Intelligent monitoring system for coal conveying system of thermal power plant based on multi-source perception and edge computing

Abstract

The application belongs to the technical field of industrial control, and discloses a coal conveying system intelligent monitoring system for thermal power plants based on multi-source sensing and edge computing; the system comprises the following steps: collecting multi-modal sensing data through a heterogeneous sensor array, using a physical constraint correlation model to perform deep collaborative extraction on cross-modal features such as vision, acoustics, vibration, temperature and current, and dynamically identifying equipment abnormal patterns accordingly; and introducing a hierarchical fusion analysis architecture to provide multi-level confidence guarantee for fault diagnosis. Through a simulation deduction mechanism driven by digital twinning, the system response can be pre-evaluated before the execution of the control strategy, the optimal solution with the best comprehensive performance is selected, and the reliability of the control decision is ensured; the application also realizes an incremental correction mechanism for real-time feedback of state deviation, avoids control deviation caused by model mismatch, and enables each abnormal symptom of the coal conveying link to be accurately captured and adaptively handled, thereby significantly improving the system safety and the intelligent level of operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and more specifically, to an intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing. Background Technology

[0002] With the deepening development of intelligent manufacturing and the Industrial Internet, thermal power plants, as critical infrastructure, are undergoing a transformation and upgrade from traditional manual operation and maintenance to intelligent monitoring. The coal conveying system, as the energy artery of a thermal power plant, integrates complex industrial scenarios involving mechanical transmission, material handling, and multi-equipment collaboration. Its safe and stable operation directly affects power generation efficiency and personnel safety. Traditional coal conveying systems rely on manual inspections and experience-based judgment, facing numerous technical bottlenecks in complex fault mode identification and preventative maintenance. In recent years, multi-source sensing technology and edge computing have achieved significant breakthroughs in the field of industrial monitoring, providing technical support for the intelligent transformation of coal conveying systems.

[0003] However, existing multi-source sensing and monitoring technologies face several challenges when applied to coal conveying systems in thermal power plants. Particularly, heterogeneous data fusion encounters feature conflicts and decision-making contradictions when dealing with the complex coupling relationships of multimodal sensor signals. As monitoring coverage expands, data collected by different modal sensors, such as visual, acoustic, vibration, temperature, and current sensors, exhibits heterogeneity and temporal misalignment, leading to a rapid expansion of the fault feature space and semantic gaps between modalities. Existing solutions, such as weighted voting fusion or centralized cloud processing, not only fail to resolve the feature semantic alignment problem but also introduce additional network latency and bandwidth consumption, making it difficult to meet the real-time requirements for rapid fault response.

[0004] In view of this, the present invention proposes an intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution:

[0006] An intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing includes:

[0007] The multi-source data acquisition module is used to acquire multi-modal data of the coal conveying link based on the deployed heterogeneous sensor array to obtain multi-source sensing data; the multi-source sensing data includes visual image sequences, equipment acoustic signals, temperature distribution data, vibration time series data, and current waveform data;

[0008] The data preprocessing module is used to perform environmental interference compensation on the multi-source sensing data to obtain multi-source purification data, and to perform deep collaborative extraction of cross-modal features on the multi-source purification data to obtain a collaborative feature set.

[0009] The edge decision module is used to build a hierarchical fusion analysis architecture and combine it with collaborative feature sets to identify and deeply analyze abnormal events, thereby obtaining the global operating status of the coal conveying link.

[0010] The intelligent control and feedback module is used to perform progressive diagnosis and propagation path deduction of the current coal conveying link operation status based on the global operation status, and generate an adaptive control strategy for the coal conveying link through simulation deduction; the adaptive control strategy is converted into control commands and sent to the control mechanism of the coal conveying link.

[0011] Furthermore, the process of acquiring multi-source purification data includes:

[0012] Based on the joint analysis of the dark channel prior distribution and spatial frequency attenuation of each visual image in the visual image sequence, the equivalent transmittance of dust concentration in the current environment is obtained; based on the equivalent transmittance of dust concentration, the visual image is decomposed into a multi-scale feature decomposition, dividing the visual image into a dust-free intrinsic layer and a dust scattering layer, and the original visual image is processed based on the dust-free intrinsic layer to obtain the purified visual image.

[0013] The environmental noise baseline of the collected equipment acoustic signal is dynamically tracked in real time to identify steady-state noise components and construct a time-varying parameter model based on it. An adaptive frequency domain notch filter bank is designed based on the obtained time-varying parameter model and used to suppress noise in the equipment acoustic signal to obtain the purified equipment acoustic signal.

[0014] Electromagnetic interference features are identified and extracted from the vibration time series data and current waveform data to distinguish between mechanical impact signals and electromagnetic pulse interference signals. Based on a pre-established interference feature library, template matching and adaptive cancellation are performed on the electromagnetic pulse interference signals to obtain purified vibration time series data and current waveform data.

[0015] Non-uniformity correction and temperature calibration compensation are performed on the temperature distribution data to obtain the purified temperature distribution data; the various data after purification are summarized to obtain the corresponding multi-source purification data.

[0016] Furthermore, the process of obtaining the collaborative feature set includes:

[0017] Spatiotemporal segmentation and multi-scale decomposition of moving targets within the purified visual image are performed to identify and reconstruct the complete spatial morphological features of the moving targets.

[0018] Based on wavelet packet decomposition and short-time Fourier transform, joint time-frequency analysis of the purified equipment acoustic signal is performed to identify and extract the fundamental frequency component and fault modulation component of the rotating parts of the equipment.

[0019] Spatial evolution analysis of thermal anomaly regions is performed on the purified temperature distribution data, and the location and intensity of heat sources are inverted to generate thermal anomaly propagation characteristics.

[0020] Empirical mode decomposition was performed on the purified vibration time series data, and the distribution characteristics of vibration energy in the time-frequency plane in the vibration time series data were determined based on it.

[0021] Load fluctuation pattern identification is performed on the purified current waveform data, and load fluctuation characteristics are extracted by combining the load power spectral density of the current waveform data.

[0022] Based on the extracted cross-modal features, a mutual information matrix is ​​constructed, a physical constraint association model is built, and a feature projection transformation is performed on the mutual information matrix based on it to obtain a collaborative feature set. The cross-modal features include complete spatial morphological features, fundamental frequency components and fault modulation components, thermal anomaly propagation features, load fluctuation features, and vibration energy distribution features in the time-frequency plane.

[0023] Furthermore, the implementation process of the edge decision-making module includes:

[0024] A hierarchical fusion analysis architecture is deployed in the edge computing nodes of the corresponding coal conveying link. The hierarchical fusion analysis architecture includes an edge fast response layer, an edge deep analysis layer, and a cloud collaborative verification layer.

[0025] The edge fast response layer is used to construct a baseline feature space under normal operation based on cross-modal features, obtain the Mahalanobis distance between it and the corresponding collaborative features, and perform fast discrimination based on the Mahalanobis distance to obtain abnormal event logs;

[0026] A refined fault diagnosis model is deployed in the edge deep analysis layer, and the collaborative feature set corresponding to the occurrence time in the abnormal event log is input into the refined fault diagnosis model to obtain the output results;

[0027] The decision evaluation is performed based on the confidence level of the output result. If the confidence level is less than the preset second threshold, the corresponding output result is marked as a low-confidence decision and the cloud collaborative verification layer is triggered. The cloud collaborative verification layer performs correlation analysis of global historical data and expert knowledge base verification, and updates the low-confidence decision based on the verification results.

[0028] The intermediate and final results of the above analysis are compressed and feature extracted to obtain the global running status.

[0029] Furthermore, the implementation process of the intelligent control and feedback module includes:

[0030] Construct a multi-level fault discrimination decision tree for the coal conveying link; based on the various discrimination nodes in the multi-level fault discrimination decision tree, and combined with the collaborative feature set and global operating status, perform fault combination discrimination on the coal conveying link to obtain the output results of the multi-level fault discrimination decision tree;

[0031] Uncertainty quantification is performed on the output results of the multi-level fault discrimination decision tree to generate the probability distribution of decision confidence and calculate the variance of decision confidence;

[0032] When the variance of the decision confidence level is greater than the preset confidence threshold, the corresponding output result is labeled with "manual review" and transmitted to the relevant management personnel for manual review; when the variance of the decision confidence level is not greater than the preset confidence threshold, an intelligent monitoring report is generated based on the discrimination results of the corresponding multi-level fault discrimination decision tree. The intelligent monitoring report includes the fault type, fault location, fault severity, fault propagation path, corresponding risk probability, and decision confidence level.

[0033] An adaptive control strategy is generated based on the severity and type of the fault. The generated adaptive control strategy is then encapsulated into standardized control commands and sent to the control mechanism of the coal conveying link.

[0034] Furthermore, the root node of the multi-level fault discrimination decision tree is a binary classification of normal and abnormal, and the subsequent hierarchical nodes are three types of decision nodes: fault category discrimination node, fault location node, and fault severity assessment node. Each decision node makes a discrimination based on different combinations of collaborative feature sets.

[0035] Furthermore, the process of fault combination discrimination includes:

[0036] Based on the fault category discrimination nodes in the multi-level fault discrimination decision tree, modal decomposition is performed on the collaborative feature set to identify and label the modal sources that dominate the fault features, and a modal combination strategy library is established for fault types to determine specific fault types.

[0037] The fault location node in the multi-level fault discrimination decision tree determines the spatial location of the fault source based on the spatial topology of the abnormal sensor array and the propagation delay analysis of fault characteristics.

[0038] The fault severity assessment node in the multi-level fault discrimination decision tree performs fault degradation curve fitting by tracking the fault evolution trend of fault features on the time axis, and predicts the remaining time for the fault features to reach the danger threshold.

[0039] Based on the predicted remaining time and fault location, a fault propagation path simulation analysis is performed, the risk probability of each fault propagation path is calculated, and the output results of a multi-level fault discrimination decision tree are generated.

[0040] Furthermore, the process of generating the adaptive control strategy includes:

[0041] Based on the fault type and severity, a set of candidate control strategies is obtained by matching strategies from a preset control strategy library.

[0042] For each candidate control strategy in the candidate control strategy set, forward simulation is performed using a digital twin model of the coal conveying link to predict the system state evolution trajectory after the execution of the candidate control strategy.

[0043] Based on the system state evolution trajectory, each candidate control strategy is scored to obtain a comprehensive score for each candidate control strategy, and the candidate control strategy with the highest comprehensive score is used as the initial control adjustment strategy.

[0044] The parameters of the initial control adjustment strategy are finely adjusted, and the adjusted initial control adjustment strategy is used as the adaptive control strategy.

[0045] Furthermore, the process of acquiring the purified visual image includes:

[0046] For the dust-free intrinsic layer, a local window is used to traverse the pixels to obtain the local contrast within each local window, and then mapped to an intuitive contrast distribution heatmap.

[0047] Based on the contrast distribution heatmap, low-contrast, medium-contrast, and high-contrast regions within the original visual image are identified.

[0048] Differential enhancement processing is performed on each contrast region. After the enhancement processing is completed, the processed contrast regions are smoothly blended to obtain a purified visual image.

[0049] Furthermore, the design process of the adaptive frequency domain notch filter bank includes:

[0050] A short-time Fourier transform is performed on the current analysis frame of the device's acoustic signal to obtain a time-frequency plane representation. The spectral ridges of environmental noise are identified on the time-frequency plane, and the local maxima are located as the noise center frequencies, while the corresponding time-series evolution sequences are obtained simultaneously. The noise center frequencies at future moments are predicted based on the time-series evolution of the noise center frequencies. Notch filters are constructed at each predicted noise center frequency, and all notch filters are cascaded to form an adaptive frequency-domain notch filter bank.

