Waste household appliance disassembly material flow direction tracking control system

Abstract

The invention relates to the technical field of solid waste resourceful treatment, and discloses a waste household appliance disassembly material flow direction tracking control system which comprises a variable-frequency feeding conveying device, an industrial crushing device, a multi-stage sorting execution device, a visual acquisition device and an edge computing node. The variable-frequency feeding conveying device is based on the material crushing resistance complexity of visual identification, the feeding speed is adjusted in a self-adaptive mode so as to enlarge the physical distance, and crushing load aliasing is prevented; the load characteristic monitoring device collects the current waveform of the main motor, and performs cross-correlation analysis with the visual data to lock the physical anchor point time so as to realize true value verification; and the edge computing node deduces the time when the material reaches the downstream based on the physical anchor point and the conveying speed, and carries out prospective parameter compensation on the multi-stage sorting execution device. Multi-dimensional sensing and flow direction tracking are achieved through cloud-side cooperation, the problems of equipment overload and separation lag caused by non-standard materials are solved, and the separation purity and operation stability of a disassembly production line are improved.

Description

Technical Field

[0001] This invention relates to the field of solid waste resource utilization technology, specifically to a tracking and control system for the flow of materials from the dismantling of waste household appliances. Background Technology

[0002] The resource recovery of waste household appliances is an important component of the circular economy. Its core lies in the efficient separation and recycling of mixed components through mechanical crushing and physical sorting. In actual industrial production lines, the materials to be processed often exhibit highly non-standard characteristics. Different brands, models, and eras of household appliances differ in structural dimensions, material composition, and assembly processes. This complexity and uncertainty in material sources often makes it difficult for traditional automated production lines to simultaneously achieve both processing efficiency and operational stability when faced with continuously changing operating conditions.

[0003] Existing waste appliance crushing lines mostly employ open-loop or simple PID feedback-based control modes. In the feeding stage, the conveying device typically operates at a constant speed or based on a simple current threshold. However, waste appliances often contain high-density, high-hardness components such as compressors and counterweights. When these high-resistance materials continuously enter the crushing chamber at short intervals, a cumulative load effect can easily form on the mechanical spindle, causing instantaneous torque to exceed the limits of the motor or hydraulic system, leading to equipment overload shutdown or fatigue damage to mechanical components. Traditional control logic struggles to predict the mechanical impact before the material enters the crusher and lacks a feedforward intervention mechanism based on material characteristics. It can only respond passively after an overload occurs, which not only reduces production efficiency but also increases equipment maintenance costs.

[0004] Furthermore, although machine vision technology has been introduced into dismantling production lines to assist in material identification, a spatiotemporal misalignment exists between visual perception data and physical crushing response in high-speed production environments. Due to interference from physical factors such as conveyor belt slippage, material tumbling, and material drop lag, it is difficult to achieve a precise correspondence between the moment the visual sensor captures the image and the moment the material actually contacts the cutter roller. This temporal asynchrony prevents the control system from accurately establishing a mapping relationship between appearance features and crushing resistance, making it difficult to translate visual data into effective control commands. Even worse, visual perception alone is insufficient to identify hidden hard materials encased in a shell. This visual deception often leads the control system to use an incorrect low-power mode to handle high-hardness materials, resulting in incomplete crushing or equipment jamming.

[0005] In the downstream sorting stage, equipment control also faces the challenge of response lag. There is a significant physical delay between receiving commands and completing actions in actuators such as hydraulic adjustment mechanisms and pneumatic valves. Traditional control methods often trigger adjustment actions only when the material arrives at the sorting equipment, causing equipment actions to always lag behind changes in the material flow. For example, if the crushing gap fails to open in advance when large materials pass through, it will cause blockage at the discharge port; conversely, if the gap fails to tighten in time when small materials pass through, it will reduce the degree of dissociation. Simultaneously, as the equipment operates over time, physical aging phenomena such as hammer wear and liner consumption lead to a decline in equipment performance. Existing control models are mostly static parameters, unable to perceive and compensate for these dynamic changes in equipment status, resulting in a gradual decrease in control accuracy and sorting purity after long-term operation. Therefore, building a control system capable of achieving end-to-end material tracking, precise spatiotemporal alignment, and adaptive compensation for equipment status has become a key requirement for improving the intelligence level of waste appliance processing. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a material flow tracking and control system for the dismantling of waste household appliances. This system solves the problem that the material flow sequence is difficult to track accurately due to the crushing and buffering processes in existing waste household appliance dismantling lines, which in turn leads to lag in cross-process equipment collaborative control and low sorting efficiency.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a material flow tracking and control system for dismantling waste household appliances. This system includes a variable frequency feeding conveyor, an industrial crushing device, a discharge conveyor, an intermediate buffer silo, and a multi-stage sorting actuator, all connected sequentially along the process route. The system also includes edge computing nodes that are communicatively connected to each device.

[0008] This invention performs cross-process tracking control through edge computing nodes. Its core lies in constructing a mapping relationship between virtual material data objects and physical material entities. Virtual material data objects are generated using image data acquired by an infeed vision acquisition device, and visual predictive load characteristics are established. Subsequently, a spatiotemporal alignment algorithm is used to match the visual predictive load characteristics with the physical response load characteristics acquired by a load characteristic monitoring device, thereby pinpointing the physical anchor time of the virtual material data object's crushing behavior within the industrial crushing device. Using this physical anchor time as a timing benchmark, combined with the dynamic retention characteristics of the intermediate buffer silo, the precise arrival time stamp of the virtual material data object's flow to the multi-stage sorting execution device is deduced. When the system time reaches this precise arrival time stamp, synchronous control commands are output based on the characteristics of the virtual material data object, coordinating the adjustment of the multi-stage sorting execution device, the variable frequency feeding conveyor, or the industrial crushing device.

[0009] Preferably, the system employs a dynamic volume substitution algorithm for calculating the residence time of materials in the intermediate buffer silo. Edge computing nodes calculate the transmission delay of materials to the intermediate buffer silo based on the physical anchor point time and the speed integral of the discharge conveyor; simultaneously, they acquire the real-time material level height monitored by the silo monitoring component and the volumetric discharge rate of the quantitative feeding device. The calculation of the dynamic residence time of materials in the intermediate buffer silo follows the volume substitution principle, i.e., the material volume corresponding to the real-time material level height divided by the volumetric discharge rate. Finally, the formula for generating the accurate arrival timestamp is: physical anchor point time superimposed with transmission delay, dynamic residence time, and a preset chute free-fall compensation time.

[0010] Preferably, the system performs adaptive feeding control through a visual monitoring device for material discharge. Edge computing nodes calculate the crushing resistance complexity index of virtual material data objects and predict the crushing and emptying time based on a historical mapping model. Combining the current instantaneous conveying speed of the variable frequency feeding conveyor with the predicted crushing and emptying time, the system calculates the safe signal separation distance. When the crushing resistance complexity index of two adjacent virtual material data objects both exceed the high-load warning threshold, and their actual physical distance is less than the safe signal separation distance, the system generates a feedforward control command containing deceleration parameters to adjust the variable frequency feeding conveyor to increase the material spacing and prevent crusher overload.

[0011] Preferably, regarding the locking logic for the physical anchor point time, the edge computing node constructs a visually predicted load sequence based on the material projection area and material hardness in the image data, and uses an energy operator to process real-time load data to lock the load event boundaries, thus constructing a physical response load sequence. The system performs sliding window normalized cross-correlation analysis on the visually predicted load sequence and the physical response load sequence, locking the time corresponding to the maximum value of the normalized cross-correlation coefficient as the physical anchor point time. If this maximum value (i.e., the physical anchor point confidence) exceeds a preset threshold, the load feature fingerprint vector in the physical response load sequence is extracted and written as a truth label into the virtual material data object, realizing the actual measurement correction of the data.

[0012] Preferably, the system has a visual perception deviation correction function. It calculates the relative difference between the physical response load sequence and the visual prediction load sequence within a time window, outputting a visual perception deviation index. When this index exceeds a preset safety deviation range, a latent hardness anomaly is identified, and a dynamic control compensation vector is generated to correct the control parameters of the industrial crushing device. Simultaneously, it calculates the ratio of the cumulative value of the physical response load sequence to the projected area of ​​the material pixels in the image data to obtain the physical inversion hardness coefficient. This coefficient is then used to update the model parameters used to calculate the characteristics of the virtual material data object online, achieving adaptive iteration of the model.

