A smart control platform and method for smart manufacturing equipment based on the Internet of Things

By deploying multiple types of sensor arrays in intelligent manufacturing equipment and performing interference identification, weighted fusion, and feature mapping, the problems of sensor signal contamination and lack of process factor mapping are solved, achieving data-driven accuracy of equipment control and consistency of product quality.

CN123113620APending Publication Date: 2026-11-17ZHEJIANG JIULING TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611502279.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent manufacturing equipment control methods lack a complete data-driven closed-loop link. Narrowband interference mixed in the sensor signals is not effectively identified and filtered out. There is a lack of reasonable weighted fusion between multi-source sensors. The quantitative mapping relationship between process factors and quality indicators is missing, making it difficult to guarantee the accuracy of equipment control and the consistency of product quality.

Method used

Multiple types of sensor arrays are deployed at various working nodes of intelligent manufacturing equipment. Interference frequency bands are identified and notch filters are removed through edge gateways. Multi-channel weighted fusion and outlier identification are performed. Multi-dimensional features are extracted and paired with quality indicators to construct a quality impact mapping expression. The allowable range of process parameters is solved in reverse and equipment control commands are generated.

Benefits of technology

It achieves quantitative reverse derivation from quality objectives to control parameters, ensuring the consistency of equipment control accuracy and product quality. Through interference identification and data fusion, it suppresses signal pollution, reduces the input dimension of the mapping expression, and improves the robustness of control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113620A_ABST
    Figure CN123113620A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent control platform and method for smart manufacturing equipment based on the Internet of Things (IoT). The method performs interference frequency band identification and notch filtering on the original signal at the edge gateway side to ensure signal reliability. It suppresses single-channel drift contamination through variance inverse weighted fusion and outlier identification. Redundancy is eliminated through process factor decoupling, and highly correlated feature subsets are screened using conditional variance decomposition. A quality influence mapping expression is constructed and solidified into a mapping relationship, based on which the allowable range of quality indicators is deduced into the allowable range of process parameters, generating control commands for the execution layer. This invention achieves quantitative reverse derivation from quality objectives to control parameters, transforming equipment control from experience-based adjustments to data-driven parameter constraints, effectively ensuring equipment control accuracy and product quality consistency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically to an intelligent control platform and method for intelligent manufacturing equipment based on the Internet of Things. Background Technology

[0002] With the rapid development of Industrial Internet of Things (IIoT) technology, smart manufacturing equipment is now widely equipped with various types of sensors to collect operating status signals such as vibration, temperature, pressure, and current. These data are then uploaded and processed by upper-level computing devices via edge gateways. Existing smart manufacturing equipment control methods typically compare the operating data collected by sensors with preset fixed thresholds or empirical parameters. When a parameter exceeds the threshold, an alarm is triggered or simple open-loop adjustments are performed to maintain the equipment's basic operating status.

[0003] However, the above control method has the following technical problems: There is a lack of a complete data-driven closed-loop link between sensor signal acquisition and control command generation. Specifically, multi-source electromagnetic interference exists at the equipment operating site, and narrowband interference mixed into the sensor signals is not effectively identified and filtered out, leading to unreliable data processing foundations. There is a lack of reasonable weighted fusion and anomaly identification methods among multiple sensors deployed at the same operating node; single-channel drift or transient anomalies directly contaminate the control data. More importantly, the quantitative mapping relationship between process factors and quality indicators is missing. Adjustments to control parameters rely on operator experience rather than data-driven reverse derivation, making it difficult to proactively constrain process parameters before quality indicators deviate, thus making it difficult to simultaneously guarantee equipment control accuracy and product quality consistency. Summary of the Invention

[0004] This invention aims to provide an intelligent control platform and method for smart manufacturing equipment based on the Internet of Things, which realizes the quantitative reverse derivation from quality objectives to control parameters, transforming equipment control from experience-based adjustment to data-driven parameter constraints, effectively ensuring the consistency of equipment control accuracy and product quality.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent control method for intelligent manufacturing equipment based on the Internet of Things, comprising: Multiple types of sensor arrays are deployed at various working nodes of intelligent manufacturing equipment to identify interference frequency bands and filter out notch filters on the collected raw signals, thereby obtaining an anti-interference raw dataset. The original anti-interference dataset is subjected to multi-channel weighted fusion and outlier identification to obtain a reliable dataset; The reliable dataset is windowed according to the processing cycle and multidimensional features are extracted. After decoupling from process factors, these features are paired with quality index values ​​to form a feature-quality paired sample set. The feature-quality paired sample set is subjected to correlation quantification analysis to screen the feature subset that is highly correlated with the quality index, forming a correlated feature set; Based on the associated feature set, a quality influence mapping expression is constructed, and the allowable tolerance zone of the quality index is set according to the quality requirements. After inspection and correction, it is solidified into a quality influence mapping relationship. Based on the quality influence mapping relationship, the allowable range of each process parameter is deduced, equipment control instructions are generated and sent to the execution layer.

[0006] Preferably, the step of identifying interference frequency bands and performing notch filtering on the acquired raw signal to obtain an anti-interference raw dataset includes: The original signal sequence after cataloging the edge gateway is subjected to a fast Fourier transform channel by channel to convert the time-domain signal into a frequency-domain representation, thereby obtaining the amplitude spectrum of each sensor channel. The transform window length is taken as the sampling frequency divided by 1.25 and then rounded up to the nearest power of 2 to ensure that the frequency resolution does not exceed 1.25 Hz. The amplitude spectrum is compared band by band with the pre-calibrated baseline spectrum. If the amplitude of a certain frequency band exceeds three times the amplitude of the corresponding frequency band of the baseline spectrum, it is marked as an interference frequency band and its lower and upper frequency limits are recorded. For each marked interference frequency band, a second-order notch filter is constructed with its center frequency as the notch frequency and the frequency band width as the notch bandwidth. When there are multiple interference frequency bands in the same channel, the corresponding number of second-order notch filters are cascaded. The number of cascaded filters in a single channel shall not exceed 5. The filtered frequency domain signal is restored to the time domain signal by performing an inverse fast Fourier transform, and together with the sensor type, sensor number, installation location and acquisition timestamp, it forms the original anti-interference dataset.

[0007] Preferably, the original anti-interference dataset is subjected to multi-channel weighted fusion and outlier identification to obtain a reliable dataset, including: The original dataset against interference is grouped and indexed according to two dimensions: sensor type and operating node. Multiple sensor channels of the same type on the same operating node are grouped together and sorted by acquisition timestamp. Time alignment and spatial registration are performed on each channel within each group; spatial attenuation correction is introduced for the vibration sensor channel based on the installation distance of each sensor from the center of the working condition, while no spatial attenuation correction is performed for the temperature sensor, pressure sensor and current sensor. For each set of multi-channel data after registration, the reciprocal of the measurement variance of each channel within a sampling window is used as the fusion weight. The measurement values ​​of each channel are weighted and summed according to the weights, and then divided by the sum of the weights to obtain the fusion value at each time step. Calculate the mean and standard deviation of the fused value at each working node over a complete processing cycle; If the fusion value at any given time deviates from the mean by more than 3 times the standard deviation, it is marked as an outlier and removed. If the number of consecutive outliers marked as outliers does not exceed 5, the values ​​of the fused values ​​of adjacent normal times are used to make linear interpolation to fill the gap; if the number of consecutive outliers exceeds 5, the outliers will be marked as invalid and the entire processing cycle in which they occur will be excluded during window segmentation. The data remaining after outlier removal constitutes the reliable dataset.

[0008] Preferably, the step of performing time alignment and spatial registration for each channel within each group includes: Align the acquisition timestamps of each channel within the same group so that each channel has a corresponding measurement value at the same time. For vibration sensor channels, based on the installation distance of each sensor from the center of the working condition, a spatial attenuation correction coefficient is introduced according to the geometric diffusion law of body wave propagation to correct the measured values ​​of each channel to a unified reference position; if the installation distance is less than 0.01 meters, no spatial attenuation correction is performed. For temperature sensor channels, pressure sensor channels, and current sensor channels, since the physical quantities they monitor do not have propagation characteristics that attenuate with distance within the operating condition node range, no spatial attenuation correction is performed.