[0051] The technical effects and advantages of this invention, which is based on multi-source sensing and edge computing, for intelligent monitoring of coal conveying systems in thermal power plants:

[0052] This invention scientifically integrates multi-source heterogeneous sensor data to ensure accurate identification and source location of each type of fault symptom, enhancing the system's safety early warning capabilities and fault handling accuracy. In practical applications, even under complex and changing operating conditions, it maintains high confidence in diagnostic results and consistency in control strategies, eliminating the need for frequent manual intervention and making fault handling processes more automated and intelligent. Furthermore, the adaptive control correction mechanism of this invention can adapt to various equipment aging states and fluctuating operating conditions, lowering the threshold for system deployment and maintenance, and enabling intelligent monitoring technology to be extended to small and medium-sized thermal power plants and other bulk material conveying applications. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing, according to the present invention.

[0054] Figure 2 This is a schematic diagram of the fault degradation curve in an embodiment of the present invention. Detailed Implementation

[0055] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] Please see Figure 1 As shown in this embodiment, the intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing includes:

[0058] The multi-source data acquisition module is used to acquire multimodal data from the coal conveying link through a deployed heterogeneous sensor array, obtaining multi-source sensing data. This multi-source sensing data includes key information such as high-resolution visual image sequences, equipment acoustic signals, temperature distribution data, vibration time-series data, and current waveform data, acquired in real time through the heterogeneous sensor array. The visual image sequences record the surface condition of the coal conveyor belt, coal distribution, and equipment operating status; the equipment acoustic signals reflect the operating status and abnormal signs of the mechanical equipment; the temperature distribution data monitors the thermal state of the equipment and potential fault points; the vibration time-series data captures the mechanical vibration characteristics of the equipment; and the current waveform data reflects changes in motor load and electrical faults. This data provides comprehensive raw information for subsequent intelligent analysis, ensuring the all-round sensing capability of the monitoring system. The heterogeneous sensor array includes protective visual sensors for dust environments, broadband acoustic sensors, triaxial vibration sensors, infrared thermal imaging sensors, and non-contact current sensors.

[0059] The data preprocessing module is used to compensate for environmental interference in multi-source sensing data, obtaining multi-source purification data, and then performing deep collaborative extraction of cross-modal features from the multi-source purification data to obtain a collaborative feature set. Environmental interference compensation eliminates interference factors in the special environment of thermal power plants, including dust scattering, industrial noise, and electromagnetic interference, through advanced signal processing algorithms. Collaborative feature extraction establishes a physical constraint correlation model for cross-modal features, deeply fusing information from different sensors to form a comprehensive feature representation with complementarity and robustness.

[0060] The edge decision-making module constructs a layered fusion analysis architecture, combining collaborative feature sets for anomaly event identification and in-depth analysis to obtain the global operational status of the coal conveying chain. The layered fusion analysis architecture includes an edge rapid response layer, an edge deep analysis layer, and a cloud-based collaborative verification layer. Through a multi-layered decision-making mechanism, it achieves a balance between rapid response and accurate analysis, ensuring that the system meets real-time requirements while providing reliable analysis results.

[0061] The intelligent control and feedback module is used to perform progressive diagnosis and propagation path deduction of the current coal conveying link's operating status based on the global operating status, and to generate an adaptive control strategy for the coal conveying link through simulation. This adaptive control strategy is then converted into control commands and sent to the control mechanisms of the coal conveying link. The intelligent control and feedback module achieves accurate fault classification, location, and assessment through a multi-level fault discrimination decision tree, and generates the optimal control strategy by combining it with a digital twin model. Control commands are then sent to the field actuators, forming a closed-loop intelligent monitoring system.

[0062] The various modules mentioned above are connected via industrial Ethernet and / or fieldbus to enable data transmission and collaborative operation between the modules.

[0063] It should be noted that during the deployment of heterogeneous sensor arrays, the physical characteristics of typical fault modes in the coal conveying link are analyzed to establish a matching relationship between fault characteristics and sensor types. For example, for belt misalignment faults, laser displacement sensors are preferred to monitor edge offset; for material blockage and coal accumulation faults, material level sensors and infrared thermal imaging sensors are preferred to monitor coal flow accumulation; and for belt tear faults, acceleration sensors are preferred to capture high-frequency impact signals. Simultaneously, based on a comprehensive evaluation of sensor cost-effectiveness, environmental adaptability, and signal anti-interference capability, the most cost-effective sensor combination is selected. Combined with the digital twin model of the coal conveying link, the optimal installation location and quantity configuration of sensors are determined. Furthermore, a graded monitoring strategy is established according to the importance level of the equipment. Multi-sensor redundant coverage is implemented for key equipment, key parameter monitoring is implemented for important equipment, and sampling monitoring is implemented for general equipment. Ultimately, the most critical equipment parts are covered with the fewest number of sensors, ensuring the optimal balance between fault detection effectiveness and economy.

[0064] In embodiments of the present invention, the process of acquiring multi-source purification data includes:

[0065] Based on the joint analysis of the dark channel prior distribution and spatial frequency attenuation of each visual image in the visual image sequence, the equivalent transmittance of dust concentration in the current environment is obtained. The equivalent transmittance of dust concentration can be used to reflect atmospheric visibility and dust scattering, directly affecting image quality and subsequent analysis accuracy. Specifically, an N×N (N is an integer) neighborhood window is defined centered on each pixel in the acquired visual image; the minimum value of the R, G, and B components in the RGB three-color channel values ​​of all pixels in the corresponding neighborhood window is obtained and used as the minimum channel value of the central pixel. After traversing all pixels in the visual image, a local minimum channel map of the same size as the original visual image is obtained; then, the process of obtaining the minimum R, G, and B values ​​is repeated on the local minimum channel map to obtain the dark channel prior map; the larger the pixel value in the corresponding dark channel prior map, the more severe the dust influence in the corresponding area; and the pixel value of the dark channel prior map can intuitively reflect the spatial distribution of dust concentration. At the same time, the visual image is converted into a grayscale image and subjected to two-dimensional discrete Fourier transform processing to obtain the spectrum matrix in the complex domain. ,in and These represent the frequency components in the horizontal and vertical directions, respectively; the zero-frequency component (matrix) within the spectrum matrix. , Points with all values ​​equal to 0 (usually located in the upper left corner of the spectrum matrix) are moved to the center of the spectrum (the geometric center of the spectrum matrix) to complete the centering process, and the amplitude spectrum of the corresponding spectrum matrix is ​​calculated. The amplitude spectrum is used to reflect the energy distribution of a visual image at different frequency components. The Euclidean distance between each frequency point and the center of the spectrum in the centered spectrum matrix is ​​obtained, and regions with Euclidean distances less than a preset distance threshold are defined as low-frequency regions; regions with Euclidean distances not less than a preset distance threshold are defined as high-frequency regions. This facilitates the effective distinction between the overall brightness distribution (low frequency) and edge texture details (high frequency) of the image. Next, the sum of squares of the amplitude spectra corresponding to each frequency point in the low-frequency and high-frequency regions is obtained, and the ratio between the sums of squares is normalized to obtain the spatial frequency attenuation rate. It should be noted that the frequency attenuation rate is calculated as a single scalar value in the global frequency domain. To achieve pixel-level fine estimation, the visual image is divided into overlapping image blocks of fixed pixel size. The local frequency attenuation rate is calculated independently for each image block, and finally, bilinear interpolation is used to reconstruct an attenuation rate distribution map with the same resolution as the original image. The dark channel prior value of each pixel is obtained based on the dark channel prior map, and combined with the local frequency attenuation rate of each pixel, the equivalent transmittance of dust concentration is obtained through weighted fusion; the mathematical formula for calculating the equivalent transmittance of dust concentration is as follows: In the formula, pixel position Dust concentration equivalent transmittance; Represented as normalized dark channel prior values; Indicates the local spatial frequency attenuation rate; and Indicates the weighting coefficient; , representing the dark channel weight factor.

[0066] Multi-scale feature decomposition of visual images is performed based on the equivalent transmittance of dust concentration, dividing the visual image into a dust-free intrinsic layer and a dust scattering layer. Image processing is then performed on the original visual image based on the dust-free intrinsic layer to obtain a purified visual image. A guided filtering pyramid is a multi-scale image decomposition tool that preserves edge characteristics. Each scale uses the original visual image as a guide map to ensure high fidelity in detail areas. Specifically, the top 0.1% of pixels with the highest Y component (i.e., brightness value) in the YUV color channels of the original visual image are selected, and the average value of the corresponding pixels in the RGB color channels is calculated. These values ​​are then combined into a three-dimensional vector as an estimate of atmospheric light. Atmospheric light estimates characterize the ambient light intensity at infinity when dust concentration is saturated, and are key parameters in atmospheric scattering physics models. A degradation model for visual images is established based on existing atmospheric scattering physics models. This degradation model describes the visual image... Dust-free intrinsic layer Dust concentration equivalent transmittance and atmospheric light estimates The physical relationship between them; the mathematical expression of the degradation model is: Next, to accurately separate the dust-free intrinsic layer from the observed image, the dust concentration equivalent transmittance distribution map corresponding to the original image is obtained based on the dust concentration equivalent transmittance T(x, y). Then, an image pyramid is obtained by performing multi-scale downsampling on the original visual image and the dust concentration equivalent transmittance distribution map. Multi-scale downsampling refers to the process of generating an image pyramid structure with proportionally decreasing resolution and decreasing scale by downsampling layer by layer, using the original visual image and the dust concentration equivalent transmittance distribution map as a reference. Then, guided filtering is performed at each level of the image pyramid. Specifically, taking a certain level of the image pyramid as an example, the original visual image is used as the guide image, and the dust concentration equivalent transmittance distribution map of the corresponding level is used as the input image. A fixed-size guided filtering window is constructed within the guide image. The mean value of the pixels within the guide window, the mean value of the pixels at the corresponding coordinate positions in the input image, and the covariance between the guide image and the input image are obtained. Based on these, filtering is performed to obtain the filtered output. ;in, ; Indicates the first in the guided filter window The pixel value of each pixel in the guide image at the corresponding pixel position; Indicates the first Input image within each guided filter window and guide map Covariance between and These represent the guide diagrams. and input image The Middle The average pixel value of each guided filter window; This represents the pixel standard deviation of the k-th guided filter window in the guided graph G; This indicates that there are preset regularization parameters and differences between different image pyramid layers; This represents the filtering result of the i-th pixel within the k-th guided filtering window;

[0067] Then, the filtering results of each level are upsampled sequentially and fused with the filtering results of the previous level until the resolution of the original visual image is restored, and the refined dust concentration equivalent transmittance is obtained. Based on the degradation model, the inverse operation is performed to recover the dust-free intrinsic layer from the original visual image; the difference between it and the original visual image is obtained to obtain the dust scattering layer.

