Bucket wheel machine operation image monitoring method and system based on neural network

Abstract

The invention discloses a bucket wheel machine operation image monitoring method and system based on a neural network, and relates to the technical field of image data processing, and the method comprises the steps: obtaining a bucket wheel machine monitoring region and a monitoring time interval, and obtaining an ith bucket wheel machine operation image after the ith monitoring time interval, obtaining an ith pixel feature map based on the image pre-training neural network and the ith bucket wheel machine operation image, wherein each pixel of the pixel feature map corresponds to a feature quantity; obtaining a preset processing number n, and obtaining a to-be-processed feature image set; acquiring a monitoring weight corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point and the to-be-processed feature map set, and acquiring an observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weight and the to-be-processed feature map set; and obtaining a monitoring result according to the observation value. The method has the advantages of real-time accurate monitoring, dynamic adaptive optimization and multi-dimensional fault distinguishing.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a method and system for monitoring the operation images of a bucket wheel stacker-reclaimer based on a neural network. Background Art

[0002] In industrial scenarios, as a core material transfer equipment, the bucket wheel stacker-reclaimer undertakes the tasks of stacking and reclaiming bulk materials such as coal and ore. Due to being exposed to harsh working conditions such as high dust, strong vibration, and drastic changes in temperature and humidity for a long time, key components of the bucket wheel stacker-reclaimer, such as the traveling mechanism, slewing bearing, and bucket wheel drive device, are extremely prone to mechanical failures such as gear wear, bearing failure, and structural cracking. Traditional monitoring means mainly rely on contact detection devices such as vibration sensors and oil fluid analyzers, and there are many technical bottlenecks.

[0003] In recent years, although the non-contact monitoring technology based on machine vision theoretically has the advantage of full-domain coverage, it still faces severe challenges in practical applications. First, the operation images taken under dynamic lighting conditions have regional overexposure / underexposure, resulting in insufficient recognition accuracy of the traditional edge detection algorithm for the bucket wheel tooth profile. The image blur caused by pulverized coal dust makes the feature matching error exceed 15 pixels, which cannot meet the millimeter-level monitoring requirements of the bucket wheel trajectory. More critically, the existing image processing methods adopt the mean filtering strategy with a fixed time window, and fail to effectively distinguish the temporal feature differences between normal vibration and abnormal displacement of the equipment. In the detection of conveyor belt deviation, the instantaneous deformation caused by material accumulation is misjudged as a mechanical failure. And when there is a conflict between the vision monitoring system and the control signal, the central control platform often needs to intervene manually for discrimination, and the average response delay reaches more than 8 minutes, which completely cannot meet the requirements of industrial scenarios for real-time perception of the health status of equipment. Summary of the Invention

[0004] Aiming at the defects in the prior art, the present invention provides a method and system for monitoring the operation images of a bucket wheel stacker-reclaimer based on a neural network.

[0005] A method for monitoring the operation images of a bucket wheel stacker-reclaimer based on a neural network includes: obtaining the monitoring area and monitoring time interval of the bucket wheel stacker-reclaimer, and in real time obtaining the i-th bucket wheel operation image corresponding to the monitoring area of the bucket wheel stacker-reclaimer after the i-th monitoring time interval, and obtaining the i-th pixel feature map based on the image pre-trained neural network and the i-th bucket wheel operation image, and each pixel of the pixel feature map respectively corresponds to the feature quantity output by the image pre-trained neural network; obtaining the preset processing quantity n, obtaining n continuously monitored pixel feature maps and using them as the to-be-processed feature map set; obtaining the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set, and obtaining the observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weights, and the to-be-processed feature map set; obtaining the monitoring result according to the observation value.

[0006] Optionally, obtaining the monitoring result according to the observed value includes: obtaining the reference threshold range corresponding to the monitoring area of the bucket wheel stacker-reclaimer; obtaining the deviation degree between the observed value and the reference threshold range, and judging whether the operation state of the bucket wheel stacker-reclaimer is abnormal according to the deviation degree; constructing a cumulative judgment result mechanism in the time dimension, and establishing a progressive alarm level according to the abnormal duration and the occurrence times.

[0007] Optionally, obtaining the monitoring weight corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set includes: obtaining the number of monitoring time intervals closest to the current time point according to the current time point; obtaining the time suppression coefficient and the time sensitivity coefficient according to the monitoring area of the bucket wheel stacker-reclaimer; obtaining the monitoring weight corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the monitoring time interval, the number of monitoring time intervals, the time suppression coefficient, and the time sensitivity coefficient.

[0008] Optionally, the weight model in obtaining the monitoring weight corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the monitoring time interval, the number of monitoring time intervals, the time suppression coefficient, and the time sensitivity coefficient is expressed as: where, w j is the monitoring weight corresponding to the j-th pixel feature map in the to-be-processed feature map set, α is the time suppression coefficient, β is the time sensitivity coefficient, Δt is the monitoring time interval, k is the number of monitoring time intervals closest to the current time point, and γ is the incremental threshold.

[0009] Optionally, obtaining the observed value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weight, and the to-be-processed feature map set includes: obtaining the number of pixels in the pixel feature map; obtaining the feature quantity corresponding to each pixel in each pixel feature map in the to-be-processed feature map set; obtaining the observed value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weight, the number of pixels, and the feature quantity corresponding to the pixels.

[0010] Optionally, the feature model in obtaining the observed value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weight, the number of pixels, and the feature quantity corresponding to the pixels is expressed as:

[0011] where, O is the observed value corresponding to the to-be-processed feature map set, w j is the monitoring weight corresponding to the j-th pixel feature map in the to-be-processed feature map set, n is the number of pixel feature maps in the to-be-processed feature map set, m is the number of pixels in the pixel feature map, is the feature quantity corresponding to the l-th pixel in the j-th pixel feature map in the to-be-processed feature map set.

