A control optimization method based on timing data of industrial equipment joint control

Abstract

This invention relates to the field of data-driven joint control technology, specifically a control optimization method for time-series data of industrial equipment joint control. It includes: collecting motor speed at the pay-off end, machining host linear speed, winding torque at the take-up end, and displacement at the buffer end to generate joint control time-series data; performing Fast Fourier Transform to calculate the frequency domain features of the operating conditions, and performing timestamp mapping to generate a time-series feature tensor; obtaining the difference from the physical limit boundary to generate tension buffer margin and adaptive weights, and performing weighted calculation to output spatiotemporal correlation features; inputting a fully connected network to calculate the host speed prediction value, generating speed deviation and displacement deviation through calculation, and fusing them to generate the speed change demand; and generating pay-off compensation instructions and take-up compensation instructions. This invention treats independent feature acquisition nodes as a holistic model of feature fusion, which helps to extract and adapt to complex coupling mapping relationships between multi-source data, and balances and smooths multi-dimensional noise disturbances in dynamic and continuous pattern recognition.

Description

Technical Field

[0001] This invention relates to the field of data-driven joint control technology, specifically a control optimization method for time-series data of industrial equipment joint control. Background Technology

[0002] In continuous manufacturing environments (such as wire and cable, optical fiber, and coil processing), multi-level coordinated control of industrial equipment, including the pay-off end, tension buffer end, processing host, and take-up end, is an important current technological application direction. Existing technologies generally employ basic time-series data monitoring combined with simple threshold matching methods. This involves collecting time-series data such as displacement or tension of tension buffer mechanisms (e.g., wire storage racks) on the production line, and performing pattern recognition of equipment operating status based on predefined fixed rules. When data exceeds limits, a preset logic architecture is automatically triggered, and an adjustment signal is output. This is currently the mainstream technology for achieving tension balance and data-driven control in industrial production lines.

[0003] However, existing technical solutions have significant shortcomings in the accuracy of time-series data feature extraction and pattern recognition under complex working conditions. First, existing joint control systems suffer from rigidity and coarseness in the preprocessing and feature representation of heterogeneous data. Faced with heterogeneous time-series signals such as high-frequency rotation and low-frequency floating generated by different physical structures on the production line, they usually use a single fixed period for sampling, which is difficult to effectively filter out complex frequency domain noise interference caused by mechanical vibration, resulting in severe dimensionality distortion of the underlying feature vectors of the input model. Second, in terms of multivariate feature fusion and feature space construction, existing solutions often treat the time-series parameters of multi-source equipment equally or statically, failing to keenly identify the nonlinear data mutation patterns of buffer nodes when approaching the mechanical physical limits, resulting in a serious sluggishness in the feature perception and weight allocation of the data-driven model for the system's critical state of collapse. Finally, their time-series prediction-based inference mechanism suffers from passive lag and singularity. The system relies excessively on the local state feedback of the current time step and lacks long-term dependency feature extraction of the global operating trend of the main traction equipment, failing to effectively decouple and coordinate the active prediction trend features and passive buffer state features. In summary, existing technologies cannot match the complex dynamic coupling characteristics between multiple devices, which can easily lead to tension loss and line breakage during continuous high-speed production.

[0004] To address this, a control optimization method based on timing data of industrial equipment joint control is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a control optimization method for industrial equipment joint control timing data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a control optimization method for industrial equipment joint control timing data, comprising: The system collects data from the edge control equipment, including the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end, to generate joint control timing data. It then performs a fast Fourier transform on the joint control timing data to calculate the operating condition frequency domain characteristics. Based on these operating condition frequency domain characteristics, it generates a target sampling period through inverse proportional mapping, and resamples according to the target sampling period to generate a timing feature tensor. The difference between the physical limit boundary calculation and the buffer end displacement is obtained from the edge control device to generate the tension buffer margin; the deviation between the safety warning threshold and the calculated tension buffer margin is obtained, and adaptive weights are generated through exponential growth; the temporal feature tensor is weighted according to the adaptive weights to output the spatiotemporal correlation features; Input the spatiotemporal correlation features into the speed prediction value of the fully connected network computing host, compare it with the linear speed of the processing host to generate the feedforward speed deviation, extract the buffer end displacement to calculate the buffer displacement deviation, and fuse the feedforward speed deviation and the buffer displacement deviation to generate the speed change demand. Based on the speed change demand, pay-off compensation commands and take-up compensation commands are generated; these commands are then converted into control electrical signals to trigger the hardware frequency conversion action.

[0007] Preferably, the specific generation process of the operating condition frequency domain features includes: extracting the joint control time series data, dividing the joint control time series data into macroscopic observation windows according to the system-set observation period; starting an asynchronous analysis process within the macroscopic observation window to perform a fast Fourier transform on the joint control time series data and outputting a frequency domain power spectrum sequence; parsing the frequency domain power spectrum sequence to locate the node with the maximum energy, and extracting the peak amplitude value corresponding to the node with the maximum energy as the dominant frequency amplitude; obtaining the stable reference value set for the initial state of equipment operation, comparing the dominant frequency amplitude and the stable reference value to calculate the difference ratio, and aggregating the difference ratio and the dominant frequency amplitude to generate the operating condition frequency domain features.

[0008] Preferably, the specific generation process of the time-series feature tensor includes: receiving the operating condition frequency domain features, extracting the differential ratio contained in the operating condition frequency domain features, constructing an inverse proportional mapping function between the differential ratio and the target sampling period; substituting the differential ratio into the inverse proportional mapping function to calculate and output the target sampling period; extracting the original timestamp contained in the joint control time-series data, performing grid resampling on the original timestamps according to the target sampling period to generate aligned time nodes; performing linear interpolation calculation on the joint control time-series data at the aligned time nodes to output a reconstructed time-series sequence; and combining all reconstructed time-series sequences to generate the time-series feature tensor.

[0009] Preferably, the specific construction process of the tension buffer margin includes: establishing a communication connection with the edge control device; reading the top and bottom coordinates of the guide rail from the internal register of the edge control device; combining the top and bottom coordinates of the guide rail as the physical limit boundary; extracting the buffer end displacement contained in the joint control timing data; comparing the buffer end displacement with the top coordinate of the guide rail to calculate the absolute difference at the top; comparing the buffer end displacement with the bottom coordinate of the guide rail to calculate the absolute difference at the bottom; and selecting the tension buffer margin by comparing the absolute difference at the top and the absolute difference at the bottom.

[0010] Preferably, the specific generation process of the adaptive weight includes: extracting the tension buffer margin and obtaining the safety warning threshold; comparing the tension buffer margin and the safety warning threshold to calculate the margin deviation; constructing an exponential growth formula between the margin deviation and the adaptive weight; when the tension buffer margin is less than the safety warning threshold, substituting the margin deviation into the exponential growth formula to perform calculation, generating a surge weight value as the adaptive weight; when the tension buffer margin is not less than the safety warning threshold, substituting the margin deviation into the exponential growth formula to perform calculation, generating a smooth weight value as the adaptive weight.

[0011] Preferably, the specific generation process of the spatiotemporal correlation feature includes: receiving the adaptive weights and the temporal feature tensor, and constructing a spatiotemporal attention matrix within the deep prediction model; mapping the adaptive weights to the buffer end displacement data channel corresponding to the spatiotemporal attention matrix to generate an updated attention matrix; extracting the temporal feature tensor and performing a matrix transpose operation to generate a transposed feature tensor; performing matrix multiplication between the transposed feature tensor and the updated attention matrix to calculate and output a spatial fusion tensor; performing a sliding scan operation along the data direction on the spatial fusion tensor to extract the dependency data between channels, and concatenating the dependency data to generate the spatiotemporal correlation feature.

[0012] Preferably, the specific generation process of the speed change demand includes: extracting the spatiotemporal correlation features, flattening the spatiotemporal correlation features into a tiled feature vector; inputting the tiled feature vector into the hidden layer nodes contained in the fully connected network to perform nonlinear activation, and outputting the host speed prediction value; comparing the host speed prediction value with the processing host linear speed to calculate the feedforward speed deviation, and extracting the buffer end displacement to calculate the buffer displacement deviation; obtaining a system-preset displacement speed coefficient, multiplying the buffer displacement deviation by the displacement speed coefficient to convert it into a feedback speed deviation; superimposing the feedforward speed deviation and the feedback speed deviation to calculate the comprehensive speed error; extracting a preset comprehensive speed adjustment coefficient, multiplying the comprehensive speed error by the comprehensive speed adjustment coefficient to perform a multiplication operation, and outputting the calculation result as the speed change demand.

