Dynamic calibration control method for detection precision of non-standard automatic detection tool

Abstract

The invention discloses a dynamic calibration control method for the detection precision of a non-standard automatic testing fixture, and particularly relates to the field of industrial automation and control, which comprises the following steps of: constructing a real-time dynamic error field model of the testing fixture; a distributed optical fiber sensing network and a miniature laser interferometer cooperatively collect real-time detection data and spatial-temporal variation errors under a multi-physics coupling effect, a precision degradation track is predicted based on spatial-temporal variation error data, a pre-compensation signal is generated, and a self-adaptive Kalman filtering algorithm is adopted to generate a precision compensation parameter. Environment temperature, vibration and electromagnetic interference data are monitored in real time, environment disturbance data are generated, a hybrid controller combining nonlinear model predictive control and reinforcement learning is adopted, multi-source data are fused, a calibration instruction is generated through multi-target optimization, and an execution mechanism is driven to achieve online dynamic calibration. And digital threads and transfer learning are integrated to realize full-life-cycle data closed-loop management and continuous optimization.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and control technology, and more specifically, to a method for dynamic calibration and control of the detection accuracy of non-standard automated inspection tools. Background Technology

[0002] In recent years, as the manufacturing industry has developed towards high precision and intelligence, non-standard automated inspection tools have played an increasingly important role in product quality control. The inspection accuracy of the tool directly affects the product pass rate and the stability of the production line. However, in the actual industrial environment, the tool is subjected to the coupling effects of multiple physical fields such as temperature changes, mechanical vibration, and electromagnetic interference for a long time, which causes its inspection accuracy to drift and degrade. Traditional periodic static calibration methods are difficult to meet the requirements of real-time performance and accuracy. Therefore, there is an urgent need for an online, dynamic, and adaptive precision calibration control method.

[0003] Current technical solutions typically set a fixed calibration cycle and conduct offline calibration in a standard environmental laboratory. The positioning error of the gauge probe is calibrated using standard gauge blocks to obtain fixed compensation parameters. Then, environmental data is collected based on a single temperature sensor or vibration sensor, and the compensation value is corrected using a linear interpolation algorithm. Finally, the position of the gauge probe or the parameters of the motion mechanism are adjusted manually. After calibration, the compensation parameters are fixed to the gauge control system, and the above process is repeated until the next calibration cycle.

[0004] However, it still has some drawbacks in practical use. For example, it adopts fixed-cycle offline calibration, which cannot respond in real time to the dynamic error changes of the fixture under multi-physics coupling, and is prone to the problem of "error exceeding the tolerance within the calibration interval". It relies on linear interpolation algorithm for static compensation, without considering the nonlinear characteristics of fixture accuracy degradation. The compensation parameters do not match the actual error evolution law, and the calibration accuracy is limited. It only monitors environmental data through a single sensor, without building multiple quantization mapping relationships, and the calibration data does not form a closed loop throughout the entire life cycle, making it impossible to continuously optimize the control strategy. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a dynamic calibration and control method for the detection accuracy of non-standard automated inspection tools, which solves the problems mentioned in the background art through the following scheme.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a dynamic calibration and control method for the detection accuracy of a non-standard automated inspection fixture, comprising S1: constructing a real-time dynamic error field model of the inspection fixture, and collecting real-time detection data and spatiotemporal variation error of the inspection fixture under the coupling effect of multiple physical fields through a distributed optical fiber sensor network and a miniature laser interferometer; S2: Based on spatiotemporal variation error data, predict the accuracy degradation trajectory of the fixture and generate a pre-compensation signal. At the same time, use an adaptive Kalman filter algorithm to filter and compensate for errors in real-time detection data to generate accuracy compensation parameters. S3: Based on multi-source sensors, real-time monitoring of temperature, vibration and electromagnetic interference data of the environment in which the fixture is located generates environmental disturbance data; S4: A hybrid controller based on nonlinear model predictive control and reinforcement learning is adopted. The pre-compensation signal, accuracy compensation parameters and environmental disturbance data are fused to solve a multi-objective optimization problem to generate calibration instructions. The reinforcement learning part is trained with a dual-delay deep deterministic policy gradient algorithm. The reward function simultaneously optimizes accuracy stability, calibration efficiency and equipment loss. S5: Based on the hybrid controller, the actuator is driven to achieve online dynamic calibration, and digital thread technology is integrated to realize closed-loop management of data throughout the entire life cycle. The transfer learning algorithm is used to continuously optimize the control strategy.