[0012] Optionally, obtaining the preset processing quantity n includes: obtaining the basic quantity according to the operation time of the bucket wheel stacker-reclaimer; obtaining the correction ratio according to the monitoring area of the bucket wheel stacker-reclaimer; obtaining the preset processing quantity n according to the basic quantity and the correction ratio.

[0013] There is also provided a bucket wheel stacker-reclaimer operation image monitoring system based on a neural network. The system includes: an acquisition module, configured to acquire the monitoring area and the monitoring time interval of the bucket wheel stacker-reclaimer, and in real time acquire the i-th bucket wheel stacker-reclaimer operation image corresponding to the monitoring area of the bucket wheel stacker-reclaimer after the i-th monitoring time interval, and acquire the i-th pixel feature map based on the image pre-trained neural network and the i-th bucket wheel stacker-reclaimer operation image, where each pixel of the pixel feature map respectively corresponds to the feature quantity output by the image pre-trained neural network; a monitoring set formation module, configured to obtain the preset processing quantity n, and obtain n continuously monitored pixel feature maps as the to-be-processed feature map set; a data processing module, configured to obtain the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set, and obtain the observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weights, and the to-be-processed feature map set; a monitoring output module, configured to obtain the monitoring result according to the observation value.

[0014] Optionally, the monitoring output module is further configured to: obtain the reference threshold range corresponding to the monitoring area of the bucket wheel stacker-reclaimer; obtain the deviation degree between the observation value and the reference threshold range, and judge whether the operation state of the bucket wheel stacker-reclaimer is abnormal according to the deviation degree; construct a time-dimensional cumulative judgment result mechanism, and establish a progressive alarm level according to the abnormal duration and the occurrence times.

[0015] Optionally, the monitoring set formation module is further configured to: obtain the basic quantity according to the operation time of the bucket wheel stacker-reclaimer; obtain the correction ratio according to the monitoring area of the bucket wheel stacker-reclaimer; obtain the preset processing quantity n according to the basic quantity and the correction ratio.

[0016] The beneficial effects of the present invention are embodied in:

[0017] In the entire neural network-based bucket wheel stacker-reclaimer operation image monitoring method, first, based on the regional division strategy differentiated by equipment functions and failure modes, combined with the adjustment of image acquisition parameters, the image quality is improved, significantly enhancing the reliability of feature extraction under dynamic lighting and dust interference. Further, the pre-trained neural network adopts a multi-scale feature fusion architecture, and the autoencoder reconstruction error mechanism constructs a pixel-level "normal benchmark". The output feature quantity can quantitatively characterize the degree of deviation of the local structure from the normal state. At the same time, for the analysis of temporal features, the dynamic time window mechanism adaptively adjusts the number of processed frames according to the regional characteristics, and cooperates with the weight model combining exponential decay and linear suppression to enhance the sensitivity to recent data while retaining the historical change law, integrating the sudden drop of single-frame features and the continuous decay of multiple frames. Further, the dual aggregation strategy of the observed values further improves the comprehensiveness of anomaly perception. Combining the dynamic threshold mechanism and the spatio-temporal confidence evaluation, it realizes the self-arbitration of conflict signals and the intelligent filtering of pseudo anomalies. Finally, through multi-level progressive alarm and fault root cause location, the average response delay is compressed from 8 minutes to the second level, greatly improving the operation and maintenance efficiency while ensuring the monitoring accuracy, providing real-time decision support for equipment predictive maintenance, and effectively avoiding the economic losses caused by unplanned shutdowns. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0019] Figure 1 It is a schematic diagram of the steps of the neural network-based bucket wheel stacker-reclaimer operation image monitoring method of the present invention in one embodiment;

[0020] Figure 2 It is a schematic diagram of a part of the steps of S4 in the neural network-based bucket wheel stacker-reclaimer operation image monitoring method of the present invention;

[0021] Figure 3 It is a schematic diagram of a part of the steps of S3 in the neural network-based bucket wheel stacker-reclaimer operation image monitoring method of the present invention;

[0022] Figure 4 It is a schematic diagram of another part of the steps of S3 in the neural network-based bucket wheel stacker-reclaimer operation image monitoring method of the present invention;

[0023] Figure 5 It is a schematic diagram of a part of the steps of S2 in the neural network-based bucket wheel stacker-reclaimer operation image monitoring method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention usually described and illustrated herein can be arranged and designed in various different configurations.

[0025] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0026] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not require further definition and explanation in subsequent drawings. In addition, the terms "first", "second", etc. are only used for descriptive distinction and cannot be construed as indicating or implying relative importance.

[0027] As Figure 1 shown, a method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network is provided, including:

[0028] S1. Obtain the monitoring area and monitoring time interval of the bucket wheel stacker-reclaimer, and in real time obtain the i-th bucket wheel stacker-reclaimer operation image corresponding to the monitoring area of the bucket wheel stacker-reclaimer after the i-th monitoring time interval, and obtain the i-th pixel feature map based on the image pre-trained neural network and the i-th bucket wheel stacker-reclaimer operation image, and each pixel of the pixel feature map corresponds to the feature quantity output by the image pre-trained neural network;

[0029] S2. Obtain the preset processing quantity n, and obtain n continuously monitored pixel feature maps and use them as the to-be-processed feature map set;

[0030] S3. Obtain the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set, and obtain the observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weights, and the to-be-processed feature map set;

[0031] S4. Obtain the monitoring result according to the observation value.