[0013] Preferably, the specific generation process of the wire release compensation command and the wire take-up compensation command includes: receiving the speed change demand, obtaining the wire release side line speed frequency conversion coefficient and the wire take-up side line speed frequency conversion coefficient built into the edge control device; multiplying the speed change demand by the wire release side line speed frequency conversion coefficient to calculate the wire release basic compensation amount, obtaining a preset frequency conversion coefficient, multiplying the tension buffer margin by the frequency conversion coefficient to calculate the wire release amplitude limiting threshold, performing amplitude limiting truncation operation on the wire release basic compensation amount according to the wire release amplitude limiting threshold to generate the wire release compensation command; multiplying the speed change demand by the wire take-up side line speed frequency conversion coefficient to calculate the wire take-up basic compensation amount, multiplying the tension buffer margin by the frequency conversion coefficient to calculate the wire take-up amplitude limiting threshold, performing amplitude limiting truncation operation on the wire take-up basic compensation amount according to the wire take-up amplitude limiting threshold to generate the wire take-up compensation command; converting the wire release compensation command and the wire take-up compensation command into the control electrical signal, and outputting it to the edge control device to trigger hardware frequency conversion action.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By performing a Fast Fourier Transform on the joint control time series data, the operating condition frequency domain characteristics are calculated, and the target sampling period is determined based on the operating condition frequency domain characteristics. The time series data is then timestamped according to the target sampling period to generate a time series feature tensor. The data sampling density is dynamically adjusted according to the actual fluctuation state of operation, which helps to capture the subtle dynamic changes between multiple devices in continuous high-speed production. This provides more suitable data feature support for subsequent control algorithms, thereby alleviating the tension fluctuation phenomenon caused by the difficulty of matching complex dynamic coupling characteristics with fixed sampling data to a certain extent.

[0015] 2. By obtaining the physical limit boundary from the edge control device, calculating the difference between the physical limit boundary and the physical limit boundary, a tension buffer margin is generated. The temporal feature tensor and the tension buffer margin are then input into the depth prediction model to generate adaptive weights. The temporal feature tensor is weighted according to the adaptive weights to output spatiotemporal correlation features. This maps the physical safety operating space of the actual equipment to the feature extraction calculation process, enabling the algorithm to have physical constraint guidance when dealing with the complex dynamic coupling characteristics between multiple devices. This helps to perceive the contraction trend of the safety boundary under continuous high-speed production conditions, and then to intervene appropriately before the equipment approaches the physical limit, reducing the risk of tension runaway and line breakage shutdown in the target production line.

[0016] 3. By inputting spatiotemporal correlation features into a fully connected network to calculate the predicted speed of the main machine, the system compares this value with the linear speed of the main machine to generate a feedforward speed deviation. This deviation is then calculated by extracting the tension buffer margin, and finally, the feedforward speed deviation and buffer displacement deviation are fused to generate the speed change demand. Based on this speed change demand, the system generates pay-off compensation and take-up compensation commands, thus constructing a composite control logic that takes into account both trend prediction and actual physical state feedback. This logic can smoothly adapt to the complex dynamic coupling characteristics between multiple devices, guiding each actuator to perform coordinated compensation actions. This helps maintain the dynamic balance of the system in continuous high-speed production and provides a more robust control strategy to prevent tension runaway and machine shutdown due to wire breakage.

[0017] 4. By collecting data from edge control devices, including the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end, a coordinated control timing data is generated. This data is then combined with feedforward speed deviation and buffer displacement deviation to generate speed change requirements. Finally, pay-off compensation commands and take-up compensation commands are generated based on these speed change requirements and output to the equipment to trigger hardware frequency conversion actions. This constructs a globally collaborative control architecture that treats independent execution components as an organic whole with interconnected data. This helps to coordinate and adapt to the complex dynamic coupling characteristics between multiple devices from a global perspective. It can balance and smooth multi-end disturbances in continuous high-speed production, providing overall operational assurance to prevent tension loss of control and wire breakage shutdowns in the target production line. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a control optimization method based on timing data for industrial equipment joint control, as proposed in an embodiment of this invention. Figure 2 This is a flowchart illustrating the generation of time-series feature tensors based on operating condition frequency domain features, as proposed in an embodiment of this invention. Figure 3 This is a flowchart illustrating the compensation instruction generation process based on composite deviation fusion proposed in an embodiment of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1-3 The present invention provides a control optimization method for industrial equipment joint control timing data, the specific steps of which are as follows: The system collects data from the edge control equipment, including the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end, to generate joint control timing data. It then performs a fast Fourier transform on the joint control timing data to calculate the operating condition frequency domain characteristics. Based on these operating condition frequency domain characteristics, it generates a target sampling period through inverse proportional mapping, and resamples according to the target sampling period to generate a timing feature tensor. The difference between the physical limit boundary calculation and the buffer end displacement is obtained from the edge control device to generate the tension buffer margin; the deviation between the safety warning threshold and the calculated tension buffer margin is obtained, and adaptive weights are generated through exponential growth; the temporal feature tensor is weighted according to the adaptive weights to output the spatiotemporal correlation features; Input the spatiotemporal correlation features into the speed prediction value of the fully connected network computing host, compare it with the linear speed of the processing host to generate the feedforward speed deviation, extract the buffer end displacement to calculate the buffer displacement deviation, and fuse the feedforward speed deviation and the buffer displacement deviation to generate the speed change demand. Based on the speed change demand, pay-off compensation commands and take-up compensation commands are generated; these commands are then converted into control electrical signals to trigger the hardware frequency conversion action.

[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1

[0023] This application discloses a control optimization method based on timing data of industrial equipment joint control. (See attached document.) Figure 1 The specific steps proposed in this invention include: S1. Collecting the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end from the edge control device to generate joint control timing data; performing a fast Fourier transform on the joint control timing data to calculate the operating condition frequency domain characteristics; generating a target sampling period based on the operating condition frequency domain characteristics through inverse proportional mapping, and resampling according to the target sampling period to generate a timing feature tensor; S2. Obtaining the difference between the physical limit boundary calculation and the buffer end displacement from the edge control device to generate a tension buffer margin; obtaining the safety warning threshold and the tension buffer margin calculation margin. The degree of deviation is determined by generating adaptive weights through exponential growth; the temporal feature tensor is weighted according to the adaptive weights to output the spatiotemporal correlation features; S3, the spatiotemporal correlation features are input to the fully connected network to calculate the host speed prediction value, compared with the processing host linear speed to generate the feedforward speed deviation, the buffer end displacement is extracted to calculate the buffer displacement deviation, and the feedforward speed deviation and buffer displacement deviation are fused to generate the speed change demand; S4, the wire feeding compensation command and wire take-up compensation command are generated based on the speed change demand; the wire feeding compensation command and wire take-up compensation command are converted into control electrical signals to trigger the hardware frequency conversion action.

[0024] Further, the speed of the motor at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end are collected from the edge control equipment to generate joint control timing data; a fast Fourier transform is performed on the joint control timing data to calculate the operating condition frequency domain characteristics; the target sampling period is determined based on the operating condition frequency domain characteristics, and a timestamp mapping is performed on the joint control timing data according to the target sampling period to generate a timing feature tensor; this corresponds to step S1 above; see [link to relevant documentation]. Figure 2 The specific implementation process includes: In specific hardware deployment scenarios, the underlying environment is configured with an edge control device based on the industrial Ethernet bus protocol. This edge control device establishes data communication connections with various actuators in the field, synchronously extracts data information from four core dimensions, and integrates them to generate initial joint control timing data. The specific input channels of the data and their physical representation are as follows: The speed of the wire feeding end motor: It represents the initial release speed when the raw material enters the machine, and is obtained by pulse counting by an incremental photoelectric encoder installed at the tail end of the wire feeding shaft motor.

[0025] Processing machine linear speed: Characterizes the actual linear speed of the material as it passes through the drawing die, and is calculated by an absolute value rotary encoder on the side of the main drawing traction wheel.

[0026] Take-up end winding torque: Characterizes the winding force state of the finished material during winding, and is obtained by mapping and calculation from the internal current feedback loop of the take-up shaft driver.

[0027] Buffer end displacement: Characterizes the current physical position of the tension absorption device on the set track, which is obtained by a linear displacement sensor installed on the tension swing or buffer guide rail.

[0028] The analog-to-digital conversion circuit and clock synchronization mechanism inside the edge control device align the signals of these four channels in time, and output a unified packaged joint control timing data matrix. The dimensions of the matrix include the four independent data channels mentioned above as well as the total number of discrete sampling points acquired within the sampling time window.