[0007] Preferably, the distributed optical fiber sensing network uses optical frequency domain reflective distributed optical fiber with a spatial resolution of 1 mm, a strain measurement accuracy of ±1 μe, and a temperature measurement accuracy of ±0.1 °C. The micro laser interferometer is a nanoscale laser interferometer with a measurement accuracy of ±0.05 μm and a sampling frequency of 1 kHz.

[0008] Preferably, the real-time detection data includes fiber optic sensing data, laser interferometer data, and the instrument's own detection data. The spatiotemporal variation error includes temperature deformation error, vibration displacement error, and probe positioning error, wherein temperature deformation error accounts for 50% of the total error, vibration displacement error accounts for 35% of the total error, and probe positioning error accounts for 15% of the total error.

[0009] Preferably, the gauge accuracy degradation trajectory is based on spatiotemporal variation error data, and dynamic trend modeling is performed using a Long Short-Term Memory (LSTM) network. Spatiotemporal variation error data from the past 24 hours are selected to construct a 3-layer LSTM network, which outputs the gauge accuracy degradation trend for the next hour, i.e., the gauge accuracy value corresponding to different time nodes.

[0010] Preferably, the pre-compensation signal is dynamically adjusted over time and is set according to the predicted increase in error within the next 30 minutes, with the setting format being "increase the compensation amount by 'a' every 10 minutes".

[0011] Preferably, the adaptive Kalman filter algorithm optimizes the error compensation effect by adjusting the filter gain matrix in real time, including constructing state equations and observation equations for adaptive gain adjustment.

[0012] Preferably, the environmental disturbance data needs to be classified into levels. The classification criteria are based on historical data and detection error thresholds to classify environmental disturbances into low disturbances, medium disturbances, and high disturbances. The determination factors are temperature fluctuations, vibration acceleration, electromagnetic interference, and corresponding error contributions.

[0013] Preferably, the hybrid controller is divided into two layers. The upper layer is an RL agent, which is responsible for dynamically adjusting the weight coefficients of the nonlinear model predictive control to balance accuracy stability, calibration efficiency and equipment loss. The lower layer is a nonlinear model predictive controller, which solves a multi-objective optimization problem based on the pre-compensation signal, accuracy compensation parameters and environmental disturbance data, and generates calibration instructions.

[0014] Preferably, the online dynamic calibration includes actuator driving, probe compensation, motion mechanism adjustment, environmental adaptive control, and real-time monitoring of calibration effect. The transfer learning algorithm continuously optimizes the control strategy, which includes three steps: data modeling, data flow, and transfer learning continuous optimization. Data modeling includes three types of core data: static data, dynamic data, and historical data. Data flow goes through three levels: acquisition layer, processing layer, and application layer. Transfer learning continuous optimization includes three core elements: source domain data, target domain data, and transfer strategy.

[0015] The technical effects and advantages of this invention are as follows: 1. By constructing a real-time dynamic error field model, this invention can monitor the spatiotemporal variation error of the gauge in real time and perform dynamic calibration based on the predicted accuracy degradation trajectory, effectively improving the detection accuracy and stability of the gauge. Compared with traditional manual calibration or offline calibration, this invention can significantly reduce downtime and improve production efficiency. 2. This invention employs a hybrid controller based on nonlinear model predictive control and reinforcement learning, which can simultaneously optimize accuracy stability, calibration efficiency, and equipment loss, achieving multi-objective optimization control. Compared with traditional methods, this invention can more effectively balance the relationship between various objectives and improve calibration results. 3. This invention achieves closed-loop management of the entire lifecycle data of the inspection tool through digital thread technology, and continuously optimizes the control strategy by adopting transfer learning algorithm. It establishes a data closed-loop management and autonomous learning optimization system covering the entire lifecycle, which enhances the system's adaptability under different working conditions and the stability of long-term operation. Compared with traditional methods, this invention can effectively utilize historical data to optimize the control strategy and improve calibration efficiency and control effect. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a schematic diagram of the hybrid controller workflow of the present invention; Figure 3This is a schematic diagram of the closed-loop data management throughout the entire lifecycle of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] As attached Figure 1 Appendix Figure 2 and appendix Figure 3 The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection fixture is shown, including S1: constructing a real-time dynamic error field model of the fixture, and collecting real-time detection data and spatiotemporal variation error of the fixture under the coupling effect of multiple physical fields through a distributed optical fiber sensor network and a micro laser interferometer.

