A cleaner for belt conveyor based on visual detection

Abstract

The present invention provides a belt conveyor cleaner based on visual detection, comprising a multimodal data acquisition unit, a multimodal data preprocessing unit, a model unit, a dynamic cleaning execution unit, and a cleaning effect feedback unit. The cleaning effect feedback unit is used to evaluate the cleaning effect. If the cleaning effect is not up to standard, the reinforcement learning controller adjusts the cleaning parameters according to the evaluation results until the preset standard is met. Accurate residue detection and adaptive cleaning are achieved through synergistic effects, covering multimodal data acquisition, model building, dynamic path planning, and cleaning effect evaluation, ensuring high-precision residue identification and efficient cleaning under complex working conditions. These effects enable the cleaner to maintain excellent performance in environments with high dust, wet materials, or extreme temperatures, while reducing operating costs and environmental impacts. The cleaner not only solves the contradiction between cleaning efficiency and belt protection, but also achieves a comprehensive improvement in intelligence and high adaptability, significantly surpassing traditional technologies.

Description

Technical Field

[0001] The present invention relates to the technical field of conveying auxiliary equipment, and in particular to a cleaner for a belt conveyor based on visual detection. Background Art

[0002] Belt conveyors are indispensable material transportation equipment in industries such as mining, ports, electric power, and chemicals, and their operating efficiency directly affects the production process. However, material adhesion and residue often cause belt surface deviation, wear, reduced conveying efficiency, and dust pollution. To address these problems, mechanical cleaning devices, such as scrapers or rotating brushes, are commonly used in the existing technology to remove residues through physical contact. However, such devices lack specificity for different types of residues, such as hard blocks, wet and sticky materials, and fine dust. Incomplete cleaning or excessive cleaning can easily cause increased belt wear and shorten the lifespan by approximately 20%-30%. In addition, mechanical cleaning devices require frequent maintenance and high component replacement costs, making it difficult to strike a balance between high-efficiency cleaning and belt protection.

[0003] With technological advancements, visual inspection technology has begun to be applied to the condition monitoring of conveyor belts to identify residues or defects on the belt surface. Existing visual inspection systems typically rely on a single industrial camera to detect belt status through image analysis. However, in complex environments such as high dust, wet materials, or low light, the detection accuracy of a single visual sensor decreases significantly, with false detection and missed detection rates as high as 15%-20%. Due to the lack of multimodal data fusion, such as ultrasonic or vibration signals, existing systems struggle to fully capture the type, thickness, viscosity, and distribution characteristics of residues. This results in a lack of precise guidance for subsequent cleaning strategies and limits the overall cleaning effect.

[0004] Existing cleaning devices also have shortcomings in their intelligence. They typically use fixed cleaning paths and parameters, making it difficult to dynamically adjust strategies based on belt speed, material type, or environmental changes. This results in significant fluctuations in cleaning efficiency under complex operating conditions, particularly with wet and sticky materials. Furthermore, existing technologies lack a data-driven, long-term optimization mechanism, preventing continuous improvement in cleaning performance over time. Maintenance often relies on manual inspections, resulting in long periods of unplanned downtime and high maintenance costs, limiting the widespread application of cleaning devices under dynamic conditions. Summary of the Invention

[0005] The purpose of the present invention is to provide a sweeper based on multimodal perception and intelligent optimization, which significantly improves cleaning efficiency, detection accuracy and environmental adaptability through collaborative structure and method.

[0006] To achieve the above objectives, the present invention proposes the following technical solution: a belt conveyor cleaner based on visual detection, comprising:

[0007] A multimodal data acquisition unit, which collects belt surface images in real time, measures the three-dimensional morphology of residues using ultrasonic data to generate point cloud data, and simultaneously records belt running vibration signals to reflect residue distribution;

[0008] a multimodal data preprocessing unit, which performs Gaussian filtering to remove noise and histogram equalization to enhance contrast on the belt surface image to generate preprocessed image data, performs time-domain filtering to remove noise on the ultrasonic data to generate preprocessed three-dimensional topography data, and simultaneously performs Fourier transform on the vibration signal to extract frequency domain features to generate preprocessed vibration spectrum data;