[0032] In this embodiment, it should be noted that in S1, by combining structured area division with feature extraction of a pre-trained neural network, an image analysis foundation that dynamically adapts to complex operating environments is constructed. First, the monitoring area of the bucket wheel machine is divided into three feature analysis domains: the drive mechanism area, the bucket wheel operation area, and the conveyor belt loading area. This division is based on the differences in equipment functions and failure modes: the drive mechanism area needs to capture the gear meshing state and bearing displacement characteristics, the bucket wheel operation area focuses on the integrity of the tooth profile and the dynamics of coal loading, and the conveyor belt loading area needs to identify lateral belt deviation and abnormal material distribution. The camera parameters for image acquisition in each area will be adjusted accordingly. For example, a short exposure strategy is adopted for the bucket wheel operation area with frequent dust emissions to freeze motion blur, while the HDR fusion technology is enabled for the conveyor belt area with drastic light changes to alleviate overexposure / underexposure problems at the hardware level.

[0033] Furthermore, the feature extraction process based on the pre-trained neural network adopts a multi-scale feature fusion mechanism, and the input image is parsed layer by layer through convolutional layers. For example, for the recognition of the bucket wheel tooth profile, the shallow network extracts edge gradient features to locate the tooth tip position, while the deep network differentiates the real tooth profile from dust artifacts through semantic segmentation; for the detection of conveyor belt deviation, the network correlates the belt edge coordinates of adjacent frames through temporal features to suppress the interference of instantaneous deformation caused by material accumulation. The feature quantity of each pixel in the feature map is essentially a probabilistic expression of the "normal degree" by the neural network. For example, when the shape of a certain tooth tip area of the bucket wheel is deformed due to wear, the feature quantity of the corresponding pixel will be significantly reduced due to deviation from the normal tooth profile feature distribution in the training set.

[0034] Specifically, during the training phase, the network will construct a "feature reference template" for each pixel position through the learning of a large number of normal operation images. For example, when using an autoencoder structure, the encoder compresses the input image into a latent feature vector, and the decoder attempts to reconstruct the original image. After the network is trained on the normal image dataset, the decoder will generate a high-precision reconstruction for the normal area (such as a complete bucket wheel tooth profile), while the reconstruction error for the abnormal area (such as a worn tooth tip) will increase significantly. At this time, the feature quantity can be defined as the pixel-level similarity between the reconstructed image and the original image - when the current pixel value is close to the typical distribution at this position during normal operation, the reconstruction result of the decoder is highly consistent with the original value, and the feature quantity approaches 1; conversely, if there is an abnormality (such as coal powder occlusion or tooth profile defect), the reconstruction error surges and the feature quantity approaches 0.

[0035] Furthermore, in response to dynamic lighting and dust interference, the network will fuse multi-level features to enhance robustness. For example, in the ResNet architecture, edge texture features captured by shallow convolution kernels (such as tooth tip contour lines) will be cross-layer fused with semantic features extracted by deep networks (such as the overall geometric shape of bucket wheels). By calculating the Mahalanobis distance of the current pixel in the multi-scale feature space, the degree of its deviation from the normal feature distribution is evaluated. Taking the conveyor belt deviation detection as an example, the texture direction, color saturation and other features of the belt edge pixels are tightly clustered in the latent space during normal operation, while the feature vectors of the edge pixels will deviate from the cluster center when deviating. At this time, the feature quantity will decay with the increase of the deviation distance, thereby accurately distinguishing material accumulation (instantaneous feature offset) from mechanical failure (continuous feature offset).

[0036] In S2, the time window length is dynamically determined to balance the timeliness and anti-interference ability of anomaly detection. Specifically, the determination of the preset processing quantity n needs to comprehensively consider the equipment operation cycle and the characteristics of the monitoring area: for example, the conveyor belt bearing area usually develops gradually due to deviation, so a larger n value needs to be set to cover the complete deviation formation cycle, so as to capture the belt deviation trend; while the bucket wheel operation area is prone to instantaneous abnormalities due to the impact of coal powder, so a smaller n value is used to ensure rapid response. The basic quantity is set according to the standard operation cycle length of the equipment (such as the time for the bucket wheel to rotate one circle), and then adjusted by the regional correction ratio - the drive mechanism area needs to shorten the window due to high-frequency vibration to avoid signal aliasing, and the conveyor belt area needs to extend the window to filter out material accumulation noise.

[0037] Furthermore, the construction of the feature atlas to be processed realizes the multi-dimensional fusion of time series features through the feature map sequence of continuous timestamps. For example, in order to solve the problem of misjudgment of instantaneous deformation of the conveyor belt, after collecting n continuous feature maps, the slope of the change of the feature quantity over time can be analyzed: the decrease of the feature quantity caused by normal material accumulation presents a short-term spike shape, while the mechanical deviation is manifested as a continuous linear attenuation. At the same time, the dynamic adjustment of the n value can adapt to the noise characteristics of different areas. For example, in the bucket wheel operation area with severe dust, the cumulative effect of blurred frames can be reduced by shortening n, while the driving mechanism area with stable lighting can extend n to enhance the detection sensitivity of small changes in gear wear.

[0038] In S3, first of all, the calculation of the monitoring weight fully considers the time decay characteristic and regional sensitivity: for the high-frequency vibration detection in the driving mechanism area, a higher weight is given to the recent feature map, and the exponential decay function is used to quickly weaken the influence of historical frames, so as to highlight the instantaneous abnormality of the bearing displacement; while for the deviation monitoring in the conveyor belt loading area, a relatively slow-changing weight distribution is adopted, and the linear suppression factor is used to retain the feature change trend within a longer time window, effectively capturing the progressive characteristics of the belt deviation. For example, when there is material accumulation on the conveyor belt, the feature quantity only drops suddenly in a single frame, and the observed value fluctuates slightly after weighted fusion; while continuous deviation will cause the continuous attenuation of the feature quantity in multiple frames, and trigger a threshold alarm under weighted accumulation.

