A multifunctional embedded intelligent cleaning control system

Abstract

The application relates to the technical field of intelligent control systems, and discloses a multifunctional embedded intelligent cleaning control system, which comprises a control execution layer, an Internet of Things transmission layer and a decision analysis layer. The control execution layer comprises intelligent variable-frequency scraping and air-water cleaning devices and executes cleaning and dust removal operations; the Internet of Things transmission layer realizes data interconnection and transmission through edge computing nodes and protocol conversion gateways; and the decision analysis layer adopts time series data analysis and a convolutional neural network algorithm, dynamically optimizes a cleaning operation model, constructs a multi-dimensional linkage cleaning control strategy model, and realizes cleaning intensity adjustment, air-water parameter collaborative control and dust removal efficiency dynamic optimization. Through multi-technology fusion and collaborative optimization, the system improves the intelligent, accurate and efficient levels of cleaning operations, is suitable for a conveyor belt cleaning scene, can improve cleaning efficiency, reduce equipment loss and dust emission, and has remarkable economic and environmental benefits.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control systems, in particular to a multifunctional embedded intelligent cleaning control system. BACKGROUND

[0002] In the field of industrial production, the surface cleaning and dust removal efficiency and effect of the conveyor belt as the key equipment for material transportation directly affect the continuity, safety and environmental protection of production. The traditional conveyor belt cleaning control system generally has low automation, poor adaptability, single control strategy and other problems, which is difficult to meet the precise cleaning demand under complex working conditions.

[0003] From the execution level of cleaning operation, the traditional device mostly adopts fixed parameter mechanical scraping and single dust removal method. For example, the traditional scraping device usually relies on mechanical structure with fixed pressure for operation, which cannot dynamically adjust the scraping force according to the change of the material adhesion strength on the surface of the conveyor belt. This not only may cause incomplete material cleaning, but also may cause excessive wear of the blade or even damage to the conveyor belt due to excessive pressure. In terms of dust removal, the traditional water and air cleaning and dust removal equipment often adopts fixed water and air mixing ratio and dust removal parameters, which is difficult to adapt to the changes of complex working conditions such as different dust particle size, environmental wind speed and conveyor belt surface humidity, resulting in low dust removal efficiency, easy dust emission exceeding standard and not meeting environmental protection requirements.

[0004] In the aspect of data transmission and control, the data interconnection ability of the traditional system is insufficient. Different communication protocols are often used between different devices, which leads to ineffective data intercommunication and forms an "information island". The transmission of control commands lacks dynamic routing scheduling and priority management, making it difficult to realize the collaborative operation of multiple devices. At the same time, the traditional system lacks the use of device operation data, and cannot optimize the cleaning operation model through historical data and real-time working condition data, so that the control strategy cannot be dynamically adjusted with the change of working conditions, and intelligent cleaning control cannot be realized.

[0005] From the aspect of model construction and optimization, the traditional cleaning control system lacks the fusion analysis and dynamic modeling ability of multi-dimensional data. It is difficult to establish an effective correlation between the physical state of the device and the control logic space, and it is difficult to convert the actual cleaning operation data into a high-precision digital control space. In terms of model training and optimization, the traditional method usually uses simple mathematical models or empirical formulas, which cannot use advanced machine learning algorithms to accurately adjust and optimize cleaning intensity, air and water parameters and dust removal efficiency, resulting in low intelligent level of cleaning operation.

[0006] In addition, with the development of industrial intelligence, higher requirements are put forward for the intelligentization, automation and integration of the cleaning control system. The traditional system is difficult to meet the state monitoring, abnormal warning and remote control requirements of large-scale equipment, and cannot realize the whole process monitoring and management of the cleaning operation. Therefore, it is urgent to develop a multifunctional embedded intelligent cleaning control system that can adapt to complex working conditions, realize precise control and has intelligent optimization capability, so as to improve the efficiency and quality of the cleaning operation of the conveying belt, reduce the equipment loss and operation cost, and meet the intelligentization and environmental protection requirements of industrial production. SUMMARY

[0007] The present application aims to provide a multifunctional embedded intelligent cleaning control system to solve the problems in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solution: a multifunctional embedded intelligent cleaning control system, the system comprising:

[0009] a control execution layer, an Internet of Things transmission layer and a decision analysis layer;

[0010] The control execution layer comprises an intelligent variable frequency scraping device, a wind-water cleaning device, a return scraping device and a micro-power dust collector, which is used to execute the conveying belt surface scraping, wind-water combined dust removal and return belt surface residual material removal operations according to the control instructions;

[0011] The Internet of Things transmission layer comprises an edge computing node, a protocol conversion gateway and a cloud server cluster, and deploys a multi-protocol adaptation engine and a load balancing mechanism, which is used to establish a data interconnection channel between devices, realize the synchronous transmission of the cleaning device state monitoring data and the operation instructions, adopt a dynamic routing scheduling strategy for priority sorting and parallel distribution of task instructions according to the data flow and device response speed, and realize the bidirectional interaction between the physical device layer and the control strategy layer through multiple source heterogeneous communication protocols;

[0012] The decision analysis layer is used to dynamically optimize the cleaning operation model by adopting time series data analysis technology combined with the device operation history library and real-time working condition data set, construct a multi-dimensional linkage cleaning control strategy model, and adjust the cleaning intensity, control the wind-water parameters and dynamically optimize the dust removal efficiency based on the strategy model by adopting the convolution neural network algorithm.

[0013] Preferably, the intelligent variable frequency scraping device at least includes a hydraulic drive mechanism, a pressure sensor, an angle adjuster and a blade wear detection module, for collecting the material adhesion strength characteristics of the conveying belt surface; the air-water cleaning device at least includes a high-pressure fan set, an atomizing nozzle array, a flow regulating valve and a water quality filtration unit, for performing air-water mixed cleaning operation under different working conditions; the micro-power dust collector at least includes a negative pressure air induction module, a filter cartridge type dust collection box, a pulse back blowing device and a dust concentration sensor, for completing the directional capture of the dust raised around the conveying belt.

[0014] Preferably, the working condition data collected by the control execution layer in the Internet of Things transmission layer is transmitted by the edge computing node to the protocol conversion gateway through the industrial bus, and then the protocol conversion gateway performs data format standardization, and then the historical operation records stored in the cloud server cluster are jointly transmitted to the decision analysis layer for model training; the model training adopts a multi-dimensional feature fusion technology, and the collected cleaning operation parameters are input into the feature space constructed by different neural network architectures for joint learning;

[0015] The bidirectional interaction between the physical device layer and the control strategy layer through multiple source heterogeneous communication protocols includes: establishing a data through channel between the device layer and the strategy layer by using Modbus-TCP, OPC-UA and MQTT protocols, exchanging device real-time operation parameters, cooperatively processing the parsed operation characteristics, simultaneously realizing dynamic matching of control parameters, completing instruction issuing, state feedback and abnormal warning, and outputting variable frequency speed regulation parameters to the control execution layer.

