A method and system for real-time monitoring and fault early warning of paper machine operating status

Abstract

This application relates to the field of industrial automation and intelligent monitoring technology, and discloses a method and system for real-time monitoring and fault early warning of paper machine operation status. It aims to solve problems in existing technologies such as weak multi-source data fusion capabilities, insufficient fault feature extraction, delayed early warning, and high false alarm rates. The method includes: simultaneously acquiring vibration, temperature, acoustic emission, current, and tension signals through multiple types of sensors; performing time alignment, resampling, and noise reduction processing on the signals; extracting and fusing the physical domain features of each signal using a dedicated sub-network; generating a dynamic state vector through joint encoding using LSTM and an attention mechanism; achieving fault identification and source tracing by combining a fault classifier with historical operating condition matching; and dynamically adjusting the early warning threshold using an exponentially weighted moving average, triggering graded early warnings based on the deviation magnitude and duration. The system is integrated into an edge computing platform, supporting low-latency, high-precision predictive maintenance. This application significantly improves the reliability and maintenance efficiency of paper machine operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial automation and intelligent monitoring, and particularly relates to a method and system for real-time monitoring and fault warning of the operating state of a paper machine. Background Art

[0002] As an important part of process manufacturing, the paper industry has strong production continuity, high equipment complexity, and variable operating conditions, posing extremely high requirements for the stability and reliability of the paper machine system. Modern large paper machines integrate multiple key subsystems such as headboxes, wire sections, press sections, dryer sections, and winding devices. Each component is coupled with each other during high-speed operation. Any minor abnormality in any link may trigger a chain reaction, resulting in quality problems such as paper breaks, wrinkles, and uneven thickness, and even causing unplanned shutdowns. Therefore, real-time, comprehensive, and accurate monitoring of the operating state of the paper machine, and early fault warning based on this, have become the core requirements for improving papermaking efficiency, ensuring product quality, and reducing operation and maintenance costs.

[0003] Among them, the technology of real-time monitoring and fault warning of the operating state of the paper machine aims to dynamically evaluate the health status of the equipment and predict potential failure risks through multi-source sensor data fusion and intelligent analysis means. This technology usually relies on operating signals collected by various sensors such as vibration, temperature, current, acoustic emission, and images, and combines signal processing, feature extraction, and pattern recognition methods to construct a state discrimination model. Its core goal is to identify the performance degradation trend or abnormal signs before the occurrence of a fault, so as to provide a time window for maintenance decision-making.

[0004] In the prior art, most paper machine monitoring systems still adopt an offline analysis method based on threshold alarm or simple statistical models, which are difficult to adapt to the actual challenges such as frequent switching of paper machine operating conditions, strong interference noise, and diverse fault modes. On the one hand, traditional methods lack the ability to deeply fuse multi-physical quantity data and cannot effectively distinguish process fluctuations from real faults; on the other hand, existing warning mechanisms are mostly driven by static rules and do not consider dynamic factors such as equipment aging and load changes, resulting in a high false alarm rate or serious missed alarms. Especially in the scenario of high-speed paper machines, the fault evolution speed is fast and the concealment is strong. If abnormal detection and risk assessment cannot be completed within the time scale of milliseconds to seconds, it will be difficult to support timely intervention. Therefore, there is an urgent need for a method and system for real-time monitoring and fault warning of the operating state of a paper machine that can achieve high timeliness, strong robustness, and adaptive ability. Summary of the Invention

[0005] This invention provides a method and system for real-time monitoring and fault early warning of paper machine operating status, aiming to solve the problems of weak multi-source heterogeneous data fusion capability, insufficient fault feature extraction, delayed early warning response, and high false alarm rate in existing paper machine condition monitoring technologies. By constructing a collaborative analysis architecture based on multimodal perception, deep encoding of temporal features, and adaptive adjustment of dynamic thresholds, high-precision, low-latency, and robust monitoring of the operating status of key paper machine components and early fault warning are achieved.