[0009] A model unit, comprising a multimodal residue identification model and a cleaning strategy optimization model, wherein the multimodal residue identification model is used to generate a residue distribution, the cleaning strategy optimization model is used to generate an optimal path and parameters, and the model unit performs intelligent residue classification and cleaning priority assessment;

[0010] Dynamic cleaning execution unit, the dynamic cleaning execution unit includes:

[0011] The cleaning head module includes a scraper, a rotating brush, a high-pressure air nozzle, an ultrasonic cleaner, a robotic arm, a torque sensor, an adaptive liquid spray device, a pressure sensor, a fan, a heating element, and a PLC controller. The PLC controller drives the robotic arm to execute the path based on the data of the model unit. The cleaning head module selects tools based on the data of the model unit. The torque and pressure sensors provide real-time feedback to dynamically adjust the angle and pressure of the cleaning head module. The fan and heating element adjust the temperature of the cleaning position.

[0012] The cleaning effect feedback unit is used to evaluate the cleaning effect. If the cleaning effect does not meet the standard, the reinforcement learning controller adjusts the cleaning parameters according to the evaluation results until the preset standard is met.

[0013] Furthermore, in the present invention, the multimodal data acquisition unit includes:

[0014] A high-definition industrial camera array, which collects a belt surface image I (x, y) in real time;

[0015] An ultrasonic sensor array, wherein the ultrasonic sensor array measures the three-dimensional morphology of the residue to generate point cloud data H (x, y, z);

[0016] Vibration sensor: The vibration sensor records the belt running vibration signal V(t) to reflect the distribution of residues;

[0017] The LED light source system is installed on the belt support frame and automatically adjusts the brightness according to the ambient light intensity to ensure clear images.

[0018] Furthermore, in the present invention, the multimodal data preprocessing unit is an edge computing node, using an ARM Cortex-A78 processor, 16GB of memory, 10TFLOPS of computing power, one per module, and supporting 5G / Wi-Fi6 communication;

[0019] The edge computing node performs Gaussian filtering denoising and histogram equalization on the image I(x, y) to enhance the contrast and generate the preprocessed image data I p (x, y), perform time domain filtering on the ultrasonic data H(x, y, z), eliminate noise, and generate pre-processed three-dimensional shape data H p (x, y, z), perform Fourier transform on the vibration signal V(t), extract frequency domain features, and generate preprocessed vibration spectrum data V p (f).

[0020] Furthermore, in the present invention, the model unit includes a multimodal data fusion unit, a reinforcement learning controller and a cloud-based digital twin platform. The multimodal data fusion unit adopts the NVIDIA Orin embedded AI chip with a computing power of 200TOPS and a memory of 32GB, and runs the Transformer model. The reinforcement learning controller adopts the NVIDIA Jetson AGX Orin with a computing power of 275TOPS and a memory of 64GB, and runs the PPO algorithm. The cloud-based digital twin platform adopts the Intel Xeon server with 128GB of memory and 10TB of storage, and supports model training and parameter optimization.

[0021] Furthermore, in the present invention, the multimodal residue recognition model is as follows:

[0022] S(x, y)=w1·φ(I p (x,y))+w2·ψ(H p (x, y, z))+w3·γ(V p (f)); S(x, y): residue distribution score, φ(I p (x, y)): visual features, ψH p (x, y, z): ultrasonic three-dimensional characteristics, γV p (f): Vibration spectrum characteristics, w1, w2, w3: adaptive weights, w1+w2+w3=1;

[0023] NVIDIA Orin embedded AI chip runs YOLOv8 to extract φ(I p (x, y)), analyze the residue type and boundary, and process the ultrasonic point cloud data to calculate ψH p (x, y, z), determine the thickness, and analyze the vibration spectrum to generate γVp (f) indirectly reflects the distribution. The Transformer model integrates features, optimizes w1w, 2, w3, calculates S(x, y), and trains the Transformer model on the cloud server to optimize the loss function:

[0024] L1: loss function of the multimodal recognition model, used to optimize S(x, y), including mean square error and KL divergence, N: number of training samples, S i : The residual score predicted by the model, i represents the i-th sample. True residue score, KL divergence, which measures the predicted distribution P(S i ) and the true distribution The difference is used for regularization, λ: regularization coefficient, which controls the weight of KL divergence.

