An intelligent operation and maintenance platform for coal-fired power plant storage equipment

Abstract

This invention provides an intelligent operation and maintenance platform for storage equipment in coal-fired power plants, relating to the field of intelligent operation and maintenance technology for coal-fired power plants. It includes a data acquisition layer that collects vibration, temperature, current, video, and dust concentration data using a multi-dimensional sensor array; a data preprocessing layer that performs denoising, interpolation, keyframe extraction, and steady-state filtering on each modal data; a feature extraction layer that generates vibration energy spectrum, temperature and current statistical features, visual semantic features, and concentration frequency domain features; an intelligent diagnosis layer that diagnoses fault types and their confidence levels based on physical mechanism degradation equations and LSTM time-series prediction algorithms; and an operation and maintenance decision-making layer that dynamically generates maintenance work orders, links the power plant information management system and spare parts inventory, and optimizes personnel scheduling paths and spare parts delivery plans through a parallel elite strategy genetic algorithm. This invention significantly improves the intelligence, automation, and accuracy of equipment condition monitoring, fault diagnosis, and maintenance scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for coal-fired power plants, and in particular to an intelligent operation and maintenance platform for storage equipment in coal-fired power plants. Background Technology

[0002] Currently, the operation and maintenance management of storage equipment in coal-fired power plants is still dominated by traditional technical systems. In terms of equipment monitoring, single physical quantity sensors (such as temperature and vibration sensors) combined with fixed threshold alarm mechanisms are commonly used. For example, single-point temperature monitoring devices have sparse sensor deployment and can only achieve basic anomaly detection, making it difficult to effectively detect complex faults in critical equipment such as the stacker-reclaimer's slewing mechanism. Manual inspections rely on offline tools such as handheld infrared thermal imagers and listening sticks, and inspection data is stored through paper records or isolated databases, resulting in fragmented equipment status information. At the operation and maintenance management software level, while mainstream systems such as MAXIMO and SAP PM can achieve work order dispatch and maintenance record management, the lack of standardized data interfaces prevents real-time data interaction with the power plant's DCS control system and supply chain management platform, creating "information silos."

[0003] With the rapid development of the Industrial Internet of Things (IIoT) and big data technologies, the intelligent upgrading of coal-fired power plants is showing a significant trend. Sensor network deployment density is gradually increasing; for example, multi-node monitoring systems use the LoRa protocol to achieve wide-area coverage, but data acquisition remains limited to structured numerical signals. Machine learning algorithms are being introduced into data analysis methods; SVM classifiers can achieve simple fault mode identification, but the accuracy rate for identifying complex faults (such as concurrent gear wear and bearing overheating) is less than 65%. Digital twin technology is gradually being applied to equipment visualization, with typical 3D models of coal conveyor belts, but their functions are mostly focused on static display, lacking accurate simulation and real-time mapping capabilities for dynamic equipment behavior, making it difficult to support in-depth operation and maintenance decisions.

[0004] The existing technological system still has fundamental flaws that urgently need to be addressed. First, at the data fusion level, there is a lack of a unified spatiotemporal alignment mechanism for cross-modal data such as mechanical vibration, electrical parameters, and video streams. For example, the platform only supports structured data access and cannot process unstructured video logs and high-dimensional spectral data. Second, fault diagnosis models rely excessively on manual feature engineering and do not integrate equipment physical degradation mechanisms (such as gearbox wear rate equations), resulting in prediction error rates exceeding 35% under complex operating conditions. Third, maintenance decisions are out of sync with the actual production needs of power plants. Existing solutions only generate maintenance work orders based on equipment status, without considering multi-dimensional constraints such as coal inventory, peak power generation load, and spare parts logistics cycles. This causes more than 30% of maintenance plans to be delayed due to resource conflicts, resulting in unplanned downtime losses. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent operation and maintenance platform for coal-fired power plant storage equipment, which significantly improves the intelligence, automation and precision of coal-fired power plant storage equipment status monitoring, fault diagnosis and maintenance scheduling.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A smart operation and maintenance platform for storage equipment in coal-fired power plants includes:

[0008] The data acquisition layer consists of a multi-dimensional sensor array deployed on the coal conveyor belt drive device, the stacker-reclaimer rotary mechanism, and the coal bunker level gauge. It is used to collect vibration data, temperature data, current data, video stream data, and environmental dust concentration data based on the multi-dimensional sensor array.

[0009] The data preprocessing layer is used to preprocess the vibration data, temperature data, current data, video stream data, and environmental dust concentration data respectively to obtain preprocessed data for each mode.

[0010] The feature extraction layer is used to extract features from the preprocessed data of each mode to obtain the fault features corresponding to each mode.