[0006] According to one aspect of the present invention, a method for real-time monitoring and fault early warning of paper machine operating status is provided, comprising: By deploying multi-type sensor arrays in key parts of the paper machine, vibration signals, temperature signals, acoustic emission signals, motor current signals, and paper web tension signals are collected simultaneously during the operation of the paper machine. The acquired multi-source raw signals are processed by timestamp alignment, sampling frequency unification, and noise suppression filtering to generate a standardized multi-channel time-series data stream. The multi-channel time-series data streams are respectively input to the corresponding dedicated feature extraction sub-networks. Among them, the vibration signal is processed by the wavelet packet decomposition sub-network to extract the frequency band energy distribution features, the temperature signal is processed by the sliding window statistics sub-network to extract the temperature rise rate and steady-state deviation features, the acoustic emission signal is processed by the short-time Fourier transform sub-network to extract the event energy and time-frequency clustering features, the motor current signal is processed by the Hilbert-Huang transform sub-network to extract the instantaneous amplitude and phase modulation features, and the paper web tension signal is processed by the differential accumulation sub-network to extract the tension fluctuation variance and trend offset features. The feature vectors output by each dedicated feature extraction subnetwork are concatenated to form a high-dimensional fusion feature vector. This high-dimensional fusion feature vector is then input into a state representation encoder jointly constructed based on a long short-term memory network and an attention mechanism to generate a dynamic state code that can reflect the overall operational health of the paper machine. Based on the dynamic state code, a pre-trained fault mode classifier determines whether the current operating state belongs to a preset abnormal category set. If it is determined to be abnormal, a fault tracing module based on historical similar working conditions is further activated. This module retrieves the top N historical state samples with the smallest Euclidean distance from the current dynamic state code from the historical database and generates preliminary fault type suggestions based on the known fault labels corresponding to these samples. Meanwhile, based on the numerical change trajectory of the dynamic state code, an adaptive threshold adjustment unit calculates the fault warning threshold under the current working condition in real time. The adaptive threshold adjustment unit uses an exponential weighted moving average algorithm to dynamically update the mean and standard deviation of the historical normal state code, and sets the upper and lower warning boundaries on this basis. When the dynamic status code exceeds the upper and lower warning boundaries, a hierarchical warning mechanism is triggered. Based on the magnitude and duration of the exceedance, a level 1 warning, level 2 warning, or level 3 warning instruction is generated, and the warning instruction, along with fault type suggestions, relevant feature contribution ranking, and recommended maintenance measures, is pushed to the human-machine interaction terminal.

[0007] In one embodiment of the present invention, the multi-type sensor array includes: a triaxial accelerometer mounted on the drying cylinder bearing housing for acquiring radial, axial, and tangential vibration data; a distributed fiber optic temperature sensor embedded inside the press roll for acquiring the temperature field distribution along the circumferential and axial directions of the roll surface; an acoustic emission sensor arranged near the vacuum roll for capturing high-frequency elastic waves caused by paper web breakage or roll surface damage; a Hall effect current sensor connected in series in the power supply circuit of the drive motor for monitoring instantaneous fluctuations in the operating current; and a tension sensor disposed between the sizing section and the winding section for measuring the real-time tension value of the paper web during the traction process.

[0008] In one embodiment of the present invention, the timestamp alignment adopts a global clock source based on hardware synchronization triggering. All sensor data acquisition modules are connected to the same field-programmable gate array controller, which emits synchronous sampling pulses at a reference frequency of 10 kHz to ensure that the data of each channel are aligned within the microsecond level. The sampling frequency unification process adopts cubic spline interpolation to resample signals with different original sampling rates to a unified 5 kHz frequency. The noise suppression filtering adopts differentiated strategies for different signal types. The vibration signal adopts a fifth-order Butterworth bandpass filter with a passband range of 10 Hz to 2 kHz. The temperature signal adopts a median filter combined with a first-order low-pass filter with a cutoff frequency of 0.5 Hz. The acoustic emission signal adopts wavelet threshold denoising, using the db4 wavelet basis function and a soft threshold rule. The motor current signal adopts a moving average filter with a window length of 50 sampling points. The paper tension signal adopts Kalman filtering for state estimation to eliminate measurement noise.

[0009] In one embodiment of the present invention, the wavelet packet decomposition subnetwork performs five-level decomposition on the vibration signal, generating thirty-two frequency bands, and calculates the energy proportion of each frequency band as a feature; the sliding window statistical subnetwork uses a sliding window with a width of 300 seconds, calculates the mean of the first derivative of the temperature signal within the window as the temperature rise rate, and calculates the difference between the mean of the last 60 seconds of the window and the mean of the first 60 seconds of the window as the steady-state deviation; the short-time Fourier transform subnetwork uses a Hanning window with a window length of 1024 points and an overlap rate of 75%, and calculates the sum of the energy of components with frequencies greater than 20 kHz within each time frame as an event. The energy is used to calculate the dispersion of the centroid coordinates of the energy in the time-frequency plane as the time-frequency clustering degree. The Hilbert-Huang transform subnetwork first performs empirical mode decomposition on the current signal, extracts the first three intrinsic mode function components, and then performs Hilbert transform on each component to obtain the instantaneous amplitude sequence. The standard deviation of the sequence is calculated as the amplitude modulation feature, and the second-order root mean square difference of the instantaneous phase sequence is calculated as the phase modulation feature. The differential accumulation subnetwork performs first-order difference on the tension signal, calculates the standard deviation of the difference sequence as the fluctuation variance, performs linear fitting on the original tension signal, and takes the cumulative sum of the fitting residuals as the trend offset.

[0010] As one embodiment of the present invention, the state representation encoder is composed of two stacked long short-term memory units, each containing 128 hidden units, followed by a multi-head attention mechanism module. This module contains four parallel attention heads, each with a dimension of 32. The high-dimensional fusion feature vector is weighted and aggregated through query, key, and value matrices, and finally outputs a 60-dimensional dynamic state encoding vector.

