Data mining method applied to hardware product control system

Abstract

The invention belongs to the field of data mining of hardware product control systems, and particularly relates to a data mining method applied to a hardware product control system, which comprises the following steps: deploying a sensor to collect data in real time and process the collected data; carrying out adaptive sampling and constructing a feature screening unit; predicting the quality of the product and deploying an early warning system to monitor the quality of the product; constructing an optimization feedback mechanism to optimize a prediction result; according to the method, real-time monitoring and high-precision quality prediction are realized, the prediction accuracy is improved from about 75% to about 86%, the false alarm rate is reduced from about 12% to about 6%, and the downtime is reduced by about 30%, so that the fault positioning and recovery time is remarkably shortened, and the rework rate is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data mining of hardware control systems, and particularly relates to a data mining method applied to a hardware control system. BACKGROUND

[0002] A data mining method and system applied to a hardware control system are disclosed in Chinese application No. CN202411455398.8. The application accurately identifies software abnormal control events that meet specific target feature control modes, generates mode control event sequences accordingly, and labels these software abnormal control events with accurate fault tags. The application further screens events highly related to target faults, constructs candidate abnormal control event sequences, integrates rich fault information and knowledge points, and visually displays the internal relationship and mutual influence between abnormal events in the form of knowledge entities and knowledge links. The application provides a powerful knowledge support platform for fault diagnosis experts.

[0003] Although in the field of computer technology, the application generates a highly focused and refined target abnormal knowledge network by fine feature screening and elimination of the basic knowledge network, directly pointing to the core abnormal control events, greatly simplifying the fault troubleshooting process and shortening the fault recovery time, the application only analyzes the hardware from a static perspective, lacks dynamic real-time monitoring and prediction of the hardware, which may lead to failure to timely discover and respond to abnormalities in the production process, thereby increasing the rejection rate and rework. In addition, the lack of a dynamic real-time early warning mechanism for real-time monitoring of the hardware may lead to failure to timely discover and respond to abnormalities, thereby expanding the problem, increasing downtime and rework costs, and prolonging the fault recovery time. Furthermore, the lack of a self-adaptive adjustment mechanism to adjust and improve the prediction model may lead to degradation of the prediction model with changes in the environment and equipment state, thereby reducing the prediction accuracy and increasing false positives and false negatives, and thus weakening the actual effectiveness of early warning and control. Therefore, there is an urgent need for a data mining method applied to a hardware control system to solve the above problems. SUMMARY

[0004] The application provides a data mining method applied to a hardware product control system, and aims to solve the following problems: only analyzing hardware from a static perspective, lacking dynamic real-time monitoring and prediction of hardware, which can cause failure to discover and respond to abnormalities in the production process in time, thereby increasing the unqualified rate and rework, in addition, lacking a dynamic real-time early warning mechanism for real-time monitoring of hardware, which can cause abnormalities to be unable to be discovered and responded to in time, thereby causing problems to be enlarged, downtime and rework costs to be increased, and fault recovery time to be prolonged, and lacking a self-adaptive adjustment mechanism for adjusting and improving the prediction model, which can cause the prediction model to be degraded with changes in the environment and the equipment state, thereby reducing the prediction accuracy and increasing false positives and false negatives, and thereby weakening the actual effectiveness of early warning and control.

[0005] The application adopts the technical scheme that a data mining method applied to a hardware product control system comprises the following steps: installing multiple types of sensors on the side surface of the main shaft, the tool handle, the feeding mechanism and the machining cavity of each production line, and determining the installation precision and sampling parameters of the sensors; processing the collected data; running an online concept drift detector on the real-time collected data; calculating the correlation degree between each feature in the sliding window and the target quality index according to the Pearson correlation coefficient; and judging the absolute value of the Pearson correlation coefficient between the features and the target quality index according to a set threshold F; performing fast Bayesian optimization using a small amount of incremental samples after drift triggering; outputting the physical benchmark calculated based on the process equation to the residual learning model, running the deviation calculated by the residual model in parallel, and jointly outputting the prediction value with the physical model; adopting exponential weighted moving average on the jointly output prediction value, and estimating the mean and standard deviation of the jointly output prediction value; constructing an optimization feedback mechanism to optimize the prediction result.