[0068] Next, local windows are used to traverse pixels in the dust-free intrinsic layer to obtain the local contrast within each local window, and a contrast distribution heatmap is generated by pseudo-color mapping. Then, based on the contrast distribution heatmap, low-contrast regions, medium-contrast regions, and high-contrast regions in the original visual image are identified. Region recognition is achieved through threshold segmentation, setting two contrast thresholds, C_low and C_high. The visual image is divided into three contrast regions: regions with contrast less than C_low are marked as low-contrast regions, typically corresponding to areas severely affected by dust or lacking detail; regions with contrast greater than C_high are marked as high-contrast regions, typically corresponding to areas with sharp edges and rich texture; and regions in between are marked as medium-contrast regions. The thresholds are usually adaptively determined based on the statistical characteristics of the contrast distribution. Differential enhancement processing is applied to each contrast region. After enhancement, the processed contrast regions are smoothly blended to obtain a purified visual image sequence. The smooth blending is achieved through a gradual transition using distance transformation. The differential enhancement strategy employs different enhancement intensities for different contrast regions to avoid over-enhancement or under-enhancement. Specifically, the processing involves: significantly improving the contrast of low-contrast regions through piecewise gamma correction; improving the contrast of medium-contrast regions through moderate contrast stretching; and maintaining or slightly enhancing high-contrast regions to avoid noise amplification caused by over-sharpening.

[0069] The environmental noise baseline of the collected equipment acoustic signals is dynamically tracked in real time to identify steady-state noise components and construct a time-varying parameter model based on it. The environmental noise baseline reflects the background noise level and spectral characteristics of the site and serves as a reference standard for noise suppression. Specifically, the collected equipment acoustic signals are divided into several overlapping analysis frames, and each analysis frame is windowed using a Hanning window to reduce spectral leakage. A Fast Fourier Transform is then performed on the windowed analysis frames to obtain a time-spectrum graph, which is a two-dimensional matrix where rows and columns correspond to frequency and time, respectively. Furthermore, a time observation window is set, and statistical quantitative indicators such as the time mean, time standard deviation, time coefficient of variation (the ratio of the time standard deviation to the time mean), and peak factor of the time series (representing the ratio of the maximum value to the time mean in the amplitude sequence of all time frames within the time observation window) are obtained based on the time-spectrum graph. The obtained quantitative indicators are compared with preset multi-condition joint judgment rules. If the quantitative indicators meet the multi-condition joint judgment rules... If the frequency points are not found to be steady-state noise components, the corresponding frequency points are marked as steady-state noise components. The multi-condition joint judgment rules include a time standard deviation less than a preset standard deviation threshold, a coefficient of variation less than 0.1, and a peak factor not exceeding 1.5 times the time mean. All frequency points marked as steady-state noise components are extracted, and their median spectral amplitudes are connected to form a spectral envelope curve, which is the environmental noise spectral envelope for the current time period. Next, the environmental noise spectral envelope is divided into several sub-bands based on frequency, each covering a specific frequency range. The energy integral of the spectral envelope is calculated in each sub-band to obtain a multi-dimensional sub-band energy vector (the specific dimension depends on the number of sub-bands). Time series modeling is performed on the energy sequences of each sub-band, and a first-order autoregressive model is used to describe its dynamic evolution, resulting in a time-series parameter model. The mathematical expression formula for the time-varying parameter model is:

[0070] In the formula, Indicates noise signal, and Indicates model parameters; and Identify the model order; It represents white noise excitation; based on the time-varying parameter model, it predicts the environmental noise spectrum envelope at the next moment, maps the predicted energy of each sub-band back to the frequency domain, and reconstructs the complete predicted spectrum envelope through spline interpolation to provide feedforward information for the design of adaptive filters, thereby improving the real-time performance and accuracy of noise suppression.

[0071] An adaptive frequency-domain notch filter bank is designed based on the obtained time-varying parameter model, and noise suppression is performed on the equipment acoustic signal to obtain a purified equipment acoustic signal. The adaptive notch filter bank consists of multiple narrowband notch filters, each notch filter suppressing a noise frequency component. Specifically, a short-time Fourier transform is performed on the current analysis frame of the equipment acoustic signal to obtain a time-frequency plane representation. The spectral ridges of environmental noise, i.e., the continuous trajectory of the noise dominant frequency over time, are identified on the time-frequency plane. Peak detection is performed on the predicted spectral envelope, and local maxima are located as the noise center frequency, and their corresponding time-series evolution sequences are obtained simultaneously. The Kalman filter method is used to predict the noise center frequency at future times for the time evolution sequence of each noise center frequency. The state variables of the Kalman filter include the center frequency and its rate of change, and the observed variable is the currently detected peak frequency. A smooth frequency trajectory prediction is obtained through recursive updates.

[0072] A notch filter is constructed at each predicted noise center frequency. The frequency domain transfer function of the notch filter adopts the form of a second-order IIR notch filter, and its gain characteristic is as follows:

[0073] In the formula, Center frequency The transfer function of the notch filter at that point; Indicates the sampling frequency. Let be the pole radius, used to control the notch bandwidth; z is a complex variable. Next, the notch bandwidth of the notch filter is adaptively controlled by dynamically adjusting the pole radius r by analyzing the spectral diffusion around the noise center frequency. When the energy of the noise frequency is concentrated, the pole radius is reduced to narrow the notch bandwidth, avoiding suppression of nearby effective signal components. When the energy of the noise frequency diffuses, the pole radius is increased to widen the notch bandwidth, ensuring the integrity of noise suppression. All notch filters are cascaded to form an adaptive frequency domain notch filter bank, and based on this bank, the acoustic signal of the equipment is filtered to achieve selective attenuation of noise components in the frequency domain. To preserve the weak acoustic characteristic components of equipment faults, the filtering... The device design incorporates a feature protection mechanism: a common fault characteristic frequency library (such as bearing fault characteristic frequency, gear meshing frequency, etc.) is pre-established. When the interval between the notch filter center frequency and the fault characteristic frequency is less than the critical protection bandwidth, the center frequency of the adaptive frequency domain notch filter is automatically adjusted or its notch depth is reduced to avoid accidentally damaging fault characteristics. The filtered frequency domain signal is subjected to inverse Fourier transform to restore the time domain waveform, and the continuous time domain signal is reconstructed by the overlapping addition method to obtain the purified equipment acoustic signal. The purified equipment acoustic signal effectively suppresses steady-state environmental noise while retaining transient acoustic characteristics such as fault impact and friction noise, providing a high signal-to-noise ratio data basis for subsequent fault diagnosis.

[0074] Electromagnetic interference (EMI) characteristics are identified and extracted from vibration time-series data and current waveform data to distinguish between mechanical shock signals and electromagnetic pulse (EMP) interference signals. Mechanical shock signals typically exhibit transient impacts in the time domain and broadband characteristics in the frequency domain, while EMP interference manifests as sharp time-domain pulses and harmonic components at specific frequencies. Specifically, the signals are first decomposed into multiple intrinsic mode components using wavelet transform or empirical mode decomposition. Then, the statistical characteristics such as kurtosis and skewness of each intrinsic mode component are calculated. Mechanical shock signals typically have high kurtosis but a wide frequency distribution, while EMP signals have extremely high kurtosis and are concentrated at the power frequency and its harmonic frequencies. Finally, based on the differences in statistical characteristics, the signal components are classified into mechanical components and EMP components.

[0075] An interference feature library is established, comprising time-domain waveform templates of interference pulses, frequency-domain harmonic fingerprints, and an interference correlation matrix among multiple sensors. The construction process begins by collecting a large number of labeled electromagnetic interference samples. Through statistical analysis, time-domain waveform features of typical interferences are extracted, including parameters such as pulse amplitude, rise time, and duration, forming a waveform template library. Then, spectral analysis is performed on the electromagnetic interference samples to extract the frequency, amplitude, and phase features of frequency-domain harmonics, forming a harmonic fingerprint library. Finally, the correlation between interference signals collected by multiple sensors is analyzed to establish a correlation matrix, which includes an amplitude correlation matrix, a waveform similarity matrix, and a delay mean matrix. This matrix is ​​used to characterize the spatial response characteristics of the sensor network to electromagnetic interference from different perspectives.

[0076] Template matching and adaptive cancellation are performed on electromagnetic pulse interference signals based on an interference feature library. Template matching identifies the location and type of interference components by calculating the similarity between the real-time electromagnetic pulse interference signal and waveform templates in the feature library. Adaptive cancellation constructs a cancellation signal based on the matching results to subtract interference from the original electromagnetic pulse interference signal. The matching process first performs a sliding window scan on the real-time electromagnetic pulse interference signal, calculating the normalized cross-correlation coefficient between the signal and each template within each window. The normalized cross-correlation coefficient is a value used to measure the similarity between the real-time electromagnetic pulse interference signal and the waveform template at different time offsets, calculated using the time-domain sliding cross-correlation method. When the cross-correlation coefficient exceeds a preset threshold (usually set to 0.8-0.9), it is determined that there is interference of the corresponding type in that window. Then, based on the matched waveform templates and the correlation matrix, the amplitude and phase parameters of the current interference component are estimated, and an adaptive cancellation signal is constructed based on this, which is then subtracted from the original signal. The resulting purified vibration time series data and current waveform data effectively eliminate the influence of electromagnetic pulse interference while preserving the true mechanical and electrical characteristics of the equipment operation.

[0077] Non-uniformity correction and temperature calibration compensation are performed on the temperature distribution data to eliminate the inconsistencies in the response of the infrared thermal imaging equipment. Non-uniformity correction is based on a two-point correction method, which establishes a linear mapping between the response value and the true temperature by measuring the equipment response at two reference points with known temperatures. Specifically, the response values ​​of a high-temperature reference blackbody and a low-temperature reference blackbody are first collected; then, a linear correction coefficient is established for each pixel; finally, the correction coefficient is applied to correct the real-time measurement value. The temperature calibration compensation process involves measuring the ambient temperature and the surface emissivity of the target object; then, a model relating radiant energy to the true temperature is established based on the Stefan-Boltzmann law, and compensation calculations are performed on the corrected temperature value based on this model. The resulting purified temperature distribution data has high measurement accuracy and spatial consistency, accurately reflecting the true thermal distribution state of the equipment.

[0078] The purified visual images, equipment acoustic signals, vibration time-series data, current waveform data, and temperature distribution data are aggregated to obtain corresponding multi-source purification data. During the aggregation process, time alignment and spatial registration are performed on the data for each modality to ensure that the data collected by different sensors are consistent in time and space, providing a synchronized data foundation for subsequent cross-modal feature extraction.

[0079] In an embodiment of the present invention, the process of obtaining the collaborative feature set includes:

[0080] Spatiotemporal segmentation and multi-scale decomposition are performed on moving targets within the purified visual images to identify and reconstruct their complete spatial morphological features. These complete spatial morphological features accurately describe the dynamic behavior patterns of moving targets such as coal flow, foreign objects, and equipment components in the coal conveying chain. Specifically, image sequences corresponding to the purified visual images are acquired, and moving target detection and spatiotemporal segmentation are performed on them. A Gaussian mixture background modeling method is used to model the grayscale value sequence of each pixel in the visual image into multiple Gaussian components. Each Gaussian component represents a possible background state, including a static background and a periodically moving background. Each pixel contains multiple Gaussian components, characterized by their mean and variance. When a new visual image is reached, the current pixel value is matched with the Gaussian components. If a Gaussian component satisfies that the absolute value of the difference between the grayscale value and the mean is less than 2.5 times the variance, then the corresponding pixel is... Points are marked as background pixels; if all Gaussian components do not match, they are determined to be foreground images, i.e., components of the moving target; connected component analysis is performed on the identified foreground pixels to obtain connected regions, and geometric features such as area, perimeter, and compactness of the connected regions are calculated. Based on these features, noise regions that are too small (e.g., area less than the preset area threshold Amin) and false detection regions that are too large (e.g., area greater than the preset area threshold Amax) are filtered out, and valid moving target regions in the coal conveying link are retained; the bounding box of the segmented moving target region is extracted, and its center coordinates, width, and height are recorded as spatial references for subsequent motion decomposition. Next, by tracking the displacement of the centroid of the moving target and image feature points between consecutive image frames, the global translation component, local rotation component, and non-rigid deformation component of the moving target are extracted, and the complete spatial morphological features are reconstructed based on them. The velocity and direction of the global translation vector, the angular velocity and rotation center of the local rotation, the spatial distribution map of the non-rigid deformation, and the geometric features (area, aspect ratio, compactness) of the moving target are combined into a high-dimensional feature vector. This feature vector comprehensively describes the spatial morphology and dynamic behavior of the moving target and serves as a representative feature of the visual modality for subsequent cross-modal fusion.

