A method for detecting abnormalities in belt conveyor rollers based on DAS and VAE

Abstract

This invention proposes a method for detecting abnormalities in belt conveyor rollers based on DAS and VAE. The method first performs time-slicing, denoising, and normalization processing on the raw roller vibration data acquired by the Φ-OTDR or DAS system, and extracts MFCC features. With limited labeled data, self-supervised learning is used to model the vibration modes under normal operating conditions. This invention trains the VAE model using only normal samples, enabling it to learn the potential distribution characteristics of normal vibration signals. In the detection phase, the reconstruction error of the test samples is standardized pixel-by-pixel with the pixel-level error baseline constructed from normal samples to generate an anomaly rate map. This map is then combined with threshold filtering and distance dimension aggregation to achieve reliable identification of abnormal events. This invention can achieve rapid and accurate detection of roller faults under conditions of noise interference and data scarcity, demonstrating strong robustness and engineering application potential.

Description

Technical Field

[0001] This invention belongs to the fields of distributed optical fiber sensing technology and artificial intelligence applications, and in particular, it is a method for detecting anomalies in belt conveyor rollers. Background Technology

[0002] Belt conveyors are widely used material handling equipment in industry. Their rollers, as key components, present numerous challenges to traditional fault detection methods due to their large number and wide distribution. Vibration and temperature monitoring technologies, in particular, are often limited by installation complexity, sensor coverage, and environmental interference, making comprehensive and effective monitoring difficult. Therefore, combining Distributed Fiber Acoustic Sensing (DAS) technology with Artificial Intelligence (AI) algorithms has become an effective solution. DAS technology, through the deployment of fiber optic sensors, achieves full-process, contactless monitoring of conveyor belt rollers, offering advantages such as high sensitivity and long-distance monitoring, capable of capturing minute vibrations and acoustic signals from the rollers in real time. Especially in early fault detection, acoustic signals can provide earlier warnings, overcoming the limitations of traditional methods that often fail to detect minor faults in a timely manner. Combined with AI algorithms such as deep learning, spatial and temporal features can be automatically extracted from massive amounts of acoustic data, achieving more accurate fault diagnosis. AI models efficiently process complex time-series data, identifying potential fault modes and overcoming the inefficiency and human error problems of manual data processing. Fault detection systems based on DAS technology and AI algorithms not only improve detection accuracy and efficiency, but also effectively reduce maintenance costs and downtime, providing a safer and more reliable monitoring method for industrial production.

[0003] In the field of drum fault detection, the combination of DAS technology and AI algorithms has played a crucial role. Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) are commonly used to extract spatial and temporal features from vibration signals, but they require a large amount of labeled data, facing the challenge of data scarcity in practical applications. The Temporally Distributed Convolutional Network (TD-CLNet) combines CNN and LSTM (Reference 1: Alharbi F, Luo S, Yang G, TD-CLNet: a time-distributed CNN-LSTM network for fault detection in belt conveyor idlers[J], Neural Computing and Applications, 2025, 37(30): 25151-25181; Alharbi F, Luo S, Yang G. TD-CLNet: A Temporally Distributed CNN-LSTM Network for Fault Detection in Belt Conveyor Rollers[J], Neural Computing and Applications, 2025, 37(30): 25151-25181). It can simultaneously process spatial features and temporal series information. Although it performs well in complex data processing, it is still constrained by large data requirements and high computational resource consumption. While liquid neural networks (LNNs) have been applied in some industrial fault detection, they are highly dependent on data preprocessing and model optimization, and require extensive experimental verification (Reference 2: Liu H, Xu Y, Qi Y, Yang H, Bi W. Rapid diagnosis of distributed acoustic sensing vibration signals using mel-frequency cepstral coefficients and liquid neural networks[J]. Sensors, 2025, 25(10): 3090; LiuH, Xu Y, Qi Y, Yang H, Bi W. Rapid diagnosis of distributed acoustic sensing vibration signals based on Mel-frequency cepstral coefficients (MFCC) and liquid neural networks[J], Sensors, 2025, 25(10): 3090).Therefore, existing distributed fiber optic vibration monitoring technology still faces several key challenges in drum anomaly detection, including high dependence on large-scale labeled anomaly samples, limited ability to identify unknown anomaly patterns, susceptibility of detection results to interference from error fluctuations in different frequency bands, and insufficient fault location accuracy. There is an urgent need to integrate artificial intelligence algorithms with traditional signal processing methods to build a more efficient and robust anomaly detection and location technology system. Summary of the Invention

[0004] To overcome the problems of existing DAS and AI-based roller anomaly detection technologies, such as difficulties in data labeling, imbalanced category distribution, and insufficient fault identification performance in complex noise environments, this invention proposes a belt conveyor roller anomaly detection method based on DAS and VAE (Variational Autoencoder). This method first performs time-slicing, denoising, and normalization on the raw roller vibration data acquired by the Φ-OTDR or DAS system, and extracts MFCC features. With limited labeled data, self-supervised learning (SSL) is used to model the vibration patterns under normal operating conditions, effectively mitigating the adverse effects of data imbalance. Specifically, this invention trains the VAE model using only normal samples, enabling it to learn the potential distribution characteristics of normal vibration signals, reducing dependence on large-scale labeled data. In the detection phase, the reconstruction error of the test samples is standardized pixel-by-pixel with the pixel-level error baseline constructed from normal samples to generate an anomaly rate map. This map is then combined with threshold filtering and distance dimension aggregation to achieve reliable identification of abnormal events. This invention enables rapid and accurate detection of roller faults under conditions of noise interference and data scarcity, demonstrating strong robustness and engineering application potential. Overall, compared to traditional artificial intelligence methods, the method proposed in this invention combines high detection accuracy, low data requirements, and good environmental adaptability, providing an efficient and robust technical solution for roller fault detection.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a method for detecting abnormalities in belt conveyor rollers based on DAS and VAE, characterized by comprising the following steps: Step 1: Acquisition of raw drum vibration signals: Fix the armored optical cable to the drum support of the belt conveyor, use a DAS system based on Φ-OTDR to acquire the raw vibration data of the belt conveyor drum, and perform time slicing of the raw vibration data according to a preset time window to obtain time-domain vibration data blocks obtained along a specific distance of the optical cable.

