Workpiece machining adaptive control method and device, terminal and medium

Abstract

The application discloses a workpiece machining self-adaptive control method and device, a terminal and a medium. The method comprises the following steps: collecting a normally open low-cost signal data, performing time sequence normalization processing on the normally open low-cost signal data, and obtaining first multi-source sensing feature data; performing state estimation based on the first multi-source sensing feature data, obtaining a trusted state, and obtaining a sensing configuration action and a machining candidate action through joint strategy reasoning based on the trusted state; performing safety constraint optimization projection solving based on the machining candidate action, obtaining a safety action, and performing workpiece machining based on the safety action. The application collects normally open low-cost signals, obtains sensing feature data through time sequence normalization, obtains a trusted state through state estimation, obtains a candidate action through joint reasoning, obtains a safety action through safety constraint optimization projection solving, and realizes workpiece machining, which can reduce the sensing hardware cost, ensure the reliability of the machining state identification, and ensure the safety of the action execution.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of numerical control machining process control, intelligent sensing and artificial intelligence control, and particularly relates to a workpiece machining adaptive control method and device, a terminal and a medium. BACKGROUND

[0002] Against the backdrop of the rapid development of high-end equipment manufacturing and precision machining industries, CNC machining, as a core process technology in the mechanical manufacturing field, is widely used in key areas such as aerospace, precision molds, and high-end components. The refined control of machining accuracy, stability, and efficiency has become a core indicator for measuring the level of CNC machining technology and ensuring the quality of high-end equipment. The CNC machining process exhibits typical nonlinear, time-varying, and multi-physics coupling characteristics. Throughout the machining process, the continuous cutting of the tool into the workpiece, the dynamic reduction of the blank material allowance, the real-time changes in the local stiffness of the workpiece, and the gradual accumulation of tool wear result in different cutting forces, vibrations, machining temperatures, and workpiece surface qualities at different machining stages using the same CNC program and fixed process parameters. Traditional CNC machining processes generally rely on process manuals, empirical formulas, or preliminary trial cut results, pre-setting fixed spindle speeds and feed rates before machining. This approach cannot adapt to various dynamic disturbances encountered in actual machining, such as material hardness fluctuations, uneven blank allowances, tool wear, and workpiece clamping deviations. This easily leads to problems like machining chatter, tool overload, workpiece dimensional errors, and substandard surface quality, making it difficult to meet the high precision, high stability, and high yield requirements of modern precision CNC machining. To overcome the limitations of traditional fixed-parameter machining, adaptive machining control technology has become a core research direction for closed-loop optimization in CNC machining. This technology collects machining status signals such as spindle current, cutting force, vibration, and temperature, analyzes the machining conditions in real time, and dynamically adjusts process parameters such as feed rate and spindle speed, or triggers alarms and shutdowns for protection, achieving dynamic control of the machining process. Classical CNC machining adaptive control systems mainly consist of a machine tool body, a CNC controller, sensor units, and an adaptive controller. Based on monitoring data such as cutting torque, tool temperature, and machine tool vibration, they calculate material removal rate, tool wear rate, and machining constraint states, thus initially establishing a closed-loop control architecture for CNC machining, effectively overcoming the static limitations of traditional fixed-parameter machining. However, these traditional adaptive control schemes generally employ fixed sensor combinations and fixed sampling configurations, failing to fully consider the cost of multi-source sensor data acquisition, communication bandwidth consumption, and equipment computational load. With the introduction of multi-dimensional sensor signals such as vibration, acoustic emission, machine vision, and high-frequency cutting forces, continuous full-data acquisition generates a large amount of redundant information, easily causing high load operation of the edge controller, reducing control real-time performance and accuracy, and hindering the improvement of adaptive control effects. In recent years, reinforcement learning, with its advantages of adapting to nonlinear systems and fitting long-term working condition feedback, has been gradually applied to the field of CNC machining adaptive control. It can iteratively learn the mapping relationship between machining states and control actions through long-term machining effects, effectively adapting to the nonlinear characteristics and quality delay feedback problems of CNC machining. However, online reinforcement learning relies on real-time working condition exploration and iteration. Unconstrained strategy output can easily lead to unreasonable actions such as excessive feed rate, large speed fluctuations, and abnormal cooling adjustments, which in turn can cause production risks such as tool overload, machining chatter, workpiece deviation, or even emergency machine tool shutdown.To mitigate the security risks of online trial and error, the industry is gradually exploring offline reinforcement learning solutions. These solutions train control strategies based on massive historical processing logs and quality inspection data accumulated in industrial settings, significantly reducing the real-time trial and error costs of machine tools. However, this approach still has significant shortcomings: it cannot effectively cover processing action scenarios outside the historical data distribution, resulting in uncertainty in strategy output and difficulty in achieving safe iterative updates after strategy deployment. Meanwhile, active sensing technology, with its advantages of on-demand data acquisition and dynamic adaptation, has been maturely applied in fields such as robot control, environmental monitoring, and industrial soft measurement. The core idea is that the decision-making system autonomously determines the acquisition target, acquisition time, and acquisition scale based on the uncertainty of the current task and operating conditions. However, its application in CNC machining is currently relatively superficial. Existing research mainly focuses on sensor placement, optimal observation locations, and data domain selection, without achieving joint optimization of active sensing actions and machining control actions, nor has it established a safe and controllable machining closed-loop deployment system. There is significant room for technological integration and practical application. Currently, several CNC machining adaptive control related technical solutions and research results have been published both domestically and internationally, but all have inherent technical defects, failing to achieve high-precision, high-safety, and low-cost intelligent CNC machining closed-loop control. Among existing mainstream technical solutions, patent CN112180833B identifies cutting forces by collecting spindle current, rotational speed, and feed rate signals, and optimizes the feed rate using a reinforcement learning model, establishing a basic closed loop for signal acquisition, model prediction, and parameter adjustment. However, the sensor monitoring mode and acquisition parameters are all fixed in advance, and sensor start / stop, sampling frequency, and feature processing methods are not included in the decision-making and control scope. It lacks a joint perception control and safety action projection mechanism based on the uncertainty of the working condition. Patent CN113985814B monitors machining signals in real time using a digital twin system, identifies machining anomalies through neural networks, and generates control actions according to preset rules, possessing anomaly identification and hierarchical control capabilities. However, it still adopts a fixed sensor monitoring mode, failing to achieve coordinated optimization of sensor configuration costs, information gain, and machining control, and cannot complete the physical safety constraint projection and closed-loop execution of reinforcement learning candidate actions. US Patent 20200166909A1 constructs a general adaptive control framework for manufacturing processes based on machine learning, which can be adapted to various subtractive machining scenarios. However, it does not form a joint output architecture for machining actions and sensor configuration actions, nor does it propose specific solutions for core issues such as multi-rate sensing in CNC machining, information timeliness, sensor budget, and action safety constraints. In addition, existing academic research has verified the feasibility of using reinforcement learning for machining data screening, industrial offline control, and adaptive regulation in batch manufacturing, but it has not yet built a complete technical system adapted to CNC machining scenarios, integrating dynamic sensor configuration, collaborative control of machining parameters, uncertainty margin compensation, safety projection, and backoff control.

[0003] In summary, existing adaptive control technologies for CNC machining generally suffer from three major technological shortcomings, severely hindering the intelligent, precise, and safe development of CNC machining. First, perception configuration and machining control are disconnected. The entire process relies on a fixed sensor acquisition mode, which easily leads to redundant waste of data and computing resources during stable machining phases, while failing to accurately capture dynamic changes in the working conditions during high-risk machining phases. Information acquisition capabilities cannot dynamically match machining control requirements. Second, reinforcement learning control mechanisms lack security and determinism. Traditional online learning carries high trial-and-error risks, while offline learning suffers from strategy uncertainty and a lack of standardized, independent safety constraint projection mechanisms, failing to guarantee the operational safety of machine tools, cutting tools, workpieces, and processes throughout the entire machining sequence. Third, the ability to fuse multi-source machining data is weak. Machine tool-embedded signals, vibrations, acoustic emissions, and visual signals exhibit asynchronous, multi-rate, and dynamically missing characteristics, resulting in delayed feedback of machining quality indicators.

[0004] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a workpiece machining adaptive control method, device, terminal and medium to address the above-mentioned defects of the prior art. The aim is to solve the problems of fixed configuration of sensing resources, inability to fully collect machining status, easy safety hazards caused by direct online trial and error of reinforcement learning, and the disconnect between machining parameters and sensing strategies during workpiece machining.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an adaptive control method for workpiece machining, wherein the method includes: Normally open low-cost signal data is collected, and time-series normalization processing is performed on the normally open low-cost signal data to obtain the first multi-source sensing feature data; Based on the first multi-source sensor feature data, state estimation is performed to obtain a credible state. Based on the credible state, sensor configuration actions and processing candidate actions are obtained through joint policy reasoning. Based on the candidate machining actions, an optimization projection solution with safety constraints is performed to obtain the safe actions, and the workpiece is machined based on the safe actions.

