Papermaking machine fault diagnosis method based on multi-channel grouping

Abstract

The invention relates to the technical field of industrial equipment fault diagnosis, and discloses a paper machine fault diagnosis method based on multichannel grouping, which comprises the following steps: firstly, preprocessing and re-sampling signals of different physical types, and grouping according to a vibration group, a pressure and current group and a temperature group; inputting each group of time sequence windows into a multi-scale time-frequency dynamic attention module to extract time-frequency attention features, obtaining gently changed time sequence representation through slow feature layer constraint, reconstructing by an intra-group long and short term memory auto-encoder, and calculating a time sequence point reconstruction error and a window aggregation error; and finally, completing fault judgment by combining the static threshold value and a self-adaptive dynamic threshold value based on normal window sliding statistics. The method realizes high-precision modeling and adaptive anomaly detection of multi-channel complex time sequence signals, has the advantages of high diagnosis precision, low false alarm rate, small detection delay and the like, and is suitable for online monitoring and intelligent operation and maintenance of the paper machine.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment fault diagnosis technology, specifically relating to a paper machine fault diagnosis method based on multi-channel grouping. Background Technology

[0002] As the core equipment of the modern papermaking industry, the papermaking machine encompasses key processes such as pulp forming, pressing and dewatering, and drying and setting. Its operational stability directly impacts product quality and production efficiency. However, the papermaking machine has a complex structure (containing thousands of components including drying cylinders, pressure rollers, drive belts, and vacuum pumps) and operates under harsh conditions of high temperature, high humidity, and high mechanical load for extended periods, making it prone to typical malfunctions such as bearing wear, drying cylinder temperature imbalance, drive belt slippage, and roller eccentricity. Therefore, achieving real-time status monitoring and accurate fault diagnosis of the papermaking machine is crucial for ensuring production safety and reducing operation and maintenance costs.

[0003] In current industrial scenarios, existing methods for fault diagnosis of papermaking machines, such as Principal Component Analysis (PCA) and Long Short-Term Memory (LSTM) network models, have the following shortcomings:

[0004] 1. Feature redundancy: Multi-source sensor data contains a large number of irrelevant features, which reduces the sensitivity of the model.

[0005] 2. Insufficient time-series dependence: Fault signals often exhibit long-period gradual changes, which are difficult for traditional models to capture.

[0006] 3. Poor adaptability of dynamic thresholds: Fixed thresholds are prone to false alarms and missed alarms.

[0007] With the popularization of Industrial Internet of Things (IIoT) technology, machine learning-based fault diagnosis methods are gradually being applied to paper machine monitoring, but the following bottlenecks still exist:

[0008] High-dimensional data noise interference: The multi-source sensor data (vibration, temperature, pressure, etc.) of papermaking machines have high dimensionality (usually more than 10 dimensions) and are significantly affected by environmental noise (such as electromagnetic interference in the workshop and mechanical vibration crosstalk). Traditional dimensionality reduction methods (such as PCA and t-distributed random neighborhood embedding (t-SNE)) are difficult to effectively distinguish key features from noise.

[0009] Insufficient modeling of long-term time-dependent features: Fault signals often exhibit long-period gradual change characteristics. Although traditional Long Short-Term Memory (LSTM) network models can alleviate the gradient vanishing problem, their memory capacity still deteriorates when faced with long sequences, making it difficult to capture long-period gradual change features.

[0010] Poor adaptability of dynamic thresholds: Existing methods mostly use fixed thresholds (such as the 3σ principle) to determine anomalies, but the operating conditions of paper machines change dynamically with production tasks (such as the difference in the temperature setting of the drying cylinder corresponding to different paper types), and fixed thresholds are prone to false alarms or missed alarms.

[0011] Ambiguous fault location: Most diagnostic models can only determine that a fault has occurred, but cannot accurately locate the faulty component (such as distinguishing between overheating of the drying cylinder and abnormality of adjacent rollers), resulting in low maintenance efficiency. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to provide a paper machine fault diagnosis method based on multi-channel grouping, which uses a multi-scale time-frequency dynamic attention long short-term memory autoencoder (MS-MTFA-LSTM-AE) model to detect multi-channel time-series data grouping of the paper machine, and combines dynamic thresholds to predict and alarm faults.