[0029] The joint control time series data is extracted and divided into macroscopic observation windows according to the system-set observation period. Within the macroscopic observation window, an asynchronous analysis process is started to perform a fast Fourier transform on the joint control time series data, outputting a frequency domain power spectrum sequence. The frequency domain power spectrum sequence is analyzed to locate the node with the maximum energy, and the peak amplitude value corresponding to the node with the maximum energy is extracted as the dominant frequency amplitude. The stable reference value set for the initial state of equipment operation is obtained, and the differential ratio is calculated by comparing the dominant frequency amplitude and the stable reference value. The differential ratio and the dominant frequency amplitude are aggregated to generate the operating condition frequency domain characteristics.

[0030] Specifically, the generation process of the operating condition frequency domain features is as follows: First, the joint control timing data is extracted, and the continuous data stream is truncated into fixed-length macroscopic observation windows according to a pre-set observation period. The selection of the observation period parameter needs to meet two conditions: on the one hand, the window length must be able to cover at least one complete basic vibration cycle at the lowest operating speed of the equipment to ensure that frequency characteristics are not lost; on the other hand, in order to meet the mathematical premise of the weak stationarity of the signal for the Fast Fourier Transform, the window must be shortened as much as possible to reduce the frequency drift caused by the acceleration and deceleration of industrial machinery. This application specifically limits the macroscopic observation window to 1.0 second. Within this extremely short physical time, the rotational inertia of the drawing machine spindle makes the evolution of the mechanical vibration frequency extremely gradual, which can be approximated as a quasi-stationary state. Combined with the endpoint smoothing processing of the Hanning window, the mathematical reliability of the transient dominant frequency amplitude extraction is effectively guaranteed.

[0031] Subsequently, single-channel data of buffer end displacement (in millimeters) was extracted from the joint control time series data, and a Hanning window function was used to perform smooth attenuation processing on the single-channel data. Specifically, within the macroscopic observation window, a Hanning window function vector of equal length was constructed for the single-channel data. A weight sequence within this window was generated by calculating the smooth rise and fall curve of the cosine function. The original sampling point values ​​in the joint control time series data were multiplied point-by-point with their corresponding weight values ​​at the time nodes in the Hanning window function vector. This multiplication mapping forces the data amplitudes at the initial and final nodes of the macroscopic observation window to be smoothly flattened to zero, thereby effectively eliminating truncation abrupt changes and suppressing spectral leakage during frequency domain conversion while preserving the signal's main frequency characteristics.

[0032] The processed data sequence is subjected to a time-decimation-based cardinality-even decomposition Fast Fourier Transform (FFT) to convert the time-domain fluctuation signal into a complex-domain frequency signal. Considering that continuous production may encounter transient conditions such as rapid acceleration, deceleration, or emergency shutdown, the spindle frequency may drift drastically within a short period, violating the weak stationarity prerequisite required for FFT calculations. Therefore, before performing the FFT, if the absolute value of the transient rate of change of the machining host linear speed exceeds a preset stationary acceleration threshold (set to 0.2 m / s² in this example), a frequency domain analysis bypass mechanism is triggered: the current macroscopic observation window skips the FFT and directly inherits and outputs the frequency domain features extracted from the previous stationary window. If the bypass mechanism is continuously activated for more than a set time (set to 3.0 seconds in this example, covering the normal rapid acceleration transient phase), the bypass mode is forcibly exited, and the target sampling period is directly locked to the base maximum sampling period for downgraded conservative sampling until the transient rate of change of the linear speed falls back to a safe range, at which point the FFT is resumed. This mechanism effectively avoids data distortion caused by severe spectrum leakage in non-stationary states.

[0033] Finally, the real and imaginary data of each frequency node of the complex domain signal are extracted, the squares of the real and imaginary data are calculated respectively, and the numerical summation operation between the two is performed. The resulting absolute energy value sequence is output as a frequency domain power spectrum sequence that records the signal intensity of each frequency node.

[0034] The spectrum analysis process parses the power spectrum sequence in the frequency domain, traverses the spectrum interval to locate the node with the maximum energy, and extracts the peak amplitude value corresponding to the node as the dominant frequency amplitude. The dominant frequency amplitude parameter characterizes the dominant mechanical vibration amplitude at the device buffer end within the current acquisition window.

[0035] The control process then obtains the stable reference value set for the initial state of equipment operation. The method for selecting this stable reference value parameter is as follows: during the standard load trial operation phase after the equipment is assembled, the average dominant frequency amplitude under continuous normal operation is recorded and used as a reference. For example, in this case, the stable reference value is set to 2.5 mm.

[0036] The comparison logic then compares the currently acquired main frequency amplitude with the stable reference value, and calculates the difference ratio through mathematical operations. The specific calculation process of the difference ratio is as follows: subtract the set stable reference value from the extracted main frequency amplitude value, calculate the absolute value of the difference, and finally divide the absolute value by the stable reference value. The difference ratio can be obtained by performing the division operation.

[0037] The calculated differential ratio and the main frequency amplitude are serialized and concatenated in a single column dimension. A one-dimensional feature vector containing two floating-point data nodes is constructed inside the storage register. This one-dimensional feature vector is then encapsulated as a whole variable to generate operating frequency domain features containing frequency domain state information.

[0038] By constructing an inverse mapping function between the differential ratio and the target sampling period and performing grid resampling, dynamic matching between the operating status and the acquisition density is achieved, which helps to provide data support with appropriate accuracy for continuous high-speed production processes while preventing computing power overload.

[0039] The system receives the operating condition frequency domain features, extracts the differential ratio contained in the operating condition frequency domain features, and constructs an inverse proportional mapping function between the differential ratio and the target sampling period; substitutes the differential ratio into the inverse proportional mapping function to calculate and output the target sampling period; extracts the original timestamp contained in the joint control time series data, performs grid resampling on the original timestamp according to the target sampling period, and generates aligned time nodes; performs linear interpolation calculation on the joint control time series data at the aligned time nodes, and outputs the reconstructed time series sequence; combines all reconstructed time series sequences and concatenates them to generate the time series feature tensor.

[0040] Specifically, the generation process of the temporal feature tensor is as follows: The differential ratio is extracted. The internal control logic constructs an inverse proportional mapping operation function between the differential ratio and the target sampling period. The specific calculation process of the target sampling period is as follows: the preset maximum sampling period of the device is extracted as the dividend, and the scaling factor is extracted. The scaling factor is multiplied by the differential ratio to obtain the product value. This product is added to the constant value of the basic value to form the divisor. Finally, the maximum sampling period of the basic value is divided by the divisor, and the output quotient is defined as the target sampling period.

[0041] The principles for determining the parameters mentioned above are as follows: The selection method for the basic maximum sampling period is based on the slowest response time constant of the mechanical structure and the minimum polling requirement of the network bus; in this embodiment, 50.0 milliseconds is selected. The selection method for the scaling factor is determined by engineers with reference to the sensitivity of amplitude-frequency changes in machine step response testing; in this embodiment, 1.2 is selected. The selection method for the basic numerical constant is to ensure that, under stable operating conditions, the calculated target sampling period is exactly equal to the basic maximum sampling period without scaling drift (i.e., maintaining a denominator of 1); this constant is set to 1.0.

[0042] Subsequently, the resampling algorithm extracts the original timestamps contained in the joint control timing data, performs grid resampling on the original timestamps according to the target sampling period step, and generates equidistant aligned time nodes.

[0043] In this embodiment, the delay fluctuation rate of data transmission in each channel is extracted. A first-order Kalman filter is used to estimate the phase deviation of the generated alignment time nodes, and a timestamp reverse shift correction is performed based on the estimated deviation value to generate the corrected alignment time nodes. Specifically, after generating equidistant initial alignment time nodes, the Ethernet arrival timestamps of the timing data of each channel's joint control at the bottom layer of the edge control device are extracted, and the transmission delay fluctuation rate sequence in milliseconds is generated by subtracting the transmission timestamps of the remote sensors from these timestamps. To eliminate clock jitter errors caused by bus network queuing, a first-order discrete Kalman filter is introduced. The initial alignment time nodes are set as observation variables, and the variance of the delay fluctuation rate is mapped to the measurement noise parameters of the state transition. The process noise covariance is set to 0.0001, and the measurement noise covariance is set to 0.001, which is obtained by statistical variance calculation and calibration of the measured transmission jitter sequence of the industrial Ethernet bus after multiple consecutive samplings. The prior estimation error covariance and dimensionless Kalman gain are initialized to 1. Through recursive calculation of single-step prediction and measurement update, state prediction is performed using the optimal time estimate from the previous moment and the prior error covariance, while measurement update is performed using the currently extracted delay volatility and its variance. This iteratively outputs the optimal phase time estimate for each channel's data. Subsequently, the optimal phase time estimate is subtracted from the initial alignment time node to obtain the phase drift deviation in the time dimension. A phase tolerance dead zone is set to 5% of the target sampling period. If the phase drift deviation exceeds this dead zone, the deviation value is multiplied by a dimensionless phase compensation coefficient of 0.85, and the product is shifted in the opposite direction and superimposed onto the initial alignment time node for fine-tuning. If it does not exceed the dead zone, the original value is retained, resulting in the corrected alignment time node. This process mitigates timing misalignment caused by random jitter in the industrial bus and improves the characteristic synchronization fidelity of multi-source heterogeneous data interpolation reconstruction.