[0081] Based on wavelet packet decomposition and short-time Fourier transform, joint time-frequency analysis is performed on the purified acoustic signal of the equipment to identify and extract the fundamental frequency component and fault modulation component of the rotating parts of the equipment. The fundamental frequency component reflects the normal operating status of the equipment, while the fault modulation component carries key information about the equipment's abnormality. Specifically, the acoustic signal from the purified equipment is subjected to both Short-Time Fourier Transform (SFT) and Wavelet Packet Decomposition (WPD) to fully utilize the complementary advantages of the two methods. SFT is suitable for analyzing periodic fundamental frequency components, while WPD is suitable for capturing transient fault impacts. Furthermore, the time spectrum of the SFT and the wavelet packet energy spectrum of the WPD are obtained and fused for analysis. The authenticity of the signal components is verified by calculating the consistency of the energy peaks detected in each frequency band. The consistency calculation uses cross-correlation analysis, and a valid signal component is confirmed when the time deviation between the corresponding energy peak positions is less than a preset time threshold. Next, the average power spectrum of the SFT time spectrum on the time axis is obtained, and local peak detection is performed to obtain the peak frequency, which is then used as a fundamental frequency candidate. The theoretical fundamental frequency is calculated based on the rated speed of the motor in the coal conveying link and compared with the fundamental frequency candidate. If the absolute value of the difference between the two is less than the fundamental frequency tolerance (typically 5% of the theoretical fundamental frequency), the corresponding fundamental frequency candidate is marked as a fundamental frequency component. The fundamental frequency component corresponds to the basic operating parameters of the equipment, such as motor speed and conveyor belt speed. Frequency serves as the benchmark for analyzing equipment operating status. The acoustic signal is then bandpass filtered and transformed using a Hilbert transform to obtain an analytical signal. The amplitude of this analytical signal is used as the envelope signal, which undergoes spectral analysis, including a Fast Fourier Transform (FFT) to obtain the envelope spectrum. The envelope spectrum reveals the modulation frequency components, i.e., fault characteristic frequencies. Significant modulation frequencies are identified through peak detection of the envelope spectrum. These modulation frequencies are then matched against a known fault characteristic frequency library, which includes bearing fault characteristic frequencies (outer ring, inner ring, rolling element, and cage fault frequencies), gear fault characteristic frequencies (meshing frequency, crack frequency, etc.). When a detected modulation frequency matches an item in the fault characteristic frequency library (deviation less than 0.1), the presence of a corresponding fault modulation component is confirmed. The intensity of the fault modulation component is calculated, defined as the ratio of the peak amplitude of the modulation frequency to the average level of the envelope spectrum. A higher fault modulation intensity indicates a more significant fault characteristic and a more abnormal equipment state. The fundamental frequency component and the fault modulation component can be used as representative features of the acoustic modes for cross-modal fusion.

[0082] Spatial evolution analysis of thermal anomaly regions is performed on the purified temperature distribution data, and the location and intensity of heat sources are inverted to generate thermal anomaly propagation characteristics. Local overheating of equipment is an important symptom of failure, and the spatial distribution and temporal evolution of overheated areas contain rich fault information. Specifically, based on statistical thresholding methods and combined with purified temperature distribution data, thermal anomaly regions with temperatures higher than normal are extracted. The spatial evolution patterns of these regions are analyzed, including expansion, contraction, drift, splitting, and merging. For example, the time derivative of the area of ​​the thermal anomaly region is calculated, with positive values ​​indicating expansion and negative values ​​indicating contraction. Simultaneously, the location and intensity of the heat source are inverted based on the temperature distribution characteristics of the thermal anomaly region. The inversion process must follow the heat conduction equation. Nonlinear least-squares fitting is performed on the temperature distribution data within the detected thermal anomaly region to optimize the solution for the heat source location and intensity. Furthermore, by integrating the spatial evolution patterns and the heat source location and intensity, thermal anomaly propagation characteristics are generated. These characteristics include the spatial location, area, average temperature, maximum temperature, temperature gradient, diffusion rate, diffusion direction, heat source location, and heat source intensity of the anomaly region. These thermal anomaly propagation characteristics comprehensively characterize the spatiotemporal dynamics of temperature anomalies.

[0083] Empirical Mode Decomposition (EMD) is performed on the purified vibration time-series data, and the distribution characteristics of vibration energy in the time-frequency plane are determined based on this. Vibration signals directly reflect the mechanical state of equipment, and their time-frequency characteristics contain rich fault information. Specifically, all local extrema of the signal are first identified, then the upper and lower envelopes are fitted using cubic spline interpolation, and the average envelope is calculated as the local mean. The first IMF component is obtained by subtracting the local mean from the original signal. This process is repeated to extract subsequent IMF components sequentially from the residual signal until the residual signal becomes a monotonic function or meets the stopping criterion. Each IMF component represents the oscillation mode of the signal at a specific time scale, corresponding to different physical processes from high frequency to low frequency. The distribution of vibration energy in the time-frequency plane is obtained through Hilbert-Huang transform (HHT). This method performs Hilbert transform on each IMF component to obtain the instantaneous frequency and instantaneous amplitude. The instantaneous frequencies and instantaneous amplitudes of all IMFs are combined to obtain the Hilbert spectrum (time-frequency energy distribution). Based on the Hilbert spectrum, the time-frequency distribution characteristics of vibration energy are extracted, including the dominant frequency component and its time-varying trajectory, energy concentration (energy entropy of the time-frequency plane), fluctuation range of instantaneous frequency, time distribution pattern of energy, and energy proportion of specific frequency bands.

[0084] Load fluctuation pattern identification is performed on the purified current waveform data, and load fluctuation characteristics are extracted by combining the load power spectral density of the current waveform data. Specifically, firstly, the three-phase currents extracted from the purified current waveform data are subjected to Clarke or Park transform to convert the three-phase currents into two-phase orthogonal components or rotating coordinate system components. Then, the amplitude and phase of the three-phase currents are calculated, and the envelope signal of the load current is extracted. Next, the envelope signal is classified into load fluctuation patterns to identify typical patterns such as steady-state load, periodic fluctuation load, abrupt load, and random fluctuation load. For example, the periodic fluctuation pattern is characterized by periodic oscillation of the effective current value. The occurrence frequency and duration of various load fluctuation patterns are statistically analyzed, and the statistical results are used as the pattern characteristics of load fluctuation. At the same time, the time series of the effective current value of the purified current waveform data is extracted to reflect the slow change trend of the load current. Frequency domain analysis is performed on the time series of the effective current value to calculate the load power spectral density, and then... Peak detection is performed on the load power spectral density to identify and extract peak drift features, including the drift velocity (defined as the time derivative of the peak frequency) and amplitude change of the peak frequency over time. The detected peak frequency is compared with the mechanical characteristic frequencies of the equipment (such as the coal feeder speed frequency and the conveyor belt operating frequency) to identify the physical source of the peak frequency. When the peak frequency matches the characteristic frequency of a certain equipment, it is confirmed that the corresponding load fluctuation originates from the corresponding equipment, providing a basis for fault location. The statistical characteristics of the current effective value, the classification results of the load fluctuation mode, the occurrence frequency, the peak frequency, peak amplitude, and peak drift velocity of the power spectral density are combined to form the load fluctuation features. The load fluctuation features comprehensively characterize the load fluctuation characteristics of the current waveform data and serve as representative features of the current modes for cross-modal fusion.

[0085] A mutual information matrix is ​​constructed based on the extracted cross-modal features, and a physical constraint correlation model is built. Based on this model, a feature projection transformation is performed on the mutual information matrix to obtain a collaborative feature set. The cross-modal features include complete spatial morphological features, fundamental frequency components and fault modulation components, thermal anomaly propagation features, load fluctuation features, and vibration energy distribution features in the time-frequency plane. The collaborative feature set integrates complementary information from multimodal data, improving the accuracy of anomaly detection and fault diagnosis. Specifically, the modal feature vectors corresponding to the extracted cross-modal features are obtained and normalized to the range [0, 1]. The mutual information between different modal feature vectors is then obtained. The mutual information is calculated based on the joint probability distribution and marginal probability distribution. First, the modal feature vectors are discretized, and the frequency of the eigenvalues ​​of the modal feature vectors falling into each feature region is counted to obtain a discrete probability distribution. Based on this, the joint probability distribution between different modal feature vectors is calculated. Marginal probability distribution of modal eigenvectors Finally, the mutual information value is calculated based on the joint probability distribution and the marginal probability distribution. In the formula, and Let m and n represent the modal feature vectors, respectively; the larger the mutual information value, the stronger the interdependence between the two modal features and the more common information they contain; calculate the mutual information for all modal feature pairs and construct a mutual information matrix;

[0086] A physical constraint correlation model for cross-modal characteristics is constructed based on the multi-physics coupling mechanism of the coal conveying system. This model describes the causal relationships and constraints between different modal characteristics, ensuring that feature fusion conforms to physical laws. The physical constraint correlation model includes mechanical-acoustic coupling constraints (e.g., mechanical vibration is converted into acoustic signals through solid and air propagation; the constraint relationship is manifested as the correspondence between vibration frequency components and acoustic frequency components, and the positive correlation between vibration amplitude and acoustic intensity) and electrical-mechanical coupling constraints (e.g., motor load fluctuations directly affect the conveyor belt speed, thus affecting the motion state of the coal flow; the constraint relationship is manifested as the temporal synchronization of load current changes and coal flow speed changes). Based on the physical constraint correlation model, physical coupling coefficients between different cross-modal characteristics are calculated (e.g., the electrical-thermal coupling coefficient can be obtained by calculating the cross-correlation coefficient between the effective current time series and the temperature time series), and a physical constraint matrix is ​​constructed based on this model. The matrix elements are the physical coupling coefficients between different cross-modal characteristics, used to reflect the strength of the physical constraints.

[0087] A weighted mutual information matrix is ​​constructed by combining the physical constraint matrix and the mutual information matrix. The elements of the weighted mutual information matrix are defined as the product of the two matrices. Eigenvalue decomposition is performed on the weighted mutual information matrix to obtain eigenvalues. The eigenvalues ​​are arranged in descending order, with the highest eigenvalues ​​corresponding to the main change direction of the cross-modal features. The eigenvectors corresponding to the top D largest eigenvalues ​​(D is an integer between 3 and 10) are selected to form a projection matrix. The eigenvectors of the cross-modal features are concatenated into a high-dimensional joint feature vector. Based on the projection matrix, the joint feature vector is projected to obtain a collaborative feature vector. Each component in the collaborative feature vector represents a cross-modal collaborative mode, fusing information from multiple modalities. The above feature extraction and fusion process is repeated for multimodal data at consecutive time points to generate a collaborative feature set. Through cross-modal collaboration, the collaborative feature set has higher robustness and discriminative ability than single-modal features, and can effectively cope with single sensor failures or local information loss.