[0006] Step 2, Signal Preprocessing: The time-domain drum vibration data block is preprocessed, including wavelet denoising and normalization, to suppress noise interference and improve the consistency of subsequent feature extraction.

[0007] Step 3, MFCC feature extraction: The preprocessed time-domain original drum vibration signal is converted into Mel frequency cepstral coefficients (MFCC) features. Specifically, the vibration signal is subjected to pre-emphasis, framing, Hamming windowing, Fast Fourier Transform (FFT) / power spectrum calculation, Mel filter bank processing, logarithmic compression, and Discrete Cosine Transform (DCT) decorrelation and dimensionality reduction to obtain the MFCC feature matrix representing the vibration mode.

[0008] Step 4: Construct and train the VAE model: Construct a VAE and train the variational autoencoder VAE using only the MFCC features of normal samples, so that the VAE learns the latent representation of normal vibration modes and their reconstruction rules.

[0009] Step 5: Establish a normal error baseline: Input normal samples into the trained VAE model to obtain the corresponding reconstructed MFCC features, and calculate the reconstruction error map of the normal samples. Based on the reconstruction error maps of all normal samples, calculate the mean and standard deviation of the error for each pixel location to form a normal error baseline, where the first... The average value of the pixel positions is denoted as The standard deviation is denoted as .

[0010] Step 6: Calculate the reconstruction error of the test sample: Input the MFCC features of the test sample into the trained VAE model to obtain the reconstructed MFCC features corresponding to the test sample. Calculate the reconstruction error map based on the difference between the input MFCC features and the reconstructed MFCC features, and denote the error as... The error at each pixel position is The larger the error, the more the location deviates from the normal vibration mode.

[0011] Step 7, Pixel-by-Pixel Standardization: Using the normal error baseline established in Step 5, the reconstruction error map of the test samples is standardized pixel-by-pixel to obtain the standardized error map. The calculation formula is as follows: , in, Indicates the test sample at the 1st The degree of deviation of each pixel position from the normal baseline.

[0012] Step 8: Generate an anomaly rate map: Perform pixel-by-pixel square processing on the standardized error map to generate the anomaly rate map. The calculation formula is: , in, Indicates the first Abnormal intensity at each pixel location.

[0013] Step 9, Pixel-level threshold filtering: Based on the percentile threshold of the chi-square distribution. The anomaly rate map is filtered, retaining only significantly anomalous pixels; for those below a threshold... For pixels, set their anomaly rate to 0 to remove insignificant abnormal responses.

[0014] Step 10: Aggregate by time location: For each time location, select several of the largest anomaly rate values ​​from all frequency bands and average them to obtain the aggregated anomaly rate for that distance location; repeat the above process for all time locations to form an aggregated anomaly rate curve.

[0015] Step 11, Anomaly Detection: Compare the aggregated anomaly rate corresponding to each time point with the preset detection threshold. Comparison; when the aggregation anomaly rate at a certain distance location exceeds a threshold When an anomaly is detected at that distance, the location is determined to be abnormal; the suspected fault location is determined based on the number of sensor channels in the corresponding aggregated anomaly rate curve.

[0016] Furthermore, the time slice length in step 1 is set according to the actual monitoring scenario to ensure that a single data block can cover the complete characteristics of the target vibration event.

[0017] Furthermore, the normalization method in step 2 is either min-max normalization or Z-score normalization.

[0018] Furthermore, step 3, MFCC feature extraction, includes the following steps: Step 3-1: Pre-weighting The input drum vibration signal is pre-emphasized to enhance the high-frequency components, resulting in a pre-emphasized signal: , The preprocessed time-domain original drum vibration signal is: , This is the pre-weighting factor, typically taken as 0.95-0.98; Step 3-2: Frame Segmentation The pre-emphasized signal is divided into frames according to the set frame length and frame shift to obtain the first frame. m Frame signal: , in, N For frame length, H For frame shift, m The frame number; Step 3-3: Adding windows to Hamming Each frame of signal is multiplied by a Hamming window to reduce spectral leakage, resulting in the windowed frame signal: , Among them, the Hamming window function for: ; Steps 3-4: Fast Fourier Transform and Power Spectrum Calculation Perform a Fast Fourier Transform on each frame of the windowed signal to obtain its frequency domain representation: , Further calculation of the power spectrum: ,in, k Frequency index; Steps 3-5: Mel filter bank processing Power spectrum L The output energy of each filter is obtained by filtering with a group of Mel filters: , in, Indicates the first l The frequency response of a Mel filter. The conversion relationship between Mel frequency and actual frequency can be expressed as: ; Steps 3-6: Logarithmic compression Logarithmic operations are performed on the output energies of each Mel filter to compress the dynamic range, yielding the logarithmic Mel energy: ; Steps 3-7: Discrete Cosine Transform (DCT) Decorrelation and Dimensionality Reduction Performing a discrete cosine transform on the logarithmic Mel energy yields the th c MFCC coefficients: , Where C is the number of MFCC coefficients retained; Steps 3-8: Construct the MFCC feature matrix The MFCC coefficients obtained from each frame are arranged in chronological order to form the MFCC feature matrix: , in, M M represents the total number of frames, and M is the MFCC characteristic matrix representing the drum vibration mode.

