Intelligent greening crusher unblocking system and method

Abstract

The embodiment of the invention provides an intelligent greening crusher unblocking system and method. The method comprises the following steps: acquiring multi-modal sensing data of the greening crusher through a multi-modal sensor array deployed in the greening crusher; the multi-mode sensing data is used for comprehensively monitoring the state of materials in the crusher; based on the multi-modal sensing data and the internal structure of the greening crusher, establishing a blockage risk model of the greening crusher through a space-time diagram neural network; dynamically adjusting a dynamic control instruction of the greening crusher based on the blockage risk model; and generating an optimal unblocking path by adopting a graph contrast loss function, and executing a differential unblocking scheme matched with the optimal unblocking path based on the dynamic control instruction. According to the method, through the blockage risk model, blockage clearing path planning and graded blockage clearing of the greening crusher can be realized, equipment abnormity caused by congestion is avoided, the crushing efficiency of the equipment is improved, and normal operation of the greening crusher is guaranteed.

Description

Technical Field

[0001] This application relates to the field of data prediction, and particularly to an intelligent green waste crusher clogging removal system and method. Background Art

[0002] A green waste crusher is a device specifically used for crushing materials such as green waste and garden garbage, and plays an important role in urban greening maintenance, garden landscape management and other fields.

[0003] Currently, a green waste crusher usually drives a cutter or a rotor to rotate at a high speed by means of a power source (such as an electric motor). When the green waste is fed into the crushing chamber, a strong shearing force, impact force and extrusion force are generated between the high-speed rotating cutter or rotor and the fixed blade, and the material is crushed into smaller particles or fragments.

[0004] In related technologies, green waste crushers mostly work outdoors, with complex material types, different structural forms and materials, which easily lead to material accumulation inside the green waste crusher, unable to be discharged normally, low crushing efficiency of the equipment, and affecting the normal operation of the green waste crusher. Therefore, there is an urgent need to design a technical solution to overcome at least one technical problem existing in the related technologies. Summary of the Invention

[0005] The main purpose of the embodiments of this application is to provide an intelligent green waste crusher clogging removal system and method, aiming to solve the technical problems of material accumulation and low crushing efficiency of the green waste crusher caused by complex material types.

[0006] In a first aspect, the embodiments of this application provide an intelligent green waste crusher clogging removal method, including:

[0007] Collect multi-modal perception data of the green waste crusher through a multi-modal sensor array deployed in the green waste crusher; the multi-modal perception data is used to comprehensively monitor the material state inside the crusher; the multi-modal sensor array contains sensors of multiple functional types, and the sensors of multiple functional types are respectively deployed in the key internal structures of the green waste crusher;

[0008] Based on the multi-modal perception data and the internal structure of the green waste crusher, establish a clogging risk model of the green waste crusher through a spatio-temporal graph neural network; the clogging risk model is marked with the clogging risk levels and corresponding clogging risk factors of each key internal structure;

[0009] Based on the clogging risk model, dynamically adjust the dynamic control instructions of the green waste crusher; the dynamic control instructions at least include: feeding rate, scheduling instructions for different particle size crushing units, temperature supplement parameters, humidity supplement parameters;

[0010] The figure contrast loss function is used to generate the optimal blockage removal path, and a differential blockage removal scheme matching the optimal blockage removal path is executed based on the dynamic control instruction to preferentially process high-risk nodes in the blockage risk model, so as to achieve hierarchical blockage removal of the greening crusher and ensure the normal operation of the greening crusher.

[0011] In a second aspect, an intelligent greening crusher blockage removal system and method provided by an embodiment of the present application include:

[0012] An acquisition module, deployed in the multi-modal sensor array of the greening crusher, is used to acquire multi-modal perception data of the greening crusher; the multi-modal perception data is used to comprehensively monitor the material state inside the crusher; the multi-modal sensor array contains sensors of multiple functional types, and sensors of multiple functional types are respectively deployed in key internal structures of the greening crusher;

[0013] A building module is used to build a blockage risk model of the greening crusher through a spatio-temporal graph neural network based on the multi-modal perception data and the internal structure of the greening crusher; the blockage risk level and corresponding blockage risk factors of each key internal structure are marked in the blockage risk model;

[0014] An adjustment module is used to dynamically adjust the dynamic control instruction of the greening crusher based on the blockage risk model; the dynamic control instruction at least includes: feed rate, scheduling instructions for different particle size crushing units, temperature supplement parameters, and humidity supplement parameters;

[0015] An execution module is used to generate an optimal blockage removal path by using the figure contrast loss function, and execute a differential blockage removal scheme matching the optimal blockage removal path based on the dynamic control instruction to preferentially process high-risk nodes in the blockage risk model, so as to achieve hierarchical blockage removal of the greening crusher and ensure the normal operation of the greening crusher.

[0016] In a third aspect, an embodiment of the present application further provides a terminal device, which includes a processor and a memory for storing a computer program; the processor is used to execute the computer program and implement the intelligent greening crusher blockage removal method described in the first aspect or any embodiment of the present application when executing the computer program.

[0017] The embodiment of the present application provides an intelligent green crusher clogging removal system and method. In this method, first, a multi-modal sensor array deployed in the green crusher is used to collect multi-modal perception data of the green crusher; the multi-modal perception data is used to comprehensively monitor the material state inside the crusher; the multi-modal sensor array includes sensors of various functional types, and the sensors of various functional types are respectively deployed in the key internal structures of the green crusher. Then, based on the multi-modal perception data and the internal structure of the green crusher, a clogging risk model of the green crusher is established through a spatio-temporal graph neural network; the clogging risk levels and corresponding clogging risk factors of each key internal structure are marked in the clogging risk model. Furthermore, based on the clogging risk model, the dynamic control instructions of the green crusher are dynamically adjusted; the dynamic control instructions at least include: feeding rate, scheduling instructions for different particle size crushing units, temperature supplement parameters, and humidity supplement parameters. Finally, a graph contrast loss function is used to generate an optimal clogging removal path, and a differentiated clogging removal plan matching the optimal clogging removal path is executed based on the dynamic control instructions to preferentially process the high-risk nodes in the clogging risk model, realizing hierarchical clogging removal of the green crusher and ensuring the normal operation of the green crusher. The embodiment of the present application can realize the clogging removal path planning and hierarchical clogging removal of the green crusher through the clogging risk model, adapt to the crushing methods required by different materials, effectively avoid equipment abnormalities caused by congestion, improve the equipment crushing efficiency, and ensure the normal operation of the green crusher. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flow chart of an intelligent green crusher clogging removal method provided by an embodiment of the present application;

[0019] Figure 2 It is a schematic structural diagram of an intelligent green crusher clogging removal system provided by an embodiment of the present application;

[0020] Figure 3 It is a schematic block diagram of the structure of a terminal device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] In the related art, green crushers mostly work outdoors, and it is easy to have the phenomenon that materials accumulate inside the green crusher due to the complex types of materials and cannot be discharged normally, which affects the normal operation of the green crusher. For example, when the water content of some material materials is relatively high, it is easy to cause the surrounding materials to stick together and form larger lumps. These lumps are not easy to be broken inside the crusher and may also adhere to the inner wall of the crushing chamber or the screen, resulting in blockage. For example, this situation is likely to occur when crushing wet ores or soil-containing materials during the rainy season. Or, materials with hard texture or large volume cannot be effectively broken inside the crusher and will accumulate in the crushing chamber, causing blockage. For example, directly feeding large stones exceeding the specified size into a jaw crusher may cause blockage.

[0022] Aiming at the technical problems existing in the related art, the embodiment of the present application proposes an intelligent green crusher blockage clearing system and method.

[0023] Specifically, first, through a multi-modal sensor array deployed in the green crusher, multi-modal perception data of the green crusher is collected; the multi-modal perception data is used to comprehensively monitor the material state inside the crusher; the multi-modal sensor array includes sensors of multiple functional types, and the sensors of multiple functional types are respectively deployed in the key internal structures of the green crusher. Then, based on the multi-modal perception data and the internal structure of the green crusher, a blockage risk model of the green crusher is established through a spatio-temporal graph neural network; the blockage risk levels and corresponding blockage risk factors of each key internal structure are marked in the blockage risk model. Furthermore, based on the blockage risk model, the dynamic control instructions of the green crusher are dynamically adjusted; the dynamic control instructions at least include: feeding rate, scheduling instructions for different particle size crushing units, temperature supplement parameters, and humidity supplement parameters. Finally, a graph contrast loss function is used to generate an optimal blockage clearing path, and a differential blockage clearing scheme matching the optimal blockage clearing path is executed based on the dynamic control instructions to preferentially process the high-risk nodes in the blockage risk model, realizing hierarchical blockage clearing of the green crusher and ensuring the normal operation of the green crusher.

[0024] The embodiment of the present application can realize the blockage clearing path planning and hierarchical blockage clearing of the green crusher through the blockage risk model, adapt to the crushing methods required by different materials, effectively avoid equipment abnormalities caused by congestion, improve the equipment crushing efficiency, and ensure the normal operation of the green crusher.