[0016] Preferably, the cleaning control strategy model of multi-dimensional linkage is constructed, including:

[0017] Establishing a parameter association relationship and a bidirectional calibration mechanism between the device physical state and the control logic space;

[0018] The actual cleaning operation data is extracted and recognized in a manner that the material adhesion strength obtained by the pressure sensor and the dust distribution data monitored by the dust concentration sensor are used as the basis to construct a reference feature library of the conveying belt cleanliness and an abnormal working condition template library, and the model weight is updated and the running trend is predicted according to the real-time operation data, so as to convert the physical operation scene into a high-precision digital control space;

[0019] The cleaning control strategy model is parameter corrected, the real-time collected operation data is input into the established strategy model, the gradient descent algorithm is used to iteratively correct the output results of the model, and the optimized cleaning control strategy model is obtained;

[0020] The cleaning control strategy model includes a physical operation space, a control logic space, a feature database and an interaction protocol between layers.

[0021] The physical work space is a data input source of the strategy model, containing the running parameters and cleaning state characteristics of the conveyor belt; the control logic space forms a mapping relationship with the physical work space, and the cleaning work characteristics are mathematically represented through multi-dimensional parameterized modeling; the feature database integrates historical data and real-time monitoring records of the equipment, and provides a benchmark data set including a working condition mode library, an abnormal event library and an equipment performance library; the interactive protocol realizes data penetration between each level, the real-time collection of feature parameters and model updating are realized between the physical work space and the feature database through a standardized interface, parameter transmission is realized between the physical work space and the control logic space through a data bus, and information synchronization is realized between the control logic space and the feature database through middleware.

[0022] Preferably, the cleaning intensity adjustment based on the strategy model adopts a convolutional neural network algorithm, including:

[0023] Based on the cleaning control strategy model, historical conveyor belt running speed data, material transportation quantity records and blade wear degree characteristics are obtained to construct a working condition sample set;

[0024] After the working condition sample set is normalized, it is divided into a training set and a test set;

[0025] A CNN-LSTM-attention mechanism hybrid model architecture is established, model hyperparameters are initialized, the training set is input into the hybrid model for end-to-end training, spatial features are extracted through a convolutional layer, time sequence dependency relationships are captured through an LSTM network, key feature dimensions are focused through an attention mechanism, and a residual connection is used to optimize the gradient propagation path until the model converges or a preset number of training times is reached;

[0026] The test set is input into the trained hybrid model, the prediction accuracy of the model is evaluated, and the optimal cleaning intensity adjustment model is selected;

[0027] Based on the output dynamic pressure set value of the scraping device of the optimal adjustment model, a frequency converter control instruction is generated in combination with the real-time speed data of the conveyor belt.

[0028] Preferably, the air and water parameter collaborative control based on the strategy model adopts a convolutional neural network algorithm, including:

[0029] Based on the cleaning control strategy model, conveyor belt surface humidity distribution data, dust particle size characteristics and environmental wind speed parameters are extracted to construct a wind and water work feature matrix;

[0030] A feature pyramid network is used to perform cross-scale fusion processing on multi-dimensional features to obtain an enhanced feature map;

[0031] A wind and water control model based on a U-Net architecture is established, and shallow detail features and deep semantic information are fused through a skip connection;

[0032] The fused feature map is input into a regression prediction layer, and a coordinated control amount of the high-pressure fan rotating speed and the atomized water pressure parameter is output.

[0033] Preferably, the dust removal efficiency dynamic optimization based on the strategy model adopts a convolutional neural network algorithm, including:

[0034] Based on the cleaning control strategy model, filter cartridge differential pressure data, pulse backwashing cycle records, and dust accumulation rate parameters are collected to construct a dust removal efficiency feature library.

[0035] The data in the dust removal efficiency feature library is divided by a sliding time window to generate a continuous operation period sample sequence.

[0036] A time series convolution network model is established, the size of the dilated convolution kernel and the range of the receptive field are set, and the time series features are extracted by causal convolution operation.

[0037] The time series features are input into a multi-layer perceptron for efficiency level classification, and the matching result of the dust removal system working state and the optimal backwashing strategy is output.

[0038] Preferably, the construction of the multi-dimensional linkage cleaning control strategy model further includes:

[0039] An overlapping section sampling mechanism is used to slice the continuous operation data, and each data slice is independently encoded.

[0040] An association matrix of data slice features and equipment energy consumption is established to record the energy consumption distribution mode corresponding to different cleaning modes.

[0041] The control model is continuously optimized by an online distillation algorithm, and the model structure adaptive reorganization mechanism is triggered when a new working condition mode is detected.

[0042] Preferably, the generation method of the abnormal working condition processing strategy includes:

[0043] A mapping relationship table of fault types and processing measures is established, including the processing mode of reverse driving corresponding to blade jamming and enhanced backwashing corresponding to filter cartridge blockage.

[0044] A Bayesian optimization algorithm is used to search for the best combination of processing parameters, including hydraulic drive pressure, reverse driving duration, and backwashing air flow intensity.

[0045] The processing effect and equipment loss are comprehensively evaluated by a multi-objective evaluation function, and the parameter reconfiguration process is triggered when the processing effect does not meet the standard.

[0046] Preferably, the training process of the hybrid model architecture includes generating a diversified training sample set by a data enhancement method, and monitoring the model overfitting risk by an early stopping method.

[0047] Compared with the prior art, the present application has the following advantages:

[0048] In terms of cleaning operation, the intelligent variable frequency cleaning device of the control execution layer can collect the material adhesion strength characteristics of the conveyor belt surface in real time through the cooperation of the hydraulic drive mechanism, pressure sensor, angle adjuster and blade wear detection module, and dynamically adjust the cleaning force and angle according to the control instructions, realizing efficient cleaning of the material on the surface of the conveyor belt, while avoiding excessive wear of the blade and damage to the conveyor belt. The air-water cleaning device can flexibly adjust the air-water mixing ratio and cleaning parameters according to different working conditions through the combination of high-pressure fan units, atomizing nozzle arrays, flow regulating valves and water quality filtration units, realizing precise cleaning of the conveyor belt surface and effectively improving the cleaning effect and adaptability. The micro-power dust collector can capture the dust around the conveyor belt through the cooperation of the negative pressure air induction module, filter cartridge dust collection box, pulse blowback device and dust concentration sensor, combined with the dynamically optimized dust removal strategy, significantly improving the dust removal efficiency and reducing dust emission, meeting the environmental protection requirements.

[0049] In terms of data transmission and control, the Internet of Things transmission layer establishes efficient data interconnection channels between devices through the deployment of edge computing nodes, protocol conversion gateways and cloud server clusters, and the application of multi-protocol adaptation engine and load balancing mechanism, realizing the synchronous transmission of cleaning device state monitoring data and operation instructions. Through dynamic routing scheduling strategy, the task instructions are prioritized and distributed in parallel, ensuring the efficiency and reliability of multi-device collaborative operation. The application of multi-source heterogeneous communication protocol realizes the bidirectional interaction between the physical device layer and the control strategy layer, so that the real-time operation parameters of the device can be uploaded to the decision analysis layer for processing, and the control parameters can be accurately issued to the control execution layer, realizing real-time monitoring and precise control of the cleaning operation.