[0013] To address the aforementioned technical problems, this invention provides a paper machine fault diagnosis method based on multi-channel grouping, comprising the following steps: real-time acquisition of vibration, pressure / current, and temperature data of the paper machine, grouping them according to signal physical type, and then processing each group of data in parallel according to the following steps:

[0014] S1. Perform data preprocessing, including filtering, resampling, missing data completion and normalization, and construct the inference window;

[0015] S2. Input the window to be inferred into the offline-trained multi-scale time-frequency dynamic attention long short-term memory autoencoder model. First, extract attention-perception features through the multi-scale time-frequency dynamic attention module. Then the slow feature layer focuses on attention-based features. Mapping yields slow feature sequences The reconstructed feature sequence is then obtained through an intra-group long short-term memory autoencoder. Furthermore, for each time step t, the time series reconstruction error and window aggregation error are calculated;

[0016] S3. If the number of normal windows is less than the preset sliding window size W and the window aggregation error is greater than the initial dynamic threshold, an alarm is triggered; if the number of normal windows is greater than or equal to the preset sliding window size W and the window aggregation error is greater than the adaptive dynamic threshold, an alarm is triggered.

[0017] As an improvement to the paper machine fault diagnosis method based on multi-channel grouping of the present invention:

[0018] The preprocessing is as follows: vibration signals are filtered using a high-pass filter to remove the DC component, and then a band-pass filter is used to retain the effective frequency band; pressure and current signals are filtered using a low-pass filter; temperature signals are smoothed using a short-time moving average; then resampling is performed; when the buffer accumulates to a complete window and the proportion of effective samples is ≥95%, the inference window is formed with a step size S=T / 2.

[0019] As a further improvement to the paper machine fault diagnosis method based on multi-channel grouping of the present invention:

[0020] The resampling includes: during downsampling, anti-aliasing processing is first performed using a low-pass filter; during upsampling, linear interpolation is used to complete short-term missing data, interpolation is used to complete medium-term missing data, and windows with long-term continuous missing data are discarded.

[0021] As a further improvement to the paper machine fault diagnosis method based on multi-channel grouping of the present invention:

[0022] The multi-scale time-frequency dynamic attention module first extracts features at three different scales from the input sequence, then adds the features at the three different scales along the channels to obtain a fused feature F; then, it calculates the time-domain attention mapping for the fused feature F. and frequency domain attention mapping The fused feature F is combined with the temporal attention map. Weighted features are obtained by element-wise multiplication. The fused feature F is mapped to the frequency domain attention map. Weighted features are obtained by element-wise multiplication. Then the features With features The data is concatenated along the channel or feature dimension, fused using convolution, and outputs attention-aware features. .

[0023] As a further improvement to the paper machine fault diagnosis method based on multi-channel grouping of the present invention:

[0024] The time-series reconstruction error is:

[0025] (7)

[0026] The window aggregation error is:

[0027] (8)

[0028] Where T is the fixed window length.

[0029] As a further improvement to the paper machine fault diagnosis method based on multi-channel grouping of the present invention:

[0030] The multi-scale time-frequency dynamic attention long short-term memory autoencoder model is obtained through offline training via the following steps:

[0031] Historical operating data of the paper machine was collected, including multi-channel time-series data of vibration, pressure / current, and temperature. The data was divided into training, validation, and test sets according to time sequence, with the training set containing only samples from normal operating conditions. Each training, validation, and test set was then preprocessed. The training sets were then input into a multi-scale time-frequency dynamic attention long short-term memory autoencoder model to obtain reconstructed sequences, which were used to calculate the time-series reconstruction error and window aggregation error based on the training set, ultimately yielding a static threshold based on the training set. ,in and These are the mean and standard deviation of the window aggregation error based on the training set, respectively. The adjustment coefficient is used; the model performance is validated on the validation set after each round of training.

[0032] Then, the test set is input into the multi-scale time-frequency dynamic attention long short-term memory autoencoder model in groups to obtain the reconstructed sequence, which is used to calculate the temporal point reconstruction error and the window aggregation error based on the test set, and then to calculate the mean μ and standard deviation σ of the window aggregation error of normal samples in the test set.

[0033] As a further improvement to the paper machine fault diagnosis method based on multi-channel grouping of the present invention:

[0034] The initial dynamic threshold is:

[0035] (9)

[0036] The adaptive dynamic threshold is:

[0037] (10)

[0038] in, and Let W be the mean and standard deviation of the window aggregation error for the most recent W normal windows, respectively. This represents the upper limit of the threshold.

[0039] The beneficial effects of this invention are mainly reflected in:

[0040] This invention achieves high-precision diagnosis of papermaking machine faults through multi-channel group parallel processing, extraction of time-frequency attention features by a multi-scale time-frequency dynamic attention module, constraint of temporal smoothing features by a slow feature layer, intra-group LSTM-AE reconstruction detection, and dynamic threshold adaptive judgment. This method fully integrates the advantages of multi-scale learning, dynamic attention, and slow feature analysis, improving the model's ability to model complex multi-channel signals. Compared to traditional methods, it has higher detection accuracy and lower false alarm rate. Specifically:

[0041] 1. This invention employs a multi-channel grouping strategy based on physical properties, grouping heterogeneous data such as vibration, pressure / current, and temperature according to their physical mechanisms for parallel processing. This effectively avoids mutual interference between signals with different dimensions and noise distributions (such as electromagnetic interference and mechanical vibration) during unified modeling. By extracting features in parallel using multi-size convolutional kernels and adaptively weighting fault-sensitive frequency bands and key time slices using a time-frequency attention mechanism, the model's ability to extract key fault features and its robustness are significantly improved even under strong background noise.

