Gypsum board production line board disorder intelligent early warning system based on edge computing

Abstract

The present application relates to the technical field of production safety management, and discloses a gypsum board production line disorder board abnormal intelligent early warning system based on edge computing, which constructs a discrimination basis by fusing three types of multi-dimensional features of geometric deviation, appearance deformation and operation time sequence, combines a self-adaptive dynamic threshold and a four-level classification discrimination model, accurately distinguishes normal process tolerance, thermal warping deformation and different grades of true disorder boards, replaces a traditional single threshold detection mode, greatly reduces false positive and false negative rates, and improves gypsum board production disorder board detection precision and scene adaptability; through an edge computing architecture, a lightweight model is deployed locally on a production line, and data acquisition, feature extraction and fault discrimination are all locally executed; in combination with a hierarchical intervention and a shock absorption smoothing mechanism, differentiated treatment is performed according to different fault grades, the fault is suppressed from expanding, unnecessary downtime loss is reduced, and the stability of gypsum board production is ensured.

Description

Technical Field

[0001] This invention relates to the technical field of production safety management, specifically to an intelligent early warning system for disordered boards in a gypsum board production line based on edge computing. Background Technology

[0002] With the continuous expansion of gypsum board production capacity, high-speed continuous production lines have become the mainstream production mode in the industry. During the multi-stage flow of gypsum board through conveying, drying, cutting, and transfer, it is susceptible to defects such as positional shifts, tilting, and stacking jams due to factors like equipment wear, process fluctuations, and environmental interference. These defects can range from increasing the defect rate of finished products to causing equipment damage and prolonged production line downtime, severely impacting production efficiency and corporate economic benefits. Existing gypsum board defect detection solutions suffer from three core defects: First, insufficient detection accuracy, often relying on fixed thresholds, which cannot accurately distinguish between thermal warping deformation after drying, normal process tolerances, and genuine defective board issues, resulting in high false alarm and false negative rates. Second, insufficient real-time performance, often employing a centralized cloud-based processing architecture with high data transmission and inference round-trip latency, unable to meet the cycle time requirements of 100-meter-level high-speed gypsum board production lines. Third, poor long-term adaptability; manual recalibration of parameters is required when changing gypsum board specifications, and the high dust and strong vibration environment of the production site continuously causes detection benchmark drift, with the system lacking closed-loop self-optimization capabilities.

[0003] Chinese invention patent application CN118736786A discloses a safety protection system for a gypsum board production line. This system includes a protective net section fixedly installed outside the production equipment; a power-off sliding door section with a limit device, fixedly located at the motor; a monitoring and alarm section on one side of the production equipment to monitor the production line and issue real-time alarms when abnormalities occur; and a safety light curtain section located in a preset key operating area to emit and receive infrared rays and determine whether to cut off the power to the corresponding equipment. This patent not only improves the physical safety protection of the gypsum board production line but also introduces an intelligent monitoring and early warning mechanism, significantly enhancing the safety and stability of the production line and achieving comprehensive protection for equipment and operators, reducing the risk of safety accidents. However, the above technical solution cannot scientifically monitor and manage the disordered state of gypsum board during the gypsum board production process.

[0004] Existing gypsum board disorder detection solutions suffer from three core defects: First, insufficient detection accuracy, often relying on fixed thresholds, which fails to accurately distinguish between thermal warping deformation after drying, normal process tolerances, and genuine disorder faults, resulting in high false alarm and false negative rates. Second, insufficient real-time performance, often employing a centralized cloud-based processing architecture, leading to high data transmission and inference round-trip latency, which cannot match the cycle time requirements of high-speed gypsum board production lines. Third, poor long-term adaptability, requiring manual recalibration of parameters when changing gypsum board specifications, and the high dust and strong vibration environment of the production site continuously causes detection benchmark drift, with the system lacking closed-loop self-optimization capabilities. Summary of the Invention

[0005] To address the aforementioned issues, this invention achieves millisecond-level detection of disordered gypsum board based on an edge computing architecture. It integrates multi-dimensional features including geometry, morphology, and temporal sequence, and combines adaptive thresholds to accurately distinguish four types of board states. Through edge-cloud incremental learning, it achieves closed-loop self-optimization of the model, adapting to drifting on-site working conditions. Combined with hierarchical intervention and smooth control, it balances detection accuracy, response speed, and production stability, effectively reducing the false alarm rate and downtime losses of disordered gypsum board production warnings.

[0006] This invention is achieved through the following technical solution: an intelligent early warning system for disordered gypsum board production line based on edge computing, the system comprising the following:

[0007] The multi-source sensing and acquisition module synchronously collects board image data, spatial location data and production line operation parameter data in each section of the gypsum board production line, and constructs a standardized detection dataset after performing front-end data preprocessing.

[0008] The edge computing processing module extracts geometric deviation features, morphological deformation features, and runtime sequence features of the board based on the standardized detection dataset, and constructs a multi-dimensional disordered board feature vector. The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state, and outputs a warning signal of the corresponding level.

[0009] The edge-cloud collaborative iteration module uploads the difficult samples with the discrimination confidence level below the set threshold and the false positive and false negative samples marked by manual review to the cloud service platform. After the cloud completes the model optimization training based on the incremental samples, it sends the updated model to each edge node to realize the closed-loop self-optimization of the discrimination model.

[0010] The early warning and linkage execution module, based on the graded early warning signal, will link the corresponding production equipment to perform graded intervention operations, and at the same time synchronize the disordered board fault data to the production management system and operation and maintenance terminal.

[0011] Preferably, image data of the gypsum board, spatial location data, and production line operating parameter data are collected simultaneously at each stage of the gypsum board production line. After performing front-end data preprocessing, a standardized inspection dataset is constructed, which specifically includes the following steps:

[0012] S11. Detection nodes are deployed in four core sections: wet board conveying section, drying kiln outlet section, cutting and separating section, and stacking and transfer section. Connect the linear array industrial camera, laser rangefinder array and roller speed encoder to collect board contour image data, board height data and production line real-time speed data respectively.

[0013] S12. Perform dust occlusion repair, illumination equalization and edge enhancement processing on the collected board image data to suppress image noise caused by high dust and illumination fluctuations in the production site;

[0014] S13. Perform vibration filtering and outlier removal processing on the collected laser ranging data, and complete the spatiotemporal registration of the plate position by combining the roller speed data.

