Intelligent inspection system and method for coal conveyor belt of thermal power unit based on artificial intelligence

Abstract

The present invention discloses an intelligent inspection system and method for the coal conveyor belt of a thermal power unit based on artificial intelligence, which relates to the field of artificial intelligence and machine learning, including: collecting multi-source data, dividing the coal conveyor belt into a grid coordinate system, and assigning an initial risk score to each grid through an artificial intelligence model to obtain a preliminary inspection path, analyzing the time variation trend of the coal flow velocity through an artificial intelligence attention mechanism, processing the preliminary inspection path, marking high-risk areas, generating an optimized inspection path, and fusing multi-source data to obtain a comprehensive feature set; the present invention analyzes the time variation trend of the coal flow velocity through the attention mechanism, achieves the optimization of the inspection path according to the real-time coal flow trend and spatial correlation, avoids the blindness of manual experience and the lag of static design, and ensures the efficient allocation of inspection resources.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and machine learning, and in particular to an artificial intelligence-based intelligent inspection system and method for coal conveyor belts of thermal power units. Background Art

[0002] As a core component for material transportation in coal-fired power plants, the monitoring and maintenance of coal conveyor belts in thermal power units are directly related to production efficiency, safety, and stability. In recent years, with the advancement of industrial automation and intelligent technologies, coal conveyor belt inspection technology has gradually evolved from traditional manual visual inspections to automated monitoring methods based on sensors and data analysis.

[0003] While existing technologies have made significant progress in coal conveyor belt inspection, many challenges remain, such as dynamic adaptability and defect recognition accuracy. First, traditional inspection methods often rely on fixed thresholds or manually preset rules, which cannot accurately capture transition states or complex defects, resulting in high rates of missed detections and false positives. Furthermore, traditional inspection route planning relies primarily on static designs or manual experience, lacking the ability to optimize in real time, resulting in very low resource utilization efficiency. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an artificial intelligence-based intelligent inspection method for the coal conveyor belt of a thermal power unit to solve the problem that traditional inspection methods cannot accurately capture transition states or complex defects and lack the ability to optimize in real time.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an artificial intelligence-based intelligent inspection method for a coal conveyor belt of a thermal power unit, which comprises:

[0008] Collect multi-source data, divide the coal conveyor belt into a grid coordinate system, and use the artificial intelligence model to assign an initial risk score to each grid to obtain a preliminary inspection route;

[0009] The artificial intelligence attention mechanism is used to analyze the temporal variation trend of coal flow velocity, process the preliminary inspection path, mark high-risk areas, generate an optimized inspection path, and fuse multi-source data to obtain a comprehensive feature set.

[0010] The optimized inspection route is sent to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization light data and classifies defects at each grid point using a polarization light model combined with a comprehensive feature set. It then generates a defect report and sends it to the thermal power plant control center.

[0011] As a preferred solution of the artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power units described in the present invention, the artificial intelligence model is built with recurrent neural networks and graph neural networks as the basic architecture, and outputs the initial risk score of each grid point after training is completed.

[0012] As a preferred solution of the artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power plants of the present invention, the method comprises: assigning priorities in descending order according to the initial risk score of each grid point;

[0013] Get the preliminary inspection path based on the priority.

[0014] As a preferred solution of the artificial intelligence-based intelligent inspection method for the coal conveyor belt of a thermal power unit described in the present invention, it utilizes an artificial intelligence attention mechanism based on a recurrent neural network to receive real-time data from a coal flow velocity sensor, analyzes the time series change trend of each grid point, adjusts the initial risk score of each grid point, divides the high-risk areas, and generates an optimized inspection path.

[0015] As a preferred solution of the artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power units described in the present invention, the weights of multi-source data are adjusted, and the multi-source data are fused through weighted average fusion to obtain a comprehensive feature set.

[0016] As a preferred solution of the artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power units described in the present invention, the adjustment of the weights of multi-source data refers to adjusting the weights of multi-source data through preset weight adjustment rules and dynamic correction based on the stability of multi-source data.

[0017] As a preferred solution of the artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power units described in the present invention, the optimized inspection path is sent to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization data, classifies defects at each grid point using a polarization model combined with a comprehensive feature set, and generates a defect report that is sent to the thermal power plant control center. The method specifically includes the following steps:

[0018] The drone, equipped with a multispectral polarization imaging system, executes the optimized inspection route and transmits the collected three-band polarized light data to the polarization light model to obtain the defect feature vector.

[0019] The defect feature vector and the comprehensive feature set are fused by weighted average to obtain the fused feature set;

[0020] The fused feature set is classified using a classification model, and a defect report is finally generated and transmitted to the thermal power plant control center.

[0021] In a second aspect, the present invention provides an artificial intelligence-based intelligent inspection system for coal conveyor belts of thermal power plants, comprising:

[0022] The path generation module collects multi-source data, divides the coal conveyor belt into a grid coordinate system, and assigns an initial risk score to each grid through an artificial intelligence model to obtain a preliminary inspection path;

[0023] The path optimization module uses an artificial intelligence attention mechanism to analyze the temporal variation trend of coal flow speed, process the preliminary inspection path, mark high-risk areas, generate an optimized inspection path, and fuse multi-source data to obtain a comprehensive feature set.

[0024] The inspection module sends the optimized inspection route to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization light data and classifies defects at each grid point using a polarization light model combined with a comprehensive feature set. It then generates a defect report and sends it to the thermal power plant control center.