[0042] 2. This invention addresses the characteristic that paper machine failures often manifest as long-term, gradual changes (such as wear and dirt buildup). This invention introduces a slow feature layer before the Long Short-Term Memory (LSTM) network. By minimizing the temporal difference squared loss, the model is forced to learn the essential features (slow features) that change slowly in the time dimension. This not only plays a noise reduction role similar to low-pass filtering, but also solves the memory decay problem of traditional LSTM when processing long sequences, significantly enhancing the model's ability to represent early, subtle failure trends.

[0043] 3. To address the problem of high false alarm rate of fixed threshold caused by dynamic changes in the operating conditions of papermaking machines with production tasks, this invention constructs an adaptive dynamic threshold mechanism, which reduces the false alarm rate caused by fluctuations in operating conditions while ensuring a high detection rate.

[0044] 4. By using independently grouped LSTM-AE reconstruction models, this invention can not only determine overall equipment anomalies, but also pinpoint specific faults based on the reconstruction error distribution of different groups (vibration group, temperature group, etc.). Simultaneously, the grouped parallel processing architecture effectively reduces the computational complexity of a single large model, meeting the timeliness requirements of real-time online monitoring in industrial settings. Attached Figure Description

[0045] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0046] Figure 1 This is a flowchart illustrating a paper machine fault diagnosis method based on multi-channel grouping according to the present invention.

[0047] Figure 2This is a schematic diagram of a paper machine fault diagnosis model based on multi-channel grouping according to the present invention.

[0048] Figure 3 This is a structural diagram of the MS-MTFA module. Detailed Implementation

[0049] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0050] Example 1: A paper machine fault diagnosis method based on multi-channel grouping, such as... Figure 1-3 As shown, the multi-channel data of the papermaking machine industrial process is preprocessed and then grouped into multiple channels according to vibration, pressure and current, and temperature groups. Subsequently, the time-series windows of each group are input into a multi-scale time-frequency dynamic attention module to extract time-frequency attention features. After passing through a slow feature layer, a gradually changing time-series representation is obtained. Then, the data is reconstructed by a long short-term memory autoencoder within the group, and the time-series point reconstruction error and window aggregation error are calculated. Finally, the fault determination is completed by combining a static threshold and an adaptive dynamic threshold based on normal window sliding statistics.

[0051] Step 1: Signal preprocessing and grouping.

[0052] In this embodiment, multi-channel sensors are deployed at key parts of the paper machine to collect vibration, pressure, current, and temperature data, and timestamps are unified to ensure time alignment across channels. The sampling frequency for vibration signals is 500–2000Hz, for pressure / current signals it is 50–500Hz, and for temperature signals it is 0.1–5Hz. The raw signals are preprocessed by filtering and resampling at edge nodes, specifically:

[0053] (1) The collected vibration signals are first filtered to remove DC, and then the bandpass is used to retain the frequency band of 5Hz–0.45×sampling rate; the collected pressure and current signals are processed by fourth-order low-pass filtering; the collected temperature signals are processed by short-time moving average smoothing with a window length of 10 sampling points.

[0054] (2) Resampling: Resampling is performed in groups, where a group refers to a collection of multiple channels and a channel refers to the signal stream acquired by a single sensor. The signals are categorized into vibration, pressure and current, and temperature groups based on their physical type. During downsampling, a low-pass filter is used for anti-aliasing; during upsampling, linear interpolation or spline interpolation is used. Short-term missing data (≤10 missing points) is supplemented using linear interpolation, medium-term missing data (10–100 missing points) is supplemented using cubic spline interpolation, and long-term continuous missing data (≥100 missing points) is considered an invalid window and discarded.

[0055] (3) Standardization

[0056] For each channel, z-score standardization is applied. The mean μ and standard deviation σ are initially obtained from the reconstruction error statistics of normal working condition samples in step 3.1 during offline training. In the online stage, as real-time data accumulates, the error of the latest normal window is periodically updated.

[0057] (4) Window division

[0058] The preferred window length for the vibration group is T=1024, and the step size is S=T / 2. Then, the three groups that have already been grouped are further subdivided. Each group is subdivided according to the correlation between the sensor installation area and the signal. Each subgroup has 2–8 channels, and then the signal is input into the MS-MTFA module.

