Belt conveyor coal gangue in-situ intelligent sorting system and method

Abstract

The invention relates to the technical field of belt conveyor coal gangue sorting, and discloses a belt conveyor coal gangue in-situ intelligent sorting system and method. The key points of the technical scheme are as follows: a three-level cloud edge-end collaborative distributed architecture of terminal layer multi-source data acquisition and action execution-edge layer localization intelligent decision and collaborative control-cloud layer global optimization and remote management is adopted; in combination with a lightweight multi-feature fusion recognition model, a short-time window coal gangue flow state sensing sorting strategy, a multi-belt speed linkage model and a distributed redundancy control mechanism, millisecond-level real-time sorting, high-robustness recognition, gangue-coal ratio maximization and continuous and stable system operation of underground coal gangue are realized, and meanwhile, network bandwidth pressure and system energy consumption are reduced.

Description

Technical Field

[0001] This invention relates to the field of coal gangue sorting technology using belt conveyors, and more specifically, to an in-situ intelligent sorting system and method for coal gangue using belt conveyors. Background Technology

[0002] In the current field of coal gangue sorting using belt conveyors, existing technologies suffer from several key bottlenecks that restrict practical application effectiveness: In terms of processing architecture, centralized systems require all image data to be transmitted back to the central processor, leading to excessive network bandwidth consumption during multi-point, multi-line deployments and sorting response delays exceeding milliseconds, making it difficult to meet real-time requirements; in terms of identification schemes, most rely on single image features or weight parameters, exhibiting weak anti-interference capabilities in complex environments such as uneven lighting, dust, and moisture underground, resulting in low coal gangue identification accuracy; in terms of sorting logic, the common "single gangue identification - single action" mode fails to consider the continuous state of coal gangue flow, easily leading to missed areas when gangue is densely packed. The existing coal gangue sorting system suffers from several shortcomings. Firstly, it has a high sorting rate and a high rate of missorting of clean coal when coal is concentrated, making it impossible to maximize the gangue-to-coal ratio. Secondly, in terms of belt control, the speeds of the feeding belt and the main conveyor belt are not integrated into a unified control framework, resulting in a mismatch between coal and gangue flow rate and belt speed, often leading to material accumulation or idling, and causing high system energy consumption. Thirdly, in terms of operational reliability, it partially relies on a "camera + cloud server" architecture. When the underground network fluctuates or is interrupted, model inference and sorting command issuance are easily interrupted, resulting in insufficient continuous system operation. These defects collectively make it difficult for the existing coal gangue sorting system to meet the needs of complex underground production scenarios in terms of efficiency, accuracy, stability, and energy consumption control.

[0003] Therefore, the present invention provides an in-situ intelligent sorting system and method for coal gangue on a belt conveyor, which improves the above-mentioned technical problems. Summary of the Invention

[0004] This disclosure aims to address the shortcomings of existing technologies by providing an in-situ intelligent sorting system and method for coal gangue on belt conveyors. The invention employs a three-tiered cloud-edge-device collaborative distributed architecture: "terminal-layer multi-source data acquisition and action execution - edge-layer localized intelligent decision-making and collaborative control - cloud-layer global optimization and remote management." It combines a lightweight multi-feature fusion recognition model, a short-time window coal gangue flow state perception sorting strategy, a multi-belt speed linkage model, and a distributed redundant control mechanism to achieve millisecond-level real-time sorting of underground coal gangue, highly robust identification, maximization of the gangue-to-coal ratio, and continuous stable system operation, while simultaneously reducing network bandwidth pressure and system energy consumption.

[0005] To achieve the above objectives, the present disclosure proposes the following technical solutions: In a first aspect, the present disclosure proposes an in-situ intelligent sorting system for coal gangue using a belt conveyor. This system adopts a cloud-edge-device collaborative distributed architecture, including a terminal layer, an edge layer, and a cloud layer. The terminal layer is used to collect multi-dimensional feature data of materials on the feeding belt in real time, and to receive instructions to perform sorting actions and adjust belt speed. The edge layer is communicatively connected to the terminal layer and the cloud layer. It is used to receive the multi-dimensional feature data of the material, perform material identification and coal gangue flow status determination locally, and generate sorting control instructions and belt coordinated speed adjustment instructions based on the determination results and send them to the terminal layer. The edge layer has local model storage and offline decision-making capabilities. The cloud layer is used to aggregate the system's global operating data, iteratively optimize the recognition model and distribute it to the edge layer, and remotely monitor and manage the system.

[0006] As a preferred embodiment of the present invention, the terminal layer includes: a binocular camera, a servo drive receiving plate, and a belt frequency converter driver, all disposed at each feeding belt. The binocular camera is equipped with an adaptive low-light enhancement module, which is based on the multi-scale Retinex algorithm. It decomposes the image brightness components through a Gaussian filter kernel and dynamically adjusts the gain coefficient in combination with feedback from the ambient light intensity sensor. The binocular camera is used to simultaneously acquire two-dimensional images, three-dimensional shapes, and belt coordinate position data of the coal gangue mixture.

[0007] As a preferred embodiment of the present invention, the edge layer includes: edge computing nodes and edge master control nodes; The edge computing node has a built-in lightweight multi-feature fusion recognition model. The model uses a parallel structure to extract color features, texture features and shape features respectively, and outputs the gangue confidence score after weighted fusion through a fully connected network. The backbone network of the model is embedded with a channel attention mechanism and uses structured pruning and quantization techniques for lightweight processing. The edge master node is used to aggregate real-time coal-gangue ratio and flow data uploaded by each edge computing node, and to run the multi-belt speed linkage model to calculate speed commands.

[0008] As a preferred technical solution of the present invention, the edge layer is configured with a coal and gangue flow state sensing and sorting strategy based on a short time window; The strategy uses a 1.0-second time window to count the frequency of gangue occurrence within the window, classifying the coal and gangue flow state into three categories: gangue-dense, mixed-dense, and coal-dense. When the state is determined to be gangue-dense, the duration of the gangue receiving plate extension is extended and the action interval is shortened. When the state is determined to be coal-dense, a short-term action is triggered only when high-confidence gangue is detected. When the state is determined to be mixed-dense, the feed belt speed and the gangue receiving plate action frequency are adjusted proportionally.