[0044] At each alignment time node, linear interpolation is performed on the 4D joint control time series data. The specific interpolation logic is as follows: For any target alignment time node, locate the two adjacent original sampling points before and after its timestamp in the original time series data; calculate the time difference between the target alignment time node and the timestamp of the previous original sampling point, and simultaneously calculate the difference in data values ​​of the subsequent original sampling point and the previous original sampling point in the same channel, dividing this difference by the time interval between the two original sampling points to obtain the rate of change of values ​​in the current interval; multiply the time difference by the rate of change of values ​​to obtain the numerical compensation amount, and finally add the numerical compensation amount to the data value of the previous original sampling point. This process is repeated to calculate the interpolated data for each of the four channels, outputting the reconstructed time series sequence. All reconstructed time series sequences are combined and concatenated according to channel and time direction to generate a two-dimensional time series feature tensor containing a feature length equal to the total number of resampled data points and the corresponding feature widths of the four channels.

[0045] By constructing an inverse mapping function between the differential ratio and the target sampling period and performing grid resampling, dynamic matching between the operating status and the acquisition density is achieved, which helps to provide data support with appropriate accuracy for continuous high-speed production processes while preventing computing power overload.

[0046] Further, the physical limit boundary is obtained from the edge control device, and the difference between the physical boundary boundary and the displacement at the buffer end is calculated to generate the tension buffer margin. The temporal feature tensor and the tension buffer margin are input into the depth prediction model to generate adaptive weights. The temporal feature tensor is weighted according to the adaptive weights to output the spatiotemporal correlation features. This corresponds to step S2 above. The specific implementation process includes: Establish a communication connection with the edge control device, read the top and bottom coordinates of the guide rail from the internal register of the edge control device, and combine the top and bottom coordinates of the guide rail as the physical limit boundary; extract the buffer end displacement contained in the joint control timing data, compare the buffer end displacement with the top coordinate of the guide rail to calculate the absolute difference at the top, compare the buffer end displacement with the bottom coordinate of the guide rail to calculate the absolute difference at the bottom; compare the absolute difference at the top and the absolute difference at the bottom to select the tension buffer margin.

[0047] Specifically, the construction process for the tension buffer margin is as follows: The communication thread establishes a handshake connection with the edge device and reads the coordinates of the top and bottom ends of the guide rail. These two coordinate parameters are selected by measuring the physical geometric boundaries of the mechanical guide rail on site using an external laser rangefinder. For example, the bottom coordinate of the guide rail is calibrated and read as 0.0 mm, and the top coordinate of the guide rail is 600.0 mm.

[0048] The read bottom coordinates of the guide rail are set as the lower limit parameter for safe operation, and the top coordinates of the guide rail are set as the upper limit parameter for safe operation. Then, the upper limit parameter and the lower limit parameter are packaged and written into a continuously stored double-precision floating-point array memory block to construct the physical limit boundary containing the complete linear operating range.

[0049] In this embodiment, the ambient temperature change at the processing site and the continuous high-load operation time of the host are obtained. The dynamic thermal expansion deformation coefficient of the mechanical guide rail is calculated, and the deformation coefficient is used to perform thermodynamic product compensation on the physical limit boundary, reconstructing and generating the calibrated physical limit boundary. Specifically, after constructing the initial physical limit boundary, the underlying thermodynamic calibration thread is started to correct the mechanical geometric deformation. The current ambient temperature value and the continuous working time of this batch of equipment are read. The solidified 25-degree Celsius mechanical guide rail reference calibration temperature and the linear thermal expansion ratio coefficient of the guide rail material of 0.0000165 degrees Celsius are extracted. The temperature difference is obtained by subtracting the reference calibration temperature from the current ambient temperature value. The structural heat storage saturation constant with a value of 0.02 per hour is extracted, multiplied by the continuous working time to obtain the exponential factor, and the negative exponent of the exponential factor is calculated with the base of the natural logarithm as the base. The dimensionless heat storage attenuation is obtained by subtracting the negative exponent from the constant 1. Subsequently, the linear thermal expansion proportionality coefficient, temperature difference, and heat storage attenuation are multiplied together, and then a constant 1 is added to the product to construct a dimensionless nonlinear thermal expansion compensation factor. The coordinates of the top edge of the guide rail in the array are extracted and multiplied by this thermal expansion compensation factor to generate outwardly extending top calibration coordinates, while the bottom calibration coordinates remain unchanged from their original calibration values. At the same time, the top and bottom calibration coordinates are used to overwrite the values ​​in the original array memory block, completing the calibration update of the physical limit boundary. The above process compensates for the geometric boundary drift error of the mechanical actuator under long-term high-speed frictional heating, enhancing the safety of the limit approximation judgment.

[0050] Extract the current buffer end displacement value from the reconstructed control timing data and perform two difference calculations. The derivation process for the absolute difference at the top is as follows: read the coordinates of the top of the guide rail, subtract the currently obtained buffer end displacement from it, and take the absolute value of the difference; the derivation process for the absolute difference at the bottom is as follows: obtain the current buffer end displacement, subtract the coordinates of the bottom of the guide rail, and take the absolute value.

[0051] The comparison logic then compares the calculated top absolute difference and bottom absolute difference, and through the minimum value optimization logic, selects the smaller value of the two and outputs it as the final tension buffer margin.

[0052] By reading the guide rail coordinates and selecting the tension buffer margin, the mechanical safety boundary is transformed into a digital constraint index, which helps to predict displacement trends in advance and provides a basic physical space guarantee for preventing tension runaway and line breakage shutdowns in the target production line.

[0053] Extract the tension buffer margin and obtain the safety warning threshold built into the depth prediction model; compare the tension buffer margin and the safety warning threshold to calculate the margin deviation; construct an exponential growth formula between the margin deviation and the adaptive weight; when the tension buffer margin is less than the safety warning threshold, substitute the margin deviation into the exponential growth formula to perform calculation and generate a surge weight value as the adaptive weight; when the tension buffer margin is not less than the safety warning threshold, substitute the margin deviation into the exponential growth formula to perform calculation and generate a smooth weight value as the adaptive weight.

[0054] Specifically, the process of generating adaptive weights is as follows: The tension buffer margin of the output is obtained from the extracted stream, and the safety warning threshold built into the depth prediction network is also obtained. The safety warning threshold parameter is selected by initially estimating the product of the device's rated maximum linear speed and the actuator's maximum signal response delay time, and then determining it by adding a safety redundancy coefficient. In this embodiment, it is set to 150.0 mm.

[0055] The specific calculation process for the degree of deviation of the margin is as follows: extract the internally calibrated safety warning threshold, subtract the tension buffer margin from the safety warning threshold to obtain the numerical deviation, and then divide the numerical deviation by the safety warning threshold to complete the normalization and obtain the degree of deviation of the margin.

[0056] For different deviation states, the control network internally constructs an exponential growth computational system of deviation degree and adaptive weights, covering two piecewise evolution logics: The first logic corresponds to a physical state where the tension buffer margin is less than the safety warning threshold. Before substituting into the formula, an overflow prevention truncation is performed on the margin deviation: when the tension buffer margin is negative (i.e., the buffer end displacement exceeds the physical limit boundary), the calculated margin deviation is forcibly limited to the maximum upper limit of 1.0. The margin deviation is then substituted into the formula to perform exponential amplification to generate a surge weight value. The complete calculation process for the surge weight value is as follows: using the base of the natural logarithm as the base, the set gain coefficient is multiplied by the margin deviation to construct an exponent, and finally, the base weight is multiplied by the calculated result of this exponent. The base weight parameter is set to a conventional benchmark constant of 1.0; the gain coefficient is calibrated according to the weight amplification requirements corresponding to the equipment's emergency braking requirements; in this example, 5.0 is used. Substituting the margin deviation into the formula and performing this calculation yields the surge weight value, which is then used as the current adaptive weight.