[0088] In an embodiment of the present invention, the implementation process of the edge decision module includes:

[0089] A layered fusion analysis architecture is deployed in the edge computing nodes of the corresponding coal conveying link. The layered fusion analysis architecture includes an edge fast response layer, an edge deep analysis layer, and a cloud collaborative verification layer. The layered fusion analysis architecture can balance real-time response speed and analysis accuracy, and minimize end-to-end latency while ensuring decision reliability, so as to meet the real-time requirements of coal conveying system safety monitoring. Specifically, a hierarchical fusion analysis architecture is deployed in the edge computing nodes of the target coal conveying link. This architecture employs a three-layer progressive decision-making model, from rapid initial screening to refined diagnosis and then to global verification, forming a complete analysis link. Edge computing nodes refer to industrial-grade computing devices pre-deployed at the coal conveying link site, capable of handling complex real-time analysis tasks. The hierarchical fusion analysis architecture includes an edge rapid response layer, an edge deep analysis layer, and a cloud collaborative verification layer. Information transmission and collaborative work are achieved between each layer through data interfaces and decision triggering mechanisms. The edge rapid response layer is deployed on edge computing nodes close to sensors (such as industrial gateways and edge servers), featuring low latency and rapid response, responsible for real-time data acquisition, preliminary anomaly detection, and emergency alarms. The edge deep analysis layer is deployed on edge nodes with strong computing capabilities (such as field workstations and edge AI servers), equipped with GPUs or dedicated AI accelerators, responsible for running deep learning models for refined fault diagnosis and status assessment. The cloud collaborative verification layer is deployed in cloud computing centers or enterprise data centers, possessing abundant computing and storage resources, responsible for global data analysis, historical data association, expert knowledge base matching, and model training and updates.

[0090] The edge fast response layer is used to construct a baseline feature space under normal operating conditions based on cross-modal features, obtain the Mahalanobis distance between it and the corresponding collaborative features, and perform fast discrimination based on the Mahalanobis distance to obtain abnormal event logs. The Mahalanobis distance is a multi-dimensional distance metric that considers the correlation between features and can effectively identify abnormal samples that deviate from the normal distribution. Specifically, under the normal operating conditions of the coal conveying link, collaborative feature samples are continuously collected for data collection. Cross-modal features are extracted from them to construct the collaborative feature vectors corresponding to the collaborative feature samples, and the mean vector (the components in the mean vector represent the expected value of a certain dimension of the collaborative feature vector under normal operating conditions) and covariance matrix (the matrix elements represent the covariance between different dimensions of features) of the corresponding collaborative feature vectors are calculated. Then, the covariance matrix is ​​regularized and the inverse matrix corresponding to the covariance matrix is ​​obtained. The mean vector, inverse matrix, and regularized covariance matrix are used as spatial parameters of the baseline feature space under normal operating conditions and are updated periodically. The Mahalanobis distance between the current collaborative feature vector and the baseline feature space is calculated and compared with a preset first... Thresholds are compared; the first threshold is determined by the Mahalanobis distance distribution of collaborative feature samples under normal operating conditions; if the Mahalanobis distance is not greater than the first threshold, it is judged as normal operation; if the Mahalanobis distance is greater than the first threshold, it is judged as an abnormal operating state, and the edge fast response layer triggers a suspected abnormality warning and triggers the edge deep analysis layer; at the same time, the occurrence time of the Mahalanobis distance exceeding the limit event, the Mahalanobis distance value, and detailed information of the collaborative feature vector are recorded to generate an abnormal event log of the fast response layer; this log serves as input data for the edge deep analysis layer and is also used for post-event statistical analysis and threshold optimization; by monitoring the time evolution trend of the Mahalanobis distance, gradual anomalies can be identified, that is, the situation where the Mahalanobis distance continues to increase but has not yet exceeded the threshold, providing early warning signals for preventive maintenance.

[0091] A refined fault diagnosis model is deployed in the edge deep analysis layer, and the collaborative feature set corresponding to the occurrence time in the abnormal event log is input into the refined fault diagnosis model to obtain the output results. The refined fault diagnosis model realizes the accurate identification and analysis diagnosis of coal conveying link faults. It is an end-to-end diagnostic system based on deep learning, which can automatically learn the high-level abstract representation of features and complex fault modes, and has strong nonlinear fitting and generalization capabilities. Specifically, the refined fault diagnosis model adopts a hierarchical deep neural network architecture, integrating convolutional neural networks and long short-term memory networks. The convolutional neural network is responsible for extracting deep semantic representations of cross-modal features, while the long short-term memory network is responsible for capturing the temporal dependencies of fault evolution, thus achieving joint modeling of static feature patterns and dynamic evolution processes. The obtained collaborative feature set is input into the refined fault diagnosis model, which divides the deep neural network into three functional layers: bottom layer, middle layer, and top layer. The bottom layer is used to extract basic features, corresponding to basic operations in signal processing and image processing, such as the temporal peak of vibration signals and the edge contours of visual images. The middle layer is used to extract fault-related features, combining basic features into discriminative features related to known fault types through multi-layer nonlinear transformations. For example, the vibration features corresponding to roller wear faults are manifested in a specific frequency band (such as 1000Hz). The energy anomaly in the 2000Hz frequency band is enhanced. The middle layer network can automatically learn the correlation pattern between energy in this frequency band and idler roller wear. The top layer is responsible for fault classification decision-making, mapping the extracted fault correlation features to the fault category space, outputting the probability distribution of each fault type, and normalizing it through the Softmax activation function to obtain the output results. The output results include fault type (such as bearing failure, belt misalignment, material accumulation, motor overload, etc.), fault confidence, and fault severity level. Simultaneously, the diagnostic process data of the edge depth analysis layer is recorded, including the input collaborative feature sequence, the intermediate activation values ​​of each layer of the model, the output probability distribution, and the final diagnostic results. This data is used for continuous model optimization and fault source analysis. It should be noted that the specific construction process of the refined fault diagnosis model is existing technology and will not be elaborated on in this application.

[0092] The decision is evaluated based on the confidence level of the output. If the confidence level is less than the preset second threshold, the corresponding output is marked as a low-confidence decision, and the cloud-based collaborative verification layer is triggered. The cloud-based collaborative verification layer performs correlation analysis of global historical data and verification of expert knowledge base, and updates the low-confidence decision based on the verification results. The global historical data includes massive multimodal data, fault event records, and maintenance logs accumulated over long-term operation of the coal conveying link, which can provide richer contextual information for the identification of complex anomalies.

[0093] Specifically, when the edge fast response layer triggers a suspected anomaly warning, the collaborative feature set at the corresponding time of occurrence is input into the refined fault diagnosis model to obtain the corresponding output result. The maximum probability value in the probability distribution of the corresponding fault type is used as the confidence level. If the confidence level is not less than the preset second threshold (usually ranging from 0.7 to 0.85), it indicates that the corresponding output result is a high-confidence decision, and it is directly output as the final result. If the confidence level is less than the second threshold, it indicates that the corresponding output result has low confidence, and it is marked as a low-confidence decision, triggering the cloud collaborative verification layer. Then, the cloud verification layer uses the current corresponding collaborative feature vector as the query sample and uses cosine similarity calculation to retrieve historical fault cases from the cloud database that are similar to the query sample. The retrieved historical cases are labeled and their fault types are analyzed. If a certain fault type dominates in the historical cases (accounting for a high proportion in the historical cases), it is inferred that the current suspected anomaly warning belongs to the corresponding fault type. If the fault type distribution of the historical cases is relatively dispersed, the temporal evolution model of the historical cases is further analyzed, and its evolution trajectory is compared with that of the current collaborative feature vector. The system performs dynamic time-warping matching to obtain the most similar historical cases; it queries the control measures taken in historical cases and the effect evaluation after the measures were implemented; if a certain type of control measure successfully solved the problem in historical cases, then the corresponding control measure is highly applicable to the current suspected anomaly warning; it establishes a mapping relationship between the current suspected anomaly warning and historical cases through association analysis, using historical experience to assist current decision-making; at the same time, it performs matching verification on low-confidence decisions based on an expert knowledge base to ensure that low-confidence decisions are in line with the physical laws of equipment operation and expert experience; the expert knowledge base contains fault diagnosis rules, equipment operation constraints, and causal relationships of anomaly patterns summarized by domain experts; the matching verification process matches the current features with the feature patterns in the knowledge base to identify the matching rule conditions; then, it applies a logical reasoning engine (such as forward reasoning or backward reasoning) to deduce the fault conclusion based on the rule base; logical reasoning verification can use expert experience and physical mechanism knowledge to perform interpretive verification and error correction on the output of the deep learning model, improving the reliability and interpretability of the diagnosis; after the cloud collaborative verification layer completes the verification, it updates the low-confidence decisions and their corresponding confidence levels based on the verification results, and repeats the corresponding process.

[0094] The intermediate and final results of the above analysis are subjected to data compression and feature extraction to obtain the global operating status. The global operating status is a comprehensive description of the current operating status of the coal conveying link, containing multi-dimensional status information. Data compression reduces the amount of data through dimensionality reduction techniques (such as PCA and autoencoders) and data encoding techniques (such as Huffman coding and arithmetic coding), facilitating storage and transmission. Feature extraction extracts key status indicators from the intermediate results of hierarchical analysis, including abnormal event logs and confidence levels. The global operating status is organized in the form of structured data, including fields such as timestamp, equipment identifier, status category (normal / abnormal / fault), detailed description, confidence level, and suggested measures. The global operating status is updated periodically and pushed to the monitoring center. Then, the monitoring center determines the current status category of the coal conveying link based on the updated global operating status. If the status category is normal, continuous monitoring is performed; if the status category is abnormal or faulty, the intelligent control and feedback module is activated.

[0095] One embodiment of the present invention further includes: real-time load detection of edge computing nodes; when the real-time load exceeds a preset load threshold, identifying computationally intensive tasks in the edge analysis process as candidates for task migration; computationally intensive tasks typically include tasks with high computational complexity but relatively low real-time requirements, such as deep learning model inference, large-scale matrix operations, and complex optimization solutions; evaluating the migration effect of each candidate based on the actual end-to-end latency and energy consumption after task migration, and dynamically adjusting the task allocation strategy based on the migration effect; recording the latency and energy consumption data of each task migration, and dynamically adjusting the task allocation strategy through reinforcement learning or heuristic algorithms to learn the optimal load balancing scheme. This dynamic task migration mechanism fully utilizes the heterogeneous resources of the edge and cloud, improving the overall processing capacity of the system while ensuring real-time performance.

[0096] In embodiments of the present invention, the implementation process of the intelligent control and feedback module includes:

[0097] A multi-level fault discrimination decision tree is constructed for the coal conveying link to achieve progressive fault diagnosis from coarse-grained to fine-grained. The multi-level fault discrimination decision tree adopts a hierarchical decision structure, and each level of decision node performs special discrimination based on different combinations of collaborative feature sets and global operating status, gradually refining the diagnosis results.

[0098] Specifically, the root node of the multi-level fault discrimination decision tree is a binary discrimination node for normal and abnormal. The root node uses the received global operating status as the initial discrimination basis and verifies the comprehensive deviation of the collaborative feature set. The comprehensive deviation is defined as the weighted Euclidean distance of each dimension of the collaborative feature vector from the benchmark value. When the global operating status label is "abnormal" and the comprehensive deviation exceeds the threshold of the root node, it is judged as an abnormal branch and enters the subsequent level of refined discrimination. When the global operating status label is "normal" or the comprehensive deviation does not exceed the threshold, it is judged as a normal branch, outputs the normal operation conclusion, and exits the decision process.

