DAS-based belt conveyor state early warning and data processing method

Abstract

The application discloses a kind of DAS-based belt conveyor state early warning and data processing method, belong to industrial equipment intelligent monitoring technical field, this method is through pulse digital coding enhanced Φ-OTDR technology acquisition sensitized optical fiber node place's distributed acoustic signal, by matching decoding and phase demodulation restores acoustic waveform and constructs space-time acoustic field matrix, extracts mel spectrum, contrast spectrum and three kinds of multimodal voiceprint features of higher-order spectrum, utilizes noise perception dynamic filtering to inhibit environmental interference, input multi-scale attention fusion network identifies carrier roller operating state, generates hierarchical early warning decision and closed-loop feedback updates acquisition parameter and identification model.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and signal processing technology for industrial equipment, and in particular to a method for early warning and data processing of belt conveyor status based on DAS. Background Technology

[0002] Belt conveyors, as core equipment for continuous transport of bulk materials, are widely used in industrial settings such as coal mines, ports, steel mills, and power plants. Under long-distance, high-capacity operation conditions, thousands of idlers along the line bear continuous dynamic loads and complex environmental stresses, and their operating status directly affects the safety and continuity of the entire conveyor line. However, idler failures exhibit diverse characteristics: cylinder wear can be caused by excessive rotational resistance or deflection friction; bearing damage often stems from seal failure or poor lubrication; and shaft bending deformation is often related to improper selection or localized overload. These failures often initially manifest as subtle changes in acoustic characteristics. If not detected and addressed promptly, they can easily escalate into serious accidents such as idler jamming, belt tearing, or even fires, causing significant economic losses and safety risks to industrial production.

[0003] In existing technologies, monitoring the operating status of idler rollers mainly relies on two methods: manual inspection and point sensors. Manual inspection requires inspectors to rely on auditory experience to judge whether the idler rollers are making abnormal sounds in harsh conditions such as high temperature, heavy dust, and high noise. This method is not only inefficient and highly subjective, but also difficult to achieve full-time coverage of large-scale distributed idler roller groups. While point sensor solutions can acquire vibration or acoustic signals in localized areas, they are limited by the number of sensors and their installation locations, making it impossible to monitor every single idler roller along a conveyor line that is thousands of meters long. Furthermore, the sensors themselves require power supply and communication cabling, posing inherent safety constraints in explosion-proof environments such as mines. In addition, traditional vibration detection methods based on contact accelerometers or microphones require high installation accuracy, experience rapid sensor performance degradation under harsh conditions, and have high maintenance costs.

[0004] In recent years, distributed optical fiber sensing technology has attracted attention due to its inherent advantages such as being passive, long-distance, and resistant to electromagnetic interference. Distributed acoustic sensing systems based on phase-sensitive optical time-domain reflectometry can transform a single ordinary single-mode optical fiber into a continuous acoustic sensor array, theoretically possessing full-coverage monitoring capabilities along the entire line. However, directly applying this technology to the monitoring of idler rollers in belt conveyors still faces several challenges. First, standard communication optical cables have limited sensitivity to acoustic signals, making it difficult to reliably collect low-amplitude sound waves generated by idler roller operation from distances of several meters. Second, the operating environment of belt conveyors includes continuous mechanical vibration noise from idler roller rotation and belt operation, material impact noise, and relatively weak natural environmental noise such as wind and rain. These noises overlap significantly with the acoustic signatures of idler roller faults in the frequency domain, making it difficult to effectively distinguish them based on a single acoustic feature. Third, most existing voiceprint recognition algorithms are trained on static datasets and lack adaptive update mechanisms for dynamic changes in operating conditions and model performance degradation, leading to a gradual decrease in recognition accuracy over long-term operation.

[0005] Furthermore, at the data processing level, existing fault identification methods based on acoustic signals typically utilize only single-dimensional spectral features for discrimination, such as extracting only Mel-frequency cepstral coefficients or analyzing only power spectral density. This single-mode representation method captures limited information dimensions and is prone to misjudgment under conditions of severe noise interference. Simultaneously, existing systems' voiceprint recognition models are usually trained all at once before deployment and lack the ability to be updated online based on actual operating data. As the wear condition of the idler rollers gradually evolves and environmental conditions change seasonally, the model's generalization performance continuously declines, leading to a gradual increase in the false negative rate, failing to meet the needs of long-term stable monitoring in industrial settings.

[0006] In summary, existing technologies have limitations to varying degrees in multiple aspects, including sensing coverage, signal sensitivity, feature representation dimensions, and long-term model maintenance. These problems are intertwined and mutually restrictive, and improving only one aspect is unlikely to achieve a substantial breakthrough in overall performance. Therefore, there is an urgent need for a comprehensive data processing method that can simultaneously achieve long-distance full-coverage sensing, high-sensitivity acoustic acquisition, multi-modal feature fusion, and online adaptive updates to enable accurate early warning of the operating status of belt conveyor idlers. Summary of the Invention

[0007] To address the technical problems in existing technologies, such as insufficient sensitivity of distributed optical fiber sensing, difficulty in separating the overlapping acoustic signature spectra of environmental noise and idler roller faults, and the lack of an online update mechanism for acoustic signature recognition models, this invention provides a DAS-based method for early warning and data processing of belt conveyor status. This invention uses distributed acoustic sensing as the perception foundation, multimodal acoustic signature feature fusion as the core processing method, and closed-loop adaptive updating as a long-term guarantee mechanism, forming a complete data processing link from physical layer signal acquisition to application layer early warning decision-making.

[0008] The technical solution adopted in this invention is: a method for early warning and data processing of belt conveyor status based on DAS, comprising the following steps: A narrow-linewidth coherent laser pulse sequence with pulse digital encoding modulation is injected into the sensing optical fiber deployed along the belt conveyor. Backscattered interference light signals generated by Rayleigh scattering at each sensitization node are received. The spatial position of each sensitization node is determined based on the round-trip time delay of the light pulse, and the distributed acoustic raw signal is obtained. Matched decoding and phase demodulation are performed on the backscattered interference light signals at each sensitization node to extract the phase change of the Rayleigh scattered light. The acoustic vibration waveforms at each sensitization node are recovered based on the phase change, and a spatiotemporal sound field matrix is ​​constructed with the spatial position of the sensitization node as the row index and the sampling time as the column index. Multimodal acoustic features are extracted node by node from the spatiotemporal sound field matrix, including Mel frequency cepstral coefficient features, contrast spectrum features and higher-order spectrum features. The three types of features are spliced ​​along the channel dimension to form a multimodal feature tensor. A noise-sensing dynamic filtering mechanism is used to adaptively suppress environmental interference in the multimodal feature tensor. The filtering parameters are dynamically adjusted according to the estimated environmental noise level to separate the effective acoustic components related to idler operation. The feature tensors after interference suppression are input into a multi-scale attention fusion network to perform idler roller running status recognition, and output the state classification results and confidence scores of the idler rollers corresponding to each enhanced node. Based on the state classification results and confidence levels, an early warning decision is generated, and the identification results are fed back to the signal acquisition stage to adjust the pulse coding parameters and sampling strategy. At the same time, the multi-scale attention fusion network is incrementally updated using online labeled samples.