[0050] In terms of model construction and optimization, the decision analysis layer uses time series data analysis techniques combined with equipment operation history library and real-time working condition data set to dynamically optimize the cleaning operation model and construct a multi-dimensional linkage cleaning control strategy model. This model converts actual cleaning operation data into high-precision digital control space by establishing parameter association between equipment physical state and control logic space and a bidirectional calibration mechanism, achieving intelligent modeling of cleaning operation. Based on the convolutional neural network algorithm, cleaning intensity adjustment, air-water parameter collaborative control and dust removal efficiency dynamic optimization can automatically adjust cleaning parameters according to different working condition characteristics, improving the precision and efficiency of cleaning operation. For example, the CNN-LSTM-attention mechanism hybrid model can effectively utilize historical data and real-time working condition data to achieve precise setting of the dynamic pressure of the scraping device; the U-Net architecture-based air-water control model can achieve optimal matching of high-pressure fan speed and atomized water pressure parameters; the time series convolution network model can dynamically optimize the dust removal efficiency and achieve precise matching of the dust removal system working state and the optimal backflushing strategy.

[0051] In terms of system intelligence and adaptability, the application of overlapping segment sampling mechanism, online distillation algorithm and model structure self-adaptive reorganization mechanism enables the cleaning control strategy model to continuously learn and adapt to new working condition modes, achieving continuous optimization and upgrading of the model. The abnormal working condition processing strategy establishes a mapping relationship table between fault types and processing measures and uses Bayesian optimization algorithm to search for the best parameter combination, which can quickly respond to equipment failures and improve the reliability and stability of the system. Meanwhile, the application of data enhancement method and early stopping method improves the quality and efficiency of model training and reduces the risk of model overfitting. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A working principle diagram of the multifunctional embedded intelligent cleaning control system described in the present application;

[0053] Figure 2 A flowchart for realizing the functions of the devices in the control execution layer;

[0054] Figure 3 A flowchart for constructing the multi-dimensional linkage cleaning control strategy model;

[0055] Figure 4 A flowchart for cleaning intensity adjustment based on convolutional neural network. DETAILED DESCRIPTION

[0056] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0057] Please refer to Figures 1-4 The present application relates to a multifunctional embedded intelligent cleaning control system, which includes a control execution layer, an Internet of Things transmission layer and a decision analysis layer, and each layer works cooperatively to realize efficient and intelligent cleaning control. The specific implementation steps are as follows:

[0058] The control execution layer includes an intelligent variable-frequency scraping device, a wind-water cleaning device, a return scraping device and a micro-power dust collector. The intelligent variable-frequency scraping device performs the scraping and cleaning operation on the surface of the conveyor belt according to the control instruction, the wind-water cleaning device performs the wind-water combined dust removal operation, the return scraping device removes the residual material on the return belt surface, and the micro-power dust collector completes the directional capture of the dust around the conveyor belt.

[0059] The Internet of Things transmission layer includes an edge computing node, a protocol conversion gateway and a cloud server cluster, and deploys a multi-protocol adaptation engine and a load balancing mechanism. Its function is to establish a data interconnection channel between devices, realize the synchronous transmission of cleaning device state monitoring data and operation instructions. Specifically, the working condition data collected by the control execution layer is transmitted to the protocol conversion gateway by the edge computing node through the industrial bus, the protocol conversion gateway standardizes the data format, and the historical operation records stored in the cloud server cluster are transmitted to the decision analysis layer for model training. At the same time, the Internet of Things transmission layer adopts a dynamic routing scheduling strategy according to the data flow and device response speed to prioritize and distribute task instructions in parallel, and realizes the bidirectional interaction between the physical device layer and the control strategy layer through multi-source heterogeneous communication protocols, such as Modbus-TCP, OPC-UA and MQTT protocols, to establish a data through channel between the device layer and the strategy layer, realize the functions of instruction issuing, state feedback and abnormal warning, and output variable-frequency speed regulation parameters to the control execution layer.

[0060] The decision analysis layer adopts time series data analysis technology combined with equipment operation history library and real-time working condition data set to dynamically optimize the cleaning operation model, and then constructs a multi-dimensional linkage cleaning control strategy model. The model includes a physical operation space, a control logic space, a feature database, and an interaction protocol between levels. The physical operation space is the data input source of the strategy model, including the running parameters of the conveyor belt and the cleaning state characteristics; the control logic space forms a mapping relationship with the physical operation space, and the cleaning operation characteristics are mathematically represented through multi-dimensional parameterized modeling; the feature database integrates historical data and real-time monitoring records of the equipment, and provides a benchmark data set including working condition mode library, abnormal event library, and equipment efficiency library; the interaction protocol realizes the data penetration between levels. The decision analysis layer adopts a convolutional neural network algorithm based on the strategy model to adjust the cleaning intensity, cooperatively control the air and water parameters, and dynamically optimize the dust removal efficiency.

[0061] The application will be further described below in combination with Examples 1 to 5:

[0062] Example 1:

[0063] The intelligent variable-frequency scraping device, the air-water cleaning device, and the micro-power dust remover are core components of the control execution layer, and through the collaborative design of hardware structure and control logic, the precise execution of the conveyor belt cleaning operation is realized. The implementation of each device will be described in detail in combination with specific application scenarios:

[0064] The hardware architecture of the intelligent variable-frequency scraping device is based on hydraulic drive, and integrates multiple types of sensors and adjusting mechanisms. The hydraulic drive mechanism adopts a closed-loop control system, including a hydraulic pump, a proportional reversing valve, and a hydraulic cylinder, wherein the hydraulic cylinder is hinged with the support arm of the scraping blade. For example, when the conveyor belt conveys viscous materials (such as wet coal slurry), the normal pressure data of the blade and the conveyor belt contact surface are collected in real time by the pressure sensor (model Huba511 series A torque sensor can also be used to indirectly determine the pressure value ), and the data is transmitted to the local controller of the control execution layer through a 4-20mA analog signal. The angle adjuster is an electric push rod structure, and its stroke range corresponds to the angle adjustment interval of the blade and the conveyor belt surface (such as 5°-45°). When the system identifies that the material adhesion strength is high (such as the pressure value exceeds the preset threshold value 20kPa), the local controller sends a command to the angle adjuster to adjust the blade angle from the initial 25° to 15°, so as to increase the scraping depth. The blade wear detection module adopts a laser displacement sensor (such as Keyence LK-G80), which measures the change of the blade edge thickness through non-contact measurement. When the wear amount exceeds 3mm is detected, a warning signal is triggered and the hydraulic drive pressure is adjusted to compensate for the scraping force attenuation caused by wear.