[0099] The decision tree hierarchy following the root node consists of three types of decision nodes: fault category identification nodes, fault location nodes, and fault severity assessment nodes. Based on these multi-level fault identification decision nodes, and combined with the collaborative feature set and global operational status, fault combination identification is performed on the coal conveying link to obtain a fault combination identification report. The process of fault combination identification includes:

[0100] Based on the fault category discrimination node, modal decomposition is performed on the collaborative feature set to identify and mark the modal sources that dominate the fault features. A modal combination strategy library is established for each fault type to determine the specific fault type and proceed to the next level. The fault category discrimination node is located in the second level. Specifically, by analyzing the projection coefficients of the collaborative feature vectors on each feature vector in the projection matrix, and calculating the modal contribution of each modal feature to the current "anomaly," the modal contribution is defined as the weight ratio of the corresponding modal feature in the formation process of the collaborative feature vector. Based on the modal contribution, the modal source that dominates the current fault feature is identified. For example, when the visual modal contribution is dominant, it is inferred that the fault is related to the visual feature. A modal combination strategy library is constructed for fault types in historical cases. The modal combination strategy defines the feature discrimination priority and feature threshold for each fault type. For example, for the "conveyor belt misalignment" fault, the discrimination strategy is: first check the coal flow deviation feature of the visual modality (e.g., deviation > 5% bandwidth), then check the lateral vibration feature of the vibration modality (e.g., lateral vibration energy ratio > 30%), and finally check the frictional noise feature of the acoustic modality (e.g., high-frequency component enhancement > 10dB). Only when multiple modal features meet the conditions simultaneously is the fault type confirmed, avoiding misjudgment based on a single modality.

[0101] The fault location node determines the spatial location of the fault source based on the spatial topology of the abnormal sensor array and the propagation delay analysis of fault characteristics; the fault location node is located at the third level of the multi-level fault discrimination decision tree. Specifically, the coal conveying link is divided into several monitoring sections, and the geometric distribution relationship of the heterogeneous sensor arrays deployed in each monitoring section is obtained; simultaneously, the raw feature data collected by each sensor in the heterogeneous sensor array is time-aligned and abnormal response analyzed to identify the time when each sensor detects an abnormal signal, and the propagation path of the fault characteristics is obtained by comparing the time sequence of the abnormal responses of the sensors; the time difference of abnormal responses between any two sensors is calculated. In the formula, Indicates sensor and sensors The time difference for receiving fault signals; and These represent the locations from the fault location to the sensor. and sensors The distance is denoted by v; the propagation speed is denoted by v; and the time difference is used to determine whether the signals detected by the two sensors originate from the same fault source. If the time difference is close to the deviation of the actual response time of the sensors, it indicates that they belong to the same fault source. The spatial location of the fault source is then calculated using the multi-sensor time-delay triangulation method, which is achieved by solving a set of time-delay equations for multiple sensors. The optimal estimate of the fault source location is obtained by solving the overdetermined set of equations using the least squares method. For fault source locations that cannot be located by time delay, a feature intensity-based location method is used, which is based on the principle that the closer to the fault source, the greater the feature intensity of the fault characteristics. The fault location is then inverted through the spatial distribution of feature intensity. The location of the fault source obtained is mapped to the equipment number and component name of the coal conveying link to obtain specific fault location information, including section number, equipment name, specific component, and location coordinates. The fault location information provides accurate target location for subsequent maintenance scheduling and control strategy formulation.

[0102] The fault severity assessment node tracks the fault evolution trend of fault characteristics on the time axis to fit the fault degradation curve and predict the remaining time for the fault characteristics to reach the danger threshold. The accurate assessment of fault severity can provide a quantitative basis for maintenance decisions, avoid economic losses and safety hazards caused by over-maintenance or maintenance delays, and is a key technical link to achieve predictive maintenance. Specifically, historical time series related to fault characteristics are extracted, including health indicators such as vibration intensity, temperature rise, and current fluctuation. The selected health indicators are normalized to eliminate dimensional differences between different physical quantities, ensuring the comparability of parameters in health indicator calculations. A comprehensive health indicator is calculated based on the normalized health indicators. Time series data of the comprehensive health indicator during equipment operation are continuously collected, and degradation trend analysis is performed to identify the main modes of fault feature evolution. The first-order difference of the comprehensive health indicator is calculated; the first-order difference reflects the degradation rate, and the second-order difference reflects the degradation acceleration. The sign distribution of the first-order difference is statistically analyzed: if the proportion of positive values ​​(increased comprehensive health indicator) exceeds a preset threshold, it is determined to be a monotonic degradation mode, indicating continuous fault deterioration; if positive and negative values ​​alternate, it is determined to be a fluctuating degradation mode, indicating that the fault exhibits non-monotonic evolution under the influence of environmental disturbances or load changes. The sign of the second-order difference is statistically analyzed: if the second-order difference is consistently positive, it is determined to be an accelerated degradation mode, indicating that the fault has entered a rapid deterioration stage; if the second-order difference is consistently positive... If the difference is close to zero, it is determined to be a linear degradation mode; if the second-order difference is negative, it is determined to be a deceleration degradation mode (rare, usually occurring in the initial break-in stage). Based on the identified degradation mode, a suitable degradation model is selected for curve fitting. This invention preferably uses the Weibull degradation model to describe the temporal evolution of fault characteristics. The Weibull model, due to its flexible shape parameters, can adapt to various degradation modes (linear, exponential, S-shaped, etc.). The Weibull cumulative failure distribution function is used, and the Weibull parameters are estimated using the nonlinear least squares method on historical data to achieve the optimal fit of the degradation curve to the historical data, thus obtaining the fault degradation curve. The goal of parameter estimation is to minimize the squared residual between the model prediction and the actual observation. Based on the fitted fault degradation curve, the remaining time for the fault characteristics to reach the danger threshold is predicted. The remaining time is determined according to the equipment failure criteria, such as the temperature exceeding the material's heat resistance limit. The danger threshold is substituted into the degradation curve equation to obtain the degradation time, and the remaining time is obtained based on its deviation from the current time. The remaining time reflects the urgency of the fault; the shorter the remaining time, the more urgent the fault.

[0103] Based on the predicted remaining time, fault evolution trajectory, fault severity, and fault location, a fault propagation path simulation analysis is performed to calculate the risk probability of each fault propagation path and generate a fault combination discrimination report. The fault propagation path describes the possible pathways and impact chains through which a fault spreads from its initial location to other equipment or systems. Specifically, an equipment association topology diagram of the coal conveying link is obtained. The association topology diagram is represented by a directed graph, where nodes represent equipment or components of the coal conveying system (such as drive motors, conveyor belts, idler rollers, cleaners, etc.), and directed edges represent physical connections, energy transfer, or functional dependencies between equipment. The attributes of the edges include association type (mechanical connection, electrical power supply, material transfer, etc.) and association strength (strong coupling, moderate coupling, weak coupling). The association topology diagram is constructed based on the design drawings and actual layout of the coal conveying system, and implicit association relationships are supplemented based on operational experience. Based on the equipment association topology map, a propagation path search is performed starting from the location coordinates within the current fault location information. The search process employs a breadth-first search method, expanding potentially affected equipment nodes layer by layer. For each expanded equipment node, existing propagation rules are used to determine whether the current fault will affect that equipment node and the propagation delay of the impact. The propagation delays along the propagation path are accumulated to obtain the estimated time for the fault to reach the equipment node. Expansion stops when the propagation path length exceeds a preset maximum depth or the propagation delay exceeds a preset time window (e.g., 24 hours). It should be noted that the existing propagation rules must be satisfied during the search process; for example, the propagation rules for conveyor belt misalignment faults are: misalignment, uneven stress on idlers, accelerated wear of idler bearings, idler jamming, and conveyor belt tearing. Risk probabilities are calculated for all propagation paths obtained from the search. The risk probability of a propagation path is defined as the likelihood that a fault will propagate along that path and cause secondary faults. The risk probability is related to the conditional probability of each propagation link on the path and is calculated using the chain rule. The risk probabilities of each propagation path are sorted to identify high-risk propagation paths. A risk probability threshold is set (typically 0.3 to 0.5), and paths with risk probabilities exceeding the threshold are marked as high-risk paths. Key propagation links on high-risk paths are marked, as these links are the best intervention points to block fault propagation. When generating control strategies, preventive measures for key propagation links are given priority, such as enhanced monitoring, early maintenance, or proactive intervention.

[0104] Based on the fault type, fault location, fault severity, fault evolution trajectory, fault propagation path, and corresponding risk probability, a structured fault combination discrimination report is generated, which is the output of a multi-level fault discrimination decision tree. The report adopts a multi-level information organization method. The top level is the fault summary (fault type, location, level), the second level is detailed diagnostic information (characteristic evidence, confidence level, diagnostic basis), the third level is predictive analysis (remaining time, propagation path, risk assessment), and the fourth level is recommended measures (monitoring priorities, maintenance suggestions, preventive measures).

[0105] Uncertainty quantification is performed on the output of a multi-level fault discrimination decision tree to generate a probability distribution of decision confidence and calculate the variance of the decision confidence. Uncertainty quantification reflects the reliability of the diagnostic results and provides a basis for decision classification and manual review. Specifically, the confidence level of each discrimination node in the multi-level fault discrimination decision tree is estimated during the discrimination process. For nodes based on threshold judgment, the confidence level is related to the degree of deviation of the feature value from the threshold; the greater the deviation, the higher the confidence level. For nodes based on model prediction, the confidence level is directly taken from the probability value output by the model. The confidence level values ​​of each level node are collected based on the discrimination path of the multi-level fault discrimination decision tree, and the node confidence level values ​​are propagated to obtain the comprehensive confidence level of the final output result. The propagation calculation process adopts the confidence propagation mechanism of Bayesian network, considering the conditional dependencies between nodes. When the relationship between nodes is serial (AND logic), the comprehensive confidence level is taken as the minimum or geometric mean of the confidence levels of each node. When the relationship between nodes is parallel (OR logic), the comprehensive confidence level is taken as the maximum of the confidence levels of each node. Simultaneously, the uncertainties in the multi-level fault discrimination decision tree process are analyzed. These uncertainties include measurement uncertainty, model uncertainty, and knowledge uncertainty. Measurement uncertainty stems from sensor noise and data quality issues, which are quantified by sensor accuracy specifications and data quality scores. Model uncertainty arises from the generalization error of the diagnostic model, which is quantified by the confusion matrix and prediction variance of the model on the validation set. Knowledge uncertainty stems from the incompleteness of the fault mode library, which is quantified by knowledge base coverage and expert confidence. A Monte Carlo sampling method is used to generate the probability distribution of decision confidence. Several perturbation samples are generated by randomly perturbing each uncertainty source. The multi-level fault discrimination decision tree process is re-executed for each perturbation sample to obtain a set of decision results and confidence values. The distribution of decision results of perturbation samples is statistically analyzed, and the occurrence frequency of each fault type is calculated as the posterior probability. The distribution of confidence values ​​is statistically analyzed to obtain the mean and probability distribution of confidence. The variance of confidence is calculated based on the distribution of confidence values.

