Self-adaptive adjustment pressure signal efficient confirmation pressing jig and method

Abstract

The invention relates to the technical field of jig control, in particular to a self-adaptive adjustment pressure signal efficient confirmation pressing jig and method, and the jig comprises a data collection module, a data preprocessing and feature extraction module, a self-adaptive control module, a quality prediction module and an equipment health management module. Compared with the prior art in which fixed control parameters are generally set by artificial experience, so that equipment cannot adapt to material fluctuation and environment change, and the crimping quality is unstable, data such as pressure, displacement and current are acquired in real time through a multi-source sensor, and the parameters are dynamically adjusted by adopting a self-adaptive closed-loop control algorithm, so that the crimping quality is improved. And extremely high crimping precision and consistency can be maintained under various working conditions, and the problem that the control precision of a traditional method is attenuated along with changes of time and working conditions is fundamentally solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of jig control, in particular to a self-adaptive pressure signal adjusting and efficient confirmation pressing jig and method. BACKGROUND

[0002] In the field of precision electronic manufacturing and automobile parts assembly, high-precision and reliable crimping process is the key link to ensure product function and quality. The accurate adjustment of pressure signal and the stability of pressing jig directly determine the crimping quality of connectors, terminals and other components, and then affect the reliability and service life of the whole product. Therefore, developing a technical solution that can self-adaptively adjust pressure and efficiently confirm crimping quality has important engineering application value.

[0003] The existing crimping process control technology mainly depends on preset fixed parameter program and simple threshold judgment, which has obvious limitations. First, the control process lacks self-adaptive ability and cannot be adjusted in real time according to material property fluctuation, equipment wear or environmental change, resulting in unstable crimping precision. Second, quality detection often uses post-inspection or single criterion method, which is not only low in efficiency but also has the risk of missing detection, and is difficult to cope with complex and variable process abnormalities. In addition, equipment maintenance mainly depends on regular maintenance or fault repair, lacks predictive management ability based on data, and is easy to cause unplanned downtime. The root cause of these problems lies in the lack of multi-source data fusion processing and intelligent decision-making ability in the existing technology, and each link is isolated, which cannot form a closed-loop optimization.

[0004] The present application solves the above problems by constructing an intelligent system integrating multi-source sensing, real-time data processing, adaptive control, quality prediction and equipment health management. The system synchronously collects multi-dimensional data such as pressure, displacement, current, etc., after preprocessing and feature extraction, on the one hand, it uses the model predictive controller optimized by reinforcement learning to realize high-precision real-time closed-loop control, on the other hand, it uses machine learning model to judge and analyze the root cause of the quality risk in real time, and monitors the equipment performance degradation trend based on long-term data. This data-driven integrated solution realizes the intelligentization of the whole chain from process control to quality assurance to equipment operation and maintenance, significantly improving the adaptability, reliability and production efficiency of the crimping process. SUMMARY

[0005] In order to overcome the problems proposed in the background art, the present application proposes a self-adaptive pressure signal adjusting and efficient confirmation pressing jig and method.

[0006] The technical solution of the present application is: a self-adaptive pressure signal adjusting and efficient confirmation pressing jig, comprising: a data acquisition module for acquiring multi-source real-time data in the crimping process, including pressure signal, servo motor current and torque feedback, displacement of pressure output end and encoder position signal; a data preprocessing and feature extraction module, configured to synchronize, filter and extract features from the collected data, to obtain real-time state features and process feature vectors; an adaptive control module, configured to receive the real-time state features and construct a system state vector to realize real-time closed-loop control of the crimping process; a quality prediction module, configured to receive the process feature vectors and make a quality risk judgment immediately after the crimping is completed; a device health management module, configured to analyze the performance degradation trend of key components based on long-term collected data and predict the remaining service life thereof.

[0007] Preferably, when the data preprocessing and feature extraction module synchronizes, filters and extracts features from the collected data to obtain real-time state features, it specifically includes the following steps: S11: receiving multi-source real-time data streams from the data acquisition module, and time-aligning all signals based on hardware timestamps to generate synchronized multi-dimensional time series data; S12: performing real-time digital filtering processing on each signal in the multi-dimensional time series data to suppress high-frequency noise and power frequency interference, wherein Kalman filtering algorithm is used for pressure signals and displacement signals, and a low-pass infinite impulse response digital filter is used for current and torque signals; S13: performing instantaneous calculation on the filtered real-time data streams to generate a real-time state feature vector for closed-loop control, wherein the real-time state feature vector includes the current pressure value, the pressure change rate, the current displacement value, the displacement change rate and the current current value; S14: outputting the real-time state feature vector to the adaptive control module.

[0008] Preferably, when the data preprocessing and feature extraction module synchronizes, filters and extracts features from the collected data to obtain process feature vectors, it specifically includes the following steps: S21: automatically identifying the starting point and the ending point of a single crimping process based on the preprocessed displacement signal sequence and the pressure signal sequence, wherein the starting point is determined when the displacement value first exceeds a preset contact threshold and the pressure value first exceeds a zero force threshold, and the ending point is determined when the displacement value returns to the vicinity of the starting point and the pressure value is lower than the zero force threshold; S22: segmenting the multi-signal synchronous data corresponding to a single workpiece from the continuous data stream according to the identified starting point and ending point; S23: calculating a preset feature group from the time domain and the frequency domain for the segmented single crimping process data segment; S24: standardize the calculated characteristic values to eliminate dimensional influence, and combine the processed characteristic values into a process characteristic vector of fixed dimension according to a predetermined order, and output to the quality prediction module.

[0009] wherein, in the calculation of the preset feature group from the single crimping process data segment obtained by segmentation in time domain and frequency domain, the preset feature group comprises: A11: a time domain statistical feature group, calculating the maximum value, minimum value, arithmetic mean value, variance, peak-to-peak value, and root mean square value of the pressure signal and displacement signal in the data segment; A12: a time domain shape feature group, comprising: A. calculating the area enclosed by the pressure-displacement curve as a measure of crimping energy consumption; B. calculating the average slope and maximum slope of the rising edge stage of the pressure signal; C. calculating the average slope of the falling edge stage of the pressure signal; D. identifying the global maximum value point of the pressure signal and recording the corresponding displacement value; E. calculating the duration of the pressure maintaining stage; F. calculating the dynamic time warping distance between the pressure signal curve and the predefined standard pressure curve as a measure of the difference between the current crimping process and the ideal process shape; A13: a time sequence feature group, calculating the total crimping time and the time required to reach the target pressure value from the starting point.

