Edge computing-based adaptive control system for winding tension of FRP pipe fittings

Abstract

This invention relates to the field of fiberglass pipe winding control technology, specifically a fiberglass pipe winding tension adaptive control system based on edge computing. The system includes: a status monitoring module that collects tension time series, underlying hardware operating status data, and timestamps to generate a multi-dimensional observation matrix; a status assessment module that calculates tension error series, control cycle jitter entropy, and computational thermal fatigue index to generate comprehensive health data; a boundary prediction module that calculates the failure probability by combining computational risk thresholds and physical failure boundary conditions; and a mode switching module that switches between a deep feature fitting compensation mode and a basic robust control suboptimal mode based on risk assessment results, issuing tension control commands to adjust the unwinding tension, thereby achieving coordinated constraints on control timeliness, hardware thermal load, and physical failure boundaries.

Description

Technical Field

[0001] This invention relates to the field of fiberglass pipe winding molding control technology, specifically to a fiberglass pipe winding tension adaptive control system based on edge computing. Background Technology

[0002] As a crucial step in composite material manufacturing, the winding production of FRP pipe fittings requires strict error constraints on the stability of unwinding tension in high-performance applications such as deep-sea high-pressure transportation. To ensure the quality of pipe fittings and structural reliability, it is usually necessary to adjust the tension in real time under high-speed continuous production conditions. This is achieved by using edge computing nodes to collect tension data, calculate control commands, and drive actuators locally, thereby improving control response speed and real-time command issuance. However, when controlling the tension under high-speed winding conditions, it is necessary not only to pay attention to the tension deviation itself, but also to consider the computing power bearing state of the edge computing nodes under continuous high load operation. Most existing tension control methods take the minimization of tension error as the main goal, and usually use fixed parameter control or complex intelligent compensation control to directly output adjustment commands. However, they lack unified constraints on the risks of data arrival time jitter, increased processor load, increased core temperature, and the resulting control delay. It is difficult to jointly evaluate the duration of tension exceeding the limit and the physical failure boundary. When high-frequency nonlinear disturbances occur continuously, although complex compensation algorithms can improve local control accuracy, they can easily induce an increase in the computational load of edge nodes, causing control cycle instability or even delay in command issuance, which in turn leads to the tension exceeding the safe range for a long time, affecting the winding quality and production stability of FRP pipe fittings. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive control system for the winding tension of FRP pipe fittings based on edge computing. This system aims to solve the problems in the prior art where it is difficult to jointly evaluate the duration of tension exceeding the limit and the physical failure boundary, and where complex compensation algorithms easily induce an increase in the computational load of edge nodes, thereby causing instability in the control cycle and delays in command issuance. An edge computing-based adaptive tension control system for fiberglass pipe winding is described. The system is deployed on an edge computing node and is communicatively connected to an external tension sensor and an external winding actuator. The system includes: The status monitoring module collects the tension time series data output by the external tension sensor and the underlying hardware operation status data of the edge computing node, records the data arrival timestamp, generates a multi-dimensional observation matrix, and transmits it to the status evaluation module. The status assessment module acquires the multi-dimensional observation matrix, calculates the tension error sequence between the current tension amplitude and the preset target tension benchmark value, analyzes the data arrival timestamp to calculate the control beat jitter entropy, and combines the underlying hardware operating status data to quantify the computing power thermal fatigue index, generate comprehensive health data, and transmit it to the boundary prediction module. The boundary prediction module calls the comprehensive health data, compares the computing power thermal fatigue index with the preset computing power danger threshold, associates and maps the tension error sequence with the preset physical failure boundary conditions, calculates the failure probability of the system within the preset time window, generates risk assessment results and transmits them to the mode switching module. The mode switching module, based on the risk assessment results, performs a control degradation judgment: when the computing power thermal fatigue index is lower than or equal to the computing power danger threshold, it maintains the deep feature fitting compensation mode to output the first tension control command; when the computing power thermal fatigue index is higher than the computing power danger threshold, it switches to the basic robust control suboptimal mode to output the second tension control command; and sends the first tension control command or the second tension control command to the external winding actuator through the communication interface to adjust the unwinding tension.

[0004] Preferably, the multidimensional observation matrix includes tension amplitude, CPU load rate of the edge computing node, memory occupancy rate of the edge computing node, and core temperature of the edge computing node; the comprehensive health data includes tension error sequence, clock jitter entropy value, and computing power thermal fatigue index; and the risk assessment results include predicted over-limit duration and failure probability.

[0005] Preferably, the state assessment module includes: The error calculation submodule calls the tension time series data in the multidimensional observation matrix to calculate the absolute value of the deviation between the current tension amplitude and the preset target tension benchmark value, and generates the tension error value in the tension error sequence. The beat entropy quantization submodule extracts the data arrival timestamps from the multidimensional observation matrix, calculates the time interval between adjacent timestamps, statistically analyzes the probability distribution of the time intervals, calculates the information entropy based on the probability distribution, and generates the control beat jitter entropy. The fatigue fusion submodule extracts the underlying hardware operating status data from the multi-dimensional observation matrix. The underlying hardware operating status data includes the CPU load rate of the edge computing node, the memory occupancy rate of the edge computing node, and the core temperature of the edge computing node. The CPU load rate and core temperature of the edge computing node are normalized, and the normalized load rate and temperature are weighted and summed with the control beat jitter entropy according to a preset static weight to generate the computing power thermal fatigue index.

[0006] Preferably, the boundary prediction module includes: The threshold comparison submodule calls the computing power thermal fatigue index in the comprehensive health data, calculates the rate of change of the computing power thermal fatigue index as it approaches the computing power danger threshold, and generates the computing power overload approach rate. The delay risk mapping submodule determines the duration for which the tension error value exceeds the preset safe tension range based on the tension error sequence in the comprehensive health data, and generates the current over-limit duration. The probability calculation submodule combines the computing power overload approach rate and the current over-limit duration, inputs a preset risk assessment model, which is an assessment model that maps the computing power overload approach rate and the current over-limit duration to a preset failure probability. Through this model, it outputs the failure probability when the system control command issuance delay exceeds the physical failure boundary condition, and generates a risk assessment result.

[0007] Preferably, the mode switching module includes: The modal arbitration submodule reads the failure probability and the computing power thermal fatigue index from the risk assessment results, performs a conditional comparison between the failure probability and the preset failure probability warning threshold, and outputs a modal switching trigger signal. The deep compensation control submodule, in response to a trigger signal that the computing power thermal fatigue index is lower than or equal to the computing power danger threshold, activates a preset edge deep reinforcement learning network, extracts high-frequency nonlinear distortion signals from the tension time series data of the multidimensional observation matrix and performs feedforward fitting compensation, and generates a first tension control command. The degradation robust control submodule, in response to the trigger signal that the computing power thermal fatigue index is higher than the computing power danger threshold, puts the edge deep reinforcement learning network into hibernation, activates the preset classical robust controller, shields the high-frequency nonlinear distortion signal, and generates a second tension control command based on the deviation between the current tension amplitude and the preset macroscopic tension tolerance range.

[0008] Preferably, when generating the second tension control command, the degradation robust control submodule performs the following operations: obtains the current tension amplitude and determines whether the current tension amplitude is within the preset macroscopic tension tolerance range; When the current tension amplitude is within the macroscopic tension tolerance range, the current tension fluctuation is determined to be an allowable physical tolerance fluctuation, and a second tension control command to maintain the current control state is output, without executing the calculation to suppress the fluctuation; when the current tension amplitude exceeds the macroscopic tension tolerance range, a proportional-integral-derivative control algorithm is used to output a second tension control command to return to the target tension reference value.