[0019] It should be specifically noted that the distributed optical fiber sensing network uses optical frequency domain reflective distributed optical fiber with a spatial resolution of 1 mm, a strain measurement accuracy of ±1 με, and a temperature measurement accuracy of ±0.1 °C. The micro laser interferometer is a nanoscale laser interferometer with a measurement accuracy of ±0.05 μm and a sampling frequency of 1 kHz.

[0020] The real-time detection data includes fiber optic sensing data, laser interferometer data, and the instrument's own detection data. The spatiotemporal variation error includes temperature deformation error, vibration displacement error, and probe positioning error, of which temperature deformation error accounts for 50% of the total error, vibration displacement error accounts for 35% of the total error, and probe positioning error accounts for 15% of the total error.

[0021] It should be further explained that the distributed fiber optic sensors simultaneously collect strain and temperature distribution data of the fixture substrate. The deployment locations include three optical fibers attached along the axial direction of the leaf-shaped detection reference beam with a spacing of 40mm, covering the area below the probe's movement trajectory; two optical fibers attached along the circumference of the Z-axis column of the 6-axis motion mechanism with an included angle of 180° to capture the bending deformation of the column; and one optical fiber attached along the diagonal of the fixture's worktable to monitor changes in the flatness of the worktable. The miniature laser interferometer collects the displacement error of the fixture's probe. The deployment locations include one unit each on the X, Y, and Z axes of the fixture's motion mechanism, aligned with the reference reflective surface of the probe, to collect the positioning error of the probe during movement in real time; and one unit deployed below the fixture's worktable to collect the vibration displacement of the worktable and correlate the impact of vibration on the detection accuracy.

[0022] The data acquisition process is divided into a "static calibration phase" and a "dynamic operation phase." Through synchronous triggering of multiple sensors (trigger accuracy ≤1μs), the following core data are acquired: The static calibration phase involves controlling the gauge probe to move along the X, Y, and Z axes of a standard gauge block under standard environmental conditions (temperature 20℃±0.5℃, vibration ≤5Hz, electromagnetic interference ≤10V / m). Each group consists of 20 gauges, and 100 sets of "fiber optic strain / temperature-laser interferometer displacement" data are collected to establish an initial mapping relationship between "physical field parameters and errors," serving as the benchmark for the dynamic error field model. The dynamic operation phase involves synchronously acquiring data at a sampling frequency of 1kHz during normal production line operation. This includes fiber optic sensor data, temperature and strain distribution of the gauge substrate (acquisition efficiency 1200 sampling points per second), laser interferometer data, probe X / Y / Z axis displacement errors and worktable vibration displacement (acquisition efficiency 1000 sampling points per second), gauge self-detection data, and probe dimensional detection values ​​of the gearbox housing (acquisition efficiency 12 seconds per set).

[0023] The dynamic error field model is constructed based on the Gaussian process regression (GPR) algorithm. The specific processing flow is as follows: Data preprocessing: Outlier removal and time synchronization are performed on the collected raw data. Outlier removal employs... The criteria include eliminating anomalies caused by transient interference from sensors, and controlling the timestamp deviation between fiber optic data and laser interferometer data within a specified range during time synchronization. Internally, feature variable selection: Through Pearson correlation analysis and random forest feature importance assessment, variables that significantly affect the detection error are selected. The temperature change of the gauge reference beam, the strain value of the Z-axis column, and the vibration acceleration of the workbench are the core influencing factors and are used as model input variables. Model training and validation: Using 100 sets of data from the static calibration phase as the training set and 50 sets of data from the dynamic operation phase as the validation set, a Gaussian process regression (GPR) dynamic error field model is constructed, and the output is "the comprehensive error value of the gauge at the current spatiotemporal position" (integrating temperature deformation error, vibration displacement error, and probe positioning error).