[0015] S14. Align the preprocessed image data, ranging data, and working condition data with a unified timestamp, encapsulate them into a standardized format, and output a standardized detection dataset.

[0016] Preferably, based on the standardized detection dataset, geometric deviation features, morphological deformation features, and runtime sequence features of the board are extracted to construct a multi-dimensional disordered board feature vector, specifically including the following steps:

[0017] S21. Based on the plate edge image and the preset position baseline, calculate the plate's lateral offset, longitudinal tilt angle, adjacent plate spacing deviation, and inter-column misalignment of multiple conveyor columns, and extract geometric deviation features.

[0018] S22. Based on the multi-point height data of the laser ranging array, calculate the maximum warping height of the board surface, the overall flatness deviation, and the size of the edge defects, and extract the morphological deformation features.

[0019] S23. Retrieve plate position data from multiple consecutive sampling periods, calculate the offset change rate, position fluctuation trend, and matching degree between plate speed and roller speed, and extract runtime sequence features.

[0020] S24. Combine the geometric deviation features, morphological deformation features, and runtime sequence features to generate a multi-dimensional random board feature vector.

[0021] Preferably, the multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state. The adaptive threshold calibration mechanism in the corresponding level of warning signal is output, and the following steps are specifically performed:

[0022] S311. Receive the current gypsum board thickness specifications, length specifications, and real-time operating speed parameters of the production line from the production system.

[0023] S312. Call the pre-stored baseline tolerance library and match the normal offset threshold, warpage allowable threshold, and disordered board judgment threshold corresponding to the current board type and line speed.

[0024] S313. Correct the judgment threshold based on the ambient temperature and humidity and the equipment wear compensation coefficient, and output the dynamic judgment threshold under the current working condition to the lightweight disordered board classification and discrimination model.

[0025] S314. When changing production lines, the reference parameters of the corresponding board type are automatically switched without manual recalibration.

[0026] Preferably, the multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state. The corresponding level of warning signal is output. The state discrimination processing of the lightweight disordered board classification and discrimination model specifically includes the following steps:

[0027] S321. Input the multi-dimensional randomized feature vector into the shared feature extraction layer of the model, extract the spatial features corresponding to geometric deviation and shape deformation through a multi-layer one-dimensional convolutional neural network, extract the dynamic trend features corresponding to the runtime sequence through a long short-term memory network, and output a high-dimensional hidden layer feature vector after fusion.

[0028] S322. Input the high-dimensional hidden layer feature vector into the classification fully connected layer, and after nonlinear mapping, normalize it through the Softmax function to output the confidence scores corresponding to the four states: normal process tolerance, thermal warpage deformation, slight board disorder, and severe board disorder.

[0029] S323. Obtain the dynamic judgment threshold output by the adaptive threshold calibration unit, compare the confidence scores of the four states with the corresponding thresholds respectively, and filter out the state categories with confidence scores higher than the judgment thresholds.

[0030] S324. Among the state categories that meet the threshold requirements, select the category with the highest confidence as the final judgment result and output the corresponding warning signal.

[0031] Preferably, the problematic samples with a discrimination confidence level below a set threshold and the false positive and false negative samples marked by manual verification are uploaded to the cloud service platform. After the cloud completes the model optimization training based on the incremental samples, it distributes the updated model to each edge node to achieve closed-loop self-optimization of the discrimination model. The specific steps include:

[0032] S41. The edge terminal calculates the classification confidence score for each discrimination result, and automatically marks the samples with the highest confidence score below the set threshold as difficult samples and stores them in the local cache. At the same time, it receives false alarm samples and missed alarm samples marked by maintenance personnel through the field terminal.

[0033] S42. According to a fixed cycle, upload the difficult samples, manually reviewed samples, along with the corresponding process information, environmental temperature and humidity, plate type and specifications, and production line speed and working condition data to the cloud service platform in encrypted form.

[0034] S43. Based on the current online model, the cloud uses incremental learning to fine-tune and optimize only the weights of the model's classification layer. After the overall accuracy of the validation set meets the standard, an updated lightweight model is generated.

[0035] S44. The updated model is distributed to all edge computing nodes via remote upgrade. The nodes complete the model update using hot loading, without interrupting the normal testing operations of the production line.

[0036] Preferably, based on the graded early warning signal, the corresponding production equipment is linked to perform graded intervention operations, and the disordered board fault data is synchronized to the graded intervention operations in the production management system and the operation and maintenance terminal. The specific logic is as follows:

[0037] When the condition is determined to be thermal warping deformation, a temperature zone fine-tuning command is sent to the drying kiln temperature control system to optimize the temperature field distribution without changing the production line operating speed.

[0038] When the board is determined to be in a state of slight board disorder, the corresponding section roller conveyor will be activated to reduce its speed. At the same time, the pneumatic guide correction mechanism will be triggered to perform correction operation. After the board posture returns to normal, the original running speed will be automatically restored.

[0039] When a severe board disorder is detected, the corresponding emergency stop command for the work section is immediately triggered, the on-site audible and visual alarm device is activated, and the fault location and fault type are pushed to the handheld terminal of the on-site maintenance personnel.

[0040] Preferably, based on the graded early warning signal, the corresponding production equipment is linked to perform graded intervention operations. Simultaneously, the disordered board fault data is synchronized to the production management system and maintenance terminal. Before performing the graded intervention operations, a smoothing process to prevent severe vibrations is also included, specifically:

[0041] Read the current value of the device control signal from the previous control cycle, and calculate the difference between the target control signal value corresponding to the action command and the current value;

[0042] If the difference exceeds the preset maximum adjustment step size limit, the control signal issued this time will be forcibly truncated to the maximum adjustment step size limit, and the value will be gradually adjusted to the target value according to the step size gradient.

[0043] If the difference does not exceed the preset maximum adjustment step size limit, the target control signal corresponding to the action command will be sent directly to the execution device.