[0009] As a preferred embodiment of the present invention, the multi-belt speed linkage model adjusts the speed based on the difficulty of receiving coal and the total clean coal flow rate: Regarding the speed of the feed belt Using gangue content The parabolic function model with the gangue content as the independent variable maintains the base speed when the gangue content approaches 0 or 1, and automatically reduces the speed when the gangue content approaches 0.5; for the speed of the main conveyor belt... The total clean coal flow rate is adopted. The model is a hyperbolic tangent function with variable denoted by , which monotonically increases with increasing flow rate; all speed adjustments are smoothly transitioned through a linear ramp function.

[0010] As a preferred embodiment of the present invention, the edge layer has a distributed redundancy control mechanism; the edge computing node locally stores a model version library containing optimized models of the most recent versions; the edge master control node monitors the communication status with the cloud layer through heartbeat packets, and when a communication interruption is detected, it issues an offline operation command, and the edge computing node switches to the locally verified model to continue running; after communication is restored, the data during the offline period is synchronized through block verification and breakpoint resume mechanism.

[0011] As a preferred technical solution of the present invention, the cloud layer deploys a multi-feature fusion model optimization system; the model optimization system is based on an incremental learning algorithm to screen the sample data uploaded from the edge layer for validity, remove blurry or severely occluded samples, and use the screened high-quality samples to iteratively fine-tune the recognition model; after the fine-tuned model passes the verification on an independent test set containing noisy, occluded and low-light images, it is sent down to the edge layer for updating.

[0012] As a preferred technical solution of the present invention, the cloud layer is also deployed with a digital twin and visualization system; the system is built on the Unity3D engine and drives the synchronous update of the three-dimensional model by receiving the structured data stream uploaded by the edge layer in real time; the system has intelligent diagnostic function, which can judge the jamming fault of the connecting plate by combining the servo motor current waveform, and judge the belt misalignment fault by monitoring the belt edge offset.

[0013] Secondly, this disclosure proposes an in-situ intelligent sorting method for coal gangue using a belt conveyor, the method comprising the following steps: Step 1: The terminal layer acquires image and shape data of the coal and gangue mixture, which is then preprocessed and transmitted to the edge layer; Step 2: The edge layer uses a lightweight multi-feature fusion recognition model for real-time identification and calculates the frequency of gangue occurrence within a short time window. Based on this, the coal and gangue flow status is determined and the action parameters of the gangue receiving plate are dynamically adjusted. Step 3: The edge layer calculates and issues target speed commands for each belt based on the real-time gangue content and material flow rate using a multi-belt speed linkage model, thereby achieving coordinated speed regulation; Step 4: The cloud layer periodically acquires sample data uploaded by the edge layer for incremental model training and optimization, and then distributes the updated model to the edge layer.

[0014] As a preferred embodiment of the present invention, the multi-belt speed linkage model in step three calculates the target speed of the feeding belt. The formula is: in, For the base speed of the sub-belt, The difficulty level is affected by the coefficient. Real-time gangue content; Calculate the target speed of the main conveyor belt The formula is: in, Main belt rated speed, This is the minimum speed proportionality coefficient. The total clean coal flow rate, As a baseline value for flow rate, For flow sensitivity coefficient, It is the hyperbolic tangent function.

[0015] In summary, the present invention has the following beneficial effects: First, it adopts a distributed architecture of "terminal-edge-cloud", where edge nodes complete image processing and recognition locally, uploading only 1-2KB of structured data, reducing the uplink bandwidth usage of a single node by more than 95%. Relying on edge local inference (≤10ms) and near-end collaborative transmission (<5ms), the overall system response latency is controlled within 30 milliseconds, completely solving the problems of high bandwidth pressure and response latency exceeding milliseconds in traditional centralized architectures.

[0016] Secondly, by using a lightweight multi-feature fusion recognition model (improved MobileNetV2), it simultaneously extracts three heterogeneous features of color, texture, and shape and adaptively weights and fuses them. Combined with the adaptive low-light enhancement module of the binocular camera, it can still maintain an average recognition accuracy of 96.7% in complex environments such as uneven lighting, dust, and water vapor in the mine. The missed selection rate is <3.5% when gangue is dense and the misselection rate of clean coal is <1.8% when coal is dense, which is significantly better than the traditional single feature recognition scheme.

[0017] Third, a 1.0-second short-window coal and gangue flow state sensing and sorting strategy is adopted. Based on the frequency of gangue occurrence, three states are divided into "gangue-dense, mixed-dense, and coal-dense". The action parameters of the gangue receiving plate (extension duration and action interval) are dynamically adjusted to avoid the problems of missed selection and misselection in the traditional "single gangue identification-single action" mode. In actual application, the comprehensive gangue and coal sorting is improved by about 18% compared with the traditional scheme.

[0018] Fourth, through a multi-belt speed linkage model, the system classifies the working conditions into three categories: "more coal, more gangue, and mixed" based on the global coal-to-gangue ratio and material flow rate. The speeds of the main conveyor belt and each feeding belt are adjusted differently, and a linear ramp function is used to achieve smooth speed regulation, avoiding material accumulation or idling. Industrial verification shows that the energy consumption of the main transport system is reduced by about 15%, the speed regulation response time is ≤2s, and the overshoot is <5%.

[0019] Fifth, the edge layer has a distributed redundancy control mechanism. The edge computing nodes locally store the three most recent optimized models, and can seamlessly switch to the local model to continue running when the network is interrupted. After communication is restored, offline data is synchronized through block verification and breakpoint resume mechanism. It can still sort normally within 45 minutes of simulated network interruption, which solves the defects of the traditional "camera + cloud server" architecture that depends on the network and is prone to interruption.