[0019] Furthermore, step 4, constructing the VAE and training the VAE model, includes the following steps: Step 4-1: Construct a normal sample training set.

[0020] The DAS vibration signal of the drum under normal operating conditions was collected, and the corresponding MFCC features were extracted to form a normal sample training set. , in, Indicates the first n MFCC feature matrix corresponding to each normal sample N This represents the total number of normal samples; For any input sample, its MFCC feature is denoted as: ; Step 4-2: Construct the variational autoencoder (VAE).

[0021] A variational autoencoder (VAE) is constructed, comprising an encoder and a decoder. The encoder maps input features to latent variable distribution parameters, and the decoder reconstructs the input features based on the latent variables. Let the latent variables be... z The encoder outputs the mean and variance parameters of the latent variable distribution and expresses the approximate posterior normal distribution as: in The vector representing the mean of the latent variable distribution. Let cc represent the variance vector of the latent variable distribution, and let cc represent the encoder parameters. The prior distribution of the latent variables is also set as follows: ; Step 4-3: Reparameterize the latent variables.

[0022] To make the latent variable sampling process differentiable, a reparameterization method is used to generate latent variables. z： in, This indicates element-wise multiplication.

[0023] Step 4-4: Reconstruct input features using the decoder The latent variables obtained in step 4-3 z The input is processed by the decoder to obtain the reconstructed features of the input: in, This represents the reconstructed MFCC features. This represents the decoder mapping function. Indicates decoder parameters; Steps 4-5: Construct the loss function and train the VAE.

[0024] The VAE is trained using the reconstruction error between the input features and the reconstructed features, and the KL divergence between the latent variable distribution and the prior distribution, as the joint loss function. The loss function is expressed as: in, Indicates the reconstruction error. Denotes KL divergence, This represents the weight coefficients. The encoder parameters are optimized through iterative training on all normal samples. and decoder parameters Its optimization objective is expressed as: ; Steps 4-6: Output the trained VAE model.

[0025] After iterative training in steps 4-5, the trained VAE model is obtained: in, and These represent the encoder and decoder parameters after training, respectively. The trained VAE model is used to learn the latent distribution characteristics and reconstruction rules of MFCC features under normal drum vibration conditions, so as to subsequently use the difference between the input features and the reconstructed features to characterize the degree to which the test sample deviates from the normal pattern.

[0026] Furthermore, the reconstruction error in step 5 is the absolute error, squared error, or other error form that can characterize the reconstruction difference between the input MFCC feature and the corresponding pixel position of the reconstructed MFCC feature.

[0027] Furthermore, the screening threshold in step 9 Determined based on the percentiles of the chi-square distribution to preserve significantly anomalous pixels.

[0028] Furthermore, the aggregation operation in step 10 can employ maximum value averaging, top K maximum value averaging, weighted averaging, or robust statistics to reduce the impact of occasional errors in a single frequency band on the detection results.

[0029] Compared with existing technologies, this invention addresses the application scenario of drum fault detection where abnormal samples are scarce, abnormal types are complex, and prior knowledge is difficult to obtain. It employs a method of training a VAE model based solely on normal samples, eliminating the need for a large amount of labeled drum fault abnormal vibration data, thus exhibiting stronger practicality and generalization potential. Simultaneously, by constructing a pixel-level error baseline for normal drum vibration samples and performing pixel-by-pixel standardization on the reconstruction error of test samples, this invention effectively weakens the interference caused by the uneven distribution of natural errors at different frequency bands and cepstral coefficient positions, thereby improving the consistency of abnormal characterization and the stability of detection results. Furthermore, this invention generates an anomaly rate map and performs pixel-level threshold screening, which can highlight significant abnormal responses, suppress false activations caused by weak noise disturbances, and reduce the risk of misjudgment. Finally, by aggregating the anomaly rate along the distance dimension and constructing an aggregated anomaly rate curve, it can not only accurately determine abnormal states but also indicate the location of suspected drum vibration faults, thus possessing both anomaly detection and spatial localization capabilities. Furthermore, this invention converts time-domain vibration signals into MFCC features, enabling vibration modes to be expressed more compactly and effectively in the cepstral domain, which is beneficial for enhancing the ability to characterize complex vibration differences and thus improving the overall anomaly identification performance. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of a distributed fiber optic sensing system. Figure 3 This is a schematic diagram of an experimental scenario for abnormal detection of conveyor rollers based on a distributed fiber optic vibration sensing system. Figure 4 The time-domain waveform of the knocking event and its corresponding MFCC characteristics; Figure 5 The time-domain waveform of the friction event and its corresponding MFCC characteristics; Figure 6 The time-domain waveform of a normal operating event and its corresponding MFCC characteristics; Figure 7 The time-domain waveform of the no-load event and its corresponding MFCC characteristics; Figure 8 This is a schematic diagram of the VAE network constructed in step 4 of the present invention; Figure 9 Standardization of sample data for normal operation of the rollers used for training; Figure 10 Standardization of sample data used for testing; Figure 11 To set and create thresholds for anomaly detection; Figure 12The standardized MFCC coefficient difference and aggregated anomaly rate for a certain no-load event; Figure 13 The standardized MFCC coefficient difference and aggregated anomaly rate for a certain knocking event; Figure 14 This represents the standardized MFCC coefficient difference and the aggregation anomaly rate for a certain friction event. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0032] Reference Figures 1 to 8 Taking a cement ore belt conveyor of about 5 km in a certain factory as an example, about 100 m of sensing optical cable is laid on its drum support column, and a distributed fiber optic acoustic sensor demodulator is used to collect the drum vibration signal. This paper describes the belt conveyor drum abnormality detection method based on distributed fiber optic acoustic sensing and variational autoencoder described in this invention.