[0044] Preferably, the process of triggering tiered intervention operations on corresponding production equipment based on tiered early warning signals, and simultaneously synchronizing the fault data of disordered boards to the production management system and maintenance terminal, also includes statistical analysis and risk tracing steps, specifically:

[0045] Statistics on the occurrence rate of disordered boards, the proportion of fault types, and the false alarm and missed alarm rates are calculated according to the dimensions of work section, shift, and board type, and multi-dimensional analysis reports are generated to support process optimization and preventive maintenance of equipment.

[0046] By aggregating inspection data from all production lines and using time-series correlation analysis, the originating production line and the cause of board malfunctions can be located, enabling root cause tracing and risk prediction for board malfunctions across the entire production line.

[0047] Beneficial effects:

[0048] This invention provides an intelligent early warning system for disordered gypsum board production lines based on edge computing. It has the following beneficial effects:

[0049] I. By integrating three multi-dimensional features—geometric deviation, morphological deformation, and operational sequence—to construct a discrimination criterion, and combining an adaptive dynamic threshold with a four-level classification discrimination model, it accurately distinguishes between normal process tolerances, thermal warping deformation, and genuine disordered boards of different grades. This replaces the traditional single threshold detection method, significantly reducing false alarms and false negatives, and improving the detection accuracy and scenario adaptability of disordered boards in gypsum board production.

[0050] Second, by using an edge computing architecture, the lightweight model is deployed locally on the production line. Data acquisition, feature extraction, and fault diagnosis are all executed locally, with single-frame inference latency controlled at the millisecond level. This perfectly matches the cycle time requirements of a 100-meter-level high-speed gypsum board production line, avoiding the transmission latency bottleneck of centralized cloud processing. Combined with a graded intervention and anti-vibration smoothing mechanism, differentiated handling is performed for different fault levels, which can curb the expansion of faults while reducing unnecessary downtime losses and ensuring the stability of gypsum board production.

[0051] Third, by constructing a closed-loop self-optimizing architecture that integrates edge and cloud, difficult samples and manually reviewed samples are collected at the edge, and the model is continuously iterated and optimized in the cloud based on incremental learning. This allows the system to adapt to the drift of detection benchmarks caused by high dust environments and equipment wear, ensuring that the system's accuracy does not decrease over long-term operation. After data from the entire production line is aggregated, multi-dimensional statistical analysis and root cause tracing can be performed to support process optimization and preventive maintenance of equipment, thereby reducing the failure rate of disordered gypsum board production from the source. Attached Figure Description

[0052] Figure 1A flowchart illustrating the operation method of an intelligent early warning system for disordered gypsum board production line based on edge computing, provided by the present invention.

[0053] Figure 2 This invention provides a schematic diagram of a module for an intelligent early warning system for disordered gypsum board production lines based on edge computing. Detailed Implementation

[0054] 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.

[0055] Example:

[0056] Please see Figures 1-2 An intelligent early warning system for disordered boards in a gypsum board production line based on edge computing, comprising the following components:

[0057] The multi-source sensing and acquisition module synchronously collects board image data, spatial location data and production line operation parameter data in each section of the gypsum board production line, and constructs a standardized detection dataset after performing front-end data preprocessing.

[0058] The edge computing processing module extracts geometric deviation features, morphological deformation features, and runtime sequence features of the board based on the standardized detection dataset to construct a multi-dimensional disordered board feature vector. The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state, and outputs the corresponding level of early warning signal.

[0059] The edge-cloud collaborative iteration module uploads problematic samples with a confidence level below a set threshold at the edge and false positives and false negatives marked by manual verification to the cloud service platform. After the cloud completes model optimization training based on incremental samples, it sends the updated model to each edge node to achieve closed-loop self-optimization of the discrimination model.

[0060] The early warning and linkage execution module, based on the graded early warning signal, will link the corresponding production equipment to perform graded intervention operations, and at the same time synchronize the disordered board fault data to the production management system and operation and maintenance terminal.

[0061] Simultaneously collect image data, spatial location data, and production line operating parameter data of the gypsum board production line at each stage. After performing front-end data preprocessing, construct a standardized inspection dataset. The specific steps include:

[0062] S11. Detection nodes are deployed in four core sections: wet board conveying section, drying kiln outlet section, cutting and separating section, and stacking and transfer section. Connect the linear array industrial camera, laser rangefinder array and roller speed encoder to collect board contour image data, board height data and production line real-time speed data respectively.

[0063] Specifically, in step S11, four core sections cover the key nodes of the entire process of gypsum board production, from molding wet boards to demolding, drying and shaping, cutting to length, and stacking of finished products. Each detection node is equipped with two 4096-pixel line array industrial cameras, arranged laterally on both sides of the roller conveyor. The frame rate is linked to the production line speed. When the line speed is 80m / min, the corresponding frame rate is 40 frames / second, ensuring that at least three complete end face and side contour images are captured for each standard 3m long gypsum board. A laser ranging array is evenly arranged with eight measuring points along the roller conveyor, with a sampling frequency of 200Hz and a ranging accuracy of ±0.3mm, used to collect board surface height and shape data. The roller conveyor speed encoder is installed on the main shaft of the drive roller, with a pulse resolution of 1000P / R, to collect the roller conveyor running speed in real time. All sensing devices are connected to the edge computing gateway of the corresponding section. The gateway adopts an ARM architecture, with a local computing power of 12TOPS, and supports multi-source data synchronous acquisition and local real-time computing.

[0064] S12. Perform dust occlusion repair, illumination equalization and edge enhancement processing on the collected board image data to suppress image noise caused by high dust and illumination fluctuations in the production site;

[0065] Specifically, in step S12, the gypsum board production site generally suffers from problems such as gypsum dust dispersion and uneven workshop ceiling illumination, which can easily lead to image blurring and insufficient edge contrast. To address image fogging caused by dust obstruction, a dark channel prior algorithm is used for image defogging to restore the edge details of the board. To address grayscale deviations caused by workshop illumination fluctuations, a contrast-limited adaptive histogram equalization algorithm is used for illumination correction to control the image grayscale standard deviation within ±5. Finally, the Canny operator is used to strengthen the board contour edges, achieving an edge positioning accuracy of up to 1 pixel, effectively improving the accuracy of subsequent geometric feature extraction.