[0009] Preferably, the system further includes: a computing power recovery adaptive module, which continuously monitors the comprehensive health data after the mode switching module switches to the basic robust control suboptimal mode, and compares the real-time updated computing power thermal fatigue index with the preset computing power recovery safety threshold; When the real-time updated computing power thermal fatigue index is higher than the computing power recovery safety threshold, the basic robust control suboptimal mode is maintained; when the real-time updated computing power thermal fatigue index is lower than or equal to the computing power recovery safety threshold, a mode reset signal is generated, triggering the mode switching module to smoothly transition to the deep feature fitting compensation mode.

[0010] Preferably, the computing power recovery safety threshold is determined by obtaining the critical temperature parameter of the underlying chip of the edge computing node that triggers frequency reduction protection, and by performing bias calculation in combination with a preset heat dissipation time constant. The computing power recovery safety threshold is strictly less than the computing power danger threshold, so as to form a hysteresis loop for mode switching.

[0011] Preferably, the physical failure boundary condition is defined as a time window in which the tension overshoot exceeds a preset structural limit threshold and lasts for more than fifty milliseconds, and the time window serves as an absolute time constraint for the system control response.

[0012] Preferably, the high-frequency nonlinear distortion signal is an excitation signal with a frequency greater than a preset frequency threshold and an amplitude change rate that does not meet the linear constraint. The excitation signal can easily induce the edge computing node to increase its computing load.

[0013] The present invention has the following beneficial effects: By uniformly collecting tension time series data, data arrival timestamps, and underlying hardware operating status data such as CPU load rate, memory usage rate, and core temperature on edge computing nodes, a multi-dimensional observation matrix is ​​constructed. Furthermore, comprehensive health data such as tension error sequence, clock jitter entropy value, and computing power thermal fatigue index are generated, so that tension control is no longer only focused on minimizing errors, but can simultaneously characterize the stability of control clock and the health status of edge computing power. In the risk prediction stage, the computing power thermal fatigue index is compared with the computing power danger threshold, and the tension error sequence is associated with the physical failure boundary conditions. By combining the computing power overload approach rate, the current over-limit duration, the predicted over-limit duration and the failure probability, the risk of the control command issuance delay exceeding the structural tolerance window can be assessed in advance, thereby solving the problem that existing technologies cannot jointly assess the tension over-limit duration and the physical failure boundary. During the control execution phase, an arbitration switch is made between the deep feature fitting compensation mode and the basic robust control suboptimal mode based on the risk assessment results. When the hardware is running normally, deep compensation is used to perform fine-grained fitting compensation for high-frequency nonlinear distortion signals. When the computing power and thermal fatigue increase, the deep network is put into hibernation and switched to degraded robust control that only adjusts around the macroscopic tension tolerance range, so as to avoid the complex compensation algorithm from continuously inducing the load of edge nodes to increase, the cycle time instability and control lag. Meanwhile, through data missing fault tolerance processing, outlier smoothing, integral limiting, modal freezing, smooth reset, and a hysteresis loop mechanism consisting of dangerous threshold and recovery safety threshold, the system can still operate stably under abnormal sampling, heat accumulation, and repeated fluctuation conditions. The tension overshoot exceeding the structural limit threshold and lasting for more than 50 milliseconds is used as the unified failure judgment boundary, thus forming a closed-loop tension adaptive control mechanism that takes into account control timeliness, hardware thermal load, physical safety boundary, and production continuity. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be conventionally introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a module structure diagram of the fiberglass pipe winding tension adaptive control system based on edge computing of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 An edge computing-based adaptive tension control system for fiberglass pipe winding is deployed on an edge computing node and is communicatively connected to an external tension sensor and an external winding actuator. The system includes: The status monitoring module collects tension time series data output by external tension sensors and underlying hardware operation status data of edge computing nodes, records the arrival timestamps of the data, generates a multi-dimensional observation matrix, and transmits it to the status assessment module. The status assessment module acquires a multi-dimensional observation matrix, calculates the tension error sequence, analyzes the data arrival timestamps to calculate the control beat jitter entropy, and combines the underlying hardware operating status data to quantify the computing power thermal fatigue index, generate comprehensive health data, and transmit it to the boundary prediction module. The boundary prediction module calls the comprehensive health data, compares the computing power thermal fatigue index with the preset computing power danger threshold, associates and maps the tension error sequence with the preset physical failure boundary conditions, calculates the failure probability of the system within the preset time window, generates risk assessment results and transmits them to the mode switching module. The mode switching module, based on the risk assessment results, performs a control degradation judgment: when the computing power thermal fatigue index is lower than or equal to the computing power danger threshold, it maintains the deep feature fitting compensation mode to output the first tension control command; when the computing power thermal fatigue index is higher than the computing power danger threshold, it switches to the basic robust control suboptimal mode to output the second tension control command; and sends the first tension control command or the second tension control command to the external winding actuator through the communication interface to adjust the unwinding tension.

[0017] This embodiment provides an edge computing-based adaptive control mechanism for the winding tension of fiberglass pipe fittings. Specifically, the system is deployed in the edge control box of the fiberglass pipe fitting production line for deep-sea high-pressure transportation. The edge control box is installed near the main shaft of the winding machine and is directly connected to the tension sensor, the unwinding motor driver, and the tension compensation actuator. The production line operates continuously at its maximum speed, aiming to maintain a preset high-speed winding speed while keeping the fiber unwinding tension under control, thus avoiding the formation of microcracks inside the pipe wall due to prolonged tension overshoot. Specifically, the status monitoring module collects two types of data according to the control cycle: one is the tension time series data output by the external tension sensor, and the other is the underlying hardware operating status data of the edge computing node, such as the CPU load rate, memory usage rate and core temperature. Record the arrival timestamp for each sampled data point; to facilitate subsequent calculations, the observations from four consecutive control cycles can be organized into a micro-matrix; at four adjacent time points... The detected tension amplitudes were 98 N, 103 N, 109 N, and 101 N, respectively; the CPU load rates were 82%, 89%, 94%, and 91%, respectively; and the core temperatures were 71°C, 74°C, 79°C, and 78°C, respectively, with corresponding timestamps of 0 ms, 2 ms, 5 ms, and 7 ms. The status monitoring module then combines these data by row or column into a multi-dimensional observation matrix and transmits it to the subsequent stage. After receiving the matrix, the state assessment module first calculates the tension error sequence based on the target tension reference value. Assuming the target tension reference value is 100 N, the absolute values ​​of the errors at the four times can be 2, 3, 9, and 1, forming an error sequence. The control beat jitter entropy is calculated using the timestamp. Taking the aforementioned timestamp as an example, the adjacent time intervals are 2 milliseconds, 3 milliseconds, and 2 milliseconds, respectively. If, within a longer time window, the proportions of intervals with a time interval of 2 milliseconds are 0.7, 3 milliseconds are 0.2, and 4 milliseconds are 0.1, then an entropy value reflecting the degree of clock cycle dispersion can be calculated. A higher entropy value indicates greater instability in the control clock cycle. Then, the CPU load rate, core temperature, and entropy value are fused to obtain the computing power thermal fatigue index. After normalization, the load rate corresponds to 0.94, the temperature to 0.79, and the entropy value to 0.42. If the weights are 0.4, 0.4, and 0.2, the fused result is... This can be used as the computing power thermal fatigue index at the current moment; the module finally generates comprehensive health data and transfers it to the boundary prediction module. The boundary prediction module performs dual predictions based on comprehensive health data; it compares the computing power thermal fatigue index with a preset computing power danger threshold; the danger threshold is set to 0.75, and the aforementioned 0.776 is already higher than the threshold, indicating that the edge node is approaching the risk of frequency reduction; it correlates the tension error sequence with the physical failure boundary conditions; if the system stipulates that the tension exceeds the structural limit threshold and lasts for more than 50 milliseconds, it is considered a failure, then the boundary prediction module will continue to estimate the probability of this failure state occurring within a preset time window if the current error continues; within the next 20 millisecond prediction window, based on the current error growth trend, the clock jitter, and the computing power status, the failure probability can be calculated to be 0.63; this result is output as the risk assessment result; The mode switching module makes a control degradation judgment based on the above risk assessment results. When the computing power thermal fatigue index does not exceed the danger threshold, the system maintains the deep feature fitting compensation mode, which performs fine-grained compensation for high-frequency nonlinear disturbances and outputs the first tension control command. It can issue a 2% deceleration and pre-compensation command to the unwinding motor. When the computing power thermal fatigue index exceeds the danger threshold, the system immediately switches to the basic robust control suboptimal mode and outputs the second tension control command. It only executes a simplified command of 1% deceleration and holds it for 5 milliseconds based on the current tension deviation to reduce the computational burden. Both types of commands are sent to the winding actuator through the communication interface to adjust the unwinding tension. Regarding anomaly handling, if the status monitoring module detects missing tension sensor data within a certain control cycle, but complete hardware status data, the effective tension value of the previous cycle can be used for that cycle with a low confidence mark attached, and subsequent modules will adjust the weight of that cycle accordingly. If hardware status data is missing but tension data is complete, the computational thermal fatigue index will be temporarily compensated based on the average of the most recent consecutive effective windows to avoid erroneous switching triggered by a single point of missing measurement. If both types of data are missing simultaneously and continue to exceed the preset tolerance length, the mode switching module will directly enter the basic robust control suboptimal mode and freeze the deep feature fitting compensation mode until the preset number of effective cycles are continuously restored before unfreezing. For example, in the third hour after the deep-sea grade FRP pipe production line entered the night shift continuous production, the resin temperature fluctuated due to environmental changes, and the fibers generated high-frequency frictional vibration when passing through the guide roller; at this time, the tension sensor reported a large number of high-frequency micro-amplitude distortion signals, and the edge control box increased the computing intensity for a short time in order to maintain tension accuracy. The status monitoring module continuously collected information on tension value, CPU load rate and core temperature rise; the status assessment module found that the beat jitter gradually expanded from a stable 2 milliseconds to a mixed distribution of 2 milliseconds and 5 milliseconds, and the computing power thermal fatigue index rapidly exceeded the dangerous threshold. The boundary prediction module further determined that if complex compensation is still insisted upon, the delay in issuing control commands will increase significantly within the next 20 milliseconds, and the probability of failure will rise. Therefore, the mode switching module abandoned the fine-grained compensation of the deep feature fitting compensation mode and switched to the basic robust control suboptimal mode to keep the spindle tension within the safe range and ensure that the current pipe fitting is within the physical structural tolerance boundary. The purpose of this step is to integrate tension control itself with the health status of edge computing power into the same closed loop for joint management, rather than focusing solely on minimizing tension error. This will enable coordinated constraints on control timeliness, hardware thermal load, and physical failure boundaries under high-speed winding conditions.