[0025] Furthermore, in the present invention, the cleaning strategy optimization model is as follows:

[0026] Opt(P, Θ): cleaning strategy optimization objective, representing the optimal cleaning path P and parameter Θ, calculated by Jetson AGX Orin using the PPO algorithm, P: cleaning path, Θ: cleaning parameters, including cleaning head type, angle, and pressure, T(P): cleaning time, E(Θ): energy consumption, W(P, Θ): belt wear, α, β, γ: dynamic weights;

[0027] Reward function: R = η1·CR-η2·T(P)-η3·E(Θ)-η4·W(P, Θ), R: reward function, which measures the comprehensive performance of the cleaning strategy. The larger the value, the better the strategy. CR: cleaning rate, which indicates the proportion of residues removed. η1, η2, η3, η4: reward function coefficients, which correspond to the weights of cleaning rate, time, energy consumption and wear respectively.

[0028] The above model optimizes the cleaning path and parameters, balances cleaning efficiency, time, energy consumption and wear, and the reward function R drives the PPO algorithm to iteratively improve the strategy.

[0029] Furthermore, in the present invention, the cleaning effect feedback unit also includes a high-definition industrial camera array, an ultrasonic sensor array, a vibration sensor and an LED light source system. The perception module of the cleaning effect feedback unit re-collects the belt surface image I(x, y), the three-dimensional morphology generates point cloud data H(x, y, z) and the vibration signal V(t), updates S(x, y), and calculates the cleaning rate CR. If CR<98%, the PPO algorithm adjusts P, Θ, and the cloud server updates L1 and R to optimize the model parameters.

[0030] Furthermore, in the present invention, the high-definition industrial camera array includes three 1080p cameras with wide-angle and zoom lenses, supports infrared imaging, and a frame rate of ≥60fps; the ultrasonic sensor array includes 12 probes with a measurement range of 0-50cm and a resolution of ≤1mm; the vibration sensor includes four high-frequency accelerometers with a sampling rate of ≥10kHz and a sensitivity of 0.1m / s 2 ;

[0031] The camera array is installed on steel brackets above and on both sides of the belt conveyor, 50-80 cm from the belt surface. The main camera is located above the center, and the cameras on both sides are tilted 45 degrees.

[0032] The ultrasonic sensor array is distributed on the support beams on both sides of the belt, 10-20 cm away from the belt;

[0033] The vibration sensor is fixed to the key nodes of the supporting roller or frame;

[0034] The camera, ultrasonic sensor, and vibration sensor are connected to the NVIDIA Orin chip via shielded cables. The LED light source is connected to the NVIDIA Orin chip via a PWM controller. The NVIDIA Orin chip is connected to the edge computing node via Ethernet.

[0035] Furthermore, in the present invention, the maximum load of the robotic arm is 50kg, the positioning accuracy is ±0.1mm, it is equipped with a torque sensor with a resolution of 0.01Nm, the nozzle flow of the adaptive spray device is 0.1-1L / min, it sprays environmentally friendly cleaners or anti-stick coatings, and the pressure sensor has a range of 0-100N and a resolution of 0.1N.

[0036] Furthermore, in the present invention, the frequency of the ultrasonic cleaner is 40kHz, which is used to process wet and sticky materials. The cleaning head is connected to the slide rail through a servo motor and supports independent lifting and switching.

[0037] Beneficial effects: The technical solution of this application has the following technical effects:

[0038] Belt conveyor cleaners based on visual inspection feature efficient cleaning performance, optimized belt protection, excellent environmental adaptability, and extended equipment life. They achieve precise residue detection and adaptive cleaning through synergistic effects, encompassing multimodal data acquisition, model building, dynamic path planning, and cleaning effect evaluation, ensuring high-precision residue identification and efficient cleaning under complex working conditions. These effects enable the cleaner to maintain excellent performance in environments with high dust, wet materials, or extreme temperatures, while reducing operating costs and environmental impact. These beneficial effects stem from the synergistic effects of the underlying mechanisms of multimodal perception, intelligent control, adaptive execution, and long-term optimization. This not only resolves the contradiction between cleaning efficiency and belt protection, but also achieves a comprehensive improvement in intelligence and high adaptability, significantly surpassing traditional technologies.

[0039] It should be appreciated that all combinations of the foregoing concepts, as well as additional concepts described in greater detail below, to the extent such concepts are not mutually inconsistent, can be considered to be part of the inventive subject matter of this disclosure.