[0007] In one implementation, the step of acquiring normally open low-cost signal data and performing time-series normalization processing on the normally open low-cost signal data to obtain first multi-source sensing feature data includes: Normally open low-cost signal data is acquired through normally open low-cost signal channels, and timestamps are added to the normally open low-cost signal data of each channel according to a unified clock to obtain the original normally open low-cost signal data of each channel carrying a unified timing mark. A sliding time window is constructed using the control decision cycle as the timing benchmark. The original normally open low-cost signal data of each channel within the window are time-aligned to obtain the basic normally open low-cost signal data of each channel after timing synchronization alignment. Based on the effective observation time of the channel, the data failure type, and the signal operating parameters, the information timeliness, missing identifier, and quality score are calculated respectively. Based on the information timeliness, missing identifier, quality score, and the basic normally open low-cost signal data of each channel, integrated normally open low-cost signal data of each channel with timeliness, missing identifier, and quality information is obtained. The integrated normally open low-cost signal data of each channel is normalized and processed to obtain the first multi-source sensing feature data.

[0008] In one implementation, the step of performing state estimation based on the first multi-source sensing feature data to obtain a reliable state includes: The trusted state is obtained by using the state estimation function to calculate the trusted state at the previous moment, the current first multi-source sensor feature data, the security action at the previous moment, and the current sensor configuration action.

[0009] In one implementation, the step of obtaining the sensing configuration action and processing candidate action based on the trusted state through joint policy reasoning includes: By using a joint reinforcement learning strategy to perform forward reasoning on the trustworthy state, the processing action component and the sensing action component are obtained. The processing action components are quantified to obtain candidate processing actions; Multi-channel parameter assignment and resource constraint feasibility screening are performed on the sensing action components to obtain the sensing configuration action.

[0010] In one implementation, the process of obtaining the sensing configuration action includes: Set the sensor acquisition configuration parameters according to the sensor configuration action, and maintain the preset monitoring level for the forced protection channel.

[0011] In one implementation, the method further includes: Collect on-demand high-cost signal data, perform time-series normalization on the on-demand high-cost signal data, and obtain the second multi-source sensing feature data; Based on the first multi-source sensing feature data and the second multi-source sensing feature data, the target multi-source sensing feature data is obtained. Based on the target's multi-source sensor feature data, the processing state is estimated and the confidence state is updated.

[0012] In one implementation, the step of performing a safety-constrained optimization projection solution based on candidate processing actions to obtain safe actions includes: Construct an optimized projection model that includes multiple process safety constraints; The candidate actions are input into the optimization projection model to perform optimization projection solution, and the actions to be verified are calculated. If the overall working condition meets the preset normal working conditions, then the action to be verified is a safe action.

[0013] Secondly, embodiments of the present invention also provide a workpiece machining adaptive control device, wherein the device includes: The first multi-source sensing feature data acquisition module is used to collect normally open low-cost signal data, perform time-series normalization processing on the normally open low-cost signal data, and obtain the first multi-source sensing feature data. The processing candidate action and sensor configuration action acquisition module is used to perform state estimation based on the first multi-source sensor feature data, obtain a credible state, and obtain the processing candidate action and sensor configuration action based on the credible state through joint policy reasoning. The workpiece machining module is used to perform optimized projection solving with safety constraints based on the candidate machining actions, obtain the safe actions, and perform workpiece machining based on the safe actions.

[0014] Thirdly, embodiments of the present invention also provide a terminal, wherein the terminal includes a memory, a processor, and a workpiece machining adaptive control program stored in the memory and executable on the processor. When the processor executes the workpiece machining adaptive control program, it implements the steps of the workpiece machining adaptive control method described in any of the above schemes.

[0015] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores a workpiece machining adaptive control program, and when the workpiece machining adaptive control program is executed by a processor, it implements the steps of the workpiece machining adaptive control method described in any of the above schemes.

[0016] Beneficial Effects: This invention provides an adaptive control method for workpiece machining. Compared with existing technologies, this invention first acquires normally open, low-cost signal data, performs time-series normalization processing on the normally open, low-cost signal data, and obtains the first multi-source sensor feature data. It abandons the traditional high-cost, high-barrier dedicated sensor data acquisition scheme, and completes basic data acquisition based on low-cost normally open signals, reducing the hardware costs and technical barriers of equipment deployment, data acquisition, and maintenance. Simultaneously, through time-series normalization processing, it cleans, organizes, and standardizes the original messy, time-misaligned, and inconsistent formatted sensor data, effectively avoiding noise in the original data. To address the issues of large data volume, inconsistent timing, and low effectiveness, this method outputs standardized and unified first-source multi-source sensor feature data. This provides high-quality, highly consistent, and highly reliable foundational data support for subsequent state analysis, inference calculations, and action decisions throughout the entire process, ensuring the accuracy and stability of subsequent technical operations from the source. Next, based on the first-source multi-source sensor feature data, state estimation is performed to obtain a reliable state. Then, based on the standardized multi-source feature data, the real-time operating status of equipment, working conditions, and workpieces is deeply mined, overcoming the limitations of existing technologies that rely on single data points and experience to judge the state. This effectively eliminates invalid data interference, corrects state deviations, and filters abnormal operations. This system accurately outputs reliable status information that aligns with actual production scenarios, solving the problems of inaccurate status identification, delayed condition judgment, and low status reliability inherent in traditional technologies. This provides a reliable basis for subsequent strategy reasoning. Based on these reliable statuses, joint strategy reasoning yields sensor configuration actions and candidate processing actions. Compared to traditional single-strategy decision-making, this system achieves coordinated linkage between sensor adjustment and processing actions, enriching the dimensions and coverage of action decisions. It avoids the problems of poor condition adaptability and insufficient action matching caused by single-strategy decisions, providing diversified and scenario-appropriate alternative action schemes for subsequent optimization solutions. Then, based on the candidate machining actions, an optimization projection solution with safety constraints is performed to obtain safe actions. Invalid actions with safety hazards, process deviations, and operational risks are eliminated, and safe actions that take into account safety, accuracy, and efficiency are accurately output. This effectively solves the defects of traditional machining action decision-making, which tends to ignore safety risks, lacks optimization, and has a low fault tolerance rate, and ensures the optimal adaptability of machining actions. Finally, based on the safe actions, the workpiece is machined, which improves the accuracy, stability, and consistency of workpiece machining, reduces the probability of machining errors, equipment failures, and production safety accidents, and effectively improves the qualified rate of machined products and the overall production efficiency.Overall, this invention constructs a fully closed-loop intelligent processing control system encompassing low-cost data acquisition, standardized data processing, precise state perception, multi-strategy joint reasoning, secure optimization solutions, and precise processing execution. It achieves high-precision and high-safety intelligent production control entirely through low-cost hardware. While effectively reducing production hardware investment and maintenance costs, it addresses numerous technical shortcomings of existing technologies, such as poor data quality, inaccurate state recognition, single decision-making dimensions, weak security, unstable processing accuracy, and poor adaptability to operating conditions. It realizes collaborative intelligent control of sensor configuration and workpiece processing, balancing economic efficiency, safety, accuracy, and efficiency in production. This enhances the intelligence level, operating condition adaptability, and production stability of industrial processing, demonstrating strong engineering practicality and broad scenario adaptability. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a specific implementation of the adaptive control method for workpiece machining provided in this embodiment of the invention.

[0018] Figure 2 The flowchart shows the adaptive control method for workpiece machining based on active perception and safety reinforcement learning provided in the embodiments of the present invention.

[0019] Figure 3 This is a diagram illustrating the overall architecture of adaptive control for workpiece machining based on active perception and safety reinforcement learning in the workpiece machining adaptive control method provided in this embodiment of the invention.

[0020] Figure 4 This is a schematic diagram illustrating the safety projection and backoff control principle of the reinforcement learning candidate action in the adaptive control method for workpiece machining provided in this embodiment of the invention.

[0021] Figure 5 This is a schematic diagram of the workpiece machining adaptive control method device provided in the embodiment of the present invention.