[0025] An embodiment of the present application provides an intelligent green crusher clogging removal system and method. Among them, the intelligent green crusher clogging removal method can be applied to a terminal device, such as electronic devices like mobile phones, virtual reality devices, tablet computers, laptop computers, desktop computers, wearable devices, etc. The terminal device can be a server connected to a tower crane device, or a server cluster. The above connection method can be implemented through a hardware circuit or through a communication module.

[0026] The following will, with reference to the accompanying drawings, elaborate on some embodiments of the present application. Without conflict, the following embodiments and the features in the embodiments can be combined with each other. Please refer to Figure 1 , Figure 1 which is a schematic flowchart of an intelligent green crusher clogging removal method provided by an embodiment of the present application.

[0027] As Figure 1 shown, the intelligent green crusher clogging removal method includes the following steps S101 to S106.

[0028] Step S101: Collect multi-modal perception data of the green crusher through a multi-modal sensor array deployed in the green crusher.

[0029] In an embodiment of the present application, the multi-modal perception data is used to comprehensively monitor the material state inside the crusher. Further, the multi-modal sensor array includes sensors of multiple functional types, and the sensors of multiple functional types are respectively deployed in key internal structures of the green crusher.

[0030] Exemplarily, in the internal structure of the green crusher, the multi-modal sensor array realizes all-round monitoring of the material state through scientific layout and precise installation. The following are specific deployment forms:

[0031] In the crushing chamber, humidity sensors are arranged as capacitive humidity sensors at key positions on the inner wall of the crushing chamber (such as near the feed inlet and in the middle of the side wall) to detect the moisture content of the material in real time. The humidity sensors adopt protective casings to prevent material impact and dust pollution and ensure measurement accuracy. Pressure sensors are installed at key stress points on the crushing wall to monitor the pressure change of the material on the crushing wall and reflect the hardness and flow rate of the material. The sensors are installed in an embedded manner, directly contacting the inner wall of the crushing chamber to obtain real pressure data. Vibration sensors are nested inside the crushing wall and are used to monitor the vibration frequency and amplitude generated during the crushing process. The vibration sensors adopt high-sensitivity MEMS accelerometers, which can capture minute vibration changes and promptly detect abnormal vibration conditions.

[0032] In the screen area, the infrared thermal imaging module is installed above the screen to scan the surface temperature distribution of the screen and identify local overheating areas caused by material adhesion. The thermal imaging module has high resolution and fast response capabilities, enabling it to capture temperature anomaly points in real time. The laser particle size analyzer is integrated at the end of the feeding conveyor belt to dynamically scan the particle size distribution of the material, ensuring that the particle size of the material entering the crushing chamber meets the requirements. The laser particle size analyzer uses non-contact measurement to avoid interfering with the material flow.

[0033] On the material conveying path, the flow sensor is installed at key positions on the conveyor belt to monitor the real-time flow of the material and ensure the stability of the feeding rate. The flow sensor uses ultrasonic or optoelectronic principles and has high precision and anti-interference capabilities. The material composition sensor deploys a near-infrared spectroscopy (NIRS) sensor in the conveying path to detect the chemical composition and moisture content of the material in real time, providing data support for subsequent processing. The NIRS sensor transmits signals through optical fibers to ensure the stability and accuracy of the data.

[0034] In addition, the motor status sensor is installed on the drive motor of the crusher to monitor the rotational speed, temperature, and vibration of the motor, reflecting the working status of the equipment. The motor status sensor sends data to the central control system through a wireless transmission module. The lubrication system sensor installs flow and pressure sensors at key nodes of the lubrication system to monitor the status of the lubricating oil and ensure the lubrication effect of the equipment. The lubrication system sensor is connected to the main control system through a data acquisition module to achieve real-time monitoring.

[0035] In the environmental parameter monitoring, the environmental temperature and humidity sensor is installed outside the crusher to monitor the temperature and humidity of the surrounding environment, providing an environmental correction basis for the internal sensor data. The environmental sensor uses high-precision digital output and has good anti-interference capabilities. The dust concentration sensor installs a dust concentration sensor at key positions inside the crusher to monitor the dust concentration in the working environment, ensuring the safety of operators and the normal operation of the equipment. The dust sensor uses the laser scattering principle and has fast response and high sensitivity.

[0036] The edge computing gateway is deployed on edge computing devices near the crusher, responsible for real-time acquisition, time synchronization, feature extraction, and preliminary processing of multi-modal sensor data. The edge computing gateway uploads the processed data to the central control system through an industrial Ethernet or wireless communication module.

[0037] The data storage and analysis platform establishes a cloud or local data storage and analysis platform to store sensor data for a long time, conduct trend analysis, and perform fault prediction. The platform uses big data technology and machine learning algorithms to achieve intelligent diagnosis and predictive maintenance of equipment status.

[0038] Through the above deployment form of the multi-modal sensor array, the material state inside the green waste crusher and the operation status of the equipment can be comprehensively and real-time monitored, providing reliable data support for the construction of the blockage risk model and the generation of dynamic control instructions, and ensuring the efficient and stable operation of the equipment.

[0039] As an optional embodiment, in step S101, through the multi-modal sensor array deployed in the green waste crusher, the multi-modal perception data of the green waste crusher is collected, including:

[0040] Through the humidity sensors deployed at the feed inlet and the inner wall of the crushing chamber, the moisture content of the material is detected in real time; through the laser particle size detection module integrated at the end of the feed conveyor belt, the particle size distribution of the material is dynamically scanned; through the pressure sensors and vibration sensors deployed on the crushing wall and the bottom layer of the screen in the crushing chamber, abnormal stress fluctuations are monitored; wherein, the warning threshold of the abnormal stress fluctuation is set to 1.2 times the upper limit of the stress fluctuation under normal working conditions; through the infrared thermal imaging module deployed at the edge of the material screen, the surface temperature distribution of the screen is scanned to identify the local overheating area caused by the adhesion; through edge computing, the time synchronization of the data of different modal sensors collected is carried out, and according to the spatio-temporal heterogeneous characteristics of the sensors, the fusion weights of the data of different modal sensors are adaptively adjusted, and the data of different modal sensors are fused and feature extracted to obtain a plurality of perception feature data, so as to construct the multi-modal perception data.

[0041] In step S101, the system collects multi-dimensional perception data in real time through the multi-modal sensor array deployed inside the green waste crusher to construct a comprehensive understanding of the equipment operation status. The humidity sensors are installed at the feed inlet and the inner wall of the crushing chamber, and the moisture content of the material is detected in real time through the capacitive or impedance sensing principle, and an alarm is triggered when the humidity exceeds the set threshold. The laser particle size detection module is integrated at the end of the feed conveyor belt and uses TOF (Time of Flight) technology to scan the particle size distribution of the material and generate a particle size distribution curve, providing a basis for the subsequent crushing unit scheduling. The pressure sensors and vibration sensors are deployed on the crushing wall and the bottom layer of the screen, and the abnormal stress fluctuations are captured through the piezoelectric effect and the MEMS accelerometer. When the vibration spectrum energy exceeds 1.2 times the upper limit of the normal working condition, it is determined as abnormal. The infrared thermal imaging module covers the edge area of the screen, and the local overheating (ΔT>5°C) caused by material adhesion is identified through non-contact temperature measurement, and the adhesion degree is analyzed in combination with the temperature gradient. The edge computing gateway synchronizes the time of the multi-modal data (IEEE 1588 protocol), and uses the dynamic graph attention network (DGAT) to adaptively adjust the sensor weights. For example, the weight of the humidity sensor is increased to 0.6 in the rainy season, and the weight of the vibration sensor is increased to 0.5 during the equipment aging period. After eliminating the noise through the Kalman filter, features such as humidity-particle size covariance and pressure-vibration mutual information are extracted, and finally fused into a perception data vector containing 20-dimensional features.

[0042] Further optionally, before synchronizing the collected multi-modal sensor data in time in the above steps, different anti-interference modes corresponding to the multi-modal sensor data can also be identified according to different sensor device attributes. Among them, the pressure sensor adopts a notch filtering mode; the vibration sensor adopts an adaptive band-pass filtering mode; the temperature sensor and the humidity sensor adopt a cross-validation mechanism and an environmental compensation mode; the laser particle size sensor adopts a spatio-temporal domain joint calibration mode. Furthermore, the Kalman filtering algorithm is used to eliminate the sensor noise in the multi-modal sensor data according to the anti-interference mode. Finally, historical trend prediction is used to supplement the uncollected multi-modal sensor data.

[0043] In the process of multi-modal sensor data acquisition, designing dedicated anti-interference modes according to different sensor characteristics is a key step to ensure data quality. The pressure sensor is vulnerable to mechanical vibration interference. The notch filtering mode can accurately eliminate the 50Hz power frequency and its harmonic interference. The principle is to dynamically adjust the filter parameters to form a zero point in the frequency domain to cancel the specific frequency noise.

[0044] The vibration sensor needs to deal with broadband impact signals. The adaptive band-pass filtering mode can track the change of the device operation frequency in real time and automatically adjust the passband range to 10 - 200Hz to retain the effective vibration information.

[0045] For the temperature and humidity sensors, the cross-validation mechanism identifies outliers by comparing the differences in adjacent sensor readings, and the environmental compensation mode dynamically calibrates the measured values according to the temperature and humidity changes in the equipment cabin. For example, when a sudden change in the ambient temperature is detected, the thermal drift error of the sensor is automatically corrected.