[0025] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the intelligent inspection method for the coal conveyor belt of a thermal power unit based on artificial intelligence as described in the first aspect of the present invention is implemented.

[0026] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent inspection method for the coal conveyor belt of a thermal power unit based on artificial intelligence as described in the first aspect of the present invention.

[0027] The beneficial effects of the present invention are as follows: by combining an artificial intelligence model of recurrent neural networks and graph neural networks, analyzing temporal trends and spatial correlations, the limitations of a single threshold are avoided, and the transitional states and complex risks can be captured, thereby improving the accuracy of initial defect screening. Dynamic analysis of transitional states and complex trends through the attention mechanism increases sensitivity to potential defects, making up for the static shortcomings of manual rules. Furthermore, the integration of multidimensional data captures complex defects, avoiding the defects of ignoring superimposed states due to a single threshold, and improving recognition accuracy. Furthermore, the attention mechanism analyzes the temporal variation trend of coal flow velocity, achieving optimization of inspection paths based on real-time coal flow trends and spatial correlations, avoiding the blindness of manual experience and the lag of static design, and ensuring efficient allocation of inspection resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 2 is a diagram of an intelligent inspection system for coal conveyor belts of thermal power units based on artificial intelligence in this embodiment.

[0030] Figure 2 Flowchart of the intelligent inspection method for coal conveyor belts of thermal power units based on artificial intelligence in this embodiment.

[0031] Figure 3 Schematic diagram of data fusion of the intelligent inspection method for coal conveyor belts of thermal power units based on artificial intelligence in this embodiment.

[0032] Figure 4 This is a defect optimization diagram of the intelligent inspection method for the coal conveyor belt of a thermal power unit based on artificial intelligence in this embodiment. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0034] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0035] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0036] In this embodiment, refer to Figures 1 to 4 This embodiment provides an artificial intelligence-based intelligent inspection method for a coal conveyor belt of a thermal power plant, comprising the following steps:

[0037] S1. Collect multi-source data and divide the coal conveyor belt into a grid coordinate system.

[0038] The specific steps include:

[0039] The specific parameters of the coal conveyor belt at the thermal power plant are determined, such as a length of 100 meters and a width of 1.5 meters, and the operating environment is an open-air or semi-enclosed coal bunker. Based on these specific parameters of the coal conveyor belt, a sensor deployment plan is developed, with 20 sets of sensors installed every 5 meters to cover the entire length of the belt.

[0040] After deciding to install a group of sensors every 5 meters, the installation of sensors began, including high-definition visible light cameras, infrared cameras, integrated multi-band cameras (also known as multispectral polarization imaging systems, which complete a three-band polarized light scan every 2 seconds to generate multi-layer images for detecting the reflection and scattering characteristics of the belt surface and interior, especially sub-surface defects, and are installed on inspection drones), coal flow velocity sensors, voiceprint sensors and environmental monitors.

[0041] Data collection begins using the installed sensors. A coal flow velocity sensor collects real-time coal flow velocity values ​​once per second. A high-definition visible light camera captures high-definition images of the coal conveyor belt surface (a continuous video stream) at 30 frames per second. An infrared camera captures thermal images of the belt surface once per second. A voiceprint sensor collects vibration waveforms and noise decibels every 10 milliseconds. An integrated multi-band camera captures three-band polarized light images (multi-layer image data) every two seconds. An environmental monitor collects dust concentration, humidity, and vibration intensity, updating every minute. Edge computing nodes receive real-time data from all sensors via industrial Ethernet, forming a continuous multi-source data stream. For example, at a specific moment, a coal flow velocity of 2.3 m / s and a dust concentration of 500 μg / m3 are recorded.

[0042] S1.1. The edge computing nodes are divided into a grid coordinate system based on the length and width of the coal conveyor belt. Taking a 100-meter-long and 1.5-meter-wide coal conveyor belt as an example, it is divided into a 100×5 grid coordinate system, with each grid on the horizontal axis being 1 meter and each grid on the vertical axis being 0.3 meters, for a total of 500 grid points. Each grid point is bound to the data of the nearest sensor group. For example, (50,3) corresponds to the 50-meter third width segment.

[0043] S2. Assign an initial risk score to each grid through the artificial intelligence model to obtain a preliminary inspection path.

[0044] The specific steps include:

[0045] First, prepare to train the AI ​​model. Pre-training preparations include: Collect historical coal flow velocity values, such as the speed values ​​recorded per second over the past six months, covering various conditions such as normal operation, sudden coal surges, and no-load conditions. Collect historical environmental data such as dust concentration, humidity, and vibration intensity to simulate different operating conditions. Collect coal conveyor belt status labels. The actual belt status, as recorded by manual inspections or equipment logs, such as coal dust accumulation, belt deviation, or wear at a certain location during a certain time period, serves as a supervisory signal for the AI ​​model. Organize historical coal flow velocity values, historical environmental data, and coal conveyor belt status labels by timestamp and location (e.g., grid coordinates (50,3)) to form a time series and spatially distributed sample set.

[0046] The AI ​​model used in this solution is a hybrid model combining a recurrent neural network (RNN) and a graph neural network (GNN). RNNs excel at processing time series data and can capture temporal trends in coal flow velocity, such as sudden increases or sustained slowdowns. GNNs, on the other hand, are better suited for analyzing spatial relationships and can identify associations between grid points, such as whether high flow rates at (50,3) and (51,4) form a localized high-load area.