[0066] S13. Perform vibration filtering and outlier removal processing on the collected laser ranging data, and complete the spatiotemporal registration of the plate position by combining the roller speed data.

[0067] Specifically, in step S13, the mechanical vibration of the production line roller conveyor during operation will cause high-frequency fluctuations and burrs in the laser ranging data. At the same time, the start-up and shutdown of the equipment and the joints of the panels will cause jump anomalies. First, the Kalman filter algorithm is used to smooth the original ranging data, filtering out the high-frequency noise caused by vibration, and the data fluctuation can be reduced by more than 70%. Then, the 3σ criterion is used to identify and remove jump anomalies to avoid abnormal data interfering with the subsequent morphological feature calculation. Finally, combined with the real-time speed data of the roller conveyor speed encoder, the ranging data at different times and different measuring points are spatially aligned according to the distance traveled by the panel, and the spatiotemporal registration of the height data of the entire surface of a single gypsum board is completed, eliminating the positional deviation caused by speed fluctuations.

[0068] S14. Align the preprocessed image data, ranging data, and working condition data with a unified timestamp, encapsulate them into a standardized format, and output a standardized detection dataset.

[0069] Specifically, in step S14, the system has a built-in clock synchronization generator that uses the sampling trigger signal of the linear industrial camera as a reference to synchronously latch the laser ranging data and roller speed data at the current moment when generating each frame of image data, and gives all data a unified millisecond-level timestamp; the preprocessed image feature data, ranging array data, and working condition speed data are mapped one-to-one according to the timestamp and encapsulated into a unified JSON format data structure to build a standardized detection dataset, providing a data foundation with unified format and time alignment for subsequent multi-dimensional feature extraction.

[0070] Based on a standardized inspection dataset, geometric deviation features, morphological deformation features, and runtime sequence features of the board are extracted to construct a multi-dimensional disordered board feature vector. The specific steps include:

[0071] S21. Based on the plate edge image and the preset position baseline, calculate the plate's lateral offset, longitudinal tilt angle, adjacent plate spacing deviation, and inter-column misalignment of multiple conveyor columns, and extract geometric deviation features.

[0072] Specifically, in step S21, the center line of the roller conveyor is used as the preset position reference line. The physical distance is calculated by converting the pixel coordinates of the two sides of the plate. The distance difference between the left and right edges of the plate and the reference line is calculated to obtain the lateral offset of the plate. The calculation accuracy can reach ±0.5mm. The longitudinal tilt angle is calculated by the ratio of the lateral offset difference between the front and rear ends of the plate to the plate length. The angle resolution can reach 0.05°. Combining the time difference of the end faces of two adjacent plates and the production line speed, the difference between the actual plate spacing and the standard plate spacing is calculated to obtain the spacing deviation between adjacent plates. For production lines with double or multiple columns of conveyor, the difference in lateral offset of each column of plates is calculated to obtain the misalignment between columns. Finally, 16-dimensional geometric deviation features are extracted to comprehensively characterize the spatial position and posture deviation of the plate.

[0073] S22. Based on the multi-point height data of the laser ranging array, calculate the maximum warping height of the board surface, the overall flatness deviation, and the size of the edge defects, and extract the morphological deformation features.

[0074] Specifically, in step S22, using the standard thickness of the corresponding specification gypsum board as the benchmark value, the difference between the highest point of the board surface and the benchmark value is calculated based on the 8-point height data of the laser ranging array to obtain the maximum warping height of the board surface; the standard deviation of all measuring point height data is calculated to obtain the overall flatness deviation, reflecting the degree of uniform deformation of the board surface; by identifying the height change area at the edge of the board, the length and depth dimensions of the defect area are calculated to obtain the edge defect characteristics, and the minimum identifiable defect is 2mm; finally, 24-dimensional morphological deformation features are extracted to accurately depict the surface morphology and structural defects of the board.

[0075] S23. Retrieve plate position data from multiple consecutive sampling periods, calculate the offset change rate, position fluctuation trend, and matching degree between plate speed and roller speed, and extract runtime sequence features.

[0076] Specifically, in step S23, the lateral offset data of the plate body for 20 consecutive sampling periods are retrieved, and the rate of change of the offset with time is calculated by linear fitting to obtain the rate of change of the offset, reflecting the development trend of the offset; by analyzing the position data fluctuation amplitude of nearly 100 periods, the position fluctuation trend is obtained, and occasional fluctuations and continuous offsets are distinguished; the ratio of the actual traveling speed of the plate body to the set speed of the roller conveyor is calculated to obtain the matching degree between the plate speed and the roller conveyor speed, and to identify early abnormalities such as plate slippage and jamming; finally, 24-dimensional running sequence features are extracted to predict the risk of plate disorder from the dynamic trend dimension.

[0077] S24. Combine the geometric deviation features, morphological deformation features, and runtime sequence features to generate a multi-dimensional random board feature vector.

[0078] Specifically, in step S24, the 16-dimensional geometric deviation features, 24-dimensional morphological deformation features, and 24-dimensional runtime sequence features are sequentially concatenated to generate a 64-dimensional multi-dimensional disordered board feature vector; and the Min-Max normalization method is used to uniformly map all feature values ​​to the [0,1] interval to eliminate the influence of differences in the dimensions of different features on subsequent model discrimination; multi-dimensional feature fusion can comprehensively characterize the spatial position, surface morphology, and running trend of the board, avoiding one-sided misjudgment caused by single feature discrimination.

[0079] The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state. The adaptive threshold calibration mechanism in the corresponding level of warning signal is output, and the following steps are specifically executed:

[0080] S311. Receive the current gypsum board thickness specifications, length specifications, and real-time operating speed parameters of the production line from the production system.

[0081] Specifically, in step S311, the adaptive threshold calibration unit connects to the production MES system via industrial Ethernet to receive the thickness, length specifications, and production line speed parameters of the currently produced gypsum board in real time. The parameter update delay does not exceed 200ms, ensuring synchronization with the production cycle and realizing real-time linkage between the threshold and the production conditions.

[0082] S312. Call the pre-stored baseline tolerance library and match the normal offset threshold, warpage allowable threshold, and disordered board judgment threshold corresponding to the current board type and line speed.