[0018] As a preferred embodiment of the present invention, the multidimensional observation matrix includes tension amplitude, node central processor load rate, node memory occupancy rate and node core temperature, the comprehensive health data includes tension error sequence, clock jitter entropy value and computing power thermal fatigue index, and the risk assessment results include the predicted value of over-limit duration and failure probability.

[0019] This embodiment provides a specific organization mechanism for a multidimensional observation matrix and comprehensive health data. Specifically, in the aforementioned continuous production scenario, if only tension amplitude and CPU load rate are used as inputs, although the control effect can be initially judged, when high temperature and memory cache congestion occur simultaneously, simple bivariate monitoring is prone to missing early signs of underlying computing power degradation. Therefore, four types of observations are introduced here: tension amplitude, CPU load rate, memory occupancy rate and core temperature, to form a unified multidimensional observation matrix. Specifically, a matrix can be constructed by using rows to represent time and columns to represent features. Assuming that within five consecutive sampling periods, the tension amplitude is 100, 104, 108, 107, and 102 Newtons respectively, the CPU load rate is 78%, 84%, 92%, 96%, and 90% respectively, the memory usage rate is 61%, 66%, 72%, 83%, and 80% respectively, and the core temperature is 68, 71, 76, 81, and 79 degrees Celsius respectively, a 5×4 observation matrix is ​​formed. In the state assessment stage, the tension error sequence can be obtained from the first column, the beat jitter entropy value can be obtained from the sampling time distribution, and the computational thermal fatigue index can be obtained from the last three columns combined with the beat jitter entropy value. Thus, the comprehensive health data output is not a single score, but a set of data that can be called by subsequent modules. The risk assessment results include at least two quantities: the predicted duration of exceeding the limit and the probability of failure. The former describes how long the tension is expected to continue evolving if it exceeds the safe range, based on the current trend. The latter characterizes the possibility that the exceeding behavior may further break through the physical failure boundary. For example, if the system detects that the current tension has been higher than the upper limit of the safe range for 12 consecutive milliseconds, and predicts that it will continue to exceed the limit for another 28 milliseconds after considering the high load state of the edge nodes, the predicted duration of exceeding the limit can be recorded as 40 milliseconds. If the control command is judged to be likely to continue to lag due to the combined effects of clock jitter and thermal state, the probability of failure can be calculated as 0.47. Regarding anomaly handling, if the memory occupancy rate sample value jumps abnormally, such as instantly rising from 70% to 100% and then falling back to 72% in the next cycle, the system can mark this point as an instantaneous anomaly point. It will not be used as the main basis for the computing power thermal fatigue index, but will be smoothed with adjacent cycles. If the core temperature sensor malfunctions for a short time, causing the reading to exceed the physically acceptable range, such as reporting 150 degrees Celsius, the value will be considered invalid and temporarily replaced by the median of the most recent valid window. If two or more of the four types of features are continuously abnormal at the same time, the system will prioritize increasing the failure probability in the risk assessment result by a safety factor to encourage subsequent switching to be more inclined to the suboptimal mode side. For example, when the same deep-sea pressure-resistant pipe is wound to the middle section, due to short-term fluctuations in the resin supply system, the tension at the unwinding end begins to swing up and down; the buffer queue running in the edge control box increases, and the memory occupancy rate increases significantly; through the four-dimensional observation matrix, the system not only monitors the deviation of the tension body, but also identifies that the memory occupancy rate and the core temperature rise together, indicating that the computing congestion is not a purely instantaneous load, but rather that the underlying resources are beginning to be limited; at this time, the output comprehensive health data includes the error sequence, the clock jitter entropy value, and the computing power thermal fatigue index, and the boundary prediction module further concludes that the predicted value of the over-limit duration is close to the upper limit of the failure duration and the failure probability is rising rapidly; The purpose of this step is to incorporate the multi-source resource status of edge nodes into the control closed loop with a unified data structure, so that subsequent judgments can both monitor the deviation of the tension ontology and assess the computing power bearing conditions, thereby achieving a more stable risk characterization.

[0020] In a preferred embodiment of the present invention, the state assessment module includes: an error calculation submodule, which calls the tension time series data in the multidimensional observation matrix, calculates the absolute value of the deviation between the current tension amplitude and the preset target tension benchmark value, and generates the tension error value in the tension error sequence; a beat entropy quantization submodule, which extracts the data arrival timestamps in the multidimensional observation matrix, calculates the time interval between adjacent timestamps, statistically analyzes the probability distribution of the time intervals, calculates the information entropy based on the probability distribution, and generates the control beat jitter entropy; and a fatigue fusion submodule, which extracts the underlying hardware operating status data in the multidimensional observation matrix, including the CPU load rate of the edge computing node, the memory occupancy rate of the edge computing node, and the core temperature of the edge computing node; normalizes the CPU load rate and the core temperature of the edge computing node, and weights and sums the normalized load rate, temperature, and control beat jitter entropy to generate a computing power thermal fatigue index.