[0040] The foregoing and other aspects, embodiments, and features of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as features and / or beneficial effects of the exemplary embodiments, will become apparent from the following description or through practice of specific embodiments according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For the sake of clarity, not every component is labeled in every figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings, in which:

[0042] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0043] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings. Various aspects of the present invention are described in this disclosure with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the present disclosure are not necessarily defined to include all aspects of the present invention. It should be understood that the various concepts and embodiments introduced above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed in the present invention are not limited to any implementation method. In addition, some aspects disclosed in the present invention can be used alone or in any appropriate combination with other aspects disclosed in the present invention.

[0044] Example 1: System implementation of a belt conveyor cleaner based on visual inspection

[0045] This example describes a specific implementation of a visual inspection-based belt conveyor cleaner, suitable for belt conveyor cleaning scenarios in mining, port, and chemical industries. By integrating multimodal data acquisition, preprocessing, model analysis, dynamic cleaning execution, and feedback, the system achieves efficient and accurate residue removal while protecting the belt and adapting to complex operating conditions.

[0046] The cleaner consists of the following main units:

[0047] Multimodal data acquisition unit: used to collect multi-dimensional information of the belt surface.

[0048] Multimodal data preprocessing unit: denoising and feature extraction of collected data.

[0049] Model unit: includes residue identification model and cleaning strategy optimization model.

[0050] Dynamic cleaning execution unit: performs adaptive cleaning operations.

[0051] Cleaning effect feedback unit: evaluates cleaning effect and optimizes parameters.

[0052] The specific implementation of the multimodal data acquisition unit is as follows. The multimodal data acquisition unit is deployed on steel brackets above and on both sides of the belt conveyor, specifically including:

[0053] High-definition industrial camera array: Utilizes three 1080p industrial cameras with wide-angle and zoom lenses, infrared imaging support, and a frame rate of 60 fps or higher. The main camera is mounted above the belt, 60 cm from the surface; two side cameras are tilted 45° and 50 cm from the surface. The cameras capture a real-time image of the belt surface I(x, y) with a resolution of 1920 × 1080.

[0054] Ultrasonic sensor array: This array includes 12 ultrasonic probes with a measurement range of 0-50 cm and a resolution of ≤1 mm. These probes are located on the support beams on both sides of the belt, 15 cm from the belt. The sensors generate a 3D point cloud of the residue's topography, H(x, y, z).

[0055] Vibration sensor, 4 high-frequency accelerometers, sampling rate ≥ 10kHz, sensitivity 0.1m / s 2 , fixed at the key nodes of the support roller and the frame, and record the belt running vibration signal V(t).

[0056] LED light source system: installed on the belt support frame, equipped with a PWM controller, automatically adjusts the brightness according to the ambient light intensity, ranging from 0-1000 lux to ensure clear image acquisition.

[0057] All sensors are connected to the NVIDIA Orin embedded AI chip via shielded cables and communicate with edge computing nodes via Ethernet.

[0058] The multimodal data preprocessing unit is implemented as follows: It uses an edge computing node with an ARM Cortex-A78 processor, 10 TFLOPS of computing power, and 16GB of memory. The communication module supports 5G and Wi-Fi 6 to ensure real-time data transmission.

[0059] The preprocessing algorithm performs Gaussian filtering to denoise the image I(x, y) and enhances the contrast through histogram equalization to generate a preprocessed image Ip(x, y).

[0060] The ultrasonic point cloud data H(x, y, z) is filtered in the time domain with a cutoff frequency of 50 Hz to eliminate noise and generate preprocessed three-dimensional shape data Hp(x, y, z).

[0061] Perform fast Fourier transform on the vibration signal V(t), extract the frequency domain features 0-5kHz, and generate the vibration spectrum data V p (f).

[0062] The preprocessed data is transmitted to the model unit via 5G / Wi-Fi6 with a delay of ≤10ms.

[0063] The model unit includes a multimodal residue recognition model, a cleaning strategy optimization model, a multimodal data fusion unit, a reinforcement learning controller and a cloud-based digital twin platform.

[0064] Multimodal residue recognition model, hardware includes NVIDIA Orin embedded AI chip, computing power of 200TOPS, and memory of 32GB.