[0033] like Figure 1 As shown in the figure, this embodiment provides a distributed optical fiber vibration anomaly detection and localization method based on MFCC and VAE, which mainly includes the following steps: signal acquisition, signal preprocessing, MFCC feature extraction, normal sample training, normal error baseline establishment, test sample detection, anomaly rate map generation, anomaly screening, anomaly aggregation, and fault localization. Through the above steps, the abnormal vibration of belt conveyor rollers can be effectively detected and the anomaly location can be located.

[0034] Step 1: Acquisition of raw drum vibration signals: Gait signals from walking on the floor are acquired using a distributed fiber optic sensing system based on phase-sensitive optical time-domain reflectometry (φ-OTDR). The principle is referenced... Figure 2The distributed fiber optic sensing system comprises an ultra-narrow linewidth laser, an acousto-optic modulator, an erbium-doped fiber amplifier, a circulator, a balanced photodetector, an optical coupler, a photodetector, a sensing fiber, and a high-speed data acquisition card, all integrated into a distributed fiber optic sensing demodulation instrument. The instrument used in the experiment has a detection range of 5 km, encompassing 960 points, a spatial resolution of 5 m, a gauge length of 5 m, a signal sampling rate of 20 kS / s, a 1000M Ethernet communication interface, and uses G.652.D single-mode fiber for sensing. The output interface is either FC or APC. The device connects to a 200m long G.652.D single-mode fiber. The device first generates continuous light from the ultra-narrow linewidth laser, which is split into two paths by the coupler. One path enters the modulator, where it is modulated into pulsed light under the control of the modulation signal. After amplification by the erbium-doped fiber amplifier, the pulsed light enters the sensing fiber through the circulator. In the sensing fiber, the propagation of pulsed light generates Rayleigh scattering signals, which are guided by a circulator and optically mixed with the local oscillator light extracted from the second output of coupler 1 to achieve heterodyne detection. A balanced photodetector converts the optical beat frequency signal into an electrical signal, which is then digitized by an analog-to-digital converter and demodulated digitally by I / Q to reconstruct the phase fluctuations caused by fiber disturbances.

[0035] Building such Figure 3 The diagram shows an experimental scenario for abnormal detection of conveyor rollers based on a distributed fiber optic vibration sensing system. Taking a cement conveyor belt of approximately 5 km in a factory as an example, about 100 m of sensing fiber optic cable is laid on the roller support column, and a distributed fiber optic acoustic sensor demodulator (DAS) is used to collect roller vibration signals (approximately 20 channels). Due to interference such as strong winds in the actual factory measurement environment, insecure fiber optic cable fixation could affect data quality. Therefore, the researchers used anchoring pins and other methods to firmly fix the fiber optic cable to the object under test to reduce the impact of wind-induced disturbances on the acquisition results. The other end of the fiber optic cable is connected to the DAS demodulator.

[0036] The experiment mainly involves four types of events: (1) Knocking / impact. This is used to simulate abnormal roller rotation or impact-related anomalies. In the experiment, knocking is used to simulate fault scenarios related to abnormal roller rotation in order to supplement the abnormal samples, such as... Figure 4 As shown. (2) Friction. Used to simulate roller jamming anomalies. Because there are few real faults and it is difficult to collect enough fault data in a short time, friction is used to artificially simulate roller jamming in the experiment, so as to obtain this type of abnormal vibration signal, such as Figure 5 As shown. (3) Normal operation. This represents the vibration state of the belt conveyor roller under normal operating conditions, serving as the baseline operating category. This category had the largest sample size in the experiment and was used to distinguish it from abnormal conditions, such as... Figure 6As shown. (4) Idle / idle. This indicates a background operating state where the equipment is not operating under normal load. It is an independent event category used to distinguish normal operating events. Its data is collected and labeled on-site along with other categories, such as... Figure 7 As shown.

[0037] The DAS system acquires raw vibration data as a time-domain vibration response distributed along the fiber optic distance, which can be understood as a vibration measurement sequence corresponding to each distance position within a continuous time window. In this embodiment, the raw vibration data is sliced ​​according to a preset time window to obtain several time-domain vibration data blocks. The length of the time window can be set according to the duration of the vibration event to be detected, the system sampling rate, and the actual scenario requirements. Here, the raw data is divided into 2-second time windows as the basic unit for subsequent processing and classification.

[0038] Step 2: Data preprocessing; To reduce the impact of background noise and amplitude scale differences between different samples, the aforementioned time-domain vibration data blocks are preprocessed. The preprocessing steps include, but are not limited to: Step 2-1: Wavelet denoising, used to suppress high-frequency noise or random disturbances in the original vibration data; The acquired signals are preprocessed, including wavelet decomposition for noise reduction. The wavelet basis function used is the db4 wavelet. Wavelet decomposition is performed using different scales. a and location k Multi-scale analysis of the signal is performed through repeated wavelet decomposition. x ( t The wavelet decomposition is divided into low-frequency approximation components and high-frequency detail components. After wavelet decomposition, a soft threshold is applied to the wavelet coefficients. That is, coefficients smaller than the threshold are set to zero, coefficients larger than the threshold are reduced, and the coefficients that have undergone thresholding are retained for inverse wavelet transform to reconstruct the signal.