[0057] The second logic corresponds to a physical state where the tension buffer margin is not less than the safety warning threshold. In this case, the corresponding margin deviation is negative. Substituting this margin deviation into the value generates a smoothing weight. The derivation process of the smoothing weight is as follows: extract the smoothing coefficient and multiply it by the absolute value of the margin deviation. Subtract this product from the unit constant to obtain the reduction ratio. Finally, extract the basic weight and multiply it by the reduction ratio to obtain the preliminary reduction value. To avoid the reduction ratio becoming negative when the buffer margin is extremely large, the result is compared with the maximum value of the set weight lower limit constant (set to 0.1 in this embodiment). The larger of the two values ​​is taken as the final smoothing weight value, ensuring that the weight value is always positive.

[0058] The smoothing coefficient is set based on the proportion of the jitter envelope of the regular data in the steady-state operating range; in this example, 0.3 is selected. The smoothing weight value is calculated and used as the adaptive weight at this time. The unit constant is a fixed constant of 1.0.

[0059] By constructing an exponential growth formula between the margin deviation and the adaptive weight to generate adaptive weight, a feature adjustment mechanism with physical sensing capability is established, which helps to strengthen safety guidance when the device approaches its limit state and avoid the risk of disconnection caused by complex coupling characteristics.

[0060] The system receives the adaptive weights and the temporal feature tensor, and constructs a spatiotemporal attention matrix within the deep prediction model. It maps the adaptive weights to the buffer displacement data channels corresponding to the spatiotemporal attention matrix, generating an updated attention matrix. It extracts the temporal feature tensor and performs a matrix transpose operation to generate a transposed feature tensor. It performs matrix multiplication between the transposed feature tensor and the updated attention matrix to calculate and output a spatial fusion tensor. It performs a sliding scan operation along the data path on the spatial fusion tensor to extract the dependency data between channels, and concatenates the dependency data to generate the spatiotemporal correlation features.

[0061] Specifically, the generation process of spatiotemporal correlation features is as follows: The attention layer receives adaptive weights and a temporal feature tensor (its initial dimension is [time step T, number of channels C=4]). To deeply extract the physical coupling relationship of multiple actuators at the same time, a feature interaction mechanism along the channel dimension is adopted inside the self-attention layer. Specifically, tensor transpose and linear projection are performed on the input features to generate a query matrix and a key matrix, both with a data dimension of [number of channels C, projected feature dimension d]. Subsequently, the query matrix and the transposed key matrix are subjected to a dot product operation and normalized by the Softmax function to construct a square matrix structure corresponding to 4 data channels (output dimension is [C, C]), which is the spatiotemporal attention matrix.

[0062] The received adaptive weight values ​​are mapped to the data channel index positions in the spatiotemporal attention matrix corresponding to the buffer end displacement. The original scores of the relevant position elements are replaced by direct multiplication to generate an updated attention matrix.

[0063] Since the original two-dimensional temporal feature tensor (with dimensions set to [number of channels C=4, time step T=total number of resampling points]) is stored in video memory with channel priority at consecutive addresses, in order to adapt to the underlying computational requirements of the subsequent self-attention mechanism for feature interaction based on time sequence, the tensor reshaping interface is called. By modifying the mapping pointer of the underlying memory step, the logical output dimension is flipped to a transposed feature tensor of [time step T, number of channels C] without moving the physical memory data.

[0064] Subsequently, a standard matrix multiplication operation is performed on the transposed feature tensor of dimension [time step T, number of channels C] and the updated attention matrix of dimension [number of channels C, number of channels C]. This allows the channel features at each time step to be linearly fused and reconstructed according to the updated attention weights, and the output spatial fusion tensor is calculated. During this process, the surge in weight values ​​in the updated attention matrix amplifies the displacement abrupt features at the corresponding time steps within the transposed feature tensor, while suppressing high-frequency background noise in other normal channels. The output tensor maintains a spatial fusion tensor with the shape [T, C] unchanged.

[0065] A sliding computation is performed on the spatial fusion tensor along the temporal direction of the data. Specifically, an dilated causal convolutional structure is used to perform scanning to extract inter-channel dependency data, such as those representing latency characteristics. Finally, the dependency data extracted from multiple feature receptive fields are concatenated along the channel dimension to generate the output spatiotemporal correlation features.

[0066] By mapping the adaptive weights to the spatiotemporal attention matrix and performing matrix multiplication, a deep integration of physical space constraints and temporal features is achieved, which helps to uncover hidden dependencies between data streams and makes the output feature matrix more adaptable to complex dynamic coupling conditions.

[0067] Furthermore, the spatiotemporal correlation features are input into the fully connected network to calculate the predicted speed value of the mainframe, which is then compared with the linear speed of the mainframe to generate a feedforward speed deviation. The tension buffer margin is extracted to calculate the buffer displacement deviation, and the feedforward speed deviation and buffer displacement deviation are fused to generate the speed change demand; this corresponds to step S3 above. The specific implementation process includes: Extract the spatiotemporal correlation features and flatten them into a tiled feature vector. Input the tiled feature vector into the hidden layer nodes contained in the fully connected network to perform nonlinear activation and output the host speed prediction value. Compare the host speed prediction value with the processing host linear speed to calculate the feedforward speed deviation, and extract the buffer end displacement to calculate the buffer displacement deviation. Obtain the system's preset displacement speed coefficient and multiply the buffer displacement deviation by the displacement speed coefficient to convert it into a feedback speed deviation. Superimpose the feedforward speed deviation and the feedback speed deviation to calculate the comprehensive speed error. Extract the preset comprehensive speed adjustment coefficient, multiply the comprehensive speed error by the comprehensive speed adjustment coefficient, perform a multiplication operation, and output the calculation result as the speed change requirement.

[0068] Specifically, the process of generating the demand for speed changes is as follows: First, spatiotemporal correlation features are extracted, and a flattening operation is performed to concatenate and compress their spatial-temporal and channel dimensions into a one-dimensional tiled feature vector. This tiled feature vector is input to the hidden layer nodes within the fully connected network. Through batch normalization and nonlinear activation mapping using a modified linear unit with leakage, the global dependency information is aggregated by a single linear output node at the end, calculating and outputting the host speed prediction value. During this process, the physical fluctuations of the winding torque at the take-up end, one of the four input channels (such as sudden torque increases indicating pre-tension tightening on the take-up side), are implicitly mapped to the host speed prediction value via spatiotemporal correlation features. When the network detects an abnormal upward trend in the winding torque, it will raise or suppress the host speed prediction value before a substantial drift occurs in the tension buffer displacement, thereby achieving advance compensation for the frequency converter through feedforward speed deviation.

[0069] The feedforward speed deviation is obtained by subtracting the predicted host speed output from the end of the fully connected network from the processing host linear speed acquired by the data stream.

[0070] The feedback compensation logic is based on spatial displacement: the buffer end displacement is extracted from the control timing data by reading the command, compared with a pre-set median physical reference point, and its actual offset distance is calculated to generate the buffer displacement deviation. The median physical reference point is selected as the arithmetic mean of the coordinates of the top and bottom ends of the physical limit boundary. Based on the aforementioned top coordinate of 600.0 mm and bottom coordinate of 0.0 mm, it is set to 300.0 mm in this embodiment.

[0071] The computing unit then obtains the pre-calibrated displacement velocity coefficient. This parameter is selected based on the mechanical geometric proportionality constant calculated from the gear reduction ratio of the drawing machine and the cross-sectional area of ​​the buffer cylinder. In this implementation environment, the base value is set to 0.02 meters per second per millimeter. Further considering the nonlinear response of linear velocity caused by the increase in the diameter of the take-up reel, this base value is multiplied by the dynamic proportionality factor of the reel diameter (i.e., the current take-up side reel diameter divided by the empty reel reference diameter) for amplification compensation, resulting in a dynamically updated displacement velocity coefficient.

[0072] Multiply the calculated buffer displacement deviation by the obtained displacement velocity coefficient to obtain the feedback velocity deviation.

[0073] In this computational example, the feedforward velocity deviation and the feedback velocity deviation are extracted, and numerical addition is performed to superimpose the physical quantities to obtain the comprehensive velocity error.

[0074] Extract the built-in comprehensive speed adjustment coefficient. This coefficient is selected based on empirical tuning parameters of the PID proportional element of the control system; in this example, 1.5 is chosen. Multiply the comprehensive speed error by the comprehensive speed adjustment coefficient, and use the result as the set speed change demand signal.