[0011] In one embodiment of the present invention, the fault mode classifier is a fully connected neural network containing three hidden layers with 256, 128, and 64 nodes respectively. The output layer uses the softmax activation function and corresponds to twelve preset typical fault modes, including wear of the drying cylinder bearing, high deformation of the press roll, blockage of the vacuum roll, scratches on the sizing roll, broken teeth of the transmission gear, short circuit between turns of the motor winding, paper web breakage, slippage of the tension roll, eccentricity of the guide roll, leakage of the hydraulic system, failure of the lubrication system, and interference of the electrical control system.

[0012] As one embodiment of the present invention, the historical database stores at least 100,000 sets of labeled samples. Each set of samples contains complete multi-channel time-series data, corresponding dynamic state codes, fault labels, operating parameters (including vehicle speed, quantity, and moisture content), and maintenance records. The Euclidean distance calculation is performed in a sixty-dimensional dynamic state code space. The ten historical samples with the smallest distance are selected. If more than seven of these samples have the same fault label, the label is used as a preliminary fault type suggestion.

[0013] In one embodiment of the present invention, the smoothing factor α of the exponentially weighted moving average algorithm is set to 0.95, which is used to update the mean μ and standard deviation σ of the normal state code. The upper and lower warning boundaries are set to μ plus or minus three times σ, respectively. When the dynamic state code exceeds the upper or lower boundary for five consecutive sampling periods, a level 3 warning is triggered. When the exceedance exceeds four times σ and lasts for three periods, a level 2 warning is triggered. When the exceedance exceeds five times σ or is accompanied by an anomaly judgment with a confidence level higher than 90% by the fault mode classifier, a level 1 warning is triggered.

[0014] According to another aspect of the present invention, a real-time monitoring and fault early warning system for paper machine operation status is provided, comprising: The multi-source signal synchronous acquisition unit is used to synchronously acquire vibration signals, temperature signals, acoustic emission signals, motor current signals, and paper web tension signals through a multi-type sensor array deployed in key parts of the paper machine; The signal preprocessing unit is used to perform timestamp alignment, sampling frequency unification, and noise suppression filtering on the acquired multi-source raw signals to generate a standardized multi-channel time-series data stream. The multimodal feature extraction unit is used to input the multi-channel time-series data streams into the corresponding dedicated feature extraction sub-networks, extract the physical domain features of various signals, and splice them to form a high-dimensional fusion feature vector; The dynamic state coding unit is used to input the high-dimensional fused feature vector into the state representation encoder jointly constructed based on a long short-term memory network and an attention mechanism to generate dynamic state codes. The fault identification and tracing unit is used to determine abnormal states based on the dynamic state code and through a fault mode classifier, and to activate the fault tracing module when an abnormality is determined to be present, and to generate fault type suggestions. An adaptive warning threshold calculation unit is used to dynamically update the warning threshold boundary based on the historical trajectory of the dynamic state code using an exponentially weighted moving average algorithm. The graded early warning decision and output unit is used to generate graded early warning instructions based on the comparison results of dynamic status codes and early warning boundaries, combined with the duration and the extent of exceedance, and push the early warning information to the human-computer interaction terminal.

[0015] In one embodiment of the present invention, each sensor in the multi-source signal synchronous acquisition unit communicates with the central data acquisition station via the industrial Ethernet protocol, with a data transmission delay of less than one millisecond. The signal preprocessing unit, multimodal feature extraction unit, dynamic state coding unit, fault identification and tracing unit, adaptive early warning threshold calculation unit, and hierarchical early warning decision and output unit are all integrated into an embedded edge computing platform. This platform adopts a multi-core processor architecture with a main frequency of not less than 2.5 gigahertz, is equipped with 8 gigabytes of running memory and 256 gigabytes of solid-state storage, and uses a real-time Linux kernel as its operating system, with a task scheduling cycle of ten milliseconds.

[0016] In summary, this application includes at least one of the following beneficial technical effects: This invention effectively overcomes the problem of missed alarms caused by the limited information from a single sensor by constructing a monitoring framework that combines multimodal perception and deep temporal modeling. A dedicated feature extraction sub-network is used to perform refined feature mining for different types of physical signals, significantly improving the ability to capture fault-sensitive features. A dynamic state encoder based on an attention mechanism is introduced, which can adaptively focus on the feature dimension most discriminative to the current operating condition, enhancing the model's generalization performance under varying operating conditions. Fault tracing is achieved through historical operating condition matching, avoiding the lack of diagnostic basis in traditional threshold alarms. The adaptive early warning threshold mechanism can dynamically evolve with equipment aging, environmental changes, and production parameter adjustments, significantly reducing the false alarm rate of fixed threshold strategies in long-term operation. The hierarchical early warning strategy, combined with the degree and duration of state deviation, provides operators with clear priorities for handling, effectively supporting predictive maintenance decisions. Overall, this invention achieves a technological leap from "passive response maintenance" to "proactive predictive maintenance," significantly improving the reliability, safety, and production efficiency of paper machine operation. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the core principle framework of the collaborative analysis architecture based on multimodal perception, temporal feature deep encoding, and dynamic threshold adaptive adjustment in this invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of specific embodiments based on the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0019] This invention provides a method and system for real-time monitoring and fault early warning of paper machine operating status, aiming to solve the problems of weak multi-source heterogeneous data fusion capability, insufficient fault feature extraction, delayed early warning response, and high false alarm rate in existing paper machine condition monitoring technologies. By constructing a collaborative analysis architecture based on multimodal perception, deep encoding of temporal features, and adaptive adjustment of dynamic thresholds, high-precision, low-latency, and robust monitoring of the operating status of key paper machine components and early fault warning are achieved.