[0065] In actual operation, the control logic of the device is divided into automatic mode and manual mode. For example, in the scenario of cleaning the ore conveying belt, in automatic mode, the system dynamically adjusts the output power of the hydraulic pump (adjustment range 0-15 MPa) according to the real-time data of the pressure sensor. When the conveying belt speed is 2 m / s and the material accumulation thickness reaches 10 mm, the hydraulic drive pressure is automatically increased to 12 MPa to ensure that the blade scrapes off the material at a constant pressure. In manual mode, the operator can directly set the blade angle, hydraulic pressure and other parameters through the local human-machine interface (HMI), which is suitable for debugging or special working conditions.

[0066] The air and water cleaning device adopts modular design. The high-pressure fan set and the atomizing nozzle array are connected through the air-water mixed pipeline. The high-pressure fan set selects a centrifugal fan (such as Model of Tuthill Vacuum & Blower Systems), with a rated air volume of 1500 m 3 / h, and the air pressure can be adjusted in the range of 5-30 kPa. The atomizing nozzle array is arranged equidistantly along the width direction of the conveying belt (spacing 200 mm). The single nozzle uses air-assisted atomizing technology, which can produce mist droplets with an average particle size of 50 μm under water pressure of 2 MPa and air pressure of 5 kPa. The flow regulating valve is an electric three-way valve, which adjusts the ratio of water flow (0-50 L / min) and air flow (0-2000 m 3 / h) through the PID controller. For example, when dealing with fine dust with particle size less than 10 μm, the air-water ratio is set to 40:1 to form a high-concentration mist curtain to adsorb dust. The water quality filtration unit includes a three-stage filtration structure: the primary filter screen (pore size 1 mm) intercepts large particle impurities, the intermediate filter core (precision 50 μm) filters suspended solids, and the reverse osmosis membrane (molecular weight cut-off 300 Da) removes soluble salts to ensure that the nozzle operates without blockage for a long time.

[0067] For example, in the scenario of cement powder conveying, when the dust concentration sensor (such as TSI9306) detects that the dust concentration in the working area exceeds 80 mg / m 3 , the air and water cleaning device is automatically started: the high-pressure fan output air pressure is increased to 25 kPa, the atomizing water pressure is adjusted to 3 MPa, and the water flow is set to 30 L / min to form an air-water mixed flow to flush the surface of the conveying belt, and the mist droplets combine with the dust to form gravity settling to reduce the suspended dust content in the air. The system monitors the operating parameters of each component in real time through the Modbus protocol. If it is found that the flow of a certain branch nozzle is abnormal (deviation from the set value ±10%), the branch is automatically closed and a fault prompt is issued.

[0068] The micro-power dust collector adopts a working mechanism combining negative pressure drainage and pulse cleaning. The negative pressure air induction module is composed of a low-noise fan (power 5 kW, air volume 800 m 3h) and the air duct, the air duct inlet is provided with a guide plate, guiding the dust-containing air flow around the conveying belt to flow to the filter cartridge type dust collector. The filter cartridge type dust collector is internally provided with multiple groups of pleated filter cartridges (filter area 20 m 2 / each, filter precision 1 pm). When the dust-containing air flow passes through the filter cartridges, the dust is intercepted on the filter surface, and the purified air flow is discharged through the fan. The pulse backflushing device includes an air pocket, an electromagnetic pulse valve, and a blowpipe. When the filter cartridge differential pressure sensor (such as Setra267) detects that the differential pressure exceeds 1200 Pa, the pulse backflushing program is triggered: the electromagnetic pulse valve is opened, and compressed air (pressure 0.4-0.6 MPa) is injected into the filter cartridge through the blowpipe to make the filter cartridge instantaneously expand and shake, and the surface dust is stripped. The dust concentration sensor (installed at the dust collector air outlet) monitors the emission concentration in real time. When the concentration exceeds 10 mg / m 3 , the system automatically prolongs the backflushing period or increases the backflushing pressure.

[0069] In the grain processing and conveying scene, for light and fine dust such as flour, the micro-power dust collector keeps the negative pressure induced air module running continuously, and the fan speed is adjusted by the frequency converter (frequency 20-50 Hz) to adapt to the change of dust emission under different conveying speeds. When the conveying capacity of the conveying belt suddenly increases, causing a sharp increase in dust emission, the system automatically increases the fan frequency from 30 Hz to 45 Hz, and shortens the pulse backflushing period (from 5 min to 3 min), ensuring that the filter cartridge always maintains a high-efficiency filtering state. If an abnormal increase in the differential pressure of a filter cartridge is detected (such as exceeding 1500 Pa), the system determines that the filter cartridge may be clogged, immediately starts the standby filter cartridge group, and issues a replacement prompt.

[0070] The above devices are connected to the edge computing nodes of the Internet of Things transmission layer through an industrial bus (such as Profinet), upload running data (such as pressure, flow, speed, etc.) in real time, and receive control instructions. For example, when the intelligent frequency conversion scraping device detects that the blade wear reaches the threshold value, the edge computing node integrates this information with historical wear data and transmits it to the decision analysis layer through the protocol conversion gateway for optimization of the scraping pressure prediction model. At the same time, the local controllers of the devices have independent computing capabilities and can execute pre-set local control strategies when the network is interrupted, ensuring uninterrupted cleaning operations.

[0071] Through the fine design of hardware components and the hierarchical coordination of control logic, the control execution layer realizes adaptive response to different material properties and environmental conditions, providing a reliable physical execution foundation for the efficient operation of the entire intelligent cleaning control system.

[0072] Example 2:

[0073] The IOT transmission layer is the core hub connecting the control execution layer and the decision analysis layer. Through the hierarchical architecture of edge computing nodes, protocol conversion gateways, and cloud server clusters, combined with multi-protocol adaptation engines and dynamic routing mechanisms, it realizes efficient collection, transmission, processing, and instruction distribution of device data. The following describes its implementation from three dimensions: data transmission process, model training mechanism, and protocol interaction logic:

[0074] The data transmission process is based on industrial bus and cloud-edge collaborative architecture design. Control execution layer devices (such as pressure sensors of intelligent variable frequency scraping devices and flow regulating valves of air-water cleaning devices) gather real-time working condition data to edge computing nodes (UNO series embedded controllers from Advantech) through field buses (such as CANopen or ModbusRTU). Taking a mine conveyor belt scene as an example, the edge computing node collects 100 groups of data per minute, including blade pressure (accuracy ±0.1 kPa), fan speed (accuracy ±1 r / min), dust concentration (accuracy ±1 mg / m 3 ), and other parameters, which are transmitted to the protocol conversion gateway through industrial Ethernet (Profinet protocol). The protocol conversion gateway deploys a multi-protocol adaptation engine, supporting real-time conversion of Modbus-TCP, OPC-UA, MQTT, and other protocols. For example, it parses the ModbusRTU data frame of the control execution layer into JSON format, and adds metadata such as timestamps and device IDs to ensure data format standardization. Standardized data is transmitted to the cloud server cluster (using a distributed storage architecture such as Hadoop HDFS) through 4G / 5G or wired networks, along with historical job records (storage period ≥36 months) to form the data source for the decision analysis layer.