[0006] As a preferred embodiment, the specific steps of installing multiple types of sensors on the side surface of the main shaft, the tool handle, the feeding mechanism and the machining cavity of each production line, and determining the installation precision and sampling parameters of the sensors are as follows: By installing three-axis accelerometers, spindle torque sensors, temperature and humidity sensors, and acoustic sensors on the side of the spindle, the tool holder, the feeding mechanism, and the machining cavity of each production line; by installing temperature and humidity sensors, three-axis accelerometers, and spindle torque sensors on the side of the spindle, the sensors are 20mm±5mm away from the tool holder and the installation angle tolerance is ±3°, among which the temperature and humidity sensors are used for environmental monitoring, the acoustic sensors are placed in the cavity of the machine tool, and the temperature sensors measuring the temperature of the workpiece are installed near the end of the workpiece; by setting the sampling rate of the three-axis accelerometer to 2kHz-9kHz, the sampling rate of the acoustic sensor to not less than 8kHz, and the sampling rate of the temperature sensor to 1Hz, and synchronizing all sampling time stamps using a precise clock synchronization protocol, the time series meets the input data requirements for time alignment in frequency domain analysis and multi-modal fusion.

[0007] As a preferred embodiment, the specific steps of processing the collected data are: Each sensor performs timestamp unification, sampling rate resampling, and adaptive Kalman filter denoising on the edge preprocessing unit when uploading data, and sets the Kalman filter time constant to a range of 0.01s to 0.8s, obtaining intermediate data streams with a time delay of not more than 200ms and a signal-to-noise ratio improved to 6dB; by implementing lightweight feature calculation on high-frequency collected signals in the edge unit, a feature vector is obtained which can be used as input for residual learning model and threshold determination.

[0008] As a preferred embodiment, the specific steps of running an online concept drift detector on the real-time collected data are: By running an online concept drift detector on the real-time collected data and statistically analyzing the feature distribution changes of the real-time collected data in a sliding window W, and comparing the statistical quantity with a preset threshold E, when the statistical quantity exceeds the threshold, a distribution mutation trigger signal is output, obtaining an immediate detection signal for data distribution mutation.

[0009] As a preferred embodiment, the specific steps of judging the absolute value of the Pearson correlation coefficient between the features and the target quality indicator according to the set threshold F are: By calculating the Pearson correlation coefficient between each feature in the sliding window and the target quality indicator, and judging whether the absolute value G of the Pearson correlation coefficient between the feature and the target quality indicator reaches a significant level according to the set threshold F, wherein the set threshold F ranges from 0.2 to 0.5, If the absolute value of the Pearson correlation coefficient between the feature and the target quality indicator reaches a significant level, the feature is retained, otherwise the feature is discarded, and model pruning is performed on the edge-embedded model to reduce the number of neural network parameters to not more than 48% of the original model, obtaining a feature subset that can only be retained at different production stages when the prediction contribution reaches a preset significant level.

[0010] As a preferred implementation, the specific step of performing fast Bayesian optimization using a small number of incremental samples after drift triggering is: By setting the optimization iteration limit to be less than or equal to 18 times, adjusting the learning rate, regularization coefficient and tree depth and other hyperparameters, and recording the initial parameters, converged parameters and validation set performance changes of each optimization, automatic recovery during slow device state changes can be achieved.

[0011] As a preferred implementation, the specific step of outputting the physical benchmark calculated based on the process equation to the residual learning model, running the deviation calculated by the residual model in parallel, and jointly outputting the predicted value with the physical model is: By outputting the physical benchmark calculated based on the process equation to the residual learning model, running the deviation calculated by the residual model in parallel, and jointly outputting the predicted value with the physical model, a quality prediction result with both physical interpretability and data-driven accuracy is obtained, wherein the algorithm formula of residual learning and online updating is: ; represents the new parameter vector obtained after applying the update rule once at time q+1, UP represents the trainable parameter vector of the model, q represents the time index of the discrete sampling time, ST represents a positive factor that controls the amplitude of each gradient update, represents the gradient vector operation with respect to the parameter UP, represents the gradient operator, is a scalar function that measures the prediction error, represents the observed value, represents the benchmark quality estimate calculated by the physical model, represents the output of the residual model.