[0020] Sixth, the cloud layer uses incremental learning algorithms to process high-quality samples uploaded from the edge layer daily and iteratively optimize the model to ensure that the model continuously adapts to the dynamic environment downhole; the Unity3D-based digital twin system achieves high-fidelity mapping of equipment status and sub-second virtual-real synchronization, and combined with intelligent diagnosis and early warning functions, it can automatically identify faults such as stuck connecting plates and belt misalignment and push disposal suggestions, which facilitates remote monitoring and maintenance by the ground dispatch center. Attached Figure Description

[0021] Figure 1 A framework diagram of an in-situ intelligent sorting system for coal gangue using a belt conveyor, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of material flow provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the material flow of a single feeding system provided in an embodiment of the present invention; Figure 4 A control logic diagram for the action control of the coal receiving plate provided in an embodiment of the present invention; Figure 5 To improve the MobileNetV2 network model diagram; In the diagram: 1. Coal block, 2. Gangue, 3. Calculation window, 4. Feeding belt, 5. Coal block bin, 6. Main conveyor belt, 7. Ground coal mine dispatch center, 8. Gangue bin, 9. Servo-driven gangue receiving plate, 10. Edge master control node, 11. Binocular camera, 12. Edge computing node, 13. Belt frequency converter driver. Detailed Implementation

[0022] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0025] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0026] This disclosure aims to address the problems in existing belt conveyor coal gangue sorting systems, such as excessive bandwidth consumption in centralized architecture, weak anti-interference capability in complex environments, high missed and false sorting rates due to sorting logic not adapting to the continuous state of coal gangue flow, high energy consumption due to lack of coordinated control of multi-belt speeds, and insufficient operational reliability due to reliance on network stability. Therefore, this disclosure proposes an in-situ intelligent sorting system and method for belt conveyors to achieve efficient and accurate sorting of underground coal gangue. This method adopts a three-level cloud-edge-end collaborative distributed architecture, combining a lightweight multi-feature fusion recognition model, a short-time window coal gangue flow state perception sorting strategy, a multi-belt speed linkage model, and a distributed redundant control mechanism. Through configurable edge-layer local intelligent decision-making and differentiated adaptation logic, it achieves millisecond-level real-time sorting of coal gangue, highly robust identification, maximization of the coal-gangue ratio, and continuous stable system operation, while reducing network bandwidth pressure and system energy consumption.

[0027] Please refer to Figure 1 , Figure 1 This diagram illustrates a framework of an in-situ intelligent coal gangue sorting system according to an embodiment of this disclosure; the overall framework mainly includes: Terminal layer: Multi-dimensional data acquisition and action execution end.

[0028] Each feeding belt conveyor is independently equipped with a set of terminal equipment, including a binocular camera (resolution ≥ 2 million pixels, frame rate 30fps), a servo-driven receiving plate (response time ≤ 100ms, single plate reciprocating stroke range of 800-1400mm), and a belt frequency converter (power level 7.5kW-22kW, adaptable to the power requirements of different feeding belt conveyors), forming a multi-source data acquisition closed loop.

[0029] The binocular camera is equipped with an adaptive low-light enhancement module, which is based on a multi-scale Retinex algorithm. This module decomposes the image brightness components using a Gaussian filter kernel (scale parameters σ=15, 80, 250) at multiple scales. Combined with real-time feedback from an ambient light sensor, it dynamically adjusts the gain coefficient (range 0.8~2.5) and automatically activates when the illuminance is below 50 Lux to improve the visibility of details in dark areas and suppress overexposure and noise. The module simultaneously acquires two-dimensional images of the coal and gangue mixture, three-dimensional shape (including contours, edges, and aspect ratio), and coordinate position data on the conveyor belt to support multi-feature recognition.

[0030] The servo-driven connecting plate covers the full width of the belt. The belt frequency converter driver can receive speed commands from the edge layer to control the feeding belt conveyor for smooth speed adjustment.

[0031] The terminal layer mainly performs real-time acquisition and preprocessing of image, shape, and position data, and transmits the data to the edge layer. It also executes sorting and receiving plate actions and speed adjustment commands.

[0032] Edge layer: Localized intelligent decision-making and collaborative control terminal.

[0033] The edge layer includes edge computing nodes and edge master control nodes, which are connected via industrial Ethernet to form a distributed control structure with offline redundant operation capabilities.

[0034] One edge computing node is deployed near the head of each feeding belt conveyor. Its hardware configuration uses an embedded AI chip (such as NVIDIA Jetson Xavier NX) with an ARMA77 architecture, providing ≥21 TOPS of INT8 computing power, and integrating dual gigabit industrial Ethernet ports to support Profinet and EtherCAT protocols for real-time transmission of control commands. The node has a built-in lightweight multi-feature fusion recognition module based on the TensorFlow Lite framework, which adopts a two-stage mechanism of feature-level weighted fusion and decision-level integration: First, color features (HSV space grayscale mean and variance), texture features (LBP neighborhood contrast), and shape features (Canny edge contour perimeter, area, and number of edges) are extracted. Then, each feature is subjected to min-max normalization, and the values ​​of each dimension are linearly scaled to the [0,1] interval and concatenated to form a fixed-dimensional feature vector with a total dimension of 256. The feature vector is input into a two-layer fully connected network containing 128 neurons (ReLU activation) and 64 neurons for weighted fusion. During training, the network adaptively learns the weights of each feature through backpropagation. The fused features are then output as gangue confidence and location information by a softmax classification layer.

[0035] The model training employs the cross-entropy loss function and introduces L2 regularization (weight coefficient λ=0.001). The optimizer uses Adam (parameters set to β1=0.9, β2=0.999) with an initial learning rate of 1e-4, and a cosine annealing strategy is used to decay the learning rate from the initial value to 1e-6 during the training period. This module establishes an independent model version repository in local solid-state storage, storing the three most recent optimized models and related files in a hierarchical directory structure. Each version directory contains the model file (.h5 format), the corresponding configuration file (.json format), and the historical validation records for that version. The system adopts a rolling update and validity verification mechanism. When a new candidate model is received from the cloud, the node first loads the model in a sandbox isolation environment and performs inference verification using a locally reserved specialized validation set (containing 500 typical samples covering noise, occlusion, and illumination changes).