[0013] Preferably, a flow feedback detection device is installed at the end of the multi-stage sorting execution device. A defect feature matrix is ​​constructed based on the collected image data. This matrix includes a morphological entanglement index characterizing the risk of geometric entanglement of materials and a material mixing index characterizing the degree of material dissociation. Based on the statistical distribution of the defect feature matrix, feedback adjustment signals are generated for either the conveying speed of the variable frequency feeding conveyor or the discharge frequency of the intermediate buffer silo, forming a closed-loop control.

[0014] Preferably, the system incorporates a cloud server for full lifecycle management of the equipment. The cloud server calculates the equipment performance degradation index based on the historical cumulative load, instantaneous active power, and vibration acceleration of the industrial crushing device. A joint input vector containing the statistical values ​​of virtual material data object characteristics and the equipment performance degradation index is constructed, generating a dynamic state vector that includes wear compensation for the actuators. This vector is then distributed to edge computing nodes to correct the execution parameters in the synchronous control commands and compensate for control errors caused by equipment aging.

[0015] Preferably, when generating control commands, the system performs an actuator response time delay estimation. Using a dynamic model, the estimated action time required to adjust to the target execution parameters is calculated based on the real-time status and drive characteristics of the multi-level sorting actuator. The look-ahead control trigger timestamp is calculated as: the accurate arrival timestamp minus the estimated action time and a preset safety margin. The system uses high-frequency clock scanning to ensure accurate transmission of control commands upon arrival of the look-ahead control trigger timestamp.

[0016] Preferably, the system implements a parameter conflict adjustment strategy. A normalized comprehensive load utilization rate is calculated based on the real-time power and vibration intensity of the multi-level sorting actuators. When the target execution parameters cause the comprehensive load utilization rate to enter a preset buffer range, a nonlinear penalty function is used to exponentially correct the target execution parameters based on the difference between the comprehensive load utilization rate and the safety benchmark threshold. A dynamic slope limiting algorithm is then applied to the adjusted parameters to limit the actuator's movement rate and prevent mechanical shock.

[0017] Preferably, the system has an instruction fusion processing mechanism. It detects the precise arrival timestamp interval between two adjacent virtual material data objects in the control queue. If this time interval is less than the adjustment dead zone time of the multi-level sorting actuator, it obtains the normalized hardness coefficient and physical anchor point confidence level of both. Using the physical anchor point confidence level as a weight, it performs a weighted calculation on the normalized hardness coefficient to obtain the fused control hardness coefficient, thereby generating a single integrated control instruction to avoid oscillations in the actuator under high-frequency instruction switching.

[0018] This invention provides a material flow tracking and control system for dismantling waste household appliances. It has the following beneficial effects:

[0019] 1. This invention solves the problem of accurately locking the material flow status during the dismantling of waste household appliances by constructing a spatiotemporal alignment mechanism between virtual material data objects and physical response load characteristics. The system uses sliding window normalized cross-correlation analysis to lock the physical anchor point time within the industrial crushing device, and calculates the dynamic residence time by combining the volume replacement principle of the intermediate buffer hopper, thereby deriving the precise arrival timestamp at the multi-stage sorting execution device. This mechanism eliminates the temporal uncertainty caused by the randomness of the crushing process and the nonlinear time delay of the buffering process, realizing precise synchronous control across processes and improving the timing matching accuracy and sorting purity of subsequent sorting actions.

[0020] 2. This invention establishes a load balancing system based on visual perception deviation correction and adaptive feed control. By calculating the physical inversion hardness coefficient and updating model parameters online, it corrects potential hidden anomalies (such as internal high-hardness components) that might exist when judging material hardness solely based on visual images. Simultaneously, based on the crushing resistance complexity index and predicted crushing and emptying time, the safe signal separation distance is calculated. When insufficient spacing between adjacent high-load materials is detected, a feedforward control command containing deceleration parameters is generated to adjust the variable frequency feed conveyor. This multi-dimensional load management strategy effectively mitigates instantaneous impact loads on industrial crushing equipment, preventing equipment overload shutdowns caused by material accumulation or hidden hard objects.

[0021] 3. This invention optimizes the response logic and command scheduling strategy of the multi-level sorting actuator, improving the robustness of the control system. The system compensates for the inherent physical delay of mechanical drives by introducing actuator response delay estimation and calculating the look-ahead control trigger timestamp. Furthermore, for materials arriving at high frequencies continuously, the system employs command fusion processing based on physical anchor confidence, using weighted calculations to generate a single comprehensive control command, and combining this with a nonlinear penalty function to dynamically limit the slope of conflict parameters. These measures effectively prevent mechanical oscillations and overshoot when the actuator receives dense or conflicting commands, extending the service life of key actuators while ensuring sorting efficiency. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the waste household appliance dismantling material flow tracking control system architecture of the present invention;

[0023] Figure 2 This is a flowchart illustrating the waste household appliance dismantling material flow tracking and control method of the present invention.

[0024] Figure 3 This is a schematic diagram of the adaptive feed control strategy of the present invention;

[0025] Figure 4 This is a schematic diagram of the physical anchoring principle based on cross-correlation analysis of the present invention;

[0026] Figure 5 This is a schematic diagram of the value priority arbitration envelope of the present invention.

[0027] Among them, 110 is a variable frequency feeding conveyor; 120 is an industrial crushing device; 130 is a discharge conveyor; 131 is a discharge visual monitoring device; 140 is an intermediate buffer silo; 141 is a silo level monitoring component; 142 is a quantitative feeding device; 210 is an infeed visual acquisition device; 220 is a load characteristic monitoring device; 230 is a flow feedback detection device; 300 is a multi-stage sorting execution device; 310 is an adjustable eddy current sorting device; 320 is a magnetic sorting device; 330 is a wind-powered sorting device; 410 is an edge computing node; and 420 is a cloud server. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See attached document Figure 1 This invention provides a material flow tracking and control system for dismantling waste home appliances. The system integrates physical transmission, crushing and processing, multi-dimensional sensing and hierarchical control functions to achieve refined processing and flow tracking of non-standard waste home appliances.

[0030] The waste household appliance dismantling material flow tracking control system includes a variable frequency feeding conveyor 110. The variable frequency feeding conveyor 110 is configured to carry and convey individual waste household appliances to be processed. A variable frequency drive controller is connected to the drive end of the variable frequency feeding conveyor 110. This variable frequency drive controller is communicatively connected to a control network and is used to change the linear speed of the conveyor belt according to received speed adjustment commands, thereby regulating the time interval and instantaneous feed throughput of adjacent waste household appliance units entering the industrial crushing device 120.

[0031] The output end of the variable frequency feeding conveyor 110 is connected to the feed inlet of the industrial crushing device 120. The industrial crushing device 120 is configured to crush waste household appliances into a mixed material stream by mechanical shearing or hammering. The industrial crushing device 120 includes a main drive motor, which is connected to an industrial power supply via a power circuit. A discharge conveyor 130 is provided below the discharge gap of the industrial crushing device 120, and the discharge conveyor 130 extends along a predetermined process route to carry the crushed mixed material stream.

[0032] A discharge visual monitoring device 131 is installed above the discharge conveying device 130 to collect the morphological characteristics of the crushed material. The output end of the discharge conveying device 130 is connected to the intermediate buffer silo 140, which is used to temporarily store materials to balance the flow rates of the upstream and downstream. At the top of the intermediate buffer silo 140, a silo level monitoring component 141 (e.g., an ultrasonic sensor or radar sensor) is installed to monitor the height of the material in the silo in real time. At the bottom outlet of the intermediate buffer silo 140, a quantitative feeding device 142 (e.g., a vibrating feeder or a belt feeder) is installed, through which the material is uniformly conveyed to the multi-stage sorting execution device 300.