[0039] Furthermore, the feature model enhances the abnormal perception ability through a dual feature aggregation strategy. The observation contribution of each feature map is jointly determined by the average feature quantity and the maximum feature quantity: the average feature quantity reflects the overall state stability and is suitable for detecting wide-area abnormalities such as uniform wear of the bucket wheel tooth shape; while the maximum feature quantity highlights local extreme abnormalities and is used to capture sudden faults such as tearing of the conveyor belt edge. Taking the bucket wheel operation area as an example, the short-term occlusion of pulverized coal will cause a sudden drop in the local feature quantity, but the maximum feature quantity mechanism can prevent this abnormality from being diluted by the regional average value; when the low feature quantity appears in the tooth tip for multiple consecutive frames, the time series weight superposition effect will drive the observed value to break through the threshold, so as to accurately identify the real wear fault. This spatio-temporal joint analysis method effectively solves the problem of confusion between instantaneous noise and continuous faults in traditional mean filtering.

[0040] In S4, a dynamic threshold response mechanism and a multi-level early warning system are established to achieve accurate discrimination and hierarchical control of the equipment state. The monitoring result is generated through the non-linear mapping relationship between the observed value and the preset threshold. The threshold range is not a fixed value, but is adaptively adjusted according to the equipment operation stage - for example, when the bucket wheel is operating at full load, the observed value is allowed to briefly drop to the lower limit of the threshold, while in the no-load state, a more stringent abnormal judgment criterion is triggered. This dynamic threshold mechanism is realized by associating equipment operating condition data (such as motor torque, material flow). When there is a conflict between the visual monitoring observed value and the PLC control signal, the threshold matching the current load state is preferentially used for arbitration to avoid the response delay caused by manual intervention.

[0041] Furthermore, the monitoring results are fused with spatio-temporal dimension confidence evaluation to improve the reliability of decision-making. Taking the detection of bucket wheel tooth profile wear as an example, when the observed value briefly exceeds the threshold within a single monitoring period, a local feature review mechanism is initiated: retrieve the sequence of feature maps of the corresponding area, and analyze whether the attenuation of the feature quantity is distributed along the tooth profile movement trajectory to distinguish true wear from pseudo-abnormalities caused by coal powder adhesion. For conveyor belt deviation, while the observed value continuously deviates from the threshold, the spatial distribution pattern of the edge feature quantity of the belt is synchronously detected - the deviation caused by mechanical failure shows a stepwise attenuation of the edge feature quantity, while material accumulation is manifested as a random patchy feature quantity mutation. The root cause of the fault is initially located through pattern matching, providing directional guidance for subsequent maintenance.

[0042] In summary, in the entire bucket wheel operation image monitoring method based on neural network, first, based on the regional division strategy differentiated by equipment functions and failure modes, combined with image acquisition parameter adjustment, the image quality is improved, significantly enhancing the reliability of feature extraction under dynamic light and dust interference; further, the pre-trained neural network adopts a multi-scale feature fusion architecture, and the autoencoder reconstruction error mechanism constructs a pixel-level "normal benchmark". The output feature quantity can quantitatively represent the degree of deviation of the local structure from the normal state. At the same time, for time series feature analysis, the dynamic time window mechanism adaptively adjusts the number of processing frames according to the regional characteristics, combined with a weight model that combines exponential decay and linear suppression, enhancing the sensitivity to recent data while retaining historical change rules, and fusing single-frame feature sudden drops and multi-frame continuous decays; further, the dual aggregation strategy of the observed value further improves the comprehensiveness of anomaly perception. Combining the dynamic threshold mechanism and spatio-temporal confidence evaluation, self-arbitration of conflicting signals and intelligent filtering of pseudo-abnormalities are realized; finally, through multi-level progressive alarm and fault root cause location, the average response delay is compressed from 8 minutes to the second level, greatly improving the operation and maintenance efficiency while ensuring the monitoring accuracy, providing real-time decision support for equipment predictive maintenance, and effectively avoiding economic losses caused by unplanned downtime.

[0043] As Figure 2 shown, in one embodiment, obtaining the monitoring result according to the observed value in S4 includes:

[0044] S41. Obtain the reference threshold range corresponding to the monitoring area of the bucket wheel;

[0045] S42. Obtain the degree of deviation between the observed value and the reference threshold range, and judge whether the operation state of the bucket wheel is abnormal according to the degree of deviation;

[0046] S43. Construct a cumulative judgment result mechanism in the time dimension, and establish a progressive alarm level according to the abnormal duration and occurrence times.

[0047] In this embodiment, it should be noted that in S41, the reference threshold range is not globally fixed, but is calibrated multi-dimensionally according to the real-time operation status and historical operation data: for example, when the bucket wheel is fully loaded, the slight deformation of the tooth shape caused by the material gravity is included in the normal elastic deformation range, and the lower limit of the threshold is automatically relaxed; while in the no-load maintenance state, the system switches to the high-sensitivity mode, and the threshold bandwidth is compressed to 30% of that in the full-load state to capture fine cracks. Further, the acquisition of the reference threshold range is a closed-loop process achieved through a multi-source data fusion and dynamic learning mechanism; the specific process includes: first, during the equipment commissioning phase, normal operation images under typical working conditions (full load, no load, start-stop, etc.) are collected, and the statistical distributions of the characteristic quantities of each monitoring area are extracted through a pre-trained neural network. For example, the mean and standard deviation of the characteristic quantities of the tooth tips in the bucket wheel operation area are combined with the mechanical tolerance parameters provided by the equipment manufacturer (such as the maximum allowable displacement of the bearing and the safe range of belt deviation) to form an initial threshold interval; second, during the operation phase, the characteristic quantities and equipment sensor data (vibration frequency, hydraulic pressure) are continuously collected, and the threshold boundary is dynamically updated using the sliding window statistical method - for example, when the standard deviation of the characteristic quantity decreases by 20% within 10 consecutive cycles of the conveyor belt, the threshold bandwidth is automatically narrowed to improve the detection sensitivity; finally, the conflict records between the PLC control signal and the visual observation value are associated through a reinforcement learning model. For example, when the torque of the bucket wheel suddenly increases but the characteristic quantity does not exceed the threshold, the system will reversely correct the threshold model to ensure the physical consistency between the equipment safety parameters and the visual monitoring results. This mechanism enables the threshold to not only reflect the performance degradation caused by equipment aging but also avoid false triggering caused by environmental noise.