[0012] As a preferred implementation, the specific step of estimating the mean and standard deviation of the jointly output prediction value is: By constructing the formula for estimating the mean of the jointly output prediction value, a baseline that can smooth high-frequency noise and gradually adjust up or down when there is a persistent offset is obtained, wherein the formula for estimating the mean of the jointly output prediction value is: ; represents the smoothed mean at time i, H represents the smoothed mean, and i represents the time index, represents the weight of the prediction value, represents the observed value at time i, and J is the observed value.

[0013] As a preferred implementation, the specific step of estimating the mean and standard deviation of the jointly output prediction value further includes: By constructing a formula for estimating the standard deviation of the joint output prediction value, a fluctuation dynamic scaling alarm threshold is obtained, wherein the formula for estimating the standard deviation of the joint output prediction value is: M represents a threshold for triggering an alarm, i represents a time index, represents a smoothed mean value at time i, H represents a smoothed mean value, represents a standard normal quantile corresponding to a confidence level K is a standard normal distribution point, is a weight coefficient of a significance level, represents an estimation of local fluctuation of a sequence at time i, and N represents an estimation of local fluctuation of a sequence.

[0014] As a preferred embodiment, the specific steps of constructing an optimized feedback mechanism to optimize the prediction result are: By forming a ternary record of each prediction result, implemented intervention action and subsequent actual quality inspection result and storing it in a versioned ontology knowledge base, a high-quality labeled data set for supervised training of a residual learning model and weight update of a causal graph is obtained; by designing an automatic control issuing module with a safety rollback strategy, the system monitors key safety constraints after automatically issuing parameter adjustments, thereby obtaining a safeguarding capability that takes into account production safety and quality control in automatic response; after online intervention verification, the labeled samples are used for periodic batch retraining and parameter fusion combined with an online fine-tuning strategy, which can retain expert knowledge while enabling the residual learning model to quickly adapt to new faults; by establishing an audit link between the knowledge base and the residual learning model change and recording the trigger conditions, input samples and effect evaluation indicators of each model parameter update, transparent update records for engineers to review and compliance check are obtained.

[0015] The beneficial effects of the present application are: 1. By connecting the physical prior equation and the residual learning model in parallel and providing the physical equation and parameter calibration method in the specification, a joint prediction result with both engineering interpretability and significantly reduced prediction error is obtained, which can further concretize abstract algorithms, enhance implementability and make it easy for quality engineers to verify technical effects.

[0016] 2. By implementing concept drift detection based on a sliding window at the edge and performing online fine-tuning for a limited number of times according to the triggering rules and ensuring that the single inference delay of the residual model is less than or equal to 95ms, the model can adapt to equipment wear and material batch changes in real time, thereby reducing the degradation of the learning residual model and reducing false positives.

[0017] ​3. By using exponential weighted moving average to estimate the dynamic mean and local fluctuation level of the joint output prediction value, and taking the mean and uncertainty quantity amplified by a predetermined confidence level together to form a dynamic alarm threshold, and in the specification, the smoothing coefficient, confidence factor and long-term false alarm rate upper limit are specifically limited, thereby obtaining a statistically significant and quantifiable alarm triggering rule. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is a flow chart of a data mining method applied to a hardware product control system.

[0019] Figure 2 It is a technical effect diagram of a data mining method applied to a hardware product control system. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the following specific embodiments are further described, but the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application. In the following examples, the experimental methods are conventional methods, and the materials and reagents used in the following examples are commercially available unless otherwise specified.