[0020] The method includes the following steps: S1, through a multi-type sensor array deployed in key parts of the paper machine, synchronously collects vibration signals, temperature signals, acoustic emission signals, motor current signals and paper web tension signals during the operation of the paper machine; S2 performs timestamp alignment, sampling frequency unification processing, and noise suppression filtering on the acquired multi-source raw signals to generate a standardized multi-channel time-series data stream. S3, the multi-channel time-series data streams are respectively input to the corresponding dedicated feature extraction sub-networks. Among them, the vibration signal is extracted by the wavelet packet decomposition sub-network to extract the frequency band energy distribution features, the temperature signal is extracted by the sliding window statistics sub-network to extract the temperature rise rate and steady-state deviation features, the acoustic emission signal is extracted by the short-time Fourier transform sub-network to extract the event energy and time-frequency clustering features, the motor current signal is extracted by the Hilbert-Huang transform sub-network to extract the instantaneous amplitude and phase modulation features, and the paper web tension signal is extracted by the differential accumulation sub-network to extract the tension fluctuation variance and trend offset features. S4. The feature vectors output by each dedicated feature extraction sub-network are concatenated to form a high-dimensional fusion feature vector. This high-dimensional fusion feature vector is then input into a state representation encoder jointly constructed based on a long short-term memory network and an attention mechanism to generate a dynamic state code that can reflect the overall operational health of the paper machine. S5. Based on the dynamic state code, a pre-trained fault mode classifier is used to determine whether the current operating state belongs to a preset abnormal category set. If it is determined to be abnormal, a fault tracing module based on historical similar working conditions is further activated. This module retrieves the top N historical state samples with the smallest Euclidean distance from the current dynamic state code from the historical database, and generates preliminary fault type suggestions based on the known fault labels corresponding to these samples. S6. Based on the numerical change trajectory of the dynamic state code, an adaptive threshold adjustment unit calculates the fault warning threshold under the current working condition in real time. The adaptive threshold adjustment unit uses an exponential weighted moving average algorithm to dynamically update the mean and standard deviation of the historical normal state code, and sets the upper and lower warning boundaries on this basis. S7. When the dynamic state code exceeds the upper and lower warning boundaries, a hierarchical warning mechanism is triggered. Based on the magnitude and duration of the exceedance, a level 1 warning, level 2 warning, or level 3 warning instruction is generated, and the warning instruction, along with fault type suggestions, relevant feature contribution ranking, and recommended maintenance measures, is pushed to the human-machine interaction terminal.

[0021] In step S1, the multi-type sensor array includes a triaxial accelerometer mounted on the dryer cylinder bearing housing for acquiring radial, axial, and tangential vibration data; a distributed fiber optic temperature sensor embedded inside the press roll for acquiring the temperature field distribution along the circumferential and axial directions of the roll surface; an acoustic emission sensor arranged near the vacuum roll for capturing high-frequency elastic waves caused by paper web breakage or roll surface damage; a Hall effect current sensor connected in series in the drive motor power supply circuit for monitoring instantaneous fluctuations in the operating current; and a tension sensor located between the sizing section and the winding section for measuring the real-time tension value of the paper web during the traction process. All sensors have industrial-grade protection and can operate stably for a long time in the high-temperature, high-humidity, and strong electromagnetic interference environment of a paper mill. The sensor data acquisition module is connected to the same field-programmable gate array controller via a hardware synchronous triggering mechanism. This controller emits synchronous sampling pulses at a 10 kHz reference frequency to ensure that the data from each channel is time-aligned within microsecond precision, avoiding phase distortion and feature misalignment caused by asynchronous sampling.

