for multi-process manufacturing.

[0076] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0077] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can 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 processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0078] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for real-time quality correction across multiple manufacturing processes, characterized in that, include: Collect process parameter data, quality inspection data and process metadata corresponding to each manufacturing process of the parts and quantize them. Perform preprocessing and alignment to generate process parameter vectors, quality feature vectors and process metadata vectors. Construct process state triplets for each manufacturing process and form a manufacturing process sequence dataset according to the order of the processes. A cross-process quality propagation model is constructed and pre-trained. The manufacturing process sequence dataset is input into the pre-trained cross-process quality propagation model to model the quality dependency relationship between each manufacturing process, extract the full-process state vector representing the quality status of the entire part manufacturing process, and generate the final quality prediction result. Construct a cross-process correction decision model, and input the full-process state vector and process position code into the cross-process correction decision model to obtain the parameter compensation vector; The process location code includes the current process index and the identifier of the subsequent adjustable process; A multi-objective optimization function is constructed using the final quality prediction result, energy consumption level, and defect risk as joint optimization objectives. The cross-process correction decision model is optimized, and the parameter compensation vector generated by the optimized cross-process correction decision model is used as the optimal compensation vector. The process parameter vector is corrected by the optimal compensation vector, and parameter instructions are generated to update the process parameters of subsequent manufacturing processes, so as to realize dynamic control and real-time correction of the parts manufacturing equipment. After completing a multi-process manufacturing process in a control cycle, the cross-process quality propagation model is fine-tuned based on incremental learning based on the measured quality results of the currently manufactured parts, and the cross-process correction decision model is fine-tuned based on reinforcement learning, so as to realize dynamic optimization of multi-process manufacturing and cross-process quality real-time correction of parts.

2. The method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The process parameter data, quality inspection data, and process metadata corresponding to each manufacturing process of the parts are collected and quantized, preprocessed and aligned to generate process parameter vectors, quality feature vectors, and process metadata vectors, constructing a process state triplet for each manufacturing process, including: A data acquisition interface is set in the equipment control system corresponding to each manufacturing process of the parts to collect process parameter data in real time or near real time, and the collected process parameter data is vectorized to obtain the original process parameter vector. The original process parameter vector includes at least the fields of cutting speed, feed rate, depth of cut, tool compensation amount, path strategy code, machining cycle time, and a uniformly introduced timestamp field, and is recorded with a uniform sampling period and timestamp format. After each manufacturing process is completed, the generated parts are inspected according to the preset quality inspection plan, the quality inspection data is obtained and quantified to form an original quality feature vector; the original quality feature vector is associated with the corresponding timestamp and the unique identifier of the part; the quality inspection data includes at least dimensional deviation, form and position deviation, critical wall thickness, hole position deviation, surface quality grade, and defect type; For each manufacturing process, collect the corresponding process metadata and quantize it to obtain the original process metadata vector; the process metadata includes at least the process type, equipment number and machine tool model, clamping method, clamping posture, workpiece batch number, and ambient temperature and humidity, as well as the process context and equipment and environmental conditions used to describe the manufacturing process when it occurs. The obtained original process parameter vector, original quality feature vector, and original process element information vector are uniformly aligned in time, outlier removal is performed, missing data is imputed, and numerical standardization is performed to obtain process parameter vector, quality feature vector, and process element information vector. Then, the process parameter vector, quality feature vector, and process element information vector of each manufacturing process are concatenated to construct process state triplet.

3. The method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The cross-process quality propagation model includes a process embedding layer, a cross-process encoding layer, a convergence layer, and a quality prediction head; the process embedding layer is a multilayer perceptron or other nonlinear mapping structure; the cross-process encoding layer is a Transformer encoder with a self-attention mechanism or other equivalent cross-sequence modeling structure. The process involves inputting the manufacturing process sequence dataset into a pre-trained cross-process quality propagation model to model the quality dependencies between various manufacturing processes, extracting a full-process state vector representing the quality status of the entire component manufacturing process, and generating the final quality prediction result, including: The process state triplet of each manufacturing process in the manufacturing process sequence dataset is input into the process embedding layer. The process parameter vector, quality feature vector, and process meta-information vector in the process state triplet are jointly encoded through a nonlinear mapping structure to obtain the process vector of each manufacturing process. The process vectors of all manufacturing processes are arranged into a process sequence according to the order of the processes. This sequence is then input into the cross-process coding layer. The mutual influence weights between different processes are calculated through multi-head self-attention, and the process vectors of each manufacturing process in the process sequence are weighted to obtain the hidden state sequence. The hidden state sequence is input into the aggregation layer, and the hidden state vectors of each manufacturing process are aggregated to obtain the full process state vector. Input the entire process state vector into the quality prediction head to obtain the final quality prediction result.

4. The method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The step of inputting the entire process state vector and the process position code into the cross-process correction decision model to obtain the parameter compensation vector includes: The entire process state vector and the process position code are concatenated at the vector level to form a joint state vector, which is then input into the cross-process correction decision model to output parameter compensation vectors for subsequent manufacturing processes. The cross-process correction decision model employs a multilayer perceptron, attention network, or other nonlinear regression structure.

5. A method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The mathematical representation of the constructed multi-objective optimization function is as follows: ； ； in, For the constructed multi-objective optimization function; This represents the parameter compensation vector generated by the cross-process correction decision model. Represents the optimal compensation vector; This indicates the final quality prediction result; Indicates energy consumption level, Indicates defect risk; , , These are the weighting parameters for the final quality prediction result, energy consumption level, and defect risk, respectively. Indicates the need to make smallest That is, the optimization objective of this multi-objective optimization function.

6. The method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The process parameter vector correction using the optimal compensation vector includes: The optimal compensation vector is superimposed onto the process parameter vector of the corresponding manufacturing process to generate the corrected process parameter vector. Set lower and upper limits for the modified process parameter vector. If the modified process parameter vector exceeds the range of these limits, it will be trimmed or rolled back to the state before overlay.

7. A method for real-time quality correction across multiple manufacturing processes according to claim 6, characterized in that: The step of correcting the process parameter vector with the optimal compensation vector also includes: By setting a gradual control curve or interpolation trajectory, the process parameter vector is gradually transitioned to the corrected process parameter vector, avoiding dynamic shocks to the machine tool, increased vibration, or instability in the machining process caused by sudden parameter changes.

8. A method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The fine-tuning of the cross-process quality propagation model based on incremental learning includes: The quality prediction error is calculated based on the final quality prediction result and the measured quality result. ; With the aforementioned quality prediction error Constructing the loss function of the prediction model The mathematical representation is as follows: ； in Represents the calculation of the L2 norm; based on the loss function of the prediction model. Incremental learning is used to fine-tune and update the parameters of the cross-process quality propagation model.

9. A method for real-time quality correction across multiple manufacturing processes according to claim 1, characterized in that: The fine-tuning of the cross-process correction decision model based on reinforcement learning includes: The full-process state vector corresponding to each control cycle is used as the reinforcement learning state, the parameter compensation vector executed in that control cycle is used as the action, and the final quality prediction result in the corresponding cycle is used as the immediate reward, thereby constructing a reinforcement learning sample triplet. Estimating cumulative returns using a discounted return method: ； in, Indicates the current control cycle The cumulative income, For subsequent control cycles Discount factor, Indicates in Instant rewards for controlling the cycle, Indicates infinity; Based on this discount return, reinforcement learning is applied to the cross-process correction decision model to optimize parameter adjustment patterns while continuously absorbing production data and quality feedback from new control cycles.

10. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the cross-process quality real-time correction method for multi-process manufacturing as described in any one of claims 1-9.

Images