[0047] The sample set is divided into a training set (80%) and a validation set (20%). For example, the training set contains 5 months of data, and the validation set contains 1 month of data. The training objective is to ensure that the AI ​​model predicts a risk score for each grid point, ensuring that the predicted risk score for each grid point is as close as possible to the manually labeled true risk level. For example, the predicted risk score for (50,3) is 0.8 when stacked. The AI ​​model parameters are adjusted iteratively. For example, each training round uses a 5-minute window of coal flow velocity sequences and corresponding grid states as input to optimize the AI ​​model's sensitivity to temporal changes and spatial distribution. Finally, the AI ​​model's performance is tested using a validation set. For example, by inputting a specific period of historical data, the AI ​​model is tested to determine whether it correctly identifies high-risk areas (such as (50,3) with a sudden increase in coal flow) and low-risk areas (such as (80,1) with no load). If misjudgments are detected, such as rating a normal area as high-risk, the AI ​​model weights are adjusted until the AI ​​model accuracy reaches above 90%. Finally, the trained model parameters are saved and deployed to the edge computing node.

[0048] Preferably, the AI ​​model of the present invention combines recurrent neural networks and graph neural networks. By training on historical data, it learns to identify the temporal variations and spatial distribution of coal flows. The recurrent neural network captures temporal trends and indicates potential accumulation risks, while the graph neural network analyzes spatial correlations to increase the accuracy of risk scoring. Furthermore, the AI ​​model of the present invention not only relies on a single threshold (e.g., speed > 2 m / s), but also integrates temporal and spatial features to avoid misjudgments, for example, distinguishing between stable operation and unstable fluctuations.

[0049] S2.1. The trained AI model receives real-time coal flow velocity values ​​from the coal flow velocity sensor, for example, measurements taken once per second across all 500 grid points, such as 2.3 m / s at (50,3) and 0.4 m / s at (80,1). The real-time coal flow velocity data is organized by time and spatial location, forming an input sequence. The AI ​​model processes the coal flow velocity at each grid point as a time series, for example, extracting the velocity record for (50,3) over the past 10 minutes, which shows an increase from 1.5 m / s to 3 m / s. The RNN component of the AI ​​model analyzes this sequence and identifies trends. For example, if the velocity is consistently above 2 m / s or increases rapidly, this indicates an increase in coal flow load, potentially leading to accumulation, and is marked as a high-risk mode. If the velocity is consistently below 0.5 m / s, such as 0.4 m / s at (80,1) for 10 minutes, this indicates an unloaded state, potentially causing wear, and is marked as a low-risk mode requiring attention. This time window can be dynamically adjusted, for example, to 5 or 15 minutes based on the frequency of coal flow fluctuations. Customize according to specific needs.

[0050] The GNN portion of the AI ​​model also processes the spatial relationships between grid points, for example, checking whether the 2.3 m / s at (50,3) correlates with the 2.8 m / s at the adjacent grid point (51,4). For example, if multiple adjacent grid points (e.g., (50,3), (51,4), and (52,3)) simultaneously display high velocities, this indicates a localized area of ​​concentrated coal flow and a high risk of accumulation. The AI ​​model will increase the risk score for these adjacent grid points. However, if a region (e.g., (80,1) to (85,1)) displays consistently low and isolated velocities, this indicates a uniform distribution of unloaded cargo and a low risk. The AI ​​model will decrease the risk score. Spatial analysis considers the distance and connectivity between grid points. For example, a velocity difference of less than 0.5 m / s between adjacent points is considered a consistent pattern.

[0051] The artificial intelligence model integrates the results of temporal changes and spatial distribution, and outputs a risk score for each grid point, which is divided into three types: high-risk mode, low-risk mode, and intermediate mode.

[0052] A high-risk pattern refers to a risk score greater than or equal to 0.8. For example, if the velocity at grid point (50,3) suddenly increases and the adjacent points have high flow simultaneously, the risk score is set to 0.8, indicating accumulation or congestion risk.

[0053] Low-risk mode means that the risk score is less than or equal to 0.2. For example, if the grid point (80,1) has a long period of low speed and no adjacent anomalies, the risk score is set to 0.2, indicating no-load or normal operation.

[0054] Risk scores between 0.3 and 0.6 represent intermediate patterns, such as those where the speed fluctuates between 0.5 and 2 m / s with no apparent spatial clustering. In these cases, the risk score will be adjusted based on specific trends. Adjustment based on specific trends means that the risk score is not fixed at a value between 0.3 and 0.6, but is instead dynamically adjusted based on the specific changing trends in the coal flow velocity. Specific trends refer to patterns within a time series, such as whether the speed is gradually increasing, decreasing, or fluctuating erratically, as well as the duration and magnitude of pattern changes within the time series. The purpose of these adjustments is to make the score more accurately reflect potential risks, rather than simply applying a static range. This is because when the risk score is between 0.3 and 0.6, the coal conveyor is very likely in a transitional state, potentially indicating normal operation or a precursor to a problem. Fixed scores cannot accurately reflect this uncertainty, so dynamic adjustments based on trends are necessary.