[0022] Figure 6 This is a block diagram illustrating the internal structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] Against the backdrop of the rapid development of high-end equipment manufacturing and precision machining industries, CNC machining, as a core process technology in the mechanical manufacturing field, is widely used in key areas such as aerospace, precision molds, and high-end components. The refined control of machining accuracy, stability, and efficiency has become a core indicator for measuring the level of CNC machining technology and ensuring the quality of high-end equipment. The CNC machining process exhibits typical nonlinear, time-varying, and multi-physics coupling characteristics. Throughout the machining process, the continuous cutting of the tool into the workpiece, the dynamic reduction of the blank material allowance, the real-time changes in the local stiffness of the workpiece, and the gradual accumulation of tool wear result in different cutting forces, vibrations, machining temperatures, and workpiece surface qualities at different machining stages using the same CNC program and fixed process parameters. Traditional CNC machining processes generally rely on process manuals, empirical formulas, or preliminary trial cut results, pre-setting fixed spindle speeds and feed rates before machining. This approach cannot adapt to various dynamic disturbances encountered in actual machining, such as material hardness fluctuations, uneven blank allowances, tool wear, and workpiece clamping deviations. This easily leads to problems like machining chatter, tool overload, workpiece dimensional errors, and substandard surface quality, making it difficult to meet the high precision, high stability, and high yield requirements of modern precision CNC machining. To overcome the limitations of traditional fixed-parameter machining, adaptive machining control technology has become a core research direction for closed-loop optimization in CNC machining. This technology collects machining status signals such as spindle current, cutting force, vibration, and temperature, analyzes the machining conditions in real time, and dynamically adjusts process parameters such as feed rate and spindle speed, or triggers alarms and shutdowns for protection, achieving dynamic control of the machining process. Classical CNC machining adaptive control systems mainly consist of a machine tool body, a CNC controller, sensor units, and an adaptive controller. Based on monitoring data such as cutting torque, tool temperature, and machine tool vibration, they calculate material removal rate, tool wear rate, and machining constraint states, thus initially establishing a closed-loop control architecture for CNC machining, effectively overcoming the static limitations of traditional fixed-parameter machining. However, these traditional adaptive control schemes generally employ fixed sensor combinations and fixed sampling configurations, failing to fully consider the cost of multi-source sensor data acquisition, communication bandwidth consumption, and equipment computational load. With the introduction of multi-dimensional sensor signals such as vibration, acoustic emission, machine vision, and high-frequency cutting forces, continuous full-data acquisition generates a large amount of redundant information, easily causing high load operation of the edge controller, reducing control real-time performance and accuracy, and hindering the improvement of adaptive control effects. In recent years, reinforcement learning, with its advantages of adapting to nonlinear systems and fitting long-term working condition feedback, has been gradually applied to the field of CNC machining adaptive control. It can iteratively learn the mapping relationship between machining states and control actions through long-term machining effects, effectively adapting to the nonlinear characteristics and quality delay feedback problems of CNC machining. However, online reinforcement learning relies on real-time working condition exploration and iteration. Unconstrained strategy output can easily lead to unreasonable actions such as excessive feed rate, large speed fluctuations, and abnormal cooling adjustments, which in turn can cause production risks such as tool overload, machining chatter, workpiece deviation, or even emergency machine tool shutdown.To mitigate the security risks of online trial and error, the industry is gradually exploring offline reinforcement learning solutions. These solutions train control strategies based on massive historical processing logs and quality inspection data accumulated in industrial settings, significantly reducing the real-time trial and error costs of machine tools. However, this approach still has significant shortcomings: it cannot effectively cover processing action scenarios outside the historical data distribution, resulting in uncertainty in strategy output and difficulty in achieving safe iterative updates after strategy deployment. Meanwhile, active sensing technology, with its advantages of on-demand data acquisition and dynamic adaptation, has been maturely applied in fields such as robot control, environmental monitoring, and industrial soft measurement. The core idea is that the decision-making system autonomously determines the acquisition target, acquisition time, and acquisition scale based on the uncertainty of the current task and operating conditions. However, its application in CNC machining is currently relatively superficial. Existing research mainly focuses on sensor placement, optimal observation locations, and data domain selection, without achieving joint optimization of active sensing actions and machining control actions, nor has it established a safe and controllable machining closed-loop deployment system. There is significant room for technological integration and practical application. Currently, several CNC machining adaptive control related technical solutions and research results have been published both domestically and internationally, but all have inherent technical defects, failing to achieve high-precision, high-safety, and low-cost intelligent CNC machining closed-loop control. Among existing mainstream technical solutions, patent CN112180833B identifies cutting forces by collecting spindle current, rotational speed, and feed rate signals, and optimizes the feed rate using a reinforcement learning model, establishing a basic closed loop for signal acquisition, model prediction, and parameter adjustment. However, the sensor monitoring mode and acquisition parameters are all fixed in advance, and sensor start / stop, sampling frequency, and feature processing methods are not included in the decision-making and control scope. It lacks a joint perception control and safety action projection mechanism based on the uncertainty of the working condition. Patent CN113985814B monitors machining signals in real time using a digital twin system, identifies machining anomalies through neural networks, and generates control actions according to preset rules, possessing anomaly identification and hierarchical control capabilities. However, it still adopts a fixed sensor monitoring mode, failing to achieve coordinated optimization of sensor configuration costs, information gain, and machining control, and cannot complete the physical safety constraint projection and closed-loop execution of reinforcement learning candidate actions. US Patent 20200166909A1 constructs a general adaptive control framework for manufacturing processes based on machine learning, which can be adapted to various subtractive machining scenarios. However, it does not form a joint output architecture for machining actions and sensor configuration actions, nor does it propose specific solutions for core issues such as multi-rate sensing in CNC machining, information timeliness, sensor budget, and action safety constraints. In addition, existing academic research has verified the feasibility of using reinforcement learning for machining data screening, industrial offline control, and adaptive regulation in batch manufacturing, but it has not yet built a complete technical system adapted to CNC machining scenarios, integrating dynamic sensor configuration, collaborative control of machining parameters, uncertainty margin compensation, safety projection, and backoff control.

[0025] In summary, existing adaptive control technologies for CNC machining generally suffer from three major technological shortcomings, severely hindering the intelligent, precise, and safe development of CNC machining. First, perception configuration and machining control are disconnected. The entire process relies on a fixed sensor acquisition mode, which easily leads to redundant waste of data and computing resources during stable machining phases, while failing to accurately capture dynamic changes in the working conditions during high-risk machining phases. Information acquisition capabilities cannot dynamically match machining control requirements. Second, reinforcement learning control mechanisms lack security and determinism. Traditional online learning carries high trial-and-error risks, while offline learning suffers from strategy uncertainty and a lack of standardized, independent safety constraint projection mechanisms, failing to guarantee the operational safety of machine tools, cutting tools, workpieces, and processes throughout the entire machining sequence. Third, the ability to fuse multi-source machining data is weak. Machine tool-embedded signals, vibrations, acoustic emissions, and visual signals exhibit asynchronous, multi-rate, and dynamically missing characteristics, resulting in delayed feedback of machining quality indicators.

[0026] To address the aforementioned issues, this embodiment provides an adaptive control method for workpiece machining. Specifically, this embodiment first acquires normally open, low-cost signal data, performs time-series normalization processing on this data, and obtains first multi-source sensor feature data. This method abandons the traditional high-cost, high-barrier dedicated sensor data acquisition scheme, and completes basic data acquisition based on low-cost normally open signals, reducing the hardware costs and technical barriers to equipment deployment, data acquisition, and maintenance. Simultaneously, time-series normalization processing cleans, organizes, and standardizes the original messy, time-misaligned, and inconsistent formatted sensor data, effectively avoiding the problems associated with the original data. To address the issues of high noise, chaotic timing, and low effectiveness, this method outputs standardized and unified first-source multi-source sensor feature data. This provides high-quality, highly consistent, and highly reliable foundational data support for subsequent state analysis, inference calculations, and action decisions throughout the entire process, ensuring the accuracy and stability of subsequent technical operations from the source. Next, based on the first-source multi-source sensor feature data, state estimation is performed to obtain a reliable state. Then, based on the standardized multi-source feature data, the real-time operating status of equipment, working conditions, and workpieces is deeply mined. This overcomes the limitations of existing technologies that rely on single data points and experience to judge the state, effectively eliminating invalid data interference, correcting state deviations, and filtering out... By accurately outputting reliable operating condition information that aligns with actual production scenarios, this approach solves the problems of inaccurate state identification, delayed operating condition judgment, and low state reliability inherent in traditional technologies. This provides a reliable basis for subsequent strategy reasoning. Subsequently, based on these reliable states, joint strategy reasoning yields sensor configuration actions and candidate processing actions. Compared to traditional single-strategy decision-making, this approach achieves coordinated linkage between sensor adjustment and processing actions, enriching the dimensions and coverage of action decisions. It avoids the problems of poor operating condition adaptability and insufficient action matching caused by single-strategy decisions, providing diversified and scenario-appropriate alternative actions for subsequent optimization solutions. The process involves several steps. First, based on candidate machining actions, a safety-constrained optimization projection solution is performed to obtain safe actions. Invalid actions with safety hazards, process deviations, and operational risks are eliminated, and safe actions that balance safety, accuracy, and efficiency are accurately output. This effectively solves the shortcomings of traditional machining action decision-making, such as neglecting safety risks, insufficient optimization, and low fault tolerance, ensuring optimal adaptability of machining actions. Finally, based on the safe actions, workpiece machining is performed, improving the accuracy, stability, and consistency of workpiece machining, reducing the probability of machining errors, equipment failures, and production safety accidents, and effectively improving the pass rate of finished products and overall production efficiency.Overall, this invention constructs a fully closed-loop intelligent processing control system encompassing low-cost data acquisition, standardized data processing, precise state perception, multi-strategy joint reasoning, secure optimization solutions, and precise processing execution. It achieves high-precision and high-safety intelligent production control entirely through low-cost hardware. While effectively reducing production hardware investment and maintenance costs, it addresses numerous technical shortcomings of existing technologies, such as poor data quality, inaccurate state recognition, single decision-making dimensions, weak security, unstable processing accuracy, and poor adaptability to operating conditions. It realizes collaborative intelligent control of sensor configuration and workpiece processing, balancing economic efficiency, safety, accuracy, and efficiency in production. This enhances the intelligence level, operating condition adaptability, and production stability of industrial processing, demonstrating strong engineering practicality and broad scenario adaptability.

