Preventive maintenance and process parameter optimization methods for continuous stainless steel welded pipe production lines

Abstract

This invention discloses a method for preventive maintenance and process parameter optimization in a continuous stainless steel welded pipe production line. The method includes: synchronously collecting time-series operational data using a multi-source sensor array deployed at key nodes of the continuous stainless steel welded pipe production line; extracting and fusing multi-scale features from the time-series operational data to construct a high-dimensional feature vector characterizing equipment health status and process stability, and inputting this vector into a multi-step predictive model to output the remaining service life prediction curve of key components and the evolution trend of process deviations; generating a preventive maintenance decision scheme including maintenance priorities and time windows based on the remaining service life prediction curve and the evolution trend of process deviations; and generating a set of collaborative adjustment instructions for process parameters according to the preventive maintenance decision scheme. Using this invention, equipment reliability and process stability can be improved, forming a data-driven closed-loop intelligent maintenance and control system.

Description

Technical Field

[0001] This invention belongs to the field of welded pipe manufacturing technology, and in particular, it is a method for preventive maintenance and process parameter optimization of a continuous production line for stainless steel welded pipes. Background Technology

[0002] Continuous production of stainless steel welded pipes is a typical high-precision, long-process manufacturing system. Its operation is affected by a combination of factors, including welding technology, forming control, and equipment health. Currently, the industry mainly uses planned maintenance based on fixed cycles and manual experience-based parameter adjustment, making it difficult to detect equipment performance degradation and process parameter deviations in real time. Existing predictive maintenance methods mostly rely on single types of data and lack fusion analysis of multi-physics time-series signals, resulting in delayed fault warnings and crude control of process fluctuations. At the same time, equipment maintenance decisions and process optimization are often disconnected, failing to dynamically adjust production parameters within the maintenance window, which can easily lead to unplanned downtime and quality losses. Summary of the Invention

[0003] The purpose of this invention is to provide a preventive maintenance and process parameter optimization method for a continuous stainless steel welded pipe production line, in order to overcome the shortcomings of the existing technology, improve equipment reliability and process stability, and form a data-driven closed-loop intelligent maintenance and control system.

[0004] One embodiment of this application provides a method for preventive maintenance and process parameter optimization of a continuous stainless steel welded pipe production line, the method comprising: The timing operation data is collected synchronously by a multi-source sensor array deployed at key nodes of the stainless steel welded pipe continuous production line. The timing operation data includes welding current and voltage waveforms, forming roll gap displacement, pipe temperature field distribution, and equipment vibration spectrum. Multi-scale feature extraction and fusion are performed on the time-series operation data to construct a high-dimensional feature vector characterizing the health status of the equipment and the stability of the process. This vector is then input into a multi-step advanced prediction model, which outputs the remaining service life prediction curve of key components and the evolution trend of process deviation. Based on the remaining useful life prediction curve and the evolution trend of process deviation, the potential failure risk and quality defect probability are dynamically evaluated through the digital twin simulation module, and a preventive maintenance decision plan including maintenance priority and time window is generated. Based on the aforementioned preventive maintenance decision-making scheme, an adaptive dynamic programming algorithm is used to optimize welding power, roll gap servo pressure, and production line cycle time parameters in real time, generating a set of process parameter collaborative adjustment instructions to form a data-driven closed-loop optimization and maintenance control loop.

[0005] Another embodiment of this application provides a preventive maintenance and process parameter optimization system for a continuous stainless steel welded pipe production line, the system comprising: The acquisition module is used to synchronously acquire time-series operation data through a multi-source sensor array deployed at key nodes of the stainless steel welded pipe continuous production line. The time-series operation data includes welding current and voltage waveforms, forming roller gap displacement, pipe temperature field distribution, and equipment vibration spectrum. The construction module is used to perform multi-scale feature extraction and fusion on the time-series operation data, construct a high-dimensional feature vector characterizing the health status of the equipment and the stability of the process, and input it into the multi-step advanced prediction model to output the remaining service life prediction curve of key components and the evolution trend of process deviation. The evaluation module is used to dynamically assess potential failure risks and quality defect probabilities based on the remaining useful life prediction curve and process deviation evolution trend through the digital twin simulation module, and generate a preventive maintenance decision plan that includes maintenance priorities and time windows. The optimization module is used to optimize welding power, roll gap servo pressure and production line cycle time parameters in real time using an adaptive dynamic programming algorithm based on the preventive maintenance decision scheme, and generate a set of process parameter collaborative adjustment instructions to form a data-driven closed-loop optimization and maintenance control loop.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides a method for preventive maintenance and process parameter optimization of a continuous stainless steel welded pipe production line, which can improve equipment reliability and process stability, and form a data-driven closed-loop intelligent maintenance and control system. Attached Figure Description

[0009] Figure 1 Hardware structure block diagram of a computer terminal for a method of preventive maintenance and process parameter optimization for a continuous stainless steel welded pipe production line provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for preventive maintenance and process parameter optimization in a continuous stainless steel welded pipe production line, provided by an embodiment of the present invention. Figure 3 This is a schematic diagram of a preventive maintenance and process parameter optimization system for a continuous stainless steel welded pipe production line, provided in an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The present invention first provides a method for preventive maintenance and process parameter optimization of a continuous production line for stainless steel welded pipes. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a method of preventive maintenance and process parameter optimization in a continuous stainless steel welded pipe production line, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The present invention provides a method for preventive maintenance and process parameter optimization of a continuous stainless steel welded pipe production line, which may include the following steps: S201, synchronously collects time-series operation data by deploying a multi-source sensor array at key nodes of the stainless steel welded pipe continuous production line. The time-series operation data includes welding current and voltage waveforms, forming roller gap displacement, pipe temperature field distribution, and equipment vibration spectrum. Specifically, a multi-source sensor array can be designed and deployed, including Hall current and voltage sensors, laser displacement sensors, infrared thermal imager arrays, and vibration acceleration sensors. The sampling frequency and installation coordinate specifications for each sensor can be determined, and a sensor array configuration scheme can be generated. The core of this step is to achieve comprehensive capture of key operational data of the production line through scientific selection and precise layout, laying the foundation for subsequent time-series analysis. The specific implementation method is as follows: Sensor selection is strictly matched to the physical characteristics of the measured object: Hall current and voltage sensors are used to capture the dynamic electrical parameters of the welding power source, with a measurement range of current 0-500A and voltage 0-50V, accuracy ±0.5%, and response time ≤1μs, adapting to the rapid changes in the welding arc; laser displacement sensors focus on the minute displacement measurement of the forming roll gap, with a range of 0-20mm and a resolution of 0.001mm, and are not affected by the reflection of the pipe surface during sampling, ensuring accurate roll gap data; the infrared thermal imager array consists of 4 thermal imagers, with a temperature measurement range of -20℃ to 1500℃, a resolution of 640×480 pixels, and a temperature measurement accuracy of ±2%, covering temperature-sensitive areas such as the welding area and sizing area; the vibration acceleration sensor is a piezoelectric type, with a measurement range of 0-50g and a frequency response of 10Hz-10kHz, used to capture vibration signals of key components such as bearings and motors, and predict mechanical wear.

[0014] The sampling frequency is determined based on the signal change rate: due to frequent arc fluctuations, the sampling frequency of welding current and voltage signals is set to 10kHz (one set of data is collected every 0.1ms) to ensure the capture of transient characteristics such as arc extinction and reignition; the change of forming roll gap displacement signal is relatively gradual, so the sampling frequency is 1kHz (one set of data is collected every 1ms); the infrared thermal imager array sampling frequency is 50Hz (one frame of temperature field is collected every 20ms) to balance temperature resolution and data volume; the vibration acceleration sensor sampling frequency is 2kHz (one set of data is collected every 0.5ms) to cover the characteristic frequency range of mechanical faults.

[0015] The installation coordinate specifications are based on the production line's mechanical coordinate system (the origin is set at the center of the production line entrance, the X-axis is along the production direction, the Y-axis is horizontal and perpendicular to the production direction, and the Z-axis is vertically upward): Hall current and voltage sensors are connected in series at the welding power supply output, with installation coordinates X=3500mm, Y=0mm, Z=1200mm, and a distance of ≤500mm from the welding electrode to reduce the impact of line loss; laser displacement sensors are symmetrically installed on both sides of the forming roller, with coordinates X=2800mm, Y=±150mm, Z=800mm, and the lens is vertically pointed to the center of the roller gap, with a measurement distance of 120mm; an infrared thermal imager array is installed above the welding area in a 2×2 layout, with coordinates X=4200mm, Y=±200mm, Z=1500mm, and the lens is at a 45° angle to the pipe surface to avoid interference from strong light reflection; vibration acceleration sensors are attached to key parts such as the motor bearing housing and the forming roller shaft end, with coordinates X=1800mm, Y=0mm, Z=600mm (motor bearing housing), and the mounting surface is polished and cleaned to ensure vibration transmission efficiency.

[0016] The sensor array configuration plan is fixed in text form, which clearly defines the model parameters, installation coordinates, sampling frequency, calibration cycle (e.g., Hall sensor is calibrated once a month, laser displacement sensor is calibrated once every two weeks) and protection requirements (e.g., thermal imager is equipped with a dust cover, vibration sensor is equipped with a waterproof cover) of each sensor, to ensure that the deployment process is feasible and traceable.

[0017] According to the sensor array configuration scheme, all sensors are synchronously triggered to collect data through the Industrial Internet of Things protocol, and each data packet is timestamped at the microsecond level using a precise clock source to generate original synchronous data packets with a unified time tag. The core of this step is to solve the time synchronization problem of multi-sensor data and ensure the temporal consistency of the collected data. The specific implementation method is as follows: The Industrial IoT protocol chosen is ProfinetIO, which supports real-time data transmission with a cycle time of ≤1ms, adapting to the needs of synchronous data acquisition from multiple sensors. During protocol configuration, all sensors are set as slaves, and the production line main control unit is set as the master. The master sends synchronization trigger signals to the slaves via real-time Ethernet, with a transmission delay of ≤10μs, ensuring that all sensors start data acquisition at the same physical moment. The synchronization trigger mechanism employs a dual guarantee of "hardware trigger + software confirmation." After the master sends the trigger pulse, each sensor sends back a confirmation signal. The master begins data reception only after receiving all confirmation signals, avoiding data asynchrony caused by delays in individual sensors.