[0055] The adjustment logic is as follows: Trend 1: The speed gradually increases but does not exceed the limit. If the speed at a grid point (50,3) slowly increases from 0.5 m / s to 1.8 m / s over 5 minutes, and the speeds at adjacent grid points also increase slightly (for example, 1.3 m / s), the AI ​​model will deem the coal flow load to be increasing. While not yet at a high risk (>2 m / s), the risk is slightly higher, so the score may be adjusted to 0.5 or 0.6, indicating a need for some attention. Trend 2: The speed stabilizes at the median. If the speed at (50,3) remains stable at around 1.2 m / s for an extended period, fluctuating within a range of only ±0.2 m / s, and the adjacent grid points are similar, the AI ​​model will deem the operation stable and without obvious issues. The risk score may be adjusted to 0.3 or 0.4, indicating a lower risk. Trend 3: The speed fluctuates erratically. If the velocity at (50,3) frequently jumps between 0.8 and 1.6 m / s, and its spatial distribution is irregular (large speed differences between adjacent points), the AI ​​model may deem instability. While not serious, it warrants caution and may assign a score of 0.5, which is somewhere in the middle. Trend 4: Approaching the boundary but not breaking it. If the velocity at (50,3) occasionally approaches 2 m / s (for example, 1.9 m / s) but quickly falls back, with a uniform spatial distribution, the AI ​​model will adjust the risk score based on the speed and duration of the fallback. For example, a rapid fallback would result in a score of 0.4, while a sustained approach might rise to 0.6, indicating a slightly higher potential risk.

[0056] Ideally, for intermediate modes, the AI ​​model of the present invention dynamically adjusts the score based on specific trends. For example, if the speed at (50,3) increases from 0.5 m / s to 1.8 m / s, the score is adjusted to 0.6, and if it stabilizes at 1.2 m / s, the score is adjusted to 0.3. This flexible adjustment reflects the transitional nature of the coal conveyor belt's state, ensuring that the risk score is more closely aligned with the actual risk. It can automatically adapt to different operating conditions and improve the precision and practicality of the risk score.

[0057] Based on the risk score output, a priority list (also known as a preliminary inspection route) is automatically generated to guide the frequency and sequence of inspections by inspection equipment (such as drones) along the belt. In layman's terms, grid points are prioritized based on their risk scores and assigned corresponding inspection intervals. For example, high-risk grid points require more frequent inspections, while low-risk grid points can be inspected less frequently.

[0058] It is further explained that compared with fixed thresholds, manual rules or single-dimensional methods, this solution has comprehensive improvements in dynamics, intelligence, adaptability and real-time performance, significantly reducing the missed detection rate (such as the potential risks of intermediate modes) and false detection rate (such as misjudgment of normal fluctuations), providing more efficient and intelligent protection for the safe operation of coal conveyors in thermal power plants.

[0059] S3. Analyze the temporal variation trend of coal flow velocity through the artificial intelligence attention mechanism, process the preliminary inspection path, mark high-risk areas, and generate an optimized inspection path.

[0060] The specific steps include:

[0061] An artificial intelligence attention mechanism module is trained based on the recurrent neural network of the artificial intelligence model, which is specifically responsible for dynamically analyzing the time variation trend of coal flow speed and optimizing it based on the preliminary inspection path.

[0062] An attention layer is added to the recurrent neural network to highlight key moments in the time series. For example, when analyzing a 10-minute sequence at (50, 3), the AI ​​attention mechanism assigns higher weight to the moment when the speed suddenly increases to 3 meters per second, rather than treating the entire sequence equally. The attention mechanism simulates the human ability to focus on shifting priorities. The input is a time series of coal flow speed at each grid point, such as a 10-minute record at (50, 3); the output is a trend score (ranging from 0 to 1) reflecting the risk level of the dynamic change, for example, a sudden increase in the trend is rated 0.8, while a slowing trend is rated 0.2.

[0063] The AI ​​attention mechanism module is trained through supervised learning. The training goal is to predict the trend score of each grid point, making the predicted trend score of each grid point as close as possible to the actual trend label. For example, the target for the sudden increase trend at (50,3) is 0.8, and the target for the slow trend at (80,1) is 0.2.

[0064] The trained AI attention mechanism module receives real-time data from the coal flow velocity sensor, such as measurements taken once per second, such as 2.3 m / s at (50,3) and 0.4 m / s at (80,1). The AI ​​attention mechanism module organizes the real-time data from the coal flow velocity sensor into a time series. For example, extracting the velocity records for (50,3) over the past 10 minutes, it finds an increase from 1.5 m / s to 3 m / s. It then analyzes the changing trends and adjusts the initial risk score: If the velocity is consistently above 2 m / s or increases rapidly (such as at (50,3)), this indicates an increased coal flow load, potentially leading to accumulation, and the initial risk score is adjusted to 0.8. If the velocity is consistently below 0.5 m / s (such as 0.4 m / s at (80,1) for 10 consecutive minutes), this indicates an unloaded state, and the initial risk score is adjusted to 0.2. If the velocity is intermediate (such as fluctuating between 0.5 and 2 m / s), the module adjusts the risk score based on the specific pattern, such as a slow increase to 1.8 m / s.

[0065] Adjusting based on specific patterns means that when the coal flow velocity is in the mid-range of 0.5-2 m / s, the AI ​​attention mechanism module does not simply assign a value. Instead, it analyzes the specific pattern of velocity change (such as rising, stable, and fluctuating) and dynamically adjusts the trend score or inspection frequency based on the trend, amplitude, and duration of the change pattern in the coal flow velocity time series. For example, if the velocity at grid point (60,2) slowly increases to 1.8 m / s, the score is adjusted from the initial value to 0.6. This approach ensures accurate assessment of transition states and improves the targeted and efficient optimization of inspection paths.