[0027] The adaptive control method for workpiece machining provided in this embodiment can be applied to smart terminals, such as... Figure 1 As shown, the specific steps include the following: Step S100: Collect normally open low-cost signal data, perform time-series normalization processing on the normally open low-cost signal data, and obtain the first multi-source sensing feature data.

[0028] In this embodiment, firstly, normally open low-cost signal data acquisition is performed. This step captures raw signal data output from various low-cost normally open sensing devices. Based on the stability, low failure rate, and compatibility with low-cost hardware characteristics of normally open signals, it can completely acquire the raw sensing information corresponding to the scenario while controlling the cost of data acquisition hardware and reducing the barriers to equipment deployment and maintenance. This provides real and effective raw data support for subsequent feature processing, while avoiding the high cost and high loss problems caused by complex signal acquisition, and is suitable for large-scale, lightweight sensing data acquisition scenarios. On this basis, the acquired normally open low-cost signal data undergoes time-series normalization processing. This unifies the fragmented and time-disordered raw signal data format and time dimension standard, correcting problems such as timestamp offsets, uneven data sampling intervals, and chaotic time-series arrangement in the raw acquired data. This transforms the disordered raw signal data into a standardized data form with continuous time dimension, unified sampling rules, and clear arrangement logic, ultimately generating the first multi-source sensing feature data. The overall process employs a closed-loop operation of first collecting data and then standardizing it. This approach retains the cost-effectiveness advantage of acquiring low-cost, always-on sensor data while resolving issues such as poor temporal consistency of raw sensor data and its inability to be directly used for multi-source data fusion, feature analysis, and model computation. It effectively transforms raw, coarse-grained data into standardized, structured sensor feature data, providing high-quality, directly accessible foundational data support for core business tasks such as in-depth mining, state analysis, and precise judgment of multi-source sensor data. This enhances the accuracy and reliability of subsequent data applications.

[0029] Specifically, step S100 includes the following steps: Step S101: Acquire normally open low-cost signal data through normally open low-cost signal channels, and add timestamps to the normally open low-cost signal data of each channel according to a unified clock to obtain the original normally open low-cost signal data of each channel carrying a unified timing mark. Step S102: Using the control decision cycle as the timing reference, a sliding time window is built, and the original normally open low-cost signal data of each channel within the window is time-aligned to obtain the basic normally open low-cost signal data of each channel after timing synchronization alignment. Step S103: Calculate information timeliness, missing data identifier, and quality score based on the effective observation time of the channel, data failure type, and signal operating parameters respectively. Based on the information timeliness, missing data identifier, quality score, and basic normally open low-cost signal data of each channel, obtain integrated normally open low-cost signal data of each channel with timeliness, missing data identifier, and quality information. Step S104: Regularize the integrated normally open low-cost signal data of each channel to obtain the first multi-source sensing feature data.

[0030] In one implementation, firstly, as Figure 2 As shown, the machining program, workpiece model, tool parameters, and safety constraints are loaded. After loading the parameters and model, the process enters the physical machining layer, whose structure is as follows: Figure 3 As shown, the process consists of a CNC machine tool, a CNC system, and a tool-workpiece machining process. The CNC system controls the CNC machine tool to drive the spindle, feed axis, and cooling actuators, performing cutting operations on the tool and workpiece according to machining parameters to remove material and generate machining states such as cutting force, vibration, temperature, and deformation. Subsequently, normally open low-cost signals are acquired through the normally open low-cost signal channel of the multi-source active sensing layer. The normally open low-cost signal channel preferably directly reads existing data from the CNC system or driver, including spindle current, spindle power, spindle speed, feed rate, axis position, servo following error, feed motor current, program segment number, and alarm status, completing the normally open signal acquisition and forming initial observations. Subsequently, the edge intelligent control layer performs multi-rate data synchronization and feature fusion on the collected normally open low-cost signal data. Specifically, the edge intelligent controller adds timestamps to the normally open low-cost signal data of each channel according to a unified clock, obtaining the original normally open low-cost signal data of each channel carrying a unified time sequence mark. Then, a sliding time window is built using the control decision cycle as the time sequence benchmark. For signals with different sampling frequencies, the original normally open low-cost signal data of each channel within the window is time-aligned using hold, interpolation, or model prediction methods, resulting in the basic normally open low-cost signal data of each channel after time synchronization and alignment. Next, the information timeliness is calculated based on the effective observation time of each sensing channel, the data failure type, and the signal operating parameters. Missing identifier Quality rating The information timeliness satisfies the expression In the formula For the first The information timeliness characterizes the time interval between the current moment and the most recent valid observation moment of the signal. The missing information identifier can distinguish different data failure situations such as sensor shutdown, communication packet loss, and abnormal invalidity. The quality score is determined by the signal-to-noise ratio, saturation state, calibration state, or abnormal detection results. Based on the information timeliness, missing information identifier, quality score, and basic normally open low-cost signal data of each channel, integrated normally open low-cost signal data of each channel with timeliness, missing information identifier, and quality information is generated. Finally, the integrated normally open low-cost signal data of each channel is normalized to obtain the first multi-source sensor feature data. The normalization process includes uniformly encapsulating the signals of all channels with attribute labels into a multi-source sensor dataset in the form of structured vectors. Based on the above time synchronization and data quality labeling process, it is possible to obtain time-uniform sensor features with complete and reliable information without the need to deploy a large number of additional sensors and relying on the machine tool's built-in normally open signals, providing a reliable input basis for subsequent processing status perception and processing control decisions.

[0031] Step S200: Based on the first multi-source sensing feature data, perform state estimation to obtain a credible state, and based on the credible state, obtain sensing configuration actions and processing candidate actions through joint policy reasoning.

[0032] In this embodiment, firstly, state estimation is performed based on the first multi-source sensor feature data to obtain a reliable state with high credibility that closely matches the actual working conditions. This provides accurate and reliable core data for subsequent action decisions, avoiding decision-making biases caused by data distortion and state misjudgment from the source. Based on the obtained reliable operating state, a joint strategy reasoning mechanism is further used, combined with a multi-dimensional strategy system including scenario operation rules, process standards, and equipment parameter thresholds, to conduct in-depth analysis of the current reliable operating state. This matches sensor configuration optimization actions and processing candidate actions that are adapted to the scenario requirements. This not only enables dynamic adaptation and precise calibration of sensor acquisition parameters, ensuring the accuracy and effectiveness of subsequent sensor data acquisition, but also provides scientific and feasible operational alternatives for the production and processing stages, supporting adaptive optimization of production conditions, stable equipment operation, and improved processing accuracy. The overall process forms a complete closed loop from state assessment to action decision-making. It improves the comprehensiveness and accuracy of state perception by integrating multi-source data, and realizes intelligent and precise decision-making through strategic joint reasoning. It effectively breaks down the barriers between sensing and production execution, improves the system's adaptive capability, working condition control accuracy and production operation stability, and adapts to complex and ever-changing production operation scenarios.

[0033] Specifically, step S200 includes the following steps: Step S201: Calculate the credible state by using the state estimation function on the previous credible state, the current first sensor feature data, the previous processing action, and the current sensor configuration action; Step S202: Perform forward reasoning on the trustworthy state through a joint reinforcement learning strategy to obtain the processing action component and the sensing action component; Step S203: Quantify and characterize the processing action components to obtain processing candidate actions; Step S204: Assign multi-channel parameters and screen resource constraints for the sensing action components to obtain the sensing configuration action.