[0075] By multiplying the buffer displacement deviation by the displacement velocity coefficient to convert it into a feedback velocity deviation and superimposing it with the feedforward deviation, a computational logic that combines advance prediction and physical correction is constructed. This helps to smooth abrupt changes in control commands and maintain the dynamic balance of multi-equipment linkage in continuous high-speed production.

[0076] Further, based on the speed change demand, pay-off compensation commands and take-up compensation commands are generated; these commands are converted into control electrical signals and output to the edge control device to trigger the hardware frequency converter action; corresponding to step S4 above; see [link to relevant documentation]. Figure 3 The specific implementation process includes: The system receives the speed change demand and obtains the line speed frequency conversion coefficients for the pay-off and take-up sides built into the edge control device. It multiplies the speed change demand by the line speed frequency conversion coefficients for the pay-off side to calculate the basic pay-off compensation amount, obtains a preset frequency conversion coefficient, multiplies the tension buffer margin by the frequency conversion coefficients to calculate the pay-off amplitude limiting threshold, and performs amplitude limiting truncation operations on the basic pay-off compensation amount according to the pay-off amplitude limiting threshold to generate the pay-off compensation command. It then multiplies the speed change demand by the line speed frequency conversion coefficients for the take-up side to calculate the basic take-up compensation amount, multiplies the tension buffer margin by the frequency conversion coefficients to calculate the take-up amplitude limiting threshold, and performs amplitude limiting truncation operations on the basic take-up compensation amount according to the take-up amplitude limiting threshold to generate the take-up compensation command. Finally, it converts the pay-off compensation command and the take-up compensation command into control electrical signals and outputs them to the edge control device to trigger hardware frequency conversion.

[0077] Specifically, the processes for issuing and receiving compensation commands are as follows: To accommodate changes in receiving speed, the communication end obtains the linear speed frequency conversion coefficients for both the pay-off and take-up sides from the addressing device's storage area. These conversion coefficients are obtained by extracting constants such as the mechanical reduction ratio and the number of motor pole pairs, combining them with the current winding diameter of the pay-off and take-up shafts, and dynamically calculating them using a conversion logic from linear speed to operating frequency. The unit of measurement is Hertz per meter per second.

[0078] Since the calculated speed change requirement is represented as a linear velocity change in physical dimensions, while the remote frequency converter receives a frequency control signal, a low-level command conversion must be performed. The main control board extracts the linear velocity change value and retrieves the current wire-laying spool diameter data read via the communication bus to dynamically calculate the wire-laying side linear velocity frequency conversion factor. Multiplying the two directly converts the product into a Hertz increment form recognizable by the frequency converter. Encapsulating this product value into a digital format yields the wire-laying base compensation.

[0079] Obtain the preset frequency conversion coefficient. The selection method for this frequency conversion coefficient parameter is to deduce the safe conversion ratio corresponding to the frequency converter control frequency based on the maximum stall torque data of the motor allowed by the frequency converter and the ultimate yield strength data of the material in the motor calibration experiment. In this example, it is set to 0.3 Hz per millimeter.

[0080] The tension buffer margin data is extracted and multiplied by the frequency conversion coefficient to obtain the basic laying-out amplitude limiting threshold. To prevent the compensation action from being completely locked when the buffer end is extremely close to the physical limit (tension buffer margin approaches 0), a minimum safety minimum frequency parameter is introduced. This parameter is selected to meet the lower limit of the frequency at which the actuator overcomes the static friction of the system and maintains the minimum correction capability; in this embodiment, it is set to 0.5 Hz. The basic laying-out amplitude limiting threshold and the minimum safety minimum frequency are compared, and the maximum value between the two is extracted to obtain the laying-out amplitude limiting threshold. At the same time, the laying-out amplitude limiting threshold is multiplied by -1 to generate the corresponding negative amplitude limiting threshold, so as to construct an absolute value symmetrical upper and lower amplitude limiting interval.

[0081] In this embodiment, the torque feedback data of the wire-laying end motor from the most recent three communication cycles is extracted, and the torque change rate is calculated using forward differential. A nonlinear attenuation and contraction factor is constructed based on the torque change rate, and multiplied by the initially derived wire-laying limit threshold to generate an adaptive contraction wire-laying limit threshold. Specifically, after calculating the basic wire-laying limit threshold, a transient stress protection mechanism to prevent yielding deformation of the wire is introduced. The actual feedback values ​​of the winding torque of the wire-laying end motor at the current moment and the two consecutive discrete moments are extracted from the data cache queue. The first-order derivative of these three consecutive values ​​in the time dimension is calculated using a multi-point discrete differential algorithm to obtain the absolute value of the transient change rate of the current torque, in Newton-meters per second. A pre-calibrated upper limit of the safe tensile change rate of 18.5 Newton-meters per second is extracted. When the absolute value of the transient rate of change of torque is less than the safety upper limit, the dimensionless attenuation and contraction factor is set to 1.0. When the absolute value of the transient rate of change of torque exceeds the safety upper limit, the excess difference is divided by the safety upper limit to obtain the stress over-limit ratio. This stress over-limit ratio is multiplied by a stress penalty coefficient of 0.45, and the product of the penalty is subtracted from the constant 1.0 to obtain the calculated value. To prevent excessive restriction on the output command, this calculated value is compared with the constant 0.65, and the larger of the two is taken, thus outputting an attenuation and contraction factor with a minimum guarantee of 0.65. Finally, the initially obtained wire-feeding limit threshold is multiplied by this attenuation and contraction factor to perform a dimensionality reduction calculation, generating a new wire-feeding limit threshold with stress adaptive contraction capability. The above process effectively avoids wire plastic thinning or wire breakage caused by rigid adjustment by increasing the tolerance of the active contraction compensation action when the equipment encounters transient abnormal tensile stress.

[0082] The limiting logic then performs a limiting truncation operation on the basic compensation amount of the laying out according to this laying-out limiting threshold, comparing the absolute value of the basic compensation amount with the laying-out limiting threshold. If the basic compensation amount exceeds the calculated positive laying-out limiting threshold, the final output value is forcibly anchored to the positive limiting threshold. Conversely, if the basic compensation amount is lower than the negative laying-out limiting threshold, the output value is forcibly anchored to the negative limiting threshold. When the basic compensation amount is exactly within the safe range formed by the positive and negative limiting thresholds, the basic compensation amount of the laying out is directly retained as the output value. Finally, the final data after this nonlinear truncation process is output as the laying-out compensation instruction.

[0083] The take-up end data is processed using symmetrical operational logic: the calculation process for the basic take-up compensation amount involves extracting the required amount of the same speed change and multiplying it by the take-up side line speed frequency conversion factor; the calculation process for the take-up limiting threshold involves extracting the tension buffer margin and multiplying it by the frequency conversion factor set for the take-up side; to prevent the compensation action from being completely locked when the buffer end is extremely close to the physical limit, a minimum safety limiting mechanism is also introduced. The basic take-up limiting threshold is compared with the aforementioned preset minimum safety minimum frequency (0.5 Hz), and the maximum value between the two is extracted as the final take-up limiting threshold. Extract the calculated positive and negative take-off limiting threshold boundaries and compare them with the absolute value of the take-off basic compensation amount. If the take-off basic compensation amount exceeds the positive take-off limiting threshold, the output value is forcibly anchored to the positive limiting threshold. If it is lower than the negative take-off limiting threshold, the output value is forcibly anchored to the negative limiting threshold. When the take-off basic compensation amount is exactly within the safe range formed by the positive and negative take-off limiting thresholds, the original value is directly retained. Finally, the value after this nonlinear limiting and truncation operation is output as the take-off compensation command.

[0084] Finally, the underlying communication interface reads the wire release compensation command and the wire take-up compensation command, converts them into control electrical signals in fieldbus data packet format, and outputs them to the associated remote drive device through the physical interface to trigger the frequency converter or servo driver to adjust the output frequency of the transistor array and execute hardware frequency conversion.

[0085] By multiplying the tension buffer margin by the frequency conversion coefficient to calculate the threshold and performing a limiting cutoff operation, the physical barrier that compensates for the flexibility of the command output link helps to prevent abnormal calculation results from triggering violent frequency conversion actions, thus ensuring the safe and stable operation of the production line under complex working conditions.

[0086] Deep prediction networks need to be pre-trained on a computing platform to determine their network weights before they can be deployed in control equipment in industrial settings to perform forward inference. The following describes the hierarchical structure and training parameters included in this embodiment.