[0048] In S42, an abnormal quantitative evaluation is achieved through a hierarchical deviation measurement mechanism. The deviation degree calculation adopts a regional adaptive strategy: for the bearing displacement detection in the drive mechanism area, the absolute difference method is used to directly compare the physical quantity deviation between the observed value and the threshold; while for the conveyor belt deviation, the relative deviation degree is introduced, and the offset amount of the observed value is converted into a percentage change relative to the belt width. For example, when the observed value of the conveyor belt deviates from the threshold range but the relative offset is less than 2%, it is determined as a recoverable deformation caused by material accumulation; when the offset continuously expands to 5% and is accompanied by a stepwise attenuation of the edge characteristic quantity, it is confirmed as a mechanical deviation fault. This mechanism enables even a small deviation of key components (such as the bucket wheel tooth tips) to trigger a high-priority alarm by introducing a regional characteristic weight matrix.

[0049] In S43, abnormal events are classified into three levels of response according to their duration and occurrence frequency: level 1 alarm is for instantaneous anomalies with a single duration less than 10 seconds, only triggering operation log records; level 2 alarm starts a warning prompt for events with cumulative anomalies exceeding 5 times within 30 minutes in the same area; level 3 alarm activates automatic shutdown protection after continuous anomalies exceed the threshold for 60 seconds. Taking the wear of bucket wheel teeth as an example, when the characteristic quantity is first detected to be lower than the threshold, it is marked as a level 1 observation event; when the tooth tip repeatedly triggers anomalies within three operation cycles, it is upgraded to a level 2 alarm and a wear trend analysis report is generated; if the characteristic quantity continues to deteriorate and is accompanied by a chain reaction of adjacent tooth profiles, a level 3 alarm is immediately triggered and the control module is linked to reduce the speed. This sequential accumulation mechanism effectively avoids the problem of missed reports of progressive faults in the traditional single-threshold mechanism.

[0050] As Figure 3 shown, in one embodiment, obtaining the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set in S3 includes:

[0051] S31. Obtain the number of monitoring time intervals closest to the current time point according to the current time point;

[0052] S32. Obtain the time suppression coefficient and the time sensitivity coefficient according to the monitoring area of the bucket wheel machine;

[0053] S33. Obtain the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the monitoring time interval, the number of monitoring time intervals, the time suppression coefficient, and the time sensitivity coefficient.

[0054] In this embodiment, it should be noted that in S31, the timeliness boundary of the time window is determined. By calculating the time interval sequence between the current time point and the historical monitoring moments, the k nearest valid monitoring nodes are identified to ensure that the weight model focuses on the most relevant time series segments of the equipment state. For example, under the condition of high-speed rotation of the bucket wheel, the feature maps within the last 5 complete operation cycles are automatically tracked to avoid signal aliasing caused by the periodic movement of the equipment.

[0055] In S32, the time suppression coefficient controls the attenuation rate of historical data. For example, a smaller suppression coefficient is set in the drive mechanism area to quickly forget old vibration data; the time sensitivity coefficient adjusts the steepness of the time attenuation curve. A low sensitivity coefficient is used for conveyor belt monitoring to make the weight transition smoothly, so as to capture the gradual process of belt deviation. These two parameters are dynamically optimized based on the equipment physical property library. For example, when monitoring the gearbox, the time-varying effect caused by thermal expansion needs to be considered synchronously.

[0056] In S33, the weight model fuses the time decay function with the regional constraint conditions to generate dynamic weights for each feature map: for recent feature maps, the weights exponentially decay with the vibration characteristics of the monitored area to ensure that instantaneous anomalies such as bearing displacement are preferentially responded to; for historical feature maps, a linear suppression factor is introduced to prevent the early data interference in the conveyor belt area from offsetting the trend judgment. The weight calculation process embeds the correction of the device operation phase. For example, the weight correlation of adjacent time windows is automatically enhanced during the starting and stopping stages of the bucket wheel.

[0057] In one implementation, the weight model in obtaining the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the monitoring time interval, the number of monitoring time intervals, the time suppression coefficient, and the time sensitivity coefficient in S33 is expressed as:

[0058]

[0059] w j is the monitoring weight corresponding to the j-th pixel feature map in the to-be-processed feature map set, α is the time suppression coefficient, β is the time sensitivity coefficient, Δt is the monitoring time interval, k is the number of monitoring time intervals closest to the current time point, and γ is the increment threshold.

[0060] In this implementation, it should be noted that the weight model dynamically allocates the monitoring weights of the feature maps through the dual mechanisms of time decay and regional constraint to solve the problem of time series feature confusion caused by fixed time windows.

[0061] Among them, is the exponential decay term used to achieve time sensitivity control; jΔt - kΔt represents the time difference between the j-th frame and the k-th frame of the current frame, reflecting the timeliness of historical data; the time difference is mapped to the weight decay coefficient through the exponential function. When j is close to k (recent data), the weight approaches 1; the greater the time difference, the exponential decline of the weight; more importantly, the time suppression coefficient α controls the decay speed, the smaller α is, the faster the decay (such as α = 1 in the drive mechanism area, quickly forgetting old data); the time sensitivity coefficient β adjusts the steepness of the decay curve, generally a positive integer, when β ≥ 2, the weights of long-distance time differences are more strongly suppressed; while for example, when β = 1 for the conveyor belt, gentle decay is allowed to capture the gradual offset.