[0036] The verification process simultaneously evaluates recognition accuracy (not lower than the currently used model) and single-frame inference time (≤15ms). If both metrics are met, the new model is deemed valid. Subsequently, the system automatically archives and backs up the earliest historical version directory and then deletes it, while moving the verified new model directory into the version repository, completing rolling updates. This mechanism supports seamless switching of nodes to the latest and verified local model in the version repository when cloud communication is interrupted. For collaborative control, edge computing nodes upload structured real-time data packets to the edge master node at a fixed frequency of 10Hz. The data includes: real-time coal-to-gangue ratio, material flow rate, gangue receiving plate operation status, actual feeder belt speed, and the node's own health status. The edge master node is deployed in the main conveyor belt drive unit, responsible for aggregating the uploaded data from all nodes and running a multi-belt speed linkage model to calculate the target speeds of the main conveyor belt and each feeder belt. System commands are issued using a priority queue mechanism, with speed adjustment commands and emergency control commands (such as "offline operation") having the highest priority, taking precedence over regular commands such as status queries and data reporting.

[0037] The edge master control node is deployed in the main conveyor belt drive unit. It is responsible for aggregating real-time data uploaded by all edge computing nodes at a frequency of 10Hz, including coal-to-gangue ratio, material flow rate, gangue receiving plate status, and feeder belt speed. It also runs a multi-belt speed linkage model to calculate the target speeds of the main conveyor belt and each feeder belt. The system command issuance adopts a priority queue mechanism, with speed adjustment commands taking precedence over status query commands to support real-time control response. Simultaneously, the node monitors the communication status with the cloud layer through periodic heartbeat packets (every 2-second interval). If no response is received from the cloud for three consecutive times, it is considered a communication interruption. The node then issues an "offline run" command to all edge computing nodes, which switch to the latest validated model stored locally and continue performing identification and sorting tasks. When the heartbeat packets resume and remain stable for five consecutive cycles, communication is considered restored, and the historical synchronization of the offline running data is automatically triggered.

[0038] The "terminal-edge" distributed processing architecture adopted in this invention is compared with the traditional "camera + cloud server" centralized processing solution in terms of key performance indicators as follows: ① In terms of bandwidth usage, traditional centralized solutions require continuous uplink of 4-8Mbps high-definition video streams to the cloud server; while this system completes image processing and recognition locally at the edge node, requiring only the upload of about 1-2KB of structured result data (such as coordinates and confidence scores), reducing the uplink bandwidth usage of a single node by more than 95%.

[0039] ② Regarding system response latency, centralized processing typically has a total latency of 150-700ms due to network transmission (50-200ms) and cloud computing (100-500ms); this system relies on local inference at edge nodes (≤10ms) and near-end collaborative transmission (<5ms) to control the overall response latency within 30 milliseconds.

[0040] ③ Regarding network dependency and operational continuity, traditional architectures are highly dependent on network stability, and interruptions will cause the sorting process to stop. This system, through edge-layer local intelligent decision-making and model redundancy mechanisms, can continue to run independently after being disconnected from the cloud, in order to support production continuity.

[0041] Cloud layer: Global optimization and remote management terminal.

[0042] The cloud layer is deployed at the ground coal mine dispatch center. It consists of a cluster of multiple high-performance servers, equipped with a multi-feature fusion model optimization system, a distributed data management system, and a digital twin and visualization system, forming a comprehensive platform covering model optimization, data management, and visualization monitoring.

[0043] The multi-feature fusion model optimization system employs an online optimization algorithm based on incremental learning, automatically processing data uploaded from the edge layer on a 24-hour cycle. The system first performs multi-level validity screening on approximately 10,000 new samples daily: the first level filters out blurry images (Laplacian variance below 500), severely occluded images (effective contour area less than 30% of the original region), or obviously mislabeled samples; the second level, based on sample diversity, uses cluster analysis to ensure that the selected samples cover different coal-gangue ratios (10%-70%), lighting conditions (20-200 Lux), and typical noise scenarios (dust, water vapor splashing), ultimately retaining approximately 6,000-8,000 high-quality samples for training. Subsequently, the system uses these filtered samples to iteratively fine-tune the existing lightweight recognition model to continuously adapt to changes in the dynamic underground environment and maintain the model's generalization ability. The optimized model must pass an automated testing and verification process on an independent test set containing 2,000 images. The test set simulates complex downhole scenarios, including: noisy images with added Gaussian noise (standard deviation 5-15) and salt-and-pepper noise (density 1%-3%), occluded images with 20%-50% random rectangular occlusion, and low-light images with illumination varying randomly between 30-150 Lux. The model must achieve the following comprehensive performance metrics on this test set: average accuracy of at least 95.5%, recall of at least 94%, and model size limited to 15MB to pass validation. Validated models are then deployed to edge nodes for updates to address the impact of downhole environmental changes on the model's generalization ability.

[0044] The distributed data management system is based on the Hadoop architecture and classifies, compresses, and persistently stores data such as material batches, sorting accuracy, equipment parameters, and fault logs.

[0045] The digital twin and visualization system, based on the Unity3D engine, enables high-fidelity mapping and intelligent interaction between the physical system and the twin model. The 3D modeling data originates from multi-source data fusion of the downhole equipment entities: using the equipment's factory CAD drawings as the basic geometric framework, and employing high-precision laser scanning point cloud data (such as Leica BLK360) to perform real-world supplementary scanning of key equipment (such as the conveyor drive unit, rollers, and slag receiving plates). The point cloud registration accuracy is better than 2mm. Model accuracy is verified through coordinate comparison using a high-precision laser tracker: the actual 3D coordinates of key equipment are measured on-site and compared with corresponding positions in the twin model; point deviations are controlled within 5mm, ensuring that the final equipment 3D model's geometric dimensions and positional errors can be controlled within 5mm using this method. The synchronization frequency between the model and the physical system is 10Hz, consistent with the edge layer data upload frequency. The system receives structured data streams (including belt speeds, coal-to-gangue ratios, equipment status words, etc.) from edge master nodes in real time. Through preset coordinate mapping and kinematic models, it drives the corresponding models in the 3D scene to synchronously update their position, speed, motion status, and material flow animation, supporting sub-second virtual-real synchronization. The system has intelligent diagnostic and early warning functions. Its diagnostic logic combines status threshold judgment, timing analysis, and pattern recognition. For example, the early warning for bridging plate jamming is not only based on a single time threshold of "not reaching the preset position within 300ms after the command is issued," but also simultaneously monitors the servo motor current waveform. If the current continuously exceeds the limit or exhibits abnormal oscillations during operation, it is comprehensively judged as mechanical jamming or abnormal load. For belt misalignment detection, the system receives signals from paired laser beam sensors or vision detection units installed on both sides of the belt conveyor frame and calculates the offset (ΔL) of the belt edge relative to the baseline in real time. Early warning and alarm thresholds are dynamically set according to the belt width: when ΔL continuously exceeds 2% (early warning threshold) of the belt width for 1 second, or momentarily exceeds 5% (alarm threshold), the system triggers the corresponding level of alarm. The system can automatically diagnose the above-mentioned fault types and push alarm information containing fault location, possible causes and preliminary handling suggestions to the dispatch terminal based on the knowledge base (such as historical fault cases and equipment maintenance manuals).