[0087] Deep networks consist of several processing layers that are connected in series and in parallel, with the specific layers as follows: Spatiotemporal Attention Layer: This layer receives a pre-processed two-dimensional input tensor (with a time step dimension of T and a feature channel dimension of C). Inside this layer, linear projection is used to map the input tensor to generate a query matrix and a key matrix. Simultaneously, the transposed input tensor is directly used as the value matrix of the standard self-attention mechanism, with dimensions [C, T]. After performing a dot product operation on the query and key matrices and normalizing them to generate an initial [C, C] attention matrix, externally input scalar weights are replaced on the specified diagonal nodes to generate an updated attention matrix. Subsequently, the updated attention matrix and the transposed value matrix are subjected to matrix multiplication with weighted summation, outputting a two-dimensional spatial fusion tensor (maintaining a shape of [T, C]) that completes the spatial channel feature reconstruction.

[0088] Channel convolutional layer: Receives the above spatial fusion tensor and performs sliding scan along the time dimension using a 4-layer cascaded dilated causal convolutional structure (dilation rates of 1, 2, 4, and 8, with the kernel size uniformly defined as 3, and a causal padding mechanism on the left to prevent future data leakage) to extract multi-scale temporal local dependency features.

[0089] Bidirectional gated recurrent dimensionality reduction layer (to resolve dimensionality mismatch): A tensor sequence with extracted local features is input into the BiGRU network. The number of internal unidirectional hidden nodes is set to 128, and after bidirectional concatenation, the total output feature dimension is expanded to 256. To align with the receiving dimension of the downstream network and reduce redundancy, a linear dimensionality reduction projection layer is added at the end to compress and map the 256-dimensional hidden state features into a fixed 64-dimensional tiled feature vector.

[0090] The fully connected network for prediction output is responsible for receiving the aforementioned 64-dimensional tiled feature vectors and performing the final nonlinear dimensionality reduction mapping. It is configured with hidden layers containing 64 nodes and then 32 nodes sequentially from front to back, ultimately converging to a single linear output node to output a single-value prediction result.

[0091] Before network deployment, the following iterative optimization steps are performed: Construct a sample dataset. Collect raw sensor data under multiple operating conditions and perform label separation. To ensure the model can generalize to various extreme production states, the dataset for multiple operating conditions covers the entire speed range of the drawing equipment from low startup speed to rated full load high speed, encompassing typical production operating states such as constant speed and stability, rapid acceleration, rapid deceleration, and emergency shutdown for disc replacement.

[0092] The actual host speed corresponding to the lead time step (in this embodiment, the lead step is set to the sum of the BiGRU block inference observation delay, inverter hardware response time, and communication delay, and is equal to the total physical feedforward delay time divided by the current target sampling period) is extracted as the training set label. In the full dataset, only the data distribution of the training set is used to calculate the mean and variance-normalized statistical parameters (mean and standard deviation), and these statistical parameters are then fixed and directly applied to the validation and test sets. The dataset partitioning ratio is selected as follows: to ensure the model learns sufficiently while possessing independent generalization evaluation capabilities, the training set accounts for 80% of the total data, the validation set 10%, and the test set 10%. The training set is only used for forward calculation of the loss and backpropagation to update the network parameter matrix; the validation set does not participate in the gradient update of the network parameters, but is only used after each training iteration to monitor the model's generalization trend and trigger an early stopping mechanism; the test set is treated as completely unknown data and is retained until the training is completely finished for final evaluation.

[0093] A mean squared error function is selected, and a regularization constraint term is added internally to construct a hybrid loss network to calculate the output error. The hybrid loss is calculated as follows: the average of the squares of the differences between the network's predicted speed and the true label is calculated, and then the product of the regularization penalty coefficient and the sum of the squares of all elements of the weight matrix is ​​added. The regularization penalty coefficient parameter is selected by referring to empirical values ​​to prevent model overfitting; in this example, 0.0001 is used. Based on this, a dynamic physical boundary penalty term is introduced: a quadratic penalty is applied when the predicted output exceeds the device's physical upper limit threshold; an exponential penalty is applied when the rate of change of the output between adjacent time steps exceeds the hardware's maximum physical response limit.

[0094] An adaptive moment estimation algorithm, adapted to the characteristics of non-stationary sequences, was selected as the optimizer to calculate the gradient update of the weights. The initial learning rate parameter was set to 0.001 to balance the requirement of rapid convergence to the extreme point in the early stage of gradient descent and the stability requirement of avoiding divergence. The batch size parameter was set according to the upper limit of the computing platform's memory capacity and the requirement of maintaining the smoothness of the gradient update for each parameter update, with 128 time window samples processed per batch.

[0095] The network performs forward computation to obtain the loss value and initiates the error backpropagation algorithm to update the matrix parameters. A gradient pruning mechanism is implemented at this stage. The gradient pruning threshold parameter is selected by observing the gradient vector norm distribution range during the oscillation of the loss function in the early pre-training phase, and a threshold upper limit of 1.5 is set. A validation set is used to monitor the loss. The early stopping tolerance parameter is selected by assessing the maximum number of evaluation cycles required to escape the local optimum trap of validation set error fluctuations, and is set to 10 iteration cycles. When the validation set error no longer decreases within the set 10 cycles, the early stopping mechanism is triggered, the training process is truncated, and the system automatically reverts and saves the optimal parameter matrix corresponding to the minimum validation set error.

[0096] Before deployment, a final generalization evaluation is performed: isolated test set data is input into the model loaded with the optimal parameter matrix to perform forward inference, and the root mean square error (RMSE) between the test set predictions and the true labels is calculated. A deployment tolerance threshold is set, which is chosen to meet the maximum tolerance of the field process for line speed fluctuations; in this example, it is set to less than 0.05 meters per second. When the RMSE calculated from the test set is less than this tolerance threshold, the model generalization is confirmed to be successful, and the deployment tool can then extract the parameter matrix and deploy it to the edge control device. A multi-threaded asynchronous concurrent control architecture is configured within the device operating system: the main control thread is dedicated to hardware bus polling and control signal distribution with the highest priority; the inference background thread runs the lightweight network model independently.

[0097] This invention provides a control optimization method for industrial equipment joint control timing data. By collecting the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end from the edge control device, joint control timing data is generated. Combined with the feedforward speed deviation and the buffer displacement deviation, speed change demand is generated. Finally, pay-off compensation command and take-up compensation command are generated based on the speed change demand and output to the edge control device to trigger hardware frequency conversion action. A globally collaborative control architecture is constructed, treating independent execution components as an organic whole with interconnected data. This helps to coordinate and adapt to the complex dynamic coupling characteristics between multiple devices from a global perspective, balance and smooth multi-end disturbances in continuous high-speed production, and provide overall operational protection to prevent tension loss of control and wire breakage shutdown in the target production line.

[0098] Example 2

[0099] This embodiment describes a control optimization method based on timing data for industrial equipment joint control, applied to a high-speed continuous drawing system for a certain type of alloy wire.

[0100] At the alloy wire processing site, edge-controlled equipment synchronously collects multi-dimensional continuous data such as the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end, generating joint control time-series data. According to the calibrated macroscopic observation period, an asynchronous analysis process is initiated to perform a Fast Fourier Transform on the data within the window, outputting a frequency domain power spectrum sequence and extracting the dominant frequency amplitude. By comparing the displacement reference value of the drawing machine in its initial stable phase with the current dominant frequency amplitude, the difference ratio is calculated and aggregated to generate the operating condition frequency domain features. The difference ratio in this feature is extracted, and the basic maximum sampling period and scaling factor calibrated according to the equipment's frequency response characteristics are substituted into the inverse proportional mapping function to calculate and output the current adaptive target sampling period. According to this dynamic step size, grid resampling and linear interpolation calculations are performed on the original timestamps, and the multi-channel sequences are combined and spliced ​​to generate a time-series feature tensor.

[0101] Subsequently, the control unit reads the top and bottom coordinates of the device from the internal registers as the physical limit boundaries. The displacement of the buffer end is extracted, and the absolute difference between this displacement and the coordinates of both ends is calculated. The end with the smaller offset is selected as the current tension buffer margin. When it is determined that the current tension buffer margin is less than the preset safety warning threshold, indicating a high-risk approach state, the control logic compares and calculates the margin deviation and substitutes it into the exponential growth formula. Under the exponential amplification effect of the base weights and the set gain coefficient, a surge weight value is generated as an adaptive weight. A spatiotemporal attention matrix is ​​constructed within the deep prediction model. This adaptive weight is mapped to the buffer end displacement data channel to generate an updated attention matrix, which is then multiplied with the transposed temporal feature tensor. Channel dependencies are then extracted through a sliding scan operation, outputting the spatiotemporal correlation features.