[0062] Through the exponential decay term, instantaneous and continuous anomalies can be distinguished. High weights for recent data can quickly respond to instantaneous anomalies (such as bearing displacement), while the conveyor belt area retains the historical trend through slow decay to avoid misjudgment of material accumulation. Further, it realizes dynamic adaptation to the device cycle. During the bucket wheel rotation cycle (such as 30 seconds), the exponential decay ensures that only the effective vibration data of the current cycle is retained, eliminating cross-cycle signal aliasing.

[0063] Furthermore, max{1, γ - (k - j)} is a linear increment term; the increment threshold γ limits the upper bound of the number of frames for obtaining increments. When γ ≥ (k - j), that is, historical frames within a certain time range are selected and the weight is multiplied by the coefficient γ - (k - j) to further enhance the contribution of recent historical frames; when γ < (k - j), max{1, γ - (k - j)} is forced to be 1 to prevent obsolete data from interfering. Generally, let γ = 2, and only retain the high-frequency vibration characteristics of the last 2 frames.

[0064] Through the linear increment term, the accumulation of noise is suppressed. In the bucket wheel operation area with serious dust, the number of historical blurred frames is limited by γ to avoid the distortion of observation values caused by the accumulation of low-quality data; the calculation efficiency is also balanced, and invalid frames are restricted from participating in the calculation.

[0065] In summary, the synergistic effect and parameter tuning are achieved. The exponential term controls the time decay, and the linear increment term restricts the historical range. The product of the two realizes the high weight of recent frames to highlight instantaneous anomalies, the moderate weight of valid historical frames to retain trend information, and the weight of invalid historical frames to be zero or the minimum value. Specifically, the parameter linkage of α, β, and γ is adopted, and joint tuning is carried out according to the characteristics of the monitoring area. For example: conveyor belt deviation detection: α = 2 (slow decay), β = 1 (gentle curve), γ = 5 (long window); gearbox vibration detection: α = 0.3 (fast decay), β = 2 (steep curve), γ = 2 (short window).

[0066] As Figure 4 shown, in one embodiment, obtaining the observation value corresponding to the to-be-processed feature map set based on the feature model, monitoring weight, and to-be-processed feature map set in S3 includes:

[0067] S34. Obtain the number of pixels in the pixel feature map;

[0068] S35. Obtain the feature quantities corresponding to each pixel in each pixel feature map in the to-be-processed feature map set;

[0069] S36. Based on the feature model, monitoring weight, the number of pixels, and the feature quantities corresponding to the pixels, obtain the observation value corresponding to the to-be-processed feature map set.

[0070] In this embodiment, it should be noted that in S34, to obtain the number of pixels in the pixel feature map, the total number of pixels in each pixel feature map is counted to provide a spatial dimension calculation benchmark for subsequent feature quantity aggregation.

[0071] In S35, the feature quantities of each pixel are obtained. In this process, the feature quantities of each pixel in the to-be-processed feature map set are extracted. These feature quantities are essentially the probabilistic expressions of the neural network for the "normal state" of the device. For example, in the bearing displacement detection in the drive mechanism area, the feature quantity maps the matching degree between the bearing contour and the standard meshing position; the feature quantity of the pixels at the edge of the conveyor belt reflects the degree of texture distortion caused by the lateral offset of the belt. By extracting pixel by pixel, a multi-dimensional feature matrix is constructed, providing a data basis for time series analysis while retaining spatial details, and effectively overcoming the smoothing effect of traditional mean filtering on local anomalies.

[0072] In S36, first, the regional average feature quantity of each frame of the feature map is calculated to reflect the overall state stability (such as the uniform wear of the bucket wheel teeth); at the same time, the maximum feature quantity of this frame is extracted to highlight local extreme anomalies (such as the tearing of the conveyor belt). After the two are weighted and fused, the influence of the time series weight is superimposed.

[0073] In one implementation, the feature model in the observation value corresponding to the to-be-processed feature map set obtained based on the feature model, the monitoring weight, the number of pixels, and the feature quantity corresponding to the pixels in S36 is expressed as:

[0074] Among them,

[0075] O is the observation value corresponding to the to-be-processed feature map set, w j is the monitoring weight corresponding to the j-th pixel feature map in the to-be-processed feature map set, n is the number of pixel feature maps in the to-be-processed feature map set, m is the number of pixels in the pixel feature map, is the feature quantity corresponding to the l-th pixel in the j-th pixel feature map in the to-be-processed feature map set.

[0076] In this implementation, it should be noted that this model fuses the overall state stability and local extreme anomalies through a spatio-temporal dual feature aggregation mechanism, and accurately distinguishes instantaneous interference from real faults.

[0077] Among them, represents the regional average feature quantity, realizing the overall state evaluation; specifically, by calculating the mean value of all pixel feature quantities in each frame of the feature map, it reflects the stability of the overall operation state of the device; it realizes wide-area anomaly detection, captures wide-area anomalies such as uniform wear (such as the overall wear of the bucket wheel teeth) and progressive deviation of the conveyor belt, and at the same time smooths the local pixel fluctuations caused by dust or sudden light changes within a single frame through the mean value.

[0078] Furthermore, To extract the maximum value of the feature quantity in each frame of the feature map and locate the most abnormal pixel points; realizing the capture of sudden faults and identifying local severe damages such as conveyor belt edge tearing and bearing cracks; at the same time, avoiding abnormal dilution and preventing local severe anomalies from being diluted by the regional average value (such as coal powder occlusion only affecting a single tooth tip).

[0079] Furthermore, Multiply the mean value of each frame by the temporal weight and add it to the maximum value, and finally accumulate the contributions of all frames; realizing temporal feature enhancement, with high weights for recent frames (such as the superposition of low feature quantities in multiple frames when the conveyor belt continuously runs off track), and low weights for historical frames (such as a sudden drop in a single frame due to material accumulation); at the same time, realizing spatio-temporal joint analysis, where the mean value captures the trend, the maximum value locates the local area, and the weight controls the timeliness, and the three cooperate to improve the detection accuracy.