[0009] Furthermore, the pulse digital code modulation adopts Golay complementary sequence coding, with a coding length of 64 bits to 256 bits, a single pulse width of 20 ns to 100 ns, and a pulse repetition frequency of 500 Hz to 2000 Hz; the sensitivity enhancement nodes are arranged along the steel frame of the belt conveyor at intervals of 5 m to 10 m, and the number of optical fiber coils in each sensitivity enhancement node is 3 to 8.

[0010] Furthermore, the phase demodulation adopts a 3×3 coupler demodulation scheme, and the matched decoding includes performing cross-correlation operation on the received backscattered interference optical signal and the transmitted coded sequence to improve the signal-to-noise ratio. The improvement in signal-to-noise ratio is proportional to the square root of the coding length.

[0011] Furthermore, in the noise perception dynamic filtering mechanism, the estimated environmental noise level is obtained by calculating the spatial correlation coefficient between adjacent sensitized nodes in the spatiotemporal sound field matrix. When the spatial correlation coefficient is lower than a preset threshold, the corresponding frequency band is determined to be the dominant frequency band of environmental noise. The preset threshold ranges from 0.3 to 0.7.

[0012] Furthermore, the extraction process of the Mel frequency cepstral coefficient features includes: performing pre-emphasis processing on the acoustic vibration waveform, with a pre-emphasis coefficient of 0.95 to 0.98; performing frame segmentation and windowing with a frame length of 20ms to 40ms and a frame shift of 8ms to 15ms; filtering through a Mel filter bank containing 20 to 40 triangular filters after fast Fourier transform; performing discrete cosine transform after taking the logarithm, and retaining the first 12 to 20 dimensions of coefficients as the Mel frequency cepstral coefficient features.

[0013] Furthermore, the higher-order spectral features are obtained through bispectral analysis, specifically including: calculating the third-order cumulant of the short-time frame signal of the acoustic vibration waveform, then obtaining the bispectral matrix through two-dimensional discrete Fourier transform, and extracting the logarithmic mean of the bispectral amplitude and the weighted variance of the bispectral phase from the bispectral matrix as higher-order spectral feature vectors.

[0014] Furthermore, the multi-scale attention fusion network includes a multi-scale feature extraction module, a channel attention module, and a classification decision module; the multi-scale feature extraction module adopts a parallel multi-branch convolutional structure with branch convolutional kernel sizes of 1×1, 3×3, and 5×5, respectively; the channel attention module performs global average pooling on the output of each branch and generates channel weight vectors through two fully connected layers, and then weights and fuses the features of each branch; the classification decision module includes a global average pooling layer and a fully connected classification layer.

[0015] Furthermore, during the construction of the spatiotemporal sound field matrix, bandpass filtering is performed on the recovered acoustic vibration waveform, with a passband frequency range of 10Hz to 8000Hz; normalization processing is performed on the bandpass filtered signal, with the normalization method being zero-mean unit variance standardization performed according to the sensitization node.

[0016] Furthermore, the generation of the early warning decision includes: setting four status levels: normal, attention, early warning, and alarm; triggering an early warning when the abnormal confidence level of the idler corresponding to a single enhanced node continuously exceeds a first threshold and the duration exceeds a first time window; and triggering an alarm when the abnormal confidence level exceeds a second threshold; the first threshold is 0.6 to 0.75, the second threshold is 0.85 to 0.95, and the first time window is 30s to 120s.

[0017] Furthermore, the incremental update adopts an experience replay strategy, maintaining a sample buffer pool of fixed capacity. Whenever a new online labeled sample is obtained, historical samples are randomly sampled from the sample buffer pool and mixed with the new samples to form a training batch. The classification decision module of the multi-scale attention fusion network is fine-tuned by decaying the learning rate.

[0018] This method first injects a narrow-linewidth coherent laser pulse sequence modulated by pulse digital coding into sensing fibers deployed along a belt conveyor, and then receives the backscattered interference light signals generated by Rayleigh scattering at each sensitization node to obtain the distributed acoustic raw signal. Pulse digital coding technology can effectively improve the detection signal-to-noise ratio without sacrificing spatial resolution, and the fiber winding structure within the sensitization node significantly enhances the coupling efficiency of the acoustic vibration signal. Subsequently, the acoustic vibration waveforms at each sensitization node are recovered through matched decoding and phase demodulation to construct a spatiotemporal sound field matrix.

[0019] In the feature extraction stage, this invention extracts three complementary types of voiceprint representations node-by-node from the spatiotemporal sound field matrix: Mel-frequency cepstral coefficient features, contrast spectrum features, and higher-order spectrum features. Mel-frequency cepstral coefficient features capture the spectral envelope information of the sound signal; contrast spectrum features highlight the dynamic differences between steady-state background and transient anomalies through short-time power spectrum difference operations; and higher-order spectrum features utilize bispectral analysis to suppress Gaussian noise and retain nonlinear coupling information. These three types of features are concatenated along the channel dimension to form a multimodal feature tensor, achieving a comprehensive representation of the voiceprint from three orthogonal dimensions: spectral envelope, time-varying dynamics, and nonlinear structure.

[0020] Furthermore, this invention employs a noise-aware dynamic filtering mechanism to adaptively suppress environmental interference in the multimodal feature tensor. It utilizes the spatial correlation between signals from adjacent enhanced nodes to distinguish between environmental noise and local idler signals, effectively solving the problem of overlapping noise and fault acoustic signature spectra under complex operating conditions. The interference-suppressed feature tensor is input into a multi-scale attention fusion network. Multi-branch parallel convolution captures local patterns at different scales, and the channel attention mechanism adaptively allocates the fusion weights of each branch, achieving accurate classification of various operating states such as normal idler operation, wear, bearing abnormality, and deformation.

[0021] This invention also constructs a closed-loop feedback path from early warning decision-making to signal acquisition and model updating. The identification results are not only used to generate tiered early warning information, but also to adjust pulse coding parameters to adapt to different monitoring accuracy requirements. Simultaneously, online labeled samples are used to perform incremental updates to the identification model, enabling the system to maintain a high identification accuracy rate over long-term operation. This closed-loop mechanism transforms the signal acquisition end and data processing end from two separate, independent links into a collaborative whole capable of autonomous adjustment and continuous optimization. When an abnormal trend is detected in a certain section of the idler rollers, the system automatically increases the sampling accuracy and coding length of that section, achieving on-demand allocation of monitoring accuracy under resource-constrained conditions.