[0009] Preferably, the reliable dataset is windowed according to the processing cycle and multidimensional features are extracted. After decoupling from process factors, these features are paired with quality index values ​​to form a feature-quality paired sample set, including: Based on the processing cycle of the equipment, the reliable dataset is divided into several working condition windows, and each working condition window corresponds to a complete action cycle of a workpiece from clamping to unloading. For each operating condition window, statistical domain features, time domain features, and frequency domain features are extracted to form a multi-dimensional feature vector; Based on the installation location of each sensor and the physical quantity monitored, the components in the multidimensional feature vector are labeled with the corresponding process factor sources. The fusion values ​​of different installation locations under the same process factor are averaged and merged into a single representative value to achieve decoupling of process factors. The decoupled multidimensional feature vectors are paired with the offline quality inspection results of the corresponding processed parts, and all feature components are dimensionless to form a feature-quality pairing sample set.

[0010] Preferably, the feature-quality paired sample set is subjected to correlation quantification analysis to screen feature subsets that are highly correlated with quality indicators, forming a correlated feature set, including: A preliminary assessment is made of the linear correlation between each feature dimension and the quality index in the feature-quality paired sample set, and feature dimensions whose absolute value of linear correlation coefficient is less than the initial screening threshold are removed. For each feature dimension retained in the initial screening, the value range of the feature dimension is divided into several intervals. The variance of the quality index is calculated in each interval. Then, the variances of each interval are weighted and summed with the sample proportion as the weight to obtain the expected conditional variance. The ratio of the expected conditional variance to the overall variance of the quality index is subtracted from 1 to obtain the nonlinear correlation degree. Each feature dimension is sorted from largest to smallest according to the nonlinear correlation degree, and feature dimensions with a correlation degree greater than the correlation degree threshold are retained; if the number of retained exceeds 15, only the top 15 are retained to form a subset of highly correlated features. The dimensionless values ​​of each feature dimension in the highly correlated feature subset are re-integrated with the corresponding quality index values ​​to form the correlated feature set.

[0011] Preferably, constructing the quality influence mapping expression based on the associated feature set includes: The highly correlated feature dimensions retained in the aforementioned correlated feature set are used as input variables, and the quality index is used as the output variable. A quality influence mapping expression containing intercept, linear term, and pairwise interaction term is established using multiple regression. Each coefficient is solved by least squares, which is to sum the squares of the differences between the measured and predicted values ​​of all samples to obtain the residual sum of squares. The partial derivatives of the residual sum of squares with respect to each coefficient are calculated and set to zero to obtain the normal equation system, which is then solved. To ensure that the normal equation system has a unique stable solution, the total number of samples must be no less than three times the number of unknowns. When there is an approximate linear correlation between the input variables, a regularization term is added to the diagonal of the coefficient matrix of the normal equation system before solving.

[0012] Preferably, the quality influence mapping relationship, after verification and correction, includes: Substitute the input variables of each sample in the feature-quality paired sample set into the quality influence mapping expression, calculate the residual between the predicted value and the measured value, and calculate the mean and standard deviation of the residual; If the absolute value of the mean of the residuals is greater than 5% of the allowable tolerance zone of the quality index, then the intercept term is corrected once. If the standard deviation of the residual is greater than 30% of the allowable tolerance band of the quality index, then the feature dimensions with correlation between the initial screening rejection threshold and the retention threshold are added to the system, and the coefficients are recalculated using the expanded correlation feature set. The addition is performed only once. If the standard deviation of the residual still does not meet the requirements after the addition, the obtained quality influence mapping expression is retained as the currently available result. The quality influence mapping expression, after verification and correction, along with the values ​​of each coefficient, is solidified into the quality influence mapping relationship.

[0013] Preferably, the step of reversing the allowable range of each process parameter based on the quality influence mapping relationship, generating equipment control commands, and issuing them to the execution layer includes: Set the allowable range of target quality indicators based on the quality requirements of the current production orders; In the quality influence mapping expression, all input variables except for one input variable to be determined are fixed as the feature values ​​corresponding to the most recently completed processing cycle. The quality influence mapping expression is degenerated into a linear equation with respect to the input variable to be determined. The upper and lower limits of the allowable interval of the quality index are substituted into the equation to solve for the two boundary values ​​corresponding to the input variable to be determined. The above inverse solution operation is performed sequentially for each input variable. For the input variables with the highest correlation ranking, perform multivariate joint validation, setting them to the upper or lower bounds to form extreme combinations, and check whether the predicted values ​​of the quality indicators still fall within the allowable range; if they exceed the range, narrow the allowable range for the variables with lower correlation ranking. The dimensionless boundary values ​​obtained from the inverse kinematics are restored to characteristic values ​​with actual physical dimensions and transformed into control commands for each actuator. Before being issued, the sum of the set power of all actuators is checked for consistency. After the check passes, the control commands are issued to the device execution layer via industrial Ethernet.

[0014] On the other hand, this invention proposes an intelligent control platform for smart manufacturing equipment based on the Internet of Things, comprising: Multiple types of sensor arrays are deployed at various working nodes of intelligent manufacturing equipment to collect raw signals; The edge gateway receives the raw signals collected by the multi-type sensor arrays through the Internet of Things communication link, and performs cataloging, fast Fourier transform, interference frequency band identification, notch filtering and inverse fast Fourier transform on the raw signals, and outputs anti-interference raw dataset. The host computing device is connected to the edge gateway via an industrial Ethernet. It is used to perform multi-channel weighted fusion and outlier identification on the anti-interference original dataset to obtain a reliable dataset. The reliable dataset is windowed according to the processing cycle and multi-dimensional features are extracted. After decoupling by process factors, it is paired with quality index values ​​to form a feature-quality paired sample set. The feature-quality paired sample set is subjected to correlation quantification analysis to form a correlation feature set. Based on the correlation feature set, a quality influence mapping expression is constructed and solidified into a quality influence mapping relationship. Based on the quality influence mapping relationship, the allowable range of each process parameter is deduced and equipment control instructions are generated. The device execution layer receives device control commands from the host computer via industrial Ethernet or fieldbus, and adjusts the operating parameters according to the set values ​​in the commands.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention identifies interference frequency bands and cascades second-order notch filters on the original signal at the edge gateway, sufficiently attenuating narrowband interference and ensuring that subsequent data processing is based on a clean signal foundation. Through variance-weighted fusion and outlier identification, it suppresses the contamination of the fusion results by single-channel drift and transient anomalies, ensuring the stability of the reliable dataset. By labeling process factors and merging the average values ​​of multiple locations for the same factor, it eliminates redundant information and reduces the input dimension of the mapping expression. It calculates nonlinear correlation through conditional variance decomposition, selecting feature subsets highly correlated with quality indicators. Based on this, a quality influence mapping expression is constructed, verified, and corrected, then solidified into a quality influence mapping relationship. Finally, based on this mapping relationship, the allowable range of quality indicators is deduced into the allowable range of each process parameter, generating equipment control commands that are sent to the execution layer. This achieves quantitative reverse derivation from quality objectives to control parameters, transforming equipment control from experience-based adjustments to data-driven parameter constraints, effectively ensuring the consistency of equipment control accuracy and product quality. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent control method for smart manufacturing equipment based on the Internet of Things according to the present invention; Figure 2 This is a block diagram of the intelligent control platform for smart manufacturing equipment based on the Internet of Things (IoT) of this invention. Detailed Implementation

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

[0018] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.

[0019] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0020] Please see Figure 1 This invention provides an intelligent control method for smart manufacturing equipment based on the Internet of Things, comprising: Multiple types of sensor arrays are deployed at various operating nodes of intelligent manufacturing equipment to identify interference frequency bands and perform notch filtering on the collected raw signals, resulting in an anti-interference raw dataset; specifically including: The original signal sequence after cataloging the edge gateway is subjected to a fast Fourier transform channel by channel to convert the time-domain signal into a frequency-domain representation, thereby obtaining the amplitude spectrum of each sensor channel. The transform window length is taken as the sampling frequency divided by 1.25 and then rounded up to the nearest power of 2 to ensure that the frequency resolution does not exceed 1.25 Hz. The amplitude spectrum is compared with the pre-calibrated baseline spectrum band by band. If the amplitude of a certain band exceeds three times the amplitude of the corresponding band of the baseline spectrum, it is marked as an interference band and its lower and upper frequency limits are recorded. For each marked interference frequency band, a second-order notch filter is constructed with its center frequency as the notch frequency and the frequency band width as the notch bandwidth. When there are multiple interference frequency bands in the same channel, the corresponding number of second-order notch filters are cascaded. The number of cascaded filters in a single channel shall not exceed 5. The filtered frequency domain signal is restored to the time domain signal by performing an inverse fast Fourier transform, and together with the sensor type, sensor number, installation location and acquisition timestamp, it forms the anti-interference original dataset.