[0080] In summary, distinguish instantaneous interference from continuous faults; in instantaneous interference (such as material accumulation), the maximum feature quantity of a single frame drops suddenly, but the mean value changes little, and the subsequent frames recover. After weighting, due to the low weight of historical frames, the observed value fluctuates limitedly; in continuous faults (such as mechanical deviation), the mean value and maximum value of multiple frames continuously decrease, and the contribution of recent frames with high weights is superimposed, and the observed value significantly deviates from the threshold. Furthermore, it also realizes the balance between anti-environmental noise and local anomalies; such as dust interference, the feature quantity of a single pixel drops suddenly (low maximum value), but the mean value does not change significantly, and the observed value is buffered by the weight and the mean value, avoiding false alarms; such as local damage, the maximum value highlights the abnormal pixel, and even if the mean value does not exceed the threshold, an alarm can still be triggered through the weighting of the maximum value. Furthermore, it also realizes dynamic adaptation to the operating state of the equipment; such as the linkage of the weight model, high weights for recent frames (such as vibration detection of the driving mechanism) ensure fast response to instantaneous anomalies, and slow attenuation of the long window (such as conveyor belt deviation) captures the gradual change trend; such as the complementarity of dual features, the mean value prevents misjudgment of local noise, and the maximum value avoids missing sudden damages, and the two are weighted and fused to improve robustness.

[0081] As Figure 5 shown, in one embodiment, obtaining the preset processing quantity n in S2 includes:

[0082] S21. Obtain the basic quantity according to the operation time of the bucket wheel machine;

[0083] S22. Obtain the correction ratio according to the monitoring area of the bucket wheel machine;

[0084] S23. Obtain the preset processing quantity n according to the basic quantity and the correction ratio.

[0085] In this embodiment, it should be noted that in S21, the basic quantity is dynamically set according to the standard operation cycle of the equipment to ensure that the time window covers the key operation stages. For example, the complete cycle duration of the bucket wheel rotating one week is used as a benchmark to determine the initial number of processed frames, so that the feature map set can completely capture the state changes in the equipment's cyclic operation. This setting avoids truncating the periodic vibration signal due to a too short window or introducing redundant data interference due to a too long window.

[0086] In S22, the correction ratio is dynamically adjusted based on the physical characteristics and failure modes of the monitoring area. In the drive mechanism area, the window needs to be shortened to prevent signal aliasing due to high-frequency vibration, and the correction ratio is less than 1; in the conveyor belt loading area, the window is extended to capture progressive deviation, and the correction ratio is greater than 1. The ratio value is generated through the quantitative evaluation of the regional noise sensitivity (such as the dust emission frequency) and the failure development speed (such as the gear wear rate).

[0087] In S23, the basic quantity is multiplied by the correction ratio to generate the region-adaptive processing quantity n. For example, in the bucket wheel operation area, it is multiplied by 0.8 on the basis of the basic quantity to quickly respond to the pulverized coal impact, and in the conveyor belt area, it is multiplied by 1.5 to cover the complete deviation cycle. This dynamic adjustment mechanism enables the time window to adapt to the equipment operation rhythm and filter out regional-specific noise, solving the misjudgment and delay problems caused by a fixed window.

[0088] A bucket wheel operation image monitoring system based on a neural network is also provided. The system includes:

[0089] An acquisition module, configured to acquire the monitoring area and monitoring time interval of the bucket wheel, and in real time acquire the i-th bucket wheel operation image corresponding to the monitoring area of the bucket wheel after the i-th monitoring time interval, and acquire the i-th pixel feature map based on the image pre-trained neural network and the i-th bucket wheel operation image, and each pixel of the pixel feature map respectively corresponds to the feature quantity output by the image pre-trained neural network;

[0090] A monitoring set formation module, configured to acquire the preset processing quantity n, and acquire n continuously monitored pixel feature maps as the to-be-processed feature map set;

[0091] A data processing module, configured to acquire the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set, and acquire the observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weights, and the to-be-processed feature map set;

[0092] A monitoring output module, configured to acquire the monitoring result according to the observation value.

[0093] In one embodiment, the monitoring output module is further configured to: obtain the reference threshold range corresponding to the monitoring area of the bucket wheel stacker-reclaimer; obtain the degree of deviation between the observed value and the reference threshold range, and determine whether the operation state of the bucket wheel stacker-reclaimer is abnormal according to the degree of deviation; construct a cumulative judgment result mechanism in the time dimension, and establish a progressive alarm level according to the duration and occurrence times of the abnormality.

[0094] In one embodiment, the monitoring set forming module is further configured to: obtain the basic quantity according to the operation time of the bucket wheel stacker-reclaimer; obtain the correction ratio according to the monitoring area of the bucket wheel stacker-reclaimer; and obtain the preset processing quantity n according to the basic quantity and the correction ratio.

[0095] In this embodiment, it should be noted that regarding the above-mentioned bucket wheel stacker-reclaimer operation image monitoring system based on neural network, the specific implementation manner of the execution operation has been described in detail in the embodiment of the bucket wheel stacker-reclaimer operation image monitoring method based on neural network, and will not be elaborated here.

[0096] The preferred embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings. However, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept scope of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all belong to the protection scope of the present disclosure.

[0097] In addition, it should be noted that, in the above specific embodiments, the various specific technical features described can be combined in any suitable manner without conflict. To avoid unnecessary repetition, the present disclosure will not further describe various possible combination manners.

[0098] In addition, any combination can be made between various different embodiments of the present disclosure, as long as it does not violate the idea of the present disclosure, and it should also be regarded as the content disclosed by the present disclosure.