[0010] A14: a frequency domain feature group, performing fast Fourier transform on the pressure signal sequence in the pressure maintaining stage to extract the dominant frequency component.

[0011] As a preferred, when the starting point and ending point of a single crimping process are automatically identified based on the preprocessed displacement signal sequence and pressure signal sequence through an algorithm, it specifically comprises: S31: setting key thresholds including a contact threshold, a zero force threshold and a regression tolerance, and initializing the system state; S32: monitoring the displacement signal and the pressure signal, when the displacement value first exceeds the contact threshold and the pressure value first exceeds the zero force threshold, determining the crimping process starting point, and recording the current time stamp; S33: from the starting point, when the displacement value falls back to the regression tolerance range of the displacement value corresponding to the starting point and the pressure value is lower than the zero force threshold, determining the crimping process ending point, and recording the current time stamp; S34: according to the starting point time stamp and the ending point time stamp, intercepting all synchronous multi-source sensor data in this time period from the continuous signal stream to form a crimping process data segment corresponding to a single workpiece.

[0012] Further comprising a timeout processing step, specifically: starting a timer from the recording of the start point timestamp; when the maximum allowed crimping time is exceeded, and the end point determination condition is still not met, then forcibly terminate this cycle division, reset the system state, and generate a timeout alarm signal.

[0013] As a preferred embodiment, the adaptive control module specifically comprises: A21: a reinforcement learning agent for autonomously learning and dynamically adjusting the core parameters of the lower-level model predictive controller by analyzing historical crimping performance; A22: a model predictive controller for calculating and outputting precise control instructions in a rolling optimization manner using the parameters adjusted by the reinforcement learning agent and the system model; Wherein, the reinforcement learning agent updates its strategy after a single crimping process is completed, and the model predictive controller runs in real time within each sampling period.

[0014] As a preferred embodiment, the reinforcement learning agent is configured to: A31: the state space is the process feature vector of the historical crimping process or the summary statistics derived from the real-time state features; A32: the action space is the adjustment amount of the core control parameters of the model predictive controller; A33: the reward function is calculated based on the performance indicators generated after crimping, including at least the pass probability obtained by the quality prediction module, the total energy consumption of the crimping process, and the total crimping time, and the principle formula of the reward function is: ; Wherein, is the quality pass rate, is the energy consumption of this crimping, is the energy consumption normalization coefficient, is the crimping time, is the time normalization coefficient, , and are weight coefficients.

[0015] As a preferred embodiment, the model predictive controller is configured to receive the adjusted control parameters output by the reinforcement learning agent, and based on a simplified system dynamics model, the current real-time state feature vector, and the target pressure-displacement trajectory, to calculate the optimal control sequence within a future period of time in each control period, and output the first control instruction in the sequence to the servo driver; the optimization problem of the model predictive controller is represented as: ; Wherein, J is the cost function, is a system state vector, is a reference trajectory, is a state error weight matrix, is a control input weight matrix, is a control input, denotes a transpose operation, denotes a system dynamics model, denotes a lower limit of the control input, denotes an upper limit of the control input.

[0016] As preferred, the adaptive control module, when in operation, specifically comprises: S41: After completion of a single crimping process, the reinforcement learning agent receives a process feature vector as a state input, calculates a reward value obtained, and updates a policy network; S42: The reinforcement learning agent outputs an action adjustment amount for core parameters of the model predictive controller according to the updated policy; S43: At each control sampling period, the model predictive controller receives the adjusted parameters, a current real-time state feature vector, and a target trajectory, solves a constrained rolling optimization problem, and calculates an optimal control sequence; S44: The first control instruction in the optimal control sequence is output to a servo driver in real time, and precise crimping is performed.

[0017] As preferred, the quality prediction module, when receiving a process feature vector and immediately making a quality risk judgment after crimping is completed, specifically comprises: S51: The received process feature vector is input into a pre-trained machine learning classification model to output a probability value indicating that the crimped product can pass subsequent functional tests, wherein the machine learning model used is a gradient boosting decision tree model; S52: The obtained probability value is compared with a preset pass threshold; if the probability value is higher than the pass threshold, the product is determined to be qualified; if the probability value is lower than the pass threshold, the product is determined to be a potential unqualified product, and an explainability analysis process is automatically triggered; S53: An integrated explainability AI tool is called to trace the decision-making process of the machine learning classification model for the crimping process, and calculate the contribution of all features in the process feature vector to the low pass probability, wherein the explainability AI tool used is the SHAP algorithm; S54: A root cause analysis report is generated based on the contribution degree analysis result.

[0018] As preferred, the equipment health management module, when in operation, specifically comprises: S61: Continuously collect and store process feature vectors and real-time state features to form a time-series health data set corresponding to the device's working time; S62: Extract key performance indicators related to device wear from the health data set, including average working current of the servo motor, zero drift value of the pressure sensor, and deviation between encoder position and displacement sensor readings; S63: Perform time-series trend analysis on key performance indicators and establish a performance degradation model using predictive maintenance algorithms; S64: Based on the performance degradation model, predict the remaining useful life of the specified key components; S65: When the predicted remaining useful life is lower than the preset life threshold, generate structured warning information and push it to the management personnel.

[0019] As a preferred, it also includes a system calibration module for performing a standardized calibration process, automatically calculating sensor calibration coefficients and transmission system compensation parameters, and updating to the adaptive control module, the process includes: S71: Guide the user to install the standard calibration piece; S72: Automatically execute the pre-defined calibration compression program; S73: Collect sensor readings and compare them with standard values; S74: Automatically calculate the gain and bias coefficients of the pressure sensor and displacement sensor; S75: Automatically calculate the back lash compensation value of the transmission mechanism; S76: Automatically update the gain and bias coefficients and back lash compensation value to the adaptive control module.