[0046] A coal and gangue flow state perception and sorting strategy based on a 1.0-second short time window: This strategy uses a 1.0-second time window to continuously perceive and assess the state of the coal and gangue flow, replacing the traditional isolated judgment mode of "single gangue identification - single action". Experimental verification shows that the 1.0-second window length strikes a balance between instantaneous fluctuations and system response delays—shorter windows (e.g., 0.5 seconds) are more susceptible to noise, while longer windows (e.g., 2.0 seconds) lead to action lag. The thresholds were set based on sorting tests of 100 actual samples with different coal-gangue ratios (10%–70%), optimized through frequency division statistics and regression analysis: gangue occurrence frequency ≥6 times / second is defined as "gangue dense", ≤2 times / second as "coal dense", and 3–5 times / second as "mixed dense". This classification, using a comprehensive score (comprehensive missed selection rate and incorrect selection rate) as the evaluation index, demonstrated the highest sorting stability in the aforementioned tests. Meanwhile, the system integrates a multi-belt speed linkage model and a distributed redundant control mechanism. Through configurable edge-layer local intelligent decision-making and differentiated coal and gangue flow status to adapt sorting actions and belt speed regulation logic, it can cope with the complex underground environment and the system complexity of multi-point deployment. This reduces network bandwidth pressure, improves sorting response speed and identification accuracy, ensures the maximization of the gangue-coal ratio and the continuous and stable operation of the system, and reduces the system implementation cost.

[0047] Please refer to Figure 4 , Figure 4 This invention discloses a flowchart of an in-situ intelligent sorting method for coal gangue according to an embodiment of the present invention; the method mainly includes the following steps: Step 1: Multi-source data acquisition and preprocessing at the terminal layer.

[0048] After the feed conveyor belt starts, the terminal layer equipment operates synchronously. A binocular camera acquires images, 3D point clouds, and location data of the coal-gangue mixture. The image data is first processed by a low-light enhancement algorithm based on multi-scale Retinex (MSR). This algorithm uses three Gaussian surround scales (σ=15, 80, 250) to decompose and fuse the brightness components, enhancing dark details while suppressing halo effects. Subsequently, a 3×3 median filter is used to suppress impulse noise introduced by dust; this window size has been experimentally verified to achieve a balance between noise suppression and edge preservation. Shape data is extracted using the Canny edge detection algorithm to determine the coal-gangue contour and the number of edges. The high and low thresholds are set to 100 and 200, respectively. This threshold combination was optimized by performing gradient histogram analysis on 500 typical downhole sample images, ensuring continuous edge extraction while suppressing false edges. Location data is converted into precise coordinates in the conveyor belt coordinate system using binocular visual parallax calculation. The preprocessed image, shape, and location data are transmitted in real-time to the corresponding local edge computing node via industrial Ethernet.

[0049] Step 2: Multi-feature identification of edge computing nodes and judgment of coal gangue flow status.

[0050] After receiving data, the edge computing node performs dynamic decision-making in two steps. In the multi-feature fusion and recognition stage, the node loads a lightweight coal gangue recognition model based on an improved MobileNetV2 architecture. This model, designed for complex underground scenarios and edge computing constraints, incorporates several improvements over the standard MobileNetV2. Its core structure is as follows: Figure 5 As shown: While retaining depthwise separable convolutions to reduce computation, the backbone network embeds a channel attention mechanism (SE module) in key bottleneck layers to enhance the weights of feature channels relevant to coal gangue identification. The architecture incorporates a parallel "multi-feature early fusion module": in the shallow layers of the backbone network (e.g., after the extension layer of the third inverted_residual module), three lightweight dedicated branches are deployed in parallel to extract color features (after dimensionality reduction via 1×1 convolution, local grayscale statistical vectors are calculated in HSV space), texture features (texture descriptors are obtained through simplified LBP operator layers and pooling), and shape features (geometric attribute vectors of the contour are calculated in real time based on Canny edge results). These three heterogeneous feature vectors are immediately concatenated and normalized after extraction, and then input into a lightweight feature recalibration and fusion module. This module consists of a 32-dimensional fully connected layer and a sigmoid activation function, used to learn adaptive weights for each feature dimension.

[0051] The weighted and fused features are then concatenated with the output feature map of the backbone network at this layer using channel dimension, and jointly input into subsequent network layers for joint representation learning. For model lightweighting, three strategies are comprehensively applied: First, the backbone network after training convergence is structurally pruned, and redundant inverted_residual modules are evaluated and removed based on the average rank of the output feature maps; second, post-training quantization (PTQ) is used to uniformly quantize model weights and activation values ​​from FP32 to INT8 format; finally, knowledge distillation is introduced, using soft labels output by a large teacher model (such as ResNet50) trained in the cloud to guide the training of this lightweight student model.