[0046] The laser particle size sensor is prone to measurement deviation in strong light or dusty environments. The spatio-temporal domain joint calibration mode combines spatial filtering to suppress the scattered light interference and eliminates the instantaneous pulse noise through time series analysis.

[0047] The Kalman filtering algorithm plays a core role in this process. The state equation is established according to the dynamic models of each sensor, and the minimum mean square error estimation is recursively calculated using the optimal estimation value at the previous moment and the current observation value. For example, in the vibration signal processing, the historical acceleration data and the current sampling value are fused to effectively suppress random noise. For example, when a sensor fails briefly, the historical trend prediction module analyzes the long-term operation rules based on the mixture density network (MDN). For example, the current missing value is inferred according to the change trend of the feeding rate in the past 1 hour to ensure data continuity. For example, in a certain mine crusher, the accuracy of the humidity sensor can be improved to ±1.2%, the signal-to-noise ratio of the vibration signal can be increased by 18dB, and 98% of the data availability can still be maintained under sudden dust interference, significantly improving the reliability of equipment condition assessment.

[0048] Furthermore, the outlier points can be filtered in real time based on the 3σ principle to ensure data validity.

[0049] Further optionally, after adaptively adjusting the respective fusion weights of the data of different modal sensors according to the spatio-temporal heterogeneous characteristics of the sensors in the above steps, the cross-modal feature alignment of the semantic information in the data of different modal sensors can also be performed. Furthermore, through the contrast learning loss function, the feature distance of the sensor data under similar clogging patterns is adjusted, and the adjusted sensor data feature distance is added to the multi-modal perception data to enhance the recognition accuracy of the implicit semantic information in the data of different modal sensors.

[0050] For example, a cross-modal attention mechanism based on Transformer is adopted to map temporal features such as the humidity change rate and the peak value of the particle size distribution to the temperature gradient in the thermal imaging and the spectral features of the vibration signal into a shared feature space. For example, when the humidity sensor detects a sudden change in the moisture content, the corresponding infrared thermal imaging area will show an abnormal increase in temperature. The feature contribution degrees of the two are dynamically adjusted through the attention weights, so that the different modal data are consistent at the semantic level.

[0051] To further improve the feature recognition ability, a contrast learning loss function based on InfoNCE is introduced. Its core is to maximize the feature similarity of positive samples (such as clogging scenarios with high humidity and overheating of the screen at the same time) and minimize the similarity of negative samples (such as the combination of normal humidity and abnormal temperature). This mechanism forces the model to learn the core semantic associations rather than simple correlations. For example, in a sand and gravel plant, in a high-humidity environment during the rainy season, it can give an early warning of screen clogging and reduce the false alarm rate. By embedding the adjusted feature distance into the perception data, the model can capture implicit associations that are difficult to discover by traditional methods, such as the phase difference between humidity fluctuations and vibration spectra indicating clogging, so as to achieve more accurate risk assessment.

[0052] Step S102: Based on the multi-modal perception data and the internal structure of the greening crusher, a clogging risk model of the greening crusher is established through a spatio-temporal graph neural network.

[0053] In the embodiments of the present application, the blockage risk model is marked with the blockage risk levels of each key internal structure and the corresponding blockage risk factors. The blockage risk model is constructed with a spatio-temporal graph neural network as the core, combining multi-modal perception data and the characteristics of the internal structure of the equipment to achieve accurate prediction and analysis of blockage risks. The model abstracts each key internal structure of the crusher, such as the feed inlet, crushing chamber, screen, etc. as nodes in the graph structure, and defines the material transmission path and mechanical drive connection between nodes as edges, thus constructing a graph model reflecting the internal operation relationship of the equipment. At the same time, multi-modal perception data such as vision, hearing, vibration, temperature, etc. are used as node features, and the time dimension is incorporated through a time sliding window to capture the dynamic changes of the equipment state over time.

[0054] In terms of the model architecture, the data first undergoes feature extraction layer to achieve multi-modal fusion, then enters the spatial graph convolution layer to mine the correlation relationships between structures, and then is analyzed by the time convolution layer for temporal changes. The attention mechanism is used to focus on key risk points, and finally the blockage risk levels of each key internal structure are obtained at the output layer, and the corresponding blockage risk factors are identified, such as foreign object entry, overload, component wear, etc. The model adopts a four-level risk classification standard, warning from normal to dangerous in sequence, and with the help of the graph attention mechanism, the risk assessment results have good interpretability, not only can predict blockage risks, but also clearly indicate the root causes of risks, providing strong data support and decision-making basis for preventive maintenance of the equipment.

[0055] As an alternative embodiment, in step S102, based on the internal structure of the greening crusher, a three-dimensional structure model of the greening crusher is constructed. Then, based on the acquisition time and the physical position information of the corresponding internal structure, each perception feature data in the multi-modal perception data is mapped into the three-dimensional structure model. Furthermore, each key internal structure in the three-dimensional structure model is used as the corresponding graph node, and the corresponding connection edges are set, where the key internal structures at least include: the material inlet and outlet of the greening crusher, the material crushing chamber, the screen, and the impact plate. Then, an adjacency matrix is constructed based on the material flow path and mechanical connection relationship to obtain the edge weights corresponding to each connection edge. The edge weights are used to indicate the blockage propagation probability between the connected graph nodes. Then, contrastive learning is used to extract spatio-temporal associated positive sample pairs from historical blockage events, and randomly select negative sample pairs without historical association, and optimize the contrastive loss function through the positive and negative sample pairs to train the risk prediction model. The node embedding features of each graph node in the three-dimensional structure model are updated through incremental graph convolution to adapt to the change of material characteristics. Finally, the trained risk prediction model is used to identify congestion risks and predict the congestion degree in the real-time updated three-dimensional structure model, and obtain the blockage risk levels and corresponding blockage risk factors of each graph node in the three-dimensional structure model, so as to construct the final output blockage risk model.

[0056] It is understandable that the construction of the greening crusher blockage risk model integrates 3D structure modeling, multi-modal data mapping, and spatio-temporal graph neural networks to achieve accurate risk prediction through the deep coupling of the physical space and the data space. Its core principle is to abstract the physical structure of the equipment into a graph structure, construct an adjacency matrix through the material flow path and mechanical connection relationship, enabling the model to capture the propagation mechanism of blockage inside the equipment. The introduction of the 3D structure model provides a spatial index for multi-modal perception data, associates time series data with physical locations, and forms a four-dimensional spatio-temporal data structure. The contrast learning mechanism mines spatio-temporal correlation patterns through historical events, and uses positive and negative sample pairs to optimize the model's sensitivity to risk features, while incremental graph convolution enables the model to adapt to changes in material characteristics and maintain long-term prediction accuracy.

[0057] First of all, constructing a 3D structure model is the foundation for establishing the entire blockage risk model. With the help of advanced 3D modeling technologies, such as laser scanning reverse modeling methods, the internal structure of the greening crusher is finely characterized, and key structures such as material inlets and outlets, crushing chambers, screening meshes, and impact plates are digitally presented. During the modeling process, the dimensions, shapes, and relative position relationships of each component are accurately restored, thus forming a highly realistic digital mapping of the physical equipment and establishing a spatial framework for subsequent data integration and analysis.

[0058] Next, based on the acquisition time and physical location information, multi-modal perception data is deeply integrated with the 3D structure model. Multi-modal perception data covers various types of information such as vibration data, temperature data, and image data generated during the operation of the equipment, and each piece of data contains the operating state characteristics of the equipment in different dimensions. By accurately matching the acquisition time of the data with the physical location of the corresponding internal structure, these perception data are precisely mapped to the corresponding structures of the 3D model. For example, the vibration data collected by a vibration sensor installed on the wall of the crushing chamber is accurately associated with the crushing chamber structure node in the 3D model according to its acquisition time and installation location. Through this process, the originally scattered multi-modal data forms a four-dimensional data set with spatio-temporal correlation characteristics, enabling the data to not only contain the characteristic information of the equipment operation but also have the dual dimensions of spatial location and time sequence, providing a richer and more three-dimensional data basis for subsequent analysis.

[0059] Then, the key structures in the three-dimensional model are transformed into graph structures. Key structures such as material inlet and outlet, crushing chamber, screen, impact plate, etc. are abstracted as graph nodes, and the flow path of materials inside the equipment and the mechanical connection relationship between the components are defined as connection edges. In order to quantify the connection relationship represented by the edge, through in-depth analysis of historical operation data, the mutual influence between the components when the equipment was blocked in the past was studied. At the same time, combined with the principles of fluid mechanics, the flow process of materials inside the crusher was simulated and analyzed. On this basis, the edge weights corresponding to each connection edge are determined. These weights can reflect the probability of blockage propagation between the connected graph nodes. Finally, an adjacency matrix is ​​constructed, which completely and accurately describes the correlation between the key structures inside the crusher and the possibility of propagation of blockage risks, providing a structured data expression for the prediction of blockage risks.