[0020] The method for efficiently confirming the pressing jig by adaptively adjusting the pressure signal includes the following steps: S81: Data acquisition, synchronously collect pressure, current, torque, displacement and position signals during the compression process through multiple source sensors; S82: Real-time feature extraction, time align, filter and denoise the collected raw data, and calculate and generate an instantaneous state feature vector for real-time control; S83: Process feature extraction, segment a single compression process from the continuous data stream, and extract a multi-dimensional feature group in time and frequency domains to form a standardized process feature vector; S84: Adaptive control, use the reinforcement learning intelligent agent optimization model to predict the parameters of the controller, and use the controller to solve the optimal control instruction in real time to drive the servo motor to complete high-precision compression; S85: Quality prediction and root cause analysis, use a machine learning model to predict the quality pass probability based on process features, and perform explainable analysis on unqualified results to generate a root cause report; S86: Equipment health management, monitor the long-term degradation trend of key performance indicators, predict the remaining life of components, and give early maintenance warning when the life is insufficient; S87: System calibration, automatically calculate and update the compensation parameters of sensors and transmission systems by performing standardized calibration procedures.

[0021] Advantages of the present application: 1. Compared with the prior art which generally uses artificial experience to set fixed control parameters, the device cannot adapt to material fluctuations and environmental changes, resulting in unstable crimping quality. The present scheme collects pressure, displacement, current and other data in real time through multiple source sensors, and dynamically adjusts parameters using an adaptive closed-loop control algorithm, which can ensure high crimping precision and consistency under various working conditions, and fundamentally solves the problem of declining control precision over time and working conditions in traditional methods; 2. Compared with the prior art which mainly relies on single threshold judgment or manual sampling inspection, there are high risk of quality misjudgment and misjudgment, low detection efficiency and other problems. The present scheme innovatively extracts process features from time domain, frequency domain and other dimensions, and uses a machine learning model for real-time quality prediction, achieving automatic and high-precision quality risk assessment, significantly improving detection accuracy and efficiency, and ensuring zero outflow of defective products; 3. Compared with the prior art which lacks collaborative optimization of energy consumption and production efficiency, and often trades off high energy consumption for high quality, resulting in serious resource waste. The present scheme designs a multi-objective reward function to drive the reinforcement learning agent to actively optimize production rhythm and energy consumption while ensuring crimping quality, achieving intelligent balance and Pareto improvement among quality, efficiency and energy consumption; 4. Compared with the prior art which needs manual intervention to segment crimping process data, has low efficiency, strong subjectivity and is prone to errors. The present scheme uses a dual threshold judgment mechanism of displacement and pressure signals, combined with timeout processing protection, to achieve automatic and high-precision crimping process data segmentation, providing a complete and reliable data basis for subsequent feature extraction and quality analysis, and significantly improving the system automation level; 5. Compared with the prior art which is difficult to effectively cope with control deviation caused by time-varying factors such as process fluctuations and equipment aging, and lacks self-optimization ability. The present scheme uses a double-layer intelligent control architecture combining reinforcement learning and model predictive control, which enables the system to have continuous self-learning and real-time dynamic optimization capabilities, significantly improving the system's adaptability and control robustness to complex working conditions; 6、Relative to the prior art, the unqualified products can only be qualitatively judged, and the quick root cause positioning ability is lacked, resulting in low problem analysis efficiency, the scheme innovatively integrates the explainable AI tool, can automatically analyze the feature contribution degree of the quality prediction result and generate a root cause report, provides strong data-driven decision support for process optimization and quality traceability, and greatly improves the problem solving efficiency; 7、Relative to the prior art, each link is operated in isolation and lacks cooperation, resulting in information fragmentation and low overall efficiency, the scheme realizes integrated intelligent operation and maintenance from real-time control to quality judgment, from health management to system calibration through the whole-chain data penetration and intelligent module cooperative working mechanism, builds a complete crimping process digital twin system, and comprehensively improves production efficiency and process management level. BRIEF DESCRIPTION OF DRAWINGS

[0022] Fig. 1 A structure schematic diagram of the adaptive pressure signal efficient confirmation pressing jig is shown. Fig. 2 A flowchart schematic diagram of the method of the adaptive pressure signal efficient confirmation pressing jig is shown. DETAILED DESCRIPTION

[0023] The application will be further described below in combination with the drawings and examples.

[0024] Please refer to Figs. 1-2 The application provides an example: an adaptive pressure signal efficient confirmation pressing jig, which comprises: A data acquisition module is used for acquiring multi-source real-time data in a crimping process, including pressure signals, servo motor current and torque feedback, pressure output end displacement, and encoder position signals. A data preprocessing and feature extraction module is used for synchronizing, filtering and feature extracting the collected data to obtain real-time state features and process feature vectors. An adaptive control module is used for receiving real-time state features and constructing a system state vector to realize real-time closed-loop control of the crimping process. A quality prediction module is used for receiving process feature vectors and immediately judging quality risks after crimping is completed. An equipment health management module is used for analyzing the performance degradation trend of key components based on long-term collected data and predicting the remaining useful life thereof.

[0025] In this embodiment, the technical solution achieves precise adaptive control and full lifecycle management of the crimping process through multi-source real-time data acquisition and intelligent processing. Its beneficial effects are as follows: on the one hand, it significantly improves product quality consistency—real-time closed-loop control ensures pressure accuracy, and combined with immediate quality risk assessment after crimping, it effectively prevents defective products from leaving the site; on the other hand, it enhances the level of intelligent equipment operation and maintenance—life prediction of key components based on long-term data supports preventive maintenance, reduces unplanned downtime, and reduces reliance on operator experience, thereby comprehensively improving production efficiency and equipment utilization.

[0026] Preferably, the data preprocessing and feature extraction module, when synchronizing, filtering, and extracting features from the collected data to obtain real-time state features, specifically includes: S11: Receives multi-source real-time data streams from the data acquisition module, and performs time alignment of all signals based on hardware timestamps to generate synchronized multi-dimensional time series data; S12: Perform real-time digital filtering on each signal in the multidimensional time series data to suppress high-frequency noise and power frequency interference. Specifically, Kalman filtering algorithm is used for pressure signal and displacement signal, and low-pass infinite impulse response digital filter is used for current and torque signal. S13: Perform instantaneous calculations on the filtered real-time data stream to generate a real-time state feature vector for closed-loop control. The real-time state feature vector includes the current pressure value, pressure change rate, current displacement, displacement change rate, and current current value. S14: Output the real-time state feature vector to the adaptive control module.