[0065] The algorithm runs YOLOv8 to extract visual features φ(Ip(x, y)) and identify the type and boundary of the residue; processes the ultrasonic point cloud data Hp(x, y, z) and calculates the thickness of the residue ψ(Hp(x, y, z)); analyzes the vibration spectrum V p (f), generating distribution features γ(V p (f)).

[0066] Feature fusion, using the Transformer model to fuse multimodal features and calculate the residual distribution score: S(x, y) = w1·φ(Ip(x, y))+w2·ψ(Hp(x, y, z))+w3·γ(Vp(f));

[0067] Among them, w1, w2, and w3 are adaptive weights, which are optimized through cloud training.

[0068] Loss function:

[0069]

[0070] Among them, N is the number of training samples, S i Score the predictions, is the true score, λ=0.1 is the regularization coefficient, D KL is the KL divergence.

[0071] Cleaning strategy optimization model, hardware NVIDIA Jetson AGX Orin, 275TOPS computing power, 64GB memory. Algorithm: Uses the PPO reinforcement learning algorithm to optimize the cleaning path P and parameters θ, as well as the cleaning head type, angle, and pressure.

[0072] Optimization goal: Opt(P, θ)=min(α·T(P)+β·E(θ)+γ·W(P, θ));

[0073] Among them, T(P) is the cleaning time, E(θ) is the energy consumption, W(P, θ) is the belt wear; α, β, γ are dynamic weights.

[0074] Reward function: R = η1·CR-η t rajetória2·T(P)-η3·E(θ)-η4·W(P, θ);

[0075] Where CR is the cleaning rate, η1 = 0.5, η2 = 0.2, η3 = 0.2, and η4 = 0.1. The PPO algorithm generates the optimal cleaning strategy by iteratively optimizing R.

[0076] Cloud-based digital twin platform, Intel Xeon server hardware, 128GB memory, and 10TB storage.

[0077] The function updates the Transformer and PPO model parameters in real time, stores historical data, trains the loss function L1 and reward function R, and supports long-term optimization.

[0078] The specific implementation of the dynamic cleaning execution unit includes a cleaning head module and a PLC controller;

[0079] The cleaning head module includes a scraper for hard blocks, a rotating brush for dust, a high-pressure air nozzle for light residues, and an ultrasonic cleaner with a frequency of 40kHz for wet and sticky materials.

[0080] The robotic arm has a maximum load of 50kg, a positioning accuracy of ±0.1mm, and is equipped with a torque sensor and a pressure sensor. The adaptive spray device, with a nozzle flow rate of 0.1-1L / min, can spray environmentally friendly cleaning agents or anti-stick coatings.

[0081] Auxiliary components: fans are used to adjust the temperature of the cleaning position, heating elements are used to handle wet materials, and servo motors and slide rails are used to support the lifting and switching of the cleaning head.

[0082] The PLC controller drives the robotic arm to perform cleaning based on the path P and parameter θ output by the model unit. Torque and pressure sensors provide real-time feedback, dynamically adjusting the cleaning head angle and pressure.

[0083] The fan and heating element regulate the temperature of the cleaning location according to the ambient temperature.

[0084] The cleaning effect feedback unit reuses the sensors of the multimodal data acquisition unit to re-collect data:

[0085] Data acquisition: High-definition industrial camera arrays capture images, ultrasonic sensors generate point clouds, and vibration sensors record signals.

[0086] Effect evaluation: Update the residue distribution score S(x, y) and calculate the cleaning rate CR:

[0087]

[0088] Among them, S before and S after The residue was scored before and after cleaning.

[0089] Optimization and adjustment: If CR < 98%, the reinforcement learning controller PPO algorithm adjusts the path P and parameter θ, and the cloud server updates the loss function L1 and reward function R to optimize the model parameters until CR ≥ 98%.

[0090] The implementation effect is as follows: the test on a mine belt conveyor with a belt width of 1.2m and a running speed of 2m / s shows that

[0091] Cleaning efficiency: Cleaning rate CR ≥ 98%, effective for hard blocks, wet sticky materials and dust, better than traditional mechanical cleaners.

[0092] Belt protection: Belt wear rate is reduced by about 25% and belt life is extended by about 30%.

[0093] Environmental adaptability: In high dust visibility <1m, wet material humidity >80% and low temperature -10℃ environment, the cleaning rate remains ≥95%.