[0024] S2: Based on spatiotemporal variation error data, predict the trajectory of gauge accuracy degradation and generate a pre-compensation signal. At the same time, use an adaptive Kalman filter algorithm to filter and compensate for errors in real-time detection data to generate accuracy compensation parameters.

[0025] It should be specifically noted that the fixture accuracy degradation trajectory is based on spatiotemporal variation error data and uses a Long Short-Term Memory (LSTM) network for dynamic trend modeling. Spatiotemporal variation error data from the past 24 hours are selected to construct a 3-layer LSTM network, which outputs the fixture accuracy degradation trend for the next hour, i.e., the fixture accuracy value corresponding to different time nodes.

[0026] The pre-compensation signal is dynamically adjusted over time and is set according to the predicted increase in error within the next 30 minutes. The setting format is "increase the compensation amount by 'a' every 10 minutes".

[0027] The adaptive Kalman filter algorithm optimizes the error compensation effect by adjusting the filter gain matrix in real time, including constructing state equations and observation equations for adaptive gain adjustment.

[0028] It should be further explained that, based on spatiotemporal variation error data, a Long Short-Term Memory (LSTM) network is used to predict the accuracy degradation trajectory of the inspection fixture. The specific implementation is as follows: Spatiotemporal variation error data from the past 24 hours are selected, with a time step of 1 minute and 1440 data points, including the hourly average error value, error fluctuation standard deviation, and the maximum, minimum, and mean values ​​of statistical characteristics of temperature / vibration / strain. A three-layer LSTM network consisting of an input layer, a hidden layer, and an output layer is constructed to output the accuracy degradation trend of the inspection fixture in the next hour. Based on the predicted degradation trajectory, a "pre-compensation signal that dynamically adjusts over time" is generated, with the format set as "compensation amount increasing by 'a' every 10 minutes" (for example, if the error is predicted to increase by 0.002 mm in the next 30 minutes, the pre-compensation signal is set as "compensation amount increasing by 0.0007 mm every 10 minutes", where 'a' is approximately 0.0007 mm).

[0029] The adaptive Kalman filter (AKF) algorithm is adopted to optimize the error compensation effect by adjusting the filter gain matrix in real time. The specific process is as follows: construct the state equation and the observation equation.

[0030] Equations of state:

[0031] in, This is the error state vector at time k (including temperature deformation error and vibration displacement error). This is the state transition matrix (updated in real time based on the dynamic error field model). For the control matrix, For pre-compensation signal, The process noise (follows a Gaussian distribution, and its variance is estimated from real-time data).

[0032] Observation equation:

[0033] in, Let k be the laser interferometer observation value at time k. For the observation matrix, The observation noise (variance is set to 0.05 μm² using the sensor accuracy parameter) is used.

[0034] Kalman gain is adjusted in real time using the fading memory method. When the deviation between the observed and predicted values ​​is large (e.g., a sudden change in vibration causes an error jump), increase... To increase the weight of observations, when the system is stable, decrease... To reduce the impact of noise, the error value after AKF filtering is fused with the pre-compensation signal to generate the final accuracy compensation parameters (for example, when the laser interferometer detects the probe). The shaft displacement error is 0.0025mm, which is corrected to 0.0023mm after AKF filtering. Combined with the pre-compensation signal of 0.0007mm, the final compensation parameter is set to 0.003mm. The position of the probe is adjusted by the servo system of the fixture to achieve error cancellation.

[0035] S3: Based on multi-source sensors, real-time monitoring of the temperature, vibration and electromagnetic interference data of the environment in which the fixture is located generates environmental disturbance data.

[0036] It should be specifically noted that the environmental disturbance data needs to be classified into levels. The classification criteria are based on historical data and detection error thresholds to divide environmental disturbances into low disturbances, medium disturbances, and high disturbances. The determining factors are temperature fluctuations, vibration acceleration, electromagnetic interference, and the corresponding error contribution.