[0033] A feed vision acquisition device 210 is installed above the variable frequency feeding conveyor 110. The feed vision acquisition device 210 is located upstream of the industrial crushing device 120. The feed vision acquisition device 210 includes a high-resolution industrial image sensor and a depth sensing component, used to acquire the appearance texture data and three-dimensional geometric data of individual waste household appliances before they enter the industrial crushing device 120. The data output port of the feed vision acquisition device 210 is connected to an edge computing node 410.

[0034] A load characteristic monitoring device 220 is connected to the power circuit of the main drive motor of the industrial crushing device 120. The load characteristic monitoring device 220 includes a high-frequency current transformer and a voltage sampling unit, and is configured to collect the three-phase current waveform data and voltage waveform data of the main drive motor in real time at a preset high sampling rate. The load characteristic monitoring device 220 is configured to capture the transient load fluctuation signals generated by the industrial crushing device 120 when crushing parts of different materials, and transmit the digitized load signals to the edge computing node 410.

[0035] A multi-stage sorting actuator 300 is arranged sequentially along the conveying direction of the discharge conveyor 130. The multi-stage sorting actuator 300 includes an adjustable eddy current separator 310. The adjustable eddy current separator 310 is used to separate non-ferrous metal components from the mixed material stream. The adjustable eddy current separator 310 is equipped with a servo control unit, which is used to adjust process parameters such as the magnetic roller speed and belt speed. The multi-stage sorting actuator 300 may also include a magnetic separator 320 and a pneumatic separator 330, used to separate ferromagnetic materials and lightweight materials, respectively. The parameter control interfaces of each sorting device are connected to a control network.

[0036] A flow feedback detection device 230 is installed at each material output end of the multi-stage sorting execution device 300. The flow feedback detection device 230 is configured to monitor the flow rate and optical characteristics of the sorted materials in real time to generate actual output data.

[0037] The waste appliance dismantling material flow tracking and control system also includes an edge computing node 410 and a cloud server 420. The edge computing node 410 establishes bidirectional communication connections with the variable frequency feeding conveyor 110, industrial crushing device 120, feed vision acquisition device 210, load characteristic monitoring device 220, multi-level sorting execution device 300, and flow feedback detection device 230 via industrial Ethernet or fieldbus. The edge computing node 410 is configured to perform real-time data acquisition, preprocessing, and control command generation and distribution.

[0038] The cloud server 420 is connected to the edge computing node 410 via a wide area communication network. The cloud server 420 contains a standard bill of materials database and a crushing efficiency mapping model database. The cloud server 420 is configured to provide model parameters and standard data support to the edge computing node 410, and to receive production statistics uploaded by the edge computing node 410 for iterative model updates.

[0039] See attached document Figure 2 This invention provides a method for tracking and controlling the flow of materials from the dismantling of waste household appliances. This method is based on the aforementioned material tracking and control system for the dismantling of waste household appliances. The method for tracking and controlling the flow of materials from the dismantling of waste household appliances may include the following steps:

[0040] In step S100, the edge computing node 410 performs adaptive discretized feeding control based on visual complexity. When a single waste appliance enters the detection area of ​​the feeding visual acquisition device 210, the device acquires image data of the appliance and transmits it to the edge computing node 410. The edge computing node 410 calculates the crushing resistance complexity index of the appliance based on the image data. Simultaneously, the edge computing node 410 monitors the physical distance between the current and subsequent waste appliances on the variable frequency feeding conveyor 110.

[0041] Edge computing node 410 executes dual threshold discrimination logic: if the crushing resistance complexity index of two adjacent waste household appliances is detected... Both exceeded the high load warning threshold, and the real-time physical distance between them on the conveyor belt ( The distance is less than the calculated safe signal separation distance. If the system immediately triggers the anti-aliasing active intervention mechanism, the edge computing node 410 generates a deceleration control command with a slope limit and sends it to the variable frequency drive controller to drive the variable frequency feeding conveyor 110 to perform controlled deceleration, thereby widening the time window for entering the industrial crushing device 120 and ensuring that the load waveforms generated by adjacent materials are effectively separated in the time domain.

[0042] In step S200, the edge computing node 410 and the cloud server 420 collaboratively perform residual value feature extraction and feedforward prediction. During the process of transferring the waste household appliance unit from the frequency conversion feeding conveyor 110 to the industrial crushing device 120, the edge computing node 410 extracts the initial appearance feature data (such as shell damage, degree of rust, etc.) of the waste household appliance unit based on the image data.

[0043] Edge computing node 410 requests corresponding standard bill of materials data and crushing efficiency mapping model parameters from cloud server 420 based on the identification information of the waste household appliance. Edge computing node 410 combines the standard bill of materials data, surface defect feature data, and crushing efficiency mapping model parameters to calculate and generate the dynamic expected material vector for the waste household appliance. Simultaneously, edge computing node 410 calculates the theoretical time window for the waste household appliance to enter the industrial crushing device 120 based on the current transmission speed and distance.

[0044] In step S300, edge computing node 410 performs event anchoring and truth verification based on load graph. When the theoretical time window is reached, load characteristic monitoring device 220 collects real-time load current waveform data of the main drive motor of industrial crushing device 120 and transmits it to edge computing node 410. Edge computing node 410 performs cross-correlation matching analysis between the real-time load current waveform data and the preset component characteristic load fingerprint.

[0045] When the matching correlation exceeds a preset confidence threshold, the edge computing node 410 marks the matching moment as the physical anchor time and confirms the validity of the dynamic expected material vector. When no matching characteristic load waveform is detected within the theoretical time window, the edge computing node 410 determines that visual deception has occurred, corrects the dynamic expected material vector by setting the material content value of the corresponding missing component to zero, and generates a corrected final control vector.

[0046] In step S400, the edge computing node 410 performs downstream device adaptive control based on spatiotemporal alignment. The edge computing node 410 calculates the estimated arrival time of the crushed material at each sorting unit in the downstream multi-stage sorting execution device 300 based on the physical anchor time, the average residence time in the industrial crushing device 120, and the operating speed of the discharge conveying device 130.

[0047] The edge computing node 410 generates operating parameter adjustment instructions for each sorting unit based on the material composition distribution data in the final control vector. The edge computing node 410 sends the operating parameter adjustment instructions to the corresponding adjustable eddy current separator 310, magnetic separator 320, or wind separator 330 at a preset lead time before the expected arrival time, so that the sorting equipment completes the switching of operating status before the material arrives.

[0048] In step S500, the system performs closed-loop feedback and model update. The flow feedback detection device 230 collects actual material throughput data at the sorting output end. The edge computing node 410 calculates the deviation between the actual material throughput data and the final control vector. The edge computing node 410 uploads this deviation value, along with the corresponding original image data and load waveform data, to the cloud server 420. The cloud server 420 uses this feedback data to train and update the crushing efficiency mapping model, and sends the updated model parameters to the edge computing node 410 for subsequent control processes. After completing the closed-loop feedback data upload, the system explicitly calls the destructor function to release the memory resources occupied by the virtual material data object, preventing memory overflow due to object accumulation during long-term continuous operation and ensuring system stability.

[0049] During the adaptive discretized feed control in step S100, the edge computing node 410 needs to perform quantitative evaluation of non-standard materials. The core of this embodiment lies in establishing a mapping relationship between visual appearance features and mechanical crushing load, and predicting the transient impact of materials on the crusher's main shaft through weighted fusion of multi-physics field features.

[0050] Edge computing node 410 performs geometric feature extraction. The raw depth point cloud data acquired by the feed vision acquisition device 210 typically contains background noise and redundant information from the conveyor belt plane. Edge computing node 410 uses a voxel mesh-based downsampling filtering algorithm to clean the point cloud and uses a random sampling consensus algorithm to fit a planar model to remove the conveyor belt background. For the extracted target object point cloud, the system calculates its minimum bounding box. In this embodiment, an oriented bounding box is used for calculation instead of an axis-aligned bounding box because an oriented bounding box can be constructed along the principal axis of inertia of the object, thus more compactly representing the actual physical contour of the object in any posture.

[0051] After obtaining the bounding box volume, the system performs a normalization mapping on it to generate volume feature values. Specifically, edge computing node 410 calculates the ratio of the bounding box volume to the maximum theoretical throughput volume of the crusher, and truncates this ratio to an upper limit (i.e., not exceeding 1.0) to prevent abnormal data overflow. The physical meaning of the volumetric characteristic value lies in characterizing the basic requirement of the material's macroscopic dimensions for crushing shear torque.