[0018] The precise clock source is built based on the IEEE 1588 Precision Time Protocol (PTP). The master clock is deployed in the main control unit with a clock accuracy of ≤10ns / day. Each sensor has a built-in slave clock module that synchronizes with the master clock via Ethernet with a synchronization period of 1 second, ensuring that the time deviation between the slave clock and the master clock is ≤5μs. The timestamp format is “YYYYMMDDHHMMSSssssss” (year-month-day-hour-minute-second-microsecond), for example, “20250501103000.123456”. The timestamp of each data packet is generated in real time by the sensor slave clock and recorded synchronously with the acquisition action to avoid deviations caused by later re-timed timestamps.

[0019] The original synchronization data packet structure contains five core fields: a unique sensor identifier (e.g., "Hall Sensor-H01", "Laser Displacement Sensor-L02"), a microsecond-level timestamp, a data type (e.g., "Current", "Displacement", "Temperature", "Vibration Acceleration"), a data value (including units, such as "205.3A", "0.852mm", "1180.5℃", "3.2g"), and a data status ("Normal", "Abnormal"). For example, the original data packet for the Hall sensor is "Sensor-H01_20250501103000.123456_Current_205.3A_Normal", and the data packet for the laser displacement sensor is "Sensor-L02_20250501103000.123456_Displacement_0.852mm_Normal". All data packets are transmitted to the data acquisition server in real time via industrial Ethernet. CRC-32 verification is enabled during transmission to ensure data integrity, with a loss rate of ≤0.01%.

[0020] The original synchronization data packets are preprocessed, wavelet threshold denoising algorithm is applied to eliminate arc interference in welding current, median filtering is used to smooth roll gap displacement signal, and non-uniform temperature field correction algorithm is used to process thermal imager data to generate preprocessed time sequence signal. The core of this step is to eliminate noise and distortion in the original data, improve data quality, and provide reliable input for subsequent feature extraction. The specific implementation method is as follows: The wavelet thresholding denoising algorithm is used to handle arc interference (high-frequency noise) in welding current. The db4 wavelet basis is selected (balancing denoising effect and computational efficiency), and the decomposition level is set to 3 levels (adapting to the frequency characteristics of the current signal). The threshold calculation uses an adaptive thresholding method, with the formula λ=σ×√(2×lnN), where σ is the noise standard deviation (estimated through the high-frequency coefficients after decomposition), and N is the data length. For example, if the length of a current data segment is N=10000 and σ=2.5A, then λ=2.5×√(2×ln10000)≈2.5×9.21≈23.03A. During the processing, the high-frequency coefficients after decomposition are subjected to soft thresholding (coefficients with absolute values ​​less than λ are set to zero, and coefficients with absolute values ​​greater than λ are subtracted from λ). Then, the signal is reconstructed through wavelet inverse transform. After denoising, the signal-to-noise ratio of the current signal is increased from 25dB to more than 35dB, and the glitches caused by arc interference are completely eliminated. For example, the original current signal fluctuates around 200A with ±5A, and after denoising, the fluctuation range is reduced to ±0.8A.

[0021] Median filtering is used to smooth mechanical vibration noise in roll gap displacement signals. The window size is set to 5 (an odd number of windows avoids signal offset). The filtering principle is to sort the 5 data points within the window by size and take the median value as the filtering result for the current data. For example, if the original roll gap displacement sequence is [0.852mm, 0.861mm, 0.855mm, 0.873mm, 0.858mm], the filtered result for the center data 0.855mm is the median value after sorting, which is 0.858mm. After filtering, the fluctuation amplitude of the signal is reduced from ±0.02mm to ±0.005mm, completely preserving the slow change trend of the roll gap without losing effective information.

[0022] The non-uniform temperature field correction algorithm is used to correct temperature deviations in infrared thermal imagers caused by lens distortion and environmental reflections. It employs a quadratic polynomial fitting method, with the correction model being T_corr = a × T_raw. 2 The formula is: +b×T_raw+c, where T_raw is the original temperature value, T_corr is the corrected temperature value, and a, b, and c are correction coefficients (obtained through standard blackbody furnace calibration, e.g., a=-0.0001, b=1.02, c=-5.2). For example, if the original temperature of a point in the welding area captured by the thermal imager is 1200℃, substituting it into the model, we get T_corr=-0.0001×1200℃. 2 +1.02×1200-5.2=-144+1224-5.2=1074.8℃. The temperature after correction deviates from the actual measured value (1075℃) by ≤0.1%, ensuring the absolute accuracy of the temperature field data is ≤±2%.

[0023] The preprocessed time-series signals are stored according to sensor type. Each signal is a continuous numerical sequence containing a timestamp and the processed data value. For example, the current time-series signal is “20250501103000.123456_205.3A,20250501103000.124456_204.9A,...”, which provides clean data for subsequent alignment and packaging.

[0024] The preprocessed time series signals are aligned and packaged according to a unified timestamp, and encapsulated into a multi-dimensional time series dataset containing current and voltage waveforms, roll gap displacement sequences, temperature field matrices, and vibration spectra. Finally, a synchronous time series running dataset that can be used for subsequent analysis is generated.

[0025] The core of this step is to achieve time unification and data integration of multi-source time-series signals to form a structured dataset. The specific implementation method is as follows: Timestamp alignment employs a "reference time axis + linear interpolation" strategy. Using a millisecond-level time reference from a precise clock source (each 1ms is a time step) as a unified standard, the preprocessed timing signals from each sensor are mapped to this reference axis. For signals with sampling frequencies higher than 1kHz (e.g., 10kHz current / voltage, 2kHz vibration), the average value is taken and integrated over each time step. For signals with sampling frequencies lower than 1kHz (e.g., a 50Hz temperature field), linear interpolation is used to supplement missing time step data, ensuring that each time step contains the corresponding data from all sensors. For example, a temperature field signal is acquired every 20ms, and 20 consecutive temperature field matrices are generated through interpolation within 20 time steps. The interpolated temperature changes are smooth with no obvious jumps.

[0026] Data is packaged and integrated according to the "time step-data type" structure. Each time step contains four core data categories: current and voltage waveforms (an array of 10 consecutive sampling points, corresponding to signal changes within 1 ms), roll gap displacement sequence (a single value representing the current roll gap size), temperature field matrix (a 640×480 pixel two-dimensional array, where each element is the corrected temperature of the corresponding pixel), and vibration spectrum (the 2kHz vibration signal is converted into a frequency-amplitude spectrum using Fast Fourier Transform (FFT), with a frequency range of 0-1000Hz and amplitude in g). For example, the vibration spectrum data for a certain time step may include information such as an amplitude of 0.8g at 50Hz, 1.2g at 100Hz, and 3.5g at 200Hz (a characteristic frequency of bearing failure).

[0027] The multidimensional time-series dataset is encapsulated in binary format. The file header contains basic dataset information: time range (e.g., "20250501103000-20250501104000", covering 10 minutes of data), time step (1ms), number of sensors (16, including 4 types of sensors), and data dimensions (current 2D, displacement 1D, temperature field 307200D, vibration spectrum 1000D). The file body stores each type of data in time step order, using a compression algorithm (LZ4) to reduce storage usage, with a compression ratio of approximately 3:1. The dataset is named in the format "ProductionLine_Data_YYYYMMDDHHMMSS.bin", for example, "ProductionLine_Data_20250501103000.bin". After generation, its integrity is verified through a checksum to ensure no data loss or corruption, and it can be directly used for subsequent multi-scale feature extraction.

[0028] S202, perform multi-scale feature extraction and fusion on the time-series operation data to construct a high-dimensional feature vector characterizing the health status of the equipment and the stability of the process, and input it into a multi-step advanced prediction model (specifically including the following attention-based long short-term memory network and Weibull proportional risk model), and output the remaining service life prediction curve of key components and the evolution trend of process deviation. Specifically, the synchronous time-series dataset can be decomposed into multiple scales. The empirical mode decomposition method can be used to extract the intrinsic mode function components from the vibration spectrum. The sliding window statistical method can be used to calculate the time-domain features including the root mean square value and peak factor from the current and voltage waveforms, generating a set of multi-scale feature components. The core of this step is to mine effective features at different frequencies and time scales through multi-scale decomposition, thereby achieving information layering and refinement of the original data and providing rich feature materials for subsequent fusion. The specific implementation method is as follows: Empirical Mode Decomposition (EMD) is an adaptive decomposition method for nonlinear and non-stationary signals. It does not require pre-defined basis functions and can adaptively separate Intrinsic Mode Function (IMF) components of different frequencies from the vibration spectrum. Each IMF component represents the vibration characteristics of the signal at a certain frequency scale and satisfies the constraints that "the number of extrema is equal to or differs from the number of zero-crossings by no more than 1" and "the upper and lower envelopes are locally symmetric about the time axis." Taking the vibration acceleration signal of a forming roller bearing as an example, the frequency range of the original vibration spectrum is 10Hz-10kHz, containing low-frequency characteristics of mechanical wear and high-frequency impact characteristics of bearing failure. The EMD decomposition process is as follows: First, identify the local maxima and minima of the original signal. Fit the upper and lower envelopes using cubic spline interpolation and calculate the mean m1(t). Subtract m1(t) from the original signal x(t) to obtain the first candidate component h1(t). Determine whether h1(t) satisfies the IMF conditions. If not (e.g., the number of extrema and zero-crossings is equal), the component is decomposed. If the phase difference is 2), then h1(t) is used as the new original signal and the above process is repeated until the IMF1 component (high frequency impact characteristics, frequency concentrated in 5kHz-8kHz) that meets the conditions is obtained; the original signal is subtracted from IMF1, and the remaining signal is decomposed repeatedly to obtain IMF2 (mid frequency characteristics, 2kHz-5kHz), IMF3 (low frequency characteristics, 10Hz-2kHz) and residual component r(t) (trend term, reflecting the overall drift of the signal), and finally a vibration multi-scale feature subset containing 3 IMF components and 1 residual component is generated.