[0022] The beneficial effects of this invention include: utilizing the synergistic effect of pulse digital coding and sensitive fiber structure, the detection signal-to-noise ratio is improved by more than 6dB without increasing the power of the light source, enabling the system to reliably acquire acoustic signals generated by idler roller operation within a 3.5m detection radius; through multimodal fusion of Mel spectrum, contrast spectrum, and higher-order spectrum, the recognition accuracy is improved by 8% to 15% compared to single feature methods, and the sensitivity to weak anomalies such as early bearing wear is significantly enhanced; with the help of cascaded processing of noise perception filtering and multi-scale attention network, the recognition accuracy can still be maintained at more than 85% under strong noise conditions with a signal-to-noise ratio of less than 0dB, effectively overcoming various environmental interferences such as continuous mechanical vibration noise, material impact noise, and wind and rain generated by idler roller rotation and belt operation; the closed-loop online update mechanism ensures that the recognition accuracy of the model decreases by no more than 3% after 6 months of continuous operation, significantly reducing the system maintenance frequency and manual intervention costs. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of the DAS-based belt conveyor status early warning and data processing method provided in this embodiment of the invention;

[0024] Figure 2 This is a flowchart of distributed acoustic signal acquisition and phase demodulation provided in an embodiment of the present invention;

[0025] Figure 3 This is a flowchart of multimodal voiceprint feature extraction and fusion provided in an embodiment of the present invention;

[0026] Figure 4 This is a diagram of the adaptive interference suppression and multi-scale attention fusion network structure provided in the embodiments of the present invention;

[0027] Figure 5 This is a flowchart of closed-loop early warning decision-making and online model update provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1 This invention provides a DAS-based method for early warning and data processing of belt conveyor status. This method is applicable to industrial sites equipped with long-distance belt conveyors, such as coal mines, ports, and steel plants. This embodiment uses a 1.2km long belt conveyor as an example for detailed explanation. The conveyor has approximately 170 sets of idlers along its length, with a spacing of about 7m between idlers. Each set of idlers includes 3 to 5 rollers (upper and lower idlers), resulting in a total of over 600 rollers along the entire conveyor line. The sensing fiber is laid along one side of the belt's steel frame, utilizing a spare single-mode fiber core from an existing communication optical cable as the sensing medium. This eliminates the need for laying dedicated optical cables, significantly reducing deployment costs and construction difficulty. The entire method comprises six core steps: distributed acoustic signal acquisition, phase demodulation and acoustic reconstruction, multimodal acoustic signature extraction, adaptive environmental interference suppression, multi-scale attention fusion recognition, and closed-loop early warning decision-making and online model update. These steps form a deeply coupled closed loop of forward data-driven and backward parameter feedback.

[0030] Step S1: Distributed acoustic signal acquisition based on PDC-enhanced Φ-OTDR. (Refer to...) Figure 2 In one embodiment of the present invention, the acquisition of distributed acoustic signals is achieved based on pulse digital coding enhanced phase-sensitive optical time-domain reflectometry (PTZ). Specifically, the system's light source employs a narrow-linewidth distributed feedback laser with an output wavelength of 1550 nm and a linewidth of less than 3 kHz to ensure that the coherence length meets the requirements for long-distance sensing. The continuous light output of the laser is pulse-modulated by a semiconductor optical amplifier to form a single-pulse optical pulse with a width of 50 ns, corresponding to a spatial resolution of approximately 5 m.

[0031] Preferably, the present invention employs pulse digital coding technology to modulate the optical pulse sequence. In this embodiment, a 128-bit Golay complementary sequence pair is selected as the coding method. The A and B sequences in the complementary sequence pair are sequentially modulated onto continuous laser pulses and injected into the sensing fiber. The core advantage of the Golay complementary sequence is that the sum of the autocorrelation functions of the A and B sequences is a perfect delta function. Therefore, by performing matched filtering on the two sets of backscattered signals at the receiving end and then summing them, the signal-to-noise ratio can be improved by approximately [missing information - likely a percentage] while eliminating sidelobes. times, of which This is the encoding length. In this embodiment... This corresponds to an improvement in signal-to-noise ratio of approximately 11.3 dB. The pulse repetition frequency is set to 1000 Hz, the channel polling time is 30 s, and the sampling rate is 16 kHz.

[0032] On the sensing fiber side, enhancement nodes are arranged at 7m intervals along the steel frame of the belt conveyor. Each enhancement node adopts a suspended fiber optic coil design, with the coil separated from the base by a damping pad, remaining suspended to reduce interference from the vibration of the belt conveyor's channel steel structure on the target acoustic signature acquisition. Preferably, the fiber optic coil in each enhancement node has 5 turns and a coil diameter of 60mm. Drawing inspiration from the diaphragm and pickup hole design of audio pickups, an acoustic coupling hole is provided on the outer shell of the enhancement head to allow external sound waves to be efficiently transmitted to the surface of the coiled fiber optic coil. The sensitivity of the enhancement sensor reaches 5rad / Pa, the detection radius is 3.5m, and the frequency response range is 10Hz to 8kHz. After the pulse peak power is amplified to 400mW by an erbium-doped fiber amplifier, it is injected into one end of the sensing fiber through an optical circulator. The backscattered Rayleigh interference light is guided by the circulator to the avalanche photodiode module at the receiving end for photoelectric conversion. The analog-to-digital converter digitizes the analog signal at a sampling rate of 100MHz. Each frame of scattering trajectory data is transmitted to the signal processing unit via a USB 3.0 bus.

[0033] The output of this step is a digitized trajectory sequence of the original backscattered interference optical signal, with a data dimension of [missing information]. ,in To increase the number of sensitive nodes, this embodiment... ; This represents the number of sampling points per frame. It is important to note that the placement of the sensor enhancement nodes is carefully planned. Each node is installed in the middle of two adjacent sets of idlers, ensuring its detection radius covers both sets simultaneously. This achieves full monitoring coverage of all idlers without increasing the number of sensor enhancement nodes. The sensor enhancement nodes are connected to the sensing fiber via fusion splicing. The bending radius of the fiber in the coiled section is not less than 30mm to avoid power loss due to excessive bending. In one embodiment of this invention, the sensor enhancement head shell is made of engineering plastic, has an IP65 protection rating, and operates within a temperature range of -30℃ to 60℃, adapting to various harsh environmental conditions such as underground coal mines and outdoor environments. This trajectory sequence serves as the direct input for step S2.

[0034] Step S2: Phase demodulation and acoustic reconstruction of the enhanced coupling signal. (Refer to...) Figure 2Upon receiving the raw backscattered interferometric optical signal, matched decoding is first performed on the two sets of Golay-encoded trajectories. Specifically, the backscattered trajectory corresponding to sequence A is cross-correlated with the time-reversed conjugate of sequence A, and the backscattered trajectory corresponding to sequence B is cross-correlated with the time-reversed conjugate of sequence B. Then, the two sets of cross-correlation results are summed point by point. Since the autocorrelation sidelobes of the Golay complementary sequence pairs cancel each other out, the equivalent impulse response obtained after summing is a narrow peak without sidelobes, thus significantly improving the signal-to-noise ratio while maintaining a 5m spatial resolution.