[0102] Next, the spatiotemporal correlation features are flattened into tiled feature vectors, input into the hidden nodes of the fully connected network to perform nonlinear activation, and the predicted host speed is calculated and output. The predicted value is compared with the actual linear speed of the processing host to generate a feedforward speed deviation; at the same time, the buffer end displacement is extracted, multiplied by a preset displacement-speed coefficient, and converted into a feedback speed deviation. The feedforward and feedback deviations are superimposed to calculate the comprehensive speed error, which is then multiplied by a comprehensive speed adjustment coefficient, and the result is output as the speed change requirement.

[0103] Finally, to determine the required change in receiving speed, the frequency conversion coefficients for the pay-off and take-up sides, calculated within the edge control device, are obtained. These coefficients are then multiplied to derive the pay-off and take-up base compensation amounts. Combining this with the obtained frequency conversion coefficient, the tension buffer margin is multiplied by this coefficient to obtain the dynamic limiting threshold. Strictly adhering to the limiting threshold boundary, a limiting truncation calculation is performed on the pay-off and take-up base compensation amounts, generating the final pay-off and take-up compensation commands. The communication interface converts these commands into control electrical signals and outputs them to the edge control device, triggering the frequency converter to execute hardware frequency conversion.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control optimization method based on timing data of industrial equipment joint control, characterized in that, include: The system collects data from the edge control equipment, including the motor speed at the pay-off end, the linear speed of the processing host, the winding torque at the take-up end, and the displacement at the buffer end, to generate joint control timing data. This data is then divided into macroscopic observation windows according to the system's set observation period. Within these macroscopic observation windows, an asynchronous analysis process is initiated to perform a fast Fourier transform on the joint control timing data, outputting a frequency domain power spectrum sequence. The frequency domain power spectrum sequence is analyzed to locate the node with the maximum energy, and the peak amplitude value corresponding to the node with the maximum energy is extracted as the dominant frequency amplitude. Obtain the stable reference value set at the initial operating state of the equipment, compare the main frequency amplitude with the stable reference value to calculate the differential ratio, and aggregate the differential ratio and the main frequency amplitude to generate the operating condition frequency domain characteristics. Based on the frequency domain characteristics of the operating conditions, a target sampling period is generated through inverse proportional mapping. Resampling is then performed according to the target sampling period to generate a time-series feature tensor. The difference between the calculated physical limit boundary and the displacement of the buffer end is obtained from the edge control device to generate the tension buffer margin; the deviation between the safety warning threshold and the calculated tension buffer margin is obtained, and an adaptive weight is generated based on the comparison result between the tension buffer margin and the safety warning threshold. The temporal feature tensor is weighted according to adaptive weights to output spatiotemporal correlation features; The specific process of generating adaptive weights includes: extracting the tension buffer margin and obtaining the safety warning threshold; comparing the tension buffer margin and the safety warning threshold to calculate the margin deviation; constructing an exponential growth formula between the margin deviation and the adaptive weights; when the tension buffer margin is less than the safety warning threshold, substituting the margin deviation into the exponential growth formula to perform calculations, generating a surge weight value as the adaptive weights; when the tension buffer margin is not less than the safety warning threshold, extracting a smoothing coefficient and multiplying it by the absolute value of the margin deviation, subtracting the product from a unit constant to obtain a reduction ratio, extracting the base weight and multiplying it by the reduction ratio to obtain a preliminary reduction value; comparing the preliminary reduction value with the set weight lower limit constant and taking the larger value to generate a smoothing weight value as the adaptive weights; Input the spatiotemporal correlation features into the speed prediction value of the fully connected network computing host, compare it with the linear speed of the processing host to generate the feedforward speed deviation, extract the buffer end displacement to calculate the buffer displacement deviation, and fuse the feedforward speed deviation and the buffer displacement deviation to generate the speed change demand. Based on the speed change demand, pay-off compensation commands and take-up compensation commands are generated; these commands are then converted into control electrical signals to trigger the hardware frequency conversion action.

2. The control optimization method for industrial equipment joint control timing data according to claim 1, characterized in that, The specific generation process of the time-series feature tensor includes: receiving the operating condition frequency domain features, extracting the differential ratio contained in the operating condition frequency domain features, and constructing an inverse proportional mapping function between the differential ratio and the target sampling period. Specifically, this involves: extracting a preset basic maximum sampling period as the divisor, simultaneously extracting a scaling factor, multiplying the scaling factor and the differential ratio to obtain a product value, adding the product value to a basic constant value to form a divisor, and dividing the basic maximum sampling period by the divisor; substituting the differential ratio into the inverse proportional mapping function to calculate and output the target sampling period; extracting the original timestamp contained in the joint control time-series data, performing grid resampling on the original timestamps according to the target sampling period to generate aligned time nodes; performing linear interpolation calculation on the joint control time-series data at the aligned time nodes to output a reconstructed time-series sequence; and combining all reconstructed time-series sequences to generate the time-series feature tensor.

3. The control optimization method for industrial equipment joint control timing data according to claim 1, characterized in that, The specific construction process of the tension buffer margin includes: establishing a communication connection with the edge control device; reading the top and bottom coordinates of the guide rail from the internal register of the edge control device; combining the top and bottom coordinates of the guide rail as the physical limit boundary; extracting the buffer end displacement contained in the joint control timing data; comparing the buffer end displacement with the top coordinate of the guide rail to calculate the absolute difference at the top; comparing the buffer end displacement with the bottom coordinate of the guide rail to calculate the absolute difference at the bottom; and selecting the tension buffer margin by comparing the absolute difference at the top and the absolute difference at the bottom.

4. The control optimization method for industrial equipment joint control timing data according to claim 1, characterized in that, The specific generation process of the spatiotemporal correlation features includes: receiving the adaptive weights and the temporal feature tensor, and constructing a spatiotemporal attention matrix within the deep prediction model; mapping the adaptive weights to the buffer end displacement data channels corresponding to the spatiotemporal attention matrix to generate an updated attention matrix; extracting the temporal feature tensor and performing a matrix transpose operation to generate a transposed feature tensor; performing matrix multiplication between the transposed feature tensor and the updated attention matrix to calculate and output a spatial fusion tensor; performing a sliding scan operation along the data path on the spatial fusion tensor to extract the dependency data between channels, and concatenating the dependency data to generate the spatiotemporal correlation features.

5. The control optimization method for industrial equipment joint control timing data according to claim 1, characterized in that, The specific generation process of the speed change demand includes: extracting the spatiotemporal correlation features, flattening the spatiotemporal correlation features into a tiled feature vector; inputting the tiled feature vector into the hidden layer nodes contained in the fully connected network to perform nonlinear activation, and outputting the host speed prediction value; comparing the host speed prediction value with the processing host linear speed to calculate the feedforward speed deviation, and extracting the buffer end displacement to calculate the buffer displacement deviation; obtaining the system preset displacement speed coefficient, multiplying the buffer displacement deviation by the displacement speed coefficient to convert it into a feedback speed deviation; superimposing the feedforward speed deviation and the feedback speed deviation to calculate the comprehensive speed error; extracting the preset comprehensive speed adjustment coefficient, multiplying the comprehensive speed error by the comprehensive speed adjustment coefficient to perform a multiplication operation, and outputting the calculation result as the speed change demand.

6. The control optimization method for industrial equipment joint control timing data according to claim 1, characterized in that, The specific generation process of the wire release compensation command and the wire take-up compensation command includes: receiving the speed change demand, obtaining the wire release side line speed frequency conversion coefficient and the wire take-up side line speed frequency conversion coefficient built into the edge control device; multiplying the speed change demand by the wire release side line speed frequency conversion coefficient to calculate the wire release basic compensation amount, obtaining a preset frequency conversion coefficient, multiplying the tension buffer margin by the frequency conversion coefficient to calculate the wire release amplitude limiting threshold, performing amplitude limiting truncation operation on the wire release basic compensation amount according to the wire release amplitude limiting threshold to generate the wire release compensation command; multiplying the speed change demand by the wire take-up side line speed frequency conversion coefficient to calculate the wire take-up basic compensation amount, multiplying the tension buffer margin by the frequency conversion coefficient to calculate the wire take-up amplitude limiting threshold, performing amplitude limiting truncation operation on the wire take-up basic compensation amount according to the wire take-up amplitude limiting threshold to generate the wire take-up compensation command; converting the wire release compensation command and the wire take-up compensation command into the control electrical signal, and outputting it to the edge control device to trigger hardware frequency conversion action.

Images