[0039] Step 2-2: Normalization processing, used to limit the amplitudes of different original vibration data after wavelet denoising to a uniform range or transform them into a uniform statistical distribution.

[0040] Step 3: Construct MFCC features; MFCC feature extraction is performed on the preprocessed time-domain vibration data. The specific process includes: Step 3-1: Pre-weighting The input drum vibration signal is pre-emphasized to enhance the high-frequency components, resulting in a pre-emphasized signal: , The preprocessed time-domain original drum vibration signal is: , This is the pre-weighting factor, typically taken as 0.95-0.98; Step 3-2: Frame Segmentation The pre-emphasized signal is divided into frames according to the set frame length and frame shift to obtain the first frame. m Frame signal: , in, N For frame length, H For frame shift, m The frame number is used; the pre-emphasized signal is divided into frames with a frame length of 25 ms and a frame shift of 15 ms. At a sampling rate of 20 kHz, this corresponds to a frame length of N = 500 and a frame shift of H = 300.

[0041] Step 3-3: Adding windows to Hamming Each frame of signal is multiplied by a Hamming window to reduce spectral leakage, resulting in the windowed frame signal: , Among them, the Hamming window function for: For each frame of signal, multiply by a Hamming window with a window length of N = 500.

[0042] Steps 3-4: Fast Fourier Transform and Power Spectrum Calculation Perform a Fast Fourier Transform on each frame of the windowed signal to obtain its frequency domain representation: , Further calculation of the power spectrum: ,in, k The frequency index is used; a Fast Fourier Transform (FFT) is performed on each frame of the windowed signal, with a frame length of 500 points, an FFT point count of 512, and a frame shift of 300 points, and then the power spectrum of each frame is calculated.

[0043] Steps 3-5: Mel filter bank processing The power spectrum was filtered through a bank of 128 Mel filters to obtain the output energy of each filter: , in, Indicates the first l The frequency response of a Mel filter. The conversion relationship between Mel frequency and actual frequency can be expressed as: ; Steps 3-6: Logarithmic compression Logarithmic operations are performed on the output energies of each Mel filter to compress the dynamic range, yielding the logarithmic Mel energy: ; Steps 3-7: Discrete Cosine Transform (DCT) Decorrelation and Dimensionality Reduction Performing a discrete cosine transform on the 128-dimensional logarithmic Mel energy yields the 20th MFCC coefficient: , Where C is the number of MFCC coefficients retained; Steps 3-8: Construct the MFCC feature matrix The 20-dimensional MFCC coefficients obtained from each frame are arranged in chronological order to form the MFCC feature matrix: , in, M The total frame value is 132, and M is the MFCC characteristic matrix representing the drum vibration mode.

[0044] Through the above process, the original vibration signal can be transformed from a time-domain representation into a cepstral feature representation with spectral structure information. Since MFCC features can characterize local spectral envelopes and mode differences to a certain extent, they are more beneficial for subsequent anomalous pattern recognition. In this embodiment, each sample is ultimately represented as a two-dimensional MFCC feature matrix, where one dimension corresponds to the time frame sequence and the other dimension corresponds to the MFCC coefficient index.

[0045] Step 4: Constructing and Training the VAE Network: Construct a VAE network and train the variational autoencoder VAE using only the MFCC features of normal samples. This allows the VAE to learn the latent representations and reconstruction patterns of normal vibration modes. Step 4, constructing and training the VAE model, includes the following steps: Step 4-1: Construct a normal sample training set.

[0046] The DAS vibration signal of the drum under normal operating conditions was collected, and the corresponding MFCC features were extracted to form a normal sample training set. in, Indicates the first n MFCC feature matrix corresponding to each normal sample N This represents the total number of normal samples.

[0047] For any input sample, its MFCC feature is denoted as: .

[0048] Step 4-2: Construct the variational autoencoder (VAE).

[0049] like Figure 8As shown, a variational autoencoder (VAE) network is constructed, comprising an encoder and a decoder. The encoder maps input features to latent variable distribution parameters, and the decoder reconstructs the input features based on the latent variables. Let the latent variables be... z The encoder outputs the mean and variance parameters of the latent variable distribution and expresses the approximate posterior normal distribution as: in The vector representing the mean of the latent variable distribution. Let cc represent the variance vector of the latent variable distribution, and let cc represent the encoder parameters. The prior distribution of the latent variables is also set as follows: .

[0050] like Figure 8 As shown, the VAE network in this embodiment consists of three parts: an encoder, a latent space, and a decoder. It achieves compressed representation and reconstruction of two-dimensional input data with a size of (64, 64, 1). The encoder first extracts high-order features step by step through two layers of two-dimensional convolutional modules. The first layer includes convolution, batch normalization, activation, and max pooling operations to compress the features to (8, 8, 32), and the second layer further extracts them to (4, 4, 64). Subsequently, a low-dimensional representation is obtained through flattening and fully connected mapping. The mean and standard deviation parameters of the latent distribution are learned through two parallel branches, and then a 4-dimensional latent variable Z is generated using reparameterized sampling. The decoder takes Z as input, restores the initial feature structure through fully connected and reshaping operations, and then gradually completes the spatial resolution reconstruction by using multiple layers of two-dimensional transposed convolutions. Finally, it outputs a reconstruction result with the same size as the original input. Overall, this network achieves effective low-dimensional modeling of input data and MFCC feature reconstruction through a collaborative mechanism of "convolutional encoding - probabilistic mapping - decoding reconstruction".