[0037] It should be further explained that by using multi-source sensors to monitor environmental parameters in real time and constructing an "environmental disturbance-error" mapping relationship, the specific implementation is as follows: A resistance temperature sensor with an accuracy of ±0.1℃ and a sampling frequency of 1Hz is selected and deployed near the gauge probe, at both ends of the reference beam, below the workbench, and inside the control cabinet to collect environmental temperature distribution data. A piezoelectric accelerometer with a measurement range of ±50g, a frequency response of 0.1Hz-10kHz, and a sampling frequency of 1kHz is selected and deployed at the four support points of the gauge base to collect vibration acceleration data of the X, Y, and Z axes. An electric field strength sensor with a measurement range of 1V / m-1000V / m, a frequency range of 10kHz-1GHz, and a sampling frequency of 10Hz is selected and deployed 1m above the gauge control cabinet to collect environmental electromagnetic interference intensity data.

[0038] Environmental disturbance data was quantified by using Min-Max standardization to convert the data to the [0,1] interval, eliminating dimensional differences. Based on historical data and the detection error threshold (±0.01mm), environmental disturbances were divided into three levels. The classification criteria were as follows: low disturbances included temperature fluctuations ≤2℃ / h, vibration acceleration ≤0.5g, electromagnetic interference ≤30V / m, and corresponding error contribution ≤0.002mm; and low disturbances included temperature fluctuations of 2℃ / h-5℃ / h, vibration acceleration of 0.5g-1g, electromagnetic interference of 30V / m-60V / m, and corresponding error contribution of 0.002mm-0.005mm. Medium disturbances are defined as temperature fluctuations > 5℃ / h, vibration acceleration > 1g, electromagnetic interference > 60V / m, and corresponding error contributions > 0.005mm. High disturbances are defined as temperature fluctuations > 5℃ / h, vibration acceleration > 1g, electromagnetic interference > 60V / m, and corresponding error contributions > 0.005mm. The influence coefficients of each environmental parameter on the detection error are calculated by multiple linear regression.

[0039] S4: A hybrid controller based on nonlinear model predictive control and reinforcement learning is adopted. It integrates pre-compensation signals, accuracy compensation parameters and environmental disturbance data to solve a multi-objective optimization problem and generate calibration instructions. The reinforcement learning part is trained with a dual-delay deep deterministic policy gradient algorithm, and the reward function simultaneously optimizes accuracy stability, calibration efficiency and equipment loss.

[0040] It should be specifically noted that the hybrid controller is divided into two layers. The upper layer is an RL agent, which is responsible for dynamically adjusting the weight coefficients of the nonlinear model predictive control to balance accuracy stability, calibration efficiency and equipment loss. The lower layer is a nonlinear model predictive controller, which solves a multi-objective optimization problem based on the pre-compensation signal, accuracy compensation parameters and environmental disturbance data, and generates calibration instructions.

[0041] It should be further explained that a hybrid controller of "Nonlinear Model Predictive Control (NMPC) + Reinforcement Learning (RL)" is adopted. The RL agent is trained by the dual-delay deep deterministic policy gradient (TD3) algorithm to achieve multi-objective optimization. The specific implementation is as follows: A two-layer hybrid controller architecture is designed, and the state space is set to include six state variables, including environmental disturbance level (low / medium / high), current detection error, gauge servo motor load rate, and historical calibration time.

[0042] Output the three weight coefficients of NMPC. Precision stability weights Calibration efficiency weights Equipment loss weights, all ranging from [0,1], and .

[0043] Constructing a multi-objective comprehensive reward function :

[0044] Among them, accuracy and stability rewards When the detection error is ≤0.001mm, =10, for every 0.0001mm of error, Decrease by 1; when the error is > 0.002 mm, =-5, calibration efficiency bonus When the calibration time is ≤0.5 seconds (without affecting the 12-second detection cycle), =8, for every 0.1 seconds increase in calibration time, Decrease by 1; when calibration time > 1 second, =-3, Equipment Loss Bonus When the servo motor load rate is ≤30%, =6, for every 5% increase in load rate Decrease by 1 when the load rate is >50%. =-4, the training process adopts an experience playback mechanism, collecting 100,000 sets of training data (covering low / medium / high perturbation scenarios), setting the learning rate to 0.001, the discount factor to 0.95, and training iterations to 5000. After training, the RL agent dynamically adjusts the weight coefficients according to real-time operating conditions. High perturbation scenario (error contribution > 0.005mm): Output =0.6, =0.2, =0.2, prioritizing accuracy, in medium disturbance scenarios (error contribution 0.002mm-0.005mm): Output =0.4, =0.3, =0.3, balancing the requirements of the three factors, low-disturbance scenario (error contribution ≤0.002mm): Output =0.2, =0.5, =0.3, prioritize improving efficiency.