[0021] This embodiment provides a detailed calculation mechanism for the state assessment module. Specifically, although the aforementioned scheme can generate comprehensive health data, if the generation process of error, beat jitter and thermal fatigue is not further broken down, it is difficult to explain whether the tension end or the computing power end deteriorates first under high-risk working conditions. Therefore, this embodiment refines the state assessment module into three sub-processes: error calculation, beat entropy quantification and fatigue fusion. Specifically, the error calculation submodule first reads the tension time series data from the multidimensional observation matrix and compares it point by point with the target tension benchmark value. If the target value is set to 100 N, and the sampled values ​​within a certain window are 99, 105, 111, and 103 N, then the absolute values ​​of the deviations are 1, 5, 11, and 3 N respectively, forming an error sequence. This error sequence can directly reflect control deviations and also provide a basis for subsequent statistics on overrun durations. The beat entropy quantization submodule extracts the corresponding data arrival timestamps. Assuming the timestamps are 0, 2, 4, 9, and 11 milliseconds, the adjacent time intervals are 2, 2, 5, and 2 milliseconds. If, within a longer window, 2 milliseconds occur 7 times, 3 milliseconds occur once, and 5 milliseconds occur twice, the probability distribution of the time intervals is approximately 0.7, 0.1, and 0.2. Based on this distribution, an entropy value can be calculated. The more concentrated the distribution, the lower the entropy value, indicating stable control beats; the more dispersed the distribution, the higher the entropy value, indicating severe beat jitter. In this way, the uniformity of data arrival can be transformed into quantifiable control beat jitter entropy. The fatigue fusion submodule further extracts the CPU load rate and core temperature from the same window; in order to facilitate unified calculation with the entropy value, normalization can be performed first; the load rate reference range is agreed to be 50% to 100%, and the temperature reference range is 40 degrees Celsius to 90 degrees Celsius; if the current load rate is 95%, the normalized value is approximately 0.90. If the current temperature is 80 degrees Celsius, the normalized value is approximately 0.80; if the normalized clock jitter entropy is 0.45, then with weights of 0.5, 0.3, and 0.2 respectively, the computing power thermal fatigue index can be... The formula for calculating the thermal fatigue index of computing power is: ,in, The thermal fatigue index is the power factor. This represents the normalized CPU load rate. The normalized core temperature. The normalized control beat jitter entropy, and are pre-set static weights, respectively.

[0022] The weights of the normalized parameters in the weighted summation are static weights predetermined by combining historical fault data with the analytic hierarchy process. In terms of configuration principles, the weight of the node's central processing unit load rate is set to be strictly greater than the weights of the node's core temperature and the control cycle jitter entropy, to ensure that the system's sensitivity to instantaneous computing power congestion is higher than its sensitivity to gradual heat accumulation. In this way, the system can form a fusion index that simultaneously reflects load, heat generation, and control cycle stability. Regarding anomaly handling, if the target tension reference value changes due to process switching, such as switching from 100 N to 108 N, the error calculation submodule will recalculate according to the new reference value from the switching time to avoid mixing the old and new references; if the timestamp is reversed or repeated, it indicates that there is an abnormal retransmission in the communication link. At this time, the cycle entropy quantization submodule can discard the abnormal point and record a link abnormality count; if the upper and lower limits of the normalization interval are touched, such as the load rate exceeding the statistical upper limit of 100% or the temperature being lower than the reference lower limit, the relevant values ​​will be truncated according to the boundary values ​​to prevent the fusion result from being distorted. For example, on the same production line, after the resin temperature rises, the adhesion state of the fiber bundle surface changes, and the friction at the guide wheel begins to intermittently increase; the error calculation submodule observes that the tension error evolves from a small fluctuation to a continuous high deviation; the cycle entropy quantization submodule finds that the timestamp changes from a uniform 2-millisecond cycle to alternating 2-millisecond, 4-millisecond, and 5-millisecond cycles; the fatigue fusion submodule simultaneously detects that the central processing unit load rate and core temperature continue to rise, eventually pushing the computing power thermal fatigue index to a dangerous area; the system can then distinguish whether the tension end problem and the computing power end problem are aggravated simultaneously or triggered sequentially; The purpose of this step is to decompose the abstract control risk into three interpretable sources, thereby enabling the traceability of the status assessment results and the basis for subsequent switching actions.

[0023] In a preferred embodiment of the present invention, the boundary prediction module includes: a threshold comparison submodule, which calls the computing power thermal fatigue index in the comprehensive health data, calculates the rate of change of the computing power thermal fatigue index approaching the computing power dangerous threshold, and generates the computing power overload approach rate; a delay risk mapping submodule, which determines the duration for which the tension error value exceeds the preset safe tension range based on the tension error sequence in the comprehensive health data, and generates the current over-limit duration; and a probability calculation submodule, which combines the computing power overload approach rate and the current over-limit duration, inputs a preset risk assessment model, outputs the failure probability that the system control command issuance delay exceeds the physical failure boundary condition, and generates a risk assessment result.

[0024] This embodiment provides a prediction mechanism oriented towards failure boundaries. Specifically, although the aforementioned state assessment has given the computational thermal fatigue index and tension error sequence, if the switch judgment is made solely based on whether the current threshold is exceeded, the response is often too late at extreme winding speeds. In particular, when the edge node has not yet truly reduced its frequency but has already rapidly approached the danger edge, the system needs to calculate the failure trigger time and whether it has crossed the physical failure boundary in advance. Therefore, three sub-processes are introduced here: threshold comparison, delay risk mapping, and probability calculation. Specifically, the threshold comparison submodule does not only determine whether the computing power thermal fatigue index is greater than the danger threshold, but also calculates its approach speed. Assuming that the computing power thermal fatigue index in four consecutive cycles is 0.62, 0.68, 0.73, and 0.78, and the danger threshold is 0.80, then although the first three cycles have not exceeded the threshold, its growth slope exceeds the preset slope tolerance limit. The rate of change can be estimated by the adjacent difference, for example, 0.06, 0.05, and 0.05, to obtain the computing power overload approach rate, which means that the edge node will soon reach the condition to trigger the frequency reduction mechanism. The delay risk mapping submodule focuses on the tension end; assuming the safe tension range is set to 95 to 105 N, and the current continuous sampling values ​​are 106, 107, 108, and 109 N, it indicates that the tension has continuously exceeded the upper limit; if the sampling period is 2 milliseconds, the current over-limit duration has reached 8 milliseconds; if the subsequent values ​​further increase, the over-limit duration continues to accumulate; the module can transform discrete error points into a continuous over-limit time that is closer to the process risk; The probability calculation submodule combines the computing power overload approach rate and the current over-limit duration, inputs a preset risk assessment model, and outputs the failure probability. This model can be a lookup table model, a piecewise linear model, or a trained lightweight probabilistic model. Preset piecewise logic rules can be set: when the computing power overload approach rate is low and the over-limit duration is less than 20 milliseconds, the failure probability is 0.1; when the approach rate is medium and the over-limit duration is between 20 and 40 milliseconds, the failure probability is 0.4; when the approach rate is high and the over-limit duration exceeds 40 milliseconds, the failure probability is 0.8. If, in this judgment, the approach rate is high and the current over-limit duration is 36 milliseconds, the output failure probability can be 0.65, and this will also form the risk assessment result. Regarding anomaly handling, if the thermal fatigue index of computing power first rises and then falls within a short window, it indicates that the edge node may have finished processing other concurrent computing tasks for a short period of time. In this case, the threshold comparison submodule can add a smoothing judgment, and only when it increases for several consecutive cycles is the approximation rate considered effective. If the tension error frequently enters and exits around the safety boundary, for example, alternating between 105 N, 104 N, 106 N, and 105 N, the delay risk mapping submodule can adopt a strategy of holding the timer when the limit is exceeded and clearing it when it falls back. Only when it is continuously in the state of exceeding the limit will the duration be accumulated. If one of the input items of the risk assessment model is missing, the failure probability is adjusted and assigned a higher value according to the preset safety bias coefficient, and an additional indicator of reduced data credibility is output to the mode switching module. For example, when the current deep-sea pressure-resistant pipe is nearing the outer layer thickening and winding stage, the material tension safety margin is at a critical state; the edge control box continuously detects a rapid increase in the computing power thermal fatigue index, indicating that the chip is about to reduce its computing power due to heat; at the same time, the tension value has exceeded the process safety limit and has continued for several sampling cycles; based on this, the boundary prediction module calculates that if complex compensation is still maintained, the actual issuance time of the control command may be later than the response window allowed by the physical system, and the probability of failure will increase significantly. Therefore, a high-risk assessment result is given to the downstream in a timely manner. The purpose of this step is to map the two types of information, namely, that computing power is deteriorating and that tension has exceeded the limit, onto the same failure boundary, thereby achieving forward-looking predictions earlier than the actual point of obsolescence.