[0052] Building upon this, edge computing node 410 extracts material-dimensional features based on a deep learning model. For the input 2D RGB image, this embodiment employs a semantic segmentation network with an encoder-decoder structure (e.g., based on U-Net or DeepLabV3+ architecture) for pixel-level classification. This semantic segmentation network specifically includes the following modules:

[0053] Encoder module: Used to extract multi-scale high-dimensional features of images, and can use ResNet or EfficientNet as the backbone network for downsampling feature extraction;

[0054] The decoder module upsamples the feature maps to restore them to the original image resolution and outputs a class probability map for each pixel. The network's output layer uses the Softmax function to classify pixels into high-resistance metal areas (corresponding to compressors and motors), medium-resistance structural areas, and low-resistance background areas. During model training, a combination of weighted cross-entropy loss and Dice loss is used to address the sample imbalance problem caused by the small proportion of high-value metal components in the image.

[0055] Based on the segmentation results, the system counts the number of metal pixels and the total number of pixels, and defines the ratio of the two as the metal distribution density value. The calculation includes zero-prevention logic; if the total number of pixels approaches zero, the metal distribution density value is forced to be zero. This indicator directly reflects the distribution probability of high-hardness materials in the fracture section.

[0056] Furthermore, the system combines geometric information with prior knowledge to construct structural stiffness characteristics. Edge computing node 410 queries the basic stiffness coefficient of the appliance category (e.g., drum washing machine) from the local attribute database based on the visually recognized appliance category ID. Simultaneously, it calculates the space fill rate of the point cloud data within the minimum bounding box, i.e., the ratio of the effective point cloud volume to the bounding box volume. The space fill rate objectively reflects whether the object is a solid block or a thin-walled hollow shell, directly affecting the engagement depth of the crushing tool. Based on this, edge computing node 410 defines the product of the basic stiffness coefficient and the space fill rate as the structural integrity coefficient. This design ensures that for the same type of home appliance, the calculated structural integrity coefficient of the undisassembled unit will be higher than that of the disassembled unit due to its higher space filling rate, thus accurately reflecting the difference in the difficulty of breakage.

[0057] Finally, edge computing node 410 synthesizes the final index based on a multivariate linear regression model. Substituting the aforementioned feature components into the following weighted equation, the fracture resistance complexity index is calculated. :

[0058] ;

[0059] in, This represents the complexity index of breaking resistance. Represents volume characteristic value, Indicates the metal distribution density value. Represents the structural integrity coefficient. Represents the volume weighting factor. Indicates the material weighting factor. Represents the structural weighting factor. This represents the bias constant, which is used as a reference value to compensate for fluctuations in the system's no-load current.

[0060] In this preferred embodiment, the material weighting factor Structural weighting factor With volume weighting factor satisfy The relationship between the magnitudes is as follows. The physical basis for this weighting is that, in actual crushing conditions, the hardness of the material itself (such as the steel shell of a compressor) contributes the most to the current peak value, followed by the compactness of the structure, and finally the simple geometric volume. The specific values ​​are obtained by the cloud server based on the multivariate regression analysis of historical current peak data and feature vectors, and the values ​​are usually between [0.1, 0.8]. Through the above calculations, the system transforms unstructured visual data into a dimensionless scalar—that is, the crushing resistance complexity index. This index not only reflects the physical properties of the object, but also directly relates to the expected load of the crusher, providing a quantitative basis for determining whether to increase the feed spacing in the subsequent step S100.

[0061] Edge computing node 410 is configured to perform adaptive feed control, and the crushing resistance complexity index of the current waste household appliance unit is obtained at edge computing node 410. Subsequently, the system enters the active intervention phase of the feed rate. The technical objective of this embodiment is to solve the load cascading effect, that is, to prevent high-resistance materials from entering the crushing chamber before subsequent materials have completely passed through, which would cause the torque of the crusher's main shaft to accumulate and lead to overload shutdown. To this end, the system utilizes the mapping relationship between the time and space dimensions to dynamically adjust the feed spacing through the following steps:

[0062] In step S110, the edge computing node 410 establishes a mapping relationship between the complexity index and the width of the crushing load pulse. The process of the crusher handling different materials is represented by a continuous current pulse in electrical parameters; the width of this pulse directly corresponds to the physical time required for the material to be torn apart and pass through the grid. The edge computing node 410 uses a nonlinear regression model built based on historical operating data to represent the dimensionless crushing resistance complexity index. This is mapped to the predicted crusher emptying time. The prediction aims to define a safe time window to ensure the crusher load current can drop from its peak to the no-load reference value. The calculation formula is as follows:

[0063] ;

[0064] in, This indicates the predicted time for breakage and emptying, in seconds. This represents the minimum mechanical throughput time, a value determined by the rated rotor speed of the industrial crusher 120, the effective volume of the crushing chamber, and the aperture of the discharge grid. It characterizes the theoretically fastest throughput time. This represents the complexity index of the breaking resistance calculated in the preceding steps. This represents the time delay factor, which is derived through regression analysis of the half-wave width of historical broken current waveforms and corresponding visual characteristics. It is used to quantify the impact of complexity on processing time. The nonlinear response index typically ranges from [1.2 to 2.0] and is used to reflect the physical characteristic that the difficulty of crushing increases exponentially but nonlinearly as the complexity of the material increases.

[0065] In step S120, the edge computing node 410 converts the time-domain prediction results into control constraints for the conveying speed. The system monitors in real time the physical distance between subsequent waste appliance units on the variable frequency feeding conveyor 110 and the waste appliance units currently entering the crusher. Based on the current conveying speed, the edge computing node 410 calculates the minimum physical distance required to ensure that subsequent material contacts the cutter roller only after the load waveform generated by the current material has completely subsided, i.e., the safety anti-overlapping distance. Specifically, the system converts the current instantaneous conveying speed of the variable frequency feeding conveyor 110 and the predicted crushing and emptying time... The safe anti-overlapping distance is obtained by multiplying the three factors together, including a preset safety redundancy factor. The reason for setting the safety redundancy factor (usually between 1.1 and 1.3) is to compensate for the positional uncertainty caused by conveyor belt slippage, material rolling, or visual ranging errors.

[0066] In step S130, the edge computing node 410 generates a variable frequency drive control command and executes closed-loop regulation. When the actual physical distance detected by the feed vision acquisition device 210 is less than the calculated safe anti-overlap distance, the edge computing node 410 determines that there is a risk of load overlap. At this time, the system must decelerate the variable frequency feeding conveyor 110 to increase the arrival time difference of subsequent materials. The edge computing node 410 executes proportional deceleration logic based on the distance ratio: calculates the ratio of the actual physical distance to the safe anti-overlap distance, and uses this ratio as an attenuation coefficient to apply to the current drive frequency, thereby generating a target drive frequency. In order to prevent the conveyor belt from stopping immediately due to the target frequency being too low (close to 0), which would cause the waste household appliances entering the crusher to get stuck at the crusher inlet or be backed by the rotor, this step sets a minimum maintenance frequency ( ).

[0067] In this process, the system introduces low-speed clamping protection logic: if the calculated target drive frequency is lower than the minimum critical value required for motor heat dissipation or torque maintenance (i.e., the aforementioned...) The system no longer continues to reduce the frequency, but instead forcibly clamps the target drive frequency to [a certain value]. To maintain a low-speed, constant feeding state on the conveyor belt, if the calculated target frequency is consistently lower than... If the preset time threshold is exceeded (indicating severe congestion), the feed belt will be stopped until the physical spacing is restored. Zero-frequency shutdown will not be easily executed unless the system detects that the spacing of subsequent materials is extremely dangerous (e.g., emergency avoidance is necessary), to ensure that materials currently in the crushing and biting stage can pass smoothly.

[0068] It is worth noting that when issuing frequency commands, the edge computing node 410 does not directly and abruptly modify the frequency, but instead uses an S-shaped speed planning curve for transition. The technical reason for adopting this flexible acceleration and deceleration strategy is that waste household appliances often contain cylindrical objects such as water heaters and fire extinguishers. If the conveyor belt stops or decelerates suddenly, the huge inertia can easily cause these objects to roll off the conveyor belt or become unstable, thus causing blockage at the feed inlet.