[0075] The model training mechanism uses multi-dimensional feature fusion and distributed computing architecture. After receiving standardized data from the IOT transmission layer, the decision analysis layer first removes outliers (such as negative pressure sensor values) through a data cleaning module, and then extracts key features through a feature engineering module. Taking cleaning operation parameters as an example, the extracted features include: conveyor belt speed (m / s), material accumulation thickness (mm), blade angle (°), air-water mixing ratio (air-water volume ratio), filter cartridge pressure difference (Pa), etc. Multi-dimensional feature fusion technology is implemented through parallel neural network architecture, such as inputting pressure data into a one-dimensional convolutional neural network (CNN) to extract spatial features, inputting time series data into a long short-term memory network (LSTM) to capture temporal dependencies, and then implementing joint learning in the feature space through a fully connected layer. The cloud server cluster distributes computing tasks through a load balancing mechanism (such as the round-robin algorithm). When processing training data from 100 devices simultaneously, the task can be split into 10 computing nodes for parallel processing, with each node responsible for feature extraction and model training for 10 devices, significantly improving training efficiency.

[0076] Protocol interaction logic enables bidirectional real-time communication between the physical device layer and the control strategy layer. The IoT transmission layer establishes a real-time data channel between the device layer and the strategy layer through the Modbus-TCP protocol. For example, the hydraulic drive pressure set value of the intelligent frequency conversion scraping device is updated 10 times per second through this protocol, ensuring the real-time nature of control commands. The OPC-UA protocol is used to transmit structured data, such as device account information and maintenance records, supporting cross-platform access and data encryption (AES-256 algorithm) to meet the security requirements of industrial scenarios. The MQTT protocol is used for lightweight data transmission, such as dust concentration overrun warning signals, which achieve second-level response through the publish-subscribe mode. When the dust concentration in a certain area exceeds the threshold, the warning information can be transmitted to the decision analysis layer and the local monitoring terminal within 200ms.

[0077] In the bidirectional interaction process, the dynamic routing scheduling strategy optimizes the transmission path according to data traffic and device response speed. For example, when the cloud server cluster detects that the network delay in a certain area exceeds 500ms, it automatically switches to a backup route (such as a satellite communication link) to transmit critical control commands, ensuring uninterrupted device control. At the same time, the load balancing mechanism dynamically allocates model training tasks based on CPU utilization (threshold ≤ 80%) and memory usage (threshold ≤ 70%) of computing nodes to avoid single-node overload. In the peak data traffic scenario, when 500 devices are simultaneously connected, the load balancer can complete route switching and task allocation within 50ms, ensuring data transmission delay ≤ 100ms.

[0078] The IoT transmission layer also has fault self-healing capability. When the edge computing node detects an industrial bus communication interruption, it automatically switches to a local cache mode, temporarily storing real-time data in the built-in SSD (capacity ≥ 256GB), and uploading them in batches to the protocol conversion gateway after communication is restored, ensuring data integrity. The protocol conversion gateway has a dual-power module (AC220V and DC24V) built-in. When the main power fails, the backup power can seamlessly switch within 10ms, ensuring continuous system operation.

[0079] Example 3:

[0080] The multi-dimensional linkage cleaning control strategy model realizes dynamic modeling and intelligent optimization of cleaning operations by constructing an organic whole of physical operation space, control logic space, feature database, and interaction protocol. The following describes its implementation from the aspects of model construction process, parameter association mechanism, feature database establishment method, and model correction algorithm:

[0081] The model construction process is data-driven. First, the parameter association between the physical state of the equipment and the control logic space is established. The physical operation space includes the running parameters of the conveyor belt (such as speed v and material transport quantity q) and the cleaning state characteristics (such as material adhesion strength F and dust concentration C), which are collected in real time by sensors in the control execution layer. The control logic space realizes the mapping of physical parameters through mathematical modeling, for example, defining the correlation function of the scraping pressure P and the material adhesion strength F and the conveyor belt speed v as P = f(F, v), where f is a nonlinear mapping relationship trained through historical data. The two-way calibration mechanism is realized by comparing the actual measured value and the model predicted value, for example, when the actual scraping pressure and the model output value deviate by more than 5%, the parameter calibration process is triggered, and the coefficients of function f are adjusted until the deviation converges.

[0082] The feature library establishment method is based on multi-source data fusion technology. Based on the material adhesion strength F obtained by the pressure sensor and the dust distribution data C(x, y) (where x and y are the conveyor belt plane coordinates) monitored by the dust concentration sensor, the feature vector is extracted by the sliding window algorithm (window size is 5 minutes of operation data). The reference feature library includes parameters such as mean value μ F , standard deviation σ F , dust concentration spatial distribution matrix M C , etc. For example, the mean value of F under normal working conditions is 15 kPa, the standard deviation is 2 kPa, the dust concentration in the center area of the conveyor belt C(0, 0) is 30 mg / m 3 , and the edge area C(±L, 0) is 10 mg / m 3 (L is the half width of the conveyor belt). The abnormal condition template library stores typical fault features, such as F suddenly increasing to 30 kPa and lasting more than 10 seconds when the blade is stuck, and the dust concentration C suddenly rising and accompanied by an abnormal increase in the current of the negative pressure induced air module when the filter cartridge is blocked. Real-time operation data is reduced in dimension by a feature extraction algorithm (such as principal component analysis PCA) and matched with the reference feature library for cosine similarity, and when the similarity is lower than a threshold (such as 0.8), it is determined as an abnormal working condition and a warning is triggered.

[0083] The model parameter correction process uses the gradient descent algorithm to iteratively optimize the model output. Let the model predicted scraping pressure be and the actual measured value be P real , the loss function is defined as:

[0084]

[0085] where N is the number of samples and i is the sample index. The gradient of the loss function with respect to the model parameters θ is calculated by the back propagation algorithm and the update formula is Adjust the parameters, where a is the learning rate (range 0.001-0.1). Taking a coal conveyor belt in a certain port as an example, the initial model prediction pressure deviation from the actual value is 8 kPa, and after 50 iterations of training, the deviation is reduced to 1.5 kPa, meeting the control accuracy requirements.

[0086] The interactive protocol enables data flow between different levels. The physical work space and the feature database are connected through a standardized interface (such as RESTAPI). Real-time feature parameters F(t) and C(t) are transmitted to the database in JSON format every second, and the delay for historical data query is ≤200 ms. The physical work space and the control logic space transmit parameters through a data bus (such as EtherCAT), and the control instruction P cmd is issued every 100 ms to ensure real-time response. The control logic space and the feature database synchronize information through middleware (such as Apache Kafka). The updated parameters after model training are pushed to the control logic space through the message queue, and the update delay is ≤500 ms.