[0011] The intelligent diagnostic layer is used to diagnose the fault characteristics based on the physical mechanism degradation equation and the LSTM time series prediction algorithm to obtain diagnostic results;

[0012] The operation and maintenance decision-making layer is used to dynamically generate maintenance work orders based on the diagnostic results, and link the power plant information management system and spare parts inventory database to optimize the scheduling path of maintenance personnel and spare parts delivery plan through genetic algorithms.

[0013] Preferably, it further includes:

[0014] An edge computing gateway is used to locally cache and real-time synchronize the vibration data, temperature data, current data, video stream data, and environmental dust concentration data collected by the multi-dimensional sensor array to ensure that the timestamp deviation of the data of each modality does not exceed 10ms.

[0015] Preferably, the data preprocessing layer includes:

[0016] A vibration processing unit is used to denoise the vibration data based on a wavelet threshold denoising method.

[0017] An interpolation point supplementation unit is used to perform linear interpolation point supplementation on the temperature data and the current data based on a sliding window;

[0018] A frame extraction unit is used to extract key frames from the video stream data based on the YOLO-X target detection model.

[0019] The concentration processing unit is used to filter the environmental dust concentration data based on a steady-state filtering method with cubic exponential smoothing.

[0020] The resampling unit is used to perform unified resampling on the multi-dimensional sensor array based on a multimodal time alignment algorithm.

[0021] Preferably, the expression for the cubic exponential smoothing steady-state filtering method is:

[0022]

[0023] in, Let be the dust concentration after filtering at time t. Let be the initial dust concentration at time t. The components are cubic exponentially smoothed, with α as the principal smoothing factor, β as the mixed weighting factor (β = 0.6 when the equipment is running, β = 0.4 when the equipment is shut down), γ as the mutation suppression coefficient, and ΔC. t-1 Let T be the concentration gradient of the previous period, ∈ be the numerical stability constant, δ be the temperature compensation coefficient, and T be the temperature gradient of the previous period. t For real-time ambient temperature, T base The reference temperature is σ, the temperature-dependent bandwidth is I. t This serves as a device start / stop indicator. When the device is running, I... t =1, when the equipment stops, I t =0.

[0024] Preferably, the feature extraction layer includes:

[0025] The vibration extraction unit is used to perform 6-level wavelet packet decomposition on the preprocessed data of vibration modes to obtain 64-dimensional energy spectrum features;

[0026] The temperature and current extraction unit is used to extract the mean, variance, skewness and kurtosis from the preprocessed data of temperature and current modes to obtain 8-dimensional statistical features.

[0027] The image extraction unit is used to extract features from keyframe images of the video modality using a Swin-Transformer encoder to obtain 256-dimensional visual semantic features.

[0028] The concentration extraction unit is used to perform a fast Fourier transform on the preprocessed dust concentration data to obtain the frequency band energy characteristics.

[0029] Preferably, it further includes:

[0030] The feature fusion layer is used to fuse the fault features corresponding to each mode at the feature level through a self-attention mechanism, forming a 512-dimensional fused fault feature.

[0031] Preferably, the network structure of the LSTM time series prediction algorithm includes two LSTM networks of 128 units each and a fully connected regression head.

[0032] Preferably, the diagnostic results include: fault type and confidence level of the fault type; the fault type includes: bearing inner ring fault, outer ring fault, belt misalignment, and stacking arm jamming.

[0033] Preferably, the genetic algorithm adopts a parallel elite strategy, with a population size of 200 and a maximum number of iterations of 300. The Pareto optimal solution set is used to determine the final maintenance plan through a normalized score based on the entropy weight method.

[0034] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0035] This invention provides an intelligent operation and maintenance platform for storage equipment in coal-fired power plants, comprising: a data acquisition layer, consisting of a multi-dimensional sensor array deployed on the coal conveyor belt drive device, the stacker-reclaimer rotary mechanism, and the coal bunker level gauge, used to collect vibration data, temperature data, current data, video stream data, and environmental dust concentration data based on the multi-dimensional sensor array; a data preprocessing layer, used to preprocess the vibration data, temperature data, current data, video stream data, and environmental dust concentration data respectively to obtain preprocessed data for each mode; a feature extraction layer, used to extract features from the preprocessed data for each mode to obtain the fault features corresponding to each mode; an intelligent diagnosis layer, used to diagnose the fault features based on the physical mechanism degradation equation and the LSTM time series prediction algorithm to obtain the diagnosis results; and an operation and maintenance decision layer: dynamically generating maintenance work orders based on the diagnosis results, and linking the power plant information management system and spare parts inventory database, optimizing the maintenance personnel scheduling path and spare parts delivery plan through a genetic algorithm. This invention significantly improves the intelligence, automation, and precision of condition monitoring, fault diagnosis, and maintenance scheduling of storage equipment in coal-fired power plants. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the platform structure provided in an embodiment of the present invention;

[0038] Figure 2 A platform workflow diagram provided for embodiments of the present invention. Detailed Implementation

[0039] 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.