[0069] Finally, the edge computing node 410 sends the smoothed frequency command to the frequency converter via an industrial fieldbus (such as Profinet or EtherCAT). The frequency converter adjusts the output torque and magnetic flux of the asynchronous motor to precisely control the arrival time of subsequent highly complex materials. Although a single-stage conveyor belt cannot change the physical distance between materials, by reducing the conveying speed, the system can linearly extend the time required for materials to travel a fixed physical distance, thereby effectively increasing the time interval between adjacent materials entering the crushing chamber. This achieves complete separation of the two load current waveforms on the time axis, reducing the risk of crusher overload caused by excessively fast feeding from the source.

[0070] In the cloud-edge collaborative architecture, to resolve the contradiction between limited uplink bandwidth and massive visual data transmission in industrial settings, while ensuring real-time feedback on the quality of crushed products, edge computing node 410 is configured to perform hierarchical processing tasks of edge feature extraction and cloud-based decision analysis. The core objective of this embodiment is to construct a defect feature matrix. Through feature dimensionality reduction, unstructured image data is transformed into structured data capable of quantifying two key process defects: incomplete dissociation and morphological entanglement. These two defects directly determine the purity upper limit of downstream eddy current sorting and the failure probability of the mechanical transmission system.

[0071] In step S210, the edge computing node 410 performs dynamic segmentation and background suppression of the region of interest (ROI) based on the data collected by the streaming feedback detection device 230. To accurately assess the quality of the broken product, the raw image stream collected by the streaming feedback detection device 230 is mapped to the HSV color space. Compared to the RGB space, the HSV space can separate the luminance and chrominance components, thereby reducing the interference of strong reflections or shadows on the metal surface on target extraction. The edge computing node 410 uses an adaptive Gaussian mixture model (GMM) to dynamically model the conveyor belt background, and achieves foreground segmentation by calculating the Mahalanobis distance between the current frame pixel and the background model. For particles that are physically stacked or adhered, the system uses a distance-transform-based watershed algorithm to calculate the segmentation mask, ensuring that each independent broken particle can be assigned a unique connected component identifier.

[0072] In step S220, the edge computing node 410 calculates the morphological entanglement index (denoted as ) for the segmented individual particles. The shredded waste household appliances often contain two types of high-risk materials: one is extremely thin copper wires or sealing strips, and the other is metal casings with hook-like structures created by extrusion deformation. The former easily becomes entangled in magnetic separator drums, while the latter easily gets stuck in the gaps of vibrating feeders. Therefore, the system performs weighted aggregation calculations based on contour geometric features, with the specific logic as follows:

[0073] First, the aspect ratio term characterizing the slender features is calculated: this term is obtained by dividing the maximum Freret diameter of the particle profile (i.e., the maximum Euclidean distance between any two points on the profile) by the equivalent minor diameter perpendicular to it. To prevent division by zero errors, a small numerical stability factor is introduced in the calculation.

[0074] Secondly, the indentation degree term, which characterizes the snagging feature, is calculated: this term is based on the difference rate between the projected area of ​​the particle and the convex hull area (i.e., the area of ​​the smallest convex polygon enclosing the contour). The greater the difference, the deeper the irregular indentation on the object's surface, and the higher the risk of snagging.

[0075] Finally, the edge computing node 410 performs a weighted sum of the aspect ratio and the concavity term to obtain the morphological entanglement index. The weight of the aspect ratio term is usually set to 0.6 to 0.8, based on the fact that historical fault data statistics show that the number of shutdowns caused by entanglement due to thin cables is greater than the number of times the cable gets stuck due to hooks.

[0076] In step S230, edge computing node 410 calculates the material mixing index (denoted as...). This is used to characterize incomplete dissociation defects. In an ideal state of fragmentation, a single particle should exhibit material homogeneity (e.g., pure copper particles exhibit a uniform purplish-red color, and pure plastic particles exhibit a uniform texture). Conversely, if the particle surface simultaneously exhibits a high-reflectivity area of ​​metal and a diffuse-reflectivity area of ​​plastic, or if the hue distribution shows a multi-peak characteristic, then the particle is determined to be a composite particle of metal and plastic. To quantify this characteristic, the system performs multi-dimensional visual feature fusion calculations: Extracting color dimension features: calculating the hue channel variance of pixels within the particle mask area; the hue channel variance directly reflects the dispersion of color on the object's surface. Extracting texture dimension features: calculating texture contrast based on the gray-level co-occurrence matrix (GLCM). Metal fracture surfaces are typically smooth, while attached insulating layers or colloids are typically rough; the coexistence of both leads to increased texture contrast. The system multiplies the hue channel variance by the texture contrast and performs a logarithmic transformation on the product to obtain the material mixing index. The purpose of using logarithmic transformation is to compress the dynamic range of the data, allowing even small undissociated impurity signals to be numerically amplified.

[0077] In step S240, the edge computing node 410 encapsulates the aforementioned features into a standardized defect feature matrix. For multiple broken particles identified within the current detection time window, the system constructs a feature tensor containing the aforementioned morphological entanglement index, material mixing index, and basic physical dimensions. The edge computing node 410 does not directly upload the original high-definition image data, but instead uploads the structured defect feature matrix only via a lightweight industrial protocol (such as MQTT). This strategy reduces the data transmission volume from megabytes to kilobytes, enabling the cloud server to receive and process the entire field of data with millisecond-level latency. The cloud server generates a feedforward control signal based on the statistical distribution of the defect feature matrix: when the average morphological entanglement index of the entire batch of materials exceeds a preset safety threshold (such as 3.5), the system automatically reduces the frequency of the upstream vibrating feeder, causing the material to spread out in a single layer on the conveyor belt, thereby actively avoiding the risk of entanglement through physical deceleration.

[0078] In the cloud-edge collaborative architecture, the cloud server not only serves as the data aggregation center but also undertakes the task of full lifecycle state modeling across time dimensions. The core technology of this embodiment lies in solving the control deviation problem caused by the time-dependent degradation of equipment performance. The crushing tools (such as hammers or cutter boxes), liners, and screens of the industrial crushing device 120 undergo irreversible physical wear over time, resulting in a gradual decrease in the degree of dissociation of the crushed products under the same feed parameters. Therefore, the system cannot rely solely on instantaneous characteristics uploaded from the edge for static control but must introduce a time-series-based degradation model to generate dynamic state vectors, thereby achieving adaptive compensation for the control strategy.

[0079] In step S250, the cloud server performs spatiotemporal alignment and fusion of multi-source heterogeneous data. The cloud server receives the defect feature matrix (denoted as ) from the edge computing node 410 via a message queue (such as Kafka). Simultaneously, the system collects real-time operating data of the crusher through an industrial IoT gateway, including instantaneous active power, effective value of rotor vibration acceleration, and cumulative operating time. Considering the physical distance between the visual acquisition point (located on the conveyor belt at the crusher outlet) and the crusher cavity, the cloud server constructs a time sliding window based on the real-time encoder speed value of the conveyor belt, accurately backtracking and matching the current crusher operating status with the visual characteristics of the corresponding material batch on the timeline. Specifically, the system uses the moment the material passes through the crushing cavity as a benchmark, plus the transmission delay calculated based on belt speed, to lock the corresponding visual detection frame, thereby constructing a time-aligned joint dataset containing material attributes and equipment status.

[0080] In step S260, the cloud server calculates the current sampling time based on the historical cumulative load. The equipment performance degradation index (denoted as ) This index is a monotonically increasing time function used to quantify the physical wear of core crushing components of a crusher (such as hammers and impact plates). For the initial state after the system's first startup or replacement of new components, the following settings are used: And load the factory-preset universal control mapping matrix. This general matrix is ​​constructed based on laboratory calibration data of the same model of equipment. This embodiment is based on the principle of energy and wear equivalence, assuming that the wear rate of the equipment is not constant, but rather non-linearly positively correlated with the average crushing resistance complexity index of the processed material and the instantaneous active power. The cloud server updates this index in real time using a discrete numerical integration method; the calculation logic is as follows:

[0081] ;

[0082] in, Indicates the current sampling time The device performance degradation index is normalized to the range [0, 1], where 0 represents a brand new state and 1 represents a critical state where the recommended replacement of parts is made. This indicates the positive deviation of the vibration amplitude from the reference value;

[0083] It represents the instantaneous active power of the main motor at the current moment, which characterizes the energy input of the crushing process. The higher the instantaneous active power, the stronger the grinding effect between the material and the liner.