[0025] In a preferred embodiment of the present invention, the mode switching module includes: The modal arbitration submodule reads the failure probability and computing power thermal fatigue index from the risk assessment results, performs a conditional comparison between the failure probability and the preset failure probability warning threshold, and outputs a modal switching trigger signal. The deep compensation control submodule, in response to the trigger signal that the computing power thermal fatigue index is lower than or equal to the computing power danger threshold, activates the preset edge deep reinforcement learning network, extracts high-frequency nonlinear distortion signals from the tension time series data of the multi-dimensional observation matrix for fitting compensation, and generates the first tension control command. The degradation robust control submodule, in response to the trigger signal that the computing power thermal fatigue index is higher than the computing power danger threshold, puts the edge deep reinforcement learning network into hibernation, activates the preset classical robust controller, shields high-frequency nonlinear distortion signals, and generates a second tension control command based on the deviation between the current tension amplitude and the preset macroscopic tension tolerance range.

[0026] This embodiment provides a dynamic switching mechanism for control modes. Specifically, the aforementioned boundary prediction module can output the failure probability, but if the same controller is always used subsequently, the risk information cannot be truly utilized. Especially when high-frequency nonlinear interference occurs, although the deep feature fitting compensation mode has high accuracy, it will consume more computing power and may accelerate the deterioration of the underlying cycle time. Therefore, this embodiment sets up three sub-processes: mode arbitration, deep compensation control, and degradation robust control, so that the system can select different control logics under different computing power health states. Specifically, the modal arbitration submodule reads the failure probability and the computing power thermal fatigue index, and performs a conditional comparison between the failure probability and a preset failure probability warning threshold. The computing power danger threshold can be set to 0.75, and the failure probability warning threshold to 0.50. When the computing power thermal fatigue index is 0.68 and the failure probability is 0.31, the system maintains the deep feature fitting compensation mode. When the computing power thermal fatigue index rises to 0.79, even if the failure probability is temporarily only 0.46, a degradation signal is triggered because the computing power is now in a more urgent state. The mode switching not only assesses process risks but also assesses the survival status of edge hardware. When the depth compensation control submodule is activated, it calls the preset edge deep reinforcement learning network to extract high-frequency nonlinear distortion signals from the tension time series for fitting compensation. Here, high-frequency distortion can be understood as a disturbance that swings up and down frequently in a short period of time and whose change pattern does not conform to a simple linear law. Within 8 sampling points, the tension values ​​are 100, 104, 98, 105, 97, 106, 99, and 103 Newtons. The system can identify that it is superimposed with high-frequency oscillation components and predict the offset trend of the next 1 to 2 sampling periods by the edge network, thereby generating the first tension control command, such as reducing the braking force in advance and simultaneously fine-tuning the unwinding speed. When the degradation robust control submodule is triggered, it first puts the edge deep reinforcement learning network into hibernation, stopping the fine fitting of high-frequency nonlinear distortion signals to avoid consuming a lot of computing power. Then it activates the classic robust controller, shielding high-frequency small disturbances and making basic adjustments only around the macroscopic tension tolerance range. If the tolerance range is 97 to 103 N, and the current tension is 107 N, the second tension adjustment command can be based on the amount exceeding the upper boundary only, without attempting to compensate for each high-frequency jitter individually. Regarding anomaly handling, if the failure probability received by the modal arbitration submodule contradicts the conclusion of the computing power thermal fatigue index (e.g., the failure probability is high but the computing power thermal fatigue index is low), the system can prioritize retaining the deep feature fitting compensation mode, but increase its control output limit to avoid overly aggressive actions. If the computing power thermal fatigue index is high but the failure probability is still low, the system can still first enter the basic robust control suboptimal mode and prioritize restoring the control cycle. If the switching command is issued when the communication interface is congested, the output of the previous cycle will be maintained in the current control cycle, and the switching will be performed in the next effective cycle to avoid control idling within half a cycle. For example, when the outer layer of the deep-sea pipe fitting enters the high-speed finishing stage, the fiber bundle generates a large amount of high-frequency excitation due to local clustering. Initially, the modal arbitration submodule observes that the computing power thermal fatigue index is still on the safe side, so the depth compensation control submodule continues to work to carefully cancel the complex disturbances, and the production line tension is kept relatively smooth. Subsequently, as the core temperature rises and the cycle jitter intensifies, the computing power thermal fatigue index exceeds the dangerous threshold. At this time, the system no longer insists on pursuing fine correction of every high-frequency fluctuation, but quickly switches to the basic robust control suboptimal mode, only retaining the tension safe range maintenance function, to buy time for the edge control box to recover the cycle. The purpose of this mechanism is to establish a switchable strategy layer between control precision and control survivability, so as to pursue better compensation when the hardware is running normally and prioritize maintaining within the physical safety boundary when computing power deteriorates.

[0027] As a preferred embodiment of the present invention, when the degradation robust control submodule generates the second tension control command, it performs the following operations: obtains the current tension amplitude and determines whether the current tension amplitude is within the preset macroscopic tension tolerance range; When the current tension amplitude is within the macroscopic tension tolerance range, the current tension fluctuation is determined to be an allowable physical tolerance fluctuation, and a second tension control command to maintain the current control state is output, without executing the calculation to suppress the fluctuation; when the current tension amplitude exceeds the macroscopic tension tolerance range, a proportional-integral-derivative control algorithm is used to output a second tension control command to return to the target tension reference value.