[0021] Example 1 combines Figure 1 As shown in the figure, a flow chart of a data mining method applied to a hardware product control system, including: deploying sensors to collect data in real time and processing the collected data; adaptive sampling and constructing a feature screening unit; quality prediction of products and deployment of early warning system to monitor the quality of products; constructing an optimization feedback mechanism to optimize the prediction results; the following are specific implementation steps: Step one, deploy sensors to collect data in real time and process the collected data; Specifically, the steps for deploying sensors to collect data in real time are as follows: Multiple types of sensors are installed on the side of the spindle, tool holder, feeding mechanism, and machining cavity of each production line, and the installation accuracy and sampling parameters of the sensors are clearly defined. This yields time-series multimodal raw data covering vibration, torque, temperature, acoustics, and workpiece electrical parameters, providing a complete and verifiable data foundation for subsequent precise feature extraction and time-series consistency analysis. The various types of sensors include: triaxial accelerometers, spindle torque sensors, temperature and humidity sensors, and acoustic sensors. The installation accuracy is achieved by installing temperature and humidity sensors, triaxial accelerometers, and spindle torque sensors on the side of the spindle, with a sensor distance of 20mm ± 5mm from the tool holder and an installation angle tolerance of ± 3°. The temperature and humidity sensors are used for environmental monitoring, the acoustic sensors are placed inside the machine tool cavity, and the temperature sensor, which measures the workpiece temperature, is installed near the workpiece. The sampling parameters are determined by setting the channel sampling rate range, such as a sampling rate of 2kHz–9kHz for the triaxial accelerometer, a sampling rate of not less than 8kHz for the acoustic sensor, and a sampling rate of 1Hz for the temperature sensor, using a precision time synchronization protocol. The Protocol (PTP) synchronizes all sampling timestamps, ensuring that the time alignment error between any two channels does not exceed 1ms, thereby obtaining input data that meets the time alignment requirements for time series analysis and multimodal fusion. The specific steps of processing the collected data are: each sensor performs time stamping, sampling rate resampling and adaptive Kalman filter denoising on the edge preprocessing unit when uploading data, and the Kalman filter time constant is controlled in the range of 0.01s to 0.8s, so that the intermediate data stream with a time delay of not more than 200ms and a signal-to-noise ratio improved to 6dB is obtained, thereby ensuring that the data input to the cloud has real-time and reliability; by implementing lightweight feature calculation on the high-frequency acquisition signal in the edge unit, for example, performing discrete Fourier transform on the acceleration signal with a SIGNAL=1024 point window, and calculating the main frequency energy ratio, energy in the predetermined band-pass frequency range and kurtosis, etc. Frequency domain features are obtained which can be used as feature vectors for residual learning model and threshold judgment input, thereby ensuring that the end-to-end feature transmission delay is less than or equal to 200ms and improving the abnormal detection and early warning performance; the residual learning model trains a lightweight residual learning model with the multi-modal feature vector extracted by the edge as input, by fitting the residual between the physical benchmark and the observed value, and adjusting the parameters in the periodic online updating strategy, thereby obtaining a joint prediction that can compensate for the unexplained bias of the physical model and output in real time; the physical model is based on process and mechanical principles and measurable process parameters, such as cutting force, cutting speed, cutting section, material hardness, etc. as input, through analytical equation calculation and output of the verifiable engineering model of the benchmark quality index; by uploading the sensor data to the central database and establishing an index in the manufacturing execution system (Manufacturing Execution System, MES), the centralized storage and retrieval capability is obtained, which supports the subsequent data management.