[0029] The sliding window statistical method is used to extract stable time-domain features from dynamically changing current and voltage waveforms. The window size is set to 1 second (corresponding to 10,000 current sampling points and 1,000 voltage sampling points) based on the signal change rate, and the window sliding step is 0.5 seconds to ensure the continuity and timeliness of the features. The calculated time-domain features include: the root mean square (RMS) value, which reflects the average energy of the signal, and the formula is RMS=√[(1 / N)Σ(x_i)] 2 [], where N is the number of data points in the window, and x_i is the sampled value. For example, the sum of the squares of the welding current sampled values ​​in a certain window is 4.2 × 10^7 A. 2If N=10000, then RMS=√(4.2×10^7 / 10000)=√4200≈64.8A; the peak factor (C_f) reflects the impulse characteristics of the signal, and the formula is C_f=peak value / x_rms, where the peak value is the maximum value within the window. For example, if the peak current is 89.6A and RMS=64.8A, then C_f=89.6 / 64.8≈1.38; features such as crest factor, kurtosis, and skewness characterize the steepness and symmetry of the signal waveform, respectively. For example, the kurtosis of a voltage signal is 3.2 (close to 3 in a normal distribution, indicating stable voltage fluctuations), and the skewness is 0.1 (approximately symmetrical). Each sliding window corresponds to a set of time-domain feature values, and the feature values ​​of all windows are arranged in chronological order to form a multi-scale time-domain feature subset of current and voltage.

[0030] Multi-scale feature extraction of temperature field data adopts a combination of spatial window statistics and time series analysis. The spatial window is set to 8×8 pixels (corresponding to an actual spatial range of 4mm×4mm). The mean temperature, standard deviation, maximum temperature, and temperature gradient (the maximum temperature difference between adjacent pixels) are calculated within each spatial window. For example, the mean temperature within a certain spatial window is 1050℃, the standard deviation is 35℃, the maximum temperature is 1120℃, and the temperature gradient is 25℃ / mm. In the time dimension, 5 consecutive frames (100ms) are used as time windows to calculate the time change rate of temperature (the mean temperature difference between frames). For example, if the temperature drops from 1050℃ to 1030℃ within a certain window, the time change rate is -0.2℃ / ms. The multi-scale characteristics of the roll gap displacement signal were calculated using statistics obtained through sliding windows of different lengths (0.5 seconds, 1 second, and 2 seconds). Short windows capture instantaneous fluctuations, while long windows reflect trend changes. For example, the standard deviation of displacement for a 0.5-second window is 0.008 mm (small instantaneous fluctuations), and the mean displacement change rate for a 2-second window is 0.002 mm / ms (stable trend). Finally, the vibration IMF components, time-domain features of current and voltage, spatiotemporal features of temperature field, and multi-window statistical features of roll gap displacement were integrated to generate a multi-scale feature component set containing 128 feature dimensions. Each component is labeled with corresponding scale information (frequency scale, time window scale, and spatial window scale).

[0031] Based on a multi-scale feature component set, a graph convolutional neural network is used to model the spatiotemporal dependencies between different physical quantity features, mapping current, displacement, temperature and vibration features to a unified latent space, generating a deeply fused high-dimensional feature vector. The core of this step is to break the independence of different physical quantity characteristics and achieve deep feature fusion by modeling their inherent spatiotemporal correlations. This transforms scattered multi-scale features into high-dimensional vectors that can comprehensively characterize equipment status and process stability. The specific implementation method is as follows: The construction of Graph Convolutional Neural Networks (GCNs) is based on a graph structure of "feature nodes - associated edges". First, the nodes and edges of the graph are defined: nodes are divided into four categories, corresponding to current feature nodes, displacement feature nodes, temperature feature nodes, and vibration feature nodes. Each node contains multi-scale feature components of the corresponding physical quantity (e.g., current nodes contain 16 feature dimensions such as root mean square value and peak factor, and vibration nodes contain 24 feature dimensions such as peak value and frequency of the three IMF components). The entire graph contains a total of 8 core nodes (2 nodes for each type of physical quantity, corresponding to short-scale and long-scale features respectively). The weight of the edge is determined by calculating the Pearson correlation coefficient between features of different nodes. The larger the absolute value of the correlation coefficient, the higher the edge weight. For example, the correlation coefficient between the root mean square value of current and the average value of roll gap displacement is 0.75, and the corresponding edge weight is 0.75. The correlation coefficient between the temperature gradient and the peak value of the vibration IMF1 component is 0.62, and the corresponding edge weight is 0.62. The correlation coefficient between unrelated feature nodes is less than 0.1, and the edge weight is set to 0 to simplify the computational complexity.

[0032] The GCN network structure consists of an input layer, a graph convolutional layer, an activation layer, and an output layer. The input layer has a dimension equal to the sum of the feature dimensions of 8 nodes (16+16+24+24+20+20+12+12=144 dimensions). The graph convolutional layer uses two stacked layers. The first layer has 64 kernels and performs feature aggregation by multiplying the adjacency matrix and the node feature matrix. The calculation process is H^1=σ(Ã×X×W^0+b^0), where à is the normalized adjacency matrix (to avoid excessively large eigenvalues), X is the input node feature matrix, and W^0 is the first layer weight matrix (144×64). b^0 is the bias term (64-dimensional), and σ is the ReLU activation function. This layer captures the spatial dependencies between features of different physical quantities by aggregating the feature information of adjacent nodes. The second layer has 128 convolutional kernels, and the calculation process is H^2=σ(Ã×H^1×W^1+b^1), which further deepens feature fusion and explores more complex spatiotemporal correlations. The output layer flattens the node feature matrix of the second layer to obtain a 128×8=1024-dimensional latent space feature vector. This vector has eliminated the feature heterogeneity of different physical quantities and realized the deep fusion of current, displacement, temperature and vibration features.

[0033] For example, in a multi-scale feature component set at a certain moment, the root mean square value of the current node is 64.8A, the mean value of the displacement node is 0.85mm, the gradient of the temperature node is 25℃ / mm, and the peak value of the vibration node is 3.8g. After aggregation through the first convolutional layer of GCN, these features are assigned different weights (the correlation weight between current and displacement is 0.75, and the correlation weight between temperature and vibration is 0.62), resulting in 64-dimensional intermediate features. The second convolutional layer further fuses the intermediate features, ultimately generating a 1024-dimensional high-dimensional feature vector [0.23, 0.18, 0.45, ..., 0.32, 0.29]. Each dimension of the vector contains comprehensive information on multiple physical quantities and multiple scales. For example, the 32nd dimension simultaneously reflects the synergistic relationship between current fluctuations and vibration impacts, and the 156th dimension characterizes the spatiotemporal coupling characteristics of temperature gradient and roll gap displacement, providing a comprehensive state representation for subsequent prediction models. For example, the training process of GCN is as follows: First, a training dataset was constructed, derived from three consecutive months of historical operating data from a continuous stainless steel welded pipe production line. Data collection was synchronized with production, resulting in 500,000 valid samples. Each sample set contains a multi-scale feature component set (corresponding to the input node features of a graph convolutional network) within a time window and a corresponding equipment health status label. The equipment health status labels were manually annotated using equipment maintenance records and quality inspection reports, categorized into four levels: "Healthy," "Attention," "Warning," and "Fault." "Healthy" corresponds to no abnormalities, "Attention" to slight performance degradation, "Warning" to significant performance decline but still productionable, and "Fault" requiring immediate shutdown for maintenance. Label annotation employed a dual-person independent annotation plus expert review mechanism to ensure label accuracy of at least 95%. The dataset was divided chronologically, with the first 70% used as the training set, the middle 15% as the validation set, and the last 15% as the test set.

[0034] Before training, the input features are standardized to ensure that the mean of each feature dimension is 0 and the variance is 1. The graph convolutional network uses cross-entropy loss for multi-class classification tasks, and an L2 regularization term is added to prevent overfitting, with the regularization coefficient set to 0.001. The optimizer is Adam, with an initial learning rate of 0.001. After every 10 training epochs, if the validation set loss no longer decreases, the learning rate is reduced to 90% of its original value. The training batch size is set to 64, with a maximum of 200 training epochs. An early stopping strategy is used: training stops if the validation set loss does not decrease for 20 consecutive epochs. During training, the node feature matrix and adjacency matrix are input into the network. The predicted classification result is calculated through forward propagation, and then the network weights are updated through backpropagation. After training, the classification accuracy on the test set reaches over 92%, meeting the requirements for engineering applications.

[0035] The high-dimensional feature vector is input into a long short-term memory network based on an attention mechanism. This network learns the mapping pattern from historical state to future state through an encoder-decoder structure, and generates a prediction sequence of equipment state and process parameters for a predetermined number of time steps in the future. The core of this step is to enhance the LSTM's ability to capture key historical information by leveraging the attention mechanism. An encoder-decoder architecture is used to achieve a precise mapping from historical high-dimensional features to future states, generating multi-time-step prediction sequences. The specific implementation is as follows: The Attention-LSTM network, based on the attention mechanism, adopts a dual-end structure of "encoder-decoder". It presets 100 future prediction time steps (corresponding to 100ms, which is suitable for the time scale of real-time control of the production line). The equipment status predicted at each time step includes bearing vibration peak, motor temperature, and roller wear. The process parameters include welding current, welding voltage, roller displacement, and pipe temperature, for a total of 7 prediction indicators.

[0036] The encoder consists of three LSTM layers and one attention layer. The hidden layer of the LSTM layer has a dimension of 512, the activation function is tanh, and the forget gate threshold is set to 0.8 (to preserve important historical information). The input is a high-dimensional feature vector sequence arranged in chronological order (one vector every 1ms, taking the feature vectors of the most recent 200 time steps as historical input, forming a 200×1024 input matrix). The attention layer adopts an additive attention mechanism to calculate the attention weight of the feature vector at each historical time step, with the formula e_ij=v^Ttanh(W_hh_i+W_ss_j), where h_i is the hidden state of the encoder at the i-th time step, s_j is the hidden state of the decoder at the j-th time step, and v, W_h, and W_s are learnable parameters. The weights α_ij are obtained by normalization through the softmax function, and the weighted sum is used to obtain the attention-enhanced historical feature representation. For example, the feature vector of a certain historical time step contains key information about abnormal bearing vibration, and its attention weight is 0.92 (far higher than 0.01-0.05 for other time steps). The features of this time step are focused on, effectively improving the model's ability to capture information about the precursors of the fault.