[0040] The purpose of this invention is to provide an intelligent operation and maintenance platform for storage equipment in coal-fired power plants, which significantly improves the intelligence, automation and precision of status monitoring, fault diagnosis and maintenance scheduling of storage equipment in coal-fired power plants.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 and Figure 2 As shown, the present invention provides an intelligent operation and maintenance platform for storage equipment in coal-fired power plants, comprising:

[0043] The data acquisition layer consists of a multi-dimensional sensor array deployed on the coal conveyor belt drive device, the stacker-reclaimer rotary mechanism, and the coal bunker level gauge. It is used to collect vibration data, temperature data, current data, video stream data, and environmental dust concentration data based on the multi-dimensional sensor array.

[0044] The data preprocessing layer is used to preprocess the vibration data, temperature data, current data, video stream data, and environmental dust concentration data respectively to obtain preprocessed data for each mode.

[0045] The feature extraction layer is used to extract features from the preprocessed data of each mode to obtain the fault features corresponding to each mode.

[0046] The intelligent diagnostic layer is used to diagnose the fault characteristics based on the physical mechanism degradation equation and the LSTM time series prediction algorithm to obtain diagnostic results;

[0047] The operation and maintenance decision-making layer is used to dynamically generate maintenance work orders based on the diagnostic results, and link the power plant information management system and spare parts inventory database to optimize the scheduling path of maintenance personnel and spare parts delivery plan through genetic algorithms.

[0048] Specifically, the data acquisition layer consists of a multi-dimensional sensor array deployed on key storage equipment such as coal conveyor belt drive units, stacker-reclaimer slewing mechanisms, and coal bunker level gauges. The sensor array specifically includes: triaxial vibration acceleration sensors installed on the coal conveyor belt drive bearings and reducer housings; temperature sensors distributed across various motors and key transmission nodes; Hall current sensors connected in series with the main power line; industrial cameras installed in the belt conveyor corridor and unloading port; and laser dust concentration sensors deployed in enclosed spaces and dust-generating points. All types of sensors are connected to the distributed edge acquisition unit via wired (e.g., RS485, CAN, Ethernet) or wireless (e.g., LoRa, WiFi) networks. The edge acquisition unit has functions such as data acquisition and temporary storage, time synchronization, and fault self-diagnosis. It performs local caching and preliminary data formatting on the raw data collected from multiple sources, and uploads data such as vibration, temperature, current, video stream, and dust concentration with a unified timestamp to the upper-level data preprocessing layer in real time. Through the above methods, full, continuous and multimodal data collection of core operating parameters of power plant storage equipment was achieved, providing a complete and reliable data foundation for subsequent intelligent diagnosis and operation and maintenance optimization.

[0049] Preferably, it further includes:

[0050] An edge computing gateway is used to locally cache and real-time synchronize the vibration data, temperature data, current data, video stream data, and environmental dust concentration data collected by the multi-dimensional sensor array to ensure that the timestamp deviation of the data of each modality does not exceed 10ms.

[0051] Specifically, in this embodiment, the edge computing gateway integrates an industrial-grade ARM platform (CPU: quad-core Cortex-A53, 1.2GHz; memory: 2GB DDR3; storage: 32GB eMMC) with a multi-port Ethernet switch module, and connects to a multi-dimensional sensor array via RS485, CAN, GigE interfaces, and a MIPI-CSI camera interface. The gateway has a built-in real-time Linux (PREEMPT_RT) operating system running on a data acquisition daemon (DataCollector). This process periodically retrieves raw data from triaxial acceleration, platinum resistance temperature, current Hall effect sensors, 4K industrial camera frame streams, and laser dust sensors based on hardware interrupts or timed scheduling. All raw data is first written to a memory circular queue (with a maximum cache of 30 seconds of data) and simultaneously stored in a local time-series database (InfluxDB, retention period ≥24h) to cope with upstream network jitter or short-term disconnections.

[0052] To ensure cross-modal data timing alignment, the edge computing gateway integrates an IEEE-1588v2 hardware timestamp network card and acts as a PTP slave node to synchronize with the workshop master clock. The clock deviation is hardware-calibrated to an accuracy of ≤1μs. In the DataCollector acquisition process, the driver layer adds a hardware timestamp (TS_PHY) to the raw samples of each sensor at the physical interface. Subsequently, the kernel-mode NTP / PTP compensation service (combined with a sliding window Kalman filtering algorithm to continuously correct local clock drift) generates a service timestamp (TS_APP), which is then written into the data packet header. Through this two-stage time calibration, the final timestamp deviation of all modal data can be stably controlled within ±10ms. The subsequent data preprocessing layer can directly use TS_APP as a unified time reference for resampling and alignment.