[0035] After matching and decoding are completed, phase demodulation is performed on the decoded trajectory sequence. In one embodiment of the present invention, a phase demodulation scheme based on a 3×3 fiber coupler is adopted. The 3×3 coupler splits the backscattered light into three interference signals with a fixed phase difference of 120°. After the three signals are converted into electrical signals by photodetectors, the phase change of the Rayleigh scattered light is extracted using the following phase recovery formula:

[0036] ,in, For position At the moment The amount of phase change, in rad; , , The three output ports of the 3×3 coupler are located at positions... place, time The light intensity measured is in mW. Indicates the spatial location of the sensitivity enhancement node along the optical fiber, in meters; The value represents the sampling time, expressed in seconds. This formula is based on the principle of three-channel 120° phase-shift interference, and can directly obtain phase information without additional modulation and demodulation devices. It has the advantages of simple structure and strong real-time performance.

[0037] After obtaining the time series of phase changes at each sensitization node, it is further converted into an acoustic vibration waveform. Since there is a linear relationship between acoustic strain and phase change in the optical fiber, and the phase change is proportional to the strain experienced by the fiber, therefore... This can be considered a linear representation of the acoustic vibration signal at that location. Preferably, the recovered acoustic vibration waveform is bandpass filtered, with the passband frequency range set from 10Hz to 8000Hz, to filter out DC offset and high-frequency noise outside the acoustic frequency band. The bandpass-filtered signal is then normalized to zero mean and unit variance at each sensitization node to eliminate amplitude inconsistencies between nodes caused by differences in fiber coupling efficiency.

[0038] Finally, the standardized acoustic vibration waveforms of all the amplification nodes are arranged into rows according to their spatial location and columns according to their sampling time to construct a spatiotemporal sound field matrix. ,in To increase the number of sensitive nodes, This represents the number of sampling points within each analysis time window. In this embodiment, the analysis window length is 1 second, and the sampling rate is 16 kHz. Each row of the spatiotemporal sound field matrix represents the complete acoustic time series of an enhanced node within the current analysis window, while each column represents a snapshot of the spatial sound field distribution of all enhanced nodes at a given sampling time. This two-dimensional matrix representation not only preserves the temporal waveform information of each node's signal but also provides a natural data structure for subsequent spatial correlation analysis. Preferably, adjacent analysis windows have a 50% overlap rate, meaning a spatiotemporal sound field matrix is ​​output every 0.5 seconds to improve the time tracking accuracy of roller state changes. After each new analysis window is output, the data from the previous window is written to a circular buffer for use in the sound playback function. The capacity of the circular buffer is set to the acoustic data of the most recent 24 hours. This spatiotemporal sound field matrix serves as the input for multimodal acoustic feature extraction in step S3.

[0039] Step S3: Multimodal voiceprint feature extraction and joint time-frequency representation. (Refer to...) Figure 3 After obtaining the spatiotemporal sound field matrix, this invention extracts three types of complementary acoustic features from the acoustic signals of the sensitized nodes corresponding to each row of the matrix, so as to achieve a comprehensive characterization of the acoustic state of the idler roller.

[0040] The first category is the Mel frequency cepstral coefficient feature. The specific extraction process is as follows: First, pre-emphasis processing is performed on the acoustic vibration waveform of each sensitization node, and the pre-emphasis coefficients... Taking 0.97, this step passes through a first-order high-pass filter. To compensate for the attenuation of high-frequency components, the signal spectrum is flattened. After pre-emphasis, the signal is framed with a frame length of 25ms and a frame shift of 10ms, and a Hamming window is added to each frame to reduce spectral leakage. Then, a 512-point Fast Fourier Transform is performed on each frame to calculate the power spectral density, and the signal is filtered through a Mel filter bank containing 26 triangular bandpass filters. The center frequency of the Mel filter bank is uniformly distributed on the Mel frequency scale, with a low-frequency end of 10Hz and a high-frequency end of 8000Hz. The conversion relationship between Mel frequency and linear frequency is as follows:

[0041] ,in, Mel frequency, measured in megahertz (mel). The linear frequency is expressed in Hz. This nonlinear mapping simulates the difference in the human ear's ability to perceive and distinguish different frequency bands, exhibiting high resolution in the low-frequency region and gradually decreasing resolution in the high-frequency region. This precisely matches the frequency distribution characteristics of idler roller fault sound patterns—the characteristic frequencies generated by early bearing wear are usually concentrated in the mid-to-low frequency range of 300Hz to 2000Hz, while abnormal noises caused by drum wear and deformation are presented over a wider frequency range. The natural logarithm of the energy of each sub-band after filtering is taken, and then a 26-point discrete cosine transform is performed, retaining 13 coefficients from the 2nd to the 14th dimensions as static Mel frequency cepstral coefficients. In one embodiment of the invention, first-order and second-order difference coefficients are further calculated to characterize the dynamic rate of change and acceleration information of the sound pattern. The first-order difference is calculated using the linear regression slope of two frames before and after, and the second-order difference is performed again on the basis of the first-order difference, ultimately obtaining a 39-dimensional Mel frequency cepstral coefficient feature vector. In this embodiment, the 1-second analysis window corresponds to approximately 99 frames; therefore, the dimension of the Mel-frequency cepstral coefficient feature formation for each sensitization node is [dimension missing]. The two-dimensional matrix. It should be noted that the reason for choosing 13-dimensional static coefficients instead of more dimensions is that, as verified by experiments, cepstral coefficients of more than 14th order contribute almost zero identification information in the idler roller acoustic print classification task, and instead introduce dimensional redundancy, increasing the risk of overfitting in subsequent classifiers.

[0042] The second category is contrastive spectral features. Contrastive spectral features aim to characterize the short-term dynamic changes of sound signals, effectively highlighting the differences between steady-state operating backgrounds and transient abnormal events. Specifically, the acoustic vibration waveform is first framed and windowed according to the same framing parameters as the Mel-frequency cepstral coefficients, and the power spectral density of each frame is calculated. ,in For frequency index, Use the frame index. Then calculate the logarithmic difference of the power spectral density between adjacent frames:

[0043] ,in, For frequency In frame The contrast spectral values ​​are dimensionless; For frequency In frame The power spectral density value, in W / Hz; The frame interval is 1 frame, corresponding to 10ms in this embodiment. Comparison spectrum. The positive values ​​correspond to frequency bands with increased energy, while the negative values ​​correspond to frequency bands with decreased energy. The physical significance of this characteristic is that the acoustic signal generated by a normally operating idler roller exhibits high temporal stability on a short time scale, with the contrast spectrum values ​​approaching zero across all frequency bands. However, when an idler roller malfunctions, such as when a peeling defect occurs in the inner ring of a bearing, the rolling elements generate brief impact pulses each time they pass over the defect location, causing a rapid increase in energy within a specific frequency band, resulting in a significant positive peak in the contrast spectrum for that band. Therefore, the contrast spectrum characteristic has a unique advantage in detecting intermittent and sudden idler roller anomalies. Preferably, the contrast spectrum is divided into four sub-bands along the frequency axis, with sub-band boundaries of 10Hz to 500Hz, 500Hz to 1500Hz, 1500Hz to 4000Hz, and 4000Hz to 8000Hz, respectively. These four sub-bands cover the characteristic frequency bands of different fault modes, such as low-frequency resonance of idler rollers, mid-frequency friction noise, high-frequency bearing impact, and ultra-high-frequency microcrack vibration. The mean and standard deviation of the contrast spectrum values ​​within each sub-band are calculated to form an 8-dimensional contrast spectrum feature vector. The contrast spectrum feature formation dimension of each sensitization node within a 1s window is... A two-dimensional matrix.