[0034] In one implementation, since variables such as tool wear, local material properties, actual clamping stiffness, and workpiece internal stress cannot be directly and completely observed, this embodiment models the machining process as a partially observable Markov decision process. The system continuously maintains a reliable state. This is used to characterize the estimation and uncertainty of hidden machining states under current observation conditions. A reliable state includes at least the current program segment and tool position, spindle and feed status, cutting load, vibration or chatter indicators, tool wear estimation, workpiece local residual material and stiffness, temperature and thermal drift estimation, enabled sensor status, timeliness of various signal information, computational and communication load, and prediction confidence interval. For example... Figure 2 and Figure 3 As shown, the process confidence state estimation and uncertainty calculation are performed. Specifically, the state estimation function is used. According to the formula: Regarding the credible state at the previous moment The current first multi-source sensor feature data Safety actions at the previous moment and current sensor configuration actions Perform calculations to obtain a reliable state Among them, the state estimation function Kalman filtering, particle filtering, Bayesian networks, recurrent neural networks, state-space neural networks, or combinations of these methods can be used. Meanwhile, according to the formula... Calculate the state uncertainty, function The calculation can be performed using covariance trace, prediction variance, model ensemble divergence, information entropy, confidence interval width, or nearest neighbor distance.

[0035] Based on the obtained trustworthy state, a joint reinforcement learning strategy is implemented to simultaneously generate processing candidate actions and sensor configuration actions. The specific execution process is as follows: First, the trustworthy state... As a joint reinforcement learning strategy The input is used for forward reasoning to obtain machining motion components and sensing motion components. Then, the machining motion components are quantified and represented within the strategy to obtain candidate machining actions. Specifically, the preferred machining actions adopt a multiplier of the CNC program reference value or bounded residual correction to maintain the stability of the original process planning. The machining actions satisfy... In the formula This indicates spindle speed or speed ratio correction. This indicates a feed rate correction. Indicates cooling flow or pressure correction. This indicates the tool position, path offset, or attitude correction for subsequent machining stages. Different embodiments may select one or more of these control variables. Simultaneously, within the strategy, multi-channel parameter assignment and resource constraint feasibility screening are performed on the sensing motion components to obtain the sensing configuration action. Specifically, the sensing action satisfies… In the formula Indicates the first The start / stop or operating level of each sensor channel Indicates the sampling frequency level. Indicates the sampling window length. This represents the feature extraction or transmission mode; sensing actions can be discrete, multiple discrete, continuous, or hybrid actions. For scenarios with limited bandwidth and computing resources, sensing actions need to meet resource constraints. In the formula Indicates the first Resource consumption of the channel under a given configuration This represents the currently available sensing, communication, or computing budget. Sensing actions primarily affect the observation quality at subsequent time steps, and their benefits are reflected in subsequent state estimation errors, quality risks, and processing rewards, thus achieving the coupling of sensing and control. Ultimately, the output is defined as... The joint operation, in which To process candidate actions, The above-mentioned joint reinforcement learning strategy can adopt a network architecture with a shared state encoder and two sets of independent action heads, or a collaborative architecture of processing policy agents and sensing policy agents that can interact and communicate with each other. The strategy is accompanied by a value network to estimate the long-term processing benefits under a credible state. Iterative optimization is carried out based on reinforcement learning algorithms that can handle mixed action spaces. Optional algorithms include proximal policy optimization, soft actor-commentator, deep deterministic policy gradient, hierarchical reinforcement learning, etc. After iterative convergence, the joint reinforcement learning strategy can be based on a single credible state input, fully embedding the entire process of forward inference, action decomposition, processing parameter quantization, sensing parameter assignment, and resource constraint verification, and stably outputting processing candidate actions that meet the process stability requirements and sensing configuration actions that meet the system resource limits.

[0036] After receiving the sensor configuration action, the sensor acquisition configuration parameters are set according to the action, while the forced protection channel maintains a preset monitoring level (minimum monitoring level). Specifically, the sensor configuration interface of the multi-source active sensing layer receives the sensor configuration action, performs enable or disable operations on each sensor channel, configures the discrete or continuous sampling frequency, window length, filtering parameters, feature extraction level, and raw data / feature data transmission method, and sets minimum hold time and cooling time during sensor switching to avoid hardware shocks and status jitter caused by frequent start-stop operations. For protection sensors that must operate continuously, the sensor configuration action only allows adjustment of the sampling or transmission level and does not support channel shutdown. The sensor action participates in processing decisions by changing the type, accuracy, timeliness, and acquisition cost of subsequent observation information, achieving coordinated optimization of sensing resource allocation and processing parameter adjustment.

[0037] In one implementation, such as Figure 2 As shown, this method also includes: acquiring on-demand high-cost signal data within a specified waiting time or window, fusing it with normally open low-cost signal data, and updating the reliability state again if necessary. The on-demand high-cost signal channels include vibration, acoustic emission, cutting force, temperature, acoustic, machine vision, or in-machine measurement signals. The meaning of "high cost" can be characterized by data bandwidth, sampling energy consumption, computation time, sensor lifespan, installation and maintenance costs, or impact on machining cycle time, and is not limited to electrical energy costs. Specifically, such as... Figure 3 As shown, on-demand high-cost signal data is collected through the on-demand high-cost signal channel in the multi-source active sensing layer, and the data is time-series normalized using time synchronization and feature fusion of the edge intelligent control layer. Specifically: a timestamp is added to the on-demand high-cost signal data of each channel according to a unified clock, resulting in the original on-demand high-cost signal data of each channel carrying a unified time series marker; a sliding time window is built using the control decision cycle as the time series benchmark, and time-domain, frequency-domain, or time-frequency-domain features are calculated for the original on-demand high-cost signal data of each channel within the window, resulting in the basic on-demand high-cost signal data of each channel after time series synchronization and alignment; information timeliness is calculated based on the effective observation time of the channel, data failure type, and signal operating parameters. Missing identifier Quality rating The information timeliness satisfies the expression In the formula For the first The information timeliness characterizes the time interval between the current moment and the most recent valid observation moment of the signal. The missing information flag distinguishes different data failure scenarios such as sensor shutdown, communication packet loss, and abnormal invalidity. The quality score is determined by the signal-to-noise ratio, saturation state, calibration state, or anomaly detection results. Based on the information timeliness, missing information flag, quality score, and basic on-demand high-cost signal data of each channel, integrated on-demand high-cost signal data of each channel with timeliness, missing information flag, and quality information is obtained. The integrated on-demand high-cost signal data of each channel is normalized to obtain the second multi-source sensor feature data. The first multi-source sensor feature data and the second multi-source sensor feature data are fused to obtain the target multi-source sensor feature data in the form of a structured vector. Processing state estimation is performed based on the target multi-source sensor feature data, and the confidence state is updated again. Based on the updated confidence state, sensor configuration actions and processing candidate actions are obtained through joint policy reasoning, providing a unified and reliable input basis for overall decision-making.

[0038] Step S300: Based on the candidate machining actions, perform optimization projection with safety constraints to obtain the safe actions, and perform workpiece machining based on the safe actions.

[0039] In this embodiment, firstly, an optimization projection solution with safety constraints is performed based on candidate machining actions. This allows for comprehensive screening and optimization of all candidate machining actions, eliminating unreasonable actions that pose safety hazards, do not comply with equipment operating specifications, or exceed machining condition limits. Simultaneously, compliant actions are iteratively optimized by considering core optimization objectives such as machining accuracy, machining efficiency, and equipment load. Safe actions that are suitable for the current machining scenario and possess both safety and optimal machining performance are selected. This avoids various safety issues such as equipment failure, workpiece damage, and operational risks during machining from the source, ensuring the feasibility and standardization of subsequent machining actions. After obtaining standardized and safe optimal machining actions, formal workpiece machining operations are carried out based on these safe actions. This allows the entire machining process to follow safety constraints and optimized action logic, effectively avoiding machining deviations, safety accidents, and workpiece scrap caused by arbitrary human operation and poor adaptability of candidate actions. It also ensures stable control of workpiece machining accuracy and quality, improving the overall stability and reliability of the machining process. The overall process achieves dual control over the safety constraints and performance optimization of processing actions through a closed-loop logic of first optimizing the solution and then implementing the processing. This not only eliminates various safety risks in the processing process, but also maximizes the accuracy, efficiency and finished product qualification rate of the workpiece processing, making the entire workpiece processing process more standardized, intelligent and safe, effectively balancing processing safety and production efficiency.