[0053] Preferably, the data preprocessing layer includes:

[0054] A vibration processing unit is used to denoise the vibration data based on a wavelet threshold denoising method.

[0055] An interpolation point supplementation unit is used to perform linear interpolation point supplementation on the temperature data and the current data based on a sliding window;

[0056] A frame extraction unit is used to extract key frames from the video stream data based on the YOLO-X target detection model.

[0057] The concentration processing unit is used to filter the environmental dust concentration data based on a steady-state filtering method with cubic exponential smoothing.

[0058] The resampling unit is used to perform unified resampling on the multi-dimensional sensor array based on a multimodal time alignment algorithm.

[0059] Specifically, the vibration processing unit runs on the edge computing gateway and uses a discrete wavelet transform algorithm to decompose the original vibration data into multiple layers. The mother wavelet used is selected from the fourth-order wavelet of the Daubechies wavelet family. After the transform, the noise threshold is automatically calculated based on the statistical characteristics of the coefficients of the highest frequency band, and a soft thresholding strategy is used to threshold the coefficients of each detailed sub-band. Then, the denoised vibration signal is reconstructed through inverse transform. The interpolation and point filling unit updates the window every thirty seconds for the sampling sequences from the temperature and current sensors, scans the missing data points in the window, and interpolates and fills in the missing data points for adjacent valid samples according to a linear relationship. Finally, it outputs time-continuous and uninterrupted temperature and current curves.

[0060] Optionally, the frame extraction unit is deployed on a GPU-accelerated inference server, loading a pre-trained YOLO-X target detection model to perform full-resolution detection on each frame of the original video stream. When a coal conveyor belt, stacker-reclaimer, or other monitored target is detected with a confidence level exceeding 50%, the frame is immediately marked as a keyframe and stored. The concentration processing unit uses a cubic exponential smoothing algorithm to update the smoothing coefficient and trend coefficient in real time on the laser dust sensor sampling data, effectively filtering out short-term fluctuations to obtain a stable concentration curve. The resampling unit aligns the denoised vibration, interpolated temperature, current, keyframe timestamps, and smoothed dust concentration to a frequency of 100 degrees per second according to a unified time base. Low sampling rate data is upsampled through cubic spline interpolation, and high sampling rate data is downsampled after anti-aliasing filtering, thereby forming a multimodal consistent synchronous time-series dataset.

[0061] Preferably, the expression for the cubic exponential smoothing steady-state filtering method is:

[0062]

[0063] in, Let be the dust concentration after filtering at time t. Let be the initial dust concentration at time t. The components are cubic exponentially smoothed, with α as the principal smoothing factor, β as the mixed weighting factor (β = 0.6 when the equipment is running, β = 0.4 when the equipment is shut down), γ as the mutation suppression coefficient, and ΔC. t-1 Let T be the concentration gradient of the previous period, ∈ be the numerical stability constant, δ be the temperature compensation coefficient, and T be the temperature gradient of the previous period. t For real-time ambient temperature, T base The reference temperature is σ, the temperature-dependent bandwidth is I. t This serves as a device start / stop indicator. When the device is running, I... t =1, when the equipment stops, I t =0.

[0064] Optionally, this invention addresses the issues of sudden dust concentration changes and signal instability during operation of warehousing equipment by proposing a cubic exponential smoothing filter method that integrates adaptive adjustment based on equipment start-up and shutdown states. This method innovatively introduces operating condition identifiers on top of traditional smoothing algorithms, assigning different values ​​to the smoothing factor and mixed weights to achieve the dual objectives of rapidly tracking concentration changes during operation and strongly suppressing environmental drift during shutdown. Simultaneously, a sudden change suppression mechanism is superimposed, adaptively amplifying the suppression coefficient when a sudden increase or decrease in concentration is detected, effectively suppressing spurious fluctuations caused by probe response delays. Furthermore, real-time temperature compensation is integrated to correct the temperature bandwidth of the original concentration data, ensuring stable output even under seasonal and diurnal temperature variations. This multi-factor coupled filtering design improves the smoothness of the concentration curve while retaining sensitivity to real-world dust emergencies.