[0052] Table 1. Performance comparison results of different coal gangue identification models As shown in Table 1, the model was finally deployed under the TensorFlow Lite framework, with an average inference time of ≤10ms on edge computing nodes. It can output high-confidence gangue identification results, accurate bounding boxes, and conveyor belt coordinates in real time. Subsequently, the system calculates the real-time gangue-to-coal ratio and gangue occurrence frequency within a short time window of 1.0 seconds. This window length was experimentally verified. In the comparative test of 100 samples with different gangue-to-coal ratios, the shorter duration (e.g., 0.5 seconds) was easily affected by instantaneous fluctuations, while the longer duration (e.g., 2.0 seconds) showed a sluggish response. In the comparative test of 100 samples with different gangue-to-coal ratios, the 1.0-second window had the highest comprehensive score for sorting performance (considering both missed and incorrect selection rates), with an average improvement of approximately 12% and 8% compared to the 0.5-second and 2.0-second windows, respectively. Subsequently, the system determines the current coal and gangue flow status based on preset thresholds: gangue dense (≥6 times / second), coal dense (≤2 times / second), and mixed dense (3~5 times / second). These frequency thresholds were optimized and determined after performing frequency division statistics and sorting effect regression analysis on the aforementioned 100 sets of sample data.

[0053] During the dynamic adjustment phase of the coal and gangue flow, the system dynamically adapts the action parameters of the gangue receiving plate based on the real-time status and the gangue-coal ratio: In the case of dense gangue, the extension duration of the receiving plate is gradually increased from the base value of 300ms to a range of 500-800ms, and the action interval is shortened from the default 150ms to 50-100ms. When the frequency consistently exceeds 8 times / second, the plate remains extended to cope with extremely dense conditions. In the case of dense coal, the plate only activates when gangue with a confidence level higher than 0.9 is detected. The system issues short-duration actions (duration ≤ 200ms). In mixed and dense conditions, the feeder belt speed and the contact plate's operating frequency are adjusted synchronously. The belt speed decreases linearly as the gangue-to-coal ratio increases (e.g., for every 0.1 increase in the gangue-to-coal ratio, the speed decreases by 5% of the rated value). The operating frequency is adjusted linearly within the range of 3-5 times / second based on the real-time gangue occurrence frequency (e.g., at 5 times / second, the operating interval is set to 100ms; at 3 times / second, it is set to 200ms). The edge computing node continuously uploads the real-time gangue-to-coal ratio, coal-gangue flow status, feeder belt speed, and contact plate operating parameters to the edge master node.

[0054] Step 3: Multi-belt collaborative speed regulation at the edge master node.

[0055] After receiving real-time data uploaded by all edge computing nodes, the edge master node executes multi-belt coordinated stepless speed regulation based on the dual principles of sensing the difficulty of coal gangue connection and matching the total clean coal flow. This system achieves smooth dynamic speed adjustment through a continuous function model, replacing traditional threshold segmented control, thus more accurately adapting to the complex changes in underground coal gangue flow.

[0056] Speed ​​regulation logic and parameter definition: The system first measures the real-time gangue content of each feed belt (sub-belt). (Percentage of gangue quality) The data is summarized, and the total clean coal flow rate is calculated. .in, Let be the material flow rate of the i-th sub-belt. The core principle of speed regulation is as follows: sub-belt speed... The difficulty of the coal receiving operation is determined by the coal content. It exhibits a non-monotonic relationship of "low at both ends and high in the middle": when →0 (More coal than gangue) or →1 (more gangue than coal), the gangue receiving plate operates simply and at a low frequency, making operation easier and allowing the sub-belt to maintain a higher speed; when When handling a mixture of coal and gangue, the gangue receiving plate requires frequent and precise movements, making the operation most challenging. The sub-belt must be slowed down to ensure sorting accuracy. Main belt speed. Total clean coal flow The results show a monotonically positive correlation: when the total clean coal flow is large, the main conveyor belt speeds up to match the transport capacity; when the total clean coal flow is small, the main conveyor belt speeds down to reduce energy consumption.

[0057] Mathematical model of continuously variable speed regulation: Sub-belt speed model (based on joint difficulty): Sub-belt target speed Its gangue content For a continuous function, a non-monotonic speed regulation with a "low in the middle and high on both sides" is achieved by using an upward-opening parabola: In the formula: This is the base speed of the sub-belt (m / s), typically taken as 70% to 90% of the rated belt speed; The difficulty influence coefficient (range 0.1~0.3) controls the rate of speed reduction under mixed operating conditions; Function characteristics: ①When =0 (pure coal) or When =1 (pure gangue), Maximum speed; ②When When =0.5 (half coal, half gangue), ( ), the speed is the smallest.

[0058] This function reflects the non-monotonic relationship between the difficulty of receiving coal and the coal content. It automatically reduces speed under mixed working conditions and automatically and steplessly adjusts speed up under working conditions with more coal or more coal.

[0059] Main conveyor belt speed model (based on total clean coal flow): Target speed of main conveyor belt Total clean coal flow rate For a monotonically increasing continuous function, smooth bounded speed regulation is achieved using the hyperbolic tangent function: In the formula: Rated speed of the main belt (m / s). σ is the baseline flow rate (t / h), usually taken as 50% of the design capacity; σ is the flow sensitivity coefficient (t / h), the rate of change of the control function; κ is the minimum speed proportional coefficient (usually taken as 0.5~0.7), to ensure that the main belt still maintains the basic operating speed when the flow rate is extremely low; It is the hyperbolic tangent function. Function characteristics: When hour, ;when hour, Achieve a smooth match between flow rate and speed.

[0060] Smooth speed transition and system protection: To ensure stable equipment operation and prevent material slippage, all speed adjustments use a linear ramp function for smooth transition. in, The maximum single-step speed change is calculated from the system acceleration limit (0.2 m / s²); `sign` is the sign function, used to determine the positive or negative sign of the input value, thereby determining the direction of speed adjustment. Simultaneously, the edge master control node monitors the speed difference between the main and sub-belts in real time to ensure it does not exceed the safe range, preventing material accumulation or breakage.

[0061] Step 4: Cloud layer optimization and redundancy collaboration.