[0021] It helps to identify and suppress narrowband interference in sensor signals in a timely manner at the edge, making the data foundation on which subsequent fusion and feature extraction depend cleaner and more reliable. At the same time, through reasonable frequency resolution settings and cascade quantity limits, a good balance is achieved between filtering effect and signal fidelity, providing more robust input conditions for the entire control link.

[0022] To obtain a reliable dataset, multi-channel weighted fusion and outlier identification are performed on the original dataset to combat interference; specifically including: The original dataset against interference is grouped and indexed according to two dimensions: sensor type and operating node. Multiple sensor channels of the same type on the same operating node are grouped together and sorted by acquisition timestamp. Time alignment and spatial registration are performed on each channel within each group; spatial attenuation correction is introduced for the vibration sensor channel based on the installation distance of each sensor from the center of the working condition, while no spatial attenuation correction is performed for the temperature sensor, pressure sensor and current sensor. For each set of multi-channel data after registration, the reciprocal of the measurement variance of each channel within a sampling window is used as the fusion weight. The measurement values ​​of each channel are weighted and summed according to the weights, and then divided by the sum of the weights to obtain the fusion value at each time step. Calculate the mean and standard deviation of the fusion value at each working condition node within a complete processing cycle; if the fusion value at a certain moment deviates from the mean by more than 3 times the standard deviation, it is marked as an outlier and removed. If the number of consecutive outliers marked as outliers does not exceed 5, the data is filled by linear interpolation of the fused values ​​of adjacent normal times. If the number of consecutive outliers exceeds 5, the data will be marked as invalid and the entire processing cycle in which the outlier is located will be excluded during window segmentation. The data remaining after outlier removal constitutes a reliable dataset.

[0023] This includes performing time alignment and spatial registration for each channel within each group, including: The acquisition timestamps of each channel within the same group are aligned so that each channel has a corresponding measurement value at the same time. For vibration sensor channels, a spatial attenuation correction coefficient is introduced based on the installation distance of each sensor from the working condition center, according to the geometric diffusion law of body wave propagation, to correct the measurement values ​​of each channel to a unified reference position. If the installation distance is less than 0.01 meters, no spatial attenuation correction is performed. For temperature sensor channels, pressure sensor channels, and current sensor channels, since the physical quantities they monitor do not have the propagation characteristics of attenuation with distance within the working condition node range, no spatial attenuation correction is performed.

[0024] This helps to reasonably suppress the impact of single-channel drift and transient anomalies on the fusion results based on multi-channel redundant acquisition. At the same time, by distinguishing the physical propagation characteristics of different sensor types for differentiated registration, the fused values ​​can better reflect the true state of the working nodes, providing more stable data support for subsequent feature extraction and quality mapping.

[0025] The reliable dataset is windowed according to the processing cycle and multidimensional features are extracted. After decoupling from process factors, these features are paired with quality index values ​​to form a feature-quality paired sample set; specifically including: Based on the processing cycle of the equipment, the reliable dataset is divided into several working condition windows. Each working condition window corresponds to a complete action cycle of a workpiece from clamping to unloading. Statistical domain features, time domain features, and frequency domain features are extracted for each working condition window to form a multi-dimensional feature vector. Based on the installation location of each sensor and the physical quantity monitored, the components in the multidimensional feature vector are labeled with the corresponding process factor sources. The fusion values ​​of different installation locations under the same process factor are averaged and merged into a single representative value to achieve decoupling of process factors. The decoupled multidimensional feature vectors are paired with the offline quality inspection results of the corresponding processed parts, and all feature components are dimensionless to form a feature-quality pairing sample set.

[0026] This helps to accurately correlate sensor data with the complete processing cycle of a single workpiece, enabling the extracted multi-domain features to naturally match the quality inspection results in the time dimension. By labeling the sources of process factors and merging the mean values ​​of the same factor at multiple locations, it eliminates information duplication caused by redundant sensor deployment to a certain extent, reducing the input dimension of subsequent mapping modeling. Combined with dimensionless processing, it makes feature components of different dimensions comparable, providing a relatively standardized and well-aligned sample basis for correlation analysis and the construction of quality mapping expressions.

[0027] A correlation metric analysis is performed on the feature-quality paired sample set to screen the feature subsets that are highly correlated with the quality indicators, forming a correlated feature set; specifically including: A preliminary assessment is made of the linear correlation between each feature dimension and the quality index in the feature-quality paired sample set, and feature dimensions whose absolute value of linear correlation coefficient is less than the initial screening threshold are removed. For each feature dimension retained in the initial screening, the value range of the feature dimension is divided into several intervals. The variance of the quality index is calculated in each interval. Then, the variances of each interval are weighted and summed with the sample proportion as the weight to obtain the expected conditional variance. The ratio of the expected conditional variance to the overall variance of the quality index is subtracted from 1 to obtain the nonlinear correlation degree. Sort each feature dimension from largest to smallest according to its nonlinear correlation degree, and retain feature dimensions with a correlation degree greater than the correlation degree threshold; if the number of retained features exceeds 15, only the top 15 are retained to form a highly correlated feature subset; the dimensionless values ​​of each feature dimension in the highly correlated feature subset are reintegrated with the corresponding quality index values ​​to form a correlated feature set.

[0028] This helps to further capture the non-linear dependencies between features and quality indicators on the basis of linear initial screening, making the screening results less susceptible to the limitations of a single correlation metric. At the same time, by constraining the correlation degree ranking and the upper limit of the number, the input dimensions are appropriately simplified while retaining key information, which helps to reduce the complexity of subsequent mapping modeling and the risk of overfitting, and provides a more focused feature foundation that is closely related to the target indicator for the construction of the quality impact expression.

[0029] A quality impact mapping expression is constructed based on the associated feature set. Allowable tolerance zones for quality indicators are set according to quality requirements. After verification and correction, the expression is solidified into a quality impact mapping relationship. Specifically, this includes: The high-correlation feature dimensions retained in the correlation feature set are used as input variables, and the quality index is used as the output variable. A quality influence mapping expression containing intercept, linear term, and pairwise interaction term is established using multiple regression. Each coefficient is solved by least squares, which is to sum the squares of the differences between the measured and predicted values ​​of all samples to obtain the residual sum of squares. The partial derivatives of the residual sum of squares with respect to each coefficient are calculated and set to zero to obtain the normal equation system, which is then solved. To ensure that the normal equation system has a unique stable solution, the total number of samples must be no less than three times the number of unknowns. When there is an approximate linear correlation between the input variables, a regularization term is added to the diagonal of the coefficient matrix of the normal equation system before solving.

[0030] Substitute the input variables of each sample in the feature-quality paired sample set into the quality influence mapping expression, calculate the residual between the predicted value and the measured value, and calculate the mean and standard deviation of the residual; if the absolute value of the mean of the residual is greater than 5% of the allowable tolerance band of the quality index, then the intercept term is corrected once. If the standard deviation of the residual is greater than 30% of the allowable tolerance zone of the quality index, the feature dimensions with correlation between the initial screening rejection threshold and the retention threshold will be added to the system, and the coefficients will be recalculated using the expanded correlation feature set. The addition will be performed only once. If the standard deviation of the residual still does not meet the requirements after the addition, the obtained quality influence mapping expression will be retained as the currently available result. The quality influence mapping expression after verification and correction, together with the values ​​of each coefficient, will be solidified as the quality influence mapping relationship.

[0031] By introducing a mapping structure between linear terms and pairwise interaction terms, we can characterize the independent contribution and interactive influence of each associated feature on the quality index to a certain extent, making the quality prediction results more intuitive and interpretable. The introduction of sample size constraints and regularization terms helps to alleviate the coefficient instability problem caused by multicollinearity and improves the robustness of the solution process. The stratified test correction mechanism based on the residual mean and standard deviation enables targeted adjustments to the mapping expression before it is solidified, addressing systematic biases or insufficient fitting accuracy. At the same time, limiting the number of supplements avoids the risk of overfitting. The resulting mapping relationship structure is relatively simple and easy to call directly in online deployment at the edge and in subsequent quality trend analysis.