[0066] While analyzing temporal trends, the AI ​​attention mechanism module combines spatial correlation to examine speed changes at adjacent grid points. For example, if the speed at (51,4) increases synchronously to 2.8 m / s, consistent with the sudden increase at (50,3), it is identified as a localized high-load area. The initial risk score for (51,4) is then adjusted to 0.75, strengthening the high-risk assessment.

[0067] Optimally, the AI ​​attention mechanism module accurately identifies and marks high-risk areas by analyzing the temporal trends and spatial correlations of coal flow velocity. This targeted adjustment allows inspection routes to prioritize potential problem areas, such as areas at risk of accumulation or blockage.

[0068] Based on temporal variation and spatial correlation analysis, the AI ​​attention mechanism module dynamically adjusts the initial inspection path. The risk priority of each grid point is reassessed: For (50,3), because the initial risk score is adjusted to 0.8 and the load is continuously high, the high-risk status of grid point (50,3) is confirmed (i.e., a high-risk area), and the 15-minute inspection frequency is maintained or strengthened. For (80,1), the initial risk score is adjusted to 0.2 and there are no abnormalities. The priority of grid point (80,1) is lowered, and the inspection frequency is adjusted from 20 minutes to 30 minutes. For intermediate-mode grid points (such as (60,2, with an initial risk score of 0.6), the frequency of drone inspections is adjusted based on the trend of the temporal variation pattern of the coal flow velocity (such as rising or stable), for example, fine-tuning from 20 minutes to 18 minutes or 22 minutes.

[0069] After adjusting priorities, optimized inspection routes are planned for high-risk areas. For example, the flight time for an inspection drone from (1,1) to (50,3) is calculated to be 24.5 seconds at a speed of 2 meters per second, a distance of 49 meters. Including hovering and acquisition time, the plan is to complete a high-risk area cycle every 14 minutes. The optimized inspection route starts at (1,1) and covers (50,3) and (51,4) in sequence. This adjustment process is set to dynamic refresh mode, reanalyzing the latest coal flow velocity data every 5 minutes. For example, if (80,1) maintains a low velocity, it will continue to be skipped. If (50,3) and (51,4) maintain high flow, they will be prioritized. The optimized inspection route is sent to the drone via the wireless network to ensure real-time execution.

[0070] Optimally, by dynamically adjusting inspection frequency, resources are concentrated in high-risk areas while redundant inspections in low-risk areas are reduced. For example, high-risk areas are inspected approximately four times per hour (60 minutes ÷ 14 minutes = 4.3 times), while low-risk areas are inspected only twice per hour (60 minutes ÷ 30 minutes = 2 times), significantly improving resource utilization. Furthermore, the inspection route is refreshed every five minutes based on the latest coal flow velocity data, ensuring that inspections are always synchronized with the current status. This achieves resource conservation while ensuring coverage of high-risk areas.

[0071] S4. Fuse multi-source data to obtain a comprehensive feature set.

[0072] The specific steps include:

[0073] Extract high-definition images of the coal conveyor belt surface collected by a high-definition visible light camera, thermal images of the belt surface collected by an infrared camera, and voiceprint data collected by a voiceprint sensor.

[0074] Environmental data collected by environmental monitors is read. This data reflects environmental conditions. For example, a dust concentration of 600 μg / m³ is above the normal range of 500 μg / m³; a vibration intensity of 6 m / s² indicates severe machine operation. Environmental conditions can affect the reliability of sensor data. For example, high dust levels can blur visible light images (referring to high-definition images of the coal conveyor belt surface captured by a high-definition visible light camera), while high vibrations can amplify voiceprint signals. Reading this environmental data provides a basis for adjusting weights in the next step, ensuring that the fusion results are adapted to actual operating conditions.

[0075] After reading the environmental data, the weights of each sensor are adjusted according to the adjustment rules.

[0076] For example, the adjustment rules are as follows: the default weighting is 33% for each sensor (visible light, infrared, and voiceprint each account for 1 / 3). In high-dust scenarios, with dust concentrations of 600 μg / m³ (>500 μg / m³), visible light images may be distorted by occlusion, so their weighting is reduced to 20%. Infrared thermal images are less affected by dust, so their weighting is increased to 40%, and the voiceprint weighting is increased to 35%. In high-vibration scenarios, with vibration intensities of 6 m / s², voiceprint signals are more likely to indicate mechanical anomalies, so their weighting is further increased to 40%, infrared signals are reduced to 35%, and visible light signals remain at 20%.

[0077] For example, in the current environment (dust 600 μg / m3, vibration 6 m / s²), the weights are adjusted to 20% for visible light, 35% for infrared, and 40% for voiceprint.

[0078] Preferably, the weights are dynamically adjusted to make the fusion results more dependent on reliable data, such as giving priority to infrared and voiceprints under high dust conditions, and highlighting voiceprints under high vibration conditions, to avoid unreliable data interfering with the evaluation.

[0079] Check the data stability of each sensor and analyze fluctuations in sensor data within a short time window (1 minute). For example, if the frequency of a (50,3) voiceprint fluctuates irregularly (possibly due to noise or sensor failure), the data is considered unstable. If an anomaly is detected, the weight of the unreliable data is reduced (for example, from 40% to 10% for voiceprints) and the weight of other stable data is increased (for example, from 35% to 45% for infrared and 20% for visible light). By checking stability, the influence of abnormal data is eliminated, ensuring that the fusion results are based on reliable data and improving the accuracy of the assessment.