[0022] In step S2, the multi-source raw signals are preprocessed. First, timestamp alignment is performed. Using the global clock source provided by the field-programmable gate array controller, the data frames of all sensor channels are reindexed according to a unified time base to eliminate the offset introduced by transmission delay and startup timing differences. Then, sampling frequency unification is performed. Cubic spline interpolation is used to resample the signals with different original sampling rates to a unified 5 kHz frequency. For vibration signals, the original sampling rate is 10 kHz, and high-frequency dynamic characteristics are preserved after interpolation; the original sampling rate of temperature signals is 1 Hz, and after upsampling, the time dimension is consistent with high-speed signals; the original sampling rate of acoustic emission signals is 50 kHz, and downsampling is required to reduce computational load; the original sampling rates of motor current and tension signals are 2 kHz and 1 kHz, respectively, and both are upsampled to the target frequency. After resampling, noise suppression filtering is performed: the vibration signal uses a fifth-order Butterworth bandpass filter with a passband range of 10 Hz to 2 kHz to effectively filter out low-frequency mechanical vibration and high-frequency electromagnetic interference; the temperature signal uses median filtering combined with a first-order low-pass filter, with a median filter window length of 15 points to remove spike outliers, and then passes through a low-pass filter with a cutoff frequency of 0.5 Hz to smooth thermal drift noise; the acoustic emission signal uses wavelet thresholding for denoising, selecting the db4 wavelet basis function for six-level decomposition, using a soft thresholding rule, with the threshold determined by a general threshold formula; the motor current signal uses moving average filtering with a window length of 50 sampling points to suppress ripple introduced by the switching power supply; the paper web tension signal uses Kalman filtering for state estimation, establishing a one-dimensional linear state-space model, with the process noise covariance set to 0.01 and the observation noise covariance set to 0.005, and the optimal tension estimate is updated recursively to effectively eliminate the coupling noise of sensor drift and mechanical vibration. After the above processing, a standardized time-series data stream with five channels, synchronization, equal frequency, and noise reduction is generated. Each channel has a data length of ten consecutive seconds, corresponding to 50,000 sampling points.

[0023] In step S3, multimodal feature extraction is performed on the standardized time-series data stream. The vibration signal is input into a wavelet packet decomposition subnetwork, which performs a five-layer complete binary tree decomposition, generating thirty-two equal-width frequency bands with center frequencies uniformly distributed from 0 Hz to 2500 Hz. The energy of the signal within each frequency band is calculated, i.e., the sum of squares of the wavelet coefficients of that band, and then divided by the total energy to obtain the normalized energy percentage, forming a thirty-two-dimensional frequency band energy distribution feature vector. The temperature signal is input into a sliding window statistical subnetwork, using a sliding window with a width of 300 seconds, sliding forward in 10-second steps. Within each window, the temperature sequence is first-order differencing, and the arithmetic mean of the difference values ​​is calculated as the temperature rise rate feature. Simultaneously, the window is divided into an initial 60-second segment, a middle 180-second segment, and a final 60-second segment, and the difference between the mean of the final segment and the mean of the initial segment is calculated as the steady-state deviation feature, reflecting the degree of deviation in the equipment's thermal equilibrium state. The acoustic emission signal is input into a short-time Fourier transform subnetwork using a Hanning window with a length of 1024 points and an overlap rate of 75%, generating a time-frequency spectrum. Energy at all frequencies greater than 20 kHz is extracted and summed to obtain the event energy characteristics. The centroid coordinates of this high-frequency band energy on the time-frequency plane are calculated, and the dispersion (standard deviation) of the centroid coordinates for each time frame is calculated as a time-frequency clustering characteristic to characterize the spatial concentration of acoustic emission events. The motor current signal is input into a Hilbert-Huang transform subnetwork. First, empirical mode decomposition is performed, iteratively selecting the top three intrinsic mode function components, covering the main dynamic components. A Hilbert transform is performed on each component to obtain the instantaneous amplitude sequence and instantaneous phase sequence. The weighted average of the standard deviations of the three amplitude sequences (weighted by the energy proportion of each component) is calculated as the amplitude modulation characteristic. The phase sequence is subjected to second-order difference, and its root mean square is taken as the phase modulation characteristic, reflecting the degree of current waveform distortion. The paper tension signal is input into a differential accumulation subnetwork. First, first-order differencing is performed, and the standard deviation of the differencing sequence is calculated as the fluctuation variance feature. Then, linear least-squares fitting is performed on the original tension sequence to obtain a trend line. The cumulative sum of the fitting residuals is calculated as the trend offset feature, which is used to detect slowly developing tension runaway. The output feature dimensions of each subnetwork are as follows: vibration (32 dimensions), temperature (2 dimensions), acoustic emission (2 dimensions), current (2 dimensions), and tension (2 dimensions), for a total of 40 feature vectors.

[0024] In step S4, the forty-dimensional feature vector is concatenated to form a high-dimensional fused feature vector. This vector is input to the state representation encoder, which consists of two stacked layers of long short-term memory units, each containing 128 hidden units. The first layer receives the forty-dimensional input and outputs a 128-dimensional hidden state; the second layer takes the output of the first layer as input and generates a similarly 128-dimensional deep temporal representation. Subsequently, a multi-head attention mechanism module is connected. This module contains four parallel attention heads, each whose query, key, and value projection matrix maps the input to a 32-dimensional subspace. Each head independently calculates attention weights, performs weighted aggregation on the input sequence, and then concatenates the outputs of the four heads and maps them back to a 60-dimensional space through a linear transformation, ultimately outputting a 60-dimensional dynamic state encoding vector. This encoding vector is updated once per sampling period, fully representing the overall health status of the current paper machine operation. Each dimension implicitly contains the nonlinear coupling relationships and operational sensitivity between different physical signals.