[0065] As an example, all smoothing factors and compensation parameters are obtained through a two-step process: on-site calibration and historical data backtesting. First, based on dust and operating condition logs from the past three months, a grid search and cross-validation are performed on the main smoothing factor and mixed weights with the goal of minimizing error. The mutation suppression coefficient is determined by analyzing the data difference before and after a typical sudden dust event to determine the optimal suppression factor. The numerical stability constant is set to a sufficiently small fixed value to ensure that the algorithm does not divide by zero at extremely low concentrations. The temperature compensation coefficient, reference temperature, and influence bandwidth are obtained through stratified testing in an environmental simulation chamber to ensure that the compensation curves for different temperature zones closely match on-site changes. Equipment start / stop indicators are directly taken from the operating condition signals of the control system, enabling automatic switching between running and stopped states within the software. All the above parameters are configured in the edge computing gateway initialization file and can be automatically or manually adjusted online according to the daily operation of the equipment to meet the filtering accuracy requirements under different operating conditions.

[0066] Preferably, the feature extraction layer includes:

[0067] The vibration extraction unit is used to perform 6-level wavelet packet decomposition on the preprocessed data of vibration modes to obtain 64-dimensional energy spectrum features;

[0068] The temperature and current extraction unit is used to extract the mean, variance, skewness and kurtosis from the preprocessed data of temperature and current modes to obtain 8-dimensional statistical features.

[0069] The image extraction unit is used to extract features from keyframe images of the video modality using a Swin-Transformer encoder to obtain 256-dimensional visual semantic features.

[0070] The concentration extraction unit is used to perform a fast Fourier transform on the preprocessed dust concentration data to obtain the frequency band energy characteristics.

[0071] Optionally, the vibration extraction unit, based on Python and the PyWavelets library, performs six-level wavelet packet decomposition on the preprocessed vibration signal acquired per second, using the Daubechies-4 mother wavelet. After each level of decomposition, the energy values ​​of each sub-band signal are calculated and summarized, and finally, the energy values ​​generated by the six levels are concatenated into a 64-dimensional energy spectrum feature vector. The temperature and current extraction unit utilizes the NumPy and SciPy libraries to calculate the first-order mean, second-order variance, third-order skewness, and fourth-order kurtosis of the interpolated temperature and current sequences within each time window, concatenating them into an 8-dimensional statistical feature vector. The statistical results at each time step are normalized to ensure uniformity of feature dimensions and stable convergence of subsequent model training.

[0072] Specifically, in this embodiment, the image extraction unit is deployed on a GPU-accelerated deep learning inference server, loading a Swin-Transformer-based encoder model. This model is pre-trained on the ImageNet dataset and then fine-tuned for keyframes in this system. The system first divides each keyframe into fixed-size image blocks, maps them to a high-dimensional embedding space, and then extracts visual semantic features through multi-level local window attention and cross-window interaction modules. Finally, a 256-dimensional image feature vector is obtained through global average pooling. The concentration extraction unit uses Python's FFT algorithm and window functions to perform frequency domain transformation on the smoothed dust concentration time-series data, extracting the sum of squared amplitudes within each preset frequency band as a frequency band energy index. These frequency band energy values ​​are then concatenated in ascending order to form the final frequency domain feature vector for subsequent multimodal fusion.

[0073] Preferably, it further includes:

[0074] The feature fusion layer is used to fuse the fault features corresponding to each mode at the feature level through a self-attention mechanism, forming a 512-dimensional fused fault feature.

[0075] In this embodiment, the feature fusion layer adopts a self-attention-based structure. First, the 64-dimensional energy spectrum features from the vibration extraction unit, the 8-dimensional statistical features from the temperature and current extraction units, the 256-dimensional visual semantic features from the image extraction unit, and the frequency domain features from the concentration extraction unit are mapped into a unified 128-dimensional embedding vector through four sets of linear projection networks. These four embedding vectors are arranged in a fixed order into a vector sequence of length four, and each vector is also embedded with a corresponding trainable modality identifier to preserve the identity information of each modality. This sequence is then input into a Transformer encoder module, which consists of alternating multi-head self-attention sublayers and feedforward network sublayers. The multi-head self-attention uses eight parallel attention branches, each capturing the intra-modal and inter-modal correlations in a 16-dimensional subspace; the feedforward network consists of two linear transformation layers and one nonlinear activation layer to enhance feature representation with contextual information mixing. External residual connections and layer normalization are applied to each sublayer to ensure network training stability, ultimately outputting four 128-dimensional fused vectors.

[0076] To obtain the final 512-dimensional fused fault features, the four 128-dimensional vectors mentioned above are directly concatenated in modal order to form a high-dimensional vector. This high-dimensional vector preserves the diversity of each modality and includes interaction information between modalities, providing unified features for the subsequent diagnostic model input. In specific implementation, the linear projection and self-attention modules can call standard components of existing deep learning frameworks, and dropout layers are added between sub-layers to prevent overfitting. The parameters of the fusion layer are jointly optimized with labeled fault and health data during the joint training phase with subsequent networks. During online inference, the fusion layer is deployed on edge devices, receiving features from each modality in real time and processing them through projection and encoder, outputting 512-dimensional fused features for use by subsequent LSTM-based health prediction and fault classification modules.