[0027] In this embodiment, the technical solution constructs a high-precision, low-latency real-time state feature vector by time alignment of multi-source signals based on hardware timestamps, targeted real-time digital filtering (Kalman filtering for pressure / displacement signals and low-pass IIR filtering for current / torque signals), and instantaneous calculation of key dynamic features (such as pressure change rate and displacement change rate). Its beneficial effects are: significantly improving the response speed and accuracy of closed-loop control, ensuring the reliability of state perception by effectively suppressing noise interference, and providing a stable and clean decision-making basis for the adaptive control module, thereby fundamentally guaranteeing the consistency and stability of the pressing process.

[0028] Preferably, the data preprocessing and feature extraction module, when synchronizing, filtering, and extracting features from the collected data to obtain the process feature vector, specifically includes: S21: Based on the preprocessed displacement signal sequence and pressure signal sequence, the algorithm automatically identifies the start and end points of a single pressing process. The start point is determined when the displacement value first exceeds the preset contact threshold and the pressure value first exceeds the zero force threshold. The end point is determined when the displacement value returns to the vicinity of the start point and the pressure value is lower than the zero force threshold. S22: Based on the identified start and end points, segment the multi-signal synchronization data segment corresponding to a single workpiece from the continuous data stream; S23: For a single crimping process data segment obtained by segmentation, calculate the preset feature group from multiple dimensions in the time domain and frequency domain; S24: Standardize the calculated feature values ​​to eliminate the influence of dimensions, and combine the processed feature values ​​into a fixed-dimensional process feature vector in a predetermined order, and output it to the quality prediction module.

[0029] In this embodiment, the technical solution achieves precise data segmentation by automatically identifying the start and end points of the crimping process (based on threshold determination of displacement and pressure), and performs multi-dimensional feature extraction and standardization processing on the multi-signal data segments of a single crimping cycle in the time and frequency domains, ultimately generating a fixed-dimensional process feature vector. Its beneficial effects are: it achieves precise and automated characterization of each crimping process, provides highly structured, dimensionless interference-free input data for the quality prediction module, greatly improves the accuracy and consistency of quality risk assessment, and lays a reliable foundation for data-driven process optimization.

[0030] Specifically, when calculating a preset feature set from multiple dimensions in the time and frequency domains for a single segment of the crimping process data obtained from the segmentation, the preset feature set includes: A11: Time-domain statistical feature group, calculates the maximum, minimum, arithmetic mean, variance, peak-to-peak value, and root mean square value of pressure and displacement signals within this data segment; A12: Temporal morphological feature group, including: A. Calculate the area enclosed by the pressure-displacement curve as a measure of the energy consumption of the crimping process; B. Calculate the average slope and maximum slope during the rising edge phase of the pressure signal; C. Calculate the average slope of the pressure signal during the falling edge phase; D. Identify the global maximum value of the pressure signal and record its corresponding displacement value; E. Calculate the duration of the pressure holding phase; F. Calculate the dynamic time warping distance between the pressure signal curve and the predefined standard pressure curve, as a measure of the difference between the current crimping process and the ideal process morphology; A13: Timing characteristic group, calculates the total pressing time and the time required from the starting point to reaching the target pressure value.

[0031] A14: Frequency domain feature group, which performs a fast Fourier transform on the pressure signal sequence during the pressure holding phase to extract its dominant frequency components.

[0032] In this embodiment, the technical solution extracts refined features from a single crimping process data segment across four dimensions: time-domain statistics, time-domain morphology, time series, and frequency domain. This constructs a comprehensive, multi-faceted feature set encompassing energy consumption, slope, pressure holding characteristics, curve morphology differences (dynamic time regularization distance), and frequency domain components. The beneficial effect is that this comprehensive feature set can comprehensively and profoundly characterize the dynamic behavior and intrinsic properties of the crimping process, providing the quality prediction module with a rich information base far exceeding that of a single indicator. This significantly improves the accuracy, robustness, and ability to identify complex anomaly patterns in quality risk assessment.

[0033] Preferably, when automatically identifying the start and end points of a single pressing process based on the preprocessed displacement signal sequence and pressure signal sequence using an algorithm, the specific steps include: S31: Set key thresholds including contact threshold, zero force threshold and regression tolerance, and initialize the system state; S32: Monitor displacement and pressure signals. When the displacement value exceeds the contact threshold for the first time and the pressure value exceeds the zero force threshold for the first time, it is determined to be the starting point of the crimping process, and the current timestamp is recorded. S33: From the starting point, when the displacement value falls back to the regression tolerance range near the displacement value corresponding to the starting point and the pressure value is lower than the zero force threshold, the pressing process is determined to be at its end, and the current timestamp is recorded. S34: Based on the start and end timestamps, extract all synchronous multi-source sensor data from the continuous signal stream within that time period to form a crimping process data segment corresponding to a single workpiece.

[0034] In this embodiment, This technical solution achieves accurate and automated identification of the start and end points of a single crimping process by setting key parameters such as contact threshold, zero-force threshold, and regression tolerance, and by using a dual condition judgment based on displacement and pressure signals (displacement and pressure exceeding the threshold together indicate the start, while displacement falling back to the origin tolerance range and pressure falling below the threshold indicate the end). Its advantages are: the method is logically rigorous, has strong anti-interference capabilities, and can reliably segment complete and synchronous multi-source sensor data segments from a continuous data stream, providing an accurate data foundation for subsequent feature extraction and quality analysis, and completely avoiding the subjectivity and inefficiency of manual segmentation.

[0035] The system also includes a timeout handling step, which is as follows: starting from the timestamp of the recording start point, the timer is started; if the end point determination condition is not met after the maximum allowable crimping time has been exceeded, the current cycle division is forcibly terminated, the system state is reset, and a timeout alarm signal is generated.