[0080] S4.1. Normalize the weighted data (normalize the weighted data to the range of 0-1) and then fuse them using weighted averaging. For example, the multimodal data of (50,3) is integrated. The weighted data of (50,3) is 20% visible light (surface texture features, such as wear marks), 45% infrared (high temperature point, 48°C, high weight, large contribution), and 10% voiceprint (vibration pattern, abnormally high, low weight, small impact). After fusion, a multidimensional feature vector (also known as a comprehensive feature set) is generated, describing the status of (50,3) as "abnormally high temperature and increased vibration," indicating potential deviation or roller problems. This comprehensive feature set is updated every second to ensure it reflects the latest status.

[0081] Preferably, the fusion of multi-source data forms a comprehensive description of the coal conveyor belt status, which can reveal more complex problems than a single data, and achieve comprehensive and accurate monitoring of the coal conveyor belt status, high reliability evaluation, real-time dynamic update, efficient inspection and maintenance coordination and resource utilization optimization, significantly improving the intelligence level of inspection.

[0082] S5. Send the optimized inspection route to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization data and classifies defects at each grid point using a polarization model combined with a comprehensive feature set. It then generates a defect report and sends it to the power plant control center.

[0083] The specific steps include:

[0084] S5.1. Transmit the optimized inspection route to two inspection drones. The coal conveyor belt is 100 meters long and divided into 500 grid points (a 100×5 grid system). To improve efficiency, the inspection task is divided into two sections: the first drone covers the first 50 meters (grids (1,1) to (50,5), focusing on the high-risk point (50,3)), and the second drone covers the last 50 meters (grids (51,1) to (100,5), focusing on (51,4)). This partitioning strategy ensures that the two drones work in parallel, covering the entire belt. After receiving the optimized inspection route, the two drones are activated. Their built-in navigation system adjusts their direction based on the inspection route coordinates to ensure they follow the optimized route.

[0085] After arriving at a high-risk grid point, the inspection drone hovers, activating a multispectral polarization imaging system and beginning to scan the coal conveyor belt. The multispectral polarization imaging system integrates three cameras: a visible light lens (for capturing light reflected from the coal conveyor belt surface), a near-infrared lens (for detecting shallow scattered signals), and a short-wave infrared lens (for penetrating into the coal conveyor belt). A polarization filter is also added to measure the polarization angle and intensity of light, enhancing the ability to detect surface features (such as textures and cracks) and internal defects. The three cameras output images separately, forming a multi-layered data structure (i.e., three-band polarization data). For example, a (50,3) visible light image reveals wear, a near-infrared image highlights scattering anomalies, and a short-wave infrared image reveals underlying problems.

[0086] S5.2 To analyze three-band polarization data, a polarization model is established. First, three-band polarization data is extracted from historical inspection records. This data is collected by a multispectral polarization imaging system onboard a drone, covering both normal and abnormal conditions. Data is collected for at least six months to one year, including thousands of image sets, such as near-infrared polarization images and visible light surface reflectance images at a (50,3) grid point. Defect annotations are also collected for each image set, and environmental conditions at the time of acquisition, such as dust concentration (e.g., 600 μg / m³) and vibration intensity (e.g., 6 m / s²), are recorded as auxiliary information to help the polarization model understand the environmental impact of the polarization data. The three-band polarization data is normalized to the 0-1 range to prevent magnitude differences between different bands from affecting training. The data is then divided into a training set (80%, e.g., eight months of data) and a validation set (20%) to ensure independence between training and evaluation.

[0087] A convolutional neural network (CNN) is used as the underlying architecture because CNNs excel at extracting spatial features from images and are suitable for processing multispectral polarized light data. The input layer receives a three-band polarized light image, for example, three layers of data (visible light, near-infrared, and short-wave infrared) in a (50,3) image. Each layer is a two-dimensional matrix (e.g., 256×256 pixels), with an input dimension of three channels. The convolutional layer extracts reflective properties (such as surface texture) and scattering properties (such as near-infrared intensity enhancement) through multiple layers of convolution and pooling operations. For example, the convolution kernel captures anomalous regions in a (50,3) near-infrared image, generating a feature map. Prior knowledge, such as the photon scattering equation (based on Monte Carlo simulations), is embedded into the polarized light model to enhance its ability to interpret the scattered light signal and help distinguish the scattering patterns of internal cracks from those of surface foreign matter. Finally, the output layer outputs a defect feature vector, for example, an internal stress or crack signal in a (50,3) image, containing defect type and quantitative parameters (such as scattering intensity).

[0088] The training set is fed into the polarization model in batches, for example, 32 images per batch. The polarization model processes the (50, 3) three-band data, using a convolutional layer to extract features, such as the near-infrared scattering enhancement region, and outputs a predicted crack signal.

[0089] Cross-entropy loss is used to measure the difference between the predicted features (e.g., a crack depth of 4 mm) and the true annotation. For example, if the predicted depth is 3.8 mm and the true depth is 4 mm, the error is calculated and back-propagated. The Adam optimizer is used to adjust the polarization model parameters, with the learning rate initially set to 0.001 and decaying dynamically with each iteration. Weights are updated with each iteration to bring the prediction closer to the true defect. Scattering constraints are also incorporated into the loss function, such as a positive correlation between near-infrared scattering intensity and crack depth, to ensure that the polarization model output conforms to optical laws. This is continued until the loss converges. Finally, validation data is input, such as a polarization image with a (50, 3) angle, to verify that the polarization model correctly predicts internal cracks with a depth of 4 mm. The accuracy (e.g., defect type accuracy) and mean squared error (e.g., depth prediction error) are calculated. The optimal parameters are saved until the polarization model reaches 90% accuracy—for example, 90% of the grid points have the correct defect type and a depth error of less than 0.5 mm.