[0106] When the variance of the decision confidence level is greater than the preset confidence threshold, it indicates that the output result of the corresponding multi-level fault discrimination decision tree has high uncertainty and there is a risk of misjudgment. In this case, the corresponding output result is labeled with "manual review" and transmitted to the relevant management personnel for manual review. When the variance of the decision confidence level is not greater than the preset confidence threshold, it indicates that the output result of the corresponding multi-level fault discrimination decision tree is a high-confidence decision. In this case, an intelligent monitoring report is generated based on the discrimination result of the corresponding multi-level fault discrimination decision tree. The intelligent monitoring report includes fault type, fault location, fault severity, fault propagation path, corresponding risk probability, and decision confidence level. The intelligent monitoring report is presented with a visual interface and supports access from multiple terminals (such as monitoring screens, PCs, and mobile devices), providing maintenance personnel with comprehensive and intuitive fault information.

[0107] An adaptive control strategy is generated based on the severity and type of the fault. The generation process of the adaptive control strategy includes:

[0108] Based on the fault type and severity, a set of candidate control strategies is obtained by matching strategies from a preset control strategy library. The control strategy library contains standardized control schemes and expert experience for various fault scenarios. Specifically, the control strategy library adopts a hierarchical organizational structure, including: the first layer categorized by fault type, such as mechanical fault strategies, material fault strategies, and electrical fault strategies; the second layer categorized by fault severity, such as minor fault strategies, moderate fault strategies, and severe fault strategies; and the third layer subdivided by fault location, such as specific strategies corresponding to different equipment or components. Each leaf node corresponds to a specific control strategy, including the control objective, control action sequence, control parameter range, and expected effect. Furthermore, the system locates the corresponding strategy sub-library based on the fault type, then filters applicable strategies based on the fault severity level, and finally refines the strategy based on the fault location. The search results may contain multiple candidate control strategies, representing different control approaches or applicable to different boundary conditions. The applicability of the retrieved candidate control strategies is verified. It is verified whether the current coal conveying link status meets the strategy's execution conditions; for example, if a strategy requires "standby equipment to be in standby mode," the current status of the standby equipment needs to be queried for confirmation. Strategies that do not meet the execution conditions are filtered out, resulting in a set of executable candidate control strategies. Candidate control strategies are initially ranked based on their historical performance, with higher success rates and shorter problem-solving times ranking higher. They are then ranked based on resource consumption, prioritizing strategies with shorter downtime and less economic loss. Furthermore, they are ranked based on security, prioritizing strategies with lower risk and fewer side effects. A comprehensive priority ranking of candidate control strategies is then obtained by considering multiple ranking dimensions.

[0109] For each candidate control strategy in the candidate control strategy set, a forward simulation is performed using a digital twin model of the coal conveying link to predict the system state evolution trajectory after the execution of the candidate control strategy. The forward simulation is performed by inputting the control parameters of each candidate control strategy into the corresponding digital twin model of the coal conveying link. Based on the corresponding digital twin model, the system state evolution trajectory is generated by tracking equipment operating parameters, material flow status, and energy consumption trends in real time. The simulation uses virtual mapping technology to evaluate the actual impact of the strategy before actual execution. The simulation process first uses the current real-time operating state of the coal conveying link as the initial simulation condition. The operating parameters of the coal conveying link at the current moment are obtained through a multi-source data acquisition module and used as the initial state vector. The initial state vector includes the speed, material flow rate, and motor torque of each belt segment, as well as the equipment temperature, vibration displacement, and motor current at each monitoring point. Environmental parameters, including ambient temperature, humidity, and coal bulk density, are also recorded.

[0110] Then, based on the control parameter configuration in the candidate control strategy, the control vector corresponding to the candidate control strategy is parsed and extracted. Each control vector contains the speed command of each coal feeder, the speed command of each belt section, and the gate opening command of each coal feeder, etc. The control vector is constrained to ensure that each control vector is within the allowable range and the single-step change does not exceed the maximum change rate limit.

[0111] Next, a discrete-time stepping algorithm is used to advance the simulation process, i.e., by setting the simulation time step and the total simulation time step; then, within each time step, the control input in the digital twin model is updated to the currently extracted control vector, and the following solution process is executed: After the control input is updated, the material conveying dynamics equations in the corresponding coal conveying link are solved; for example, the differential equation for belt speed is solved using the fourth-order Runge-Kutta method, and the state value at the next moment is obtained by calculating the intermediate values ​​of four slopes and averaging them; for the feeder flow equation, the differential equation for gate opening is solved using the fourth-order Runge-Kutta method, and then the material flow rate is calculated based on the updated gate opening; then, the equipment response equations in the coal conveying link are solved; for example, the differential equation for motor dynamic response is solved using the fourth-order Runge-Kutta method, and after updating... The angular velocity is used to calculate the motor current; the forward Euler method is used to solve the heat balance equation for temperature evolution because the thermal inertia of temperature is relatively large and the change is relatively slow. The temperature at the next moment is equal to the current temperature plus the time step divided by the heat capacity multiplied by the power loss minus the convective heat dissipation power. The energy consumption calculation equation in the coal conveying link is solved simultaneously. For example, the instantaneous power of each motor is equal to the electromagnetic torque multiplied by the angular velocity divided by the motor efficiency. The total power is the sum of the power of all motors and the power of auxiliary equipment. The cumulative energy consumption is calculated using the trapezoidal integral method, which is equal to the cumulative energy consumption at the previous moment plus half of the time step multiplied by the sum of the total power at the previous moment and the total power at the current moment. If there is a fault, the fault characteristic evolution state is updated synchronously. For example, for belt misalignment, the offset is updated; for bearing fault, the vibration amplitude is updated; for abnormal temperature, the temperature rise rate at the fault point is updated.

[0112] By capturing the dynamic response process of the digital twin model during simulation and recording the relevant time-series data of the state variables during the simulation at a fixed sampling frequency, a complete system state evolution trajectory is formed. The recorded state variables include the speed of each belt segment, the material flow rate of each piece of equipment, the temperature of each monitoring point, the vibration displacement of each monitoring point, the current of each motor, and the cumulative energy consumption. The state evolution trajectory is stored in matrix form. At the same time, the time series of fault characteristic residues is recorded. The fault characteristic residues include four components: deviation residue, vibration residue, temperature residue, and current residue. Deviation residue is equal to the absolute value of the current deviation divided by the critical deviation. Vibration residue is equal to the maximum value of the vibration displacement of all monitoring points divided by the normal vibration reference value. Temperature residue is equal to the maximum value of the temperature of all monitoring points exceeding the rated temperature divided by the temperature margin. Current residue is equal to the maximum value of the current of all motors deviating from the rated current divided by the current margin.

[0113] Performance metrics are extracted from the system state evolution trajectory. These extracted metrics include: fault elimination speed, system stability metrics, cumulative energy consumption during the simulation process, and equipment protection level. It should be noted that the fault elimination speed is obtained by finding the moment when the L2 norm of the fault characteristic residual is first less than a preset safety threshold. The fault elimination time is equal to that moment minus the simulation start time. The moment when the L2 norm of the fault characteristic residual is first less than the preset safety threshold indicates that the system fault level has been suppressed to a safe operating level at that corresponding moment.

[0114] The system stability index is obtained by calculating the fluctuation amplitude of each state variable during the simulation process (the fluctuation amplitude is measured by the coefficient of variation, i.e., the standard deviation divided by the mean), and taking the average of the fluctuation amplitudes of all state variables to obtain the overall stability index. At the same time, the overshoot is calculated (the overshoot is the maximum percentage deviation of each belt speed from the target speed).

[0115] The cumulative energy consumption is obtained by directly reading the cumulative energy consumption at the end of the simulation and subtracting the cumulative energy consumption at the beginning of the simulation. At the same time, the energy efficiency ratio is calculated. The energy efficiency ratio is obtained by calculating the total work of material transportation (the time integral of the material flow rate of each segment multiplied by the transportation distance) and dividing it by the energy consumed.

[0116] The equipment protection level includes peak stress (the maximum ratio of all motor torques to rated torque), temperature margin (the minimum value of the actual temperature minus the upper limit of all monitoring points), and fatigue accumulation (using the Miner linear accumulation criterion, i.e., the sum of the actual number of cycles under each stress level and the fatigue life).

[0117] It should be further noted that the solution and index calculation processes involved in the above simulation are all existing technologies, therefore this application will not elaborate on the specific calculation process.

[0118] It should be noted that a digital twin model refers to a three-dimensional simulation model of a coal conveying link constructed based on digital simulation technology. This digital twin model uses a hybrid method that integrates physical mechanism modeling and historical data-driven modeling to reproduce the dynamic operating characteristics, equipment response behavior, and fault evolution patterns of the coal conveying link in high fidelity in digital space. The digital twin model takes multi-source sensing data as input and outputs equipment operating parameters, material flow status, energy consumption change trends, and fault propagation paths. It provides a virtual trial-and-error environment for the simulation verification, optimization and adjustment, and risk assessment of control strategies, ensuring that control commands undergo sufficient feasibility verification and performance prediction before actual issuance. The construction process of the digital twin model is existing technology and will not be elaborated on in this application.

[0119] The extracted fault elimination speed, system stability index, cumulative energy consumption, and equipment protection level are input into a pre-constructed multi-objective optimization function to obtain the scores of each component of the multi-objective optimization function. A comprehensive score for each candidate control strategy is obtained by weighted summation of these scores, and the candidate strategy with the highest comprehensive score is selected as the initial control adjustment strategy. The comprehensive score calculation refers to transforming multi-dimensional performance indicators into a unified, comparable standard through quantitative evaluation. Based on the extracted fault elimination speed, system stability index, cumulative energy consumption, and equipment protection level, the corresponding components of the multi-objective optimization function are obtained. Each component is normalized and mapped to the [0, 1] interval. Then, weight coefficients are dynamically allocated according to the priority requirements of the current operation and maintenance scenario, and the comprehensive score is calculated through weighted summation. The scoring process fully considers the multiple objectives of fault handling: rapidly eliminating faults to ensure production continuity, maintaining stable system operation to avoid secondary disturbances, and simultaneously considering economy and equipment lifespan. The candidate control strategy with the highest comprehensive score was selected as the initial control and adjustment strategy, laying a solid foundation for subsequent refined optimization. The multi-objective optimization function is a multi-dimensional performance quantification model used to comprehensively evaluate the effectiveness of candidate control strategies. By integrating four core evaluation dimensions—fault elimination speed, system stability index, energy consumption, and equipment protection level—the multi-objective optimization function transforms the simulation results of the control strategy in the digital twin model into comparable scalar scores, providing a quantitative decision-making basis for intelligent selection and parameter optimization of control strategies. The construction of the multi-objective optimization function follows the actual needs of coal conveying system operation and maintenance. Specifically, the sub-score for fault elimination speed quantifies the timeliness performance of the control strategy in eliminating fault characteristics; the sub-score for system stability quantifies the fluctuation degree and overshoot amplitude of system operating parameters during the execution of the control strategy; the sub-score for cumulative energy consumption quantifies the energy economy during the execution of the control strategy; and the sub-score for equipment protection level quantifies the impact of the control strategy on the mechanical life and health status of the coal conveying equipment.