[0036] In this embodiment, the technical solution introduces a timeout handling mechanism. After recording the starting point, a timer is started. If the pressing time exceeds the preset maximum allowable duration and the termination condition is not met, the system will forcibly terminate the current cycle, reset the state, and trigger an alarm. Its beneficial effects are: effectively preventing indefinite system waiting or freezing due to equipment failure or process abnormalities, greatly improving the automation level and robustness of the production line, avoiding production interruptions, equipment damage, or batch quality accidents, and ensuring the continuity and safety of the production process.

[0037] Preferably, the adaptive control module specifically includes: A21: A reinforcement learning agent used to autonomously learn and dynamically adjust the core parameters of the lower-level model predictive controller by analyzing historical crimping performance; A22: Model Predictive Controller, used to calculate and output precise control commands in a rolling optimization manner using parameters and system model adjusted by reinforcement learning agents; In this system, the reinforcement learning agent updates its policy after a single pressing process is completed, and the model prediction controller runs in real time during each sampling period.

[0038] In this embodiment, the technical solution constructs a two-layer architecture consisting of a reinforcement learning agent and a model predictive controller to achieve adaptive optimization control of the pressing process: after each pressing, the upper-layer reinforcement learning agent autonomously learns and adjusts the core parameters of the lower-layer controller based on historical performance data; the lower-layer model predictive controller uses the updated parameters and system model to perform rolling optimization and output real-time control commands in each sampling period. Its beneficial effects are: this architecture cleverly combines the advantages of long-term strategy optimization and instantaneous real-time control, enabling the system to not only quickly respond to dynamic disturbances during the process but also continuously learn from historical data and improve itself, thereby significantly improving the control accuracy, adaptability, and overall process stability of complex pressing processes.

[0039] As a preferred configuration, the reinforcement learning agent is configured as follows: A31: The state space is the process feature vector of the historical pressing process or the summary statistics derived from the real-time state features; A32: Action space is the adjustment amount of the core control parameters of the model predictive controller; A33: The reward function is calculated based on the performance indicators generated after crimping is completed. The performance indicators include at least the pass probability, total energy consumption of the crimping process, and total crimping time obtained from the quality prediction module. The principle formula of the reward function is as follows: ; in, For quality pass rate, The energy consumption for this crimping process, This is the energy consumption normalization coefficient. This refers to the crimping time. The time normalization coefficient is... , and These are the weighting coefficients.

[0040] In this embodiment, the technical solution defines the state space of the reinforcement learning agent as historical process features and the action space as the adjustment amount of the controller's core parameters. It innovatively designs a multi-objective reward function that integrates quality pass probability, pressing energy consumption, and time to drive the agent in strategy optimization. Its beneficial effects are: this design enables the agent to automatically and continuously find the optimal adjustment strategy for control parameters, ensuring high-quality output while proactively optimizing production cycle time and energy consumption. This achieves an intelligent balance and Pareto improvement among quality, efficiency, and energy consumption, significantly reducing the system's reliance on manual adjustments and improving the overall process adaptability and economy.

[0041] Preferably, the model predictive controller is configured to receive adjusted control parameters output by the reinforcement learning agent, and based on a simplified system dynamics model, the current real-time state feature vector, and the target pressure-displacement trajectory, to perform rolling optimization calculations for the optimal control sequence over a future period within each control cycle, and output the first control command in the sequence to the servo driver; the optimization problem of the model predictive controller is expressed as: ; Where J is the cost function, Let be the system state vector. For reference trajectory, Here is the state error weight matrix. To control the input weight matrix, To control the input, This indicates the transpose operation. Represents the system dynamics model. This indicates the lower limit of the control input. This indicates the upper limit of the control input.

[0042] In this embodiment, the technical solution configures a Model Predictive Controller (MPC) to solve a constrained rolling optimization problem (aiming to minimize state error and control input) within each control cycle based on parameters adjusted by reinforcement learning, a simplified system model, real-time state, and target trajectory, and outputs the optimal control command. Its advantages are: this design integrates prediction, optimization, and feedback correction, not only proactively compensating for system dynamic delays and disturbances, but also achieving precise pressure tracking under strict physical constraints, thereby greatly improving the control accuracy, response speed, and stability of the pressing process, while ensuring operational safety.

[0043] Preferably, the adaptive control module includes the following components during operation: S41: After a single pressing process is completed, the reinforcement learning agent receives the process feature vector as the state input, calculates the obtained reward value, and updates the policy network. S42: The reinforcement learning agent outputs the action adjustment amount for the core parameters of the model predictive controller based on the updated policy; S43: In each control sampling period, the model predicts that the controller receives the adjusted parameters, the current real-time state feature vector and the target trajectory, solves the constrained rolling optimization problem, and calculates the optimal control sequence. S44: Outputs the first control command in the optimal control sequence to the servo driver in real time to perform precise pressing.

[0044] In this embodiment, the technical solution precisely couples the policy update of the reinforcement learning agent (based on features and rewards after the process is completed) with the real-time rolling optimization of the model predictive controller (based on adjusted parameters and the current state) in a temporal sequence, constructing a two-layer collaborative workflow of "slow learning-fast control". Its beneficial effects are: this mechanism enables the system to continuously self-optimize its control strategy in the long term, pursuing the comprehensive optimality of quality, efficiency, and energy consumption, while also enabling high-precision, constrained real-time decision-making and execution for each sampling period in the instantaneous dimension. This achieves unprecedented control adaptability, process stability, and overall production efficiency in dynamically changing production environments.

[0045] Preferably, when the quality prediction module receives the feature vector during the process and immediately performs a quality risk assessment after crimping, it specifically includes: S51: Input the received process feature vector into a pre-trained machine learning classification model, and output a probability value representing the probability that the crimped product can pass the subsequent functional test. The machine learning model used is a gradient boosting decision tree model. S52: Compare the obtained probability value with the preset pass threshold; if the probability value is higher than the pass threshold, the product is determined to be qualified; if the probability value is lower than the pass threshold, the product is determined to be a potential non-conforming product, and the interpretability analysis process is automatically triggered. S53: Call the integrated interpretable AI tool to perform source analysis on the decision-making process of the machine learning classification model for this crimping process, and calculate the contribution of all feature pairs in the process feature vector to the low pass probability. The interpretable AI tool used is the SHAP algorithm. S54: Generate a root cause analysis report based on the contribution analysis results.