[0059] Step 2: Construct the MS-MTFA-LSTM-AE model

[0060] The multi-scale time-frequency dynamic attention long short-term memory autoencoder (MS-MTFA-LSTM-AE) model includes a multi-scale time-frequency dynamic attention module, a slow feature layer, and an intra-group long short-term memory autoencoder connected in sequence. In this embodiment, there are three independent MS-MTFA-LSTM-AE models used for inference of data grouped by vibration, pressure / current, and temperature, respectively.

[0061] Step 2.1: Multi-Scale Multi-channel Temporal-Frequency Attention Module (MS-MTFA)

[0062] Vibration, pressure and current, and temperature data are input into the MS-MTFA module for feature extraction. The MS-MTFA module (comprising a multi-scale feature extraction module (MS) and a time-frequency attention module (MTFA)) has the following structure: Figure 3 As shown, MS represents Multi-Scale, corresponding to Figure 3 The module consists of two parts: multi-scale convolutional branch and multi-scale feature fusion. MTFA stands for Multi-channel Temporal-Frequency Attention, corresponding to... Figure 3 The architecture combines time-frequency attention computation with attention weighting and fusion. The multi-scale feature extraction (MS) module includes multiple convolutional branches with different receptive fields, using 3×3, 5×5, and 7×7 one-dimensional convolutional kernels to convolve the input sequence respectively, yielding three features F1, F2, and F3 at different scales. These three features are then summed along the channels to obtain the fused feature F. Next, the time-frequency attention module (MTFA) computes the temporal attention mapping A on the fused feature F.t and frequency domain attention mapping A f Specifically, attention can be calculated using the attention weighting formula:

[0063] (1)

[0064] in, and These represent the global average pooling and global max pooling operations on the fused feature F along a specified dimension, respectively. This represents vector concatenation. It is a mapping consisting of two convolutional layers and a nonlinear function. The first layer uses a 3×3 convolutional kernel, the second layer uses a 1×1 convolutional kernel, and the nonlinear function is ReLU. This is the Sigmoid activation function.

[0065] First, the fused feature F is subjected to global average pooling and global max pooling along the frequency direction according to formula (1). After concatenation, the temporal attention map A is obtained by two layers of convolution and Sigmoid. t Similarly, the fused feature F is pooled along the time direction according to formula (1) to obtain the frequency domain attention map A. f With A t and A f The weighted features are obtained by element-wise multiplication. and Then and The features are concatenated along the channel or feature dimension and fused using a 1×1 convolution to output attention-aware features. The MS-MTFA module incorporates the concepts of multi-scale learning and time-frequency joint attention, enabling the network to focus on significant time segments and frequency bands of fault signals.

[0066] Step 2.2: Slow feature layer.

[0067] MS-MTFA output characteristics The input is fed into the slow feature layer. The slow feature layer guides feature learning by introducing Slow Feature Analysis (SFA) constraints, ensuring that the output features change as smoothly as possible over time. Let the output slow feature sequence be... This layer calculates the squared difference loss between adjacent time steps:

[0068] (2)

[0069] Here, T represents the total time length of the slow feature sequence, i.e., the window length, and t is the time step index, traversing each adjacent time step from 1 to T-1. Simultaneously, zero mean and unit variance constraints are imposed on each feature channel to ensure a reasonable distribution of the output slow features, thereby avoiding the generation of trivial solutions. This design makes the network more inclined to learn the features that change most slowly in the time domain, enhancing its ability to express fault indicators (typically long-term trends or low-frequency components).

[0070] Step 2.3: Intragroup Long Short-Term Memory Auto-Encoder (LSTM-AE) Reconstruction.

[0071] The slow feature sequence output by the slow feature layer As input to the LSTM-AE, the LSTM-AE consists of an encoder and a decoder: the encoder uses a two-layer LSTM unit to map the input sequence to a low-dimensional temporal hidden representation; the decoder uses the hidden representation as the starting state and gradually reconstructs the output sequence, outputting the reconstructed sequence. .

[0072] The network is trained using input sequences acquired under normal operating conditions, learning normal signal patterns by minimizing the reconstruction error. Reconstruction error Defined in the form of mean squared error (MSE):

[0073] (3)

[0074] in, and These are the i-th sampling points of the input sequence and the reconstructed sequence, respectively. After training, LSTM-AE exhibits low reconstruction error for normal data, but for anomalous data, the reconstruction error increases significantly due to pattern shift.

[0075] Step 3: Offline Training and Testing

[0076] Step 3.1: Constructing Training Data

[0077] Multi-channel time-series data of vibration, pressure / current, and temperature are collected from historical data of the paper machine or from a simulation platform. The data is then divided into vibration, pressure / current, and temperature groups according to time sequence. Each group is further divided into training, validation, and test sets, i.e., three different training, validation, and test sets for vibration, pressure / current, and temperature, respectively. The recommended ratio of training, validation, and test sets is 70%:10%:20%. The training set contains only samples under normal operating conditions and is used for self-supervised reconstruction learning. After grouping and preprocessing each sample as in step 1, sample windows are obtained, segmented with a fixed window length T=1024 and a step size S=T / 2. Each sample window stores its timestamp and group; training windows are labeled "Normal," and test windows are labeled "Normal / Fault" for evaluation.