[0045] Based on the weight coefficients output by RL, construct the optimization objective function for NMPC. : in, This is the detection error after calibration. To calibrate the time, For servo motor load rate, constraint settings include error constraints. ≤0.001mm, time constraint ≤1 second, load constraint ≤50%.

[0046] The optimization problem is solved using the Sequential Quadratic Programming (SQP) algorithm to generate calibration instructions.

[0047] S5: Based on the hybrid controller, the actuator is driven to achieve online dynamic calibration, and digital thread technology is integrated to realize closed-loop management of data throughout the entire life cycle. The transfer learning algorithm is used to continuously optimize the control strategy.

[0048] It should be specifically noted that the online dynamic calibration includes actuator driving, probe compensation, motion mechanism adjustment, environmental adaptive control, and real-time monitoring of calibration results. The transfer learning algorithm continuously optimizes the control strategy, which includes three steps: data modeling, data flow, and transfer learning continuous optimization. Data modeling includes three types of core data: static data, dynamic data, and historical data. Data flow goes through three levels: acquisition layer, processing layer, and application layer. Transfer learning continuous optimization includes three core elements: source domain data, target domain data, and transfer strategy.

[0049] It should be further explained that after receiving the calibration command from the hybrid controller, the fixture PLC drives the following actuators to achieve dynamic adjustment: The probe position is adjusted via a piezoelectric ceramic micro-displacement stage, with a displacement resolution of 0.001µm and a response time ≤1ms, achieving X / Y / Z three-axis compensation; the speed and acceleration of the 6-axis motor are adjusted via a servo driver to reduce motion hysteresis error; and the workshop air conditioning system and vibration-damping platform are linked to respond to environmental disturbances. 总 When the error is greater than 0.005mm, the air conditioning temperature compensation (accuracy ±0.5℃) and platform active vibration reduction (reduction rate ≥90%) are automatically activated. The calibrated data is collected in real time by distributed fiber optic sensors and laser interferometers to determine whether the target of "error ≤0.002mm" is met. If it is met, the detection task continues and the accuracy is rechecked every 10 blades. If it is not met, a second calibration is triggered and the hybrid controller regenerates the calibration command until the error meets the target (maximum 3 times to avoid getting stuck in an infinite loop).

[0050] Digital thread technology constructs a full lifecycle data chain for inspection fixtures through three stages: "data modeling, flow, and analysis." The specific implementation is as follows: A unified product lifecycle data exchange standard is used to establish a "digital twin model" of the inspection fixture, containing three core data categories: static data, including fixture design parameters (material, dimensions, sensor model) and blade inspection standards (tolerance range, inspection points); dynamic data, including real-time inspection data (blade size values, error values), calibration data (calibration instructions, compensation parameters), and environmental data (temperature, vibration, electromagnetic interference); and historical data, including accuracy degradation curves, fault records, and calibration effect statistics for the past 12 months. Data flow across systems is achieved through an industrial internet platform. The acquisition layer uploads sensor data to the platform in real time. The processing layer cleans and fuses the data to generate a "precision-environment-calibration" related report. The application layer synchronizes the data to the design software for model iteration, the MES system for production scheduling, and the quality management system for blade traceability. At the same time, transfer learning is used for continuous optimization. When the application scenario of the inspection tool changes, the transfer learning algorithm is used to reuse the historical calibration strategy to avoid training from scratch. Specifically, this includes source domain data, 50,000 sets of calibration data from the original high-pressure turbine blade inspection, and target domain data, 5,000 sets of initial calibration data under the new scenario. The transfer strategy freezes the first 3 layers of the TD3 network, fine-tunes the output layer, and reduces the number of training rounds from 5,000 to 1,000.