[0094] Operating costs: Maintenance frequency reduced by 50%, energy consumption reduced by 20%, and detergent usage reduced by 30%. Intelligence level: The system is continuously optimized through a cloud-based digital twin platform. After six months of operation, cleaning time was reduced by 15% and energy consumption by 10%.

[0095] This embodiment achieves high-precision residue identification and efficient cleaning through the synergy of multimodal perception, intelligent models, and adaptive execution, addressing the low efficiency, severe belt wear, and poor environmental adaptability of traditional cleaners. The system performs exceptionally well under complex operating conditions, significantly reducing operating costs and environmental impact, and possesses broad industrial application prospects.

[0096] Example 2: Optimized application of cleaning wet sticky materials

[0097] This embodiment optimizes the parameters of the cleaner to improve the processing efficiency for the cleaning scenario of wet sticky materials such as clay or wet coal.

[0098] System configuration: Same as Example 1, except that the cleaning head module preferably uses an ultrasonic cleaner with a frequency of 40 kHz and an adaptive liquid spray device, spraying an environmentally friendly cleaning agent at a flow rate of 0.5 L / min. The heating element is turned on to maintain the cleaning position temperature at 40°C.

[0099] Data acquisition and processing, camera array: Enable infrared imaging mode to enhance wet material boundary recognition.

[0100] Ultrasonic sensor: Increase the sampling frequency to 20Hz and accurately measure the thickness of wet materials from 0 to 30mm.

[0101] Vibration sensor: focuses on analyzing low-frequency vibration (0-500Hz) to reflect the distribution of wet materials.

[0102] Model optimization, residue identification model: Increase the weight of wet material features and improve the contribution of ultrasonic data to S(x, y). Cleaning strategy optimization model: Adjust the reward function, increase the cleaning rate coefficient, reduce the energy consumption weight, and prioritize cleaning results.

[0103] During cleaning, the cleaning head module mainly uses an ultrasonic cleaner, combined with a high-pressure air nozzle to assist in removing residual materials.

[0104] The robot arm path P is optimized to an S-shaped trajectory, covering the wet material concentration area, with an angle adjustment range of ±20° and a pressure maintenance of 50N.

[0105] The adaptive spray device sprays the anti-stick coating to reduce secondary adhesion.

[0106] Effect evaluation, tested on a wet coal conveyor, humidity 85%, material thickness 5-20mm:

[0107] The cleaning rate CR reaches 99%, which is better than traditional cleaners.

[0108] Reduce cleaning time by 10% and use 20% less detergent.

[0109] The belt surface adhesion rate is reduced by 40%, and the maintenance cycle is extended by 2 times.

[0110] By optimizing the cleaning strategy for wet and sticky materials, this embodiment significantly improves the cleaning efficiency and belt protection effect, and verifies the high adaptability of the sweeper under specific working conditions.

[0111] Example 3: Cleaning application in high dust environment

[0112] This embodiment optimizes the performance of the sweeper for high dust environments, such as cement plants, and ensures stable operation under low visibility conditions.

[0113] The system configuration is the same as in Example 1, but with an increased LED light source brightness of 1000 lux and a camera array that preferentially uses a wide-angle lens. The cleaning head module primarily features a rotating brush and a high-pressure air nozzle.

[0114] Data collection and processing:

[0115] Camera array: Enhanced histogram equalization to improve image contrast in dusty environments.

[0116] Ultrasonic sensor: Reduce the noise threshold SNR>10dB to ensure the accuracy of point cloud data.

[0117] Vibration sensor: Analyzes high-frequency vibrations (1-5kHz) to identify areas of dust accumulation.

[0118] Model optimization, residue recognition model: Increase the weight of visual features to improve YOLOv8's detection accuracy for dust boundaries.

[0119] Cleaning strategy optimization model: adjust the optimization target, reduce the weight of cleaning time, and increase the priority of cleaning rate.

[0120] During cleaning, the cleaning head module mainly uses a rotating brush speed of 500rpm and a high-pressure air nozzle pressure of 6bar to assist cleaning.

[0121] The robot arm path P was optimized to a grid trajectory to cover the dust distribution area, and the pressure was adjusted to 30N.

[0122] The fan turns on to blow away the dust and maintain visibility in the cleaning area.