[0062] The cloud and edge layers interact in real time, while also handling communication anomalies. Model iteration and deployment: The cloud-based model optimization system automatically collects approximately 10,000 newly uploaded valid samples daily from the edge layer, randomly selecting 80% as the training set and 20% as the validation set for incremental fine-tuning of the existing lightweight recognition model. The fine-tuning process uses the cross-entropy loss function, employs the Adam optimizer with an initial learning rate of 1e-4, and iterates on the training set for 5 epochs. After each epoch, the model performance is evaluated on the validation set. If the validation set accuracy improves and the overfitting risk is controllable (the ratio of training loss to validation loss is stable), the current model parameters are saved. After fine-tuning, the system deploys the optimized model to all edge computing nodes via Ethernet. Upon receiving the model, the nodes, based on their built-in rolling update and validity verification mechanisms, automatically replace the old model after successful verification and back it up to a locally stored rolling version repository. Data synchronization and monitoring: The cloud-based data management system receives global operational data uploaded by the edge master control node at a frequency of 10Hz, classifies, compresses, and persists the data, and generates production reports; the digital twin system updates the status of downhole equipment in real time, and dispatchers can monitor global operational information and receive fault warnings through a large screen.

[0063] Offline Redundancy Processing and Data Synchronization: When the edge master node detects a communication interruption with the cloud via a heartbeat mechanism, it immediately sends an "offline run" command to all edge computing nodes. Each edge computing node then switches to the latest validated optimized model in its local model repository and continues to independently execute the identification and sorting decision logic, ensuring production continuity. During the communication interruption, the edge master node opens a dedicated offline data cache in its local storage (such as an industrial-grade solid-state drive), encapsulating and storing all real-time data packets from the edge computing nodes (including coal-gangue ratio, flow rate, equipment status, etc.) and its own speed adjustment command logs in a time-series format. Each data record is appended with a timestamp (1ms precision) and a data hash checksum. When the heartbeat detects that the network link has been restored and communication has been stable for 5 consecutive cycles (10 seconds), the edge master node determines that communication has been restored and automatically triggers the offline data synchronization process. The synchronization process employs a block-based verification and breakpoint resumption mechanism: The edge master node first packages the data in the cache according to time windows (e.g., one data block every 5 minutes), calculates the hash value of each data block, generates a synchronization list, and uploads it to the cloud. The cloud data management system receives the list, verifies its integrity, and provides feedback on missing or retransmitted data blocks. The edge master node retransmits the necessary data blocks based on the feedback until the cloud confirms that all offline data has been received completely. For potential model or parameter conflicts during synchronization (e.g., the cloud generates a new model version during offline processing, while the edge node generates a large amount of data based on the old model), the system follows a "data-first, model-progressive" conflict resolution strategy: First, it ensures that all offline production data is completely synchronized to the cloud as real feedback for model iteration; when processing this new data, the cloud model optimization system merges it with the existing training set and considers it in the new round of incremental training, thereby absorbing the experience from offline operation into the next generation model to support continuous data-driven model evolution. Once synchronization is complete, the cloud-based data management system updates the global database, and the model optimization system can supplement or adjust the model fine-tuning plan based on the new synchronized dataset to maintain the consistency and traceability of system data and model versions.

[0064] Example: To more specifically illustrate the implementation methods and technical effects of the present invention, the following description is based on the actual application scenario of a large-scale mine (annual production of 3 million tons, thick coal seam fully mechanized mining face). After being crushed, the raw coal in this working face is transferred via three parallel feeder belts (1.2 meters wide, rated capacity 800 tons / hour) to a main conveyor belt (1.4 meters wide, rated capacity 2500 tons / hour) for transport. The underground environment presents typical complex working conditions such as uneven illumination, high dust concentration, and local network instability.

[0065] When applying the system of this invention, an integrated device is installed at the terminal layer approximately 2 meters behind the unloading point of each feeding belt conveyor head. The binocular camera is an industrial model with 5 megapixels, a 30fps frame rate, and a built-in adaptive low-light enhancement module. The servo drive connecting plate uses a servo system with a response time of 80 milliseconds and a stroke of 1200 mm. The belt frequency converter driver is adapted to an 18.5 kW power rating. At the edge layer, an edge computing node based on a high-performance embedded AI chip is deployed near each feeding belt. This node has a built-in lightweight optimized multi-feature fusion recognition model (initial size approximately 12.3 MB) and a model version library built into local solid-state storage, supporting the use of verified local models for recognition in case of network anomalies. An industrial-grade edge master control node is set up at the main conveyor belt drive unit to run the multi-belt speed coordination model. The cloud layer is located at the ground dispatch center, where a dual-server cluster supports model optimization, data management, and the 3D digital twin system.

[0066] After a week of parameter calibration following deployment, the system was put into continuous operation. During operation, terminal devices collect images and location information of the coal-gangue mixture in real time, which are then preprocessed and transmitted to local edge computing nodes for real-time identification and status judgment, with an average inference latency of less than 10 milliseconds. The edge computing nodes dynamically adjust the action parameters of the receiving plate based on the real-time coal-gangue flow status, and simultaneously upload data such as the coal-gangue ratio and material flow rate to the edge master control node. The edge master control node calculates the target speed of each belt through the multi-belt speed linkage model based on the aggregated global data (such as the average coal-gangue ratio and total clean coal flow rate), and performs smooth adjustment using a ramp function to complete coordinated speed regulation. The cloud system automatically collects approximately 10,000 new samples daily for incremental learning and model optimization, and iteratively distributes the validated models. The digital twin system simultaneously realizes three-dimensional real-time monitoring and intelligent early warning of the underground scene.

[0067] During three consecutive months of operation, the system performance data is as follows: the average gangue identification accuracy reached 96.7%, the recall rate was 95.9%, the gangue detection rate exceeded 97.5% during peak gangue periods, the missorting rate of clean coal during peak coal periods was less than 1.8%, and the overall gangue-to-coal ratio improved by approximately 18% compared to before application; the overall system response latency was controlled within 30 milliseconds, the speed coordination adjustment response time was approximately 1.5 seconds, and the overshoot was less than 4%; through belt speed optimization, the energy consumption of the main transportation system was reduced by approximately 15%; during a 45-minute network simulation interruption, the system continued to operate relying on the edge layer local model version library and redundant decision-making mechanism without any sorting interruption. This embodiment demonstrates that the system of the present invention can effectively adapt to the complex underground environment, achieving efficient, accurate, and reliable sorting of coal gangue, and has significant practical value and promotion prospects.