[0087] Under complex working conditions, the model adapts to environmental changes through real-time weight updating mechanism. For example, when the conveyor belt transports materials from ore to coal powder, the dust particle size distribution characteristics change significantly. The model automatically adjusts the weight of the wind and water parameter collaborative control branch, increases the atomizing water pressure from 2 MPa to 3 MPa, and increases the high-pressure fan air pressure by 10% to meet the adsorption needs of fine dust. The running trend prediction module is based on the time difference model (TD model) and predicts the material adhesion strength change in the next 10 minutes based on the F and v data in the last 30 minutes. The blade angle and hydraulic pressure are adjusted in advance to avoid insufficient cleaning or excessive wear.

[0088] Through the above implementation, the cleaning control strategy model converts the physical work scene into a calculable digital space, realizing a closed-loop control from data acquisition, feature modeling, parameter optimization to instruction execution. The multi-dimensional linkage characteristics of the model are reflected in: physical data driving logical decision, logical output feedback to physical equipment, historical data optimizing model parameters, forming a self-learning, self-adjusting intelligent control system, suitable for coal mine, power plant, building material and other multi-dust, high-load industrial cleaning scenes.

[0089] Example 4:

[0090] The application of the convolutional neural network algorithm based on the cleaning control strategy model in the cleaning intensity adjustment realizes the adaptive control of the scraping device through the whole process design of working condition data modeling, hybrid model training and dynamic parameter output. The following describes its implementation in detail in combination with the scene of a sintering ore conveyor belt in a steel plant:

[0091] The working condition sample set is constructed by integrating multi-source historical data. The system extracts the running data of the conveyor belt from the cloud server cluster for nearly 6 months, including the running speed v (range 0.8-2.5 m / s), the material transportation quantity q (collected by a weighing sensor, accuracy ±1%), and the blade wear degree w (measured by a laser displacement sensor to measure the change of the blade thickness, unit mm). Taking each complete cleaning operation cycle (about 30 minutes) as a sample unit, a total of 2000 samples are obtained. For example, a sample record is: v = 1.8 m / s, q = 500 t / h, w = 1.2 mm, and the actual pressure setting value of the scraping device is 10 MPa. All sample data are mapped to the [0, 1] interval through normalization processing to eliminate dimensional differences, such as dividing the pressure value by the maximum set pressure 15 MPa for normalization.

[0092] The mixed model architecture is built and trained using a three-layer structure of CNN-LSTM-attention mechanism. The convolutional layer includes 2 one-dimensional convolution kernels (size 5) for extracting spatial features such as the association mode of material transportation quantity and blade wear; the LSTM layer is set to 64 memory cells to capture the time sequence dependence relationship, such as the lagging effect of conveyor belt speed change on material adhesion strength (lag time about 5 minutes); the attention mechanism module weights the hidden state output by LSTM to focus on key feature dimensions (such as blade wear). The end-to-end training is performed with a batch size of 32 and a training period of 20 times. During the training process, the gradient vanishing problem is alleviated by residual connection (Residual Connection). For example, in the early stage of training, the model has a large deviation in pressure prediction for high wear conditions (w>2mm), and through residual connection, the original input is directly transmitted to the subsequent layer, so that the prediction error of the model is significantly reduced after the 10th period.

[0093] The model evaluation and optimization divide the sample set into training set and test set according to 7:3. The test set contains 600 samples, and the model performance is evaluated by root mean square error (RMSE) and mean absolute error (MAE). In a certain training, the RMSE of the test set is 0.8 MPa, and the MAE is 0.6 MPa, which meets the industrial control accuracy requirement (allowable error ≤1 MPa). If the evaluation result does not meet the standard, the system automatically adjusts the hyperparameters (such as increasing the LSTM layer to 2 layers, adjusting the learning rate from 0.001 to 0.0005), and re-trains the model. The final selected optimal model can output dynamic pressure setting value in the range of 0-15 MPa according to the input working condition parameters, with an error controlled within ±0.9 MPa.

[0094] The dynamic control instruction generation realizes closed-loop control in combination with the real-time data of the conveying belt. When the system detects that the current conveying belt speed v = 2.2 m / s and the material transportation quantity q = 600 t / h, the data is input into the optimal adjustment model through the edge computing node, and the dynamic pressure set value of the scraping device is output as 12 MPa. At the same time, according to the real-time signal of the conveying belt speed sensor (accuracy ± 0.05 m / s), the local controller generates a variable frequency controller control instruction: if the speed exceeds 2 m / s, the variable frequency controller will increase the motor frequency from 50 Hz to 55 Hz to improve the operating efficiency of the scraping device; if the speed is lower than 1 m / s, the frequency is reduced to 40 Hz to reduce the equipment wear.

[0095] The abnormal working condition processing mechanism integrates the wear warning and parameter compensation logic. The blade wear detection module monitors w in real time, and when w ≥ 2.5 mm, the system automatically increases the pressure set value output by the model by 15% as compensation (for example, the original output of 10 MPa is adjusted to 11.5 MPa), and at the same time triggers a blade replacement reminder. If the pressure after compensation still cannot effectively remove the material (for example, the pressure set value reaches the upper limit of 15 MPa for 3 consecutive periods), the system determines that the blade is invalid, and forces the machine to stop and sends out a fault alarm.

[0096] The engineering application scenario is extended in the grain storage conveying belt scene. For granular materials such as wheat and corn, the model input parameters increase the material humidity h (measured by a capacitive sensor, range 5%-15%). Due to the influence of humidity on the viscosity of the material, when h > 12%, the model automatically reduces the pressure set value by 10%-20% to avoid excessive scraping that causes material breakage. At the same time, according to the seasonal changes (such as high temperature and high humidity in summer), the system automatically adjusts the training set weight to enhance the learning proportion of the humidity feature, so that the model is more suitable for periodic working condition changes.

[0097] Through the above implementation, the convolutional neural network algorithm and the cleaning control strategy model are deeply integrated, realizing the intelligent upgrading from historical data learning to real-time control. When the scheme is deployed in the industrial field, the cleaning intensity can be dynamically adjusted according to different material characteristics (such as particle size, humidity, and viscosity) and equipment states (such as blade wear degree), which not only avoids the problems of conveying belt deviation and material residue caused by insufficient cleaning, but also prevents equipment wear and energy waste caused by excessive cleaning, providing an efficient solution for intelligent cleaning under complex working conditions.