[0032] Based on the quality impact mapping relationship, the allowable range of each process parameter is determined, equipment control commands are generated, and sent to the execution layer; specifically including: Set the allowable range of target quality indicators based on the quality requirements of the current production orders; In the quality influence mapping expression, all input variables except for one of the input variables to be determined are fixed as the corresponding feature values ​​within the most recent completed processing cycle. The quality influence mapping expression is degenerated into a linear equation with respect to the input variables to be determined. The upper and lower limits of the allowable interval of the quality index are substituted into the equation to solve for the two boundary values ​​corresponding to the input variables to be determined. The above inverse solution operation is performed sequentially for each input variable. For the input variables with the highest correlation, perform multivariate joint validation, setting them to the upper or lower bounds to form extreme combinations, and check whether the predicted values ​​of the quality indicators still fall within the allowable range; if they do, narrow the allowable range for the variables with lower correlation. The dimensionless boundary values ​​obtained from the inverse kinematics are restored to characteristic values ​​with actual physical dimensions and transformed into control commands for each actuator. Before issuing the commands, the sum of the set power of all actuators is checked for consistency. After the check passes, the control commands are issued to the device execution layer via industrial Ethernet.

[0033] By leveraging the reversible derivation of mapping relationships, abstract quality objectives are transformed into concrete and executable boundary constraints for each process parameter, enabling quality control to extend from passive detection to proactive prevention. The joint verification of variable-wise inverse solution combined with extreme combinations of multiple variables helps to simplify the solution while taking into account the potential impact of parameter coupling effects on the quality boundary, reducing the overly optimistic risk that may be brought about by independent inverse solution of single variables. The dimensional restoration and power consistency verification before issuance provide a basic guarantee for the physical executability of the instructions and the safe operation of the equipment. This ensures that the generated control instructions not only meet the quality requirements of the current order, but also take into account the actual carrying capacity of the execution layer, making it easier to achieve timely and reliable closed-loop parameter adjustments within the production line cycle time.

[0034] On the other hand, this invention proposes an intelligent control platform for smart manufacturing equipment based on the Internet of Things, such as... Figure 2 As shown, it includes: Multiple types of sensor arrays are deployed at various working nodes of intelligent manufacturing equipment to collect raw signals; The edge gateway receives the raw signals collected by the multi-type sensor arrays through the Internet of Things communication link, and performs cataloging, fast Fourier transform, interference frequency band identification, notch filtering and inverse fast Fourier transform on the raw signals, and outputs anti-interference raw dataset. The host computing device is connected to the edge gateway via an industrial Ethernet. It is used to perform multi-channel weighted fusion and outlier identification on the anti-interference original dataset to obtain a reliable dataset. The reliable dataset is windowed according to the processing cycle and multi-dimensional features are extracted. After decoupling by process factors, it is paired with quality index values ​​to form a feature-quality paired sample set. The feature-quality paired sample set is subjected to correlation quantification analysis to form a correlation feature set. Based on the correlation feature set, a quality influence mapping expression is constructed and solidified into a quality influence mapping relationship. Based on the quality influence mapping relationship, the allowable range of each process parameter is deduced and equipment control instructions are generated. The device execution layer receives device control commands from the host computer via industrial Ethernet or fieldbus, and adjusts the operating parameters according to the set values ​​in the commands.

[0035] Furthermore, the IoT-based intelligent manufacturing equipment intelligent control platform in this embodiment can also implement other steps of the above-mentioned IoT-based intelligent manufacturing equipment intelligent control method, as follows: Step 1: Deploy multiple types of sensor arrays at various operational nodes of the intelligent manufacturing equipment, and perform interference frequency band identification and notch filtering on the collected raw signals; specifically including: Step 1.1: Arrange temperature sensors, vibration sensors, pressure sensors, and current sensors in the processing area, transmission area, temperature control area, and material conveying area of ​​the intelligent manufacturing equipment, respectively.

[0036] At each working condition node, no fewer than three sensors of the same type are arranged and installed in different directions along the equipment body to form a spatially redundant data acquisition pattern.

[0037] Each sensor, at a sampling frequency of no less than 1,000 times per second, converts the sensed analog electrical signals into digital signals through analog-to-digital conversion and transmits them to the edge gateway deployed near the device via the Internet of Things communication link.

[0038] The edge gateway performs preliminary cataloging of the received digital signals according to sensor type, sensor number, installation location number, and acquisition timestamp to form the raw signal sequence for subsequent processing.

[0039] Step 1.2: Perform a Fast Fourier Transform on each channel of the original signal sequence after cataloging the edge gateway to convert the time-domain signal into a frequency-domain representation and obtain the amplitude spectrum of each sensor channel.

[0040] During the transformation, the window length is calculated by dividing the sampling frequency by 1.25 and rounding up to the nearest power of 2. For example, at a sampling frequency of 1000 times per second, 1000 divided by 1.25 equals 800, which, rounded up to the nearest power of 2, is 1024, meaning the window length is 1024 sampling points, corresponding to a frequency resolution of approximately 0.98 Hz; at a sampling frequency of 5000 times per second, the window length is 4096 sampling points, corresponding to a frequency resolution of approximately 1.22 Hz; and at a sampling frequency of 12000 times per second, the window length is 16384 sampling points, corresponding to a frequency resolution of approximately 0.73 Hz.

[0041] The above method ensures that the frequency resolution does not exceed 1.25 Hz at any sampling frequency, allowing narrowband interference to be adequately resolved. The baseline spectrum mentioned here is pre-calibrated by on-site personnel during the equipment installation and commissioning phase, when the equipment is in normal no-load operation. The signal is continuously collected for no less than 30 no-load cycles and averaged. The calibration results are stored in the read-only memory area of ​​the edge gateway and are automatically read by the edge gateway each time the equipment is powered on.

[0042] The spectrum is divided into equal-width bands according to the frequency resolution described above, and the amplitude of each band is compared with the amplitude of the corresponding band in the baseline spectrum. If the amplitude of a certain band exceeds three times the amplitude of the corresponding band in the baseline spectrum, the band is marked as an interference band, and its lower and upper frequency limits are recorded. At the same time, the sensor channel number corresponding to the interference band is marked for use in step 1.3.

[0043] It should be noted that the above Fourier transform is based on the premise that the signal within the window is approximately stationary. If there are transient switching events such as tool entry or reversal during the machining process, which cause obvious non-stationary segments in the signal within the window, the window can be further subdivided into several short windows for spectral analysis before merging the results. This will not be elaborated here.

[0044] Step 1.3: Based on the interference frequency bands identified and recorded in Step 1.2, design a notch filter for each sensor channel in the edge gateway.

[0045] Specifically, for each marked interference frequency band, a second-order notch filter is constructed using its center frequency as the notch frequency and the width of that frequency band as the notch bandwidth. If multiple interference frequency bands exist in the same channel, the corresponding number of second-order notch filters are cascaded so that each interference frequency band is suppressed independently. The number of second-order notch filters cascaded in a single channel does not exceed 5; if more than 5 interference frequency bands are identified, they are sorted from largest to smallest amplitude, and the first 5 are filtered out, while the remaining interference frequency bands with smaller amplitudes are left unprocessed.

[0046] For the portion of the effective signal frequency band that does not contain interfering frequency bands, no filtering is applied to avoid unnecessary attenuation of the effective signal. After the above notch filtering, the signal components in the interfering frequency band are sufficiently attenuated, while the components in the effective signal frequency band are completely preserved.

[0047] Step 1.4: Perform an inverse fast Fourier transform on the filtered frequency domain signal to restore it to the time domain signal.

[0048] Since the second-order notch filter only produces amplitude attenuation in a very narrow frequency band near the notch frequency, it has almost no impact on the effective signal components far from the notch frequency. Therefore, no additional compensation is needed for the amplitude in the passband. For the small amplitude fluctuations that may occur at the edge of the notch frequency band, the attenuation is no more than 0.5 dB, which is within an acceptable range and no further correction is made.

[0049] The processed time-domain signal, along with its corresponding sensor type, sensor number, installation location, and acquisition timestamp, forms an anti-interference raw dataset. This dataset is temporarily stored by the edge gateway and transmitted to the host computing device via industrial Ethernet. Here, the edge gateway handles real-time operations such as signal acquisition, cataloging, frequency domain transformation, and notch filtering; the host computing device handles subsequent operations involving large-scale data processing and mapping relationship construction. Data exchange between the two is completed via industrial Ethernet.

[0050] Step Two: Perform multi-channel weighted fusion and outlier identification on the anti-interference original dataset output from Step One to obtain a reliable dataset, specifically including: Step 2.1: The host computing device receives the anti-interference raw dataset transmitted by the edge gateway in Step 1.

[0051] Since the cataloging in step one already includes four pieces of information: sensor type, sensor number, installation location, and data acquisition timestamp, the data is grouped and indexed here according to two dimensions: sensor type and operating node. Multiple sensor channels of the same type on the same operating node are grouped together. The number of channels in each group is consistent with the number of sensors of the same type deployed in step 1.1. The channels in each group are sorted by data acquisition timestamp to ensure that the data of each channel can correspond one-to-one in the time dimension during subsequent fusion processing.