[0037] The decoder consists of three LSTM layers and one fully connected layer. The hidden layer dimension of the LSTM layer is the same as that of the encoder (512). The initial hidden state and cell state are initialized by the output of the encoder's last time step, and the state transition rules are learned through attention-enhanced historical feature vectors. The output dimension of the fully connected layer is 7 (corresponding to 7 prediction indicators), and the activation function is linear activation (to ensure the continuity of prediction values). Each time step outputs a set of predicted values ​​for equipment status and process parameters. During training, mean squared error (MSE) is used as the loss function, Adam is used as the optimizer, and the learning rate is initially set to 1e-4. When the validation set loss does not decrease for 5 consecutive rounds, it is decayed to 1e-5. The number of training iterations is 300 rounds. The prediction error of the final model is controlled within the following ranges: vibration peak ≤ 0.1g, temperature ≤ 5℃, current ≤ 2A, and displacement ≤ 0.005mm, which meets the accuracy requirements for real-time prediction.

[0038] For example, after inputting a high-dimensional feature vector sequence from the most recent 200ms, the decoder generates a predicted sequence for the next 100ms: At time step 1 (t+1ms), the welding current is 64.5A, the roll gap displacement is 0.848mm, the bearing vibration peak is 3.7g, and the pipe temperature is 1045℃; at time step 50 (t+50ms), the welding current is 63.2A, the roll gap displacement is 0.842mm, and the bearing vibration peak is 4.2g (vibration gradually increases, indicating increased wear); at time step 100 (t+100ms), the welding current is 62.8A, the roll gap displacement is 0.838mm, the bearing vibration peak is 4.8g, and the pipe temperature is 1032℃. The predicted sequence fully presents the evolution trend of the future equipment state and the changing patterns of process parameters. For example, the training process of a long short-term memory network based on an attention mechanism is as follows: A dedicated time-series prediction dataset was constructed, using the same historical data as the graph convolutional network, but employing a sliding window approach to construct samples. The sliding window length was set to 200 time steps (corresponding to 200 milliseconds), and the prediction step size was set to 100 time steps (corresponding to the next 100 milliseconds). Each sample's input was a sequence of high-dimensional feature vectors (200×1024 dimensions) spanning 200 consecutive time steps, and the output labels were a sequence of actual equipment status and process parameters for the next 100 time steps. Equipment status included bearing vibration peak value, motor temperature, and roll gap wear; process parameters included welding current, welding voltage, roll gap displacement, and pipe temperature, totaling seven prediction indicators. Label values ​​were directly extracted from historical data and smoothed using Kalman filtering to reduce measurement noise. The dataset generated 400,000 samples, which were chronologically divided into a training set (70%), a validation set (15%), and a test set (15%).

[0039] Before training, the input features and label sequences are normalized, mapping each index to the 0-1 range. Both the encoder and decoder of the attention mechanism use a 3-layer LSTM with a hidden layer dimension of 512. The mean squared error loss function is used to measure the difference between the predicted and true sequences. The optimizer is Adam, with an initial learning rate of 0.0001. The learning rate decays to 90% of its original value every 10 epochs when the validation set loss stagnates. The training batch size is set to 32, with a maximum of 300 training epochs. An early stopping strategy is employed (training stops if the validation set loss does not decrease after 20 consecutive epochs). During training, the input sequence is sequentially fed into the encoder LSTM to obtain the hidden states at each time step. Weighted historical features are then calculated through the attention layer, and the decoder LSTM generates the predicted sequence step-by-step based on these weighted features. A gradient clipping threshold of 1 is set during backpropagation to prevent gradient explosion. After training, the results were evaluated on the test set: the root mean square error of the vibration peak prediction was less than 0.1g, the temperature error was less than 5℃, the current error was less than 2A, and the displacement error was less than 0.005mm, which met the requirements for real-time prediction accuracy.

[0040] Based on the predicted sequence, the Weibull proportional hazards model is used to calculate the probability distribution of the remaining useful life of key mechanical components. At the same time, Kalman filtering is used to track the deviation trend of process parameters relative to the set values. Finally, the remaining useful life prediction curve and process deviation evolution trend report are output.

[0041] The core of this step is to transform the predicted sequence into remaining useful life and process deviation information with engineering significance, providing a quantitative basis for subsequent preventive maintenance decisions. The specific implementation method is as follows: The Weibull Proportional Hazards Model is a classic model for calculating remaining useful life (RUL) in reliability analysis. It is suitable for predicting the life of mechanical components. Its core formula is h(t|z(t))=h0(t)exp(β^Tz(t)), where h(t|z(t)) is the conditional risk rate (instantaneous probability of failure) at time t, h0(t)=(m / η)(t / η)^(m-1) is the baseline risk rate (Weibull distribution, where m is the shape parameter and η is the scale parameter), z(t) is the covariate vector at time t (equipment state characteristics extracted from the prediction sequence, such as peak vibration, temperature, and wear), and β is the risk coefficient of the covariate (obtained through training with historical failure data).

[0042] Model parameter calibration was completed using historical fault data: Fault records of forming roller bearings from the past three years (50 sets in total) were collected, including the fault occurrence time and the equipment state sequence before the fault. The maximum likelihood estimation method was used to obtain the shape parameter m = 2.8 (m > 1, indicating that the risk rate increases over time, consistent with the fault pattern of mechanical wear), the scale parameter η = 800 hours (baseline life), and the covariate risk coefficient β = [0.6 (peak vibration), 0.3 (temperature), 0.1 (wear)], indicating that the peak vibration has the greatest impact on fault risk. The covariate sequence z(t) for the next 100ms was extracted from the predicted sequence (e.g., peak vibration 4.8g, temperature 85℃, wear 0.012mm at t+100ms), substituted into the model to calculate the conditional risk rate at each moment, and then the probability distribution of the remaining service life P(RUL>t) = exp(-∫0^th(τ|z(τ))dτ) was calculated through integration. For example, the probability of a residual useful life (RUL) of 120 hours is calculated to be 60%, 100 hours to be 85%, and 80 hours to be 98%. The RUL values ​​at different confidence levels (50%, 80%, and 95%) are arranged in chronological order to generate a residual useful life prediction curve. The curve shows a decreasing trend, and the confidence interval gradually widens as time goes on (the further into the future, the greater the uncertainty). For example, at a 50% confidence level, it takes 20 hours for the RUL to decrease from the current 120 hours to 100 hours, and 45 hours to decrease to 80 hours.

[0043] Kalman filtering is used to track the deviation trend of process parameters relative to the setpoint in real time. Its core is to achieve the optimal estimation of deviation and suppress noise interference through iterative updates of the state equation and observation equation. Taking welding current as an example, the setpoint is 65A. The state equation is x_k=Ax_{k-1}+Bu_{k-1}+w_{k-1}, where x_k is the current deviation at time k (x_k=actual current-setpoint), A=0.99 (time correlation coefficient of deviation), B=0.01 (control input coefficient, u=0 here), and w_{k-1} is the process noise (following N(0,Q), Q=0.01). The observation equation is z_k=Hx_k+v_k, where z_k is the observation deviation at time k (extracted from the prediction sequence, such as z_k=64.5-65=-0.5A), H=1 (observation matrix), and v_k is the observation noise (following N(0,R), R=0.02).

[0044] The filtering iteration process is as follows: First, a prediction step is performed, calculating the prior estimate x^-k=Ax{k-1}+Bu_{k-1} at time k, and the prior covariance P^-k=AP{k-1}A^T+Q; then, an update step is performed, calculating the Kalman gain K_k=P^-_kH^T(HP^-_kH^T+R)^-1, the posterior estimate x_k=x^-_k+K_k(z_k-Hx^-k), and the posterior covariance P_k=(I-K_kH)P^-k. For example, the deviation estimate at time k-1 is x{k-1}=-0.3A, P{k-1}=0.015, and the observation deviation at time k is z_k=-0.5A. We can calculate x^-_k=0.99×(-0.3)=-0.297A, P^-_k=0.99×0.015×0.99+0.01≈0.0247, K_k=0.0247×1×(1×0.0247+0.02)^-1≈0.0247 / 0.0447≈0.553, x_k=-0.297+0.553×(-0.5+0.297)≈-0.297-0.112≈-0.409A. That is, the optimal deviation estimate at time k is -0.409A (the current is lower than the set value of 0.409A).

[0045] The aforementioned Kalman filtering process was applied to all process parameters, including welding voltage, roll gap displacement, and pipe temperature, to track their deviation trends. For example, with a roll gap displacement setpoint of 0.85 mm, the deviation gradually increased from 0.002 mm to 0.012 mm after filtering (displacement too small). With a pipe temperature setpoint of 1050℃, the deviation increased from -5℃ to -18℃ (temperature too low). Finally, the remaining service life (RUL) prediction curve and the deviation trends of each process parameter were integrated to generate an evolution trend report. The report includes: the RUL prediction curve (50%, 80%, and 95% confidence intervals), the failure risk level of key components (currently medium risk, rising to high risk after 100 hours), the deviation time series of each process parameter, the critical point of accelerated deviation increase (e.g., welding current deviation exceeding -0.5A at t+80ms is the critical point), and the potential impact of deviation on product quality (e.g., low temperature may lead to insufficient weld penetration). This provides a comprehensive and quantitative basis for subsequent preventive maintenance decisions.

[0046] S203, based on the remaining service life prediction curve and the evolution trend of process deviation, the potential failure risk and quality defect probability are dynamically evaluated through the digital twin simulation module, and a preventive maintenance decision scheme including maintenance priority and time window is generated. Specifically, a high-fidelity dynamic model of the production line can be built in a digital twin simulation environment. The remaining service life prediction curve and the evolution trend of process deviation can be imported as boundary conditions to simulate the equipment operating status and product quality at different future periods and generate a set of virtual operating scenarios. The core of this step is to replicate the physical characteristics of the production line through high-fidelity modeling, combine it with predictive data to set boundaries, and achieve accurate simulation of future states, providing a virtual experimental environment for risk assessment. The specific implementation method is as follows: The high-fidelity dynamic model of the production line is constructed using a multiphysics coupling modeling method, integrating three core systems: mechanical, electrical, and thermal, covering the entire process from raw material forming to welding and sizing. The mechanical system modeling focuses on the forming roller, traction machine, and sizing machine, establishing a multibody dynamic model based on the Newton-Euler equations. Contact mechanics theory is introduced to describe the interaction between the roller gap and the pipe material. Key parameters include the forming roller stiffness (2.5 × 10⁵ N / m) and the roller shaft moment of inertia (0.8 kg·m). 2 The friction coefficient (0.12, sliding friction between steel and roll surface) was used to calibrate the model using roll gap displacement data measured by a laser tracker, ensuring that the roll gap variation error was ≤0.005mm. The electrical system modeling focused on the welding power supply and servo driver. A voltage and current transfer function was constructed based on the circuit topology. The output characteristic parameters of the welding power supply were set as follows: rated power 60kW, current regulation response time ≤5ms, voltage fluctuation range ±0.5V. The consistency of the model output was verified using historical welding current waveform data. The thermal system modeling adopted the finite volume method to simulate the conduction and radiation heat dissipation of the temperature field in the welding zone. Parameters such as thermal conductivity (16.3W / (m·K) for stainless steel at 20℃ and 28.5W / (m·K) at 1000℃) and radiation heat transfer coefficient (0.85) were obtained through inversion from measured data using an infrared thermal imager, ensuring that the temperature field simulation deviation was ≤5℃.