[0060] During the model training phase, a contrastive learning strategy is used to mine potential patterns in the data. From a large amount of historical blockage event data, sample pairs with spatiotemporal correlation are selected as positive samples. For example, in a certain blockage event, the crushing chamber first experiences abnormal vibration, and then the degree of blockage of the screen gradually increases. Such a combination of abnormal data of the crushing chamber and the screen constitutes a positive sample pair, which reflects the true propagation path and correlation relationship of the blockage risk inside the equipment. At the same time, samples that are not associated in historical data are randomly selected as negative sample pairs to enhance the model's ability to discern risk features. Through positive and negative sample pairs, the contrast loss function of the model is optimized, so that the model can better learn the key features and correlation patterns when the blockage risk occurs, significantly improve the model's sensitivity to risk features, and thus more accurately identify potential blockage risks.

[0061] In order to enable the model to adapt to changes in the properties of different materials, an incremental graph convolution mechanism is designed. During the actual operation of the equipment, the properties of the processed materials such as humidity, hardness, and shape will continue to change, and these changes will affect the operating status and blockage risk of the equipment. The incremental graph convolution mechanism can update the node embedding features of each graph node in the three-dimensional structure model in real time according to the newly input data without retraining the entire model. For example, when the humidity of the processed material suddenly increases, the model will automatically adjust the node embedding features related to humidity and re-evaluate the correlation between each node and the probability of blockage propagation, so that the model can adapt to changes in material properties in a timely manner and maintain the accuracy and effectiveness of the prediction of blockage risk.

[0062] Finally, the model trained through the above steps can efficiently analyze the real-time collected data. Based on the input real-time data and combined with the learned knowledge and patterns, the model conducts a blockage risk assessment on each graph node in the three-dimensional structure model and accurately outputs the blockage risk level of each node. Moreover, by analyzing the model calculation process and internal logic, it is possible to determine the specific reasons for the corresponding risk levels of each node, providing detailed and targeted risk warning information for equipment maintenance personnel, helping them take effective preventive and treatment measures in a timely manner, reducing the occurrence probability of equipment blockage failures, and ensuring the stable operation of the green waste crusher.

[0063] All in all, in terms of the spatial dimension, the model can accurately locate blockage risk points, such as key structures like material inlets and outlets, and crushing chambers. In the time dimension, by capturing temporal anomalies such as sudden increases in humidity, it can early warn of potential risks. At the interpretability level, the dynamic update mechanism of edge weights and node embedding features enables the model to output interpretable risk factors, such as the impact of material property changes on the crushing chamber. This spatio-temporal aware graph neural network model not only improves the accuracy of blockage prediction but also enhances the generalization ability of the model through physical structure constraints, making it particularly suitable for the intelligent operation and maintenance scenarios of complex industrial equipment.

[0064] Further optionally, after constructing the final output blockage risk model in the above steps, a blockage risk heat distribution model matching the blockage risk model can be generated through digital twin technology. The blockage risk heat distribution model is used to quantify the regional congestion risk values and corresponding potential congestion reasons corresponding to each graph node and each connecting edge.

[0065] After completing the construction of the green waste crusher blockage risk model, introducing digital twin technology to generate a blockage risk heat distribution model realizes the leap from abstract risk prediction to visual and quantifiable analysis. This process is based on the three-dimensional structure model as the basic framework, and maps the node risk levels output by the blockage risk model and the blockage propagation probability of the edges to the virtual three-dimensional space in real time through the digital twin engine. Specifically, using color gradients (such as cold colors representing low risks and warm colors representing high risks) and transparency changes, the risk values of each key structure (graph nodes such as material inlets and outlets, and crushing chambers) are visually presented. At the same time, the visual expression of the connecting edges is enhanced by line thickness or light and shadow effects, intuitively showing the blockage propagation paths and probabilities, forming a dynamic and three-dimensional risk heat map.

[0066] This link breaks through the limitation of traditional digital twins which are only used for equipment status monitoring, deeply binds the prediction results of spatio-temporal graph neural networks with the 3D physical model, enabling the virtual model to not only reflect the equipment operation status, but also prospectively display the risk evolution trend. Through the original heat value calculation algorithm, abstract parameters such as the risk level of graph nodes and edge weights are converted into quantifiable and comparable heat data, realizing the accurate visual expression of the risk level. Combining knowledge graphs and digital twin technology, potential congestion causes are automatically associated with each heat area. When the heat value of a certain area suddenly increases, explanatory information such as "the material humidity exceeds the standard, resulting in increased viscosity" and "the impact plate is worn, causing a decrease in crushing efficiency" can be immediately displayed, forming a complete analysis chain of "risk positioning - degree quantification - cause tracing".

[0067] Through this heat distribution model of clogging risk, maintenance personnel can not only quickly identify high-risk areas, but also predict the risk diffusion path through the heat change trend. Combining the cause analysis provided by the system, precise preventive maintenance strategies can be formulated. This way of combining data-driven risk prediction with visual digital twin technology significantly improves the intuitiveness and decision-making efficiency of equipment failure warning, providing a new solution for the intelligent maintenance of industrial equipment.

[0068] Further optionally, after using the trained risk prediction model to identify the congestion risk and predict the congestion degree of the multi-modal perception data in the above steps, a structural causal model (SCM) can be used to identify the key factors of the current green crusher's clogging in the heat distribution information of the clogging risk, and transmit them to the cloud for sharing with other green crushers. Furthermore, based on the key clogging factors shared by multiple green crushers in the cloud, through counterfactual reasoning, the causal relationships of each key clogging factor under different working conditions and different equipment are verified, and the verification results are applied to the dynamic parameter optimization of the risk prediction model to avoid the interference of pseudo-correlation of a single equipment.

[0069] Suppose in a certain garden waste treatment plant, a green crusher frequently experiences screen clogging when processing wet leaves. The risk prediction model based on multi-modal perception data (such as vibration, temperature, images, etc.) identifies a high congestion risk at the screen. At this time, a structural causal model (SCM) is introduced for in-depth analysis. The SCM constructs a causal graph containing variables such as material humidity, crushing speed, and screen aperture, and analyzes that "excessively high material humidity" is the key factor leading to screen clogging, and uploads this conclusion to the cloud.

[0070] The cloud collects data on the key factors causing blockages shared by multiple green waste crushers in different regions. For example, some equipment gets blocked due to "too fast feeding speed", while some due to "severely worn sieve mesh". Based on this data, counterfactual reasoning is used for verification. Taking the key factor of "too high material humidity" as an example, counterfactual reasoning will simulate whether blockages will still occur when the material humidity is reduced while other equipment is under the same working conditions (such as processing wet leaves). Suppose a certain piece of equipment gets blocked due to material humidity during actual operation. By simulating the situation when the humidity is reduced through counterfactual reasoning, if the blockage no longer occurs, it further verifies the causal relationship between material humidity and blockage. On the contrary, if it is still blocked, it indicates that there are other hidden influencing factors.

[0071] In this way, it is possible to effectively eliminate the interference of spurious correlations caused by special working conditions or individual differences of a single piece of equipment. For example, if a certain piece of equipment gets blocked because the operator made a mistake and mixed stones during feeding, the causal relationship resulting from this special situation will be identified as a spurious correlation in the verification of multi - equipment data. Finally, the verified and reliable causal relationships are applied to the dynamic parameter optimization of the risk prediction model. For example, in the risk prediction model, the weight of the parameter of material humidity is increased. When an increase in humidity is detected, a more accurate early blockage warning can be issued, thereby improving the accuracy and generalization ability of the entire risk prediction model.

[0072] It should be noted that counterfactual reasoning is a hypothesis-based reasoning method used to evaluate the possible outcomes of an event or variable under hypothetical conditions contrary to the actual situation. First, a causal model needs to be constructed to describe the causal relationships between variables. This model is represented in the form of a directed acyclic graph (DAG), where nodes represent variables, edges represent the causal relationships between variables, and the arrow direction indicates the direction of causal influence. For example, in the case of a green waste crusher, the causal model may include variables such as material humidity, crushing speed, screen aperture, and feeding speed, as well as the causal relationships between them. For instance, too high material humidity can cause screen clogging, and too fast crushing speed may affect the material crushing effect and thus influence the clogging situation. Determine the counterfactual hypothesis according to the specific problem, that is, imagine conditions different from the actual situation. For example, in the above example, the actual situation is that high material humidity has caused screen clogging, then the counterfactual hypothesis can be "What would happen if the material humidity were reduced to a certain lower level". Under the counterfactual hypothesis, infer the possible outcomes based on the relationships between variables in the causal model. This usually involves intervening in the causal model, that is, assuming that the value of a certain variable has changed, and then observing the changes in other variables. For example, under the assumption that the material humidity is reduced, based on the relationship between material humidity and screen clogging in the causal model, as well as the influence of other relevant variables, infer whether the screen will still be clogged or whether the probability of clogging will be reduced. Evaluate the causal effect by comparing the counterfactual result with the actual result. If, under the counterfactual hypothesis, the result has changed significantly, for example, the screen is no longer clogged or the degree of clogging is significantly reduced, then it can be considered that there is a causal relationship between the original variable (such as material humidity) and the result (screen clogging). On the contrary, if the result has not changed significantly, it indicates that there may be other factors affecting the result, or the causal relationship between the original variable and the result is not strong. Counterfactual reasoning, through steps such as constructing a causal model, determining a counterfactual hypothesis, reasoning based on the model, and evaluating the result, understands the causal relationships between variables in a complex system, so as to analyze problems more accurately, predict results, and provide a strong basis for decision-making.