[0077] Preferably, the network structure of the LSTM time series prediction algorithm includes two LSTM networks of 128 units each and a fully connected regression head.

[0078] Specifically, the steps of the offline modeling stage of the intelligent diagnostic layer in this embodiment are as follows:

[0079] (1) Data preparation: Collect no less than 12 months of historical multimodal preprocessed data (vibration, temperature, current, video keyframes, dust concentration); generate samples according to a 60s sliding window and a 30s step size, and the labels are uniformly corrected by manual inspection records and fault logs.

[0080] (2) Construction of physical mechanism degradation equation: For the key transmission chain, the gear wear rate equation is established: dw / dt=k·F·v / (σH·HB); where w is the wear amount, k is the material constant, F is the load, v is the sliding speed, σH is the contact stress limit, and HB is the Brinell hardness; the equation is discretized into the state space form xk+1=Ak·xk+Bk·uk+wk, which is convenient for subsequent Kalman filtering fusion with LSTM output.

[0081] (3) Network training: The network structure in this embodiment is as follows: the input dimension is consistent with the output of the feature extraction layer (512), containing two LSTM layers of 128 units each and a fully connected regression head. The loss function is mean squared error (MSE) + 1e-4 L2 regularization; the optimizer is Adam, with an initial learning rate of 1e-3, and cosine annealing is used. The training stops after 10 consecutive rounds of no improvement on the validation set; the final model predicts a health index MAE ≤ 0.05 on the validation set.

[0082] (4) The LSTM inference subgraph is quantized into INT8 and compiled into an edge computing gateway executable file using ONNXRuntime. The physical degradation equation and the Kalman filter constant matrix are stored in Redis, supporting online hot updates.

[0083] Furthermore, the steps of the online diagnostic stage of the intelligent diagnostic layer in this embodiment are as follows:

[0084] (1) Input fault features. The feature extraction layer outputs a 512-dimensional fused feature vector Fk every 1 second. Fk is sent to the message queue of the intelligent diagnosis layer.

[0085] (2) Take the feature sequence {Fk-59…Fk} of the most recent 60s as the input of the sliding window. Complete the forward inference and output a 1×1 health prediction value hLSTM,k∈[0,1].

[0086] (3) Physical mechanism degradation calculation: real-time reading of DCS operating conditions (torque, speed, oil temperature, etc.) constitutes the control variable uk. The health prediction value hPHY,k is calculated based on the discretized degradation equation.

[0087] (4) Fusion estimation (Kalman filtering), the specific steps of which are as follows:

[0088] State prediction:

[0089] Variance prediction: Pk|k-1=Ak-1·Pk-1|k-1·Ak-1T+Q

[0090] Gain calculation: Kk=Pk|k-1·(Pk|k-1+R)-1

[0091] Status Update:

[0092] Variance update: Pk|k=(I-Kk)·Pk|k-1

[0093] The final unified health index is

[0094] (5) Fault type determination: Fk and HI_k are concatenated and fed into the Softmax classification head to output the confidence vector Pk = [p1, p2, p3, p4] for the four types of faults. If max(Pk) ≥ 0.5, the corresponding fault is considered to have occurred; otherwise, the status is "pending confirmation".

[0095] (6) Diagnostic results are published. (HI_k, Pk) is published to the operations and maintenance decision-making level via the MQTT topic "diag_out" in JSON format. An alarm event is triggered when HI_k < 0.6 or max(Pk) ≥ 0.5, and the result is written to the PostgreSQL historical database.

[0096] (7) Adaptive update: the diagnostic accuracy rate of the past day is calculated every 24 hours; if it is <90%, the encrypted gradient is uploaded to the cloud federated learning server, the incremental weight is issued after the cloud completes parameter aggregation, and the edge LSTM inference model is automatically hot-replaced.

[0097] Preferably, the diagnostic results include: fault type and confidence level of the fault type; the fault type includes: bearing inner ring fault, outer ring fault, belt misalignment, and stacking arm jamming.

[0098] In this embodiment, the intelligent diagnostic layer first inputs the 512-dimensional fused fault features into a time-series prediction module composed of two layers of long short-term memory networks. Each network unit uses 256-dimensional hidden states and is configured with a dropout layer to suppress overfitting. The module updates its internal memory and output state simultaneously when receiving fused features at each time step. When the preset time window ends, the time-series prediction module outputs the hidden state vector at the last time step. This vector is processed by a linear mapping network and a normalized activation function to generate a four-dimensional diagnostic output vector. The four output nodes correspond to bearing inner race fault, bearing outer race fault, belt misalignment, and stacker arm jamming, respectively. The output value of each node represents the confidence level of the corresponding fault type. The system selects the fault type according to the maximum confidence principle and transmits the selected fault type and its confidence level together as the final diagnostic result to the operation and maintenance decision layer. The diagnostic model is trained end-to-end on a historical multimodal time-series dataset labeled with four types of faults, using cross-entropy as the loss function and the Adam optimization algorithm. After training, it can quickly infer from the real-time fusion features collected online, achieving efficient identification and confidence assessment of the four types of faults.