[0084] This represents the average crushing resistance complexity index of the batch of materials entering the crusher at the current moment. This term is introduced to differentiate the wear caused by processing soft materials (such as plastic shells) and hard materials (such as compressors) in a weighted manner.

[0085] This represents the effective value of the rotor vibration acceleration at the current moment;

[0086] This represents the unloaded reference vibration value. Only additional vibration impacts exceeding the unloaded reference vibration value are included in abnormal losses.

[0087] This indicates the sampling time interval (usually between 100ms and 500ms).

[0088] The power wear coefficient (value range 1.0 × 10⁻⁶) -7 Up to 5.0×10 -7 ), The vibration impact coefficient (value range 2.0 × 10⁻⁶) -6 Up to 8.0×10 -6 The above coefficients need to be calibrated during the equipment commissioning phase by comparing the actual weight loss curve of the high manganese steel hammerhead.

[0089] In step S270, the cloud server constructs a mapping relationship and generates a dynamic state vector (denoted as ) for downstream control. To standardize the input spatial dimensions, the cloud server first performs statistical aggregation on the defect feature matrix (N×K), extracting the mean and variance of each dimension to construct a batch feature vector with a dimension of 1×2K (denoted as...). Subsequently, the system utilizes nonlinear functions to construct device state characteristic terms ( This item is determined by the equipment performance degradation index ( The system constructs a joint input vector by performing vector concatenation operations and left-multiplying it by the global control mapping matrix ( ). To calculate the control parameters. The calculation logic is modified as follows:

[0090] ;

[0091] ;

[0092] in, For equipment status characteristics, This represents the output dynamic state vector, which includes belt speed correction values ​​for the downstream eddy current separator and jet valve delay correction values ​​for the induction separator.

[0093] This indicates a vector concatenation operation that integrates material characteristics and equipment status into a unified input space.

[0094] This represents the nonlinear compensation gain function, and the basis for setting this function is: when wear is in the initial stage (e.g., ... When the wear period begins, the impact on the crushing effect is small, and the gain remains at 1.0; when the wear period begins (e.g., during the severe wear period), the effect on the crushing effect is small, and the gain remains at 1.0. When this happens, the gain needs to increase exponentially in order to implement a more aggressive parameter compensation strategy to offset the particle size degradation caused by the increase in the crushing gap.

[0095] This represents the global control mapping matrix, which is pre-constructed through multivariate linear regression analysis of historical production data (including input material characteristics, equipment status, and final sorting purity). As a technically equivalent alternative, those skilled in the art can also use a pre-trained multilayer perceptron (MLP) network to replace this matrix multiplication operation in order to capture more complex high-dimensional nonlinear relationships.

[0096] In step S280, the system performs integrity verification and distribution of the dynamic state vector. To prevent abnormal calculation of the equipment performance degradation index due to sensor failure (such as vibration sensor reading drift) from causing oscillations in the control system, the cloud server incorporates a safety confidence interval logic. The system monitors the rate of change of each component of the dynamic state vector in real time. If the Euclidean distance between two adjacent time windows exceeds a preset safety threshold (e.g., 15% of the rated control amplitude), the cloud server will automatically clamp the output, keeping the control parameters unchanged from the previous moment, and trigger a model parameter calibration alarm, prompting manual intervention to check the physical state of the crusher. After confirmation, the dynamic state vector is distributed to the PLC control layer via a lightweight message transmission protocol and directly superimposed on the setpoint of the basic PID control loop, thereby achieving a leap from passive reception to active predictive compensation in closed-loop control.

[0097] To construct a physical reference truth value at the physical level that can verify the visual recognition results, the system is configured to extract load feature fingerprints characterizing the material's hardness and impact properties from the equipment's electromechanical signals. While visual sensors can provide surface texture information of materials, they cannot accurately penetrate the crushing chamber to detect the actual crushing resistance in high-dust environments. However, the main motor current and machine vibration signals directly reflect the energy transfer and momentum exchange during the crushing process. This embodiment addresses the cognitive bias problem of a single visual modality under complex working conditions by establishing a multi-domain signal analysis pipeline and constructing a load feature fingerprint strictly aligned with the visual time axis.

[0098] In step S310, the edge computing node performs high-frequency acquisition of multi-channel analog signals and hardware synchronization. The system uses a Hall current sensor installed on the stator side of the main motor to acquire instantaneous stator current and a piezoelectric accelerometer rigidly connected to the crusher bearing housing to acquire machine vibration signals. Considering that the stress release during material crushing usually occurs on the microsecond to millisecond scale, in order to fully capture this transient impact characteristic, the sampling frequency of the machine vibration signal is set to no less than 10kHz, while the sampling frequency of the relatively sluggish instantaneous stator current is set to no less than 1kHz. Before the analog signal enters the analog-to-digital converter (ADC), the hardware circuit performs anti-aliasing filtering based on Butterworth topology, and its cutoff frequency is strictly set to half of the sampling rate (i.e., the Nyquist frequency) to physically eliminate the interference of high-frequency noise folding on the effective frequency band.

[0099] In step S320, the system performs adaptive noise reduction processing based on wavelet transform on the original signal. Given that high-power frequency converters in industrial environments generate wideband electromagnetic interference, and that this interference spectrum often covers the characteristic frequency band of the fracture impact signal, traditional linear Fourier filtering tends to smooth out critical impact peaks while filtering out noise. Therefore, this embodiment uses discrete wavelet transform (DWT) to decompose the signal into approximation coefficients and detail coefficients at different scales, and uses a soft thresholding function to shrink the detail coefficients. The calculation logic is as follows:

[0100] ;

[0101] in, This represents the processed wavelet detail coefficients. These represent the original wavelet detail coefficients obtained from the original decomposition. Represents a symbolic function. This represents the adaptive denoising threshold. To avoid setting the threshold too high due to an excessively large standard deviation of the noisy signal, this embodiment uses a robust estimation method based on the median absolute difference (MAD) to determine the threshold:

[0102] ;

[0103] in, These are the first-level wavelet detail coefficients. The signal length is denoted by . The physical significance of this calculation method lies in using the statistical properties of the median to eliminate the interference of abnormal impulses on the noise level estimation, thereby obtaining a more accurate silence threshold.

[0104] In step S330, the system performs load event anchoring based on the energy operator. The continuous fracturing process is microscopically composed of discrete single fracturing events. To accurately interlock the load signal with the visually captured particles in the spatiotemporal dimension, the start and end points of each fracturing action must be accurately segmented from the continuous time series. This embodiment introduces the Tigger energy operator to enhance the transient impact component in the signal. Utilizing the operator's dual sensitivity to instantaneous amplitude and frequency changes in the signal, the boundaries of the fracturing events are locked. The discretization calculation of the energy operator is as follows:

[0105] ;

[0106] in, Represents the discrete vibration signal sequence at time t. The instantaneous energy characteristic value; This represents a discrete vibration signal sequence after noise reduction processing. This indicates the discrete vibration signal sequence at the previous sampling time. The value; This indicates the discrete vibration signal sequence at the next sampling time. The value.

[0107] Based on the calculated instantaneous energy characteristic value, the system executes a dual-threshold Schmitt triggering logic: When the instantaneous energy characteristic value exceeds the event trigger threshold (set to 5 times the no-load energy baseline) for three consecutive sampling points, the system marks the current moment as the event start point; when the instantaneous energy characteristic value falls back to the event release threshold (set to 1.2 times the no-load energy baseline) and remains there for more than 50ms, the system marks the current moment as the event end point. This hysteresis comparison logic effectively prevents a single event from being incorrectly fragmented into multiple segments due to signal jitter.