[0052] Step 2.2: Perform time alignment and spatial registration on the multi-channel data of each group after sorting in the first step. Time alignment means compensating for the time offset caused by slight differences in the sampling start time of each channel within the same group, so that the data points of each channel at the same time correspond one-to-one; for cases where the offset is less than one sampling interval, linear interpolation is used to fill in the missing data points.

[0053] Spatial registration is performed only for vibration sensor channels. When vibration waves propagate in the device body as body waves, their amplitude attenuates approximately at 1 / r with the propagation distance. Therefore, for vibration sensors installed at the same operating point but with slightly different spatial positions, spatial attenuation correction is introduced based on the installation distance between each sensor and the operating point center.

[0054] The specific method is as follows: multiply the measured value of the j-th vibration sensor channel by the distance ratio, which is equal to the installation distance of the j-th vibration sensor from the center of the working condition. (Unit: meters) Divide by the installation distance from the nearest vibration sensor to the center of the working condition within that group. (Unit: m). If If all vibration sensors in the group are located near the center of the working condition, no spatial attenuation correction is made, and the original measurement values ​​of each channel are directly used for subsequent fusion.

[0055] The aforementioned distance ratio is the correction factor required to convert the measurement value of the distant vibration sensor to the equivalent level at the nearest point. The dimensions of the measurement value remain unchanged after correction. If the vibration sensor is installed on a rigid connector and the difference between the maximum and minimum installation distances does not exceed 10% of the minimum installation distance, then the aforementioned distance ratio is approximately 1 and no correction is required.

[0056] For temperature, pressure, and current sensors, since the temperature field is approximately uniform within the sensor spacing range, the pressure is approximately uniformly transmitted within a closed or semi-closed space, and the current is equal everywhere in a series circuit, none of these sensors experience spatial propagation attenuation. Therefore, no spatial attenuation correction is performed, and only time alignment is executed. After time alignment and spatial registration, the data from each channel within each group are comparable in both time and space.

[0057] Step 2.3: After completing time alignment and spatial registration, the multi-channel data of each group are fused and calculated using the inverse variance weighting method.

[0058] Specifically, for M sensor channels of the same type at a certain working condition node, the fused value at time t Calculate using the following formula: (1); In the formula, Let be the measurement value of the j-th sensor channel after registration in step 2.2 at time t. The dimension is the physical dimension of the measured quantity; for example, temperature is expressed in Kelvin (K). M is the total number of channels of the same type of sensor at this operating node, dimensionless. ; The fusion weight for the j-th channel is the reciprocal of the measurement variance of that channel within a sampling window of 10 seconds.

[0059] The measurement variance is calculated by dividing the sum of the squares of the differences between each sampled value within the window and the window mean by the number of sampling points minus 1. As can be seen from the above definition, the dimensions of measurement variance are... (Right now ), The dimensions are (Right now In formula (1), the molecule The dimensions are denominator The dimensions are After dividing the two The dimensions are K, which is consistent with the dimensions of the original measurement value.

[0060] The physical meaning of this fusion method is as follows: the smaller the measurement variance of a channel, the less interference it is affected by and the more stable its data is, thus giving it a larger weight in the fusion. Conversely, the weight of channels with large variances is reduced, thereby suppressing the adverse effects of interference on the fusion results. If a channel has a variance of 0 within a 10-second window (i.e., all sampled values ​​are exactly the same), then the weight of that channel is set to 10 times the sum of the weights of the other channels in the group, making it dominant in the fusion. At the same time, the operating status of that channel is recorded for subsequent troubleshooting.

[0061] Step 2.4: Combine the values ​​obtained in Step 2.3 for each working condition node and each time point. Outlier identification and removal are performed.

[0062] Specifically, the average value of the fusion value at each working condition node over a complete processing cycle is calculated. The fusion value at any given time deviates from the mean by more than three times the standard deviation, and the fusion value satisfies the condition s. If the fusion value at that moment is not an outlier, it will be marked as an outlier and removed.

[0063] If the number of consecutive outliers does not exceed 5, the missing data is filled in by linear interpolation of the fused values ​​of adjacent normal times. If the number of consecutive outliers exceeds 5, the entire data segment is marked as invalid, and the processing cycle containing that segment is excluded in the subsequent window segmentation in step three. After outlier removal, the remaining data constitutes a reliable dataset, which is stored and output by the host computing device.

[0064] Step 3: Perform multi-dimensional feature extraction and process factor decoupling on the reliable dataset output from Step 2 to form a feature-quality paired sample set; specifically including: Step 3.1: Receive the reliable dataset output from Step 2, and divide the reliable dataset into windows according to the processing cycle of the equipment.

[0065] The processing cycle time used here is the same concept as the processing cycle time used for outlier detection in step 2.4, both based on the complete cycle of the equipment completing one workpiece from clamping to unloading. The start and end times of each window correspond to one processing cycle time, and the window contains the fused value sequence of all working condition nodes and all sensor channels within that cycle time.

[0066] After segmentation, each window represents an independent processing step, and the data within the window fully reflects the changes in various operating parameters during that processing step. The windows are numbered sequentially according to time to facilitate subsequent correlation with quality inspection results. Processing cycles marked as invalid in step 2.4 are skipped directly without entering window segmentation.

[0067] Step 3.2: Extract multidimensional features for each segmented operating condition window. Statistical domain features include the mean, standard deviation, peak value, and root mean square value of the fused values ​​within the window, used to characterize the central tendency and dispersion of the signal within that window.

[0068] The time-domain characteristic is the sum of the absolute values ​​of the differences between adjacent sampling points, divided by the total number of sampling points in the window minus 1. That is, first calculate the absolute value of the difference between every two adjacent sampling points in the window, sum all the absolute values, and then divide by the total number of sampling points minus 1. The resulting quotient reflects the average rate of change of the signal in the time dimension. This normalization process makes windows of different lengths comparable.

[0069] Frequency domain features include the amplitude and spectral centroid of the dominant frequency after performing a Fast Fourier Transform on the fused values ​​within the window, reflecting the energy distribution characteristics of the signal in the frequency dimension. These features together constitute the multidimensional feature vector of the operating condition window, with the vector's dimension equal to the total number of extracted feature components.

[0070] Step 3.3: Label the extracted multidimensional feature vectors to indicate the process factors.

[0071] Specifically, based on the installation location of each sensor and the physical quantity being monitored, each component in the feature vector is labeled with its corresponding process factor source: the feature component corresponding to the temperature sensor is labeled as the temperature control factor, the feature component corresponding to the vibration sensor is labeled as the transmission factor, the feature component corresponding to the pressure sensor is labeled as the pressurization factor, and the feature component corresponding to the current sensor is labeled as the driving factor.

[0072] It should be noted that in step 2.3, weighted fusion of sensors of the same type and operating condition node has been completed, and each node-type combination generates only one fused value at each time. Therefore, the attribution label in this step refers to the fused values ​​of the same process factor monitored on different operating condition nodes.

[0073] For example, two bearing seats in the transmission area are each equipped with a vibration sensor. After fusion in step two, each sensor produces a vibration fusion value sequence. In this step, the feature components of these two sequences are extracted separately, averaged, and then merged into a single representative value for the transmission factor within the processing cycle. This eliminates redundant information between different installation positions under the same process factor, achieving decoupling of the process factor. After decoupling, each process factor retains only one representative value in the multidimensional feature vector of each working condition window.

[0074] Step 3.4: Pair the multidimensional feature vectors of each working condition window obtained after decoupling with the actual quality index values ​​of the corresponding workpiece obtained in offline quality inspection.

[0075] Quality indicators can include dimensional deviations, surface roughness, or geometric tolerances of the machined parts, measured by quality inspection equipment after processing. It should be noted that the data acquired through offline quality inspection is used to establish mapping relationships and belongs to the offline training phase; while the control commands issued in step six based on this mapping relationship act on the real-time processing phase.

[0076] The mapping relationship is established based on the offline detection data of the current batch and is applicable to the production control of the current batch. When the production batch is switched or the equipment has undergone a major overhaul, the operator decides whether to recollect data and establish a new mapping relationship. During pairing, the feature vector is matched one-to-one with the quality index value using the operating condition window number as the index.

[0077] In addition, all components in the feature vector are dimensionless, specifically by subtracting the mean of each component from the mean of all windows. Then divide by the standard deviation of that component across all windows. After dimensionless transformation, the mean of all characteristic components is 0 and the standard deviation is 1, eliminating the dimensional differences between different physical quantities and making the components comparable. The paired and dimensionless dataset, along with the mean of each component... and standard deviation Together, they constitute a feature-quality paired sample set.