[0028] This embodiment provides a refined execution mechanism for degraded robust control. Specifically, in the previous layer switching logic, the system can already revert from the deep feature fitting compensation mode to the basic robust control suboptimal mode. However, if all fluctuations are still calculated after degrading, the computational overhead cannot be truly reduced. Therefore, this embodiment further stipulates that under the basic robust control suboptimal mode, it is first determined whether the tension is within the macroscopic tension tolerance range. Only when it exceeds the range is the proportional-integral-derivative control algorithm activated. Specifically, the upper and lower limits of the preset macroscopic tension tolerance range are determined based on a joint evaluation of the resin matrix characteristics and glass fiber tensile strength of the currently processed pipe fitting; the absolute span of the tolerance range is strictly constrained between the preset target tension reference value and the preset structural limit threshold, and a safety margin of at least a preset percentage is reserved to ensure that the allowable fluctuations that occur within the range will absolutely not trigger physical failure mechanisms such as microcracks. Based on this, assuming the macroscopic tension tolerance range is 97 to 103 N; the degraded robust control submodule first reads the current tension amplitude in each cycle; if the reading is 101 N, it is within the tolerance range, the system judges the fluctuation as an allowable fluctuation, and directly outputs the second tension control command to maintain the current control state, such as maintaining the current speed and braking torque of the unwinding motor, without additional calculation of suppression action; this can significantly reduce invalid calculations. If the current tension amplitude is 106 N, it exceeds the upper limit of the tolerance range. At this time, the proportional-integral-derivative (PID) control algorithm is activated for basic correction. This can be illustrated by a basic calculation example: assume the proportional coefficient is 0.6, the integral coefficient is 0.2, and the derivative coefficient is 0.1. Furthermore, in order to avoid the control system only correcting the tension to the tolerance boundary, which would lead to frequent critical jitter due to excessive limits, the error calculation at this time should use the process target tension reference value of 100 N as the reference, rather than the tolerance boundary of 103 N. Therefore, the current error is 106 minus 100 equals 6 N; if the tension value in the previous cycle was 105 N, then the error in the previous cycle was 5 N; assuming the current cumulative integral term is 15, the control output can be adjusted accordingly. The calculation is then converted into the actual adjustment amount for the unwinding speed or braking force. This design based on the central target value ensures that once the tension exceeds the limit, it is quickly and thoroughly adjusted back to the vicinity of the target reference value. Since the algorithm only handles large macroscopic deviations and does not track high-frequency distortion details, the overall computational resource overhead is much lower than that of the deep feature fitting compensation mode. Furthermore, if the tension is below the lower limit of the tolerance range, such as 95 N, the negative tension error (95-100) = -5 N can be calculated based on the target benchmark value in the same way, and a second tension control command is output to increase the braking force or decrease the unwinding speed, so that the tension returns to the macro tension tolerance range; regardless of whether the tension exceeds the limit upward or downward, the basic robust control suboptimal mode can provide a reliable and clearly targeted callback capability; Regarding anomaly handling, if the current tension amplitude is exactly equal to the tolerance interval boundary value, such as 103 N or 97 N, the system can treat it as being within the interval, avoiding unnecessary actions caused by repeated boundary jitter. If the integral term in the proportional-integral-derivative calculation becomes too large due to long-term deviation, an integral limit can be set, for example, restricting the integral term to a certain value. Within the range, prevent overshoot of the control output; if the sampling period changes temporarily, for example from 2 milliseconds to 4 milliseconds, the differential term is recalculated according to the actual period to avoid miscalculation due to changes in the clock cycle; For example, on the aforementioned deep-sea pipe fitting that was about to be completed, the system had entered the suboptimal mode of basic robust control due to increased computational power and thermal fatigue. At this time, there were still high-frequency micro-oscillations on the surface of the fiber bundle, but most of the oscillation range only caused the tension amplitude to fluctuate between 99N and 102N, which was still within the macroscopic tolerance range. Therefore, the system no longer repeatedly calculated for these minor disturbances. Only when resin accumulation caused the tension to suddenly rise to 106 Newtons did the suboptimal mode of basic robust control activate the proportional-integral-derivative control algorithm to pull back the tension to the safe range. This controlled the system's computational resource consumption while satisfying process safety constraints. The purpose of this step is to ensure that the degraded controller truly reflects the characteristics of prioritizing computing power conservation, thereby achieving tolerance for permissible fluctuations and limited but effective intervention in dangerous deviations.

[0029] In a preferred embodiment of the present invention, the system further includes: a computing power recovery adaptive module, which continuously monitors comprehensive health data after the mode switching module switches to the basic robust control suboptimal mode, and compares the real-time updated computing power thermal fatigue index with the preset computing power recovery safety threshold; when the real-time updated computing power thermal fatigue index is higher than the computing power recovery safety threshold, the basic robust control suboptimal mode is maintained; when the real-time updated computing power thermal fatigue index is lower than or equal to the computing power recovery safety threshold, a mode reset signal is generated to trigger the mode switching module to smoothly transition to the deep feature fitting compensation mode.

[0030] This embodiment provides an adaptive reset mechanism after computing power recovery. Specifically, in the aforementioned degradation logic, the system can revert to basic robust control when computing power deteriorates. However, without a reset mechanism, the system may remain in a suboptimal control state for a long time, causing the production line to continuously operate in the suboptimal mode of basic robust control. Therefore, this embodiment further sets up a computing power recovery adaptive module to continuously evaluate whether the conditions for returning to the deep compensation mode are met after degradation is completed. Specifically, during the suboptimal mode operation of basic robust control, the computing power recovery adaptive module continuously reads the real-time computing power thermal fatigue index from the comprehensive health data. Assuming the system has degraded from a high-risk state of 0.82 to basic robust control, and subsequently observes the computing power thermal fatigue index decreasing to 0.76, 0.71, 0.67, and 0.62 in several windows, while the preset computing power recovery safety threshold is 0.65, the module maintains basic robust control unchanged when the index is higher than 0.65. When it decreases to 0.62 and remains stable for several consecutive cycles, the module generates a modal reset signal. The reset action is not an instantaneous hard cut, but a smooth transition; a short transition window with two control outputs existing in parallel can be used, for example, lasting 10 milliseconds; within this window, the output weight of the basic robust control gradually decreases from 1 to 0, while the output weight of the deep compensation control gradually increases from 0 to 1; for example, if the first transition cycle uses weights of 0.8 and 0.2, the second cycle uses 0.5 and 0.5, and the third cycle uses 0.2 and 0.8, then the secondary impact caused by suddenly switching from coarse adjustment back to fine adjustment can be avoided; In terms of anomaly handling, if the computing power thermal fatigue index rebounds after just dropping to near the recovery safety threshold, for example, from 0.64 to 0.68, the computing power recovery adaptive module cancels the current reset and continues to maintain basic robust control; if the tension error suddenly expands and approaches the danger range during the reset transition, the transition is immediately stopped and basic robust control takes over again; if repeated situations of resetting and degrading occur multiple times, the system can temporarily extend the minimum dwell time for maintaining basic robust control to reduce frequent switching. For example, in the final stage of this deep-sea pressure-resistant pipe fitting, the system has avoided a computing power crisis induced by high-frequency frictional vibration through degradation robust control; as the heat dissipation of the edge control box gradually recovers and the load of the central processor decreases, the computing power thermal fatigue index begins to fall; the computing power recovery adaptive module continuously monitors that the index has fallen below the recovery safety threshold and remains stable, so it issues a modal reset signal, allowing the system to gradually return to the depth compensation mode without disturbing the current tension, so as to perform more precise control on the last few layers of winding; The purpose of this mechanism is to enable the system to have closed-loop recovery capabilities for state degradation and adaptive reset, thereby achieving a balance between production continuity and winding quality.

[0031] As a preferred embodiment of the present invention, the computing power recovery safety threshold is determined by obtaining the critical temperature parameter of the underlying chip of the edge computing node that triggers frequency reduction protection, and by performing bias calculation in combination with the preset heat dissipation time constant. The computing power recovery safety threshold is strictly less than the computing power danger threshold, so as to form a hysteresis loop for mode switching.