[0047] The import of boundary conditions requires dynamic correlation between the predicted data and model parameters: the remaining useful life (RUL) prediction curve serves as the time scale input for equipment performance degradation. For example, the RUL prediction curve for the forming roller bearing shows a current remaining useful life of 120 hours, with wear increasing from 0.012 mm to 0.025 mm in the next 20 hours. The model simulates performance degradation by dynamically adjusting the roller friction coefficient (the friction coefficient increases by 0.02 for every 0.01 mm increase in wear). The evolution trend of process deviation serves as the input perturbation for the model. For example, the welding current deviation gradually increases from -0.5A to -2.8A, and the roller gap displacement deviation increases from 0.002 mm to 0.012 mm. These are directly imported into the corresponding system modules to simulate the impact of process fluctuations on product quality.

[0048] The virtual operation scenario set is divided into three time periods according to the time dimension: short-term (future 1-6 hours), medium-term (6-24 hours), and long-term (24-120 hours). Each time period contains 3-5 typical scenarios, covering operating conditions such as normal equipment degradation, deviation acceleration, and potential failures. Example Scenario 1 (Short-term, next 3 hours): Forming roller bearing wear is 0.015mm, welding current deviation is -1.2A, roller gap displacement deviation is 0.006mm, and the simulated result is a pipe temperature of 1040℃ and a weld width of 6.2mm (standard 6.0±0.3mm), which is of acceptable quality. Example Scenario 2 (Medium-term, next 12 hours): Bearing wear is 0.020mm, current deviation is -2.0A, roller gap deviation is 0.009mm, and the simulated result is a weld width of 5.7mm (close to the lower limit) and a pipe diameter ellipticity of 0.4% (standard ≤0.3%), indicating a minor defect. Example Scenario 3 (Long-term, next 72 hours): Bearing wear is 0.035mm, current deviation is -2.6A, roller gap deviation is 0.015mm, and the simulated result is a bearing vibration peak of 5.8g (fault threshold 6.0g), a weld incomplete penetration defect probability of 85%, and the equipment is on the verge of failure. All scenarios record equipment operating parameters (vibration, temperature, current), product quality indicators (melt width, ellipticity, defect type), and fault warning status, forming a virtual operating scenario set containing 12 scenarios.

[0049] Based on a set of virtual operating scenarios, the frequency of failures of key components is statistically analyzed using the Monte Carlo simulation method. At the same time, the probability of pipe defects caused by process deviations is calculated, and a failure risk probability distribution and a quality defect probability matrix are generated. The core of this step is to simulate uncertainty through random sampling, quantify the probability of faults and quality defects, and provide a quantitative basis for subsequent optimization. The specific implementation method is as follows: The core of the Monte Carlo simulation method is to cover parameter uncertainties through extensive random sampling, with the number of samplings set at 10,000 (verified to ensure a probability error ≤2%). Sampling objects include performance degradation coefficients of key components (e.g., bearing wear rate fluctuations of ±15%), random disturbances from process deviations (e.g., current deviations with added ±0.3A of normally distributed noise), and environmental interference parameters (e.g., workshop temperature fluctuations of ±3℃). During the simulation, each sample randomly selects a basic scenario from the virtual operating scenario set, superimposes random disturbances, and then substitutes it into a high-fidelity dynamic model to determine whether a fault or quality defect is triggered.

[0050] The failure frequency statistics for key components focus on forming roller bearings, welding power supplies, and servo drives. Failure judgment criteria are based on industry standards and equipment manuals: forming roller bearing failure is judged by a vibration peak ≥ 6.0g or a temperature ≥ 95℃; welding power supply failure is judged by output current fluctuation ≥ ±5A or voltage deviation ≥ ±3V; servo drive failure is judged by roller gap displacement control accuracy ≤ ±0.01mm. For example, in 10,000 samplings, the forming roller bearing failure frequency was 120 times in a short-term scenario (3 hours), 850 times in a medium-term scenario (12 hours), and 4200 times in a long-term scenario (72 hours). The failure risk probabilities for different time periods were statistically obtained: 1.2% for short-term, 8.5% for medium-term, and 42.0% for long-term. Arranging the probabilities by time step (1 hour as the unit) generates a failure risk probability distribution curve. The curve shows an exponential growth trend, with the failure probability exceeding 40% after 72 hours.

[0051] The quality defect probability matrix is ​​constructed using process deviation type and defect type as two dimensions. Process deviations include welding current deviation, roll gap displacement deviation, and temperature deviation. Defect types include incomplete weld penetration, weld overheating, excessive pipe diameter ellipticity, and surface scratches. The matrix dimension is 3×4, and each element represents the conditional probability that a certain process deviation leads to a certain defect. During the calculation, product quality under different combinations of process deviations is simulated, and the ratio of the number of defect occurrences to the total number of simulations is statistically analyzed. For example, when the welding current deviation is ≤-2.0A, the number of times incomplete weld penetration occurs is 3200 out of a total of 10000 simulations, resulting in a probability of 32%; when the roll gap displacement deviation is ≥0.01mm, the probability of excessive pipe diameter ellipticity is 45%; and when the temperature deviation is ≤-20℃, the probability of weld overheating is 8%. The final generated quality defect probability matrix is ​​shown in the example: [[32%,5%,12%,3%],[18%,10%,45%,7%],[6%,28%,15%,9%]]. The rows represent process deviation types, and the columns represent defect types, clearly showing the correlation strength between process deviations and quality defects.

[0052] By combining the probability of failure risk, the probability of quality defects, and the estimated production downtime losses and maintenance costs, a multi-objective optimization model is established. With the goal of minimizing the overall cost and maximizing the quality pass rate, the optimal intervention point is solved to generate a set of preliminary maintenance timing and measures. The core of this step is to balance cost and quality through multi-objective optimization, determine the optimal timing and measures for maintenance intervention, and achieve the scientific nature of preventive maintenance. The specific implementation method is as follows: The objective function of the multi-objective optimization model contains two core objectives: minimizing the overall cost and maximizing the quality pass rate. The overall cost C consists of three parts: downtime loss C1, maintenance cost C2, and quality defect loss C3, with the formula C = C1 + C2 + C3. Downtime loss C1 is calculated based on production efficiency and downtime duration. The production line has an hourly capacity of 120 welded pipes, with a profit of 80 yuan per pipe. Therefore, the hourly downtime loss is 120 × 80 = 9600 yuan. Downtime duration includes maintenance operation time (e.g., 2 hours for replacing bearings) and downtime due to malfunctions (calculated using a weighted average based on failure probability). Maintenance cost C2 is calculated based on the type of maintenance measure. The spare parts cost for replacing the forming roller bearing is 2800 yuan, and the labor cost is 800 yuan, totaling 3600 yuan. The labor cost for calibrating the roller gap is 500 yuan, and there is no spare parts cost. Quality defect loss C3 is calculated based on the defect rate and the cost of handling defects per pipe. The cost of handling a single pipe with incomplete weld penetration is 150 yuan, and the cost of handling a single pipe with excessive pipe diameter ellipticity is 100 yuan. The formula for calculating the quality pass rate Q is Q=1-Σ(defect probability × defect impact weight). The defect impact weight is set according to the severity: incomplete penetration 0.4, overheating 0.3, excessive ovality 0.2, surface scratches 0.1. For example, if the defect probabilities in a certain period are 32%, 5%, 12%, and 3%, then Q=1-(0.32×0.4+0.05×0.3+0.12×0.2+0.03×0.1)=1-0.164=83.6%.

[0053] The model's constraints include: maintenance intervention time t must meet production plan constraints (avoiding the daily production peak of 10:00-16:00), i.e., t∉[10,16]; the execution time of maintenance measures ≤ 80% of the equipment's remaining service life (to avoid re-failure in the short term after maintenance), i.e., t≤0.8×current RUL value (current RUL=120 hours, so t≤96 hours); and the quality pass rate Q≥85% (enterprise quality standard). The optimization variables are maintenance intervention time t (unit: hours) and maintenance measure type M (M1: replace forming roller bearing; M2: calibrate roller gap; M3: replace bearing + calibrate roller gap).

[0054] The model was solved using the Non-Dominated Sorting Genetic Algorithm (NSGA-II), with the following parameters: population size 100, number of iterations 200, crossover probability 0.8, and mutation probability 0.05. The Pareto optimal solution set was obtained through algorithm search, from which the intervention point with the best overall performance was selected: intervention time t = 72 hours (avoiding peak production periods, when the failure probability is 42%, not reaching the high-risk threshold), and maintenance measure M3 (replacing the bearing + calibrating the roll gap). The comprehensive cost C corresponding to this intervention point is calculated as follows: downtime loss (2 hours × 9600 = 19200 yuan) + maintenance cost (3600 + 500 = 4100 yuan) + defect loss ((0.05 × 0.4 + 0.02 × 0.3 + 0.03 × 0.2 + 0.01 × 0.1) × 120 × 72 × 80 = 0.033 × 69120 = 2280.96 yuan), totaling 25580.96 yuan. The quality pass rate Q = 96.7%, meeting the dual-objective optimization requirements. The final preliminary maintenance timing and measures set includes three candidate solutions: Solution 1 (t = 72 hours, M3), Solution 2 (t = 60 hours, M2), and Solution 3 (t = 84 hours, M1). Each solution is labeled with its comprehensive cost, quality pass rate, and applicable scenarios.

[0055] Resource constraints are verified on the initial maintenance timing and measures set, taking into account spare parts inventory, maintenance team availability and production plan, to determine the executable time window and execution order of each maintenance task, and finally generate a preventive maintenance decision plan that includes specific maintenance items, priority order and time window.