[0078] Step 4: Perform multidimensional correlation quantification analysis on the feature-quality paired sample set formed in Step 3 to screen feature subsets that are highly correlated with quality indicators; specifically including: Step 4.1: Receive the feature-quality paired sample set output from Step 3, denoted as N for the total number of samples and P for the number of feature dimensions. First, for each feature dimension... A preliminary assessment of the linear correlation between the index and the quality indicator Y is conducted. The specific procedure is as follows: calculate The Pearson linear correlation coefficient is obtained by dividing the covariance of Y by the product of their respective standard deviations. Since all feature components have been dimensionless in step three, Dimensionless; Y retains its original dimensions, taking dimensional deviation as an example, the unit is millimeters. The value of is between -1 and 1, and it is dimensionless. The larger its absolute value, the greater the value of . The stronger the linear correlation with Y, the better. Set the initial screening threshold to 0.05. Feature dimensions that are not selected are directly discarded and will not participate in the nonlinear correlation calculation in step 4.2. After the initial screening, the remaining feature dimensions are denoted as follows: Proceed to step 4.2.

[0079] Step 4.2: Since the influence of process factors on quality indicators during the manufacturing process is often not a simple linear relationship, the linear correlation coefficient in Step 4.1 alone is insufficient to characterize the full relationship between the two.

[0080] Therefore, the initial screening retained in step 4.1 Based on the existing feature dimensions, conditional variance decomposition is further employed to calculate the variance of each feature dimension. Nonlinear correlation with quality index Y : (2); In the formula, Y is a quality index with the dimension [L]. Taking dimensional deviation as an example, the unit is millimeters. The k-th feature dimension has been dimensionless in step three; Var(Y) is the variance of the quality index Y across all N samples, with dimensions [missing information]. ,Right now ; For a given The conditional variance of Y given the given values, with the same dimensions. ; For conditional variance pairs The weighted expected value of the values, with the same dimensions. .

[0081] In formula (2), the numerator and denominator have the same dimensions. Dimensionless, with values ​​ranging from 0 to 1. The closer to 1, the better. The stronger the explanatory power of Y; If it is close to 0, it indicates It has almost no explanatory power for Y. If Var(Y)=0 (that is, the quality index values ​​of all samples are exactly the same), then the denominator of formula (2) is 0. At this time, it means that the quality index has no fluctuation and no correlation analysis is needed, and the process is terminated.

[0082] The specific calculation process of the expected conditional variance in formula (2) is as follows: The range of values ​​of N is divided into B equal intervals, where B is the integer part of the square root of N. Let the b-th interval be denoted as . Falling into the range The number of samples within is Then the expected conditional variance is equal to the weighted sum of the variances of the quality indicators in each interval, weighted by the sample proportion.

[0083] Among them, the proportion of samples Dimensionless, used as weighting coefficients; the dimensions of the variance of the quality index within each interval are... If there is only one sample in a certain interval (i.e. If the variance of the interval is less than 2, then the variance of the interval cannot be calculated. The interval is then merged with the adjacent interval and recalculated until the number of samples in the merged interval is not less than 2.

[0084] This calculation process will The continuous values ​​of Y are discretized into several intervals. Within each interval, the degree of dispersion of Y is examined, and then a weighted sum is calculated based on the sample proportion. If If there is a strong correlation with Y, then when When the value of Y is limited to a small interval, the fluctuation of Y will be significantly reduced, and the conditional variance will decrease accordingly. It approaches 1; conversely, if If it is not related to Y, then the conditional variance is equivalent to the population variance. Approaching 0.

[0085] Step 4.3: Calculate the nonlinear correlation degree of each feature dimension based on the results obtained in Step 4.2. For all Each feature dimension is as follows Sort by size from largest to smallest.

[0086] Set the correlation threshold to 0.3, and... Feature dimensions greater than 0.3 are preserved. Feature dimensions with a value less than or equal to 0.3 are removed. If the number of retained feature dimensions still exceeds 15, only the top 15 feature dimensions are retained to control the complexity of subsequent mapping expressions and avoid overfitting caused by too many input variables. After sorting and filtering, the retained feature dimensions constitute a subset of highly correlated features. Each retained feature dimension has its correlation value and corresponding process factor source label recorded.

[0087] Step 4.4: Reintegrate the dimensionless values ​​of each feature dimension in the highly correlated feature subset obtained in Step 4.3 with the corresponding quality index values ​​to form a correlated feature set.

[0088] The associated feature set contains only feature dimensions with high correlation to quality indicators, eliminating redundant dimensions with low correlation. This provides a concise and informative set of input variables for constructing the quality impact mapping expression in step five. The associated feature set, along with the correlation values ​​of each dimension and the annotations of process factor sources, is stored by the host computing device.

[0089] Step 5: Based on the associated feature set output in Step 4, construct a quality impact mapping expression, and solidify the output after verification and correction. Specifically, this includes: Step 5.1: Receive the associated feature set output from Step 4 and determine the input and output variables of the quality influence mapping expression.

[0090] The input variables are the dimensions of the highly correlated features retained in the correlated feature set, denoted as... Where K is the number of feature dimensions retained after filtering in step 4.3, and K is no greater than 15. The output variable is the quality index Y, with the dimension [L]. Taking size deviation as an example, the unit is millimeters. Since all feature components have been dimensionless in step three, each input variable... All of them are dimensionless quantities, which provides a premise for the unification of dimensions of each coefficient in the subsequent mapping expression.

[0091] Step 5.2: For the determined input variables A quality impact mapping expression is established using multiple regression with respect to the output variable Y. This expression includes an intercept term, a linear term, and pairwise interaction terms, and its specific form is as follows: (3); In the formula, The predicted value of the quality indicator is [L]. The intercept term represents the baseline value of the quality index when all input variables are 0, with the dimension [L]. For the k-th input variable The corresponding coefficient of the linear term, since Dimensionless Dimensions and Consistent, represented by [L]; The coefficient of the cross term between the k-th and l-th input variables is given by... and All are dimensionless. The dimension of is also [L]; K is the total number of input variables, which is dimensionless. In formula (3), all terms on the right-hand side have the dimension [L], which is consistent with the dimension of the left-hand side. The dimensions are consistent.

[0092] The coefficients in formula (3) are solved using the least squares method. Specifically, the input variable values ​​and measured quality index values ​​of all N samples in the associated feature set are substituted into formula (3), and the sum of the squares of the differences between the measured values ​​and the predicted values ​​of all samples is calculated, denoted as the residual sum of squares, with dimensions of: .

[0093] The sum of squares of the residuals are respectively with respect to ,each and each Find the partial derivatives and set them to 0 to obtain a set of linear equations (i.e., normal equations). The number of unknowns in the set of equations is equal to 1 + K + K(K-1) / 2, which is the total number of intercept terms, coefficients of linear terms, and coefficients of cross terms.

[0094] To ensure a unique stable solution to the system of equations, the total number of samples N must be at least three times the number of unknowns. When K=15, the number of unknowns is 121, so N≥363 is required. The coefficient matrix of the normal equations is a symmetric positive semi-definite matrix. When there is no strict linear correlation between the input variables (i.e., the design matrix has full column rank), this matrix is ​​positive definite and can be solved using the Joleski decomposition method. If there is an approximate linear correlation that causes the matrix to approach singularity, ridge regression is used. A small regularization term (regularization coefficient of 0.01) is added to the diagonal of the coefficient matrix before solving to ensure numerical stability.

[0095] Step 5.3: Perform residual verification and one-time correction on the quality influence mapping expression obtained in Step 5.2.

[0096] Specifically, the input variables of each sample in the associated feature set are substituted into formula (3) to calculate the predicted value. Compared with measured values The residuals between And calculate the mean of the residuals. and standard deviation .

[0097] If the mean of the residuals If the absolute value of the value is greater than 5% of the allowable tolerance zone for the quality index, it indicates a systematic shift in the expression. A one-time correction is then performed on the intercept term, resulting in the corrected intercept term. .

[0098] If the residual standard deviation If the correlation degree exceeds 30% of the allowable tolerance range of the quality index, it indicates that the feature dimensions retained in the current set of related features are insufficient to explain the quality index. In this case, the correlation degree in step 4.2 will be adjusted. Feature dimensions with values ​​between 0.2 and 0.3 that were previously removed in step 4.3 are added back in, and the coefficients of steps 5.1 and 5.2 are recalculated using the expanded set of associated features.