[0068] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A belt conveyor-based intelligent in-situ coal gangue sorting system, characterized in that, The system adopts a cloud-edge-device collaborative distributed architecture, including a terminal layer, an edge layer, and a cloud layer; The terminal layer is used to collect multi-dimensional feature data of materials on the feeding belt in real time, and to receive instructions to perform sorting actions and adjust belt speed. The edge layer is communicatively connected to the terminal layer and the cloud layer. It is used to receive the multi-dimensional feature data of the material, perform material identification and coal gangue flow status determination locally, and generate sorting control instructions and belt coordinated speed adjustment instructions based on the determination results and send them to the terminal layer. The edge layer has local model storage and offline decision-making capabilities. The cloud layer is used to aggregate the system's global operating data, iteratively optimize the recognition model and distribute it to the edge layer, and remotely monitor and manage the system.

2. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 1, characterized in that, The terminal layer includes: a binocular camera, a servo drive receiving plate, and a belt frequency converter driver installed at each feeding belt; The binocular camera is equipped with an adaptive low-light enhancement module, which is based on the multi-scale Retinex algorithm. It decomposes the image brightness components through a Gaussian filter kernel and dynamically adjusts the gain coefficient in combination with feedback from the ambient light intensity sensor. The binocular camera is used to simultaneously acquire two-dimensional images, three-dimensional shapes, and belt coordinate position data of the coal gangue mixture.

3. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 1, characterized in that, The edge layer includes: edge computing nodes and edge master nodes; The edge computing node has a built-in lightweight multi-feature fusion recognition model. The model uses a parallel structure to extract color features, texture features and shape features respectively, and outputs the gangue confidence score after weighted fusion through a fully connected network. The backbone network of the model is embedded with a channel attention mechanism and uses structured pruning and quantization techniques for lightweight processing. The edge master node is used to aggregate real-time coal-gangue ratio and flow data uploaded by each edge computing node, and to run the multi-belt speed linkage model to calculate speed commands.

4. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 1, characterized in that, The edge layer is configured with a coal and gangue flow state sensing and sorting strategy based on a short time window; The strategy uses a 1.0-second time window to count the frequency of gangue occurrence within the window, classifying the coal and gangue flow state into three categories: gangue-dense, mixed-dense, and coal-dense. When the state is determined to be gangue-dense, the duration of the gangue receiving plate extension is extended and the action interval is shortened. When the state is determined to be coal-dense, a short-term action is triggered only when high-confidence gangue is detected. When the state is determined to be mixed-dense, the feed belt speed and the gangue receiving plate action frequency are adjusted proportionally.

5. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 3, characterized in that, The multi-belt speed linkage model adjusts the speed based on the difficulty of receiving coal and the total clean coal flow rate. Regarding the speed of the feed belt Using gangue content The parabolic function model with the gangue content as the independent variable maintains the base speed when the gangue content approaches 0 or 1, and automatically reduces the speed when the gangue content approaches 0.5; for the speed of the main conveyor belt... The total clean coal flow rate is adopted. The hyperbolic tangent function model with the independent variable monotonically increases with increasing flow rate; All speed adjustments are smoothly transitioned using a linear ramp function.

6. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 3, characterized in that, The edge layer has a distributed redundancy control mechanism; the edge computing nodes locally store a model version library containing optimized models of the most recent versions; the edge master control node monitors the communication status with the cloud layer through heartbeat packets, and when a communication interruption is detected, it issues an offline operation command, and the edge computing node switches to the locally verified model to continue running; after communication is restored, the data during the offline period is synchronized through block verification and breakpoint resume mechanism.

7. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 1, characterized in that, The cloud layer deploys a multi-feature fusion model optimization system; the model optimization system is based on an incremental learning algorithm, which performs validity screening on the sample data uploaded from the edge layer, removes blurry or severely occluded samples, and uses the screened high-quality samples to iteratively fine-tune the recognition model; After the fine-tuned model passes validation on an independent test set containing noisy, occluded, and low-light images, it is deployed to the edge layer for updates.

8. The in-situ intelligent sorting system for coal gangue on a belt conveyor according to claim 1, characterized in that, The cloud layer also deploys a digital twin and visualization system; this system is built on the Unity3D engine and drives the synchronous update of the 3D model by receiving structured data streams uploaded from the edge layer in real time; the system has intelligent diagnostic functions, which can determine the jamming fault of the connecting plate by combining the servo motor current waveform, and determine the belt misalignment fault by monitoring the belt edge offset.

9. A method for in-situ intelligent sorting of coal gangue using a belt conveyor, characterized in that, The method is used to implement the in-situ intelligent sorting system for coal gangue on a belt conveyor as described in any one of claims 1 to 8, and the method includes the following steps: Step 1: The terminal layer acquires image and shape data of the coal and gangue mixture, which is then preprocessed and transmitted to the edge layer; Step 2: The edge layer uses a lightweight multi-feature fusion recognition model for real-time identification and calculates the frequency of gangue occurrence within a short time window. Based on this, the coal and gangue flow status is determined to dynamically adjust the action parameters of the gangue receiving plate. Step 3: The edge layer calculates and issues target speed commands for each belt based on the real-time gangue content and material flow rate using a multi-belt speed linkage model, thereby achieving coordinated speed regulation; Step 4: The cloud layer periodically acquires sample data uploaded by the edge layer for incremental model training and optimization, and then distributes the updated model to the edge layer.

10. The in-situ intelligent sorting method for coal gangue using a belt conveyor according to claim 9, characterized in that, The multi-belt speed linkage model described in step three calculates the target speed of the feeding belt. The formula is: in, For the base speed of the sub-belt, The difficulty level is affected by the coefficient. Real-time gangue content; Calculate the target speed of the main conveyor belt The formula is: in, Main belt rated speed, This is the minimum speed proportionality coefficient. The total clean coal flow rate, As a baseline value for flow rate, For flow sensitivity coefficient, It is the hyperbolic tangent function.

Images