[0078] Step 3.2: Loss Function

[0079] An end-to-end training process is employed using a joint loss function for MS-MTFA, the slow feature layer, and the in-group LSTM-AE. A common form is:

[0080] (4)

[0081] Where: regularization term Includes weight decay and gated sparsity regularization. , All are hyperparameters, reconstruction loss for:

[0082] (5)

[0083] in, This represents the input data at time step t in the m-th training window. This represents the decoded and reconstructed value at time step t in the m-th training window. M is the number of training sample windows, and T is the time length of each window.

[0084] Slow feature loss for:

[0085] (6)

[0086] Step 3.3: Offline training and testing process

[0087] The offline training process involves independently training and testing the MS-MTFA-LSTM-AE model offline using the vibration, pressure / current, and temperature training, validation, and test sets constructed in step 3.1, resulting in three independent models that can be used online.

[0088] I. Offline Training Process

[0089] The grouped training sets of vibration, pressure / current, and temperature constructed in step 3.1 are input into the MS-MTFA module to extract attention-aware features. Then use a slow feature layer to... Perform mapping to output slow feature sequences. Then use in-group LSTM-AE to process slow feature sequences. Perform reconstruction and output the reconstruction sequence. For each time step t, the formula is used... The reconstruction error of time series points is calculated by averaging the reconstruction errors of all time series points within a time window (length T=1024) using the formula. The windowed aggregated error is obtained and used to statistically analyze the error distribution under normal operating conditions. Calculate the windowed aggregated error. mean and standard deviation Through formula Obtain static thresholds based on the training set , The adjustment coefficient is set to 3 in this embodiment. This invention sets an insurmountable upper limit for the dynamic threshold based on the maximum error of normal samples statistically analyzed during the offline phase. The calculation method is as follows Compared to the static threshold A deviation one standard deviation higher allows for drift under normal operating conditions while preventing the infinite adaptation of fault errors. Then, a joint loss function is used. The model was trained iteratively for 50 epochs using the Adam optimizer (initial learning rate 0.001, weight decay 0.0001). After each epoch, the model performance was validated on the validation set, and the optimal weights were saved. Finally, the optimal weights for each MS-MTFA-LSTM-AE sub-model were output, and a static threshold was determined. This is used for threshold initialization in subsequent online inference.

[0090] II. Testing Process

[0091] Input the grouped test set (with the same distribution and preprocessing as the training set) and the optimal model weights for each group output during the training process.

[0092] Load the optimal model weights for the corresponding group, and perform MS-MTFA feature extraction → slow feature mapping → LSTM-AE reconstruction on the test set samples to obtain the reconstruction sequence for each group of samples. The error calculation logic of the repeated training process is used to obtain the temporal point reconstruction error and window aggregation error of the test set.

[0093] For normal samples in the test set, statistically analyze the distribution of their window aggregation errors to verify whether the vast majority of window aggregation errors are less than or equal to... ; Calculate the window aggregation error for fault samples in the test set and verify whether it is greater than 1. .

[0094] After the model is trained offline, the window aggregation error of normal samples in the test set is statistically analyzed, and its mean μ and standard deviation σ are calculated. Then, μ and σ are compared with the static threshold. upper limit of threshold The dynamic threshold initialization is stored in the model library for online inference. During training and testing, the optimal weights for each dedicated MS-MTFA-LSTM-AE sub-model are output; static thresholds are also included. The mean μ and standard deviation σ of the temporal reconstruction error of normal samples in the test set. Model weights are used for model loading during online inference; static thresholds... With μ, σ, and upper threshold Used for calculating the initial dynamic threshold.

[0095] Step 4: Use online

[0096] Step 4.1: Real-time data acquisition

[0097] Multi-channel sensors are deployed at key parts of the paper machine, and vibration, pressure and current, and temperature data are continuously collected in strict accordance with the sampling rates determined in step 1 (500–2000Hz for vibration, 50–500Hz for pressure and current, and 0.1–5Hz for temperature). After collection, the data are divided into three groups according to the physical type of the signal (vibration group, pressure and current group, and temperature group) to ensure that the timestamps of each channel are synchronized.

[0098] Step 4.2: Preprocessing and Window Construction

[0099] Each raw sample entering the buffer undergoes preprocessing: the vibration group uses a high-pass filter (cutoff 0.1Hz) to remove DC components and a band-pass filter (5Hz – 0.45 × sampling rate, stopband attenuation ≥40dB) to retain effective frequency bands; the pressure and current group uses a fourth-order low-pass filter; and the temperature group uses a short-time moving average smoothing with a window length of 10 sampling points. Resampling is then performed, with upsampling using linear interpolation, missing data completion, and z-score normalization. When the buffer accumulates to a complete window and the proportion of effective samples is ≥95%, an inference window is formed with a step size S = T / 2.