[0120] The initial control strategy involves fine-tuning the control parameters to identify key parameters significantly impacting system performance (including belt speed and material flow rate). Gradient optimization algorithms are used to iteratively adjust these key parameters, and the adjustment effect is verified in real-time using a digital twin model. This process continues until the overall score converges or a preset number of iterations is reached. The optimized control strategy is then output as the final adaptive control strategy. Parameter fine-tuning transforms the initial chosen scheme into the optimal configuration for the current scenario. The adjustment process begins by slightly perturbing each control parameter of the initial strategy. Simulation is used to evaluate the impact of parameter changes on the overall score, and the sensitivity gradient is calculated. Then, parameters are prioritized based on their sensitivity, concentrating computational resources on optimizing highly sensitive parameters. Next, gradient optimization algorithms such as sequential quadratic programming or particle swarm optimization are used to search for the optimal parameter combination within the feasible region. After each parameter update, a digital twin model is used for rapid verification to evaluate the overall performance of the new parameter configuration. Finally, when the optimization algorithm converges to a local optimum or reaches the maximum number of iterations, the finely tuned control strategy is output. The parameter adjustment process strictly adheres to the physical constraints and safety boundaries of the coal conveying equipment, ensuring that the optimized parameters remain within the practically feasible range. The optimization algorithm, through an iterative optimization mechanism, systematically explores the parameter space, overcoming the performance limitations of the initial strategy and achieving further improvement in the overall score.

[0121] The generated control and regulation strategy is encapsulated into standardized control instructions, which are then distributed to the control mechanism of the coal conveying link via edge computing nodes. The adaptive control strategy is parsed, converting control parameter configurations into standardized control instructions for specific actuators. These standardized control instructions include equipment identification codes, action type codes, target setpoints, and execution timing markers. The encapsulation process transforms the adaptive control strategy into operation instructions recognizable by the equipment. This process first assigns an equipment identification code to each controlled object, then converts control parameters into standard action types and target values, and finally adds timing markers to define the execution order. The encapsulated instructions use an industrial communication protocol format, including an instruction header, instruction body, and checksum, ensuring reliable transmission. The standardized control instructions are loaded onto the edge computing nodes and distributed to the corresponding control mechanisms in an encrypted manner via industrial fieldbus or wireless communication networks. A handshake confirmation and retransmission mechanism is enabled to ensure accurate instruction delivery. Upon receiving the instructions, the control mechanism performs legality verification and security checks. After successful verification, it drives the actuator to complete the control action and feeds back the execution status to the edge computing node, achieving a closed-loop control system.

[0122] like Figure 2 The figure shown is a schematic diagram of the failure degradation curve. The health index is defined by normalizing the comprehensive vibration, temperature and acoustic characteristics, and the value range is [0, 1]. Among them, 0 indicates that the equipment is completely healthy and 1 indicates that the equipment is completely failed.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0124] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0125] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing, characterized in that, include: The multi-source data acquisition module is used to acquire multi-modal data from the coal conveying link based on the deployed heterogeneous sensor array to obtain multi-source sensing data; The multi-source sensing data includes visual image sequences, device acoustic signals, temperature distribution data, vibration time sequence data, and current waveform data. The data preprocessing module is used to perform environmental interference compensation on the multi-source sensing data to obtain multi-source purification data, and to perform deep collaborative extraction of cross-modal features on the multi-source purification data to obtain a collaborative feature set. The process of obtaining the collaborative feature set includes: Spatiotemporal segmentation and multi-scale decomposition of moving targets within the purified visual image are performed to identify and reconstruct the complete spatial morphological features of the moving targets. Based on wavelet packet decomposition and short-time Fourier transform, joint time-frequency analysis of the purified equipment acoustic signal is performed to identify and extract the fundamental frequency component and fault modulation component of the rotating parts of the equipment. Spatial evolution analysis of thermal anomaly regions is performed on the purified temperature distribution data, and the location and intensity of heat sources are inverted to generate thermal anomaly propagation characteristics. Empirical mode decomposition was performed on the purified vibration time series data, and the distribution characteristics of vibration energy in the time-frequency plane in the vibration time series data were determined based on it. Load fluctuation pattern identification is performed on the purified current waveform data, and load fluctuation characteristics are extracted by combining the load power spectral density of the current waveform data. A mutual information matrix is ​​constructed based on the extracted cross-modal features, a physical constraint correlation model is constructed, and a feature projection transformation is performed on the mutual information matrix based on it to obtain a cooperative feature set; the cross-modal features include complete spatial morphological features, fundamental frequency components and fault modulation components, thermal anomaly propagation features, load fluctuation features, and vibration energy distribution features in the time-frequency plane; The edge decision module is used to build a hierarchical fusion analysis architecture and combine it with collaborative feature sets to identify and deeply analyze abnormal events, thereby obtaining the global operating status of the coal conveying link. The intelligent control and feedback module is used to perform progressive diagnosis and propagation path deduction of the current coal conveying link operation status based on the global operation status, and generate an adaptive control strategy for the coal conveying link through simulation deduction; the adaptive control strategy is converted into control commands and sent to the control mechanism of the coal conveying link.

2. The intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing as described in claim 1, characterized in that, The process of acquiring multi-source purification data includes: Based on the joint analysis of the dark channel prior distribution and spatial frequency attenuation of each visual image in the visual image sequence, the equivalent transmittance of dust concentration in the current environment is obtained; based on the equivalent transmittance of dust concentration, the visual image is decomposed into a multi-scale feature decomposition, dividing the visual image into a dust-free intrinsic layer and a dust scattering layer, and the original visual image is processed based on the dust-free intrinsic layer to obtain the purified visual image. The environmental noise baseline of the collected equipment acoustic signal is dynamically tracked in real time to identify steady-state noise components and construct a time-varying parameter model based on it. An adaptive frequency domain notch filter bank is designed based on the obtained time-varying parameter model and used to suppress noise in the equipment acoustic signal to obtain the purified equipment acoustic signal. Electromagnetic interference features are identified and extracted from the vibration time series data and current waveform data to distinguish between mechanical impact signals and electromagnetic pulse interference signals. Based on a pre-established interference feature library, template matching and adaptive cancellation are performed on the electromagnetic pulse interference signals to obtain purified vibration time series data and current waveform data. Non-uniformity correction and temperature calibration compensation are performed on the temperature distribution data to obtain the purified temperature distribution data; the various data after purification are summarized to obtain the corresponding multi-source purification data.

3. The intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing as described in claim 1, characterized in that, The implementation process of the edge decision module includes: A hierarchical fusion analysis architecture is deployed in the edge computing nodes of the corresponding coal conveying link. The hierarchical fusion analysis architecture includes an edge fast response layer, an edge deep analysis layer, and a cloud collaborative verification layer. The edge fast response layer is used to construct a baseline feature space under normal operation based on cross-modal features, obtain the Mahalanobis distance between it and the corresponding collaborative features, and perform fast discrimination based on the Mahalanobis distance to obtain abnormal event logs; A refined fault diagnosis model is deployed in the edge deep analysis layer, and the collaborative feature set corresponding to the occurrence time in the abnormal event log is input into the refined fault diagnosis model to obtain the output results; The decision evaluation is performed based on the confidence level of the output result. If the confidence level is less than the preset second threshold, the corresponding output result is marked as a low-confidence decision and the cloud collaborative verification layer is triggered. The cloud collaborative verification layer performs correlation analysis of global historical data and expert knowledge base verification, and updates the low-confidence decision based on the verification results. The intermediate and final results of the above analysis are compressed and feature extracted to obtain the global running status.

4. The intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing as described in claim 3, characterized in that, The implementation process of the intelligent control and feedback module includes: Construct a multi-level fault discrimination decision tree for the coal conveying link; based on the various discrimination nodes in the multi-level fault discrimination decision tree, and combined with the collaborative feature set and global operating status, perform fault combination discrimination on the coal conveying link to obtain the output results of the multi-level fault discrimination decision tree; Uncertainty quantification is performed on the output results of the multi-level fault discrimination decision tree to generate the probability distribution of decision confidence and calculate the variance of decision confidence; When the variance of the decision confidence level is greater than the preset confidence threshold, the corresponding output result is labeled with "manual review" and transmitted to the relevant management personnel for manual review; when the variance of the decision confidence level is not greater than the preset confidence threshold, an intelligent monitoring report is generated based on the discrimination results of the corresponding multi-level fault discrimination decision tree. The intelligent monitoring report includes the fault type, fault location, fault severity, fault propagation path, corresponding risk probability, and decision confidence level. An adaptive control strategy is generated based on the severity and type of the fault. The generated adaptive control strategy is then encapsulated into standardized control commands and sent to the control mechanism of the coal conveying link.

5. The intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing as described in claim 4, characterized in that, The root node of the multi-level fault discrimination decision tree is a binary classification of normal and abnormal. Subsequent hierarchical nodes are three types of decision nodes: fault category discrimination node, fault location node, and fault severity assessment node. Each decision node makes a judgment based on different combinations of collaborative feature sets.

6. The intelligent monitoring system for coal conveying system of thermal power plant based on multi-source sensing and edge computing according to claim 4, characterized in that, The process of fault combination identification includes: Based on the fault category discrimination nodes in the multi-level fault discrimination decision tree, modal decomposition is performed on the collaborative feature set to identify and label the modal sources that dominate the fault features, and a modal combination strategy library is established for fault types to determine specific fault types. The fault location node in the multi-level fault discrimination decision tree determines the spatial location of the fault source based on the spatial topology of the abnormal sensor array and the propagation delay analysis of fault characteristics. In a multi-level fault discrimination decision tree, the fault severity assessment node tracks the fault evolution trend of fault features over time to fit a fault degradation curve and predicts the remaining time before the fault features reach a dangerous threshold. Based on the predicted remaining time and fault location, fault propagation path deduction analysis is performed to calculate each fault propagation path. The risk probability is calculated, and the output of a multi-level fault discrimination decision tree is generated.

7. The intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing as described in claim 4, characterized in that, The process of generating an adaptive control strategy includes: Based on the fault type and severity, a set of candidate control strategies is obtained by matching strategies from a preset control strategy library. For each candidate control strategy in the candidate control strategy set, forward simulation is performed using a digital twin model of the coal conveying link to predict the system state evolution trajectory after the execution of the candidate control strategy. Based on the system state evolution trajectory, each candidate control strategy is scored to obtain a comprehensive score for each candidate control strategy, and the candidate control strategy with the highest comprehensive score is used as the initial control adjustment strategy. The parameters of the initial control adjustment strategy are finely adjusted, and the adjusted initial control adjustment strategy is used as the adaptive control strategy.

8. The intelligent monitoring system for coal conveying systems in thermal power plants based on multi-source sensing and edge computing according to claim 2, characterized in that, The process of acquiring purified visual images includes: For the dust-free intrinsic layer, a local window is used to traverse the pixels to obtain the local contrast within each local window, and then mapped to an intuitive contrast distribution heatmap. Based on the contrast distribution heatmap, low-contrast, medium-contrast, and high-contrast regions within the original visual image are identified. Differential enhancement processing is performed on each contrast region. After the enhancement processing is completed, the processed contrast regions are smoothly blended to obtain a purified visual image.

9. The intelligent monitoring system for coal conveying system of thermal power plant based on multi-source sensing and edge computing according to claim 2, characterized in that, The design process of an adaptive frequency domain notch filter bank includes: A short-time Fourier transform is performed on the current analysis frame of the device's acoustic signal to obtain a time-frequency plane representation. The spectral ridges of environmental noise are identified on the time-frequency plane, and the local maxima are located as the noise center frequencies, while the corresponding time-series evolution sequences are obtained simultaneously. The noise center frequencies at future moments are predicted based on the time-series evolution of the noise center frequencies. Notch filters are constructed at each predicted noise center frequency, and all notch filters are cascaded to form an adaptive frequency-domain notch filter bank.

Images