[0083] Specifically, in step S312, the system has a pre-stored reference tolerance library covering all types of gypsum board. The library stores the normal offset threshold, warping allowable threshold, and disordered board judgment threshold corresponding to different board types and different line speeds. For example, for a 12mm thick, 3m long gypsum board at a line speed of 80m / min, the normal lateral offset threshold is 5mm, the warping allowable threshold is 2mm, the slight disordered board offset range is 5mm-15mm, and the severe disordered board offset threshold is above 15mm. After receiving the current production parameters, the corresponding reference threshold is quickly retrieved by table lookup matching as the initial reference for state judgment.

[0084] S313. Correct the judgment threshold based on the ambient temperature and humidity and the equipment wear compensation coefficient, and output the dynamic judgment threshold under the current working condition to the lightweight disordered board classification and discrimination model.

[0085] Specifically, in step S313, the on-site temperature and humidity sensors and equipment runtime statistics are connected, and the temperature and humidity compensation coefficient and equipment wear compensation coefficient are built in. When the ambient temperature and humidity deviate from the standard working conditions, the warpage allowable threshold is slightly corrected to offset the normal deformation interference caused by temperature and humidity changes. When the cumulative running time of the roller conveyor reaches the wear threshold, the offset judgment threshold is appropriately compensated to offset the reference drift caused by equipment wear. Finally, a dynamic judgment threshold that fits the actual working conditions on site is output, effectively reducing misjudgments caused by environmental and equipment aging.

[0086] S314. When changing production lines, the reference parameters of the corresponding board type are automatically switched without manual recalibration.

[0087] Specifically, in step S314, when the production line executes a changeover instruction to switch gypsum board specifications, the adaptive threshold calibration unit automatically receives the new board type parameters issued by the MES system and automatically switches the corresponding board type's reference tolerance library. No manual on-site calibration and debugging are required, and the parameter switching time does not exceed 1 second. This adapts to the flexible production needs of multiple varieties and significantly shortens the changeover debugging time.

[0088] The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state. The corresponding warning signal is output. The state discrimination processing of the lightweight disordered board classification and discrimination model includes the following steps:

[0089] S321. Input the multi-dimensional randomized feature vector into the shared feature extraction layer of the model, extract the spatial features corresponding to geometric deviation and shape deformation through a multi-layer one-dimensional convolutional neural network, extract the dynamic trend features corresponding to the runtime sequence through a long short-term memory network, and output a high-dimensional hidden layer feature vector after fusion.

[0090] Specifically, in step S321, the lightweight random board classification and discrimination model adopts a lightweight architecture of "one-dimensional convolutional neural network + long short-term memory network". The 64-dimensional input feature vector first passes through two layers of one-dimensional convolutional layers to extract local spatial features corresponding to geometric deviations and morphological deformations. Then, it is connected to a long short-term memory network layer containing 64 hidden units to extract dynamic trend features corresponding to the runtime sequence. The two types of features are spliced ​​and fused to output a 128-dimensional high-dimensional hidden layer feature vector. After INT8 quantization, the model is deployed on an edge computing gateway. The inference latency of a single feature vector is ≤15ms, which meets the real-time requirements of high-speed production lines.

[0091] S322. Input the high-dimensional hidden layer feature vector into the classification fully connected layer, and after nonlinear mapping, normalize it through the Softmax function to output the confidence scores corresponding to the four states: normal process tolerance, thermal warpage deformation, slight board disorder, and severe board disorder.

[0092] Specifically, in step S322, the 128-dimensional high-dimensional hidden layer feature vector is input into two fully connected classification layers. After nonlinear mapping by the ReLU activation function, it is connected to the Softmax function for normalization processing, and the confidence scores of four states are output: normal process tolerance, thermal warpage deformation, slight board disorder, and severe board disorder. The sum of all scores is 1, realizing the probabilistic output of the four states.

[0093] S323. Obtain the dynamic judgment threshold output by the adaptive threshold calibration unit, compare the confidence scores of the four states with the corresponding thresholds respectively, and filter out the state categories with confidence scores higher than the judgment thresholds.

[0094] Specifically, in step S323, the dynamic judgment thresholds of the four states output by the adaptive threshold calibration unit are obtained, and the confidence scores of the four states output by the model are compared with the corresponding thresholds respectively. If the confidence score of a certain state is lower than the corresponding judgment threshold, the state category is excluded, and only the states with qualified confidence scores are retained as candidates. Low confidence judgment results are filtered by threshold screening to further reduce the probability of misjudgment.

[0095] S324. Among the state categories that meet the threshold requirements, select the category with the highest confidence as the final judgment result and output the corresponding warning signal.

[0096] Specifically, in step S324, among the candidate state categories filtered by the threshold, the category with the highest confidence is selected as the final judgment result; the normal process tolerance state outputs no warning prompt, the thermal warping deformation state outputs a process prompt level signal, the slightly disordered board state outputs a warning level signal, and the severely disordered board state outputs an emergency alarm level signal; different levels correspond to different subsequent handling strategies to achieve hierarchical and precise control.

[0097] By integrating three multi-dimensional features—geometric deviation, morphological deformation, and runtime sequence—to construct a discrimination criterion, and combining an adaptive dynamic threshold with a four-level classification discrimination model, it accurately distinguishes between normal process tolerances, thermal warping deformation, and genuine disordered boards of different grades. This replaces the traditional single threshold detection method, significantly reducing false alarms and false negatives, and improving the detection accuracy and scenario adaptability of disordered boards in gypsum board production.

[0098] Difficult samples with a discrimination confidence level below a set threshold at the edge nodes, along with false positives and false negatives marked by manual verification, are uploaded to the cloud service platform. After the cloud platform completes model optimization training based on incremental samples, it distributes the updated model to each edge node, realizing closed-loop self-optimization of the discrimination model. The specific steps include:

[0099] S41. The edge terminal calculates the classification confidence score for each discrimination result, and automatically marks the samples with the highest confidence score below the set threshold as difficult samples and stores them in the local cache. At the same time, it receives false alarm samples and missed alarm samples marked by maintenance personnel through the field terminal.