[0073] In this application, the improvement of counterfactual reasoning is mainly reflected in the combination of a multi-device causal verification mechanism and a pseudo-correlation filtering strategy.

[0074] Traditional counterfactual reasoning usually conducts hypothesis verification based on the causal model of a single system (such as "If the material humidity of this device is reduced, will it still be clogged?"), but it is vulnerable to the special working conditions of a single device (such as individual wear differences, operating habits) or accidental noise interference, resulting in pseudo-correlation conclusions (such as attributing the clogging caused by accidentally mixed stones to humidity).

[0075] The improvement of this application lies in sharing the key factors of blockage of multiple green crushers (such as "too high material humidity", "too fast feeding speed", "screen wear", etc.) through the cloud, and constructing a cross-device causal verification dataset. Another improvement of this application is cross-condition generalization reasoning. Taking a certain key factor (such as "material humidity") as an example, counterfactual reasoning is no longer limited to a single device, but simulates hypothetical conditions (such as "the material humidity of all devices is reduced by 20%") in device clusters with different regions and usage frequencies, and observes whether the blockage results are generally weakened. For example, if a certain device is blocked due to the operator's long-term overfeeding, the association between its "feeding speed" and "blockage" may be misjudged as a causal relationship in single-device counterfactual reasoning. However, in multi-device verification, if the blockage does not decrease after other devices reduce the feeding speed, this association can be identified as a pseudo-correlation of a single device.

[0076] In addition, compared with traditional counterfactual reasoning that often relies on experience to set causal relationships, lacks explicit modeling of the hierarchical relationships between complex system variables, and is difficult to quantify the interactive effects of multiple factors.

[0077] In the embodiments of this application, a causal graph including "material characteristics → crushing parameters → structural load → blockage result" (such as material humidity → torque of the crushing cavity → screen passing rate) is also pre-constructed to clarify the direct or indirect causal paths between variables. During counterfactual reasoning, specific variables are intervened based on the SCM (such as fixing the crushing speed and only adjusting the humidity) to exclude the interference of confounding factors. Furthermore, the reliable causal relationships verified across devices (such as "for every 10% increase in humidity, the screen blockage probability increases by 25%") are converted into dynamic parameter weights of the risk prediction model. For example, when the cloud verifies that "humidity" is a universal key factor, the model will automatically enhance the attention to the data of the humidity sensor to avoid misjudgment caused by sensor errors of a single device.

[0078] In summary, the embodiments of this application perform double verification in terms of time and space. That is, in the time dimension, it is required that the key factor and the blockage event show a consistent temporal association in multiple devices (such as "sudden increase in humidity → blockage after 2 hours" appears repeatedly in ≥80% of the devices), excluding single-occurrence coincidences. In the space dimension, through horizontal comparison between devices, "personalized associations" that only hold in specific devices are identified (such as a strong correlation between "impact plate angle" and blockage in a certain device due to design defects), avoiding misjudging individual device differences as universal causal relationships.

[0079] Furthermore, the causal effect value (CATE, Conditional Average Treatment Effect) is introduced as a verification index, and only when the absolute value of CATE of multiple devices exceeds a threshold (such as the change in blockage probability > 15%) is it considered an effective causal relationship.

[0080] Step S103: Dynamically adjust the dynamic control instructions of the green waste crusher based on the blockage risk model.

[0081] In the embodiments of the present application, the dynamic control instructions at least include: feed rate, scheduling instructions for different particle size crushing units, temperature supplement parameters, and humidity supplement parameters.

[0082] Step S104: Generate an optimal blockage clearing path using the graph contrastive loss function, and execute a differentiated blockage clearing plan that matches the optimal blockage clearing path based on the dynamic control instructions to preferentially process the high-risk nodes in the blockage risk model, so as to achieve hierarchical blockage clearing of the green waste crusher and ensure the normal operation of the green waste crusher.

[0083] As an alternative embodiment, in step S103, dynamically adjusting the dynamic control instructions of the green waste crusher based on the blockage risk model includes performing the following steps:

[0084] Step 201: Through the spatio-temporal attention mechanism, use dilated causal convolution to obtain the congestion risk feature matrix at each moment in the blockage risk model.

[0085] Among them, the congestion risk feature matrix at least includes: the congestion risk features of each node, and the cross-node congestion risk association features between multiple nodes. The cross-node congestion risk association features at least include: material particle size features, feed rate features, and vibration frequency features.

[0086] Step 201 focuses on the risk associations of different moments and different structural nodes through the spatio-temporal attention mechanism, combines dilated causal convolution to expand the receptive field in the time dimension, and captures the time lag of risk propagation (such as the causal relationship between the increase in the screen pressure 2 minutes after the feed inlet is blocked). The congestion risk feature matrix not only contains the risk values of individual nodes (such as abnormal temperature in the crushing chamber), but also models cross-node associations through the graph attention mechanism (such as the combined impact of material particle size changes on the crushing chamber and the screen).

[0087] For example, when it is detected that the material particle size at the feed inlet suddenly increases, the model uses dilated causal convolution to identify the historical association between this event and the screen blockage 30 minutes later, and enhances the cross-node association weight of "feed inlet - screen" in the feature matrix.

[0088] Thus, it breaks through the limitation of traditional control that only relies on real-time data of a single node, captures the risk propagation chain in advance, and enables the control strategy to have causal prediction ability. For example, when the feed inlet is abnormal, the vibration parameters of the downstream screen are adjusted in advance, rather than waiting for the blockage to occur and then responding passively.

[0089] Step 202: For the graph nodes associated with the material granularity characteristics in the congestion risk characteristic matrix, a hierarchical control strategy is adopted to identify the current granularity control mode and generate corresponding hierarchical granularity control parameters.

[0090] The classification particle size control parameter includes one of the following: an impact plate angle control instruction, a jaw crusher start instruction in a coarse particle size priority mode, and a cone crusher start instruction in a fine particle size compensation mode.

[0091] Step 202: According to the risk level of nodes (such as crushing chamber and impact plate) associated with the particle size characteristics of the material, control modes such as "coarse particle size crushing" and "fine particle size compensation" are divided. The material characteristics are dynamically matched through classification parameters (such as impact plate angle, jaw / cone crusher start and stop). The coarse particle size mode increases the crushing chamber space and reduces the blockage of large particles. The fine particle size mode starts the secondary crushing equipment to avoid the adhesion of too fine materials.

[0092] For example, when the model detects that the "material particle size > 5cm" feature of the crushing chamber node continues to strengthen and the wear of the impact plate causes a decrease in crushing efficiency, it automatically generates a "jaw crusher start command" to form graded crushing with the original impact crushing, and at the same time adjusts the impact plate angle to 45° (the optimal angle for coarse-grained mode).

[0093] Therefore, the present application can solve the defect of the "one-size-fits-all" crushing mode of traditional crushers. For example, when the fine-grained compensation mode is used for easily sticky materials such as wet leaves, the blockage incidence rate is reduced by 40%, which is more adaptable to fluctuations in material characteristics than fixed parameter control.

[0094] Step 203: For the graph nodes associated with the feed rate characteristics in the congestion risk characteristic matrix, identify whether the feed rate reaches the rate range corresponding to the current congestion risk level, and output the corresponding adjustment instruction to link the frequency converter to dynamically adjust the belt conveyor speed to reduce the feed rate under high congestion risk levels.

[0095] Step 203, associate the feed rate characteristics with the node risk level, and set a multi-level rate range (such as normal / warning / dangerous levels corresponding to different feed upper limits). The current risk level is matched in real time by linking the belt conveyor with the frequency converter: when the risk is high (such as the risk of screen blockage>80%), the feed rate is proportionally reduced (such as from 50t / h to 30t / h) to reduce the instantaneous load. For example, when the vibration frequency characteristics of the screen node show "abnormal high-frequency vibration" (indicating possible blockage), the model recognizes that it is currently in a "dangerous level" and immediately outputs a command to reduce the feed rate to 60% of the baseline value. At the same time, the belt conveyor speed feedback closed-loop control is used to ensure that the rate adjustment error is less than 5%. In this way, the hysteresis of the traditional manual setting of the rate is avoided. For example, in a sudden load scenario, it can respond 2 minutes earlier than the fixed rate control, reducing the risk of blockage caused by excessive feeding by 65%.

[0096] Step 204: for the graph nodes associated with the vibration frequency characteristics in the congestion risk characteristic matrix, identify the current anti-sticking vibration mode, generate corresponding vibration motor control instructions, and adaptively adjust the vibration motor power.

[0097] Among them, the vibration motor control instructions are: low-frequency switching instructions in anti-blocking mode, or high-frequency switching instructions in flow promotion mode; the vibration motor power is adjusted as: the product of the reference power and the difference power between the upper and lower pressure limits, divided by the real-time load power; the difference power between the upper and lower pressure limits is the difference between the peak safety power and the no-load power. For example, when the risk score is >50%, the vibration motor in the crushing chamber switches to the 5-10Hz low-frequency mode, and the amplitude is increased to 3mm to prevent material adhesion. When the risk score is <30%, it switches to the 15-20Hz high-frequency mode, and the amplitude is reduced to 1mm to improve the crushing efficiency.