[0098] Example 5:

[0099] The convolutional neural network algorithm based on the cleaning control strategy model is applied in the wind and water parameter collaborative control and dust removal efficiency dynamic optimization, realizing the multi-parameter linkage control of cleaning operation through the whole process collaboration of feature matrix construction, model architecture design and strategy matching. The following describes the implementation in combination with the coal ash conveying belt scene in a thermal power plant:

[0100] The implementation process of the wind-water parameter collaborative control starts with multi-dimensional feature collection. The system obtains the humidity distribution data H(x, y) on the surface of the conveyor belt (where x and y are the coordinates of the conveyor belt plane) through a humidity sensor (accuracy ± 2% RH), measures the dust particle size distribution D(d) (d is the particle size, unit: μm) online through a laser particle size analyzer (such as Malvern Mastersizer 3000), and collects the environmental wind speed u through a wind speed meter (accuracy ± 0.1 m / s). Taking fly ash conveying as an example, when it is detected that the proportion of dust with a particle size of d < 10 μm is 60%, the environmental wind speed u = 3 m / s, and the surface humidity H of the conveyor belt = 10% RH, a wind-water operation feature matrix containing 12 feature dimensions is constructed, such as humidity average, particle size median, wind speed vector, etc.

[0101] The feature pyramid network (FPN) performs cross-scale fusion on multi-dimensional features. Low-level features retain detailed information of humidity distribution (such as lower humidity on the edge of the conveyor belt), and high-level features extract semantic information of dust particle size distribution (such as high proportion of fine dust). Through upsampling and horizontal connection operations, enhanced feature maps containing different scale information are generated. For example, after fusing the particle size feature map (resolution 16 × 16) and the humidity feature map (resolution 64 × 64), a feature map representing both fine dust distribution and local dry areas is obtained, providing more comprehensive input for the subsequent control model.

[0102] The wind-water control model based on the U-Net architecture fuses shallow and deep features through a skip connection. The encoder part consists of 4 convolutional blocks (each block contains 2 3 × 3 convolutional layers and 1 max pooling layer), which gradually extract the associated features of dust, humidity, and wind speed; the decoder part restores the spatial resolution through deconvolution and skip connection, and finally outputs the collaborative control quantities of high-pressure fan speed n (unit: r / min) and atomized water pressure p (unit: MPa). In the fly ash scenario, the model outputs n = 2800 r / min and p = 4 MPa, forming a wind-water mixed flow with a wind speed of 12 m / s and a droplet size of 30 μm, which effectively adsorbs fine dust and suppresses secondary dust raising.

[0103] The implementation process of dynamic optimization of dust removal efficiency takes filter cartridge pressure difference monitoring as the core. The system collects filter cartridge pressure difference data ΔP (accuracy ± 5 Pa), pulse backflushing period T (unit: min), and dust accumulation rate r (calculated by weighing sensor to measure the weight change of dust collection tank, unit: kg / h) in real time. When the increase in fly ash conveying volume leads to r = 15 kg / h and ΔP = 1100 Pa, a dust removal efficiency feature library containing data from the past 2 hours is constructed, and 12 continuous operation period sample sequences are generated through a sliding time window (window length 10 minutes), each sequence containing time series features such as pressure difference change curve and backflushing period record.

[0104] The TCN model extracts one-way time series dependence through causal convolution. By setting the size of the dilated convolution kernel to 3 and the receptive field range to 20 minutes, the first layer of causal convolution captures the pressure fluctuation trend within 10 minutes, and the second layer captures the influence of the blowback cycle within 20 minutes. For example, when detecting that ΔP has risen at a rate of 50 Pa / min within the past 15 minutes and no blowback has been triggered, the model determines that the risk of filter clogging is rising, outputs the dust removal system operating state as "inefficient operation" through a multi-layer perceptron (MLP), and matches the best blowback strategy: increasing the blowback pressure from 0.4 MPa to 0.5 MPa and extending the blowback duration from 5 s to 8 s.

[0105] The model cooperative optimization mechanism is realized through overlapping segmented sampling and online distillation. Continuous operation data is sampled at 5-minute intervals with an overlap rate of 50%, and each data slice is independently encoded to avoid interference from abnormal data at a single time point. For example, a slice contains interference data with a sudden change in conveyor belt speed, which can be reduced by averaging the features of adjacent slices. The correlation matrix of data slice features and equipment energy consumption (such as fan power and water pump power) records the energy consumption distribution of different cleaning modes, such as the "strong wind and strong water" mode, which is 25% higher than the "standard mode", providing a basis for the system to select energy-saving strategies.

[0106] When a new working condition is detected (such as a sudden increase in fly ash humidity leading to dust clumping), the online distillation algorithm triggers model structure adaptive reorganization. For example, the original model lacks learning of the pressure difference mutation characteristics of clumped dust, and through the knowledge of the distillation teacher model (a pre-trained wide neural network), the student model (a lightweight TCN) quickly adjusts the convolution kernel parameters to enhance the capture ability of step-type features. The abnormal working condition processing strategy is realized through the combination of a fault mapping table and Bayesian optimization: if the filter pressure difference exceeds 1500 Pa and does not decrease after blowback, it is determined that the filter is clogged, and the system automatically switches to the standby filter group and issues a replacement prompt; the Bayesian optimization algorithm searches for the optimal combination in the parameter space of hydraulic drive pressure (8-15 MPa) and reverse drive duration (10-30 s), for example, for a blade jamming fault, after 5 iterations, the optimal parameters are determined as pressure 12 MPa and duration 20 s, which balance the cleaning effect and equipment load.

[0107] Cross-scenario application extension In the cement raw material conveying scenario, the dust particle size distribution is mainly 50-100 μm, and the environmental wind speed often reaches 5 m / s. The wind-water control model outputs n = 3200 r / min and p = 5 MPa according to the characteristic matrix (humidity 8% RH, particle size median 70 μm, wind speed 5 m / s), forming a mixed flow of high-speed airflow and large particle size droplets, effectively impacting the residual blocky material. The dust removal system identifies the sinusoidal fluctuation characteristics (period about 15 minutes) of the dust accumulation rate under high wind speed through the time sequence convolution network, automatically adjusts the back blowing period to 12 minutes, and avoids the filter cartridge overload caused by the periodic peak of dust raising.