[0025] In step S5, the dynamic state code is input into a pre-trained fault mode classifier. This classifier is a fully connected neural network containing three hidden layers with 256, 128, and 64 nodes respectively. The activation function is a modified linear unit, and the output layer uses a softmax activation function, corresponding to twelve typical fault modes, including dry cylinder bearing wear, high deformation of the press roll, vacuum roll blockage, sizing roll scratches, broken transmission gear teeth, short circuit between motor winding turns, paper web breakage, tension roll slippage, guide roll eccentricity, hydraulic system leakage, lubrication system failure, and electrical control system interference. If the maximum output probability exceeds a preset threshold of 0.85, it is determined to be an abnormal state. Once an abnormality is determined, the fault tracing module is immediately activated. This module accesses a historical database containing at least 100,000 labeled samples. Each sample contains complete multi-channel time-series data, a corresponding 60-dimensional dynamic state code, fault labels, operating parameters (vehicle speed, quantity, moisture content), and maintenance records. The module calculates the Euclidean distance between the current code and all normal and abnormal sample codes in the 60-dimensional coding space, and selects the top ten historical samples with the smallest distance. If seven or more of these samples have the same fault label, that label is output as a preliminary fault type suggestion.

[0026] In step S6, the adaptive threshold adjustment unit updates the warning boundary in real time based on the historical trajectory of the dynamic state code. This unit maintains two exponentially weighted moving average variables: the mean μ and the standard deviation σ. The initial values ​​are determined from normal operating data during the equipment's factory commissioning phase. In subsequent operation, whenever a new dynamic state code z_t is received, it is updated according to the following formula: The smoothing factor α is set to 0.95, giving higher weight to recent data. The upper and lower warning boundaries are set to μ_t ± 3σ_t, respectively. This mechanism allows the threshold to evolve dynamically with equipment aging, changes in environmental temperature and humidity, and production parameters (such as increased vehicle speed leading to a rise in the vibration baseline), preventing fixed thresholds from becoming ineffective during long-term operation.

[0027] In step S7, the hierarchical early warning decision logic is as follows: If the dynamic state code exceeds the upper or lower boundary for five consecutive sampling periods, a level 3 early warning is triggered, indicating a potential abnormal trend that requires enhanced monitoring; if the exceedance exceeds four times σ_t and continues for three consecutive periods, a level 2 early warning is triggered, indicating a significant anomaly and suggesting planned maintenance; if the exceedance exceeds five times σ_t, or is accompanied by an anomaly judgment with a confidence level higher than 90% from the fault mode classifier, a level 1 early warning is triggered, indicating a serious fault risk that requires immediate shutdown for inspection. After the early warning command is generated, the system synchronously extracts the contribution of each feature dimension to the current state code, determines the dominant features through gradient backpropagation or attention weight analysis, sorts them, and encapsulates them together with fault type suggestions and recommended maintenance measures (such as "check the lubrication of the drying cylinder bearing" and "calibrate the position of the tension roller") into a structured message, which is then pushed to the human-machine interface terminal in the central control room via industrial Ethernet for operators to make decisions.

[0028] The system comprises a multi-source signal synchronous acquisition unit, a signal preprocessing unit, a multi-modal feature extraction unit, a dynamic state coding unit, a fault identification and tracing unit, an adaptive early warning threshold calculation unit, and a hierarchical early warning decision and output unit. Each sensor in the multi-source signal synchronous acquisition unit communicates with the central data acquisition station via the industrial Ethernet protocol, with a data transmission latency of less than one millisecond. All processing units are integrated into an embedded edge computing platform. This platform employs a multi-core processor architecture with a clock speed of at least 2.5 GHz, equipped with 8 gigabytes of RAM and 256 gigabytes of solid-state storage. The operating system is a real-time Linux kernel, with a task scheduling cycle of ten milliseconds, ensuring that the end-to-end latency from data acquisition to early warning output does not exceed one hundred milliseconds. The system supports online model updates and can receive optimized classifier weights or feature extraction parameters from the cloud via a secure channel, enabling continuous model evolution.

[0029] This embodiment, through the above-described method and system, achieves comprehensive perception, in-depth analysis, and intelligent early warning of the paper machine's operating status, effectively solving problems such as information silos, shallow features, and rigid thresholds in traditional monitoring methods, and providing reliable technical support for the intelligent operation and maintenance of the paper industry.