[0056] The core of this step is to ensure the feasibility of the maintenance plan. Feasible solutions are selected through resource constraint verification, execution details are clarified, and a highly implementable decision-making plan is formed. The specific implementation method is as follows: Resource constraint verification is conducted from three dimensions: spare parts inventory, maintenance team availability, and production plan. Spare parts inventory verification is performed by querying the spare parts ledger in the enterprise's ERP system to confirm the available quantity of spare parts required for the maintenance measures: Option 1 requires 2 sets of forming roller bearings, and the current inventory is 5 sets (meeting the requirements); Option 2 requires no spare parts (calibration only, meeting the requirements); Option 3 requires 2 sets of bearings, and the inventory is sufficient, but there is no calibration tool (the tool is sent for repair and will be returned in 3 days; for Option 3, t=84 hours ≥ 3 days, which meets the requirements). The availability verification of maintenance teams is based on the team's shift schedule. The maintenance team is divided into day shift (8:00-18:00) and night shift (18:00-8:00). In Scheme 1, t=72 hours corresponds to 0:00 on the 3rd day (night shift). The night shift team has 2 senior technicians (with qualifications for bearing replacement and roll gap calibration), which is feasible. In Scheme 2, t=60 hours corresponds to 12:00 on the 2nd day (day shift production peak). The team needs to prioritize production and has no available manpower (not feasible). In Scheme 3, t=84 hours corresponds to 12:00 on the 3rd day (day shift production peak). Similarly, there is no available manpower (not feasible).

[0057] Production plan verification requires ensuring that maintenance time windows do not conflict with production tasks. A review of the production plan shows that within the next 7 days, the planned downtime for maintenance is from 0:00 to 4:00 on the 3rd day (originally used for routine equipment checks). The maintenance time for Scheme 1 (72 hours, i.e., 0:00 to 2:00 on the 3rd day) can be embedded within this timeframe without additional production stoppage. Schemes 2 and 3 both have maintenance times that conflict with production tasks, requiring adjustments to the production plan (higher cost, lower priority). Based on the verification results, only Scheme 1 fully meets the resource constraints. Schemes 2 and 3 require time window adjustments: t for Scheme 2 is adjusted to 18:00 on the 2nd day (off-peak night shift), and t for Scheme 3 is adjusted to 0:00 on the 4th day (planned maintenance period). After re-verification, both schemes meet the constraints.

[0058] The determination of the executable time window takes into account resource constraints and equipment status. Option 1's time window is 0:00-2:00 on day 3 (2 hours), during which there are no production tasks, spare parts are sufficient, and the work team is available. Option 2's time window is 18:00-19:00 on day 2 (1 hour). Option 3's time window is 0:00-2:00 on day 4 (2 hours). The execution order is prioritized based on "risk urgency + cost-effectiveness ratio." Option 1 has the highest failure risk urgency (failure probability rises rapidly after 72 hours) and the best cost-effectiveness ratio (lowest overall cost and highest quality pass rate), so its priority is set to 1. Option 3 has the second highest risk urgency, so its priority is set to 2. Option 2 has the lowest risk urgency, so its priority is set to 3.

[0059] The final preventative maintenance decision plan comprises four core parts: a detailed list of maintenance items (replacing the forming roller bearing, calibrating the forming roller gap, and checking the stability of the welding power supply output), a priority ranking (1. Replace the bearing + calibrate the roller gap; 2. Replace the bearing; 3. Calibrate the roller gap), a time window (Priority 1: Day 3, 0:00-2:00; Priority 2: Day 4, 0:00-2:00; Priority 3: Day 2, 18:00-19:00), and execution requirements (replacing the bearing requires the use of a torque wrench (torque value 45 N·m), calibrating the roller gap requires controlling the displacement deviation within ±0.003 mm, and the equipment status must be verified by running it for 30 minutes after maintenance). The plan also includes a risk contingency plan: if bearing wear exceeds expectations during maintenance, the spare bearing will be immediately activated, the maintenance time will be extended to 3 hours, and the production plan will be adjusted accordingly.

[0060] S204. Based on the preventive maintenance decision scheme, an adaptive dynamic programming algorithm is used to optimize welding power, roll gap servo pressure, and production line cycle time parameters in real time, generating a set of process parameter collaborative adjustment instructions to form a data-driven closed-loop optimization and maintenance control loop.

[0061] Specifically, it can analyze preventive maintenance decision-making schemes, extract equipment information that is about to be maintained and the expected performance degradation of the equipment, quantify the equipment information and performance degradation into constraints for process parameter adjustment, and generate a set of parameter optimization constraints. The core of this step is to transform qualitative maintenance decisions into quantitative process parameter constraints, ensuring that parameter optimization is within the acceptable range of equipment performance degradation, and avoiding equipment overload or quality deterioration due to parameter adjustments. The specific implementation method is as follows: The analysis of the preventive maintenance decision-making plan focuses on the core maintenance objects and performance degradation characteristics. First, the equipment information to be maintained is extracted: the key maintenance equipment is the forming roller bearing (priority 1 maintenance item), the maintenance time window is 0:00-2:00 on the 3rd day, the current equipment status is 72 hours of remaining service life, and the expected performance degradation before maintenance is: the bearing wear will increase linearly from the current 0.012mm to 0.025mm before maintenance, the corresponding roller rotation friction coefficient will increase from 0.12 to 0.145, and the roller gap control accuracy will decrease from ±0.005mm to ±0.008mm; the secondary maintenance object is the welding power source (with inspection items), and the expected performance degradation is that the output power fluctuation range will expand from ±1kW to ±1.5kW.

[0062] The quantification of constraints needs to be combined with the limitations of equipment performance degradation on process parameters, and the boundary ranges of each decision variable need to be clarified: For welding power (P), considering the performance degradation of the welding power source and the requirements for weld quality, a lower limit of 55kW is set (to avoid insufficient power leading to incomplete penetration), and an upper limit of 60kW (rated power supply power, to prevent overload). Simultaneously, due to increased power supply fluctuations, an additional fluctuation constraint of |ΔP|≤1.5kW / second is added (to avoid sudden power changes). For roll gap servo pressure (F), based on the decrease in load-bearing capacity caused by wear of the forming roll bearings, a pressure range of 30MPa-40MPa (60%-80% of the rated bearing load pressure) is set, and the pressure change rate constraint of |ΔF|≤2MPa / second (to reduce the impact on worn bearings). For production line cycle time (T), considering the decrease in roll gap control accuracy and production efficiency requirements, a cycle time range of 0.4 pieces / second-0.6 pieces / second (corresponding to an hourly capacity of 1440-2160 pieces) is set, with a cycle time adjustment step of 0.05 pieces / second (to ensure smooth adjustment).

[0063] In addition, coupling constraints are added: the synergistic constraint between welding power and roll gap pressure, when the welding power is ≥58kW, the roll gap pressure must be ≥35MPa (to compensate for the deformation of the pipe under high power); the correlation constraint between cycle time and power, when the cycle time is ≥0.55 pieces / second, the welding power must not be less than 57kW (to ensure the weld penetration depth during high-speed production). All constraints are organized according to the structure of "decision variable-constraint type-numerical range-constraint basis" to generate a parameter optimization constraint set. For example: "Welding power P: 55kW≤P≤60kW, |ΔP|≤1.5kW / second (based on welding power performance attenuation and weld quality requirements); Roll gap servo pressure F: 30MPa≤F≤40MPa, |ΔF|≤2MPa / second (based on the bearing capacity after wear of the forming roller bearing); Production line cycle time T: 0.4 pieces / second≤T≤0.6 pieces / second, adjustment step size 0.05 pieces / second (based on roll gap control accuracy and production efficiency); Coupling constraint: P≥58kW→F≥35MPa, T≥0.55 pieces / second→P≥57kW".

[0064] A comprehensive evaluation function with the goals of lowest energy consumption, most stable quality, and highest production efficiency is constructed. Welding power, roll gap servo pressure, and production line cycle time are used as decision variables to establish a real-time optimization model for process parameters. The core of this step is to balance the multi-objective optimization requirements. By using a weighted comprehensive evaluation function, energy consumption, quality, and efficiency are transformed into a single optimization objective, providing a clear direction for the algorithm's solution. The specific implementation method is as follows: The comprehensive evaluation function (J) adopts a weighted linear combination form, with the formula J=ω1×J1+ω2×J2+ω3×J3, where J1 is the energy consumption optimization objective (minimization), J2 is the quality stability objective (minimization), and J3 is the production efficiency objective (maximization). ω1, ω2, and ω3 are weight coefficients, determined by the analytic hierarchy process. Five production line process experts and three equipment maintenance experts were invited to score the evaluation. After consistency testing (CR=0.06<0.1), the final weights are ω1=0.3, ω2=0.4, and ω3=0.3 (quality has the highest weight, which conforms to the production principle of "quality first").

[0065] The quantitative definitions of each sub-objective function are as follows: Energy consumption objective J1 is characterized by the sum of instantaneous energy consumption of welding power and energy consumption of production line operation, with the formula J1=P×T+k×T, where P is welding power (kW), T is cycle time (pieces / second), and k is a fixed energy consumption coefficient (valued at 5kW·second / piece, representing the unit energy consumption of auxiliary equipment such as conveyor rollers and cooling systems). For example, when P=58kW and T=0.5 pieces / second, J1=58×0.5+5×0.5=31.5kW·second / piece. The smaller J1 is, the lower the energy consumption. Quality stability objective J2 is characterized by the weighted sum of squares of process parameter deviations, with the formula J2=a×(P-P0). 2 +b×(F-F0) 2 +c×(T-T0) 2 Where P0=57kW, F0=34MPa, T0=0.5 pieces / second are standard process parameters, and a=0.02, b=0.01, c=2 are deviation weights (welding power deviation has the greatest impact on quality). For example, when P=58kW, F=35MPa, T=0.5 pieces / second, J2=0.02×(1) 2 +0.01×(1) 2 +2×(0) 2 =0.03, the smaller J2 is, the more stable the quality; the production efficiency target J3 is directly represented by the cycle time T, J3=T, the larger J3 is, the higher the efficiency, for example, when T=0.55 pieces / second, J3=0.55.