[0090] Optimally, the acquisition of three-band polarized light data through a multispectral polarization imaging system, combined with precise analysis of polarized light models, significantly improves the comprehensiveness and accuracy of defect detection. This multidimensional data analysis is more comprehensive than traditional single-sensor methods and provides a reliable basis for diagnosing complex defects (such as deviation or roller problems).

[0091] S5.3. After the trained polarization model outputs a defect feature vector (for example, an internal stress or crack signal at (50,3)), it is fused with the comprehensive feature set through weighted averaging. The defect feature vector provides internal defect clues (such as enhanced near-infrared scattering), and the comprehensive feature set provides external state information (high temperature and vibration). The two are combined to generate a more comprehensive feature vector (also known as the fused feature set). For example, the fusion result of (50,3) may indicate that high temperature and vibration have caused increased internal stress, and the polarization anomaly is further confirmed as a crack risk. This multi-source information integration improves the accuracy of defect judgment and avoids the misjudgment that may occur with a single data source (such as polarization alone).

[0092] The classification model uses a support vector machine. The fused feature set is input into the support vector machine to determine the defect type and severity of each grid point. Before this, a defect library needs to be preset. This defect library is a classification set that contains common coal conveyor belt defect types, such as "internal cracks", "surface wear", "deviation", "fiber breakage", "accumulation blockage", etc. Each defect type is associated with a set of typical features, which are used for comparison and classification matching between the support vector machine and the fused feature set. For example, for internal cracks, the polarized light scattering intensity increases, accompanied by high temperature and vibration; for surface wear, the visible light reflectivity decreases, and there is no obvious internal scattering change; for deviation, the vibration frequency is abnormal, and the polarized light shows local stress concentration.

[0093] Further explanation: The weighted averaging method that integrates a comprehensive feature set with polarization data improves the reliability of defect assessment. The comprehensive feature set incorporates external conditions (such as the high temperature of 48°C and increased vibration in the (50,3) pattern), while the polarization data provides internal clues (such as a scattering intensity of 0.8). The resulting fused feature set comprehensively reflects the synergistic effects of high temperature, vibration, and cracks, avoiding the potential bias in judgment caused by single modal data and more accurately reproducing actual operating conditions.

[0094] The fused feature set (for example, the abnormally high temperature with increased vibration and polarized light scattering at (50, 3)) is compared one by one with the feature templates in the defect library using Euclidean distance to find the most matching defect feature type. The type with the highest matching score is selected as the preliminary classification result, for example, (50, 3) is determined to be an internal crack.

[0095] After feature matching, the support vector machine assesses the severity based on preset defect thresholds, further confirming the defect classification and quantifying it. For example, internal cracks are confirmed based on a scattering intensity > 0.7 and a depth > 1 mm. The depth is estimated using scattering value mapping. (Mapping estimation establishes a mapping relationship between scattering intensity and depth based on historical data. Scattering intensities are measured on crack samples of known depths. For example, a scattering intensity of 0.5 indicates a depth of approximately 1 mm, a scattering intensity of 0.7 indicates a depth of approximately 2 mm, and a scattering intensity of 0.8 indicates a depth of approximately 4 mm.) For surface wear, a reflectivity drop of > 10% without high temperature is considered mild, and a drop of > 30% is considered moderate. (Preset defect thresholds can be customized based on individual needs and actual business requirements.)

[0096] For example, at (50,3), the scattering intensity is 0.8 (>0.7), and the depth is estimated to be 4 mm (>1 mm), indicating internal cracks with a severity of moderate to severe. At (80,1), the reflectivity decreases by 15%, but there are no scattering anomalies, indicating surface wear with a severity of mild.

[0097] Each grid point's severity is checked against a preset severity threshold (e.g., a crack depth of 3 mm). For example, at (50, 3), a crack depth of 4 mm exceeds the severity threshold, triggering an automatic alert at the edge computing node. However, for the minor wear at (80, 1), since it does not meet the verification threshold, it is simply recorded without triggering an alert, demonstrating the flexibility of priority management.

[0098] Further explanation: Using support vector machines and a pre-set defect library for classification enables rapid defect quantification and priority management. For (50,3), a scattering intensity of 0.8 and a depth of 4 mm exceeding the 3 mm threshold trigger an early warning, promptly indicating a moderate to severe crack risk. Meanwhile, mild wear (a 15% decrease in reflectivity) at (80,1) is only recorded without an alert, demonstrating flexibility and resource optimization.

[0099] S5.4. Generate a defect report, including the defect location (e.g., grid (50,3), type: internal crack, severity: 4 mm depth), and recommended repair time: belt replacement within 72 hours. The report is formatted as text with coordinates. This detailed report provides maintenance personnel with a clear action guide, ensuring that defect handling is based on evidence.

[0100] The analysis results and warning status of all grid points are combined into a complete data packet. For example, the internal crack at (50, 3), with a depth of 4 mm, triggering a warning, and the surface wear at (80, 1), with minor severity and no warning, are packaged together with a timestamp and priority tag. This data is then transmitted to the power plant control center via industrial Ethernet.