[0051] Step 4-3: Reparameterize the latent variables.

[0052] To make the latent variable sampling process differentiable, a reparameterization method is used to generate latent variables. z： in, This indicates element-wise multiplication.

[0053] Step 4-4: Reconstruct the input features using the decoder.

[0054] The latent variables obtained in step 4-3 z The input is processed by the decoder to obtain the reconstructed features of the input: in, This represents the reconstructed MFCC features. This represents the decoder mapping function. This represents the decoder parameters.

[0055] Steps 4-5: Construct the loss function and train the VAE.

[0056] During the training phase, only normal samples are selected as training data, and abnormal samples are not introduced. The MFCC feature matrix of normal samples is input into the VAE model for training. The joint loss function is used as the reconstruction error between the input features and the reconstructed features, and the KL divergence between the latent variable distribution and the prior distribution. The loss function is expressed as: in, Indicates the reconstruction error. Denotes KL divergence, This represents the weight coefficients. The encoder parameters are optimized through iterative training on all normal samples. and decoder parameters Its optimization objective is expressed as: ; The training objective of a VAE is to minimize the sum of the reconstruction loss and the latent distribution regularization term, enabling the model to accurately reconstruct normal samples while learning their latent probability distribution. Since only normal samples are used during the training phase, after training, a VAE typically exhibits strong reconstruction capabilities for normal samples but poor reconstruction performance for abnormal samples. In this example, the input consists of 3000 normal samples of 64×64 drum vibration; the training parameters are set to batch size=32 and epoch=500.

[0057] Steps 4-6: Output the trained VAE model.

[0058] After iterative training in steps 4-5, the trained VAE model is obtained: in, and These represent the encoder and decoder parameters after training, respectively. The trained VAE model is used to learn the latent distribution characteristics and reconstruction rules of MFCC features under normal drum vibration conditions, so as to subsequently use the difference between the input features and the reconstructed features to characterize the degree to which the test sample deviates from the normal pattern.

[0059] Step 5: Establish a normal error baseline.

[0060] The normal samples are then input again into the already trained VAE model to obtain the reconstructed MFCC features for each normal sample. Based on the difference between the input MFCC features and the reconstructed MFCC features, a reconstruction error map is calculated for each normal sample. Figure 9As shown, in this embodiment, the reconstruction error map can be represented as a two-dimensional error matrix with the same dimension as the MFCC feature matrix. The error values ​​of all normal samples at each pixel position are statistically analyzed to obtain the mean error and standard deviation of the error corresponding to each pixel position, denoted as [reference to 1]. and Where: i represents the row index of the MFCC feature matrix; j represents the column index of the MFCC feature matrix; This represents the average reconstruction error of a normal sample at the corresponding position; This indicates the degree of error fluctuation in normal samples at corresponding locations. Through the above statistical process, a normal error baseline can be constructed. This baseline reflects the typical reconstruction error distribution of normal samples at each feature location, providing a reference for error normalization of subsequent test samples.

[0061] Step 6: Calculate the reconstruction error of the test sample.

[0062] During the detection phase, the MFCC features of the test sample are input into the trained VAE model to obtain the reconstructed MFCC features corresponding to the test sample. Then, the input features and reconstructed features of the test sample are compared to calculate the reconstruction error map of the test sample. For the ... The error is denoted as (pixel position). The closer the test sample is to the normal pattern, the smaller its reconstruction error is usually; the more the test sample deviates from the normal pattern, the larger its reconstruction error is usually.

[0063] Step 7: Pixel-by-pixel normalization.

[0064] Because the mean error and fluctuation of different pixel positions may not be consistent under normal circumstances, directly comparing the errors at different positions of the test sample can easily lead to interference from certain naturally high error regions in the detection results. Therefore, if... Figure 10 As shown, this embodiment uses a normal error baseline to standardize the test sample error pixel by pixel. Specifically, for each pixel position in the test sample error map, the standardized error value is calculated using the following formula: After standardization, a standardized error plot is obtained. Standardized error values. This can be understood as the degree to which the test sample deviates from the normal error level at the corresponding position. If A larger value indicates a more significant abnormal deviation at that location compared to the normal pattern.

[0065] Step 8: Generate an anomaly rate map.

[0066] To further enhance the significance of outliers, the standardized error map is squared pixel by pixel to obtain the outlier rate map, calculated using the following formula: , in, Indicates the first The anomaly rate or anomaly intensity at each pixel location. After squaring, locations with larger deviations are further amplified, and regardless of whether the deviation direction is positive or negative, a positive value can be used to uniformly measure the degree of anomaly.

[0067] Step 9: Pixel-level threshold filtering After generating the anomaly rate map, based on the preset threshold Pixel-level anomaly rates are filtered. In this embodiment, a threshold is used. This can be set based on the percentiles of the chi-square distribution. For values ​​below the threshold... For pixel locations, set their anomaly rate to zero; for those exceeding the threshold... If the pixel location is not specified, the original anomaly rate value is retained. Through the above filtering process, false responses caused by slight fluctuations or noise can be effectively suppressed, retaining only significantly abnormal pixels. In this embodiment, such as... Figure 11 As shown, determined according to the "Three Sigma Rule" T 1 = 0.96.

[0068] Step 10: Aggregate by time position.

[0069] After pixel-level filtering, the anomaly rate needs to be mapped from the MFCC feature domain back to the anomaly intensity at each time point. Specifically, for each time point, several maximum values ​​are selected from the anomaly rates of multiple frequency bands or cepstral dimensions corresponding to that time point, and their average is calculated as the aggregated anomaly rate for that time point. It should be noted that selecting the top K maximum values ​​for averaging helps retain the main anomaly responses while reducing the dilution effect of a large amount of low-value background on the results. The value of K can be set according to the number of frequency bands in the system and the actual experimental results. Figure 11 As shown, in this embodiment, K is set to 5.