[0032] This embodiment provides a mechanism for determining the computing power recovery safety threshold. Specifically, although the previous layer scheme introduced a recovery safety threshold, if the threshold is set too close to the danger threshold, the system is prone to repeatedly switching near the boundary, resulting in control jitter. Therefore, this embodiment uses the frequency reduction protection critical temperature and heat dissipation time constant of the underlying chip to perform bias calculation on the recovery threshold and makes it strictly lower than the danger threshold to form a hysteresis loop. Specifically, the chips used in edge control boxes typically have a clearly defined critical temperature for frequency reduction protection; this critical temperature can be 85 degrees Celsius. The system then combines this with the measured heat dissipation time constant, for example, it takes an average of 20 seconds for the chip to naturally drop from 85 degrees Celsius to 75 degrees Celsius. Based on this, it can be determined that even if the current temperature is just away from the frequency reduction edge, the internal heat accumulation has not been completely released. Therefore, a bias can be added to the recovery judgment. For example, if the computing power thermal fatigue index corresponding to the dangerous threshold is 0.75, then the recovery safety threshold is not directly set to 0.75, but is reduced to 0.65 by combining the temperature bias and the heat dissipation time constant. The specific rules for the bias calculation are as follows: computing power recovery safety threshold = preset computing power danger threshold - (critical temperature parameter - current core temperature) × attenuation coefficient corresponding to heat dissipation time constant; in this way, the system is only allowed to return to deep compensation mode when the chip state has obviously recovered to a safer range. The formation of the hysteresis loop can be illustrated by a simple numerical example: when the computing power thermal fatigue index rises from 0.70 to 0.76, the system triggers a degradation; even if the index falls back from 0.76 to 0.72, it is still insufficient to reset; only when it continues to fall to 0.65 or below is recovery triggered; this design, where the degradation threshold is higher than the recovery threshold, avoids switching back and forth between edge fluctuations such as 0.74, 0.75, 0.74, and 0.76. In terms of anomaly handling, if the critical temperature parameter changes due to a change in the chip model, the recovery safety threshold needs to be recalibrated; if the heat dissipation airflow is blocked or the ambient temperature is abnormal, causing the heat dissipation time constant to increase significantly, the system can automatically lower the recovery safety threshold and postpone the reset timing; if the chip's critical temperature parameter cannot be read, the recovery safety threshold will temporarily call the preset safe conservative default value and lock it at a position that is at least one preset safety margin lower than the dangerous threshold. For example, during the aforementioned night shift production, the edge control box approaches the chip's thermal protection zone due to continuous high-speed operation. Although the core temperature begins to decrease after degradation, the ambient temperature around the cabinet remains high. If deep compensation is immediately restored once the indicator shows a slight drop, it is very easy to return to the dangerous edge due to heat accumulation. Therefore, based on the chip's critical temperature parameter of 85 degrees Celsius and the measured heat dissipation time constant, the system sets the recovery safety threshold significantly lower than the danger threshold. In the subsequent winding of the same pipe, the mode switching is more stable and will not frequently jump due to short-term indicator fluctuations. The purpose of this mechanism is to provide a thermal safety buffer for mode switching through a hysteresis loop, thereby achieving stability and predictability of the reset action.

[0033] As a preferred embodiment of the present invention, the physical failure boundary condition is defined as a time window in which the tension overshoot exceeds a preset structural limit threshold and lasts for more than fifty milliseconds. The time window serves as the absolute time constraint for the system control response.

[0034] This embodiment provides a mechanism for defining physical failure boundary conditions. Specifically, in the aforementioned modules, risk assessment and mode switching both require a clear and unified final failure judgment criterion. Otherwise, different modules may use excessive tension, excessive delay, or excessive temperature as the judgment criteria, resulting in a fragmented control strategy. Therefore, this embodiment clearly defines the physical failure boundary as: the tension overshoot exceeds the preset structural limit threshold and the duration exceeds 50 milliseconds. This time window serves as the highest priority control constraint boundary for the entire system. Specifically, the process target tension can be set to 100 N and the structural limit threshold to 112 N. If, starting at a certain moment, the tension is continuously maintained at 113, 114, 116, and 115 N, and the sampling period is 5 milliseconds, then after this over-limit state lasts for 10 cycles, the cumulative duration will reach 50 milliseconds. The system will recognize this as reaching the absolute failure boundary. Here, it is not only judged whether the overshoot exceeds the limit value, but also whether the duration of the overshoot exceeds the tolerance window. If it only momentarily reaches 113 N but falls back quickly after 5 milliseconds, it is not considered a final failure, but only counted as a high-risk warning. In terms of control implementation, the 50-millisecond window will inversely constrain boundary prediction and mode switching. When the boundary prediction module estimates that the control command may be delayed within the next 20 milliseconds, and the current tension has exceeded the limit by 30 milliseconds, the sum of the two will approach the 50-millisecond upper limit. The system must immediately generate a trigger signal to switch to the basic robust control suboptimal mode. The 50-millisecond window is not only a result criterion, but also becomes the absolute time budget of the front-end control algorithm. Regarding anomaly handling, if the production line requires different structural limit thresholds due to different specifications of pipe fittings, the threshold can be issued along with the process formula before each pipe fitting is started; if the sampling period changes, the duration calculation is based on the actual period accumulation, rather than a fixed number of samples; if the tension briefly drops back to near the critical value during the over-limit period, such as fluctuating around 112 Newtons, the system can use a method of continuously timing above the limit value and resetting to zero when it drops back to the safe zone to avoid ambiguity in boundary judgment; For example, in this high-pressure-bearing fiberglass pipe used for deep-sea transportation, the structural tolerance of the outer layer winding stage is lower than that of the inner layer; the system pre-sets 112 N for 50 milliseconds as the absolute failure boundary; when a frictional excitation superimposed with a computational delay, the tension exceeds 112 N for several consecutive cycles, and the boundary prediction module immediately maps it to the lower limit of the time window approaching the absolute failure determination; thus, the mode switching module prioritizes protecting the physical boundary and quickly suppresses the tension, rather than continuing to attempt complex fine compensation; The purpose of this step is to establish a unified and executable failure determination boundary for the entire system, thereby achieving consistency in the determination criteria between risk prediction, control switching, and process failure determination.

[0035] As a preferred embodiment of the present invention, the high-frequency nonlinear distortion signal is an excitation signal with a frequency greater than a preset frequency threshold and an amplitude change rate that does not meet the linear constraint. The excitation signal can easily induce edge computing nodes to increase their computing load.

[0036] This embodiment provides a mechanism for identifying high-frequency nonlinear distortion signals. Specifically, in the aforementioned deep compensation control, the system needs to perform fitting compensation for high-frequency disturbances. However, if it is not clear which signals belong to the high-frequency nonlinear distortions that should be prioritized for processing, the deep compensation network may confuse ordinary fluctuations with abnormal excitations, thereby indiscriminately increasing the computational load. Therefore, this embodiment limits such signals to excitation signals whose frequency is higher than a preset frequency threshold and whose amplitude change rate does not meet the linear constraint. Specifically, the system can first estimate the fluctuation frequency within a short sliding window; assuming the sampling period is 1 millisecond, if the tension values ​​in the most recent 10 sampling points are 100, 104, 99, 105, 98, 106, 99, 104, 100, and 103 Newtons, and multiple rapid reversals occur within the window, then the signal meets the high-frequency condition; if the preset frequency threshold corresponds to at least 4 reversals within a 10-millisecond window, then the signal meets the high-frequency condition. Next, the system determines whether the rate of change of amplitude satisfies the linear constraint. A simplified deduction can be made: if the linear constraint requires that the rate of change of adjacent sampling points should be approximately stable within ±2 N / millisecond, and the continuous changes in the current sequence are +4, -5, +6, -7, and +8 N, then the rate of change not only has a large amplitude, but also alternates between positive and negative values ​​with increasing amplitude, which obviously does not conform to the linear law; at this time, the fluctuation is marked as a high-frequency nonlinear distortion signal. Once identified, these signals serve two purposes: in deep compensation mode, the system attempts to fit them to improve the fine-grained tension control; from a risk management perspective, these signals often induce edge nodes to increase their computational load because the deep compensation network will allocate more inference resources to track these abnormal fluctuations; this signal is not only a source of disturbance at the tension end but also a trigger for the deterioration of computing power; therefore, when the computing power thermal fatigue index increases, degradation robust control will choose to block these signals and no longer consume a lot of resources to track their details. Regarding anomaly handling, if the detected frequency is higher than the threshold, but the amplitude change rate is basically stable and meets the linear constraint, the signal can be classified as a normal high-frequency disturbance and does not need to be treated as a nonlinear distortion. If the frequency is slightly lower than the threshold, but the amplitude change rate is extremely abnormal and has caused the tension to exceed the limit rapidly, the system can include it in the processing as a dangerous disturbance in advance. If the sensor noise itself causes false high-frequency reversals, the occasional noise can be eliminated through multi-cycle consistency checks. For example, during the continuous processing in the latter half of the night on the aforementioned production line, due to batch differences in resin, irregular vibrations are generated when the fiber bundle passes through the guide roller. These vibrations do not immediately push the tension to the scrap level, but instead form a dense, high-reversal, nonlinear wave sequence at the sensor end. If the depth compensation control continues to fit this type of signal, it will significantly increase the load and temperature of the edge control box. Therefore, the system uses depth compensation to finely suppress it when the hardware is operating normally, and shields it when the computing power deteriorates, retaining only the basic response to low-frequency large deviations, thereby preventing the edge nodes from being induced to a state of computing power instability by this type of vibration. The purpose of this mechanism is to clearly distinguish between complex disturbances that deserve fine compensation and high-cost disturbances that should be actively abandoned during dangerous periods, thereby achieving dynamic matching between edge computing resources and control objectives.