[0040] Specifically, step S300 includes the following steps: Step S301: Construct an optimized projection model that includes multiple process safety constraints; Step S302: Input the candidate processing action into the optimization projection model to perform optimization projection solution and calculate the action to be verified; Step S303: If the overall working condition meets the preset normal working condition, then the action to be verified is a safe action.

[0041] In one implementation, such as Figure 2 As shown, a safe projection is performed based on the predicted state, constraints, and uncertainties, i.e., as follows: Figure 3 The safety projection and rollback control are shown. Specifically, as... Figure 4 As shown, based on the aforementioned reinforcement learning strategy outputting candidate machining actions containing parameters such as spindle speed, feed rate, and cooling parameters, and collecting information such as cutting force, chatter, temperature, workpiece deformation, and tool status, state prediction and uncertainty solving are completed. A safety constraint set, composed of parameter boundaries, rate of change, cutting force and spindle power, upper limit of tool temperature, servo following error limit, and machine tool protection rules, is integrated. These three types of inputs are then fed into a safety projection solver for processing. This solver constructs a quadratic programming-based optimization projection model with the optimization objective of minimizing the correction amount for candidate actions. The specific model is expressed by the formula: The executable actions to be verified after safe projection are obtained by solving, where W is the action correction weight matrix. During the optimization process, three types of hard constraints are applied simultaneously: the first type is the boundary constraint of the processing parameter amplitude. The first type is used to limit the reasonable range of values ​​for parameters such as rotational speed and feed; the second type is the constraint on the maximum rate of change of action within a single decision cycle. To avoid machine tool shocks and machining instability caused by sudden parameter changes, among other things... and These are the upper and lower limits of the processing parameters. The third category is the maximum allowable change within a unit decision-making cycle; the fourth category is process constraints with embedded uncertainty safety margins. ,in Representing the The process safety constraints cover actual production limitations such as predicted cutting force not exceeding tool and workpiece load limits, spindle operating power not exceeding the rated ratio, cutting chatter probability below the safety threshold, tool operating temperature not exceeding limits, servo system following error not exceeding standards, and predicted workpiece deformation not exceeding the remaining machining tolerance. This is the predicted value of the system state. The prediction uncertainty of the corresponding state variables, As an uncertainty safety margin, it can compensate for state prediction deviations and improve the robustness of the constraint boundary.

[0042] After completing the safe projection solution with margin constraints, the process enters the credibility and feasibility gating stage. This stage verifies whether a feasible solution exists for the projection optimization problem and whether the credibility of the reinforcement learning strategy output reaches a preset threshold. If both conditions are met, the overall solution process is considered to be under preset normal conditions. The verified actions are then designated as safe actions (safe machining parameters) and sent to the CNC system, which performs workpiece cutting based on these safe actions. If either gating condition is not met—i.e., the projection has no feasible solution or the strategy credibility is insufficient—a rollback logic is executed, outputting a rollback action. This action may involve switching to the basic controller, reducing feed rates or other conservative machining parameters, or maintaining the verified safe actions from the previous cycle. The basic controller can be the original CNC program settings, a constant cutting force controller, a proportional-integral controller, a model predictive controller, or a rule-based controller verified by process engineers. Furthermore, the numerical solution for the safe projection can be flexibly implemented using quadratic programming, model predictive control, control barrier functions, reachability set constraints, rule pruning, or a combination of lookup tables and optimization algorithms. Furthermore, the system automatically triggers a rollback control mechanism when any of the following abnormal situations occur: timeout in projection solving, no feasible solution to the optimization problem, severely outdated sensor data, failure of key detection sensors, input of reinforcement learning strategy exceeding the distribution range of model training data, or policy credibility falling below a set threshold. The entire process covers two closed-loop logic branches: constraint optimization projection solving with uncertainty margin, dual-condition gating judgment, and normal safe execution and abnormal degradation rollback. It can not only optimize the machining process based on reinforcement learning, but also comprehensively ensure the operational safety of machine tools, cutting tools, and workpieces through multi-layered physical rules, dynamic limits, predictive uncertainty constraints, and fallback degradation strategies.

[0043] In one implementation, after completing the credibility and feasibility gating judgment stage, such as Figure 2 As shown, the instant reward for each step is calculated and the current processing segment is stored. The instant reward is calculated using the formula: Quantitative calculations are performed, among which This refers to cutting load deviation or load fluctuation. For vibration / flutter risk, For processing time increment, For energy consumption, For tool wear increment, For sensing costs, For communication and computing costs, For safety projection corrections, rollback events, or constraint proximity. Weighting coefficients before each term. It can be flexibly configured according to workpiece type, processing stage, and product quality level. To incentivize the system to actively sense the processing status, an information gain term can be added to the reward formula. Furthermore, information gain cannot be maximized independently of the processing quality objective; it needs to be jointly evaluated with quality risk and sensor investment costs to avoid resource waste caused by indiscriminately activating all sensors.

[0044] After completing the immediate reward calculation, determine whether the current processing segment or workpiece has been fully processed. If not, jump back to the "Perform credible state estimation and uncertainty calculation" step. If already completed, such as... Figure 2 As shown, the process involves fusing delayed quality feedback such as size and roughness to update the dataset, validating the limited online update in the digital twin, releasing the strategy after passing the verification, and then jumping back to the "acquiring normally open signals and forming initial observations" step. That is, executing... Figure 3The offline training and slow update of actions in the layers are implemented. Specifically, firstly, based on processing and quality feedback, an offline dataset is constructed from historical processing records, quality inspection data, and digital twin data. Each sample in the offline dataset contains a credible state, processing action, sensor configuration, next observation, immediate process indicators, terminal quality label, safety projection result, and anomaly marker. For cases where sensor configuration is not recorded in historical data, sensor actions can be reconstructed based on the actual sensors used at the time, sampling frequency, and data availability. Then, based on the offline dataset, conservative offline reinforcement learning training is performed to obtain a policy. The training process employs behavioral constraints or conservative value estimation schemes to weaken the model's overestimation of actions outside the data distribution domain. Specifically, a penalty term for actions that have not occurred can be added to the value function optimization objective, or the deviation distance between the new policy and the historical behavior policy can be constrained after iteration. At the same time, the state prediction model adopts an integrated architecture composed of multiple sets of independently initialized models or probabilistic models, and the uncertainty of the processing state is quantified by the degree of divergence of the output results of different models. Subsequently, digital twin replay and candidate strategy verification are conducted. The strategies trained offline must undergo replay testing in the digital twin, covering different materials, allowances, tool wear, sensor failures, and communication loads. Only when the safety violation rate, quality risk, and performance indicators meet the release threshold is a deployable strategy version generated. Then, based on the deployable strategy version, restricted online updates and version releases are performed. Specifically, a multi-timescale structure is preferred: the machine tool's original servo control and emergency protection operate at the fastest cycle; safety projection and machining parameter updates operate at a relatively fast cycle; sensor configuration updates are performed at a slower cycle based on state changes; and strategy parameter updates are executed at the end of a machining segment, the end of a workpiece, or the end of a batch. For example, the servo control cycle can be in the millisecond range, the joint decision-making cycle can be tens to hundreds of milliseconds, the sensor configuration retention cycle can be hundreds of milliseconds to several seconds, and the strategy update cycle can be a machining segment or a workpiece, but these ranges are not limitations. Online updates use only trajectories that have passed data quality checks and are subject to at least one of the following limitations: the distance between the old and new policies does not exceed a threshold; the change in policy parameters does not exceed a threshold; only the last few layers are updated; updates are only released when digital twin playback performance improves and the security violation rate does not increase; parallel evaluation is conducted using dual-version or shadow mode; and the previous stable version is automatically restored when a policy anomaly occurs. Furthermore, the obtained policy parameters can be used to further update sensor actions.