[0098] Step 204, according to the risk mode of the nodes associated with the vibration frequency characteristics (such as the screen and the discharge port), dynamically switch between the "anti-blocking low-frequency mode" (10-20Hz, suitable for the blockage of large particles in the crushing chamber) and the "promoting flow high-frequency mode" (50-80Hz, suitable for the adhesion of wet materials). The power of the vibration motor is based on the dynamic calculation results of the real-time load, which can ensure the optimal vibration intensity under different blockage risks. For example, when the screen node detects "humidity>70% and material retention time>10s", it is judged to enter the "bonding risk mode", automatically switch to high-frequency vibration (60Hz), and calculate the power compensation to 120% of the baseline value according to the current load (such as motor current 15A), and quickly remove the bonded materials on the screen surface. Compared with fixed-frequency vibration, adaptive control shortens the screen blockage time and reduces energy consumption, solving the problem of energy consumption imbalance in traditional vibration control.

[0099] Step 205: For the graph nodes in the congestion risk feature matrix associated with the temperature feature, generate temperature compensation parameters that match each internal structure under the current working conditions, and initiate the corresponding temperature compensation operation.

[0100] Step 205: For the nodes associated with the temperature feature (such as bearings and motors), generate temperature compensation parameters according to the current working conditions: In a low-temperature environment (<0°C), start the heating device to prevent the lubricating oil from solidifying; in a high-temperature environment (>60°C), start the cooling fan to reduce component losses. The compensation parameters are combined with the real-time load of the equipment (for example, the higher the load, the temperature threshold is dynamically lowered by 10%) to avoid misjudgment of a single temperature threshold. For example, in the early morning of winter, when the temperature of the crushing chamber bearing is detected to be <5°C and the equipment has just started with a low load, automatically turn on the heating tape of the bearing seat (power 500W) to maintain the temperature at 15 - 20°C; when the temperature of the motor is >70°C and the load is continuously high in summer, link the cooling system to control the temperature below 65°C. Thus, the incidence rate of temperature-related component failures (such as bearing jams and motor overheating) is reduced, which is more in line with the actual working conditions compared to traditional constant temperature control. For example, in outdoor scenarios with large temperature differences between day and night, the energy consumption is reduced.

[0101] Through steps 201 - 205, in the past, relying on sensor alarms after blockages occurred (such as sudden increases in current) led to large shutdown losses. Through spatio-temporal risk feature prediction, early intervention is carried out at the budding stage of the risk (such as abnormal particle size and vibration frequency changes). Traditional control only adjusts a single variable (such as the feeding rate), ignoring the combined effects of multiple factors (such as the combined action of particle size + humidity + vibration). Through feature matrix modeling and cross-node association, multi-dimensional collaborative control is achieved (such as simultaneously adjusting the crushing mode and vibration frequency when the particle size is abnormal). Fixed control parameters are difficult to cope with the dynamic changes of material properties (humidity, particle size) and the environment (temperature, load); this solution uses a hierarchical strategy and an adaptive algorithm to match the working conditions in real time. For example, for wet materials, automatically switch to a combined strategy of high-frequency vibration + fine particle size crushing. Through causal prediction and multi-parameter collaborative adjustment, the blockage incidence rate at key nodes (crushing chamber, screen) is reduced compared to traditional control. The dynamic control strategy reduces the shutdown time caused by blockages (the single blockage handling time is shortened from 30 minutes to 5 minutes), and the production capacity is improved. Adaptive power adjustment (such as vibration motors, frequency converters) reduces the comprehensive energy consumption of the equipment, especially showing significant advantages under variable load working conditions.

[0102] Based on the above embodiments, in step S104, the reward function set by the multi-objective reinforcement learning mechanism is used to dynamically update the policy parameters in the graph contrast loss function. Furthermore, through the updated policy parameters, determine the priority of the blockage clearing operations for each internal structure in the green crusher under the current working conditions, so as to establish the optimal blockage clearing path under the current working conditions; perform blockage clearing operations that match the dynamic control instructions on different internal structures based on the optimal blockage clearing path.

[0103] The graph contrastive loss function is an objective function used to optimize the node embeddings of graph neural networks. The core idea is to force the model to learn more discriminative risk feature representations by contrasting positive sample pairs (such as nodes or edges with real blockage associations) and negative sample pairs (unrelated or randomly generated nodes or edges). In the multi-objective reinforcement learning framework, this function is coupled with the reward function to dynamically adjust the policy parameters, enabling the model to adapt to blockage removal decisions under complex working conditions.

[0104] The optimal blockage removal path is the optimal operation sequence from the current risk state to the safe state solved by multi-objective reinforcement learning with the goal of minimizing the blockage removal cost (time, energy consumption, equipment wear, etc.) in the graph structure model of the green crusher. Each node in the path represents a key internal structure (such as the crushing chamber, screen), and the edge represents a blockage removal operation (such as adjusting the angle of the impact plate, starting the vibration motor). The path weight is jointly determined by the operation cost and the risk reduction effect.

[0105] The differentiated blockage removal plan is a targeted operation strategy formulated based on the risk characteristics, physical properties, and correlation relationships of each node in the optimal blockage removal path. For example, mechanical adjustment is used for the crushing chamber blocked by high-rigidity materials, and a combination of high-frequency vibration and humidity compensation is used for the screen adhered by wet materials to avoid inefficiency or equipment damage caused by one-size-fits-all operations.

[0106] For example, the optimal operation is matched for different blockage causes (such as jamming, adhesion, overload) to improve the blockage removal efficiency. For example, low-frequency strong vibration is used for the jamming of the screen, and high-frequency micro-vibration and heating are used for adhesion. Or, the blockage removal resources are dynamically allocated according to the current load of the equipment. When multiple nodes give early warnings simultaneously, full vibration power is preferentially allocated to the nodes with a high propagation probability, and the secondary nodes use hierarchical power output.

[0107] In addition, excessive vibration is avoided for vulnerable components (such as the screen), and a larger range of angle adjustment is allowed for wear-resistant components (such as the impact plate) to extend the equipment life.

[0108] Through this linkage, the most cost-effective blockage removal plan can be automatically selected under complex working conditions. For example, when dealing with suddenly high-humidity materials, the crushing mode of the crushing chamber is preferentially adjusted (reducing the blockage source), and the vibration power of the screen is simultaneously increased (preventing adhesion) to improve the blockage removal efficiency of the equipment.

[0109] As an alternative embodiment, in step S103, after dynamically adjusting the dynamic control instruction of the green crusher based on the blockage risk model, multiple upstream and downstream green crushers can be connected through the OPC-UA protocol. Based on the blockage risk model, the linkage control effect of multiple green crushers under the dynamic control instruction can be simulated online; the linkage control effect can be evaluated through a global benefit function, and the material flow rate among multiple green crushers can be dynamically allocated to optimize the dynamic control instruction in real time; the optimized dynamic control instruction can be sent to multiple green crushers through the OPC-UA protocol.

[0110] Specifically, after dynamically adjusting the dynamic control instruction of the green crusher based on the blockage risk model, multiple upstream and downstream green crushers can be connected through the OPC-UA protocol to construct a multi-device linkage control system. First, using high-precision digital twin modeling technology, the operating states of each crusher, such as the motor load rate and the screen blockage index, are synchronized in real time through OPC UA to construct a virtual mirror model, ensuring that the model error rate is less than 2%. On this basis, a discrete event simulation engine (such as SimPy) is introduced to simulate the flow path of materials among multiple crushers, accurately calculate the blockage propagation time delay, with an accuracy of ±50 ms. At the same time, the graph neural network model is extended to construct a cross-device graph structure. The nodes include local risk scores and the states of adjacent devices, and the edge weights reflect the material transmission rate. Combining with the time series prediction algorithm, based on historical blockage event data, the risk diffusion trend within the next 10 seconds is predicted to provide a scientific basis for linkage decision-making.

[0111] In terms of the optimal design of the global benefit function, a multi-objective benefit evaluation model is adopted, and the formula is as follows:

[0112]

[0113] Among them, the parameters , , , can be dynamically adjusted according to different scenarios. is used to measure the contribution degree of the reduction of blockage risk to the system benefit, and respectively represent the risk score and the corresponding risk change amount of the th green crusher at the previous moment. For example, in the high-risk emergency mode, can be increased to 0.7. The larger , the more the risk is reduced, and the more significant the contribution to the total benefit. is reduced to 0.3 to minimize energy consumption. represents the energy consumption optimization weight, reflecting the importance of energy consumption control to the system benefit. In the emergency mode, Reduce it to 0.1 to preferentially reduce the risk of blockage. In the energy consumption priority mode, it can be increased to 0.5. For example, during the low electricity price period at night, the system extends the idling cycle and reduces the motor frequency to save electricity. represents the change in the current energy consumption of the th green crusher compared with the energy consumption at the previous moment. represents the energy consumption reference value of the th green crusher at the previous moment. characterizes the relative change range of the equipment energy consumption of the th green crusher, reflecting the influence efficiency of the control strategy on energy consumption. represents the actual processing capacity of the th green crusher. represents the production capacity optimization weight, which can usually be set to a fixed value. For example, it is set to 0.1. represents the comprehensive risk score of the th green crusher, which is calculated by the GCCN model based on the congestion risk level from multi-modal sensor data. represents the penalty term weight. Through this dynamic weight adjustment mechanism, the system can adapt to the requirements of different production scenarios and achieve comprehensive optimization of risk, energy consumption, and production capacity.