[0108] In step S340, the system constructs a multi-dimensional load feature fingerprint vector (denoted as ) for each anchored load event segment. The purpose of constructing this vector is to quantify the breakability and material properties of the material being broken from a physical perspective. Since brittle components (such as phenolic resin / glass) generate high-frequency short pulses upon breakage, while ductile components (such as metal casings) generate low-frequency long-tailed oscillations, the system extracts the signal statistical kurtosis and spectral centroid frequency as core fingerprint features. The calculation formula is as follows:

[0109] ;

[0110] in, This represents the output load feature fingerprint vector, where the first component is the signal statistical kurtosis and the second component is the spectral centroid frequency. This represents the instantaneous vibration amplitude within a load event segment. This represents the arithmetic mean of the signal within that segment. This represents the standard deviation of the signal within that segment. This indicates the total number of event data points in this event segment. This indicates that the segment, after undergoing a Fast Fourier Transform (FFT), has a frequency... The power spectral density amplitude at that point, This represents the total number of spectral lines in the frequency domain analysis. Represents the numerical stability constant (with a value of 10). -8 The technical purpose of introducing this constant is to prevent division-by-zero overflow errors when the device is in an absolutely steady state (standard deviation approaches 0) or a quiescent state (total energy approaches 0), thus ensuring the robustness of the calculation process.

[0111] Based on the above calculations, the signal statistical kurtosis obtained by the system can sensitively reflect whether there are extremely difficult-to-break solid metal foreign objects (such as motor rotors) mixed in the material (exhibiting extremely high kurtosis values ​​far exceeding the normal distribution), while the spectral centroid frequency can effectively distinguish the material hardness (the higher the hardness, the richer the high-frequency components, and the higher the spectral centroid frequency). The load feature fingerprint vector composed of these two physical quantities will serve as an objective physical reference truth value, used in subsequent steps to perform closed-loop verification and correction on the dynamic state vector generated in the cloud.

[0112] After acquiring the load characteristics of the high-frequency electromechanical signal through the aforementioned steps, the core task of the system is to solve the spatiotemporal asynchrony problem between visual perception and mechanical response. Since the visual acquisition unit is installed above the conveyor belt, while physical crushing occurs within the main machine cavity, there is a fixed spatial distance between the two. Furthermore, the transmission time exhibits nonlinear dynamic fluctuations due to random factors such as material sliding and rolling on the belt and accumulation in the hopper. Directly using a fixed time delay parameter would lead to a mismatch between the visually recognized high-hardness compressor and the low-load waveform acquired by the mechanical end, resulting in training samples containing severe noise. Therefore, this embodiment employs a sliding window-based cross-correlation analysis technique. Utilizing the waveform topological similarity of the material flow in the visually predicted load sequence dimension and the physical response load sequence dimension, it dynamically searches for the optimal time lag, thereby achieving microsecond-level physical anchoring of heterogeneous data.

[0113] In step S350, the edge computing node constructs the first The visual prediction load sequence values ​​corresponding to the frames (denoted as ) The system analyzes the video stream acquired by the visual sensor frame by frame, and maps the two-dimensional image information into a one-dimensional time series based on the volume-material integration principle. To eliminate waveform distortion caused by conveyor belt speed changes, the system does not simply stack data based on frame rate, but rather maps the spatial domain distribution into a time domain sequence based on the real-time speed integration of the conveyor belt. Specifically, the system records the instantaneous belt speed at the time of each frame acquisition, calculates the material displacement increment through integration, and uses this as the time axis reference for the sequence, thereby ensuring that the visual prediction sequence does not undergo nonlinear distortion relative to the physical signal on the time axis. To accurately assess the contribution of material material to the crushing load, this embodiment integrates a lightweight convolutional neural network (e.g., based on the MobileNetV3 backbone architecture) as a material hardness estimator. Specifically, the network receives the material region of interest image (normalized to a 224×224 pixel RGB three-channel matrix) cropped from the video frame, passes it through a feature extraction layer and a global average pooling layer, and then regresses it through a fully connected output layer to obtain a normalized hardness coefficient (range [0, 1]). Based on this model output, the system generates a sequence through the following logic:

[0114] ;

[0115] in, Indicates the first The visual prediction load sequence values ​​corresponding to the frames. This indicates the total number of valid material objects identified within the current frame. Indicates the first The pixel projection area of ​​a material object is physically represented as a two-dimensional mapping of the material's geometric volume. This represents the output of the convolutional neural network. The normalized hardness coefficient for each material object. This coefficient is introduced to reflect the different crushing resistances produced by different materials (such as porous plastics and dense motor components) under the same volume.

[0116] In step S360, the system is constructed and the first Frame physical response load sequence (denoted as Since the sampling rate of the original electromechanical signal (e.g., 10kHz) is much higher than that of the visual sampling rate (e.g., 30Hz), directly performing correlation analysis would result in wasted computational power due to the curse of dimensionality and introduce high-frequency noise interference. As a preferred approach, the system first processes the instantaneous energy characteristic values ​​(…) obtained from the aforementioned steps. A moving average filter is performed to extract the low-frequency energy envelope, followed by anti-aliasing downsampling processing using a multiphase decimation filter, strictly aligning its time resolution to the visual sampling rate. This processing step filters out the structural resonance noise of the crusher itself while preserving the load fluctuation trend.

[0117] In step S370, the system performs spatiotemporal alignment analysis based on normalized cross-correlation (NCC). The system slides the visual prediction load sequence and the physical response load sequence within a preset physical delay search window, calculating the waveform similarity between the two under different time lags. The range of values ​​for this search window is... The normalized cross-correlation function is determined by the physical constraints of the conveyor belt length and the limiting belt speed. The calculation logic is as follows:

[0118] ;

[0119] in, This indicates that the time lag is The normalized cross-correlation coefficient at time t is in the range of [-1, 1]. This represents the discrete-time lag, and its search step size corresponds to the visual frame interval. This represents the numerical values ​​of the visual prediction load sequence within the calculation window. This represents the numerical value of the physical response load sequence after translation. This represents the arithmetic mean of the visual prediction load sequence within the current calculation window. This represents the arithmetic mean of the physical response load sequence within the current calculation window. Subtracting the mean aims to eliminate differences in the DC component (i.e., the no-load reference value), focusing only on the morphological matching of load fluctuations. Indicates the length of the correlation analysis window (e.g., a value of 300 corresponds to 10 seconds of data). Represents the numerical stability constant (with a value of 10). -7 The purpose of introducing this constant is to prevent the denominator from becoming zero when the equipment is in an absolutely idle and stationary state (where the sequence variance approaches 0), which could lead to computational overflow or program crash, thus ensuring the completeness of the algorithm under all operating conditions.

[0120] In step S380, the system performs physical anchor point locking and truth value interlocking verification. The system iterates through all possible lags within the search window to find the optimal lag that maximizes the normalized cross-correlation coefficient. To eliminate false matches caused by random noise, the system introduces a confidence locking threshold (empirically set at 0.75). In response to the calculated maximum cross-correlation coefficient exceeding this threshold, the system determines a causal relationship between the current visual observation window and the physical response window, and then... The current system transmission delay truth value is locked. At this time, the system triggers the truth value interlocking mechanism: first, it determines whether the current load event's time window contains only the visual prediction waveform of a single material object (i.e., no overlapping breakage). If the single-object independence condition is met, then the load feature fingerprint vector extracted from the physical end ( As an objective truth label, the physical anchor point is automatically associated with the corresponding material object ID identified by the visual system, generating self-supervised training samples that establish a mapping relationship between visual image features and load feature fingerprint vectors. Conversely, if the maximum cross-correlation number is lower than a threshold, the system identifies it as a transmission anomaly (such as violent material tumbling or sensor data loss), automatically discarding data for that time period to prevent incorrect physical labels from contaminating the training sample library of the visual model. Through this mechanism, the system can achieve autonomous iteration and evolution of the model by utilizing the constraints of physical laws without manual annotation. The system defines the maximum normalized cross-correlation number as the physical anchor confidence level (denoted as ). And store it in a virtual material data object.

[0121] In response to the maximum normalized cross-correlation coefficient falling below the confidence threshold, the system invokes the theoretical transmission time (i.e., the purely mechanical deduction time) based on the conveyor belt encoder integral as a backup anchor point, forcing entry into step S390 for deviation verification. If the visual perception deviation index calculated under the backup anchor point... If the value is extremely high (exceeding the safety threshold), it is determined to be unrelated noise and a latent hardness anomaly, triggering the correction logic.