[0099] Preferably, the genetic algorithm adopts a parallel elite strategy, with a population size of 200 and a maximum number of iterations of 300. The Pareto optimal solution set is used to determine the final maintenance plan through a normalized score based on the entropy weight method.

[0100] In this embodiment, the operation and maintenance decision-making layer first generates several alternative maintenance strategies for each fault type and its confidence level, encoding them as chromosome sequences. Each chromosome contains information such as the maintenance task order, personnel allocation, and spare parts retrieval plan. The genetic algorithm adopts a parallel elitist strategy, maintaining several subpopulations in parallel on a multi-core edge server. Each subpopulation has a size of 200, with a maximum iteration count of 300 generations. During population initialization, diverse solutions are generated based on the equipment location and personnel skill matrix. The selection operation adopts a tournament strategy, the crossover operation prioritizes two-point crossover to preserve the integrity of task blocks, the mutation operation randomly swaps maintenance tasks or adjusts the spare parts source, and the elitist strategy retains the chromosomes with the highest fitness in each generation for parallel replication to the next generation.

[0101] After the algorithm converges or reaches its iteration limit, a Pareto optimal solution set is extracted from all subpopulations and subjected to multi-index normalization. Based on four indicators—total maintenance time, personnel load balancing, spare parts delivery cost, and downtime loss—the entropy weight method is used to calculate the weights of each indicator: first, the entropy values ​​of each solution set's indicator distribution are statistically analyzed; then, weights are assigned based on the amount of information reflected by the entropy values; finally, each Pareto solution is scored using a weighted sum method. The system automatically selects the chromosome with the highest score as the final maintenance plan, balancing efficiency and cost.

[0102] Based on the aforementioned maintenance plan, the operations and maintenance decision-making level calls the power plant information management system interface to dynamically generate digital maintenance work orders. These orders include equipment number, fault type and confidence level, maintenance time window, required spare parts list, and estimated working hours, and are synchronized to the work order management module. Simultaneously, the system sends a spare parts pre-reservation request to the spare parts inventory database, checks inventory, and automatically locks the required spare parts. If inventory is insufficient, the system can also trigger emergency procurement or transfer from nearby warehouses to ensure the smooth execution of maintenance tasks.

[0103] During the maintenance execution phase, the decision-making team further optimized personnel scheduling routes and spare parts delivery plans. Based on the geographical coordinates of the maintenance task, the location of available personnel on-site, and vehicle capacity, the system invokes a lightweight route planning engine to generate the shortest or fastest delivery route and sends the scheduling results to mobile terminals. Technicians, following the work order instructions, carry spare parts to the site along the optimized route, while the distribution center dispatches and delivers spare parts according to the same route plan. The entire process is tracked, verified, and reported in real time, achieving an optimal balance between maintenance efficiency and resource utilization.

[0104] The beneficial effects of this invention are as follows:

[0105] (1) This invention collects multi-source data such as vibration, temperature, current, video and dust concentration of key storage equipment such as coal conveyor belts and stacker-reclaimers in real time and comprehensively through a multi-dimensional sensor array. Compared with traditional single-point monitoring methods, it realizes all-round dynamic perception of the operating status of key equipment, which greatly improves the accuracy and real-time performance of intelligent monitoring of storage equipment.

[0106] (2) This invention uses advanced data preprocessing and feature extraction technology to denoise, align and model deep features of data of different modalities, which can effectively filter out environmental interference, improve the quality of input data, and provide a solid data foundation for subsequent fault diagnosis and health assessment.

[0107] (3) This invention proposes an intelligent diagnostic method that combines a physical mechanism degradation model with an LSTM time-series prediction model. It combines the high interpretability of the physical model with the ability of the data-driven model to capture complex nonlinear faults, thereby achieving high-precision prediction of equipment operating status and automatic identification of multiple fault types, effectively improving the reliability and foresight of the diagnosis.

[0108] (4) This invention links the operation and maintenance decision-making layer with the power plant MIS system and spare parts inventory system, and uses a genetic algorithm to achieve global optimization of maintenance work orders, personnel and spare parts scheduling, which significantly reduces equipment failure response time and maintenance costs, and improves the intelligence and automation level of warehouse equipment management.