[0022] Step two, adaptive sampling and feature screening unit construction; Specifically, the specific steps of adaptive sampling and feature screening unit construction are: running an online concept drift detector on the real-time collected data and statistically analyzing the feature distribution change of the real-time collected data in a sliding window W, and comparing the statistical quantity with a preset threshold E, when the statistical quantity exceeds the threshold, the distribution mutation trigger signal is output, thereby obtaining an immediate detection signal of the data distribution mutation, which can prompt the residual learning model to trigger adaptive update; the range of the sliding window is 1 hour to 8 hours; after triggering the distribution mutation event, the sampling rate of the non-target channel is dynamically reduced to according to the feature importance weight A , where B is the channel sampling rate function, c represents the channel number, max is the peak channel sampling rate, min is the channel trough sampling rate, and the key frequency band sampling rate is not less than the specific frequency range defined by the physical engineering experience beforehand The sampling efficiency can retain the sampling strategy of the spectral components necessary for quality prediction even when the computing resources are limited, thereby maintaining the integrity of the prediction input; the feature importance weight represents the relative contribution of a single input feature to the target quality indicator, and the weight is calculated by a weighted sum of the computable indicators of the Pearson correlation coefficient of the feature and the target, which is used to determine feature retention, sampling rate allocation, and priority sorting; the available computing cycles, memory, or network bandwidth of the edge device with limited computing resources in the current time window are insufficient to support all high-frequency sampling at the same time, so the system performs downsampling and feature screening according to the preset strategy to ensure the real-time performance of the key prediction function and safety constraints; the Pearson correlation coefficient between each feature in the sliding window and the target quality indicator is calculated, and whether the absolute value G of the Pearson correlation coefficient between the feature and the target quality indicator reaches a significant level is determined according to a set threshold F, where the threshold F ranges from 0.2 to 0.5; when the feature is retained, otherwise the feature is discarded, and model pruning is performed on the edge embedded model to reduce the number of neural network parameters to no more than 48% of the original model, thereby being able to retain only the feature subset that reaches the preset significance level in different production stages, thereby reducing the computational complexity and ensuring the real-time inference requirement; after the drift is triggered, a small amount of incremental samples are used to perform fast Bayesian optimization, where the number of optimization iterations is limited to less than or equal to 18 to adjust the learning rate, regularization coefficient, and tree depth, and the initial parameters, converged parameters, and validation set performance changes of each optimization are recorded, thereby being able to automatically recover during slow changes in device state, and thereby maintaining the stability of the model performance without human intervention; the model pruning of the edge embedded model gradually deletes insignificant decision nodes according to feature importance, weight amplitude, or parameter sensitivity, and after each pruning, the model performance is reviewed with the validation set to ensure that the number of parameters is less than or equal to 100k, the single inference delay is less than or equal to 100ms, and the prediction accuracy decreases by less than or equal to 1%, thereby obtaining a pruned lightweight embedded model that can guarantee real-time inference ability and maintain acceptable prediction performance under limited computing resources; the insignificant decision nodes in the model inference process cause a relative decrease in key performance indicators such as root mean square error and mean absolute error on the validation set that does not exceed the preset tolerance, and the corresponding feature contribution is lower than the preset threshold; the feature subset that reaches the preset significance level is calculated by using the Pearson correlation coefficient to calculate the contribution of each candidate feature to the target quality indicator, and comparing the contribution with the preset significance threshold, and retaining the features with a contribution greater than the threshold to form the feature subset that reaches the preset significance level.

[0023] Step three, quality prediction of the product and deployment of the early warning system to monitor the quality of the product; Specifically, the steps for product quality prediction are as follows: A physical baseline calculated based on the process equations is output to a residual learning model. Deviations calculated by the residual model are processed in parallel and jointly output with the physical model to produce predicted values. This yields a quality prediction result that combines physical interpretability with data-driven accuracy, making the prediction easy to verify in engineering and capable of error compensation. The algorithm formulas for residual learning and online updates are as follows: ; Let UP represent the new parameter vector obtained after applying the update rule once at time q+1, where UP represents the trainable parameter vector of the model, q represents the time index of the discrete sampling time, and ST represents the positive factor controlling the magnitude of each gradient update. This represents the gradient vector operation with respect to the parameter UP. Represents the gradient operator. A scalar function to measure prediction error. Represents the observed value. This represents the baseline mass estimate calculated from the physical model. This represents the output of the residual model; The specific steps for deploying an early warning system to monitor product quality are as follows: First, by applying an exponentially weighted moving average to the predicted values ​​of the joint output, the mean and standard deviation of the predicted values ​​are estimated, thereby obtaining a real-time statistic that is weighted against the latest samples and suppresses short-term noise. This provides a stable and reliable statistical basis for dynamic threshold calculation and anomaly detection. Second, by constructing a formula to estimate the mean of the predicted joint output, a method is obtained that can smooth high-frequency noise and gradually adjust the baseline upwards or downwards when there is a sustained shift. This baseline is then used to determine whether a systematic shift has occurred. The formula for estimating the mean of the predicted joint output is as follows: ; Let H represent the smoothed mean at time i, where H represents the smoothed mean and i represents the time index. Indicates the weight of the predicted value. Let J represent the observation value at time i, and J be the observation value. By constructing a formula to estimate the standard deviation of the joint output prediction value, a dynamically scaled alarm threshold based on volatility is obtained. This allows the threshold to be automatically widened during periods of high volatility to reduce false alarms, and tightened during periods of low volatility to improve sensitivity. The formula for estimating the standard deviation of the joint output prediction value is as follows: ; This represents the threshold calculated at time i to trigger the alarm, where M is the threshold for triggering the alarm, and i is the time index. denotes the smoothed mean at time i, H denotes the smoothed mean, denotes the standard normal quantile corresponding to the confidence level K is the standard normal quantile, is the weight coefficient of the significance level, denotes the estimate of the local volatility of the sequence at time i, N denotes the estimate of the local volatility of the sequence; After the formula for estimating the mean and standard deviation of the joint output prediction value is constructed, the causal candidate screening is locally performed at time M when the prediction value of the joint output is greater than the warning threshold, the pre-trained structured causal graph is used, the sampling approximation is calculated, the sampling number is set to 1000, the candidate factor list is output in descending order of the marginal contribution value, and the priority correction suggestion for each candidate factor is generated, so that the executable correction order and the factor contribution degree can be provided at the moment of alarm, thereby providing direct decision basis for the on-site operator or the automatic controller; after the candidate factors are output, small-scale online intervention is performed, for example, the cutting speed is reduced by 4%, the intervention group and the control group data are collected within a 28s observation window, the sample number of each group is greater than or equal to 10, the average treatment effect is calculated according to the observation results to update the edge weight of the causal graph, thereby obtaining the engineering path for converting the correlation judgment into controlled intervention verification, and the verification result is written back to the knowledge base, thereby the candidate factors can be included in the causal judgment after engineering verification and the confidence of root cause determination is improved.