[0114] In terms of the dynamic material flow distribution strategy, a flow scheduling method based on reinforcement learning is introduced. The state space includes 12-dimensional information such as the water level in the silos of each device, the conveyor belt load rate, and the crusher load. The action space defines 6 flow distribution modes, such as "full speed direct through" and "graded buffering".

[0115] In terms of the dynamic material flow distribution strategy, a flow scheduling method based on reinforcement learning is introduced. The state space includes 12-dimensional information such as the water level in the silos of each device, the conveyor belt load rate, and the crusher load. The action space defines 6 flow distribution modes, such as "full speed direct through" and "graded buffering".

[0116] In terms of enhancing the OPC UA communication protocol, a security enhancement mechanism is mainly considered. The X.509 certificate is used for two-way authentication of devices to prevent illegal access, and AES-256 encryption is used for transmitting key control instructions (such as vibration frequency adjustment) to ensure communication security. At the same time, the information model in the OPC UA address space is extended, and new "linked logistics" nodes are added, including information such as the real-time material flow rate (m³ / s), the capacity of the buffer silos between devices (t), and the linked control mode (manual / automatic). Through the subscription / notification mechanism, when the risk score of a certain device exceeds 70%, an inter-device event notification is automatically triggered, and the delay is controlled within 100 ms to ensure fast response. The introduction of an adaptive learning and fault tolerance mechanism further improves the robustness of the system.

[0117] Further optionally, a three - level fusing strategy can also be designed. When abnormal situations such as single - device communication interruption, multi - device data conflict, or model prediction deviation exceeding 20% occur, the system can automatically switch to the locally preset mode, degrade to single - machine control, or pause linkage and trigger the intervention of the expert system to ensure the safety and reliability of the system.

[0118] Further optionally, in step S103, the wear degrees of various internal structures of the green crusher (such as the inner wall of the crushing chamber, the screen, the impact plate, the hammer head, etc.) can also be predicted based on the blockage risk model, and corresponding dynamic control instructions can be generated based on the wear degrees of various internal structures to reduce the damage caused by material congestion to the internal structures of the green crusher, extend the service life of the equipment, and improve the operation smoothness and production efficiency at the same time.

[0119] Specifically, the blockage risk model is not only used to evaluate the current and future blockage risks, but can also predict the wear rates and remaining lives of each internal component by analyzing historical data and real - time sensor data (such as vibration spectrum, temperature distribution, pressure fluctuation, etc.), combined with machine learning algorithms (such as random forest regression, LSTM time - series prediction, or wear prediction algorithms based on physical models). For example, the wear of the inner wall of the crushing chamber may be closely related to the hardness, impact frequency, and humidity of the material, while the blockage and wear of the screen are directly related to the particle size distribution and humidity of the material. By real - time monitoring the change trends of these parameters, the model can predict the wear degree in advance and adjust the control strategy accordingly.

[0120] When generating dynamic control instructions, the system will take targeted measures according to the predicted wear degree. For example, if it is predicted that the wear rate of the inner wall of the crushing chamber accelerates, the system can automatically reduce the impact frequency of the crushing chamber or adjust the angle of the impact plate to reduce the impact force on the inner wall; if the wear risk of the screen is high, the system can increase the vibration frequency of the screen or adjust the tension of the screen to prevent blockage and slow down the wear; for vulnerable parts such as hammer heads, the system can give an early warning according to the wear prediction result and automatically adjust the load distribution of the crushing unit when necessary to avoid premature damage of components caused by over - load operation.

[0121] In addition, the system can also synchronize the wear prediction results and dynamic control instructions to multiple green crushers upstream and downstream in real - time through the OPC - UA protocol to achieve collaborative optimization among multiple devices. For example, in the scenario of multi - crusher linkage, if a certain component of a certain device is severely worn, the system can dynamically adjust the material flow distribution, reduce the load pressure on this device, and optimize the operation parameters of other devices to make up for the production capacity loss, so as to balance the production efficiency and equipment maintenance cost globally.

[0122] Through this dynamic control strategy based on wear prediction, the system can not only effectively reduce the damage caused by material congestion, but also significantly improve the operation smoothness and reliability of the green crusher, extend the service life of the equipment, and reduce the downtime and maintenance costs caused by equipment failures. This optimization solution is particularly suitable for high-load and continuous industrial scenarios, and can bring significant economic benefits and competitive advantages to enterprises.

[0123] In the embodiments of this application, the internal material state data of the green crusher is collected in real time through a multi-modal sensor array, and a blockage risk model is constructed using a spatio-temporal graph neural network to achieve accurate risk assessment and dynamic control. The multi-modal sensors cover key parameters such as humidity, particle size, and pressure. Combining with the spatio-temporal graph neural network, the material flow path and blockage propagation law can be dynamically captured, high-risk nodes (such as the edge of the screen and the dead corner of the impact plate) can be marked, the risk level of each area can be quantified, and the problem of insufficient recognition of hidden blockages by traditional methods can be solved. Based on the blockage risk model, the system automatically adjusts the feeding rate, particle size scheduling, and temperature and humidity compensation parameters. For example, when the risk is high, the feeding speed is reduced and atomized drying is started to avoid equipment overload. The dynamic control instructions make the equipment operate more smoothly and reduce sudden failures. A graph contrast loss function is used to generate the optimal blockage clearing path, giving priority to processing high-risk nodes, and combining with a differentiated blockage clearing scheme (such as a combination of high-frequency vibration and air flow impact) to quickly clear stubborn blockages. The hierarchical blockage clearing strategy ensures the overall efficiency of the production line and shortens the single blockage clearing time. This application realizes the full-link intelligence from perception to decision-making, effectively extends the equipment life, reduces the maintenance cost, and is applicable to green crushing operations under complex working conditions.

[0124] Please refer to Figure 2 , Figure 2 an intelligent green crusher blockage clearing system and method provided by the embodiments of this application. The intelligent green crusher blockage clearing system and method include the following modules:

[0125] The acquisition module is deployed in the multi-modal sensor array of the green crusher and is used to acquire the multi-modal perception data of the green crusher; the multi-modal perception data is used to comprehensively monitor the material state inside the crusher; the multi-modal sensor array contains sensors of multiple functional types, and the sensors of multiple functional types are respectively deployed in the key internal structures of the green crusher;

[0126] The establishment module is used to establish a blockage risk model of the green crusher through a spatio-temporal graph neural network based on the multi-modal perception data and the internal structure of the green crusher; the blockage risk level and corresponding blockage risk factors of each key internal structure are marked in the blockage risk model;

[0127] An adjustment module, configured to dynamically adjust the dynamic control instructions of the green waste crusher based on the blockage risk model; the dynamic control instructions at least include: a feeding rate, a scheduling instruction for different particle size crushing units, a temperature supplement parameter, and a humidity supplement parameter;

[0128] An execution module, configured to generate an optimal blockage clearing path by using a graph contrast loss function, and execute a differentiated blockage clearing scheme matching the optimal blockage clearing path based on the dynamic control instructions, so as to preferentially process high-risk nodes in the blockage risk model, realize hierarchical blockage clearing of the green waste crusher, and ensure the normal operation of the green waste crusher.

[0129] In some embodiments, the intelligent green waste crusher blockage clearing system and method can be applied to a terminal device. It should be noted that, for the convenience and conciseness of description, the specific working process of the intelligent green waste crusher blockage clearing system described above can refer to the corresponding process in the foregoing embodiments of the intelligent green waste crusher blockage clearing method, and will not be elaborated here.

[0130] Please refer to Figure 3 , Figure 3 , which is a schematic block diagram of the structure of a terminal device provided by an embodiment of the present application. As Figure 3 shown, the terminal device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected through a bus 303, and this bus is, for example, an I2C bus. Specifically, the processor 301 is used to provide computing and control capabilities to support the operation of the entire terminal device. The processor 301 can be a central processing unit, and the processor 301 can also be other general-purpose processors, digital signal processors, application-specific integrated circuits, field programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or this processor can also be any conventional processor, etc. Specifically, the memory 302 can be a Flash chip, a read-only memory disk, an optical disk, a USB flash drive or a mobile hard disk, etc.

[0131] Those skilled in the art can understand that Figure 3 the structure shown in is only a block diagram of some structures related to the solution of the embodiment of the present application, and does not constitute a limitation on the terminal device to which the solution of the embodiment of the present application is applied. Specifically, the server may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements. Among them, the processor is used to run the computer program stored in the memory, and when executing the computer program, implement any one of the intelligent green waste crusher blockage clearing methods provided by the embodiments of the present application. It should be noted that those skilled in the art can clearly understand that, for the convenience and conciseness of description, the specific working process of the terminal device described above can refer to the foregoing embodiments of the intelligent green waste crusher blockage clearing method, and will not be elaborated here.