[0100] Step 4.3: Load the offline model

[0101] During system initialization or model update, the pre-trained model weights (including MS-MTFA module, slow feature layer, and in-group LSTM-AE), static thresholds, and initial values ​​μ and σ obtained from step 3.3 are loaded from the model library.

[0102] Step 4.4: Model Inference (Parallel for Each Group)

[0103] For the inference window in group g, perform the following operations: The front-end MS-MTFA module extracts and outputs attention-aware features. Slow feature layers for attention-based feature perception By mapping, a slow feature sequence that changes gradually over time is obtained. After reconstruction by the long short-term memory autoencoder within the group, the reconstructed sequence is obtained. This is used to calculate the time-series reconstruction error sequence:

[0104] (7)

[0105] Then, calculate the window aggregation error:

[0106] (8)

[0107] The fixed window length is T=1024.

[0108] The time-series reconstruction error sequence is the slow feature sequence output by the slow feature layer at a single time step. Reconstructing feature sequences with LSTM-AE The difference measure reflects the model's deviation in reconstructing features at that time point; the window aggregation error is the average of the errors of all time points within a window containing T time steps. It is an aggregation of local errors in the time dimension and reflects the overall degree of reconstruction anomaly within the window.

[0109] Step 4.5: Reconstruction Error Determination and Dynamic Threshold

[0110] This step is the core of online inference for fault determination. It combines the initial dynamic threshold from offline statistics with the online dynamic threshold update mechanism to construct an adaptive threshold to cope with signal drift and accurately determine faults, as detailed below:

[0111] (1) Initial dynamic threshold setting (data accumulation) <W)

[0112] After the model is trained offline, an initial dynamic threshold is calculated based on the mean μ and standard deviation σ of the window aggregation error of normal samples in the test set. This initial dynamic threshold is used as the default decision threshold for the initial stage of the system. The formula for the initial dynamic threshold is:

[0113] (9)

[0114] In this embodiment, the adjustment coefficient k=3. This is the static threshold obtained in step 3.3 during the offline training phase.

[0115] The system maintains a count and buffer of normal windows. Before the number of recently active normal windows accumulated in the buffer reaches the set sliding window size W, the initial dynamic threshold is used directly because the online statistics are not yet stable. For fault diagnosis, W is typically set between 10 and 20; here, we choose 15. The current window aggregation error is greater than the initial dynamic threshold. If the condition is abnormal, a fault alarm will be triggered. For long-cycle gradual fault scenarios, W is reduced to 10 to shorten the statistical period of the sliding window, reduce the accumulation of pseudo-normal windows, and thus... and It is more sensitive to increases in error and avoids over-adaptation of the threshold.

[0116] (2) Adaptation period (data accumulation ≥ W)

[0117] After accumulating W normal windows, the system begins to update the threshold in real time based on the current operating conditions to adapt to signal drift.

[0118] Adaptive dynamic threshold calculation:

[0119] (10)

[0120] in, , This is the mean and standard deviation of the window aggregation error based on the most recent W normal windows. As the upper limit of the threshold, k=3 in this embodiment.

[0121] If the current window aggregation error is greater than the dynamic threshold If this occurs, the fault alarm for that group will be triggered.

[0122] (3) Closed-loop update:

[0123] If the result is deemed normal, the window aggregation error of the current window will be included in the historical statistics and updated periodically. and This ensures that the threshold continuously follows environmental changes.

[0124] Step 4.6: Output and Recording

[0125] The system outputs fault group / specific component, alarm time, error value, attention weight snapshot, and historical record path.

[0126] experiment:

[0127] 1. To verify the effectiveness of the proposed method, simulation experiments were conducted using simulated paper machine fault data. The experimental setup was as follows: Three common fault modes were simulated (e.g., bearing wear, uneven coating, and abnormal tension), corresponding to changes in the characteristics of multi-channel sensor signals. Time-series data were collected under normal and fault conditions, with a fixed signal length of T=1024 and a sampling rate of 1kHz. The MS-MTFA-LSTM-AE model was implemented using PyTorch, with multi-scale convolutional kernel sizes of 3, 5, and 7 for each signal group. The LSTM encoder and decoder each had two layers, and the hidden unit dimension was set to 50. The learning rate during training was... Each channel group is trained independently. The dynamic threshold coefficient k is set to 3.

[0128] 2. Ablation test

[0129] Online inference was performed on all windows of the test set, the time-series reconstruction error was calculated, and the results of the ablation experiment were evaluated according to the following metrics. The results are shown in Table 1.