[0046] In this embodiment, the technical solution integrates a gradient boosting decision tree model for real-time quality probability prediction and combines it with an interpretable AI tool based on the SHAP algorithm to construct a quality risk assessment system that integrates automatic judgment and intelligent diagnosis. Its beneficial effects are: it can not only achieve high-accuracy conformity determination instantly after pressing, greatly improving inspection efficiency; but also automatically trace the root causes of potentially non-conforming products, accurately locating key process parameter characteristics that lead to quality risks. This provides powerful data-driven decision support for rapid production process adjustments, process optimization, and quality traceability, significantly improving the intelligence level and problem-solving efficiency of the production line.

[0047] Preferably, the equipment health management module includes the following features during operation: S61: Continuously collect and store process feature vectors and real-time status features to form a time-series health dataset corresponding to the device's working duration; S62: Extract key performance indicators related to equipment wear from the health dataset, including the average operating current of the servo motor, the zero drift value of the pressure sensor, and the deviation between the encoder position and the displacement sensor reading; S63: Perform time-series trend analysis on key performance indicators and establish a performance degradation model using predictive maintenance algorithms; S64: Based on a performance degradation model, predict the remaining service life of specified critical components; S65: When the predicted remaining useful life is lower than the preset life threshold, generate a structured early warning message and push it to the management personnel.

[0048] In this embodiment, the technical solution constructs a time-series health dataset by continuously collecting process and status data. It focuses on extracting key performance indicators such as servo motor current, pressure sensor zero drift, and encoder and displacement sensor deviations for time-series trend analysis and degradation modeling. This enables accurate monitoring of performance degradation of key components and prediction of their remaining useful life. Its beneficial effects include: upgrading equipment maintenance strategies from reactive and periodic maintenance to predictive maintenance; generating structured early warnings based on data-driven models; significantly reducing unplanned downtime, lowering maintenance costs, preventing production interruptions or quality risks due to component failures, and comprehensively improving overall equipment efficiency and the reliability of production plans.

[0049] Preferably, a system calibration module is also included, which performs a standardized calibration procedure, automatically calculates sensor calibration coefficients and transmission system compensation parameters, and updates them to the adaptive control module. The execution procedure includes: S71: Guides users to install standard calibration kits; S72: Automatically executes predefined calibration crimping procedures; S73: Collect sensor readings and compare them with standard values; S74: Automatically calculates the gain and bias coefficients of the pressure sensor and displacement sensor; S75: Automatically calculates the backlash compensation value of the transmission mechanism; S76: Automatically update the gain, bias coefficient, and backlash compensation value to the adaptive control module.

[0050] In this embodiment, the technical solution integrates a fully automated system calibration module, guiding the user to complete the installation of standard calibration components, execute preset programs, and automatically calculate sensor gain / bias coefficients and transmission mechanism backlash compensation values, ultimately updating the parameters to the control system. Its advantages include: completely eliminating the subjective errors and operational inconsistencies of traditional manual calibration; ensuring long-term stability of sensor measurement accuracy and transmission positioning accuracy through standardized processes; significantly improving the overall control accuracy and product quality consistency of the equipment; and substantially reducing the technical threshold and time cost of equipment maintenance.

[0051] A method for efficiently confirming the pressure jig by adaptively adjusting the pressure signal includes the following steps: S81: Data acquisition, synchronously acquiring pressure, current, torque, displacement and position signals during the crimping process through multi-source sensors; S82: Real-time feature extraction, which performs time alignment, filtering and noise reduction on the collected raw data, and calculates and generates instantaneous state feature vectors for real-time control; S83: Process feature extraction, segmenting a single crimping process from a continuous data stream, and extracting multi-dimensional feature groups in the time and frequency domains to form a standardized process feature vector; S84: Adaptive control, which uses a reinforcement learning agent to optimize the model and predict the parameters of the controller, and the controller solves the optimal control command in real time to drive the servo motor to complete high-precision pressing. S85: Quality Prediction and Root Cause Analysis. This involves using machine learning models to predict the probability of passing quality based on process characteristics, performing interpretability analysis on nonconforming results, and generating a root cause report. S86: Equipment health management, monitors the long-term degradation trend of key performance indicators, predicts the remaining life of components, and issues maintenance warnings in advance when the life is insufficient. S87: System calibration, automatically calculates and updates compensation parameters for sensors and drive systems by performing a standardized calibration procedure.

[0052] In this embodiment, the technical solution integrates multi-source sensing, real-time data processing, adaptive control, AI quality prediction, equipment health management, and automatic calibration to construct a data-driven, closed-loop optimized intelligent crimping system. Its core benefits lie in achieving end-to-end intelligent management of the crimping process, from real-time control to long-term management. It not only ensures extremely high crimping accuracy and consistency through adaptive algorithms but also instantly identifies potential quality issues and accurately pinpoints their root causes. Furthermore, predictive maintenance extends equipment lifespan, ultimately significantly improving production efficiency and product yield while substantially reducing maintenance costs and reliance on human experience.

[0053] Example 1: High-precision electronic connector crimping application This embodiment applies to the crimping production process of precision electronic connectors. This scenario demands extremely high precision, consistency, and traceability of the crimping force; even the slightest deviation can lead to signal transmission interruption.

[0054] Detailed implementation process: System Startup and Data Acquisition: The operator places the wire harness and connector to be crimped into the fixture. After the system is started, the data acquisition module immediately begins operation. A high-frequency pressure sensor monitors the pressure changes of the indenter in real time; the servo motor driver and encoder continuously provide feedback on current, torque, and precise position information; a high-precision displacement sensor synchronously measures the actual displacement of the indenter. All of these are based on hardware timestamps for synchronous data acquisition.

[0055] Real-time control and feature extraction: The data preprocessing and feature extraction module processes the raw data. First, Kalman filtering is used to reduce noise in the pressure and translation signals, and a low-pass IIR filter is used to process the current and torque signals to obtain a clean data stream. Then, the real-time state feature vector (including instantaneous pressure value, pressure change rate, displacement, displacement change rate, and current value) is calculated and output to the adaptive control module in real time.