[0123] Effect evaluation, tested on a cement plant conveyor belt, dust concentration > 500mg / m 3 :

[0124] The cleaning rate CR reaches 97%, which is better than traditional cleaners.

[0125] Detection accuracy increased by 15% and the false detection rate dropped to 5%.

[0126] The system runs stably without any downtime due to dust interference.

[0127] This embodiment significantly improves the cleaning performance in high-dust environments by optimizing the sensor configuration and cleaning strategy, verifying the robustness and environmental adaptability of the cleaner.

[0128] While the present invention has been disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Persons skilled in the art will readily appreciate that various modifications and variations can be made without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A belt conveyor cleaner based on visual detection, characterized by: include: A multimodal data acquisition unit, which collects belt surface images in real time, measures the three-dimensional morphology of residues using ultrasonic data to generate point cloud data, and simultaneously records belt running vibration signals to reflect residue distribution; a multimodal data preprocessing unit, which performs Gaussian filtering to remove noise and histogram equalization to enhance contrast on the belt surface image to generate preprocessed image data, performs time-domain filtering to remove noise on the ultrasonic data to generate preprocessed three-dimensional topography data, and simultaneously performs Fourier transform on the vibration signal to extract frequency domain features to generate preprocessed vibration spectrum data; A model unit, comprising a multimodal residue identification model and a cleaning strategy optimization model, wherein the multimodal residue identification model is used to generate a residue distribution, the cleaning strategy optimization model is used to generate an optimal path and parameters, and the model unit performs intelligent residue classification and cleaning priority assessment; Dynamic cleaning execution unit, the dynamic cleaning execution unit includes: The cleaning head module includes a scraper, a rotating brush, a high-pressure air nozzle, an ultrasonic cleaner, a robotic arm, a torque sensor, an adaptive liquid spray device, a pressure sensor, a fan, a heating element, and a PLC controller. The PLC controller drives the robotic arm to execute the path based on the data of the model unit. The cleaning head module selects tools based on the data of the model unit. The torque and pressure sensors provide real-time feedback to dynamically adjust the angle and pressure of the cleaning head module. The fan and heating element adjust the temperature of the cleaning position. A cleaning effect feedback unit is used to evaluate the cleaning effect. If the cleaning effect does not meet the standard, the reinforcement learning controller adjusts the cleaning parameters according to the evaluation results until the preset standard is met; The model unit includes a multimodal data fusion unit, a reinforcement learning controller and a cloud-based digital twin platform. The multimodal data fusion unit uses the NVIDIA Orin embedded AI chip with a computing power of 200TOPS and a memory of 32GB, and runs the Transformer model. The reinforcement learning controller uses the NVIDIA Jetson AGX Orin with a computing power of 275TOPS and a memory of 64GB, and runs the PPO algorithm. The cloud-based digital twin platform uses an Intel Xeon server with 128GB of memory and 10TB of storage to support model training and parameter optimization. The multimodal residue recognition model is as follows: ; S(x,y): residue distribution score, ϕ(I p (x, y)): visual features, ψH p (x, y, z): ultrasonic three-dimensional characteristics, γV p (f): vibration spectrum characteristics, w1, w2, w3: adaptive weights, w1+w2+w3=1; NVIDIA Orin embedded AI chip runs YOLOv8 to extract ϕ(I p (x, y)), analyze the residue type and boundary, and process the ultrasonic point cloud data to calculate ψH p (x, y, z), determine the thickness, and analyze the vibration spectrum to generate γV p (f) indirectly reflects the distribution. The Transformer model integrates features, optimizes w1, w2, w3, calculates S(x, y), and trains the Transformer model on the cloud server to optimize the loss function: ; L1: loss function of the multimodal recognition model, used to optimize S(x, y), including mean square error and KL divergence, N: number of training samples, S i : The residual score predicted by the model, i represents the i-th sample, : True residue score, : KL divergence, measuring the predicted distribution P(S i ) and the true distribution P( ), used for regularization, λ: regularization coefficient, controlling the weight of KL divergence; The cleaning strategy optimization model is as follows: ;Opt(P, ): Cleaning strategy optimization goal, representing the optimal cleaning path P and parameters , calculated by NVIDIA Jetson AGX Orin through the PPO algorithm, P: cleaning path, : Cleaning parameters, including cleaning head type, angle, pressure, T (P): cleaning time, E ( ): Energy consumption, W (P, ): belt wear, α, β, γ: dynamic weights; Reward function: ,R: reward function, which measures the comprehensive performance of the cleaning strategy. The larger the value, the better the strategy. ,CR: cleaning rate, which indicates the proportion of residues removed. ,η1,η2,η3,η4,: reward function coefficients, which correspond to the weights of cleaning rate, time, energy consumption and wear respectively; The cleaning strategy optimization model optimizes the cleaning path and parameters, balances cleaning efficiency, time, energy consumption and wear, and the reward function R drives the PPO algorithm to iteratively improve the strategy.