[0099] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. A method for monitoring the operation image of a bucket wheel stacker based on a neural network, characterized in that, Including: Obtain the monitoring area and monitoring time interval of the bucket wheel stacker-reclaimer, and in real time obtain the i-th bucket wheel stacker-reclaimer operation image corresponding to the monitoring area of the bucket wheel stacker-reclaimer after the i-th monitoring time interval, and obtain the i-th pixel feature map based on the image pre-trained neural network and the i-th bucket wheel stacker-reclaimer operation image, where each pixel of the pixel feature map corresponds to the feature quantity output by the image pre-trained neural network; Obtain the preset processing quantity n, and obtain n continuously monitored pixel feature maps as the to-be-processed feature map set; Based on the weight model, the current time point, and the to-be-processed feature map set, obtain the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set, and based on the feature model, the monitoring weights, and the to-be-processed feature map set, obtain the observation value corresponding to the to-be-processed feature map set; Obtain the monitoring result according to the observation value.

2. The method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network according to claim 1, wherein The obtaining the monitoring result according to the observation value includes: Obtain the reference threshold range corresponding to the monitoring area of the bucket wheel stacker-reclaimer; Obtain the deviation degree between the observation value and the reference threshold range, and judge whether the operation state of the bucket wheel stacker-reclaimer is abnormal according to the deviation degree; Construct a time dimension cumulative judgment result mechanism, and establish a progressive alarm level according to the abnormal duration and occurrence times.

3. The method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network according to claim 1, wherein The obtaining the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the current time point, and the to-be-processed feature map set includes: According to the current time point, obtain the number of monitoring time intervals closest to the current time point; According to the monitoring area of the bucket wheel stacker-reclaimer, obtain the time suppression coefficient and the time sensitivity coefficient; Based on the weight model, the monitoring time interval, the number of monitoring time intervals, the time suppression coefficient, and the time sensitivity coefficient, obtain the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set.

4. The method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network according to claim 3, characterized in that, The weight model in the obtaining the monitoring weights corresponding to each pixel feature map in the to-be-processed feature map set based on the weight model, the monitoring time interval, the number of monitoring time intervals, the time suppression coefficient, and the time sensitivity coefficient is expressed as: w j is the monitoring weight corresponding to the j-th pixel feature map in the feature map set to be processed, α is the time suppression coefficient, β is the time sensitivity coefficient, Δt is the monitoring time interval, k is the number of monitoring time intervals closest to the current time point, and γ is the increment threshold.

5. The method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network according to claim 1, characterized in that The obtaining the observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weights, and the to-be-processed feature map set includes: Obtain the number of pixels in the pixel feature map; Obtain the feature quantities corresponding to each pixel in each pixel feature map in the to-be-processed feature map set; Based on the feature model, the monitoring weights, the number of pixels, and the feature quantities corresponding to the pixels, obtain the observation value corresponding to the to-be-processed feature map set.

6. The method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network according to claim 5, characterized in that, The feature model in the obtaining the observation value corresponding to the to-be-processed feature map set based on the feature model, the monitoring weights, the number of pixels, and the feature quantities corresponding to the pixels is expressed as: Among them, O is the observation value corresponding to the feature map set to be processed, w j is the monitoring weight corresponding to the j-th pixel feature map in the feature map set to be processed, n is the number of pixel feature maps in the feature map set to be processed, and m is the number of pixels in the pixel feature map. is the feature quantity corresponding to the l-th pixel in the j-th pixel feature map in the feature map set to be processed.

7. The method for monitoring the operation image of a bucket wheel stacker-reclaimer based on a neural network according to claim 1, wherein The obtaining the preset processing quantity n includes: According to the operation time of the bucket wheel stacker-reclaimer, obtain the basic quantity; According to the monitoring area of the bucket wheel stacker-reclaimer, obtain the correction ratio; According to the basic quantity and the correction ratio, obtain the preset processing quantity n.

8. An image monitoring system for the bucket wheel stacker-reclaimer operation based on a neural network, characterized in that, The system includes: An obtaining module, configured to obtain the monitoring area and monitoring time interval of the bucket wheel stacker-reclaimer, and in real time obtain the i-th bucket wheel stacker-reclaimer operation image corresponding to the monitoring area of the bucket wheel stacker-reclaimer after the i-th monitoring time interval, and obtain the i-th pixel feature map based on the image pre-trained neural network and the i-th bucket wheel stacker-reclaimer operation image, where each pixel of the pixel feature map corresponds to the feature quantity output by the image pre-trained neural network; A monitoring set formation module, configured to obtain a preset processing quantity n, and obtain n continuously monitored pixel feature maps as a feature map set to be processed; A data processing module, configured to obtain the monitoring weights corresponding to the pixel feature maps in the feature map set to be processed based on a weight model, a current time point, and the feature map set to be processed, and obtain an observation value corresponding to the feature map set to be processed based on a feature model, the monitoring weights, and the feature map set to be processed; A monitoring output module, configured to obtain a monitoring result according to the observation value.

9. The image monitoring system for the bucket wheel stacker-reclaimer operation based on neural network according to claim 8, wherein The monitoring output module is further configured to: Obtain a reference threshold range corresponding to the monitoring area of the bucket wheel stacker-reclaimer; Obtain the degree of deviation between the observation value and the reference threshold range, and determine whether the operation state of the bucket wheel stacker-reclaimer is abnormal according to the degree of deviation; Construct a time dimension cumulative judgment result mechanism, and establish a progressive alarm level according to the abnormal duration and the occurrence times.

10. The bucket wheel stacker-reclaimer operation image monitoring system based on a neural network according to claim 8, characterized in that, The monitoring set formation module is further configured to: Obtain a basic quantity according to the operation time of the bucket wheel stacker-reclaimer; Obtain a correction ratio according to the monitoring area of the bucket wheel stacker-reclaimer; Obtain the preset processing quantity n according to the basic quantity and the correction ratio.

Images