[0109] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0110] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An intelligent operation and maintenance platform for storage equipment in a coal-fired power plant, characterized in that, include: The data acquisition layer consists of a multi-dimensional sensor array deployed on the coal conveyor belt drive device, the stacker-reclaimer rotary mechanism, and the coal bunker level gauge. It is used to collect vibration data, temperature data, current data, video stream data, and environmental dust concentration data based on the multi-dimensional sensor array. The data preprocessing layer is used to preprocess the vibration data, temperature data, current data, video stream data, and environmental dust concentration data respectively to obtain preprocessed data for each mode. The feature extraction layer is used to extract features from the preprocessed data of each mode to obtain the fault features corresponding to each mode. The intelligent diagnostic layer is used to diagnose the fault characteristics based on the physical mechanism degradation equation and the LSTM time series prediction algorithm to obtain diagnostic results; The operation and maintenance decision-making layer is used to dynamically generate maintenance work orders based on the diagnostic results, and link the power plant information management system and spare parts inventory database to optimize the scheduling path of maintenance personnel and spare parts delivery plan through genetic algorithms. The data preprocessing layer includes: A vibration processing unit is used to denoise the vibration data based on a wavelet threshold denoising method. An interpolation point supplementation unit is used to perform linear interpolation point supplementation on the temperature data and the current data based on a sliding window; A frame extraction unit is used to extract key frames from the video stream data based on the YOLO-X target detection model. The concentration processing unit is used to filter the environmental dust concentration data based on a steady-state filtering method with cubic exponential smoothing. A resampling unit is used to perform unified resampling on the multi-dimensional sensor array based on a multimodal time alignment algorithm; The expression for the triple exponential smoothing steady-state filtering method is: in, Let be the dust concentration after filtering at time t. Let be the initial dust concentration at time t. It is a cubic exponentially smoothed component. The main smoothing factor As a mixed weighting factor, when the device is running, When the equipment stops, , This is the mutation suppression coefficient. For the previous periodic concentration gradient, It is the numerical stability constant. This is the temperature compensation coefficient. For real-time ambient temperature, As the reference temperature, Bandwidth is affected by temperature. This serves as a start / stop indicator for the equipment; it is used when the equipment is running. When the equipment stops, ; The construction of physical mechanism degradation equations includes: For the critical transmission chain, the gear wear rate equation is established as: dw / dt=k·F·v / (σH·HB); where w is the wear amount, k is the material constant, F is the load, v is the sliding speed, σH is the contact stress limit, and HB is the Brinell hardness.

2. The intelligent operation and maintenance platform for coal-fired power plant storage equipment according to claim 1, characterized in that, Also includes: An edge computing gateway is used to locally cache and real-time synchronize the vibration data, temperature data, current data, video stream data, and environmental dust concentration data collected by the multi-dimensional sensor array to ensure that the timestamp deviation of the data of each modality does not exceed 10ms.

3. The intelligent operation and maintenance platform for coal-fired power plant storage equipment according to claim 1, characterized in that, The feature extraction layer includes: The vibration extraction unit is used to perform 6-level wavelet packet decomposition on the preprocessed data of vibration modes to obtain 64-dimensional energy spectrum features; The temperature and current extraction unit is used to extract the mean, variance, skewness and kurtosis from the preprocessed data of temperature and current modes to obtain 8-dimensional statistical features. The image extraction unit is used to extract features from keyframe images of the video modality using a Swin-Transformer encoder to obtain 256-dimensional visual semantic features. The concentration extraction unit is used to perform a fast Fourier transform on the preprocessed dust concentration data to obtain the frequency band energy characteristics.

4. The intelligent operation and maintenance platform for coal-fired power plant storage equipment according to claim 1, characterized in that, Also includes: The feature fusion layer is used to fuse the fault features corresponding to each mode at the feature level through a self-attention mechanism, forming a 512-dimensional fused fault feature.

5. The intelligent operation and maintenance platform for coal-fired power plant storage equipment according to claim 1, characterized in that, The network structure of the LSTM time series prediction algorithm includes two LSTM networks, each with 128 units, and a fully connected regression head.

6. The intelligent operation and maintenance platform for coal-fired power plant storage equipment according to claim 1, characterized in that, The diagnostic results include: fault type and confidence level of the fault type; the fault types include: bearing inner ring fault, outer ring fault, belt misalignment and stacking arm jamming.

7. The intelligent operation and maintenance platform for coal-fired power plant storage equipment according to claim 1, characterized in that, The genetic algorithm employs a parallel elite strategy, with a population size of 200 and a maximum number of iterations of 300. The Pareto optimal solution set is used to determine the final maintenance plan through normalized scoring based on the entropy weight method.

Images