[0099] This supplement is performed only once. If the residual standard deviation still does not meet the requirements after the supplement, the current residual level is recorded, the obtained mapping expression is retained as the currently available result, and operators are reminded to pay attention to quality fluctuations in subsequent use. This step does not involve multiple rounds of iterative processing.

[0100] Step 5.4: The quality influence mapping expression, verified and corrected in Step 5.3, along with the values ​​of each coefficient, the numbers of the input variables, and the corresponding explanations of the sources of the process factors, is solidified into a quality influence mapping relationship. This mapping relationship is stored by the host computer.

[0101] Step Six: Based on the quality impact mapping relationship output in Step Five, generate equipment control commands and send them to the execution layer, specifically including: Step 6.1: Receive the quality impact mapping relationship solidified in Step 5. The operator, based on the quality requirements of the current production order, sets the allowable range of the target quality index in the host computer. This allowable range is determined by the nominal value of the quality index and its upper and lower deviations.

[0102] For example, if the nominal value of the dimensional deviation is 0, the upper deviation is +0.05 mm, and the lower deviation is -0.05 mm, then the allowable range is [-0.05, +0.05] mm. This allowable range, along with the quality influence mapping relationship output in step five, serves as the input condition for subsequent inverse kinematics calculations.

[0103] Step 6.2: Based on the allowable range of the target quality index set in Step 6.1 and the quality influence mapping relationship output in Step 5, solve the allowable range of each process parameter.

[0104] Specifically, in formula (3), one of the input variables to be determined is excluded. The remaining input variables are fixed as the feature values ​​corresponding to the most recent completed processing cycle. At this time, formula (3) is related to It degenerates into a linear equation, where the constant term C is the intercept term, and the remaining linear terms and no other terms are present. The sum of the intersection terms, The comprehensive coefficient A is With all The cross term coefficients are multiplied by the sum of their corresponding eigenvalues.

[0105] Will Take the upper and lower limits of the allowed interval respectively, substitute them into the solution, and then solve. The two corresponding boundary values. If When the comprehensive coefficient approaches 0, it indicates that the impact of this variable on the quality index has been offset by the other interaction terms under the current operating conditions. In this case, no control constraints need to be applied to this variable, and it can be skipped. After performing the above inverse solution operation on each input variable in sequence, a multivariate joint verification is also required.

[0106] During joint validation, the input variables with the highest correlation (min(K,3)) from step 4.3 are selected (if K is less than 3, all K are selected; if K=1, there are no cross terms, and joint validation is not required). These variables are then set to their upper or lower bounds, forming a total of [number of variables]. Each extreme combination is substituted into formula (3) one by one to verify. Whether it still falls within the allowable range.

[0107] If a certain combination If the range exceeds the allowable range, the allowable range for variables ranked lower in correlation should be appropriately narrowed until all extreme combinations are reached. All values ​​meet the allowable range requirements. Since the input variables have been dimensionless in step three, the boundary values ​​obtained from the inverse solution are dimensionless and need to be based on the mean values ​​recorded in step three. and standard deviation This is done by restoring the dimensionless boundary values ​​to process parameter values ​​with actual physical dimensions. The restoration method is to multiply the dimensionless boundary values ​​by the standard deviation. Add the mean For example, it can restore dimensionless temperature boundary values ​​to temperature values ​​in Kelvin.

[0108] Step 6.3: Convert the allowable ranges of each process parameter obtained from the inverse solution in Step 6.2 into control commands for each actuator.

[0109] Specifically, for heating or cooling actuators corresponding to temperature control factors, the allowable range of temperature parameters is converted into a set range of heating power or cooling power based on the equipment's thermal balance relationship.

[0110] For heating actuators, the heating power is the product of the mass of the workpiece being heated (in kilograms), the specific heat capacity of the workpiece (in J / kg·K), and the target temperature rise (in K), divided by the heating cycle time (in seconds). The resulting quantity is J / s, which is equivalent to watts (W).

[0111] The workpiece mass and specific heat capacity are input by the operator into the host computer based on the material and size of the workpiece being processed. After input, the host computer stores the data for use in each processing cycle.

[0112] For cooling actuators, the cooling power is taken as the coolant volume flow rate (unit: Coolant density (unit: The product of the following four values ​​is J / s, which is equivalent to watts (W): specific heat capacity of coolant (in J / kg·K), temperature difference between coolant inlet and outlet (in K). For air-cooled actuators, replace the above quantities with the corresponding air parameters. The physical properties of the coolant or air are input by the operator into the host computer. For hydraulic actuators with pressurized parameters, multiply the allowable range of the pressure parameter by the piston area of ​​the hydraulic cylinder (unit: ...). The range of force is obtained, and then the set range of pump displacement is calculated based on the hydraulic pump displacement-pressure characteristic curve that comes with the equipment at the factory.

[0113] For the motor actuator corresponding to the driving factor, the allowable range of the current parameter is multiplied by the motor torque constant (in N·m / A) to obtain the torque range, and then converted into the speed setting range according to the speed-torque characteristic curve provided with the motor at the factory.

[0114] The control instructions for each actuator include a set value, an allowable deviation range, and an execution priority. The execution priority is determined based on the correlation value of the process factor recorded in step four; the higher the correlation, the higher the priority.

[0115] Step 6.4: Send the control commands for each actuator generated in Step 6.3 to the device execution layer via industrial Ethernet.

[0116] Before issuing the commands, the host computer performs a consistency check on each control command. The check criteria are: the sum of the set power of all actuators must not exceed 95% of the rated power of the equipment, and the heating power setting and cooling power setting must meet the safety margin that the net heat input does not exceed the heat capacity of the workpiece in terms of thermal balance.

[0117] Specifically, in terms of heating power (Unit: W, obtained according to the conversion method for heating actuators in step 6.3) and cooling power. (The unit is W, obtained according to the conversion method of the cooling actuator in step 6.3) For example, the difference between the two is required to be no more than 90% of the product of the workpiece mass, the workpiece specific heat capacity, and the workpiece maximum allowable temperature rise (in K) divided by the heating cycle time (in s). The meaning of each symbol is the same as in step 6.3.

[0118] After successful verification, the control command is transmitted to the control unit of each actuator via industrial Ethernet or fieldbus. The actuator adjusts its operating parameters according to the set values ​​in the command. Upon receiving the command, the actuator transmits the current actual operating value back to the host computer via industrial Ethernet. The host computer verifies whether the value falls within the set range specified in the command; this verification is a one-time confirmation and does not form a continuous adjustment loop.

[0119] The above description is merely an exemplary embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention specification and drawings under the technical concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A smart control method for intelligent manufacturing equipment based on the Internet of Things, characterized in that, include: Multiple types of sensor arrays are deployed at various working nodes of intelligent manufacturing equipment to identify interference frequency bands and filter out notch filters on the collected raw signals, thereby obtaining an anti-interference raw dataset. The original anti-interference dataset is subjected to multi-channel weighted fusion and outlier identification to obtain a reliable dataset; The reliable dataset is windowed according to the processing cycle and multidimensional features are extracted. After decoupling from process factors, these features are paired with quality index values ​​to form a feature-quality paired sample set. The feature-quality paired sample set is subjected to correlation quantification analysis to screen the feature subset that is highly correlated with the quality index, forming a correlated feature set; Based on the associated feature set, a quality influence mapping expression is constructed, and the allowable tolerance zone of the quality index is set according to the quality requirements. After inspection and correction, it is solidified into a quality influence mapping relationship. Based on the quality influence mapping relationship, the allowable range of each process parameter is deduced, equipment control instructions are generated and sent to the execution layer.

2. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 1, characterized in that, The process of identifying interference frequency bands and filtering out notch filters on the acquired raw signals yields an anti-interference raw dataset, including: The original signal sequence after cataloging the edge gateway is subjected to a fast Fourier transform channel by channel to convert the time-domain signal into a frequency-domain representation, thereby obtaining the amplitude spectrum of each sensor channel. The transform window length is taken as the sampling frequency divided by 1.25 and then rounded up to the nearest power of 2 to ensure that the frequency resolution does not exceed 1.25 Hz. The amplitude spectrum is compared band by band with the pre-calibrated baseline spectrum. If the amplitude of a certain frequency band exceeds three times the amplitude of the corresponding frequency band of the baseline spectrum, it is marked as an interference frequency band and its lower and upper frequency limits are recorded. For each marked interference frequency band, a second-order notch filter is constructed with its center frequency as the notch frequency and the frequency band width as the notch bandwidth. When there are multiple interference frequency bands in the same channel, the corresponding number of second-order notch filters are cascaded. The number of cascaded filters in a single channel shall not exceed 5. The filtered frequency domain signal is restored to the time domain signal by performing an inverse fast Fourier transform, and together with the sensor type, sensor number, installation location and acquisition timestamp, it forms the original anti-interference dataset.

3. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 1, characterized in that, The original, anti-interference dataset is subjected to multi-channel weighted fusion and outlier identification to obtain a reliable dataset, including: The original dataset against interference is grouped and indexed according to two dimensions: sensor type and operating node. Multiple sensor channels of the same type on the same operating node are grouped together and sorted by acquisition timestamp. Time alignment and spatial registration are performed on each channel within each group; spatial attenuation correction is introduced for the vibration sensor channel based on the installation distance of each sensor from the center of the working condition, while no spatial attenuation correction is performed for the temperature sensor, pressure sensor and current sensor. For each set of multi-channel data after registration, the reciprocal of the measurement variance of each channel within a sampling window is used as the fusion weight. The measurement values ​​of each channel are weighted and summed according to the weights, and then divided by the sum of the weights to obtain the fusion value at each time step. Calculate the mean and standard deviation of the fused value at each working node over a complete processing cycle; If the fusion value at any given time deviates from the mean by more than 3 times the standard deviation, it is marked as an outlier and removed. If the number of consecutive outliers marked as outliers does not exceed 5, the values ​​of the fused values ​​of adjacent normal times are used to make linear interpolation to fill the gap; if the number of consecutive outliers exceeds 5, the outliers will be marked as invalid and the entire processing cycle in which they occur will be excluded during window segmentation. The data remaining after outlier removal constitutes the reliable dataset.

4. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 3, characterized in that, The time alignment and spatial registration of each channel within each group includes: Align the acquisition timestamps of each channel within the same group so that each channel has a corresponding measurement value at the same time. For vibration sensor channels, based on the installation distance of each sensor from the center of the working condition, a spatial attenuation correction coefficient is introduced according to the geometric diffusion law of body wave propagation to correct the measured values ​​of each channel to a unified reference position; if the installation distance is less than 0.01 meters, no spatial attenuation correction is performed. For temperature sensor channels, pressure sensor channels, and current sensor channels, since the physical quantities they monitor do not have propagation characteristics that attenuate with distance within the operating condition node range, no spatial attenuation correction is performed.

5. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 1, characterized in that, The reliable dataset is windowed according to the processing cycle and multidimensional features are extracted. After decoupling from process factors, these features are paired with quality index values ​​to form a feature-quality paired sample set, including: Based on the processing cycle of the equipment, the reliable dataset is divided into several working condition windows, and each working condition window corresponds to a complete action cycle of a workpiece from clamping to unloading. For each operating condition window, statistical domain features, time domain features, and frequency domain features are extracted to form a multi-dimensional feature vector; Based on the installation location of each sensor and the physical quantity monitored, the components in the multidimensional feature vector are labeled with the corresponding process factor sources. The fusion values ​​of different installation locations under the same process factor are averaged and merged into a single representative value to achieve decoupling of process factors. The decoupled multidimensional feature vectors are paired with the offline quality inspection results of the corresponding processed parts, and all feature components are dimensionless to form a feature-quality pairing sample set.

6. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 1, characterized in that, The feature-quality paired sample set is subjected to correlation quantification analysis to screen feature subsets that are highly correlated with quality indicators, forming a correlated feature set, including: A preliminary assessment is made of the linear correlation between each feature dimension and the quality index in the feature-quality paired sample set, and feature dimensions whose absolute value of linear correlation coefficient is less than the initial screening threshold are removed. For each feature dimension retained in the initial screening, the value range of the feature dimension is divided into several intervals. The variance of the quality index is calculated in each interval. Then, the variances of each interval are weighted and summed with the sample proportion as the weight to obtain the expected conditional variance. The ratio of the expected conditional variance to the overall variance of the quality index is subtracted from 1 to obtain the nonlinear correlation degree. Each feature dimension is sorted from largest to smallest according to the nonlinear correlation degree, and feature dimensions with a correlation degree greater than the correlation degree threshold are retained; if the number of retained exceeds 15, only the top 15 are retained to form a subset of highly correlated features. The dimensionless values ​​of each feature dimension in the highly correlated feature subset are re-integrated with the corresponding quality index values ​​to form the correlated feature set.

7. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 1, characterized in that, The construction of the quality impact mapping expression based on the associated feature set includes: The highly correlated feature dimensions retained in the aforementioned correlated feature set are used as input variables, and the quality index is used as the output variable. A quality influence mapping expression containing intercept, linear term, and pairwise interaction term is established using multiple regression. Each coefficient is solved by least squares, which is to sum the squares of the differences between the measured and predicted values ​​of all samples to obtain the residual sum of squares. The partial derivatives of the residual sum of squares with respect to each coefficient are calculated and set to zero to obtain the normal equation system, which is then solved. To ensure that the normal equation system has a unique stable solution, the total number of samples must be no less than three times the number of unknowns. When there is an approximate linear correlation between the input variables, a regularization term is added to the diagonal of the coefficient matrix of the normal equation system before solving.

8. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 7, characterized in that, The quality impact mapping relationship, which has been verified and corrected, includes: Substitute the input variables of each sample in the feature-quality paired sample set into the quality influence mapping expression, calculate the residual between the predicted value and the measured value, and calculate the mean and standard deviation of the residual; If the absolute value of the mean of the residuals is greater than 5% of the allowable tolerance zone of the quality index, then the intercept term is corrected once. If the standard deviation of the residual is greater than 30% of the allowable tolerance band of the quality index, then the feature dimensions with correlation between the initial screening rejection threshold and the retention threshold are added to the system, and the coefficients are recalculated using the expanded correlation feature set. The addition is performed only once. If the standard deviation of the residual still does not meet the requirements after the addition, the obtained quality influence mapping expression is retained as the currently available result. The quality influence mapping expression, after verification and correction, along with the values ​​of each coefficient, is solidified into the quality influence mapping relationship.

9. The intelligent control method for intelligent manufacturing equipment based on the Internet of Things according to claim 1, characterized in that, The step of reversing the permissible range of each process parameter based on the quality influence mapping relationship, generating equipment control commands, and issuing them to the execution layer includes: Set the allowable range of target quality indicators based on the quality requirements of the current production orders; In the quality influence mapping expression, all input variables except for one input variable to be determined are fixed as the feature values ​​corresponding to the most recently completed processing cycle. The quality influence mapping expression is degenerated into a linear equation with respect to the input variable to be determined. The upper and lower limits of the allowable interval of the quality index are substituted into the equation to solve for the two boundary values ​​corresponding to the input variable to be determined. The above inverse solution operation is performed sequentially for each input variable. For the input variables with the highest correlation ranking, perform multivariate joint validation, setting them to the upper or lower bounds to form extreme combinations, and check whether the predicted values ​​of the quality indicators still fall within the allowable range; if they exceed the range, narrow the allowable range for the variables with lower correlation ranking. The dimensionless boundary values ​​obtained from the inverse kinematics are restored to characteristic values ​​with actual physical dimensions and transformed into control commands for each actuator. Before being issued, the sum of the set power of all actuators is checked for consistency. After the check passes, the control commands are issued to the device execution layer via industrial Ethernet.

10. An intelligent control platform for smart manufacturing equipment based on the Internet of Things, characterized in that, The method for implementing the intelligent control method for IoT-based smart manufacturing equipment as described in any one of claims 1-9 includes: Multiple types of sensor arrays are deployed at various working nodes of intelligent manufacturing equipment to collect raw signals; The edge gateway receives the raw signals collected by the multi-type sensor arrays through the Internet of Things communication link, and performs cataloging, fast Fourier transform, interference frequency band identification, notch filtering and inverse fast Fourier transform on the raw signals, and outputs anti-interference raw dataset. The host computing device is connected to the edge gateway via an industrial Ethernet. It is used to perform multi-channel weighted fusion and outlier identification on the anti-interference original dataset to obtain a reliable dataset. The reliable dataset is windowed according to the processing cycle and multi-dimensional features are extracted. After decoupling by process factors, it is paired with quality index values ​​to form a feature-quality paired sample set. The feature-quality paired sample set is subjected to correlation quantification analysis to form a correlation feature set. Based on the correlation feature set, a quality influence mapping expression is constructed and solidified into a quality influence mapping relationship. Based on the quality influence mapping relationship, the allowable range of each process parameter is deduced and equipment control instructions are generated. The device execution layer receives device control commands from the host computer via industrial Ethernet or fieldbus, and adjusts the operating parameters according to the set values ​​in the commands.