[0044] The third category is higher-order spectral features. Higher-order spectral analysis can preserve the nonlinear coupling information in the signal while naturally suppressing the influence of Gaussian white noise. This is particularly effective for detecting the nonlinear acoustic components generated by mechanical faults in idler rollers under strong noise backgrounds. From the perspective of signal processing theory, all cumulants above the second order in a Gaussian random process are zero. Therefore, the bispectral domain is naturally immune to Gaussian noise. The acoustic signals generated by mechanical faults usually contain rich nonlinear harmonic coupling components, which manifest as significant peak structures in the bispectral domain. Specifically, the third-order cumulant is calculated for each frame of the acoustic vibration waveform. Then, a two-dimensional discrete Fourier transform is performed to obtain the bispectrum:

[0045] ,in, Dual frequency The bispectral value at that location, in units of (W / Hz). ; For delay The third-order cumulant at the location; and These are the first and second frequency variables, respectively, in Hz; and The time delay variable is expressed in seconds. In actual calculations, to reduce computational load, this embodiment takes 256 sampling points for each frame of signal for bispectral estimation, and uses an indirect method to estimate the third-order cumulant before transforming it to the frequency domain. Two scalars, the logarithmic mean of the bispectral amplitude and the weighted variance of the bispectral phase, are extracted from the bispectral matrix as the higher-order spectral features of that frame. The logarithmic mean of the bispectral amplitude reflects the overall intensity level of the nonlinear coupling components in the frame signal, while the weighted variance of the bispectral phase characterizes the dispersion of the nonlinear coupling relationship in the frequency plane. Preferably, when the idler roller bearing malfunctions, strong second-order phase coupling occurs between the characteristic frequency and its harmonics, resulting in a significant increase in the logarithmic mean of the bispectral amplitude and a significant decrease in the weighted variance of the bispectral phase. This change pattern is distinctly different from the characteristics under normal operating conditions. The higher-order spectral features formed by each enhancement node within a 1-second window have the following dimensions: A two-dimensional matrix.

[0046] Finally, the three types of features are concatenated along the channel dimension to form a multimodal feature tensor. 49 represents the total number of channels for the 39+8+2 dimensional features. The three types of features complementarily characterize the idler roller acoustic signature from three orthogonal dimensions: spectral envelope, time-varying dynamics, and nonlinear structure. The absence of any one type of feature in subsequent steps will lead to a significant decrease in recognition performance. This synergistic relationship will be quantified in the effect verification section. This multimodal feature tensor serves as the input for step S4.

[0047] Step S4: Adaptive environmental interference suppression and signal-to-noise ratio enhancement. (Refer to...) Figure 4 Various background noises inevitably exist in the operating environment of belt conveyors. Among them, the continuous mechanical vibration noise generated by the rotation of idlers and the operation of the belt is the dominant component, with a significantly higher intensity than other interference sources and running throughout the entire conveyor line. Other noises include broadband noise caused by the impact of falling materials and relatively weak natural environmental noises such as wind and rain. These interferences overlap to varying degrees with the sound patterns of idler faults in the frequency domain, and if not effectively suppressed, they will seriously affect the accuracy of subsequent identification. This invention proposes a noise-sensing dynamic filtering mechanism. Its core idea is to utilize the spatial distribution characteristics of the sensitive node array to distinguish between spatially correlated environmental noise and spatially localized idler signals.

[0048] Specifically, for multimodal feature tensors For each time frame and feature channel, calculate the spatial correlation coefficient between adjacent enhancement nodes:

[0049] ,

[0050] in, Spatial interval Each node, frame ,aisle The spatial correlation coefficient at the location is dimensionless, with a value range of [-1, 1]. For the characteristic tensor, the first Each node, frame ,aisle Eigenvalues ​​at; The length of the sliding window is 5 nodes in this embodiment; This represents the mean value within the window. The node spacing is 1 in this embodiment.

[0051] when The value exceeds the preset threshold When it is determined that the signal components in the corresponding time frame and feature channel mainly originate from spatially correlated environmental noise, attenuation processing is applied to these components. Preferably, Take 0.5. Attenuation coefficient. Determined by the following adaptive formula:

[0052] ,

[0053] in, For frames ,aisle The attenuation coefficient at the point has a value range of (0, 1] and is dimensionless. The attenuation sensitivity parameter has a value range of 1.0 to 2.0, and is set to 1.5 in this embodiment; The spatial correlation coefficient threshold is set to 0.5 in this embodiment. When the spatial correlation coefficient does not exceed the threshold, the signal remains unchanged; when it exceeds the threshold, the larger the excess, the stronger the attenuation, thus achieving soft suppression of environmental noise. The filtered feature tensor is... Its elements .

[0054] The technical effect of this step is that environmental noise (mainly the continuous mechanical vibration noise generated by the rotation of idlers and the operation of belts along the entire route, as well as material impact noise and relatively weak natural environmental interference such as wind and rain) usually has a large spatial correlation range, meaning that multiple adjacent sensitizing nodes will simultaneously perceive similar noise levels. However, the fault sound pattern of a single idler has strong spatial locality, only showing significant effects on the 1 to 2 sensitizing nodes closest to the idler. Therefore, the spatial correlation coefficient can effectively distinguish between the two types of signal components. Preferably, in another embodiment of the invention, to further improve the filtering effect, a multi-scale spatial correlation analysis strategy can be adopted, that is, simultaneously calculating the spacing... , and The spatial correlation coefficients were calculated, and the weighted average of the three factors was taken as the final environmental noise index, with weights set to 0.5, 0.3, and 0.2, respectively. This multi-scale strategy can take into account both short-range and long-range spatial correlation characteristics, and is particularly effective for complex working conditions where both local impact noise and continuous mechanical vibration noise coexist along the entire route. Furthermore, in specific engineering implementations, the attenuation sensitivity parameter... The optimal value depends on the specific operating conditions. For high-intensity mechanical noise scenarios with high-speed belt operation, densely distributed idlers, and windy / rainy weather, a larger value, such as 1.8 to 2.0, is recommended to enhance noise suppression. For relatively quiet scenarios with low-speed belt operation and large idler spacing, a smaller value, such as 1.0 to 1.2, is recommended to avoid excessive attenuation leading to loss of effective signal. The feature tensor after noise-perceptual filtering... As input for step S5.