[0045] The adaptive control method for workpiece machining proposed in this invention constructs a complete intelligent machining control system integrating dynamic active sensing, reliable state construction, joint decision-making, safety constraint execution, robust strategy deployment, and multi-scale closed-loop updates. To optimize sensing resource allocation, this invention employs an active sensing method based on state uncertainty and machining risk: basic monitoring is maintained using normally open, low-cost internal signals; based on state uncertainty, machining area characteristics, quality risk, and resource occupancy, on-demand sensing channels such as vibration, acoustic emission, cutting force, and vision are dynamically activated, and sampling and processing levels are adjusted; simultaneously, constraints such as bandwidth, computational load, energy consumption, minimum hold time, and minimum monitoring level for protection channels are set to avoid continuous full-scale data acquisition or frequent sensor switching. Simultaneously, a reliable state construction method for asynchronous, multi-rate, and incomplete observations is proposed: unified time stamping and multi-rate fusion are applied to multi-source data such as CNC system signals, vibration, acoustic emission, cutting force, temperature, and vision, and each signal is supplemented with information on timeliness, missing data, and quality scores. Combining historical states, machining actions, sensor configurations, and workpiece material removal states, a reliable state containing state estimation results and their uncertainties is formed, providing a unified and reliable decision input for the joint strategy. Furthermore, through a joint decision-making mechanism of machining actions and sensor actions, machining actions and sensor actions are simultaneously defined within the same machining control strategy. Machining actions include control variables such as spindle speed, feed rate, cooling parameters, tool offset, or tool posture, while sensor actions include sensor start / stop, operating level, sampling frequency, sampling window, feature extraction mode, and data transmission mode. Sensor actions participate in machining decisions by changing the type, accuracy, timeliness, and acquisition cost of subsequent observation information, achieving coordinated optimization of sensing resource allocation and machining parameter adjustment. To address the safety shortcomings of reinforcement learning in machining control, an independent safety projection and rollback mechanism for reinforcement learning candidate actions is implemented: Machining candidate actions generated by the reinforcement learning strategy are not directly sent to the CNC system. Instead, they are safely projected based on constraints such as machining parameter boundaries, action change rate, cutting force, spindle power, chatter risk, tool temperature, servo error, and workpiece deformation. A safety margin corresponding to the state prediction uncertainty is added to the safety constraints. When the safety projection has no feasible solution, the solution times out, key sensors fail, the strategy input exceeds the training distribution, or the strategy's credibility is insufficient, the system automatically switches to the basic controller, the previous safe action, or a conservative machining mode. At the strategy deployment level, a strategy deployment method combining conservative offline training, digital twin verification, and restricted online updates is adopted: an offline dataset is constructed using historical processing records, trial cutting data, digital twin trajectories, and processing quality results; the overestimation of unseen actions by the strategy is reduced through behavioral constraints, conservative value estimation, or out-of-distribution action penalties; the offline strategy is deployed after multi-condition playback and security verification in the digital twin environment; in the online phase, updates are only made when the strategy changes, performance improvements, and security violation rates meet the specified conditions; and the controllability of the update process is ensured through version management, shadow running, and automatic rollback.In addition, a multi-timescale closed-loop execution and delayed quality feedback update mechanism is established: machine tool servo control, safety projection, machining parameter adjustment, sensor configuration and strategy update are run according to different time scales; after the machining segment or workpiece is completed, delayed quality results such as dimensional error, surface roughness, and shape error are associated with the corresponding program segment, tool position, workpiece area and state-action segment for strategy evaluation, dataset update and subsequent training, thereby forming a complete closed loop of "active perception - joint decision-making - safe execution - quality feedback - strategy update".

[0046] The adaptive control method for workpiece processing proposed in this invention brings the following effects: (1) Improve the efficiency of sensing resource utilization: This invention dynamically configures sensors and sampling parameters according to the uncertainty of processing state and quality risk. When the processing state is stable, it reduces the acquisition and processing of high-frequency and high-bandwidth signals. In high-risk stages such as thin walls, corners, material hard points or tool wear, it enhances the sensing ability, thereby reducing redundant data, communication load and computing overhead, so that the sensing resources match the actual processing needs. (2) Improve the identification and control capabilities under complex processing conditions: By fusing asynchronous and multi-rate sensing data and explicitly introducing information timeliness, missing state, quality score and prediction uncertainty, this invention can maintain reliable state estimation under the circumstances of dynamic sensor start-up and shutdown, communication packet loss or partial signal failure. The processing parameters and sensing configuration are jointly determined by the same strategy, so that the system can coordinate between directly adjusting the processing parameters and enhancing information acquisition, and improve the adaptability to tool wear, chatter tendency, material changes and workpiece stiffness changes. (3) Improve the safety of reinforcement learning processing control: Existing methods usually rely on rewards and penalties to limit dangerous actions, which cannot guarantee that the actual physical constraints are met at each control moment. This invention sets up an independent safety projection unit outside the strategy network and sets a safety margin in combination with the prediction uncertainty, so that the actions sent to the CNC system meet the boundaries of the machine tool, tool, workpiece and process; in abnormal situations, it can also automatically fall back to the verified basic controller, thereby reducing the risk of tool overload, machining chatter, workpiece out of tolerance and machine tool protection shutdown. (4) Reduce the cost of trial and error on real machine tools and strategy online: This invention makes full use of existing machining logs, trial cutting data and quality inspection results to carry out offline training, reducing the number of times reinforcement learning is explored on real machine tools; through digital twin playback, abnormal working condition testing, strategy version verification and restricted online updates, potential risks can be found before the strategy is officially executed, reducing scrap, tool wear and debugging costs. (5) Improve the stability of machining quality and operational robustness: This invention incorporates factors such as machining time, load fluctuation, chatter risk, energy consumption, tool wear, perception cost and safety intervention into a unified evaluation, and uses the size, roughness and shape error after machining to correct the strategy evaluation of related machining segments. Therefore, it can balance processing efficiency and quality stability while meeting safety boundaries, and maintain strong robustness under conditions such as material fluctuations, allowance changes, sensor failures and communication and computing anomalies. (6) It is easy to integrate with existing CNC systems: The present invention preferably adopts a multiple or bounded residual correction relative to the original CNC program reference parameters, without changing the underlying servo control and emergency protection structure of the machine tool; the normally open signal can directly read the existing data of the CNC system or driver, and active perception, safety projection and strategy reasoning can be deployed on the edge intelligent controller, thus having good engineering compatibility and practical deployment feasibility.

[0047] In summary, this embodiment first acquires normally open low-cost signal data, then performs time-series normalization processing on the normally open low-cost signal data to obtain the first multi-source sensor feature data. It abandons the traditional high-cost, high-barrier dedicated sensor data acquisition scheme, and completes basic data acquisition based on low-cost normally open signals, reducing the hardware costs and technical barriers of equipment deployment, data acquisition, and maintenance. Simultaneously, through time-series normalization processing, it cleans, organizes, and standardizes the original messy, time-disordered, and inconsistent formatted sensor data, effectively avoiding the problems of high noise, chaotic timing, and low effectiveness of the original data, and outputting a regular and uniform data. The first multi-source sensor feature data provides high-quality, highly consistent, and highly reliable foundational data support for subsequent end-to-end state analysis, inference calculations, and action decisions, ensuring the accuracy and stability of subsequent technical operations from the source. Next, based on the first multi-source sensor feature data, state estimation is performed to obtain a reliable state. Then, based on standardized multi-source feature data, the real-time operating status of equipment, working conditions, and workpieces is deeply mined, overcoming the limitations of existing technologies that rely on single data and experience to judge the state. This effectively eliminates invalid data interference, corrects state deviations, filters abnormal working condition information, and accurately outputs data that aligns with actual production. The reliable state of the scenario solves the problems of inaccurate state identification, delayed condition judgment, and low state reliability in traditional technologies, providing a real and reliable basis for subsequent policy reasoning. Subsequently, based on the reliable state, joint policy reasoning yields sensing configuration actions and processing candidate actions. Compared to the traditional single-policy judgment mode, this achieves coordinated linkage between sensing adjustment and processing actions, enriching the dimensions and coverage of action decision-making, and avoiding the problems of poor condition adaptability and insufficient action matching caused by single-policy decisions. This provides diversified and scenario-appropriate alternative action schemes for subsequent optimization solutions. Then, based on the addition... The process involves selecting candidate actions and performing optimized projection solutions with safety constraints to obtain safe actions. Invalid actions that pose safety hazards, process deviations, or operational risks are eliminated, resulting in the precise output of safe actions that balance safety, accuracy, and efficiency. This effectively addresses the shortcomings of traditional machining action decision-making, such as neglecting safety risks, insufficient optimization, and low fault tolerance, ensuring optimal adaptability of machining actions. Finally, based on these safe actions, workpiece machining is performed, improving the accuracy, stability, and consistency of workpiece machining, reducing the probability of machining errors, equipment failures, and production safety accidents, and effectively improving the finished product qualification rate and overall production efficiency.Overall, this invention constructs a fully closed-loop intelligent processing control system encompassing low-cost data acquisition, standardized data processing, precise state perception, multi-strategy joint reasoning, secure optimization solutions, and precise processing execution. It achieves high-precision and high-safety intelligent production control entirely through low-cost hardware. While effectively reducing production hardware investment and maintenance costs, it addresses numerous technical shortcomings of existing technologies, such as poor data quality, inaccurate state recognition, single decision-making dimensions, weak security, unstable processing accuracy, and poor adaptability to operating conditions. It realizes collaborative intelligent control of sensor configuration and workpiece processing, balancing economic efficiency, safety, accuracy, and efficiency in production. This enhances the intelligence level, operating condition adaptability, and production stability of industrial processing, demonstrating strong engineering practicality and broad scenario adaptability.