[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any conventional modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An adaptive control system for winding tension of fiberglass pipe fittings based on edge computing, characterized in that, The system is deployed on an edge computing node and is communicatively connected to an external tension sensor and an external winding actuator. The system includes: The status monitoring module collects the tension time series data output by the external tension sensor and the underlying hardware operation status data of the edge computing node, records the data arrival timestamp, generates a multi-dimensional observation matrix, and transmits it to the status evaluation module. The status assessment module acquires the multi-dimensional observation matrix, calculates the tension error sequence between the current tension amplitude and the preset target tension benchmark value, analyzes the data arrival timestamp to calculate the control beat jitter entropy, and combines the underlying hardware operating status data to quantify the computing power thermal fatigue index, generate comprehensive health data, and transmit it to the boundary prediction module. The boundary prediction module calls the comprehensive health data, compares the computing power thermal fatigue index with the preset computing power danger threshold, associates and maps the tension error sequence with the preset physical failure boundary conditions, calculates the failure probability of the system within the preset time window, generates risk assessment results and transmits them to the mode switching module. The mode switching module, based on the risk assessment results, performs a control degradation judgment: when the computing power thermal fatigue index is lower than or equal to the computing power danger threshold, it maintains the deep feature fitting compensation mode to output the first tension control command; when the computing power thermal fatigue index is higher than the computing power danger threshold, it switches to the basic robust control suboptimal mode to output the second tension control command; and sends the first tension control command or the second tension control command to the external winding actuator through the communication interface to adjust the unwinding tension; The status assessment module includes: The error calculation submodule calls the tension time series data in the multidimensional observation matrix to calculate the absolute value of the deviation between the current tension amplitude and the preset target tension benchmark value, and generates the tension error value in the tension error sequence. The beat entropy quantization submodule extracts the data arrival timestamps from the multidimensional observation matrix, calculates the time interval between adjacent timestamps, statistically analyzes the probability distribution of the time intervals, calculates the information entropy based on the probability distribution, and generates the control beat jitter entropy. The fatigue fusion submodule extracts the underlying hardware operating status data from the multi-dimensional observation matrix. The underlying hardware operating status data includes the CPU load rate of the edge computing node, the memory occupancy rate of the edge computing node, and the core temperature of the edge computing node. The CPU load rate and core temperature of the edge computing node are normalized, and the normalized load rate and temperature are weighted and summed with the control beat jitter entropy according to a preset static weight to generate the computing power thermal fatigue index.

2. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 1, characterized in that, The multidimensional observation matrix includes tension amplitude, CPU load rate of the edge computing node, memory occupancy rate of the edge computing node, and core temperature of the edge computing node. The comprehensive health data includes tension error sequence, clock jitter entropy value, and computing power thermal fatigue index. The risk assessment results include predicted over-limit duration and failure probability.

3. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 1, characterized in that, The boundary prediction module includes: The threshold comparison submodule calls the computing power thermal fatigue index in the comprehensive health data, calculates the rate of change of the computing power thermal fatigue index as it approaches the computing power danger threshold, and generates the computing power overload approach rate. The delay risk mapping submodule determines the duration for which the tension error value exceeds the preset safe tension range based on the tension error sequence in the comprehensive health data, and generates the current over-limit duration. The probability calculation submodule combines the computing power overload approach rate and the current over-limit duration, inputs a preset risk assessment model, which is an assessment model that maps the computing power overload approach rate and the current over-limit duration to a preset failure probability. Through this model, it outputs the failure probability when the system control command issuance delay exceeds the physical failure boundary condition, and generates a risk assessment result.

4. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 1, characterized in that, The mode switching module includes: The modal arbitration submodule reads the failure probability and the computing power thermal fatigue index from the risk assessment results, performs a conditional comparison between the failure probability and the preset failure probability warning threshold, and outputs a modal switching trigger signal. The deep compensation control submodule, in response to a trigger signal that the computing power thermal fatigue index is lower than or equal to the computing power danger threshold, activates a preset edge deep reinforcement learning network, extracts high-frequency nonlinear distortion signals from the tension time series data of the multidimensional observation matrix and performs feedforward fitting compensation, and generates a first tension control command. The degradation robust control submodule, in response to the trigger signal that the computing power thermal fatigue index is higher than the computing power danger threshold, puts the edge deep reinforcement learning network into hibernation, activates the preset classical robust controller, shields the high-frequency nonlinear distortion signal, and generates a second tension control command based on the deviation between the current tension amplitude and the preset macroscopic tension tolerance range.

5. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 4, characterized in that, When generating the second tension control command, the degradation robust control submodule performs the following operations: obtains the current tension amplitude and determines whether the current tension amplitude is within the preset macroscopic tension tolerance range; When the current tension amplitude is within the macroscopic tension tolerance range, the current tension fluctuation is determined to be an allowable physical tolerance fluctuation, and a second tension control command to maintain the current control state is output, without executing the calculation to suppress the fluctuation; when the current tension amplitude exceeds the macroscopic tension tolerance range, a proportional-integral-derivative control algorithm is used to output a second tension control command to return to the target tension reference value.

6. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 1, characterized in that, The system also includes a computing power recovery adaptive module, which continuously monitors the comprehensive health data after the mode switching module switches to the basic robust control suboptimal mode, and compares the real-time updated computing power thermal fatigue index with the preset computing power recovery safety threshold. When the real-time updated computing power thermal fatigue index is higher than the computing power recovery safety threshold, the basic robust control suboptimal mode is maintained; when the real-time updated computing power thermal fatigue index is lower than or equal to the computing power recovery safety threshold, a mode reset signal is generated, triggering the mode switching module to smoothly transition to the deep feature fitting compensation mode.

7. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 6, characterized in that, The computing power recovery safety threshold is determined by obtaining the critical temperature parameter of the underlying chip of the edge computing node that triggers frequency reduction protection, and by performing bias calculation in combination with a preset heat dissipation time constant. The computing power recovery safety threshold is strictly less than the computing power danger threshold to form a hysteresis loop for mode switching.

8. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 1, characterized in that, The physical failure boundary condition is defined as a time window in which the tension overshoot exceeds a preset structural limit threshold and lasts for more than fifty milliseconds. This time window serves as the absolute time constraint for the system control response.

9. The edge computing-based adaptive tension control system for fiberglass pipe winding as described in claim 4, characterized in that, The high-frequency nonlinear distortion signal is an excitation signal whose frequency is greater than a preset frequency threshold and whose amplitude change rate does not meet the linear constraint. The excitation signal can easily induce the edge computing node to increase its computing load.

Images