[0130] Accuracy: TP (True Positive) represents the number of windows correctly identified as "faulty"; TN (True Negative) represents the number of windows correctly identified as "healthy"; FP (False Positive) represents the number of healthy windows that were mistakenly identified as "faulty"; and FN (False Negative) represents the number of windows that were missed being identified as "faulty".

[0131] Precision: This reflects the proportion of truly faulty samples among all those judged as "faulty".

[0132] Recall: Also known as the detection rate, it reflects the proportion of all real fault samples that are correctly detected.

[0133] F1 score: The harmonic mean of precision and recall is used to comprehensively evaluate model performance.

[0134] Average detection delay (Delay): Typically measured in seconds, it represents the average speed at which the model issues a warning after a fault occurs. For M fault events labeled in the test set, let the actual occurrence time of the i-th fault be... The model first detected the fault at the time when The average detection delay is then defined as:

[0135] (11)

[0136] Table 1 Results of the ablation experiment

[0137] Model Accuracy (%) Accuracy (%) Recall rate (%) F1 score Average detection delay (s) LSTM-AE 90.2 88.5 89.0 0.887 4.8 MS-AE 93.1 91.8 92.5 0.921 3.9 MS-LSTM-AE 94.5 93.2 94.0 0.936 3.0 MTFA-LSTM-AE 95.8 94.7 95.0 0.948 2.5 MS-MTFA-LSTM-AE 97.4 96.8 98.1 0.974 2.1

[0138] Among them, LSTM-AE is the basic model, serving as a "benchmark" for subsequent complex models. It consists of a "Long Short-Term Memory Network (LSTM) and an autoencoder (AE)". LSTM is used to capture temporal dependencies, while AE reconstructs the temporal signal and detects faults through reconstruction errors.

[0139] MS-AE: A model structure including a multi-scale feature extraction module (Multi-Scale, MS) and an autoencoder (AE). It extracts features from time-series signals through multiple sets of operations at different scales (such as convolutional kernels and pooling layers of different sizes). The autoencoder reconstructs the features and calculates the error between the reconstructed signal and the original signal. Faults are determined by whether the error exceeds a threshold. Compared with LSTM-AE, it enhances multi-scale feature extraction.

[0140] MS-LSTM-AE: Includes a multi-scale feature extraction module (MS), LSTM and AE, combining multi-scale feature extraction with LSTM temporal modeling.

[0141] MTFA-LSTM-AE: Includes the Time-Frequency Attention Module (MTFA), LSTM, and AE. By introducing the MTFA module, the relationship with time-frequency domain features (such as abrupt changes and periodicity) is strengthened.

[0142] The results show that introducing the time-frequency attention module and the slow feature layer significantly improves the model's accuracy and F1 score, while removing either module leads to a performance decrease, demonstrating the effectiveness of these modules. Furthermore, the dynamic threshold strategy drastically reduces the false alarm rate and correspondingly reduces detection latency.

[0143] 3. Comparative Experiment

[0144] The same publicly available multi-channel industrial monitoring dataset (6 sensors, 100Hz, with fault annotations) was used as in the ablation experiments. For each model, the 99th percentile of the error distribution reconstructed from normal samples was used as the anomaly detection threshold during the offline phase. The following table shows the experimental results compared to mainstream networks, using five evaluation metrics: accuracy, precision, recall, F1 score, and average detection delay. The results are shown in Table 2.

[0145] Table 2 Results of the comparative experiment

[0146] Model Accuracy (%) Accuracy (%) Recall rate (%) F1 score Average detection delay (s) PCA+3σ 71.2 69.8 72.5 0.711 8.2 One-ClassSVM 78.4 80.2 76.9 0.785 7.6 GCN-LSTM-AE 89.1 91.0 87.4 0.891 4.5 Transformer-AE 93.2 94.1 92.3 0.932 3.3 This invention 97.4 96.8 98.1 0.974 2.1

[0147] in, Traditional fault detection methods first use Principal Component Analysis (PCA) to reduce dimensionality, and then use the 3σ criterion to detect anomalies.

[0148] One-ClassSVM: A single-class support vector machine, an unsupervised anomaly detection method.

[0149] GCN-LSTM-AE: Includes Graph Convolutional Network (GCN), LSTM, and AE.

[0150] Transformer-AE: Includes Transformer and autoencoder.

[0151] In summary, the experimental results fully demonstrate the effectiveness of the MS-MTFA module and the slow feature layer and dynamic thresholding strategy of this invention. They outperform existing mainstream methods in all key performance indicators, verifying the innovation and practical value of this invention, which can be used to solve:

[0152] 1. The problem of effective modeling of complex multi-channel signals

[0153] Traditional methods often simply stack or average data from multi-channel sensors in papermaking machines, making it difficult to fully utilize the correlations and differences between different channels.