[0024] Step four, constructing an optimized feedback mechanism to optimize the prediction result; Specifically, the steps for constructing an optimization feedback mechanism to optimize prediction results are as follows: First, by creating a ternary record of each prediction result, the implemented intervention action, and the subsequent actual quality inspection result, and storing it in a versioned ontology-based knowledge base, a high-quality labeled dataset is obtained that can be used for supervised training of residual learning models and causal graph weight updates. This provides a reliable basis for long-term improvement of supervised learning models and enhancement of root cause confidence. Second, by designing an automatic control distribution module with a safety backoff strategy, the system monitors key safety constraints after automatically distributing parameter adjustments. For example, if any constraint is triggered, it automatically restores to the most recent safety parameters and generates an abnormal work order, thus achieving a capability to balance production safety and quality control in automated response. This prevents equipment damage caused by malfunctions; key safety constraints include torque thresholds and workpiece temperature rise limits; after online intervention verification, labeled samples are used for periodic batch retraining and parameter fusion is performed in conjunction with online fine-tuning strategies, thereby enabling the residual learning model to quickly adapt to new faults while retaining expert knowledge, thus shortening the recovery time of residual learning model degradation and improving prediction accuracy; by establishing an audit link between the knowledge base and residual learning model changes and recording the triggering conditions, input samples, and effect evaluation indicators for each model parameter update, a transparent update record that can be reviewed by engineers and checked for compliance can be obtained, thereby improving system traceability and facilitating quality engineers to verify the improvement effect.

[0025] like Figure 2 A technical effect diagram of a data mining method applied to a hardware product control system is shown, where the black bars represent the technical effect of the present invention, and the gray bars represent the technical effect of the prior art. Figure 2 It can be seen that the technology of the present invention is superior to the existing technology.

[0026] Example 2: A data mining method for a hardware product control system, based on Example 1. The specific solution is as follows: Step 1: Data collection; Specifically, the data acquisition steps are as follows: By installing a triaxial accelerometer, a spindle torque sensor, a temperature sensor, and an acoustic sensor on the side of the spindle of each hardware production line at a distance of 20mm±5mm from the tool holder, on the surface of the tool holder, on the feeding end, and on the inner wall of the processing cavity, respectively, multi-channel raw time-series data covering vibration, torque, temperature, and acoustics are obtained; by performing timestamp unification, sampling rate reshaping, and bandpass filtering on each edge gateway, and by attaching a Precision Time Protocol (PTP) time stamp and SHA-256 digest to each record, intermediate data records with time alignment and integrity verification are obtained.

[0027] Step two, adaptive optimization of collected data; Specifically, the specific steps of adaptive optimization of collected data are: through the edge gateway, calculate the statistics of the arriving data stream in a one-hour sliding window and run the drift detector to obtain an update signal triggered when a statistical distribution mutation is detected; when the update signal is triggered, adjust the sampling rate by channel importance, filter features by correlation, and perform structural pruning on the edge model, so that the optimization of key frequency band information and real-time inference ability can still be ensured under the condition of limited bandwidth and computing power.