Claims

1. A method for clearing blockages in an intelligent green crusher, characterized in that, Including: Collecting multi-modal perception data of the green crusher through a multi-modal sensor array deployed in the green crusher; The multi-modal perception data is used to comprehensively monitor the material state inside the crusher; The multi-modal sensor array contains sensors of multiple functional types, and the sensors of multiple functional types are respectively deployed in the key internal structures of the green crusher; Based on the multi-modal perception data and the internal structure of the green crusher, establishing a blockage risk model of the green crusher through a spatio-temporal graph neural network; The blockage risk levels and corresponding blockage risk factors of each key internal structure are marked in the blockage risk model; Dynamically adjusting the dynamic control instructions of the green crusher based on the blockage risk model; The dynamic control instructions at least include: feeding rate, scheduling instructions for different particle size crushing units, temperature supplement parameters, and humidity supplement parameters; Generating an optimal blockage clearing path using a graph contrast loss function, and executing a differentiated blockage clearing plan matching the optimal blockage clearing path based on the dynamic control instructions to preferentially process high-risk nodes in the blockage risk model, realizing hierarchical blockage clearing of the green crusher and ensuring the normal operation of the green crusher.

2. The method according to claim 1, characterized in that, The collecting multi-modal perception data of the green crusher through a multi-modal sensor array deployed in the green crusher includes: Real-time detecting the moisture content of the material through humidity sensors deployed at the feeding port and the inner wall of the crushing chamber; Dynamically scanning the particle size distribution of the material through a laser particle size detection module integrated at the end of the feeding conveyor belt; Monitoring abnormal stress fluctuations through pressure sensors and vibration sensors deployed on the crushing wall and the bottom layer of the screen in the crushing chamber; Among them, the warning threshold of abnormal stress fluctuations is set to 1.2 times the upper limit of stress fluctuations under normal working conditions; Scanning the surface temperature distribution of the screen through an infrared thermal imaging module deployed at the edge of the material screen to identify local overheating areas caused by adhesives; Through edge computing, synchronizing the time of different modal sensor data collected, adaptively adjusting the respective fusion weights of different modal sensor data according to the spatio-temporal heterogeneous characteristics of the sensors, fusing and extracting features from different modal sensor data to obtain multiple perception feature data, so as to construct the multi-modal perception data.

3. The method according to claim 2, wherein Before the time synchronization of different modal sensor data collected through edge computing, it further includes: Identifying the anti-interference modes corresponding to different modal sensor data according to different sensor device attributes; Among them, the pressure sensor adopts a notch filtering mode; the vibration sensor adopts an adaptive band-pass filtering mode; the temperature sensor and the humidity sensor adopt a cross-validation mechanism and an environmental compensation mode; the laser particle size sensor adopts a spatio-temporal domain joint calibration mode; Using the Kalman filter algorithm to eliminate sensor noise in different modal sensor data according to the anti-interference mode; Using historical trend prediction to supplement uncollected different modal sensor data.

4. The method according to claim 2, characterized in that, After adaptively adjusting the respective fusion weights of different modal sensor data according to the spatio-temporal heterogeneous characteristics of the sensors, it further includes: Performing cross-modal feature alignment on the semantic information in different modal sensor data; Adjust the feature distance of sensor data in similar blockage patterns through a contrastive learning loss function, and add the adjusted feature distance of sensor data to the multi-modal perception data to enhance the recognition accuracy of the implicit semantic information in different-modal sensor data.

5. The method according to claim 1, wherein Based on the multi-modal perception data and the internal structure of the green waste crusher, establish a blockage risk model of the green waste crusher through a spatio-temporal graph neural network, including: Construct a three-dimensional structure model of the green waste crusher based on its internal structure; Map each perception feature data in the multi-modal perception data to the three-dimensional structure model based on the acquisition time and the physical position information of the corresponding internal structure; Take each key internal structure in the three-dimensional structure model as a corresponding graph node and set corresponding connection edges, where the key internal structures at least include: the material inlet and outlet, the material crushing chamber, the screen, and the impact plate of the green waste crusher; Construct an adjacency matrix based on the material flow path and mechanical connection relationship to obtain the edge weights corresponding to each connection edge; where the edge weights are used to indicate the blockage propagation probability between the connected graph nodes; Adopt contrastive learning to extract spatio-temporally associated positive sample pairs from historical blockage events, and randomly select negative sample pairs without historical association, and optimize the contrastive loss function through the positive and negative sample pairs to train the risk prediction model; Update the node embedding features of each graph node in the three-dimensional structure model through incremental graph convolution to adapt to the change of material properties; Adopt the trained risk prediction model to identify the congestion risk and predict the congestion degree in the three-dimensional structure model updated in real time, and obtain the blockage risk level and the corresponding blockage risk factors of each graph node in the three-dimensional structure model, so as to construct the finally output blockage risk model.

6. The method according to claim 5, wherein After constructing the finally output blockage risk model, it further includes: Generate a blockage risk heat distribution model matching the blockage risk model through digital twin technology, and the blockage risk heat distribution model is used to quantify the regional congestion risk values and the corresponding potential congestion causes corresponding to each graph node and each connection edge.

7. The method according to claim 5, wherein After adopting the trained risk prediction model to identify the congestion risk and predict the congestion degree of the multi-modal perception data, it further includes: Adopt a structural causal model SCM to identify the key blockage factors of the current green waste crusher in the blockage risk heat distribution information and transmit them to the cloud for sharing with other green waste crushers; Based on the key blockage factors shared by multiple green waste crushers in the cloud, verify the causal relationships of each key blockage factor under different working conditions and different equipment through counterfactual reasoning, and apply the verification results to the dynamic parameter optimization of the risk prediction model to avoid the pseudo-correlation interference of a single device.

8. The method according to claim 1, wherein Based on the blockage risk model, dynamically adjust the dynamic control instructions of the green waste crusher, including: Through the spatiotemporal attention mechanism, the dilated causal convolution is used to obtain the congestion risk feature matrix at each moment in the congestion risk model; the congestion risk feature matrix at least includes: the congestion risk features of each node, and the cross-node congestion risk association features between multiple nodes; the cross-node congestion risk association features at least include: material particle size features, feed rate features, and vibration frequency features; For the graph nodes associated with the material granularity characteristics in the congestion risk characteristic matrix, a hierarchical control strategy is adopted to identify the current granularity control mode and generate corresponding hierarchical granularity control parameters; For the graph nodes associated with the feed rate characteristics in the congestion risk characteristic matrix, identify whether the feed rate reaches the rate interval corresponding to the current congestion risk level, and output the corresponding adjustment instruction, linking the frequency converter to dynamically adjust the belt conveyor speed to reduce the feed rate under the high congestion risk level; For the graph nodes associated with the vibration frequency characteristics in the congestion risk characteristic matrix, the current anti-adhesion vibration mode is identified, a corresponding vibration motor control instruction is generated, and the vibration motor power is adaptively adjusted; For the graph nodes associated with the temperature characteristics in the congestion risk characteristic matrix, generating temperature supplementary parameters matching the internal structures under the current working conditions, and starting the corresponding temperature compensation operation; For the graph nodes associated with humidity features in the congestion risk feature matrix, humidity supplementary parameters matching each internal structure under the current working condition are generated, and corresponding humidity compensation operations are started.

9. The method according to claim 1, characterized in that, After dynamically adjusting the dynamic control instructions of the greening crusher based on the blocking risk model, the method further includes: Connecting multiple upstream and downstream greening crushers through the OPC-UA protocol, and online simulating the linkage control effect of multiple greening crushers under the dynamic control instruction based on the blockage risk model; The linkage control effect is evaluated through a global benefit function, and the material flow between multiple greening crushers is dynamically allocated to optimize the dynamic control instructions in real time; The optimized dynamic control instructions are sent to multiple greening crushers through the OPC-UA protocol.

10. An intelligent green crusher blockage clearing system, characterized in that, The system comprises: The acquisition module is deployed in the multimodal sensor array of the greening crusher, and is used to acquire multimodal sensing data of the greening crusher; the multimodal sensing data is used to comprehensively monitor the material state inside the crusher; the multimodal sensor array includes sensors of various functional types, and the sensors of various functional types are respectively deployed in the key internal structures of the greening crusher; An establishment module is used to establish a congestion risk model of the greening crusher through a spatiotemporal graph neural network based on the multimodal perception data and the internal structure of the greening crusher; the congestion risk model is annotated with the congestion risk level of each key internal structure and the corresponding congestion risk factor; An adjustment module is used to dynamically adjust the dynamic control instructions of the greening crusher based on the blocking risk model; the dynamic control instructions at least include: feed rate, scheduling instructions for crushing units of different particle sizes, temperature supplementary parameters, and humidity supplementary parameters; An execution module, configured to generate an optimal blockage removal path by using a graph contrastive loss function, and execute a differentiated blockage removal scheme matching the optimal blockage removal path based on the dynamic control instruction, so as to preferentially process high-risk nodes in the blockage risk model, realize hierarchical blockage removal of the green crusher, and ensure the normal operation of the green crusher.

Images