[0100] Specifically, in step S41, the edge device calculates the highest classification confidence score for each state discrimination result, sets the confidence score threshold to 0.7, and automatically marks samples with the highest confidence score below 0.7 as difficult samples and stores them in the local cache queue. At the same time, on-site maintenance personnel can review historical warning records through touch terminals, mark false alarms and missed alarms, and add them to the queue of samples to be uploaded. The sample data includes the corresponding feature vector, discrimination result, real label and working condition information to ensure that the sample data is complete and traceable.

[0101] S42. According to a fixed cycle, upload the difficult samples, manually reviewed samples, along with the corresponding process information, environmental temperature and humidity, plate type and specifications, and production line speed and working condition data to the cloud service platform in encrypted form.

[0102] Specifically, in step S42, the system selects the off-peak period of the production line in the early morning of each day and automatically uploads the accumulated difficult samples and manually reviewed samples, along with the corresponding work section, ambient temperature and humidity, plate type and specifications, production line speed and other working condition data, to the cloud service platform via industrial internet encryption. The method of uploading during off-peak hours does not occupy the bandwidth during production hours, avoids affecting the normal operation of the production system, and the data volume of a single upload is ≤300MB. The transmission process uses SSL encryption to ensure data security.

[0103] S43. Based on the current online model, the cloud uses incremental learning to fine-tune and optimize only the weights of the model's classification layer. After the overall accuracy of the validation set meets the standard, an updated lightweight model is generated.

[0104] Specifically, in step S43, the cloud uses the currently running online model as a basis and adopts an incremental learning method to fine-tune and optimize only the weights of the fully connected classification layer of the model, without modifying the weights of the underlying feature extraction network, to avoid catastrophic forgetting and retain the general features that the model has learned. After training, the model is tested on a dedicated validation set. When the overall recognition accuracy is ≥98%, it is judged to have met the standard, and an updated version of the INT8 lightweight model is generated to ensure that the updated model has both accuracy and performance.

[0105] S44. The updated model is distributed to all edge computing nodes via remote upgrade. The nodes complete the model update using hot loading, without interrupting the normal testing operations of the production line.

[0106] Specifically, in step S44, the updated lightweight model is distributed to all edge computing nodes on the production line via OTA remote upgrade; the nodes use hot loading to complete the model switching, with a switching latency of ≤1s, without interrupting the normal testing business of the production line, achieving seamless upgrade; through incremental iteration of edge-cloud collaboration, the model can adapt to changes in the field environment and equipment aging drift, maintaining high detection accuracy in the long term.

[0107] By deploying a lightweight model locally on the production line through an edge computing architecture, the entire process of data acquisition, feature extraction, and fault diagnosis is executed locally, with single-frame inference latency controlled at the millisecond level. This perfectly matches the cycle time requirements of a 100-meter-level high-speed gypsum board production line, avoiding the transmission latency bottleneck of centralized cloud processing. Coupled with a graded intervention and anti-vibration smoothing mechanism, differentiated handling is performed for different fault levels, which can curb the expansion of faults while reducing unnecessary downtime losses and ensuring the stability of gypsum board production.

[0108] Based on the tiered early warning signal, the corresponding production equipment is linked to execute tiered intervention operations. At the same time, the disordered board fault data is synchronized to the tiered intervention operations in the production management system and the operation and maintenance terminal. The specific logic is as follows:

[0109] When the condition is determined to be thermal warping deformation, a temperature zone fine-tuning command is sent to the drying kiln temperature control system to optimize the temperature field distribution without changing the production line operating speed.

[0110] When the board is determined to be in a state of slight board disorder, the corresponding section roller conveyor will be activated to reduce its speed. At the same time, the pneumatic guide correction mechanism will be triggered to perform correction operation. After the board posture returns to normal, the original running speed will be automatically restored.

[0111] When a severe board disorder is detected, the corresponding emergency stop command for the work section is immediately triggered, the on-site audible and visual alarm device is activated, and the fault location and fault type are pushed to the handheld terminal of the on-site maintenance personnel.

[0112] Specifically, in the hierarchical intervention logic of the embodiment, differentiated handling strategies are implemented for different levels of abnormal states. When the state is determined to be thermal warping deformation, a temperature fine-tuning command for the corresponding temperature zone is sent to the DCS temperature control system of the drying kiln, with an adjustment range of ±3℃. This optimizes the temperature field distribution to alleviate thermal warping of the plates without changing the production line's operating speed, ensuring production efficiency. When the state is determined to be mild plate disorder, the corresponding section's roller conveyor is reduced to 70% of its original speed, and the pneumatic guide correction mechanism on both sides is activated to correct the plate's posture. After the plate offset returns to the normal range for five consecutive detection cycles, the original operating speed is automatically restored, minimizing the impact on production while curbing the expansion of the fault. When the state is determined to be severe plate disorder, a safety emergency stop command for the corresponding section is immediately triggered, and the on-site audible and visual alarm is activated, emitting an alarm sound of over 85dB and flashing warning lights. The fault location and fault type information are pushed to the handheld terminal of the maintenance personnel, reminding them to quickly arrive at the scene to handle the situation and avoid equipment damage and prolonged downtime.

[0113] Based on the tiered early warning signal, corresponding production equipment is linked to execute tiered intervention operations. Simultaneously, the disordered board fault data is synchronized to the production management system and maintenance terminal. Before executing the tiered intervention operation, a smoothing process to prevent severe vibrations is also included, specifically:

[0114] Read the current value of the device control signal from the previous control cycle, and calculate the difference between the target control signal value corresponding to the action command and the current value;

[0115] If the difference exceeds the preset maximum adjustment step size limit, the control signal issued this time will be forcibly truncated to the maximum adjustment step size limit, and the value will be gradually adjusted to the target value according to the step size gradient.

[0116] If the difference does not exceed the preset maximum adjustment step size limit, the target control signal corresponding to the action command will be sent directly to the execution device.

[0117] Specifically, in the anti-severe vibration smoothing process of the embodiment, in order to avoid secondary displacement of the plate and equipment impact damage caused by sudden changes in control parameters, smoothing is performed before the action command is issued; the current value of the equipment control signal of the previous control cycle is read, and the difference between the target value corresponding to the current action command and the current value is calculated; a maximum single adjustment step size is set, for example, the single adjustment step size of the roller speed does not exceed 5m / min, and the single adjustment step size of the temperature does not exceed 2℃; if the difference exceeds the set maximum single adjustment step size, the control signal issued this time is forcibly truncated to the maximum step size limit, and gradually adjusted to the target value in multiple cycles; if the difference does not exceed the maximum single adjustment step size, the target control signal is directly issued; through smoothing, production fluctuations caused by sudden changes in parameters are effectively avoided, ensuring the stable operation of the equipment.