[0055] Step S5: Identification of idler roller operating status based on a multi-scale attention fusion network. (Refer to...) Figure 4 After obtaining the feature tensor after interference suppression Then, for each enhanced node, its corresponding two-dimensional feature map (dimension 1) is generated. The input is a multi-scale attention fusion network for identifying the running status of idler rollers. The overall architecture of this network consists of three cascaded modules: a multi-scale feature extraction module, a channel attention fusion module, and a classification decision module.

[0056] The multi-scale feature extraction module employs a parallel three-branch convolutional structure. Specifically, each of the three branches uses a convolutional kernel with a size of [missing information]. , and The two-dimensional convolutional layer has 64 output channels in each branch, followed by a batch normalization layer and a ReLU activation function. Convolutional branches are used to capture point-by-point nonlinear combination relationships between channels. The convolutional branch is used to extract local patterns within a moderate receptive field. The convolutional branch focuses on macroscopic structural features within a larger receptive field. The design of the three branches enables the network to simultaneously perceive acoustic patterns at different time-frequency scales, which is particularly important for distinguishing wear-related faults (mainly manifested as a slow increase in energy in a specific frequency band) from bearing pitting faults (mainly manifested as periodic impact pulses).

[0057] The channel attention fusion module is responsible for adaptively determining the fusion weights of the outputs of each branch. Preferably, the outputs of the three branches are concatenated along the channel dimension to obtain a 192-channel intermediate feature map. Then, global average pooling is performed to compress the spatial dimension to 1×1, resulting in a 192-dimensional channel description vector. This vector is then passed through two fully connected layers. The first layer compresses the dimension to 48 (compression ratio of 4) and uses ReLU activation; the second layer restores the dimension to 192 and uses Sigmoid activation, outputting the channel weight vector. ,in Each element takes a value in the range (0, 1). The channel weight vector is multiplied channel by channel of the 192-channel intermediate feature map to obtain the weighted fused feature map. The key technical effect of this attention mechanism is that, under different operating conditions and noise conditions, the contribution of each branch to the effective voiceprint information is different. The channel attention can automatically adjust the importance weight of each branch according to the actual situation of the input data, thereby avoiding the performance fluctuation problem of traditional fixed weight fusion schemes when switching operating conditions.

[0058] The classification decision module includes a global average pooling layer, a fully connected hidden layer with 128 neurons (using ReLU activation and regularization with a Dropout rate of 0.5), and an output dimension of... A fully connected classification layer. In this embodiment... These correspond to four idler roller status categories: normal operation, drum wear, bearing malfunction, and bending deformation. The classification layer uses the Softmax activation function, outputting the posterior probability of each category as the confidence level. The entire network is trained using the cross-entropy loss function, with the Adam optimizer selected. The initial learning rate is 0.001, the batch size is 32, and the training epochs are 100. Early stopping is performed when the validation set loss does not decrease for 10 consecutive epochs.

[0059] Preferably, during model training, to address the class imbalance problem where the number of fault samples is far less than that of normal samples, this invention employs the following data augmentation and sampling balancing strategies. First, temporal data augmentation is performed on minority fault samples. This includes superimposing Gaussian white noise with an amplitude of 0.01 to 0.05 times the original signal's root mean square value onto the original acoustic vibration waveform to simulate different noise levels; performing random stretching of the time axis by 0.9 to 1.1 times to simulate fine-tuning of rotational speed; and performing random masking of the spectrum, i.e., randomly masking 1 to 3 consecutive frequency bands on the Mel spectrogram to enhance the model's robustness to missing local features. After augmentation, the number of training samples for each class is maintained in a balanced ratio of 1:1:1:1. Second, a class weight factor is introduced into the loss function, with the normal class weight set to 1.0 and the weights for all three fault classes set to 2.0, allowing the model to impose a greater penalty on misclassification of fault samples during training. The training data originates from labeled samples accumulated during the initial deployment phase. The labeling method involves manually inspecting and confirming the roller status category of the corresponding sensitive node in the system. The training set and validation set are randomly divided in an 8:2 ratio. During the division, it is ensured that samples from different time windows of the same enhancement node will not appear in the training set and validation set at the same time to avoid data leakage.

[0060] The output of step S5 is the state classification result (one of four categories) of the idler roller corresponding to each sensitive node, along with the corresponding confidence value (the maximum posterior probability output by Softmax). In this embodiment, each analysis window (1 second duration, 0.5 second increment) generates 170 state classification results and confidence values ​​for sensitive nodes, resulting in 2 sets of 340 state records per second. These records are encapsulated into structured state data frames, containing fields such as timestamp, sensitive node number, spatial coordinates, classification result encoding, confidence vectors for each category, and estimated current environmental noise level. These frames are transmitted in real-time to the backend platform database for persistent storage via TCP protocol. Simultaneously, to support remote inspection, the system caches the acoustic vibration waveforms of each sensitive node within the last 5 minutes in memory. When maintenance personnel click on any sensitive node on the platform, they can replay the sound reconstruction results at that node with a single click, enabling remote auscultation based on sound reconstruction. This output serves as the input for step S6.

[0061] Step S6: Closed-loop early warning decision-making and online model iteration. (Refer to...) Figure 5 After obtaining the state classification results and confidence levels of each enhanced node, this step performs hierarchical early warning decision-making and constructs a dual-channel closed-loop feedback from the early warning end to the data acquisition end and the model end.

[0062] In terms of early warning decision-making, this invention sets four status levels: normal, attention, early warning, and alarm. The specific judgment rules are as follows: when the classification result of the idler corresponding to a certain sensitivity enhancement node is normal operation and the confidence level is higher than 0.8, the status level is judged as normal; when the classification result is an abnormal category (any of wear, bearing abnormality, or deformation) but the confidence level is lower than 0.6, the status level is judged as attention, and the system only logs in the background without triggering an alarm; when the abnormal confidence level continuously exceeds the first threshold... And the duration exceeds the first time window When the status level is upgraded to an alert, the system pushes a yellow alert to the operations and maintenance platform and activates nearby video equipment to capture images; when the anomaly confidence level exceeds the second threshold... When the alarm is triggered, regardless of the duration, the system immediately sends a red alarm message and notifies the inspection personnel via SMS. Preferably, a sliding time window smoothing mechanism is introduced into the judgment logic to perform a moving average of 10 sampling periods on the confidence sequence to suppress false jumps caused by instantaneous noise pulses.

[0063] Regarding parameter feedback, when the system detects consecutive occurrences of events of interest or warnings in a certain section, it automatically increases the pulse code length of that section from 128 bits to 256 bits, corresponding to a further improvement in the signal-to-noise ratio of approximately 3 dB, thereby acquiring the acoustic signals of the idler rollers in that section with higher sensitivity. Simultaneously, the channel polling priority for that section can be optionally increased, and the polling interval shortened to obtain higher temporal resolution. Once the idler rollers in that section have been repaired and returned to normal operation, the code length and polling priority are restored to their default values. This adaptive acquisition strategy allows the system to dynamically focus monitoring accuracy on areas requiring special attention while maintaining overall resource availability, demonstrating the collaborative optimization between the acquisition and processing ends.