[0108] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A multifunctional embedded intelligent cleaning control system, characterized in that, It includes: Control execution layer, IoT transmission layer, and decision analysis layer; The control execution layer includes an intelligent frequency conversion scraper, a wind and water cleaning device, a return scraper, and a micro-power dust collector, which are used to perform scraping and cleaning of the conveyor belt surface, combined wind and water dust removal, and removal of residual materials on the return belt surface according to control commands. The IoT transmission layer includes edge computing nodes, protocol conversion gateways, and cloud server clusters, and deploys a multi-protocol adaptation engine and load balancing mechanism to establish data interconnection channels between devices, realize the synchronous transmission of cleaning device status monitoring data and operation instructions, adopt a dynamic routing scheduling strategy to prioritize and distribute task instructions in parallel according to data traffic and device response speed, and realize bidirectional interaction between the physical device layer and the control strategy layer through multi-source heterogeneous communication protocols. The decision analysis layer is used to dynamically optimize the cleaning operation model by combining time series data analysis technology with the equipment operation history database and real-time working condition dataset, and then construct a multi-dimensional linkage cleaning control strategy model. Based on this strategy model, a convolutional neural network algorithm is used to adjust the cleaning intensity, coordinate the control of air and water parameters, and dynamically optimize the dust removal efficiency. The method of adjusting cleaning intensity using a convolutional neural network algorithm based on this strategy model includes: Based on the cleaning control strategy model, historical conveyor belt operating speed data, material transport volume records, and blade wear characteristics are obtained to construct a set of working condition samples. After normalizing the set of working condition samples, it is divided into a training set and a test set; Establish a hybrid model architecture of CNN-LSTM-attention mechanism, initialize the model hyperparameters, input the training set into the hybrid model for end-to-end training, extract spatial features through convolutional layers, capture temporal dependencies through LSTM network, focus on key feature dimensions through attention mechanism, and optimize gradient propagation path using residual connection until the model converges or reaches the preset number of training times. Input the test set into the trained hybrid model, evaluate the model's prediction accuracy, and select the optimal cleaning intensity adjustment model; Based on the optimal adjustment model, the dynamic pressure setpoint of the scraper is output, and the frequency converter control command is generated by combining the real-time speed data of the conveyor belt. The method of using a convolutional neural network algorithm for coordinated control of feng shui parameters based on this strategy model includes: Based on the cleaning control strategy model, data on surface humidity distribution, dust particle size characteristics, and environmental wind speed parameters of the conveyor belt are extracted to construct a wind and water operation feature matrix. A feature pyramid network is used to perform cross-scale fusion processing on multi-dimensional features to obtain an enhanced feature map; A feng shui control model based on the U-Net architecture is established, which integrates shallow detailed features and deep semantic information through skip connections; The fused feature map is input into the regression prediction layer, and the output is the coordinated control quantity of the high-pressure fan speed and atomized water pressure parameters; The method of dynamically optimizing dust removal efficiency using a convolutional neural network algorithm based on this strategy model includes: Based on the cleaning control strategy model, filter cartridge differential pressure data, pulse backflushing cycle records, and dust accumulation rate parameters are collected to construct a dust removal efficiency feature library. The data in the dust removal efficiency feature library is segmented by a sliding time window to generate a sample sequence of continuous operation periods; A temporal convolutional network model is established, the size of the dilated convolutional kernel and the receptive field are set, and temporal features are extracted through causal convolution operations. The time-series features are input into a multilayer perceptron for performance level classification, and the matching results of the dust removal system's working status and the optimal backflushing strategy are output.

2. The multifunctional embedded intelligent cleaning control system according to claim 1, characterized in that, The intelligent variable frequency scraper includes at least a hydraulic drive mechanism, a pressure sensor, an angle adjuster, and a blade wear detection module, used to collect the adhesion strength characteristics of the material on the conveyor belt surface; the air-water cleaning device includes at least a high-pressure blower unit, an atomizing nozzle array, a flow regulating valve, and a water filtration unit, used to perform air-water mixed cleaning operations under different working conditions; the micro-power dust collector includes at least a negative pressure induced draft module, a cartridge dust collection box, a pulse backflushing device, and a dust concentration sensor, used to complete the directional collection of dust around the conveyor belt.

3. The multifunctional embedded intelligent cleaning control system according to claim 1, characterized in that, In the IoT transmission layer, the working condition data collected by the control execution layer is transmitted from the edge computing node to the protocol conversion gateway via the industrial bus. After the protocol conversion gateway standardizes the data format, it is transmitted together with the historical operation records stored in the cloud server cluster to the decision analysis layer for model training. The model training adopts multi-dimensional feature fusion technology, which inputs the collected cleaning operation parameters into the feature space constructed by different neural network architectures for joint learning. The method of achieving bidirectional interaction between the physical device layer and the control strategy layer through multi-source heterogeneous communication protocols includes: establishing a data communication channel between the device layer and the strategy layer using Modbus-TCP, OPC-UA and MQTT protocols; exchanging real-time operating parameters of the devices and coordinating the processing of the parsed operation characteristics; simultaneously achieving dynamic matching of control parameters; completing instruction issuance, status feedback and abnormal warning; and outputting frequency conversion speed regulation parameters to the control execution layer.

4. The multifunctional embedded intelligent cleaning control system according to claim 1, characterized in that, The construction of the multi-dimensional linkage cleaning control strategy model includes: Establish the parameter correlation between the physical state of the equipment and the control logic space, and a two-way calibration mechanism; By extracting features and recognizing patterns from actual cleaning operation data, and based on the material adhesion strength obtained by pressure sensors and the dust distribution data monitored by dust concentration sensors, a benchmark feature library and an abnormal working condition template library for conveyor belt cleanliness are constructed. The model weights are updated and the operating trend is predicted based on real-time operation data, transforming the physical operation scenario into a high-precision digital control space. The cleaning control strategy model is parameter-calibrated by inputting real-time collected operation data into the established strategy model and using the gradient descent algorithm to iteratively correct the output of the model to obtain the optimized cleaning control strategy model. The cleaning control strategy model includes a physical workspace, a control logic space, a feature database, and interaction protocols between different levels. The physical workspace serves as the data input source for the strategy model, encompassing conveyor belt operating parameters and cleaning status characteristics. The control logic space and physical workspace form a mapping relationship, with multi-dimensional parametric modeling used to mathematically represent the cleaning operation characteristics. The feature database integrates historical equipment data and real-time monitoring records, providing a benchmark dataset including a working condition mode library, an abnormal event library, and an equipment performance library. The interaction protocol enables data flow between different levels. The physical workspace and feature database utilize a standardized interface for real-time acquisition of feature parameters and model updates. The physical workspace and control logic space transmit parameters via a data bus, and the control logic space and feature database synchronize information through middleware.

5. A multifunctional embedded intelligent cleaning control system according to claim 1, characterized in that, The construction of the multi-dimensional linkage cleaning control strategy model also includes: An overlapping segmented sampling mechanism is used to slice continuous operation data, and each data slice is independently encoded. Establish a correlation matrix between data slice features and equipment energy consumption, and record the energy consumption distribution patterns corresponding to different cleaning modes; The control model is continuously optimized through an online distillation algorithm, and an adaptive reorganization mechanism for the model structure is triggered when a new operating condition is detected.

6. The multifunctional embedded intelligent cleaning control system according to claim 1, characterized in that, The construction of the multi-dimensional linkage cleaning control strategy model also includes an abnormal working condition handling strategy, the method for generating the abnormal working condition handling strategy includes: Establish a mapping table between fault types and handling measures, including reverse drive for blade jamming and enhanced backflushing for filter cartridge blockage; The optimal combination of processing parameters, including hydraulic drive pressure, reverse drive duration, and backflush airflow intensity, is searched using a Bayesian optimization algorithm. The processing effect and equipment wear are comprehensively measured by a multi-objective evaluation function. When the processing effect fails to meet the standard, the parameter reconfiguration process is triggered.

7. A multifunctional embedded intelligent cleaning control system according to claim 1, characterized in that, The training process of the CNN-LSTM-attention mechanism hybrid model architecture includes: generating diverse training sample sets using data augmentation methods, and monitoring the risk of model overfitting using early stopping.

Images