[0066] The real-time optimization model for process parameters aims to maximize the comprehensive evaluation function J (by taking the negative of J1 to achieve a shift from minimization to maximization, i.e., J'=-ω1×J1+ω2×(-J2)+ω3×J3). The decision variable is X=[P,F,T]^T, and the constraints are the parameter optimization constraint set generated in step one. The boundary conditions of the model are dynamically updated. Each time new time-series running data is received (1ms / time), the constraint range and objective function weights are adjusted according to the real-time status of the equipment (such as bearing vibration value, power fluctuation). For example, when the bearing vibration value exceeds 4.5g, ω2 is temporarily increased to 0.5 to prioritize ensuring quality stability.

[0067] An adaptive dynamic programming algorithm is used to solve the optimization model. This algorithm dynamically updates the state transition equation and reward function based on real-time collected process data, searches online for the optimal parameter combination that satisfies the parameter optimization constraint set, and generates a real-time optimal process parameter vector. The core of this step is to leverage the online learning capabilities of Adaptive Dynamic Programming (ADP) to quickly search for the optimal parameter combination, adapting to dynamic changes in device status and ensuring the real-time performance and reliability of the optimization results. The specific implementation method is as follows: The core framework of the adaptive dynamic programming algorithm includes a state module, a value network, a policy network, and an update module. The state module defines the system state vector S=[P_prev,F_prev,T_prev,V_bear,ΔP_power]^T, where P_prev, F_prev, and T_prev are the process parameters at the previous moment, V_bear is the peak vibration value of the forming roller bearing (g), and ΔP_power is the power fluctuation of the welding power source (kW). The state vector has a dimension of 5, comprehensively representing the current equipment and process state. For example, at a certain moment, the state S=[57.5kW,34.2MPa,0.5 roots / second,4.2g,1.2kW].

[0068] The state transition equation describes the evolution from the current state to the next state, with the formula S_next=f(S,X)+W, where X=[P,F,T]^T is the decision variable, f(·) is the deterministic transition function, constructed based on the production line dynamics model, for example P_next=P+ΔP×Δt (Δt=1ms is the sampling period), F_next=F+ΔF×Δt, T_next=T; W is the random disturbance term (following N(0,Q), Q is the covariance matrix, and the diagonal elements are [0.01,0.01,0.001,0.02,0.01]), simulating the random fluctuations of the equipment state.

[0069] The reward function is directly related to the comprehensive evaluation function, with the formula R(S,X)=J'=-ω1×J1+ω2×(-J2)+ω3×J3. A constraint penalty term is added; if the decision variable exceeds the constraint range, the penalty term is -100, ensuring the algorithm's search does not deviate from the feasible region. The value network uses a single-hidden-layer neural network (5-dimensional input layer, 64-dimensional hidden layer, 1-dimensional output layer) with ReLU activation function, used to approximate the state value function V(S)=E[Σγ^kR(S_k,X_k)] (γ=0.95 is the discount factor). The policy network is also a single-hidden-layer neural network (5-dimensional input layer, 64-dimensional hidden layer, 3-dimensional output layer), outputting the optimal estimate of the decision variable X.

[0070] The algorithm's adaptive update mechanism is reflected in the following: each time real-time process data is collected (such as vibration value V_bear=4.3g, power fluctuation ΔP_power=1.3kW), the state vector S is updated, and the weights of the value network are adjusted through temporal difference learning (TD error). The TD error formula is δ=R+γ×V(S_next)-V(S). The weight update adopts the gradient descent method with a learning rate η=0.001. At the same time, the output range of the policy network is adjusted according to the constraint satisfaction in the new state. For example, when V_bear>4.5g, the upper limit of P output by the policy network is automatically reduced to 58kW, and the upper limit of F is reduced to 38MPa.

[0071] Example of online search process: Initial state S0 = [57kW, 34MPa, 0.5 roots / second, 4.0g, 1.0kW], policy network outputs initial parameter combination X0 = [57.2kW, 34.1MPa, 0.5 roots / second], calculate reward function R0 = 28.5; after collecting new data, update state S1 = [57.2kW, 34.1MPa, 0.5 roots / second, 4.1g, 1.1kW], value network estimates V(S1) = 29.2, TD error δ = 28.5 + 0.95 × 29.2 - V (S0) = 28.5 + 27.74 - 28.0 = 28.24, update the value network weights; the policy network outputs the optimized parameter combination X1 = [57.5kW, 34.3MPa, 0.52 roots / second] based on the new state and value estimation, with a reward function R1 = 29.8, which is an improvement over the initial value; after 100 iterations, the algorithm converges to the optimal parameter vector X* = [57.8kW, 34.8MPa, 0.53 roots / second], which satisfies all constraints, and the comprehensive evaluation function J' reaches its maximum value of 32.6.

[0072] The real-time optimal process parameter vector is converted into control commands that can be issued and synchronously sent to the welding power supply, servo driver and production line main controller via industrial bus for execution. The new timing data after execution is fed back to the data acquisition terminal, ultimately forming a closed-loop optimization and maintenance control loop from monitoring, prediction, decision-making to execution.

[0073] The core of this step is to implement the optimization results in engineering. Through instruction conversion, reliable transmission, and data feedback, the final link in closed-loop control is completed, ensuring the accurate execution and continuous optimization of process parameters. The specific implementation method is as follows: The command conversion for real-time optimal process parameter vectors needs to be adapted to the control protocols and signal formats of each actuator: the control command for the welding power supply uses a 4-20mA standard current signal. The signal value corresponding to the welding power P=57.8kW in the parameter vector is calculated as I=4+(57.8-55) / (60-55)×(20-4)=12.96mA (linear mapping relationship, 55kW corresponds to 4mA, 60kW corresponds to 20mA). The command includes additional information such as power setpoint, allowable fluctuation range, and start-up delay (10ms); the control of the roll gap servo pressure F=34.8MPa... The control command uses a digital signal from the ProfinetIO protocol, which is converted into a 16-bit binary number (0 corresponds to 30MPa, 65535 corresponds to 40MPa). The calculated binary value is approximately 31457 ((34.8-30) / (40-30))×65535. The command also includes a pressure change rate limit (2MPa / second). The control command with a production line cycle of T=0.53 pieces / second is converted into a pulse signal with a pulse frequency f=T×60=31.8Hz (corresponding to the speed control signal of the main controller). The pulse duty cycle is set to 50% to ensure stable conveyor roller speed.

[0074] Control commands are transmitted via Profinet industrial Ethernet at a rate of 100Mbps with a latency of ≤1ms, meeting real-time control requirements. A triple-protection mechanism is implemented during transmission: a CRC-32 checksum is added during command encapsulation (e.g., the checksum for command data is 0x12345678), and the receiving end returns an acknowledgment frame after verification; a "master-slave" communication mode is adopted, with the control host acting as the master station, issuing commands sequentially in the order of welding power supply, servo driver, and main controller; the slave station receives the commands and provides feedback on the execution status ("received successfully," "executing," "execution completed"); if the master station does not receive an acknowledgment frame, it automatically retransmits the command, with a retransmission count of ≤3. If the retransmission still fails, an audible and visual alarm is triggered (1kHz alarm frequency, flashing red light), and a fault log is recorded (including fault time, actuator identification, and command content).

[0075] The actuator's command execution process is as follows: After receiving a 12.96mA current signal, the welding power supply stabilizes its output power at 57.8kW through its internal PID adjustment module, with power fluctuations controlled within ±1.5kW. Upon receiving binary commands, the servo driver drives the hydraulic system to adjust the roller gap pressure. Real-time feedback from the pressure sensor ensures precise pressure control at 34.8MPa, with a pressure change rate ≤2MPa / second. Upon receiving pulse signals, the production line main controller adjusts the conveyor roller speed to stabilize the cycle time at 0.53 pieces / second, synchronously coordinating the timing of welding, cooling, and sizing processes. After execution, each actuator feeds back its actual process parameters (e.g., actual welding power supply output 57.7kW, servo pressure 34.7MPa, cycle time 0.528 pieces / second) and equipment status (e.g., power supply temperature 45℃, driver current 2.3A) to the control host via the industrial bus.

[0076] Feedback and closed-loop formation of new time-series data: The control host integrates the feedback data into a new time-series operation dataset and transmits it synchronously to the data acquisition terminal to update the data collected by the multi-source sensors; the data acquisition terminal pushes the new data to the feature extraction module, recalculates the high-dimensional feature vector, and inputs it into the prediction model to update the remaining service life and process deviation trend; the digital twin module dynamically adjusts the simulation model based on the new data and reassesses the failure risk and quality defect probability; the preventive maintenance decision scheme updates the maintenance priority and time window according to the new equipment status; the adaptive dynamic programming algorithm optimizes the process parameters again based on the new data and generates the next round of control commands, ultimately forming a closed-loop optimization and maintenance control loop of "monitoring-feedback-prediction-decision-optimization-execution-feedback", realizing the continuous and stable operation of the stainless steel welded pipe production line.

[0077] It is evident that by synchronously collecting time-series operational data through a multi-source sensor array deployed at key nodes of a continuous stainless steel welded pipe production line; performing multi-scale feature extraction and fusion on the time-series operational data to construct a high-dimensional feature vector characterizing the equipment's health status and process stability, and inputting it into a multi-step advanced prediction model, the model outputs the remaining service life prediction curves of key components and the evolution trend of process deviations; based on the remaining service life prediction curves and the evolution trend of process deviations, a preventive maintenance decision scheme including maintenance priorities and time windows is generated; according to the preventive maintenance decision scheme, a set of instructions for the coordinated adjustment of process parameters is generated, thereby improving equipment reliability and process stability, and forming a data-driven closed-loop intelligent maintenance and control system.

[0078] Another embodiment of the present invention provides a preventive maintenance and process parameter optimization system for a continuous stainless steel welded pipe production line, see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to synchronously acquire time-series operation data through a multi-source sensor array deployed at key nodes of the stainless steel welded pipe continuous production line. The time-series operation data includes welding current and voltage waveforms, forming roller gap displacement, pipe temperature field distribution, and equipment vibration spectrum. The construction module 302 is used to perform multi-scale feature extraction and fusion on the time-series operation data, construct a high-dimensional feature vector characterizing the health status of the equipment and the stability of the process, and input it into the multi-step advanced prediction model to output the remaining service life prediction curve of key components and the evolution trend of process deviation. The evaluation module 303 is used to dynamically evaluate potential failure risks and quality defect probabilities based on the remaining useful life prediction curve and process deviation evolution trend through the digital twin simulation module, and generate a preventive maintenance decision plan that includes maintenance priorities and time windows. The optimization module 304 is used to optimize welding power, roll gap servo pressure and production line cycle time parameters in real time using an adaptive dynamic programming algorithm according to the preventive maintenance decision scheme, generate a set of process parameter collaborative adjustment instructions, and form a data-driven closed-loop optimization and maintenance control loop.