2. A belt conveyor cleaner based on visual detection according to claim 1, characterized in that: The multimodal data acquisition unit includes: A high-definition industrial camera array, which collects a belt surface image I (x, y) in real time; An ultrasonic sensor array, wherein the ultrasonic sensor array measures the three-dimensional morphology of the residue to generate point cloud data H (x, y, z); Vibration sensor: The vibration sensor records the belt running vibration signal V(t), reflecting the distribution of residues; The LED light source system is installed on the belt support frame and automatically adjusts the brightness according to the ambient light intensity to ensure clear images.

3. A belt conveyor cleaner based on visual detection according to claim 2, characterized in that: The multimodal data preprocessing unit is an edge computing node, using an ARM Cortex-A78 processor, 16GB of memory, and a computing power of 10 TFLOPS. Each module is equipped with one and supports 5G / Wi-Fi 6 communication. The edge computing node performs Gaussian filtering to remove noise and histogram equalization to enhance contrast on the image I (x, y) to generate preprocessed image data I p (x, y), perform time domain filtering on the ultrasonic data H (x, y, z), eliminate noise, and generate pre-processed three-dimensional shape data H p (x, y, z), perform Fourier transform on the vibration signal V(t), extract frequency domain features, and generate pre-processed vibration spectrum data V p (f).

4. The belt conveyor cleaner based on visual detection according to claim 1, characterized in that: The cleaning effect feedback unit also includes a high-definition industrial camera array, an ultrasonic sensor array, a vibration sensor and an LED light source system. The perception module of the cleaning effect feedback unit recollects the belt surface image I (x, y), the three-dimensional shape to generate point cloud data H (x, y, z) and the vibration signal V (t), updates S (x, y), and calculates the cleaning rate CR. If CR < 98%, the PPO algorithm adjusts P, , the cloud server updates L1 and R and optimizes the model parameters.

5. The belt conveyor cleaner based on visual detection according to claim 2, characterized in that: The high-definition industrial camera array includes three 1080p cameras with wide-angle and zoom lenses, supports infrared imaging, and a frame rate of ≥60fps. The ultrasonic sensor array includes 12 probes with a measurement range of 0-50cm and a resolution of ≤1mm. The vibration sensor includes four high-frequency accelerometers with a sampling rate of ≥10kHz and a sensitivity of 0.1m / s. 2 ; The camera array is installed on steel brackets above and on both sides of the belt conveyor, 50-80 cm from the belt surface. The main camera is located above the center, and the cameras on both sides are tilted 45 degrees. The ultrasonic sensor array is distributed on the support beams on both sides of the belt, 10-20 cm away from the belt; The vibration sensor is fixed to the key nodes of the supporting roller or frame; The camera, ultrasonic sensor, and vibration sensor are connected to the NVIDIA Orin chip via shielded cables. The LED light source is connected to the NVIDIA Orin chip via a PWM controller. The NVIDIA Orin chip is connected to the edge computing node via Ethernet.

6. The belt conveyor cleaner based on visual detection according to claim 1, characterized in that: The robotic arm has a maximum load of 50kg and a positioning accuracy of ±0.1mm. It is equipped with a torque sensor with a resolution of 0.01Nm, an adaptive spray device with a nozzle flow of 0.1-1L / min, and can spray environmentally friendly cleaning agents or anti-stick coatings. It also has a pressure sensor with a range of 0-100N and a resolution of 0.1N.

7. The belt conveyor cleaner based on visual detection according to claim 1, characterized in that: The frequency of the ultrasonic cleaner is 40kHz and is used to process wet and sticky materials. The cleaning head is connected to the slide rail through a servo motor and supports independent lifting and switching.

Images