[0048] like Figure 5 As shown in the illustration, this embodiment also provides a workpiece machining adaptive control device, which includes: a first multi-source sensor feature data acquisition module 10, a machining candidate action and sensor configuration action acquisition module 20, and a workpiece machining module 30. Specifically, the first multi-source sensor feature data acquisition module 10 is used to collect normally open low-cost signal data, perform time-series normalization processing on the normally open low-cost signal data, and obtain the first multi-source sensor feature data. The machining candidate action and sensor configuration action acquisition module 20 is used to perform state estimation based on the first multi-source sensor feature data to obtain a reliable state, and based on the reliable state, obtain the machining candidate action and sensor configuration action through joint strategy reasoning. The workpiece machining module 30 is used to perform optimization projection solution with safety constraints based on the machining candidate actions to obtain a safe action, and perform workpiece machining based on the safe action.

[0049] In one implementation, the first multi-source sensing feature data acquisition module 10 includes: The raw normally open low-cost signal data acquisition unit for each channel is used to acquire normally open low-cost signal data through the normally open low-cost signal channel, and add a timestamp to the normally open low-cost signal data of each channel according to a unified clock to obtain the raw normally open low-cost signal data of each channel carrying a unified timing mark. Each channel's basic normally open low-cost signal data acquisition unit is used to build a sliding time window with the control decision cycle as the timing reference, and to perform timing alignment on the original normally open low-cost signal data of each channel within the window to obtain the timing-synchronized and aligned basic normally open low-cost signal data of each channel. The integrated normally open low-cost signal data acquisition unit for each channel is used to calculate information timeliness, missing identifier, and quality score based on the effective observation time of the channel, data failure type, and signal operating parameters, and to obtain integrated normally open low-cost signal data for each channel with timeliness, missing identifier, quality score and basic normally open low-cost signal data of each channel. The first multi-source sensor feature data acquisition unit is used to regularize and process the integrated normally open low-cost signal data of each channel to obtain the first multi-source sensor feature data.

[0050] In one implementation, the processing candidate action and sensing configuration action acquisition module 20 includes: The trusted state acquisition unit is used to calculate the trusted state of the previous moment, the current first multi-source sensor feature data, the security action of the previous moment, and the current sensor configuration action using the state estimation function to obtain the trusted state; The component acquisition unit is used to perform forward reasoning on the trusted state through a joint reinforcement learning strategy to obtain the processing action component and the sensing action component. The processing candidate action acquisition unit is used to quantify and characterize the processing action components to obtain processing candidate actions; The sensor configuration action acquisition unit is used to perform multi-channel parameter assignment and resource constraint feasibility screening on the sensor action components to obtain the sensor configuration action.

[0051] In one implementation, the workpiece processing module 30 includes: An optimized projection model building unit is used to build an optimized projection model that includes multiple process safety constraints; The action to be verified acquisition unit is used to input the candidate processing actions into the optimization projection model to perform optimization projection solution and calculate the action to be verified. The safety action acquisition unit is used to identify the action to be verified as a safe action if the overall state of the solved working condition meets the preset normal working condition conditions.

[0052] The working principle of each module in the workpiece machining adaptive control device of this embodiment is the same as that of each step in the above method embodiment, and will not be repeated here.

[0053] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 6 As shown. The terminal may include one or more processors 100 ( Figure 6(Only one is shown in the image), memory 101, and a computer program 102 stored in memory 101 that can run on one or more processors 100, such as a workpiece machining adaptive control program. When one or more processors 100 execute computer program 102, they can implement the various steps in the workpiece machining adaptive control method embodiment. Alternatively, when one or more processors 100 execute computer program 102, they can implement the functions of various modules / units in the workpiece machining adaptive control method embodiment, which is not limited here.

[0054] In one embodiment, the processor 100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0055] In one embodiment, memory 101 may be an internal storage unit of an electronic device, such as a hard drive or RAM. Memory 101 may also be an external storage device of the electronic device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, memory 101 may include both internal and external storage units. Memory 101 is used to store computer programs and other programs and data required by the terminal. Memory 101 can also be used to temporarily store data that has been output or will be output.

[0056] Those skilled in the art will understand that Figure 6 The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0057] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, operational databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual operating data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A workpiece machining adaptive control method, characterized in that, The method includes: Normally open low-cost signal data is collected, and time-series normalization processing is performed on the normally open low-cost signal data to obtain the first multi-source sensing feature data; Based on the first multi-source sensor feature data, state estimation is performed to obtain a credible state. Based on the credible state, sensor configuration actions and processing candidate actions are obtained through joint policy reasoning. Based on the candidate machining actions, an optimization projection solution with safety constraints is performed to obtain the safe actions, and the workpiece is machined based on the safe actions.

2. The adaptive control method for workpiece machining according to claim 1, characterized in that, The process of acquiring normally open low-cost signal data and performing time-series normalization on the normally open low-cost signal data to obtain first multi-source sensing feature data includes: Normally open low-cost signal data is acquired through normally open low-cost signal channels, and timestamps are added to the normally open low-cost signal data of each channel according to a unified clock to obtain the original normally open low-cost signal data of each channel carrying a unified timing mark. A sliding time window is constructed using the control decision cycle as the timing benchmark. The original normally open low-cost signal data of each channel within the window are time-aligned to obtain the basic normally open low-cost signal data of each channel after timing synchronization alignment. Based on the effective observation time of the channel, the data failure type, and the signal operating parameters, the information timeliness, missing identifier, and quality score are calculated respectively. Based on the information timeliness, missing identifier, quality score, and the basic normally open low-cost signal data of each channel, integrated normally open low-cost signal data of each channel with timeliness, missing identifier, and quality information is obtained. The integrated normally open low-cost signal data of each channel is normalized and processed to obtain the first multi-source sensing feature data.

3. The adaptive control method for workpiece machining according to claim 1, characterized in that, The process of estimating the state based on the first multi-source sensing feature data to obtain a reliable state includes: The trusted state is obtained by using the state estimation function to calculate the trusted state at the previous moment, the current first multi-source sensor feature data, the security action at the previous moment, and the current sensor configuration action.

4. The adaptive control method for workpiece machining according to claim 1, characterized in that, The process of obtaining sensor configuration actions and processing candidate actions based on trusted states and through joint policy reasoning includes: By using a joint reinforcement learning strategy to perform forward reasoning on the trustworthy state, the processing action component and the sensing action component are obtained. The processing action components are quantified to obtain candidate processing actions; Multi-channel parameter assignment and resource constraint feasibility screening are performed on the sensing action components to obtain the sensing configuration action.

5. The adaptive control method for workpiece machining according to claim 1, characterized in that, The process of obtaining the sensor configuration action then includes: Set the sensor acquisition configuration parameters according to the sensor configuration action, and maintain the preset monitoring level for the forced protection channel.

6. The adaptive control method for workpiece machining according to claim 1, characterized in that, The method further includes: Collect on-demand high-cost signal data, perform time-series normalization on the on-demand high-cost signal data, and obtain the second multi-source sensing feature data; Based on the first multi-source sensing feature data and the second multi-source sensing feature data, the target multi-source sensing feature data is obtained. Based on the target's multi-source sensor feature data, the processing state is estimated and the confidence state is updated.

7. The adaptive control method for workpiece machining according to claim 1, characterized in that, The step of performing an optimization projection solution with safety constraints based on candidate processing actions to obtain safe actions includes: Construct an optimized projection model that includes multiple process safety constraints; The candidate actions are input into the optimization projection model to perform optimization projection solution, and the actions to be verified are calculated. If the overall working condition meets the preset normal working conditions, then the action to be verified is a safe action.

8. A workpiece machining adaptive control device, characterized in that, The device includes: The first multi-source sensing feature data acquisition module is used to collect normally open low-cost signal data, perform time-series normalization processing on the normally open low-cost signal data, and obtain the first multi-source sensing feature data. The processing candidate action and sensor configuration action acquisition module is used to perform state estimation based on the first multi-source sensor feature data, obtain a credible state, and obtain the processing candidate action and sensor configuration action based on the credible state through joint policy reasoning. The workpiece machining module is used to perform optimized projection solving with safety constraints based on the candidate machining actions, obtain the safe actions, and perform workpiece machining based on the safe actions.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a workpiece machining adaptive control program stored in the memory and executable on the processor. When the processor executes the workpiece machining adaptive control program, it implements the steps of the workpiece machining adaptive control method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a workpiece machining adaptive control program, which, when executed by a processor, implements the steps of the workpiece machining adaptive control method as described in any one of claims 1-7.

Images