[0064] Regarding online model updates, this invention employs an incremental learning strategy based on experience replay. The system maintains a fixed capacity... A sample replay buffer pool stores historical labeled samples proportionally for each category. Whenever operations personnel manually confirm or correct the labeling of an alert event during the alert handling process, a new labeled sample is added to the buffer pool (if the pool is full, the earliest added sample of the same category is randomly replaced). The online update cycle is set to trigger a fine-tuning every 50 newly added labeled samples. During fine-tuning, 200 historical samples are randomly sampled from the buffer pool and mixed with 50 new samples to form a training batch of 250 samples. Only the parameters of the fully connected layer of the classification decision module are updated, and the learning rate is set to 0.1 times the initial learning rate, i.e., 0.0001. The fine-tuning rounds are 5 epochs.

[0065] The technical advantages of online updates are as follows: the experience replay mechanism effectively alleviates the catastrophic forgetting problem in incremental learning by mixing historical and new samples; the strategy of fine-tuning only the classification layer while freezing the feature extraction layer reduces the risk of overfitting while maintaining stable feature representation capabilities; and the low learning rate and few rounds of fine-tuning ensure the smoothness of model updates. Preferably, in terms of the sample management strategy of the buffer pool, this invention adopts a class balance replacement mechanism, that is, when a new sample enters the pool, only the oldest sample of the same class as the new sample is replaced, ensuring that the number of samples of each class in the buffer pool is always balanced. When the proportion of samples of a certain class in the buffer pool is less than 15% of the total capacity, the system automatically reduces the probability of replacing old samples of that class to maintain class diversity. In addition, to prevent abnormal degradation of model performance caused by online updates, the accuracy of the updated model is evaluated on an independent validation set after each fine-tuning. If the accuracy drops by more than 2 percentage points compared to before the update, the model parameters before the update are rolled back and an anomaly log is recorded. The decision on whether to accept the update is made after manual confirmation. This safety rollback mechanism provides an additional layer of quality assurance for online updates, preventing model quality degradation due to labeling errors or abnormal data. In actual deployment verification, after enabling the online update mechanism, the system's recognition accuracy only decreased from the initial 96.2% to 93.8% within 6 months of continuous operation, while the control group without the update mechanism decreased from 96.2% to 87.5% within the same period, verifying the long-term effectiveness of the closed-loop update strategy.

[0066] To verify the effectiveness of the method of this invention, a six-month actual deployment test was conducted on a 1.2km long main conveyor belt in a coal mine. The test environment included various typical scenarios such as normal operation, heavy-load impact, rainy and humid weather, and fluctuating dust concentration. During the test, the conveyor belt operated 24 hours a day without interruption, with a belt speed of 3.15m / s and a full load capacity of approximately 800t / h. A total of 170 sensitive nodes were installed along the line, covering approximately 1190 idler sets.

[0067] In terms of signal acquisition performance, PDC encoding enhancement improves the system's signal-to-noise ratio (SNR) by approximately 11.3 dB compared to the non-encoded baseline scheme. Combined with the enhanced head structure, the SNR of acoustic signal acquisition within a 3-meter range of the idler roller is better than 15 dB. In the control section without the enhanced head, the SNR at the same distance is only 3 dB to 5 dB, which is insufficient to support reliable extraction of subsequent acoustic features. Preferably, regarding the selection of encoding length, three encoding lengths—64-bit, 128-bit, and 256-bit—were tested, with corresponding SNR improvements of 8.0 dB, 11.3 dB, and 14.1 dB, respectively. However, the larger the encoding length, the longer the time window, and the lower the pulse repetition frequency. Therefore, considering both SNR and time resolution, this embodiment preferably uses 128-bit as the default encoding length, dynamically switching to 256-bit only in key areas of interest.

[0068] Regarding the multimodal feature fusion performance, the single-mode scheme using only Mel frequency cepstral coefficients, the dual-mode scheme using Mel frequency cepstral coefficients plus contrast spectrum, and the trimodal scheme of this invention were compared on the same test set, with recognition accuracies of 82.1%, 89.7%, and 96.2%, respectively. The introduction of higher-order spectral features improved the recall rate for early bearing pitting faults from 78.3% to 91.6%. This is because the periodic impact signal generated by bearing pitting has significant nonlinear phase coupling characteristics, which higher-order spectral analysis can effectively separate from Gaussian background noise. Furthermore, in the comparison of classification performance for various fault categories, the precision and recall rates of this invention for cylinder wear faults were 94.8% and 93.2%, respectively; for bearing anomaly faults, 97.1% and 91.6%, respectively; for bending deformation faults, 93.5% and 89.7%, respectively; and for normal operation faults, 96.8% and 98.5%, respectively. The results show that the three-modal fusion strategy has a relatively balanced ability to identify various types of faults and does not have a serious bias towards specific fault types.

[0069] Regarding noise immunity, the recognition accuracy was tested under signal-to-noise ratios (SNRs) of 10 dB, 5 dB, 0 dB, and -5 dB. The accuracy rates of the proposed method were 97.8%, 95.6%, 91.2%, and 85.4%, respectively, while the accuracy rates of the control scheme without noise-perceived filtering under the same conditions were 96.1%, 90.3%, 79.5%, and 62.8%, respectively. Under strong noise conditions of 0 dB and -5 dB, the accuracy gains from noise-perceived filtering were 11.7% and 22.6%, respectively, verifying the significant advantage of the spatial correlation filtering strategy under harsh conditions. Notably, even under extreme noise conditions of -5 dB, the proposed method still maintains a recognition accuracy of 85.4%. This is mainly due to the noise-perceived filtering mechanism utilizing the spatial distribution characteristics of the enhanced node array. Even with extremely low SNRs for individual nodes, spatial correlation analysis can still effectively separate localized fault signals from the spatially distributed environmental noise.

[0070] In the ablation experiment of the multi-scale attention fusion network, the complete network was compared with a control network after removing the channel attention module. The results showed that the recognition accuracy decreased from 96.2% to 92.7% after removing the attention mechanism, a decrease of 3.5 percentage points. Further analysis of the distribution of channel attention weights revealed that under normal operating conditions, the network tends to give higher weights to the Mel-frequency cepstral coefficient channel, while under strong noise conditions, the network automatically increases the weights of higher-order spectral feature channels to take advantage of their natural resistance to Gaussian noise. This adaptive adjustment behavior verifies the necessity and effectiveness of the channel attention mechanism.

[0071] Regarding long-term operational stability, the system with the closed-loop online update mechanism achieved monthly average recognition accuracy rates of 96.2%, 95.8%, 95.1%, 94.7%, 94.2%, and 93.8% over a six-month operating period, with a cumulative decrease of 2.4 percentage points. In contrast, the control system without the update mechanism achieved monthly average accuracy rates of 96.2%, 94.5%, 92.1%, 90.6%, 89.0%, and 87.5% during the same period, with a cumulative decrease of 8.7 percentage points. The closed-loop update mechanism reduced the accuracy decay rate by approximately 72%, effectively ensuring the long-term reliability of the system. The main reasons for performance degradation included changes in acoustic characteristic distribution after roller replacement, alterations in acoustic propagation characteristics due to seasonal temperature and humidity changes, and long-term trends in belt load. The online update mechanism, by continuously incorporating labeled samples under new operating conditions, allows the model to gradually adapt to these slowly changing factors, while the experience replay strategy prevents the model from forgetting existing knowledge during the adaptation process.