[0079] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0080] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0081] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0082] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for preventive maintenance and process parameter optimization in a continuous stainless steel welded pipe production line, characterized in that, The method includes: The timing operation data is collected synchronously by a multi-source sensor array deployed at key nodes of the stainless steel welded pipe continuous production line. The timing operation data includes welding current and voltage waveforms, forming roll gap displacement, pipe temperature field distribution, and equipment vibration spectrum. Multi-scale feature extraction and fusion are performed on the time-series operation data to construct a high-dimensional feature vector characterizing the health status of the equipment and the stability of the process. This vector is then input into a multi-step advanced prediction model, which outputs the remaining service life prediction curve of key components and the evolution trend of process deviation. Based on the remaining useful life prediction curve and the evolution trend of process deviation, the potential failure risk and quality defect probability are dynamically evaluated through the digital twin simulation module, and a preventive maintenance decision plan including maintenance priority and time window is generated. Based on the aforementioned preventive maintenance decision-making scheme, an adaptive dynamic programming algorithm is used to optimize welding power, roll gap servo pressure, and production line cycle time parameters in real time, generating a set of process parameter collaborative adjustment instructions to form a data-driven closed-loop optimization and maintenance control loop.

2. The method according to claim 1, characterized in that, The process involves synchronously collecting time-series operational data via a multi-source sensor array deployed at key nodes of the stainless steel welded pipe continuous production line. This time-series operational data includes welding current and voltage waveforms, forming roller gap displacement, pipe temperature field distribution, and equipment vibration spectrum. Design and deploy a multi-source sensor array, including Hall current and voltage sensors, laser displacement sensors, infrared thermal imager arrays, and vibration acceleration sensors; formulate sampling frequency and installation coordinate specifications for each sensor; and generate a sensor array configuration scheme. According to the sensor array configuration scheme, all sensors are synchronously triggered to collect data through the Industrial Internet of Things protocol, and each data packet is timestamped at the microsecond level using a precise clock source to generate original synchronous data packets with a unified time tag. The original synchronization data packets are preprocessed, wavelet threshold denoising algorithm is applied to eliminate arc interference in welding current, median filtering is used to smooth roll gap displacement signal, and non-uniform temperature field correction algorithm is used to process thermal imager data to generate preprocessed time sequence signal. The preprocessed time series signals are aligned and packaged according to a unified timestamp, and encapsulated into a multi-dimensional time series dataset containing current and voltage waveforms, roll gap displacement sequences, temperature field matrices, and vibration spectra. Finally, a synchronous time series running dataset that can be used for subsequent analysis is generated.

3. The method according to claim 2, characterized in that, The process involves multi-scale feature extraction and fusion of the time-series operational data to construct a high-dimensional feature vector characterizing equipment health status and process stability. This vector is then input into a multi-step forward prediction model, outputting the remaining service life prediction curves of key components and the evolution trend of process deviations. This includes: Multi-scale decomposition is performed on the synchronous time-series running dataset. The empirical mode decomposition method is used to extract the intrinsic mode function components from the vibration spectrum. The sliding window statistical method is used to calculate the time-domain features including the root mean square value and peak factor from the current and voltage waveforms, generating a set of multi-scale feature components. Based on a multi-scale feature component set, a graph convolutional neural network is used to model the spatiotemporal dependencies between different physical quantity features, mapping current, displacement, temperature and vibration features to a unified latent space, generating a deeply fused high-dimensional feature vector. The high-dimensional feature vector is input into a long short-term memory network based on an attention mechanism. This network learns the mapping pattern from historical state to future state through an encoder-decoder structure, and generates a prediction sequence of equipment state and process parameters for a predetermined number of time steps in the future. Based on the predicted sequence, the Weibull proportional hazards model is used to calculate the probability distribution of the remaining useful life of key mechanical components. At the same time, Kalman filtering is used to track the deviation trend of process parameters relative to the set values. Finally, the remaining useful life prediction curve and process deviation evolution trend report are output.

4. The method according to claim 3, characterized in that, Based on the remaining useful life prediction curve and the evolution trend of process deviations, the system dynamically assesses potential failure risks and quality defect probabilities through a digital twin simulation module, generating a preventative maintenance decision plan that includes maintenance priorities and time windows, including: In a digital twin simulation environment, a high-fidelity dynamic model of the production line is constructed. The remaining service life prediction curve and the evolution trend of process deviation are imported as boundary conditions to simulate the equipment operating status and product quality at different future time periods and generate a set of virtual operation scenarios. Based on a set of virtual operating scenarios, the frequency of failures of key components is statistically analyzed using the Monte Carlo simulation method. At the same time, the probability of pipe defects caused by process deviations is calculated, and a failure risk probability distribution and a quality defect probability matrix are generated. By combining the probability of failure risk, the probability of quality defects, and the estimated production downtime losses and maintenance costs, a multi-objective optimization model is established. With the goal of minimizing the overall cost and maximizing the quality pass rate, the optimal intervention point is solved to generate a set of preliminary maintenance timing and measures. Resource constraints are verified on the initial maintenance timing and measures set, taking into account spare parts inventory, maintenance team availability and production plan, to determine the executable time window and execution order of each maintenance task, and finally generate a preventive maintenance decision plan that includes specific maintenance items, priority order and time window.

5. The method according to claim 4, characterized in that, The process involves using an adaptive dynamic programming algorithm to optimize welding power, roll gap servo pressure, and production line cycle time parameters in real time, based on the preventive maintenance decision-making scheme. This generates a set of collaborative adjustment instructions for process parameters, forming a data-driven closed-loop optimization and maintenance control loop. This includes: The preventive maintenance decision-making scheme is analyzed, the equipment information to be maintained and the expected performance degradation of the equipment are extracted, and the equipment information and performance degradation are quantified into constraints for process parameter adjustment to generate a parameter optimization constraint set. A comprehensive evaluation function with the goals of lowest energy consumption, most stable quality, and highest production efficiency is constructed. Welding power, roll gap servo pressure, and production line cycle time are used as decision variables to establish a real-time optimization model for process parameters. An adaptive dynamic programming algorithm is used to solve the optimization model. This algorithm dynamically updates the state transition equation and reward function based on real-time collected process data, searches online for the optimal parameter combination that satisfies the parameter optimization constraint set, and generates a real-time optimal process parameter vector. The real-time optimal process parameter vector is converted into control commands that can be issued and synchronously sent to the welding power supply, servo driver and production line main controller via industrial bus for execution. The new timing data after execution is fed back to the data acquisition terminal, ultimately forming a closed-loop optimization and maintenance control loop from monitoring, prediction, decision-making to execution.

6. A preventive maintenance and process parameter optimization system for a continuous stainless steel welded pipe production line, characterized in that, The system includes: The acquisition module is used to synchronously acquire time-series operation data through a multi-source sensor array deployed at key nodes of the stainless steel welded pipe continuous production line. The time-series operation data includes welding current and voltage waveforms, forming roller gap displacement, pipe temperature field distribution, and equipment vibration spectrum. The construction module is used to perform multi-scale feature extraction and fusion on the time-series operation data, construct a high-dimensional feature vector characterizing the health status of the equipment and the stability of the process, and input it into the multi-step advanced prediction model to output the remaining service life prediction curve of key components and the evolution trend of process deviation. The evaluation module is used to dynamically assess potential failure risks and quality defect probabilities based on the remaining useful life prediction curve and process deviation evolution trend through the digital twin simulation module, and generate a preventive maintenance decision plan that includes maintenance priorities and time windows. The optimization module is used to optimize welding power, roll gap servo pressure and production line cycle time parameters in real time using an adaptive dynamic programming algorithm based on the preventive maintenance decision scheme, and generate a set of process parameter collaborative adjustment instructions to form a data-driven closed-loop optimization and maintenance control loop.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Design and deploy a multi-source sensor array, including Hall current and voltage sensors, laser displacement sensors, infrared thermal imager arrays, and vibration acceleration sensors; formulate sampling frequency and installation coordinate specifications for each sensor; and generate a sensor array configuration scheme. According to the sensor array configuration scheme, all sensors are synchronously triggered to collect data through the Industrial Internet of Things protocol, and each data packet is timestamped at the microsecond level using a precise clock source to generate original synchronous data packets with a unified time tag. The original synchronization data packets are preprocessed, wavelet threshold denoising algorithm is applied to eliminate arc interference in welding current, median filtering is used to smooth roll gap displacement signal, and non-uniform temperature field correction algorithm is used to process thermal imager data to generate preprocessed time sequence signal. The preprocessed time series signals are aligned and packaged according to a unified timestamp, and encapsulated into a multi-dimensional time series dataset containing current and voltage waveforms, roll gap displacement sequences, temperature field matrices, and vibration spectra. Finally, a synchronous time series running dataset that can be used for subsequent analysis is generated.

8. The system according to claim 7, characterized in that, The building module is specifically used for: Multi-scale decomposition is performed on the synchronous time-series running dataset. The empirical mode decomposition method is used to extract the intrinsic mode function components from the vibration spectrum. The sliding window statistical method is used to calculate the time-domain features including the root mean square value and peak factor from the current and voltage waveforms, generating a set of multi-scale feature components. Based on a multi-scale feature component set, a graph convolutional neural network is used to model the spatiotemporal dependencies between different physical quantity features, mapping current, displacement, temperature and vibration features to a unified latent space, generating a deeply fused high-dimensional feature vector. The high-dimensional feature vector is input into a long short-term memory network based on an attention mechanism. This network learns the mapping pattern from historical state to future state through an encoder-decoder structure, and generates a prediction sequence of equipment state and process parameters for a predetermined number of time steps in the future. Based on the predicted sequence, the Weibull proportional hazards model is used to calculate the probability distribution of the remaining useful life of key mechanical components. At the same time, Kalman filtering is used to track the deviation trend of process parameters relative to the set values. Finally, the remaining useful life prediction curve and process deviation evolution trend report are output.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images