[0070] Step 11, Anomaly Detection.

[0071] The aggregation anomaly rate at each time point in the aggregation anomaly rate curve is compared with the threshold. Comparison is performed. When the aggregated anomaly rate corresponding to a certain location at a certain time exceeds a threshold... When an anomaly is detected at that distance, it is determined that an abnormal event exists. Further, the suspected fault location can be determined based on the number of sensor channels corresponding to the aggregated anomaly rate curve. If multiple significant peaks exist on the curve, multiple anomalies can be identified by combining peak height, peak width, or continuity between adjacent locations. For example... Figure 11 As shown, in this embodiment, while TThe value of 2=1.5 is determined based on the maximum aggregation anomaly rate observed in the results of random 3-day testing of drum vibration normal samples for more than 4 hours. Through the above steps, this embodiment not only achieves anomaly identification of test samples, but also provides the spatial location of the abnormal event along the optical fiber, thereby meeting the actual fault monitoring and location requirements.

[0072] right Figure 3 The on-site experimental setup shown collected data for 3 days using randomized experiments. Normal samples were collected for 4 hours each day, and abnormal samples were manually added for approximately 2 hours, resulting in 5000 normal samples and 1000 abnormal samples. By aggregating the abnormality rates under normal and abnormal conditions, the results showed that the aggregated abnormality rate of almost all abnormal samples exceeded the threshold T² = 1.5, while the rate of most normal samples was below the threshold. Therefore, normal and abnormal states can be clearly distinguished. Ultimately, 463 abnormal samples were successfully detected from the 500 abnormal samples, including the unloaded sample, resulting in a detection accuracy of 92.6%. Figure 12 and Figure 13 and Figure 14 The results show the standardized MFCC coefficient differences and aggregated anomaly rates for typical no-load, knocking, and friction events. Overall, these results demonstrate that this method still possesses good detection capabilities for several types of abnormal faults when faced with real-world conditions, real-noise levels, and real-installation environments.

[0073] The above provides a detailed description of the method for detecting abnormalities in belt conveyor rollers applied to distributed optical fiber sensing, as described in this invention. The examples above are only for helping to understand the method and core ideas of this invention and are not intended to limit it. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of this invention should be considered equivalent substitutions and are included within the protection scope of this invention.

Claims

1. A method for detecting abnormalities in belt conveyor rollers based on DAS and VAE, characterized in that, Includes the following steps: Step 1: Acquisition of raw drum vibration signal: Fix the armored optical cable to the drum support of the belt conveyor, use the DAS system based on Φ-OTDR to acquire the raw vibration data of the belt conveyor drum, and perform time slicing of the raw vibration data according to the preset time window to obtain the time domain vibration data block obtained along a specific distance of the optical cable. Step 2, Signal preprocessing: The time-domain vibration data block is preprocessed, including wavelet denoising and normalization, to suppress noise interference and improve the consistency of subsequent feature extraction. Step 3, MFCC feature extraction: The preprocessed time-domain original drum vibration signal is converted into Mel frequency cepstral coefficients (MFCC) features. Specifically, the vibration signal is subjected to pre-emphasis, framing, Hamming windowing, Fast Fourier Transform (FFT) / power spectrum calculation, Mel filter bank processing, logarithmic compression, and Discrete Cosine Transform (DCT) decorrelation and dimensionality reduction to obtain the MFCC feature matrix representing the vibration mode. Step 4: Construct and train the VAE model: Construct a VAE and train the variational autoencoder VAE using only the MFCC features of normal samples, so that the VAE learns the latent representation of normal vibration modes and their reconstruction rules. Step 5: Establish a normal error baseline: Input normal samples into the trained VAE model to obtain the corresponding reconstructed MFCC features, and calculate the reconstruction error map of normal samples; based on the reconstruction error maps of all normal samples, calculate the mean and standard deviation of the error for each pixel location to form a normal error baseline, where the first... The average value of the pixel positions is denoted as The standard deviation is denoted as ; Step 6: Calculate the reconstruction error of the test samples: Input the MFCC features of the test samples into the trained VAE model to obtain the reconstructed MFCC features corresponding to the test samples; calculate the reconstruction error map based on the difference between the input MFCC features and the reconstructed MFCC features, and denote the error as follows: The error at each pixel position is The larger the error, the more the location deviates from the normal vibration mode; Step 7, Pixel-by-Pixel Standardization: Using the normal error baseline established in Step 5, the reconstruction error map of the test samples is standardized pixel-by-pixel to obtain the standardized error map. The calculation formula is as follows: ， in, Indicates the test sample at the 1st The degree of deviation of each pixel position from the normal baseline; Step 8: Generate an anomaly rate map: Perform pixel-by-pixel square processing on the standardized error map to generate the anomaly rate map. The calculation formula is: ， in, Indicates the first The intensity of the anomaly at each pixel location; Step 9, Pixel-level threshold filtering: Based on the percentile threshold of the chi-square distribution. The anomaly rate map is filtered, retaining only significantly anomalous pixels; for those below a threshold... For pixels, set their anomaly rate to 0 to remove insignificant anomalous responses; Step 10: Aggregate by time location: For each time location, select several of the largest anomaly rate values ​​from the anomaly rates corresponding to all frequency bands and average them to obtain the aggregated anomaly rate for that distance location; repeat the above process for all time locations to form an aggregated anomaly rate curve. Step 11, Anomaly Detection: Compare the aggregated anomaly rate corresponding to each time point with the preset detection threshold. Comparison; when the aggregation anomaly rate at a certain distance location exceeds a threshold When an anomaly is detected at that distance, the location is determined to be abnormal; the suspected fault location is determined based on the number of sensor channels in the corresponding aggregated anomaly rate curve.