[0118] Based on the tiered early warning signals, corresponding production equipment is linked to execute tiered intervention operations. Simultaneously, the disordered board fault data is synchronized to the production management system and maintenance terminal, including statistical analysis and risk tracing steps, specifically:

[0119] Statistics on the occurrence rate of disordered boards, the proportion of fault types, and the false alarm and missed alarm rates are calculated according to the dimensions of work section, shift, and board type, and multi-dimensional analysis reports are generated to support process optimization and preventive maintenance of equipment.

[0120] By aggregating inspection data from all production lines and using time-series correlation analysis, the originating production line and the cause of board malfunctions can be located, enabling root cause tracing and risk prediction for board malfunctions across the entire production line.

[0121] Specifically, in the statistical analysis and risk tracing of the embodiment, all early warning events and handling results are simultaneously uploaded to the production MES system. The system automatically calculates the occurrence rate of disordered boards, the proportion of fault types, and the false alarm and missed alarm rates by daily, weekly, monthly, work section, shift, and board type, generating multi-dimensional visual analysis reports. At the same time, it gathers the detection data of the entire work section and locates the starting work section and root cause of disordered board failures through time-series correlation analysis. For example, it identifies that the warping of the drying kiln outlet is the main cause of disordered boards in the subsequent cutting section, supporting the optimization of production processes and preventive maintenance of equipment, and reducing the probability of disordered board failures from the source.

[0122] By constructing a closed-loop self-optimizing architecture that integrates edge and cloud, difficult samples and manually reviewed samples are collected at the edge, and the model is continuously iterated and optimized in the cloud based on incremental learning. This allows the system to adapt to the drift of detection benchmarks caused by high dust environments and equipment wear, ensuring that the system's accuracy does not decrease over long-term operation. After data from the entire production line is aggregated, multi-dimensional statistical analysis and root cause tracing can be performed to support process optimization and preventive maintenance of equipment, thereby reducing the incidence of disordered gypsum board production failures from the source.

[0123] For example, taking a gypsum board production line with an annual output of 25 million square meters as an example, producing gypsum boards with a thickness of 12mm and a length of 3m, the production line operates at a speed of 80m / min. The detection node at the outlet of the drying kiln collects the board data in real time. After feature extraction and model discrimination, it is found that the board warping height reaches 5mm but the lateral offset is only 3mm. The time-series characteristics do not show a continuous deterioration trend, and it is determined to be a state of thermal warping deformation. The system sends a fine-tuning command to the temperature control system of the drying kiln to reduce the temperature of the corresponding zone by 2℃. After 8 minutes, the board warping height drops back to within 1.8mm, restoring the normal process tolerance range, and the production line speed is not reduced throughout the process. If the cutting and slab segment detection node detects a lateral offset of 9mm and a tilt angle of 1.1° for the slab, with no stacking or jamming, it is determined to be a minor scrambling condition. The system immediately activates the roller conveyor for that section, lowering its speed to 56m / min and simultaneously initiating a pneumatic guide correction mechanism for attitude correction. After 2.5 seconds, the slab offset falls back to within 4mm, the attitude returns to normal, and the roller conveyor automatically resumes its original speed of 80m / min. If the stacking and transfer section detection node detects a slab offset of 18mm and signs of stacking, it is determined to be a severe scrambling condition. The system immediately triggers an emergency stop for that section, simultaneously activating on-site audible and visual alarms and pushing fault information to the maintenance terminal. Maintenance personnel arrive within 4 minutes to resolve the fault, quickly restoring production and minimizing downtime losses.

[0124] 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. An intelligent early warning system for disordered boards in a gypsum board production line based on edge computing, characterized in that, The system includes the following: The multi-source sensing and acquisition module synchronously collects board image data, spatial location data and production line operation parameter data in each section of the gypsum board production line, and constructs a standardized detection dataset after performing front-end data preprocessing. The edge computing processing module extracts geometric deviation features, morphological deformation features and runtime sequence features of the board based on the standardized detection dataset, and constructs a multi-dimensional disordered board feature vector. The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state, and outputs the corresponding level of warning signal. The edge-cloud collaborative iteration module uploads the difficult samples with the discrimination confidence level below the set threshold and the false positive and false negative samples marked by manual review to the cloud service platform. After the cloud completes the model optimization training based on the incremental samples, it sends the updated model to each edge node to realize the closed-loop self-optimization of the discrimination model. The early warning and linkage execution module, based on the graded early warning signal, will link the corresponding production equipment to perform graded intervention operations, and at the same time synchronize the disordered board fault data to the production management system and operation and maintenance terminal.

2. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, Simultaneously collect image data, spatial location data, and production line operating parameter data of the gypsum board production line at each stage. After performing front-end data preprocessing, construct a standardized inspection dataset. The specific steps include: S11. Detection nodes are deployed in four core sections: wet board conveying section, drying kiln outlet section, cutting and separating section, and stacking and transfer section. Connect the linear array industrial camera, laser rangefinder array and roller speed encoder to collect board contour image data, board height data and production line real-time speed data respectively. S12. Perform dust occlusion repair, illumination equalization and edge enhancement processing on the collected board image data to suppress image noise caused by high dust and illumination fluctuations in the production site; S13. Perform vibration filtering and outlier removal processing on the collected laser ranging data, and complete the spatiotemporal registration of the plate position by combining the roller speed data. S14. Align the preprocessed image data, ranging data, and working condition data with a unified timestamp, encapsulate them into a standardized format, and output a standardized detection dataset.

3. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, Based on the standardized detection dataset, geometric deviation features, morphological deformation features, and runtime sequence features of the board are extracted to construct a multi-dimensional disordered board feature vector, which specifically includes the following steps: S21. Based on the plate edge image and the preset position baseline, calculate the plate's lateral offset, longitudinal tilt angle, adjacent plate spacing deviation, and inter-column misalignment of multiple conveyor columns, and extract geometric deviation features. S22. Based on the multi-point height data of the laser ranging array, calculate the maximum warping height of the board surface, the overall flatness deviation, and the size of the edge defects, and extract the morphological deformation features. S23. Retrieve plate position data from multiple consecutive sampling periods, calculate the offset change rate, position fluctuation trend, and matching degree between plate speed and roller speed, and extract runtime sequence features. S24. Combine the geometric deviation features, morphological deformation features, and runtime sequence features to generate a multi-dimensional random board feature vector.

4. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state. The adaptive threshold calibration mechanism in the corresponding level of warning signal is output, and the following steps are specifically executed: S311. Receive the current gypsum board thickness specifications, length specifications, and real-time operating speed parameters of the production line from the production system. S312. Call the pre-stored baseline tolerance library and match the normal offset threshold, warpage allowable threshold, and disordered board judgment threshold corresponding to the current board type and line speed. S313. Correct the judgment threshold based on the ambient temperature and humidity and the equipment wear compensation coefficient, and output the dynamic judgment threshold under the current working condition to the lightweight disordered board classification and discrimination model. S314. When changing production lines, the reference parameters of the corresponding board type are automatically switched without manual recalibration.

5. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, The multi-dimensional disordered board feature vector is input into the lightweight disordered board classification and discrimination model. Combined with the adaptive threshold calibration mechanism, it distinguishes four states: normal process tolerance state, thermal warpage deformation state, mild disordered board state, and severe disordered board state. The corresponding level of warning signal is output. The state discrimination processing of the lightweight disordered board classification and discrimination model includes the following steps: S321. Input the multi-dimensional randomized feature vector into the shared feature extraction layer of the model, extract the spatial features corresponding to geometric deviation and shape deformation through a multi-layer one-dimensional convolutional neural network, extract the dynamic trend features corresponding to the runtime sequence through a long short-term memory network, and output a high-dimensional hidden layer feature vector after fusion. S322. Input the high-dimensional hidden layer feature vector into the classification fully connected layer, and after nonlinear mapping, normalize it through the Softmax function to output the confidence scores corresponding to the four states: normal process tolerance, thermal warpage deformation, slight board disorder, and severe board disorder. S323. Obtain the dynamic judgment threshold output by the adaptive threshold calibration unit, compare the confidence scores of the four states with the corresponding thresholds respectively, and filter out the state categories with confidence scores higher than the judgment thresholds. S324. Among the state categories that meet the threshold requirements, select the category with the highest confidence as the final judgment result and output the corresponding warning signal.

6. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, Difficult samples with a discrimination confidence level below a set threshold at the edge nodes, along with false positives and false negatives marked by manual verification, are uploaded to the cloud service platform. After the cloud platform completes model optimization training based on incremental samples, it distributes the updated model to each edge node, realizing closed-loop self-optimization of the discrimination model. The specific steps include: S41. The edge terminal calculates the classification confidence score for each discrimination result, and automatically marks the samples with the highest confidence score below the set threshold as difficult samples and stores them in the local cache. At the same time, it receives false alarm samples and missed alarm samples marked by maintenance personnel through the field terminal. S42. According to a fixed cycle, upload the difficult samples, manually reviewed samples, along with the corresponding process information, environmental temperature and humidity, plate type and specifications, and production line speed and working condition data to the cloud service platform in encrypted form. S43. Based on the current online model, the cloud uses incremental learning to fine-tune and optimize only the weights of the model's classification layer. After the overall accuracy of the validation set meets the standard, an updated lightweight model is generated. S44. The updated model is distributed to all edge computing nodes via remote upgrade. The nodes complete the model update using hot loading, without interrupting the normal testing operations of the production line.

7. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, Based on the tiered early warning signal, the corresponding production equipment is linked to execute tiered intervention operations. At the same time, the disordered board fault data is synchronized to the tiered intervention operations in the production management system and the operation and maintenance terminal. The specific logic is as follows: When the condition is determined to be thermal warping deformation, a temperature zone fine-tuning command is sent to the drying kiln temperature control system to optimize the temperature field distribution without changing the production line operating speed. When the board is determined to be in a state of slight board disorder, the corresponding section roller conveyor will be activated to reduce its speed. At the same time, the pneumatic guide correction mechanism will be triggered to perform correction operation. After the board posture returns to normal, the original running speed will be automatically restored. When a severe board disorder is detected, the corresponding emergency stop command for the work section is immediately triggered, the on-site audible and visual alarm device is activated, and the fault location and fault type are pushed to the handheld terminal of the on-site maintenance personnel.

8. The intelligent early warning system for disordered gypsum board production line based on edge computing according to claim 1, characterized in that, Based on the tiered early warning signal, corresponding production equipment is linked to execute tiered intervention operations. Simultaneously, the disordered board fault data is synchronized to the production management system and maintenance terminal. Before executing the tiered intervention operation, a smoothing process to prevent severe vibrations is also included, specifically: Read the current value of the device control signal from the previous control cycle, and calculate the difference between the target control signal value corresponding to the action command and the current value; If the difference exceeds the preset maximum adjustment step size limit, the control signal issued this time will be forcibly truncated to the maximum adjustment step size limit, and the value will be gradually adjusted to the target value according to the step size gradient. If the difference does not exceed the preset maximum adjustment step size limit, the target control signal corresponding to the action command will be sent directly to the execution device.

9. The intelligent early warning system for disordered boards in a gypsum board production line based on edge computing according to claim 1, characterized in that, Based on the tiered early warning signals, corresponding production equipment is linked to execute tiered intervention operations. Simultaneously, the disordered board fault data is synchronized to the production management system and maintenance terminal, including statistical analysis and risk tracing steps, specifically: Statistics on the occurrence rate of disordered boards, the proportion of fault types, and the false alarm and missed alarm rates are calculated according to the dimensions of work section, shift, and board type, and multi-dimensional analysis reports are generated to support process optimization and preventive maintenance of equipment. By aggregating inspection data from all production lines and using time-series correlation analysis, the originating production line and the cause of board malfunctions can be located, enabling root cause tracing and risk prediction for board malfunctions across the entire production line.

Images