[0101] This embodiment also provides an artificial intelligence-based intelligent inspection system for coal conveyor belts of thermal power plants, including:

[0102] The path generation module collects multi-source data, divides the coal conveyor belt into a grid coordinate system, and assigns an initial risk score to each grid through an artificial intelligence model to obtain a preliminary inspection path;

[0103] The path optimization module uses an artificial intelligence attention mechanism to analyze the temporal variation trend of coal flow speed, process the preliminary inspection path, mark high-risk areas, generate an optimized inspection path, and fuse multi-source data to obtain a comprehensive feature set.

[0104] The inspection module sends the optimized inspection route to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization light data and classifies defects at each grid point using a polarization light model combined with a comprehensive feature set. It then generates a defect report and sends it to the thermal power plant control center.

[0105] This embodiment also provides a computer device, which is suitable for the case of an intelligent inspection method for coal conveyor belts of thermal power units based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the intelligent inspection method for coal conveyor belts of thermal power units based on artificial intelligence proposed in the above embodiment.

[0106] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0107] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the artificial intelligence-based intelligent inspection method for the coal conveyor belt of a thermal power unit as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0108] In summary, the present invention combines the artificial intelligence models of recurrent neural networks and graph neural networks to analyze time trends and spatial correlations, thus avoiding the limitations of a single threshold, being able to capture transition states and complex risks, and improving the accuracy of initial defect screening. Dynamically analyzing transition states and complex trends through the attention mechanism improves sensitivity to potential defects, making up for the static deficiencies of manual rules, and integrating multidimensional data to capture complex defects, thus avoiding the defects of ignoring superimposed states due to a single threshold, and improving recognition accuracy. In addition, the attention mechanism analyzes the temporal variation trend of coal flow velocity, achieving the optimization of the inspection path based on real-time coal flow trends and spatial correlations, avoiding the blindness of manual experience and the lag of static design, and ensuring efficient allocation of inspection resources.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power plants, characterized by: include, Collect multi-source data, divide the coal conveyor belt into a grid coordinate system, and use the artificial intelligence model to assign an initial risk score to each grid to obtain a preliminary inspection route; The artificial intelligence attention mechanism is used to analyze the temporal variation trend of coal flow velocity, process the preliminary inspection path, mark high-risk areas, generate an optimized inspection path, and fuse multi-source data to obtain a comprehensive feature set. The optimized inspection route is sent to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization data and classifies defects at each grid point using a polarization model combined with a comprehensive feature set. A defect report is then generated and sent to the power plant control center. The AI ​​model is built on a recurrent neural network and graph neural network architecture, and outputs an initial risk score for each grid point after training. Generating an optimized inspection path refers to utilizing an artificial intelligence attention mechanism based on a recurrent neural network to receive real-time data from coal flow velocity sensors, analyze the time series change trend of each grid point, adjust the initial risk score of each grid point, and divide high-risk areas to generate an optimized inspection path; The optimized inspection route is sent to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization data and classifies defects at each grid point using a polarization model combined with a comprehensive feature set. A defect report is generated and sent to the thermal power plant control center. The specific steps include the following: The drone, equipped with a multispectral polarization imaging system, executes the optimized inspection route and transmits the collected three-band polarized light data to the polarization light model to obtain the defect feature vector. The defect feature vector and the comprehensive feature set are fused by weighted average to obtain the fused feature set; The fused feature set is classified using a classification model, and a defect report is finally generated and transmitted to the thermal power plant control center.

2. The method for intelligent inspection of coal conveyor belts of thermal power plants based on artificial intelligence according to claim 1, characterized in that: According to the initial risk score of each grid point, priority is assigned in descending order; Get the preliminary inspection path based on the priority.

3. The method for intelligent inspection of coal conveyor belts of thermal power plants based on artificial intelligence according to claim 2, characterized in that: The weights of multi-source data are adjusted, and the multi-source data are fused through weighted average fusion to obtain a comprehensive feature set.

4. The method for intelligent inspection of coal conveyor belts of thermal power plants based on artificial intelligence according to claim 3, characterized in that: The adjusting of the weights of the multi-source data refers to adjusting the weights of the multi-source data by using preset weight adjustment rules and dynamic correction based on the stability of the multi-source data.

5. An artificial intelligence-based intelligent inspection system for coal conveyor belts of thermal power plants, based on the artificial intelligence-based intelligent inspection method for coal conveyor belts of thermal power plants according to any one of claims 1 to 4, characterized in that: include, The path generation module collects multi-source data, divides the coal conveyor belt into a grid coordinate system, and assigns an initial risk score to each grid through an artificial intelligence model to obtain a preliminary inspection path; The path optimization module uses an artificial intelligence attention mechanism to analyze the temporal variation trend of coal flow speed, process the preliminary inspection path, mark high-risk areas, generate an optimized inspection path, and fuse multi-source data to obtain a comprehensive feature set. The inspection module sends the optimized inspection route to a drone equipped with a multispectral polarization imaging system. The drone collects three-band polarization light data and classifies defects at each grid point using a polarization light model combined with a comprehensive feature set. It then generates a defect report and sends it to the thermal power plant control center.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the artificial intelligence-based intelligent inspection method for the coal conveyor belt of a thermal power plant are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the artificial intelligence-based intelligent inspection method for the coal conveyor belt of a thermal power plant are implemented.

Images