[0028] Step three, quality prediction and early warning of hardware products; Specifically, the specific steps of quality prediction and early warning of hardware products are: through the edge node, run the physical model in parallel to output the reference quality value and run the residual learning model to output the compensation amount, so as to obtain the joint final quality prediction value and provide an interpretable prediction result; after each prediction, update the sliding window statistics and calculate the threshold according to the confidence interval rule, and when the prediction value exceeds the threshold, trigger the local causal candidate screening and output the correction suggestions sorted by contribution, so as to obtain timely alarm and executable fault disposal sequence.

[0029] Step four, optimization feedback and closed-loop iteration; Specifically, the specific steps of optimization feedback and closed-loop iteration are: through the edge node, run the physical model in parallel to output the reference quality value and run the residual learning model to output the compensation amount, so as to obtain the joint final quality prediction value and provide an interpretable prediction result; after each prediction, update the sliding window statistics and calculate the threshold according to the confidence interval rule, and when the prediction value exceeds the threshold, trigger the local causal candidate screening and output the correction suggestions sorted by contribution, so as to obtain timely alarm and executable fault disposal sequence.

[0030] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A data mining method applied to a hardware product control system, characterized in that: Multiple types of sensors are installed on the side of the spindle, tool holder, feeding mechanism, and machining cavity of each production line, and the installation accuracy and sampling parameters of the sensors are determined; the collected data is then processed. An online concept drift detector is run on the real-time acquired data; the Pearson correlation coefficient and the correlation between each feature and the target quality index are calculated within the sliding window; the absolute value of the Pearson correlation coefficient between the feature and the target quality index is judged according to the set threshold F; and fast Bayesian optimization is performed using a small number of incremental samples after drift is triggered. The physical baseline calculated based on the process equation is output to the residual learning model. The deviation calculated by the residual model is run in parallel and jointly output with the physical model to produce the predicted value. An exponentially weighted moving average is used to estimate the mean and standard deviation of the joint output forecasts. An optimization feedback mechanism is constructed to optimize the prediction results.

2. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for installing various types of sensors on the side of the spindle, tool holder, feeding mechanism, and machining cavity of each production line, and for specifying the installation accuracy and sampling parameters of the sensors, are as follows: By installing triaxial accelerometers, spindle torque sensors, temperature and humidity sensors, and acoustic sensors on the side of the spindle, tool holder, feeding mechanism, and machining cavity of each production line; by installing the temperature and humidity sensors, triaxial accelerometers, and spindle torque sensors on the side of the spindle, with the sensors 20mm±5mm from the tool holder and an installation angle tolerance of ±3°, the temperature and humidity sensors are used for environmental monitoring, the acoustic sensors are placed in the machine tool cavity, and the temperature sensor for measuring the workpiece temperature is installed near the workpiece; by setting the sampling rate of the triaxial accelerometer to 2kHz–9kHz, the sampling rate of the acoustic sensor to no less than 8kHz, and the sampling rate of the temperature sensor to 1Hz, and by using a precision clock synchronization protocol to synchronize all sampling timestamps, input data that meets the time-series alignment requirements for frequency domain analysis and multimodal fusion is obtained.

3. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for processing the collected data are as follows: When each sensor uploads data, it performs timestamp unification, sampling rate resampling, and adaptive Kalman filtering noise reduction on the edge preprocessing unit, and controls the Kalman filter time constant within the range of 0.01s to 0.8s to obtain an intermediate data stream with a latency of no more than 200ms and a signal-to-noise ratio improved to 6dB. By performing lightweight feature calculations on the high-frequency acquired signals in the edge unit, feature vectors that can be used as inputs for residual learning models and threshold determination are obtained.

4. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for running an online concept drift detector on real-time acquired data are as follows: By running an online concept drift detector on the real-time acquired data and statistically analyzing the characteristic distribution changes of the real-time acquired data in units of a sliding window W, and comparing the statistical quantity with a preset threshold E, a distribution change trigger signal is output when the statistical quantity exceeds the threshold, thus obtaining an instantaneous detection signal for data distribution changes.

5. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for judging the absolute value of the Pearson correlation coefficient between the feature and the target quality index based on the set threshold F are as follows: The correlation between Pearson correlation coefficients and target quality indicators is calculated for each feature within a sliding window. A threshold F is then used to determine whether the absolute value G of the Pearson correlation coefficient between the feature and the target quality indicator reaches a significance level. The threshold F ranges from 0.2 to 0.

5. If the condition is met, the feature is retained; otherwise, it is removed. At the same time, model pruning is performed on the edge-embedded model to reduce the number of neural network parameters to no more than 48% of the original model, resulting in a subset of features that can retain only those that contribute to the prediction at a preset significance level in different production stages.

6. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for performing fast Bayesian optimization using a small number of incremental samples after drift triggering are as follows: By setting the number of optimization iterations to less than or equal to 18, hyperparameters such as learning rate, regularization coefficient, and tree depth can be adjusted, and the initial parameters, convergence parameters, and validation set performance changes of each optimization are recorded, enabling automatic recovery during periods of slow device state changes.

7. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for outputting the physical baseline calculated based on the process equation to the residual learning model, having the deviation calculated by the residual model run in parallel, and jointly outputting the predicted value with the physical model are as follows: By outputting the physical baseline calculated based on the process equations to the residual learning model, the deviations calculated by the residual model are run in parallel and jointly output with the physical model to produce predicted values, resulting in quality predictions that combine physical interpretability with data-driven accuracy. The algorithm formulas for residual learning and online updates are as follows: ， Let UP represent the new parameter vector obtained after applying the update rule once at time q+1, where UP represents the trainable parameter vector of the model, q represents the time index of the discrete sampling time, and ST represents the positive factor controlling the magnitude of each gradient update. This represents the gradient vector operation with respect to the parameter UP. Represents the gradient operator. A scalar function to measure prediction error. Represents the observed value. This represents the baseline mass estimate calculated from the physical model. This represents the output of the residual model.

8. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for estimating the mean and standard deviation of the joint output predicted values ​​are as follows: By constructing a formula to estimate the mean of the joint output predictions, a method is obtained that can smooth high-frequency noise and gradually adjust the baseline upwards or downwards when there is a persistent offset. The formula for estimating the mean of the joint output predictions is as follows: ， Let H represent the smoothed mean at time i, where H represents the smoothed mean and i represents the time index. Indicates the weight of the predicted value. Let i represent the observed value at time i, and J represent the observed value.

9. The data mining method for a hardware product control system according to claim 8, characterized in that: The specific steps for estimating the mean and standard deviation of the joint output predictions also include: By constructing a formula to estimate the standard deviation of the joint output prediction value, the dynamic scaling alarm threshold based on volatility is obtained. The formula for estimating the standard deviation of the joint output prediction value is as follows: ， This represents the threshold calculated at time i to trigger the alarm, where M is the threshold for triggering the alarm, and i is the time index. Let represent the smoothed mean at time i, and H represent the smoothed mean. Indicates the corresponding confidence level The standard normal quantiles, where K is the standard normal distribution point. The weighting coefficients represent the significance level. Let represent the estimate of the local volatility of the sequence at time i, and N represent the estimate of the local volatility of the sequence.

10. The data mining method applied to a hardware product control system according to claim 1, characterized in that: The specific steps for constructing an optimization feedback mechanism to optimize the prediction results are as follows: By creating a ternary record of each prediction result, the implemented intervention action, and the subsequent actual quality inspection results and storing it in a versioned ontological knowledge base, a high-quality labeled dataset is obtained that can be used for supervised training of residual learning models and causal graph weight updates. By designing an automatic control distribution module with a safety backoff strategy, the system monitors key safety constraints after automatically distributing parameter adjustments, thus achieving the ability to ensure both production safety and quality control in automated responses. After online intervention verification, by using labeled samples for periodic batch retraining and combining them with online fine-tuning strategies for parameter fusion, the residual learning model can quickly adapt to new faults while retaining expert knowledge. By establishing an audit link between the knowledge base and the residual learning model changes, and recording the triggering conditions, input samples, and performance evaluation indicators for each model parameter update, a transparent update record is obtained that can be reviewed by engineers and used for compliance checks.

Images