[0154] 2. The problem of effective extraction of time-frequency features and slowly varying features

[0155] Existing Convolutional Block Attention Modules (CBAMs) only perform attention weighting in a single scale and single domain (temporal / spatial), making it difficult to simultaneously capture multi-scale abrupt changes and frequency-domain periodic features, and also lacking modeling for long-term stable trends (slow features).

[0156] 3. High false alarm rate and detection delay issues in fault detection

[0157] Fixed threshold methods are prone to false alarms or missed alarms when operating conditions change or sensors drift, and there is often a certain time lag in fault detection, which affects the effectiveness of practical applications.

[0158] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A paper machine fault diagnosis method based on multi-channel grouping, characterized in that... The process includes: real-time acquisition of vibration, pressure / current, and temperature data from the paper machine, grouping the data according to signal physical type, and then processing each group of data in parallel using the following steps: S1. Perform data preprocessing, including filtering, resampling, missing data completion and normalization, and construct the inference window; S2. Input the window to be inferred into the offline-trained multi-scale time-frequency dynamic attention long short-term memory autoencoder model. First, extract attention-perception features through the multi-scale time-frequency dynamic attention module. Then the slow feature layer focuses on attention-based features. Mapping yields slow feature sequences The reconstructed feature sequence is then obtained through an intra-group long short-term memory autoencoder. Furthermore, for each time step t, the time series reconstruction error and window aggregation error are calculated; S3. If the number of normal windows is less than the preset sliding window size W and the window aggregation error is greater than the initial dynamic threshold, an alarm is triggered; if the number of normal windows is greater than or equal to the preset sliding window size W and the window aggregation error is greater than the adaptive dynamic threshold, an alarm is triggered.

2. The paper machine fault diagnosis method based on multi-channel grouping according to claim 1, characterized in that: The preprocessing is as follows: vibration signals are filtered using a high-pass filter to remove the DC component, and then a band-pass filter is used to retain the effective frequency band; pressure and current signals are filtered using a low-pass filter; temperature signals are smoothed using a short-time moving average; then resampling is performed; when the buffer accumulates to a complete window and the proportion of effective samples is ≥95%, the inference window is formed with a step size S=T / 2.

3. The paper machine fault diagnosis method based on multi-channel grouping according to claim 2, characterized in that: The resampling includes: during downsampling, anti-aliasing processing is first performed using a low-pass filter; during upsampling, linear interpolation is used to complete short-term missing data, interpolation is used to complete medium-term missing data, and windows with long-term continuous missing data are discarded.

4. A paper machine fault diagnosis method based on multi-channel grouping according to any one of claims 1 to 3, characterized in that: The multi-scale time-frequency dynamic attention module first extracts features at three different scales from the input sequence, then adds the features at the three different scales along the channels to obtain a fused feature F; then, it calculates the time-domain attention mapping for the fused feature F. and frequency domain attention mapping The fused feature F is combined with the temporal attention map. Weighted features are obtained by element-wise multiplication. The fused feature F is mapped to the frequency domain attention map. Weighted features are obtained by element-wise multiplication. Then the features With features The data is concatenated along the channel or feature dimension, fused using convolution, and outputs attention-aware features. .

5. The paper machine fault diagnosis method based on multi-channel grouping according to claim 4, characterized in that: The time-series reconstruction error is: (7) The window aggregation error is: (8) Where T is the fixed window length.

6. The paper machine fault diagnosis method based on multi-channel grouping according to claim 5, characterized in that: The multi-scale time-frequency dynamic attention long short-term memory autoencoder model is obtained through offline training via the following steps: Historical operating data of the paper machine was collected, including multi-channel time-series data of vibration, pressure / current, and temperature. The data was divided into training, validation, and test sets according to time sequence, with the training set containing only samples from normal operating conditions. Each training, validation, and test set was then preprocessed. The training sets were then input into a multi-scale time-frequency dynamic attention long short-term memory autoencoder model to obtain reconstructed sequences, which were used to calculate the time-series reconstruction error and window aggregation error based on the training set, ultimately yielding a static threshold based on the training set. ,in and These are the mean and standard deviation of the window aggregation error based on the training set, respectively. The adjustment coefficient is used; the model performance is validated on the validation set after each round of training. Then, the test set is input into the multi-scale time-frequency dynamic attention long short-term memory autoencoder model in groups to obtain the reconstructed sequence, which is used to calculate the temporal point reconstruction error and the window aggregation error based on the test set, and then to calculate the mean μ and standard deviation σ of the window aggregation error of normal samples in the test set.

7. A paper machine fault diagnosis method based on multi-channel grouping according to claim 6, characterized in that: The initial dynamic threshold is: (9) The adaptive dynamic threshold is: (10) in, and Let W be the mean and standard deviation of the window aggregation error for the most recent W normal windows, respectively. Upper limit of threshold: .

Images