2. The detection method according to claim 1, characterized in that, The time slice length in step 1 is set according to the actual monitoring scenario to ensure that a single data block can cover the complete characteristics of the target vibration event.

3. The detection method according to claim 1, characterized in that, The normalization method in step 2 is either min-max normalization or Z-score normalization.

4. The detection method according to claim 1, characterized in that, Step 3, MFCC feature extraction, includes the following steps: Step 3-1: Pre-weighting The input drum vibration signal is pre-emphasized to enhance the high-frequency components, resulting in a pre-emphasized signal: ， The preprocessed time-domain original drum vibration signal is: , This is the pre-weighting factor, typically taken as 0.95-0.98; Step 3-2: Frame Segmentation The pre-emphasized signal is divided into frames according to the set frame length and frame shift to obtain the first frame. m Frame signal: ， in, N For frame length, H For frame shift, m The frame number; Step 3-3: Adding windows to Hamming Each frame of signal is multiplied by a Hamming window to reduce spectral leakage, resulting in the windowed frame signal: ， Among them, the Hamming window function for: ； Multiply each frame of signal by a Hamming window, with a window length of N = 500; Steps 3-4: Fast Fourier Transform and Power Spectrum Calculation Perform a Fast Fourier Transform on each frame of the windowed signal to obtain its frequency domain representation: ， Further calculation of the power spectrum: ,in, k Frequency index; Steps 3-5: Mel filter bank processing Power spectrum L The output energy of each filter is obtained by filtering with a group of Mel filters: ， in, Indicates the first l The frequency response of a Mel filter. The conversion relationship between Mel frequency and actual frequency can be expressed as: ; Steps 3-6: Logarithmic compression Logarithmic operations are performed on the output energies of each Mel filter to compress the dynamic range, yielding the logarithmic Mel energy: ; Steps 3-7: Discrete Cosine Transform (DCT) Decorrelation and Dimensionality Reduction Performing a discrete cosine transform on the logarithmic Mel energy yields the th c MFCC coefficients: ， Where C is the number of MFCC coefficients retained; Steps 3-8: Construct the MFCC feature matrix The MFCC coefficients obtained from each frame are arranged in chronological order to form the MFCC feature matrix: ， in, M M represents the total number of frames, and M is the MFCC characteristic matrix representing the drum vibration mode.

5. The detection method according to claim 1, characterized in that, Step 4, constructing and training the VAE model, includes the following steps: Step 4-1: Construct a normal sample training set The DAS vibration signal of the drum under normal operating conditions was collected, and the corresponding MFCC features were extracted to form a normal sample training set. ， in, Indicates the first n MFCC feature matrix corresponding to each normal sample N This represents the total number of normal samples; For any input sample, its MFCC feature is denoted as: ; Step 4-2: Constructing a Variational Autoencoder (VAE) A variational autoencoder (VAE) is constructed, comprising an encoder and a decoder. The encoder maps input features to latent variable distribution parameters, and the decoder reconstructs the input features based on the latent variables. Let the latent variables be... z The encoder outputs the mean and variance parameters of the latent variable distribution and expresses the approximate posterior normal distribution as: in The vector representing the mean of the latent variable distribution. Let cc represent the variance vector of the latent variable distribution, and let cc represent the encoder parameters. The prior distribution of the latent variables is also set as follows: ; Step 4-3: Reparameterize sampling of latent variables To make the latent variable sampling process differentiable, a reparameterization method is used to generate latent variables. z： in, This represents element-wise multiplication; Step 4-4: Reconstruct input features using the decoder The latent variables obtained in step 4-3 z The input is processed by the decoder to obtain the reconstructed features of the input: in, This represents the reconstructed MFCC features. This represents the decoder mapping function. Indicates decoder parameters; Steps 4-5: Construct the loss function and train the VAE The VAE is trained using the reconstruction error between the input features and the reconstructed features, and the KL divergence between the latent variable distribution and the prior distribution, as the joint loss function. The loss function is expressed as: in, Indicates the reconstruction error. Denotes KL divergence, This represents the weight coefficients. The encoder parameters are optimized through iterative training on all normal samples. and decoder parameters Its optimization objective is expressed as: ； Steps 4-6: Output the trained VAE model After iterative training in steps 4-5, the trained VAE model is obtained: in, and These represent the encoder parameters and decoder parameters after training, respectively. The trained VAE model is used to learn the potential distribution characteristics and reconstruction rules of MFCC features under normal drum vibration conditions, so as to use the difference between the input features and the reconstructed features to characterize the degree to which the test sample deviates from the normal mode.

6. The detection method according to claim 1, characterized in that, The reconstruction error in step 5 is the absolute error, squared error, or other error form that can characterize the reconstruction difference between the input MFCC feature and the corresponding pixel position of the reconstructed MFCC feature.

7. The detection method according to claim 1, characterized in that, The filtering threshold in step 9 Determined based on the percentiles of the chi-square distribution to preserve significantly anomalous pixels.

8. The detection method according to claim 1, characterized in that, The aggregation operation in step 10 can employ maximum value averaging, top K maximum value averaging, weighted averaging, or robust statistics to reduce the impact of occasional errors in a single frequency band on the detection results.

Images