[0030] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0031] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for real-time monitoring and fault early warning of paper machine operating status, characterized in that, include: By deploying multi-type sensor arrays in key parts of the paper machine, vibration signals, temperature signals, acoustic emission signals, motor current signals, and paper web tension signals are collected simultaneously during the operation of the paper machine. The acquired multi-source raw signals are processed by timestamp alignment, sampling frequency unification, and noise suppression filtering to generate a standardized multi-channel time-series data stream. The multi-channel time-series data streams are respectively input to the corresponding dedicated feature extraction sub-networks. Among them, the vibration signal is processed by the wavelet packet decomposition sub-network to extract the frequency band energy distribution features, the temperature signal is processed by the sliding window statistics sub-network to extract the temperature rise rate and steady-state deviation features, the acoustic emission signal is processed by the short-time Fourier transform sub-network to extract the event energy and time-frequency clustering features, the motor current signal is processed by the Hilbert-Huang transform sub-network to extract the instantaneous amplitude and phase modulation features, and the paper web tension signal is processed by the differential accumulation sub-network to extract the tension fluctuation variance and trend offset features. The feature vectors output by each dedicated feature extraction subnetwork are concatenated to form a high-dimensional fusion feature vector. This high-dimensional fusion feature vector is then input into a state representation encoder jointly constructed based on a long short-term memory network and an attention mechanism to generate a dynamic state code that can reflect the overall operational health of the paper machine. Based on the dynamic state code, a pre-trained fault mode classifier determines whether the current operating state belongs to a preset abnormal category set. If it is determined to be abnormal, a fault tracing module based on historical similar working conditions is further activated. This module retrieves the top N historical state samples with the smallest Euclidean distance from the current dynamic state code from the historical database and generates preliminary fault type suggestions based on the known fault labels corresponding to these samples. Based on the numerical change trajectory of the dynamic state code, an adaptive threshold adjustment unit calculates the fault warning threshold under the current working condition in real time. The adaptive threshold adjustment unit uses an exponential weighted moving average algorithm to dynamically update the mean and standard deviation of the historical normal state code, and sets the upper and lower warning boundaries on this basis. When the dynamic status code exceeds the upper and lower warning boundaries, a hierarchical warning mechanism is triggered. Based on the magnitude and duration of the exceedance, a level 1 warning, level 2 warning, or level 3 warning instruction is generated, and the warning instruction, along with fault type suggestions, relevant feature contribution ranking, and recommended maintenance measures, is pushed to the human-machine interaction terminal.

2. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 1, characterized in that, The acquired multi-source raw signals are subjected to timestamp alignment, sampling frequency unification, and noise suppression filtering to generate a standardized multi-channel time-series data stream, including: Using the global clock source provided by the field programmable gate array controller, a synchronous sampling pulse is emitted at a reference frequency of 10 kHz to re-index the data frames of all sensor channels according to a unified time base to complete the timestamp alignment. The cubic spline interpolation method is used to resample signals with different original sampling rates to a uniform 5 kHz frequency; For vibration signals, a fifth-order Butterworth bandpass filter is used for noise suppression, with a passband range of 10 Hz to 2 kHz. For temperature signals, a median filter combined with a first-order low-pass filter is used, with a cutoff frequency of 0.5 Hz. For acoustic emission signals, wavelet thresholding is used for noise reduction, with a db4 wavelet basis function and a soft thresholding rule. For motor current signals, a moving average filter is used, with a window length of fifty sampling points. For paper web tension signals, Kalman filtering is used for state estimation to eliminate measurement noise.

3. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 2, characterized in that, The multi-channel time-series data streams are respectively input into the corresponding dedicated feature extraction sub-networks, including: The vibration signal is subjected to five-level wavelet packet decomposition to generate thirty-two frequency bands. The energy proportion of each frequency band is calculated as the frequency band energy distribution feature. A sliding window with a width of 300 seconds is used for the temperature signal. The average value of the first derivative of the temperature signal within the window is calculated as the temperature rise rate. The difference between the average value of the last 60 seconds of the window and the average value of the first 60 seconds of the window is calculated as the steady-state deviation. The acoustic emission signal was subjected to a short-time Fourier transform using a Hanning window with a window length of 1024 points and an overlap rate of 75%. The total energy of components with frequencies greater than 20 kHz was calculated as the event energy, and the centroid coordinate dispersion of this energy on the time-frequency plane was calculated as the time-frequency clustering. The motor current signal is first subjected to empirical mode decomposition to extract the first three intrinsic mode function components. Then, Hilbert transform is performed on each component, and the standard deviation of the instantaneous amplitude sequence is calculated as the amplitude modulation feature. The second-order root mean square difference of the instantaneous phase sequence is calculated as the phase modulation feature. After performing first-order difference on the paper web tension signal, the standard deviation of the difference sequence is calculated as the fluctuation variance. The original tension signal is then linearly fitted, and the cumulative sum of the fitting residuals is taken as the trend offset.

4. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 3, characterized in that, The feature vectors output by each dedicated feature extraction subnetwork are concatenated to form a high-dimensional fused feature vector. This high-dimensional fused feature vector is then input into a state representation encoder jointly constructed based on a long short-term memory network and an attention mechanism to generate dynamic state encoding, including: The forty-dimensional feature vectors are concatenated through channels to form a high-dimensional fused feature vector. The high-dimensional fused feature vector is input into a state representation encoder consisting of two stacked long short-term memory units, each containing 128 hidden units, to generate a 128-dimensional deep temporal representation. The deep temporal representation input includes a multi-head attention mechanism module with four parallel attention heads, each with a dimension of thirty-two. The input is weighted and aggregated through query, key, and value matrices, and finally outputs a sixty-dimensional dynamic state encoding vector as the dynamic state encoding.

5. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 4, characterized in that, Based on the dynamic state encoding, a pre-trained fault mode classifier determines whether the current operating state belongs to a preset set of abnormal categories. If it is determined to be abnormal, the fault tracing module is activated, including: The dynamic state code is input into a fault mode classifier in the form of a fully connected neural network. The classifier contains three hidden layers with 256, 128 and 64 nodes respectively. The output layer uses the softmax activation function and corresponds to twelve typical fault modes. If the maximum output probability exceeds the preset threshold of 0.85, it is determined to be an abnormal state; After determining that it is an abnormal state, retrieve the top ten historical state samples with the smallest Euclidean distance in sixty-dimensional space from the historical database that are the current dynamic state code; If seven or more samples have the same fault label, then that label will be output as a preliminary fault type suggestion.

6. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 5, characterized in that, Based on the numerical change trajectory of the dynamic state code, the fault warning threshold under the current operating condition is calculated in real time by the adaptive threshold adjustment unit, including: Maintain two exponentially weighted moving average variables: mean μ and standard deviation σ, with initial values ​​determined from normal operating data during the equipment's factory commissioning phase; Whenever a new dynamic state code z_t is received, according to the formula and Update the mean and variance, where the smoothing factor α is 0.95; Set the upper and lower warning boundaries as follows: .

7. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 6, characterized in that, When the dynamic state code exceeds the upper and lower warning boundaries, a tiered warning mechanism is triggered, including: If the dynamic state code exceeds the upper or lower boundary for five consecutive sampling periods, a level 3 warning is triggered. If the amplitude exceeds four times σ_t and continues for three consecutive cycles, a level two warning will be triggered; If the magnitude exceeds five times σ_t, or if the fault mode classifier outputs an anomaly judgment with a confidence level higher than 90%, a level one warning will be triggered. The warning instructions, along with fault type suggestions, ranking of relevant feature contributions, and recommended maintenance measures, are packaged into a structured message and pushed to the human-machine interaction terminal.

8. The method for real-time monitoring and fault early warning of paper machine operating status according to claim 7, characterized in that, The multi-type sensor array includes: a triaxial accelerometer mounted on the bearing housing of the drying cylinder; a distributed fiber optic temperature sensor embedded inside the press roll; an acoustic emission sensor arranged near the vacuum roll; a Hall effect current sensor connected in series in the power supply circuit of the drive motor; and a tension sensor disposed between the sizing section and the paper winding section.

9. A real-time monitoring and fault early warning system for paper machine operation status, characterized in that, include: The multi-source signal synchronous acquisition unit is used to synchronously acquire vibration signals, temperature signals, acoustic emission signals, motor current signals, and paper web tension signals through a multi-type sensor array deployed in key parts of the paper machine; The signal preprocessing unit is used to perform timestamp alignment, sampling frequency unification, and noise suppression filtering on the acquired multi-source raw signals to generate a standardized multi-channel time-series data stream. The multimodal feature extraction unit is used to input the multi-channel time-series data streams into the corresponding dedicated feature extraction sub-networks, extract the physical domain features of various signals, and splice them to form a high-dimensional fusion feature vector; The dynamic state coding unit is used to input the high-dimensional fused feature vector into the state representation encoder jointly constructed based on a long short-term memory network and an attention mechanism to generate dynamic state codes. The fault identification and tracing unit is used to determine abnormal states based on the dynamic state code and through a fault mode classifier, and to activate the fault tracing module when an abnormality is determined to be present, and to generate fault type suggestions. An adaptive warning threshold calculation unit is used to dynamically update the warning threshold boundary based on the historical trajectory of the dynamic state code using an exponentially weighted moving average algorithm. The graded early warning decision and output unit is used to generate graded early warning instructions based on the comparison results of dynamic status codes and early warning boundaries, combined with the duration and the extent of exceedance, and push the early warning information to the human-computer interaction terminal.

10. The real-time monitoring and fault early warning system for paper machine operation status according to claim 9, characterized in that, The signal preprocessing unit is used for: Using the global clock source provided by the field programmable gate array controller, a synchronous sampling pulse is emitted at a reference frequency of 10 kHz to re-index the data frames of all sensor channels according to a unified time base to complete the timestamp alignment. The cubic spline interpolation method is used to resample signals with different original sampling rates to a uniform 5 kHz frequency; For vibration signals, a fifth-order Butterworth bandpass filter is used for noise suppression, with a passband range of 10 Hz to 2 kHz. For temperature signals, a median filter combined with a first-order low-pass filter is used, with a cutoff frequency of 0.5 Hz. For acoustic emission signals, wavelet thresholding is used for noise reduction, with a db4 wavelet basis function and a soft thresholding rule. For motor current signals, a moving average filter is used, with a window length of fifty sampling points. For paper web tension signals, Kalman filtering is used for state estimation to eliminate measurement noise.

Images