[0051] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool, characterized in that, Includes the following steps: S1: Construct a real-time dynamic error field model of the fixture, and collect real-time detection data and spatiotemporal variation error of the fixture under the coupling effect of multiple physical fields through a distributed fiber optic sensor network and a miniature laser interferometer. S2: Based on spatiotemporal variation error data, predict the accuracy degradation trajectory of the fixture and generate a pre-compensation signal. At the same time, use an adaptive Kalman filter algorithm to filter and compensate for errors in real-time detection data to generate accuracy compensation parameters. S3: Based on multi-source sensors, real-time monitoring of temperature, vibration and electromagnetic interference data of the environment in which the fixture is located generates environmental disturbance data; S4: A hybrid controller based on nonlinear model predictive control and reinforcement learning is adopted. The pre-compensation signal, accuracy compensation parameters and environmental disturbance data are fused to solve a multi-objective optimization problem to generate calibration instructions. The reinforcement learning part is trained with a dual-delay deep deterministic policy gradient algorithm. The reward function simultaneously optimizes accuracy stability, calibration efficiency and equipment loss. S5: Based on the hybrid controller, the actuator is driven to achieve online dynamic calibration, and digital thread technology is integrated to realize closed-loop management of data throughout the entire life cycle. The transfer learning algorithm is used to continuously optimize the control strategy.

2. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The distributed optical fiber sensing network uses optical frequency domain reflective distributed optical fiber with a spatial resolution of 1 mm, a strain measurement accuracy of ±1 με, and a temperature measurement accuracy of ±0.1 °C. The micro laser interferometer is a nanoscale laser interferometer with a measurement accuracy of ±0.05 μm and a sampling frequency of 1 kHz.

3. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The real-time detection data includes fiber optic sensing data, laser interferometer data, and the instrument's own detection data. The spatiotemporal variation error includes temperature deformation error, vibration displacement error, and probe positioning error, of which temperature deformation error accounts for 50% of the total error, vibration displacement error accounts for 35% of the total error, and probe positioning error accounts for 15% of the total error.

4. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The fixture accuracy degradation trajectory is based on spatiotemporal variation error data and uses a Long Short-Term Memory (LSTM) network for dynamic trend modeling. Spatiotemporal variation error data from the past 24 hours are selected to construct a 3-layer LSTM network, which outputs the fixture accuracy degradation trend for the next hour, i.e., the fixture accuracy value corresponding to different time nodes.

5. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The pre-compensation signal is dynamically adjusted over time and is set according to the predicted increase in error within the next 30 minutes. The setting format is "increase the compensation amount by 'a' every 10 minutes".

6. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The adaptive Kalman filter algorithm optimizes the error compensation effect by adjusting the filter gain matrix in real time, including constructing state equations and observation equations for adaptive gain adjustment.

7. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The environmental disturbance data needs to be classified into levels. The classification criteria are based on historical data and detection error thresholds to classify environmental disturbances into low disturbances, medium disturbances, and high disturbances. The determining factors are temperature fluctuations, vibration acceleration, electromagnetic interference, and corresponding error contributions.

8. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The hybrid controller consists of two layers. The upper layer is an RL agent, which is responsible for dynamically adjusting the weight coefficients of the nonlinear model predictive control to balance accuracy stability, calibration efficiency and equipment loss. The lower layer is a nonlinear model predictive controller, which solves a multi-objective optimization problem based on pre-compensation signals, accuracy compensation parameters and environmental disturbance data, and generates calibration instructions.

9. The method for dynamic calibration and control of the detection accuracy of a non-standard automated inspection tool according to claim 1, characterized in that: The online dynamic calibration includes actuator driving, probe compensation, motion mechanism adjustment, environmental adaptive control, and real-time monitoring of calibration results. The transfer learning algorithm continuously optimizes the control strategy, which includes three steps: data modeling, data flow, and transfer learning continuous optimization. Data modeling includes three types of core data: static data, dynamic data, and historical data. Data flow goes through three levels: acquisition layer, processing layer, and application layer. Transfer learning continuous optimization includes three core elements: source domain data, target domain data, and transfer strategy.

Images