[0072] In terms of actual early warning effectiveness, the system detected 23 abnormal events on idlers during the 6-month operation period. Of these, 21 were confirmed as genuine faults by manual inspection, with 1 missed and 2 false alarms, resulting in an overall detection rate of 91.3% and a false alarm rate of 8.7%. The average response time from the first abnormal sound signature to the system triggering an early warning was 47 seconds, approximately 72 hours earlier than the average detection time by manual inspection, providing ample time for advance scheduling of maintenance plans. The missed case involved minor wear on the end idler roller located approximately 5 meters away from the nearest sensing node. This was due to insufficient signal acquisition caused by the idler roller being more than the detection radius away from the nearest sensing node, a problem that was subsequently resolved by optimizing the spacing of the sensing nodes. Both false alarms occurred during heavy rain, where transient impact noise from raindrops hitting the belt conveyor was mistakenly identified as bearing abnormalities. This was corrected by adding heavy rain condition samples to the training set.

[0073] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for early warning and data processing of belt conveyor status based on DAS, characterized in that, Includes the following steps: A narrow-linewidth coherent laser pulse sequence with pulse digital encoding modulation is injected into the sensing optical fiber deployed along the belt conveyor. Backscattered interference light signals generated by Rayleigh scattering at each sensitization node are received. The spatial position of each sensitization node is determined based on the round-trip time delay of the light pulse, and the distributed acoustic raw signal is obtained. Matched decoding and phase demodulation are performed on the backscattered interference light signals at each sensitization node to extract the phase change of the Rayleigh scattered light. The acoustic vibration waveforms at each sensitization node are recovered based on the phase change, and a spatiotemporal sound field matrix is ​​constructed with the spatial position of the sensitization node as the row index and the sampling time as the column index. Multimodal acoustic features are extracted node by node from the spatiotemporal sound field matrix, including Mel frequency cepstral coefficient features, contrast spectrum features and higher-order spectrum features. The three types of features are spliced ​​along the channel dimension to form a multimodal feature tensor. A noise-sensing dynamic filtering mechanism is used to adaptively suppress environmental interference in the multimodal feature tensor. The filtering parameters are dynamically adjusted according to the estimated environmental noise level to separate the effective acoustic components related to idler operation. The feature tensors after interference suppression are input into a multi-scale attention fusion network to perform idler roller running status recognition, and output the state classification results and confidence scores of the idler rollers corresponding to each enhanced node. Based on the state classification results and confidence levels, an early warning decision is generated, and the identification results are fed back to the signal acquisition stage to adjust the pulse coding parameters and sampling strategy. At the same time, the multi-scale attention fusion network is incrementally updated using online labeled samples.

2. The method according to claim 1, characterized in that, The pulse digital code modulation adopts Golay complementary sequence coding, with a coding length of 64 bits to 256 bits, a single pulse width of 20 ns to 100 ns, and a pulse repetition frequency of 500 Hz to 2000 Hz; the sensitivity enhancement nodes are arranged along the steel frame of the belt conveyor at intervals of 5 m to 10 m, and the optical fiber in each sensitivity enhancement node is wound 3 to 8 times.

3. The method according to claim 1, characterized in that, The phase demodulation adopts a 3×3 coupler demodulation scheme, and the matched decoding includes performing cross-correlation operation on the received backscattered interference optical signal and the transmitted coded sequence to improve the signal-to-noise ratio. The improvement in signal-to-noise ratio is proportional to the square root of the coding length.

4. The method according to claim 1, characterized in that, In the noise perception dynamic filtering mechanism, the estimated environmental noise level is obtained by calculating the spatial correlation coefficient between adjacent sensitized nodes in the spatiotemporal sound field matrix. When the spatial correlation coefficient is lower than a preset threshold, the corresponding frequency band is determined to be the dominant frequency band of environmental noise. The preset threshold ranges from 0.3 to 0.

7.

5. The method according to claim 1, characterized in that, The extraction process of the Mel frequency cepstral coefficient features includes: performing pre-emphasis processing on the acoustic vibration waveform, with a pre-emphasis coefficient of 0.95 to 0.98; performing frame segmentation and windowing with a frame length of 20ms to 40ms and a frame shift of 8ms to 15ms; filtering through a Mel filter bank containing 20 to 40 triangular filters after fast Fourier transform; performing discrete cosine transform after taking the logarithm, and retaining the first 12 to 20 dimensions of coefficients as the Mel frequency cepstral coefficient features.

6. The method according to claim 1, characterized in that, The higher-order spectral features are obtained through bispectral analysis, specifically including: calculating the third-order cumulant of the short-time frame signal of the acoustic vibration waveform, then obtaining the bispectral matrix through two-dimensional discrete Fourier transform, and extracting the logarithmic mean of the bispectral amplitude and the weighted variance of the bispectral phase from the bispectral matrix as higher-order spectral feature vectors.

7. The method according to claim 1, characterized in that, The multi-scale attention fusion network includes a multi-scale feature extraction module, a channel attention module, and a classification decision module. The multi-scale feature extraction module adopts a parallel multi-branch convolutional structure with branch convolutional kernel sizes of 1×1, 3×3, and 5×5, respectively. The channel attention module performs global average pooling on the output of each branch and then generates a channel weight vector through two fully connected layers to perform weighted fusion of the features of each branch. The classification decision module includes a global average pooling layer and a fully connected classification layer.

8. The method according to claim 1, characterized in that, During the construction of the spatiotemporal sound field matrix, bandpass filtering is performed on the recovered acoustic vibration waveform, with a passband frequency range of 10Hz to 8000Hz; normalization processing is performed on the bandpass filtered signal, and the normalization method is to perform zero-mean unit variance standardization according to the sensitization node.

9. The method according to claim 1, characterized in that, The generation of the early warning decision includes: setting four status levels: normal, attention, early warning, and alarm. When the abnormal confidence level of the idler corresponding to a single enhanced node continuously exceeds the first threshold and the duration exceeds the first time window, an early warning is triggered. When the abnormal confidence level exceeds the second threshold, an alarm is triggered. The first threshold is 0.6 to 0.75, the second threshold is 0.85 to 0.95, and the first time window is 30s to 120s.

10. The method according to claim 1, characterized in that, The incremental update adopts an experience replay strategy, maintaining a fixed-capacity sample buffer pool. Whenever a new online labeled sample is obtained, historical samples are randomly sampled from the sample buffer pool and mixed with the new samples to form a training batch. The classification decision module of the multi-scale attention fusion network is fine-tuned by decaying the learning rate.

Images