[0056] The adaptive control module begins operation. The model predictive controller receives the current state feature vector and the target pressure-displacement trajectory. The core parameters of the model predictive controller (such as the weight matrix and prediction time domain) are optimized by the reinforcement learning agent based on data from historical high-quality pressing processes. Within each millisecond-level control cycle, the model predictive controller solves a constrained optimization problem, calculates the optimal motor control command, and immediately sends it to the servo driver, achieving precise and compliant control of the pressing force.

[0057] Process Analysis and Quality Assessment: After the crimping action is completed, the data preprocessing and feature extraction module initiates process analysis. First, based on the preset "contact threshold" and "zero force threshold," it automatically identifies the start and end points of this crimping operation, accurately segmenting the data segment of this crimping from the continuous data stream. Then, it performs in-depth feature mining on the data segment, calculating process feature vectors including time-domain statistical features (maximum value, variance, etc.), time-domain morphological features (crimping energy, rise / fall slope, holding time, DTW distance from the standard curve), time-series features (total crimping time), and frequency-domain features (dominant frequency during the holding phase).

[0058] Then, it is immediately sent to the quality prediction module. Based on this rich feature set, the pre-trained gradient boosting decision tree model outputs a probability value (e.g., 99.5%) that the connector passes the electrical test. Since this value is higher than the preset pass threshold (e.g., 98%), the system determines the product to be qualified. If the probability value is lower than the threshold, the system will automatically trigger the SHAP algorithm for interpretability analysis, immediately generating a report indicating the specific reason, such as "insufficient pressure rise slope" or "excessive fluctuation during the pressure holding stage," guiding the operator to make adjustments.

[0059] Health Management and System Calibration: The equipment health management module runs continuously in the background. It collects characteristic data from all crimping processes, focusing on trends in key indicators such as average servo motor current and pressure sensor zero drift. When the analysis model predicts that the remaining useful life of a critical component is about to end, it sends an early warning to maintenance personnel, scheduling planned maintenance to avoid sudden downtime.

[0060] Simultaneously, the system calibration module is activated at regular intervals or when the system detects a slight deviation. The system guides the operator to install standard calibration components and automatically executes the calibration crimping procedure. By comparing sensor readings with standard values, the module automatically calculates new gain, bias coefficient, and backlash compensation values, and silently updates them to the adaptive control module, ensuring the long-term accuracy of the system.

[0061] Example 2: Application of Pin Crimping for Automotive Airbag Controller (ECU) This embodiment applies to pin crimping in automotive safety system ECUs. This scenario places extreme demands on reliability and 100% quality traceability, while also needing to adapt to the high-intensity pace of mass production.

[0062] Detailed implementation process: System Startup and Data Acquisition: On the automotive parts production line, the fixture automatically picks up the ECU board and pins. After system startup, the data acquisition module synchronously acquires all signals during the crimping process at extremely high frequency.

[0063] Real-time control and timeout handling: The data preprocessing and feature extraction module filters and generates state feature vectors in real time. The MPC controller in the adaptive control module uses parameters optimized by a reinforcement learning agent to perform rolling optimization and output control commands to ensure that the crimping force and depth of each pin are completely consistent.

[0064] Meanwhile, if a crimping operation exceeds the maximum allowable time due to foreign objects or other reasons, the system will forcibly terminate and issue an alarm to prevent damage to the ECU board. The system will also immediately mark the workstation as requiring inspection, thus avoiding batch defects.

[0065] Full-process quality traceability and root cause analysis: After each pin is crimped, the system generates a process feature vector containing dozens of dimensions, and the quality prediction module quickly determines its pass / fail status.

[0066] Predictive maintenance and zero manual calibration: The equipment health management module analyzes the trend of gradually increasing servo motor current and predicts two weeks in advance that the motor may have insufficient torque due to wear. It then orders spare parts in advance and arranges for replacement on the weekend to avoid unplanned downtime.

[0067] When the system prompts that calibration is required, they only need to follow the instructions of the system calibration module, put in the standard block, press the confirmation button, and the rest of the process is completed automatically, ensuring a high degree of consistency in equipment accuracy after maintenance by different shifts and different operators.

[0068] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An adaptive pressure signal-adjusting, high-efficiency confirmation pressing fixture, characterized in that: The method comprises the following steps: A data acquisition module is used to collect multi-source real-time data during the crimping process, including pressure signals, servo motor current and torque feedback, pressure output end displacement, and encoder position signals; A data preprocessing and feature extraction module is used to synchronize, filter, and extract features from the collected data to obtain real-time state features and process feature vectors; An adaptive control module is used to receive real-time state features and construct a system state vector to achieve real-time closed-loop control of the crimping process; A quality prediction module is used to receive the process feature vector and make a quality risk judgment immediately after the crimping is completed; A device health management module is used to analyze the performance degradation trend of key components based on long-term collected data and predict their remaining useful life.

2. The adaptive pressure signal high-efficiency confirmation press jig according to claim 1, wherein: When the data preprocessing and feature extraction module synchronizes, filters, and extracts features from the collected data to obtain real-time state features, it specifically includes the following steps: S11: Receive multi-source real-time data streams from the data acquisition module, and time-align all signals based on hardware timestamps to generate synchronized multi-dimensional time series data; S12: Perform real-time digital filtering on each signal in the multi-dimensional time series data to suppress high-frequency noise and power frequency interference, wherein Kalman filtering algorithm is used for pressure signals and displacement signals, and low-pass infinite impulse response digital filter is used for current and torque signals; S13: Perform instantaneous calculation on the filtered real-time data stream to generate a real-time state feature vector for closed-loop control, wherein the real-time state feature vector includes current pressure value, pressure change rate, current displacement value, displacement change rate, and current current value; S14: Output the real-time state feature vector to the adaptive control module.

3. The adaptive pressure signal high-efficiency confirmation press jig according to claim 2, characterized in that: When the data preprocessing and feature extraction module synchronizes, filters, and extracts features from the collected data to obtain process feature vectors, it specifically includes the following steps: S21: Automatically identify the start point and end point of a single crimping process based on the preprocessed displacement signal sequence and pressure signal sequence, wherein the start point is determined when the displacement value first exceeds a preset contact threshold and the pressure value first exceeds a zero force threshold, and the end point is determined when the displacement value returns to the vicinity of the start point and the pressure value is lower than the zero force threshold; S22: According to the identified start point and end point, segment the continuous data stream to obtain multi-signal synchronous data corresponding to a single workpiece; S23: Calculate a preset feature group from the segmented single crimping process data segment in time and frequency domains; S24: Standardize each feature value to eliminate dimensional influence, and combine the processed feature values in a predetermined order into a fixed-dimensional process feature vector, and output to the quality prediction module.

4. The adaptive pressure signal high-efficiency confirmation press jig according to claim 3, characterized in that: When the start point and end point of a single crimping process are automatically identified based on the preprocessed displacement signal sequence and pressure signal sequence, it specifically includes the following steps: S31: Set key thresholds including contact threshold, zero force threshold, and regression tolerance, and initialize the system state; S32: Monitor the displacement signal and the pressure signal, and when the displacement value first exceeds the contact threshold value and the pressure value first exceeds the zero force threshold value at the same time, determine that the crimping process starting point is reached, and record the current timestamp; S33: From the starting point, when the displacement value falls to the vicinity of the displacement value corresponding to the starting point within the regression tolerance range and the pressure value is lower than the zero force threshold value at the same time, determine that the crimping process ending point is reached, and record the current timestamp; S34: According to the starting point timestamp and the ending point timestamp, all synchronized multi-source sensor data in the time period is intercepted from the continuous signal stream to form a crimping process data segment corresponding to a single workpiece.

5. The adaptive pressure signal high-efficiency confirmation press jig according to claim 4, characterized in that: The adaptive control module specifically includes: A21: Reinforcement learning agent, for autonomously learning and dynamically adjusting the core parameters of the lower layer model predictive controller by analyzing historical crimping performance; A22: Model predictive controller, for calculating and outputting accurate control instructions in a rolling optimization manner by using the parameters adjusted by the reinforcement learning agent and the system model; Wherein, the reinforcement learning agent updates its strategy after a single crimping process is completed, and the model predictive controller is running in real time in each sampling period.

6. The adaptive pressure signal high-efficiency confirmation press jig according to claim 5, characterized in that: The reinforcement learning agent is configured as: A31: The state space is the process feature vector of the historical crimping process or the abstract statistics derived from the real-time state features; A32: The action space is the adjustment amount of the core control parameters of the model predictive controller; A33: The reward function is calculated based on the performance indicators generated after crimping is completed.

7. The adaptive pressure signal high-efficiency confirmation press jig according to claim 6, characterized in that: When the adaptive control module is working, it specifically includes: S41: After a single crimping process is completed, the reinforcement learning agent receives the process feature vector as the state input, calculates the reward value, and updates the strategy network; S42: The reinforcement learning agent outputs the action adjustment amount of the core parameters of the model predictive controller according to the updated strategy; S43: In each control sampling period, the model predictive controller receives the adjusted parameters, the current real-time state feature vector and the target trajectory, solves the rolling optimization problem with constraints, and calculates the optimal control sequence; S44: The first control instruction in the optimal control sequence is output to the servo driver in real time to perform accurate crimping.

8. The adaptive pressure signal high-efficiency confirmation press jig according to claim 7, characterized in that: When the quality prediction module receives the process feature vector and immediately judges the quality risk after crimping is completed, it specifically includes: S51: Input the received process feature vector into the pre-trained machine learning classification model to output a probability value indicating that the crimped product can pass the subsequent functional test, wherein the machine learning model used is a gradient boosting decision tree model; S52: Compare the obtained probability value with the preset pass threshold value; if the probability value is higher than the pass threshold value, the product is determined to be qualified; if the probability value is lower than the pass threshold value, the product is determined to be a potential unqualified product, and an explainability analysis process is automatically triggered; S53: Call the integrated explainability AI tool to trace the decision-making process of the machine learning classification model for this crimping process, and calculate the contribution degree of all features in the process feature vector to the low pass probability, wherein the explainability AI tool used is the SHAP algorithm; S54: generating a root cause analysis report based on the contribution degree analysis result.

9. The adaptive adjustment pressure signal high-efficiency confirmation press jig according to claim 8, characterized in that: The device health management module, when in operation, specifically comprises: S61: continuously collecting and storing process feature vectors and real-time state features to form a time series health data set corresponding to the device working time; S62: extracting key performance indicators related to device wear from the health data set, including average working current of the servo motor, pressure sensor zero drift value, and deviation between encoder position and displacement sensor readings; S63: performing time series trend analysis on the key performance indicators, and establishing a performance degradation model using a predictive maintenance algorithm; S64: predicting the remaining useful life of the specified key components based on the performance degradation model; S65: generating structured warning information and pushing it to the management personnel when the predicted remaining useful life is lower than the preset life threshold.

10. A method for adaptive adjustment of a pressure signal to efficiently confirm a press jig, characterized by: The method comprises the following steps: S81: data acquisition, synchronously collecting pressure, current, torque, displacement and position signals during the crimping process through multi-source sensors; S82: real-time feature extraction, time alignment, filtering and noise reduction of the collected raw data, and calculation of instantaneous state feature vectors for real-time control; S83: process feature extraction, segmenting individual crimping processes from continuous data streams, and extracting multi-dimensional feature groups in time and frequency domains to form standardized process feature vectors; S84: adaptive control, using a reinforcement learning agent optimization model to predict the parameters of the controller, and using the controller to solve the optimal control instructions in real time to drive the servo motor to complete high-precision crimping; S85: quality prediction and root cause analysis, using a machine learning model to predict the probability of quality passing based on process features, and performing explainability analysis on unqualified results to generate a root cause report; S86: device health management, monitoring the long-term degradation trend of key performance indicators, predicting the remaining life of components, and issuing maintenance warnings in advance when the life is insufficient; S87: system calibration, automatically calculating and updating the compensation parameters of sensors and transmission systems by executing a standardized calibration process.

Images