[0122] Although the aforementioned steps have achieved microsecond-level time alignment of multimodal data, misjudgments may occur under complex incoming material composition conditions. For example, a high-hardness compressor or large motor tightly wrapped in insulating foam or a plastic shell may be misjudged by the vision system as pure plastic or a thin-walled casing. Without correction, the control system may operate in a low-power mode for handling soft materials, leading to blade jamming or motor overload. To address this, this embodiment constructs a closed-loop correction mechanism based on physical truth feedback. This mechanism utilizes the objectivity of mechanical response to quantify the confidence level of visual perception in real time and dynamically generates compensation vectors to mitigate perception noise.

[0123] In step S390, the edge computing node performs instantaneous perception bias calculation. The system calls the visual prediction load sequence and physical response load sequence, which were previously time-aligned in the physical anchor locking step, and calculates their relative difference within the current sliding window. Considering that the difference at a single moment is susceptible to random noise and that the severity of the bias needs to be evaluated relative to the current load base, this embodiment defines a normalized dimensionless index—the visual perception bias index (denoted as ). The calculation logic is as follows:

[0124] ;

[0125] in, Indicates at time The visual perception bias index, whose positive and negative signs represent the direction of the bias (positive value represents physical load higher than visual expectation, negative value represents physical load lower than visual expectation).

[0126] The energy visual mapping coefficient is obtained through linear regression fitting of historical data and is used to map the dimensions of visual features to the dimension space of physical load energy.

[0127] Represents the physical response load sequence at historical moments The value;

[0128] Indicates the visual prediction load sequence at historical moments The value;

[0129] This indicates the length of the deviation assessment window (set to 50 sampling points, covering approximately 1-2 seconds of the breakup process).

[0130] This represents the load reference stability constant (taken as 1% of the system's rated load). This constant is introduced to prevent division by zero overflow when the value of the visual predicted load sequence is zero (such as no-load recognition), while suppressing noise amplification under low-load conditions.

[0131] Represents the time decay weighting function (in exponential form) ,in Its physical significance lies in giving higher weight to recent data to ensure that the system can respond quickly to sudden changes in lithology.

[0132] In step S392, the system performs operational condition classification and control strategy arbitration based on the visual perception deviation index. The system presets a positive deviation lockout threshold ( ) and negative deviation blocking threshold ( ). Response to the visual perception bias index The system determines that the current operating condition is one of latent hardness. For example, the material may appear as a regular washing machine casing, but internally it contains unremoved weighted cement blocks, stainless steel inner drum bearings, or high-strength cast iron components. In this case, visual data alone cannot accurately predict the load; it is necessary to combine this with the abrupt changes in the current waveform for identification; this is in response to the visual perception deviation index. The system determines that the current condition is a falsely dense working condition (e.g., the material is loose and porous). In this case, the physical response load sequence value is lower than the visually predicted load sequence value, indicating excessive energy waste due to excessive breakage. For conditions under these conditions... The system identifies the state of the interval as a reliable perception zone, and the control weights are completely tilted towards the feedforward prediction results of the visual model.

[0133] In step S393, the system generates a dynamic control compensation vector (denoted as ) based on the arbitration result. To optimize crushing efficiency while ensuring equipment mechanical safety, the control system employs a nonlinear gain scheduling strategy to fine-tune the crushing gap (Gap) and the main motor frequency. The correction formula is as follows:

[0134] ;

[0135] in, This represents the execution control vector that is ultimately output to the actuator. This represents the baseline control vector generated solely based on predictions from a visual model. This represents the control correction gain matrix, whose diagonal elements define the maximum allowable physical adjustment range for each control channel (such as the toolbox adjusting cylinder and the feeder frequency converter). This represents the hyperbolic tangent activation function. Utilizing the soft saturation property of this function, the correction amount is limited to the interval [−1, 1], preventing abrupt changes in control parameters due to extreme visual perception bias, thereby protecting the mechanical structure from impact. This represents the sensitivity adjustment coefficient;

[0136] Represents the security constraint mask vector. This represents the Hadamard product. This vector is generated by the hardwired safety logic of the underlying PLC. When the equipment is in an overload protection or fault state, the corresponding mask bit is forcibly set to zero, physically cutting off the correction channel of the higher-level algorithm to ensure the safety baseline of the equipment operation.

[0137] It is worth noting that this dynamic compensation not only applies to the current control cycle but also serves as a short-term memory correction for the control parameters of subsequent adjacent materials. That is, when the system detects that the current material has a hidden hardness, it will temporarily raise the system's alert level and apply pre-protective parameter compensation to the material that immediately follows into the crusher (even if it is visually judged to be ordinary material), thereby preventing hysteresis-induced cumulative overload caused by a continuous flow of implicitly high-hardness material.

[0138] In step S394, the system performs hard example mining and online model correction. Given that the visual model outputs a normalized hardness coefficient, while the physical signal reflects the total load, the system first performs physical equivalent hardness inversion to construct a dimensionally consistent supervision signal. Based on the cumulative value of the physical response load sequence within the current window and the sum of the identified pixel projection areas, the system calculates the physical inversion hardness coefficient (…). ):

[0139] ;

[0140] in, This represents the physically inverted hardness coefficient calculated based on physical signals. The values ​​represent the physical response load sequence within the window. For the first The pixel projection area of ​​each material object. Any time segment that triggers a control correction (i.e., the absolute value of the visual perception bias index exceeds a set threshold) is marked as a high-value case sample by the system. During the idle periods of the edge computing nodes, the system uses this set of difficult cases to fine-tune the fully connected classification head of the aforementioned material hardness estimator (convolutional neural network). The training process uses mean squared error (MSE) as the loss function.

[0141] ;

[0142] in, This represents the training loss function value. Indicates the number of samples in the batch. This represents the predicted normalized hardness coefficient output by the visual model.

[0143] By updating network weights through backpropagation, the visual model can gradually learn to recognize the subtle texture features of camouflaged materials, thereby continuously reducing the visual perception bias index in subsequent operations and achieving lifelong evolution of the intelligent system.

[0144] In a multi-stage crushing and screening process, materials of a specific material (such as the aforementioned high-hardness motor rotor or oil-impregnated compressor) identified by the upstream coarse shredder (e.g., a twin-shaft shredder) must pass through a variable-frequency conveyor belt and an intermediate buffer hopper 140 before reaching the downstream equipment. Since the conveyor belt is constantly under variable-frequency speed control, and the material level in the buffer hopper fluctuates in real time according to supply and demand, the material transport lag time exhibits time-varying nonlinear characteristics. If only a fixed delay setting is used, the parameter adjustment of the downstream equipment will inevitably be misaligned with the actual arrival time of the material (e.g., prematurely tightening the crushing gap before the arrival of the high-hardness component, leading to material blockage). Therefore, this embodiment constructs a virtual material tracing model based on velocity integral and volume conservation. This model does not rely on expensive online isotope tracking but generates a timestamped virtual data object for each physically anchored material unit based on a Lagrange perspective, and performs real-time simulation of its transport path in a digital twin space.

[0145] In step S410, the system activates and updates the spatial tracking status of the virtual material data object. In the aforementioned steps S100 or S200, the object was initialized and assigned a unique index ID when the vision system first detected the material. In this step, the system will use the physical anchor timestamp confirmed in step S380 (… ), normalized hardness coefficient ( The physical response load sequence characteristics are then appended to the object. Subsequently, this data object is logically mapped to the geometric center of the discharge port of the coarse shredder, serving as the spatial zero point of the transmission stroke (x=0), marking the start of lifecycle tracking for this batch of materials.

[0146] In step S420, the system performs a kinematic integral time-delay deduction of the conveyor belt link. Considering that the linear velocity of the conveyor belt is a variable dynamically adjusted with the feed rate, the residence time of the material on the belt cannot be obtained by simply dividing the distance by the velocity; a rigorous time-domain integration is necessary. The system uses a sampling interval of the control cycle (… The system collects the conveyor belt speed feedback value from the frequency converter in real time (preferably 50ms-100ms). For each active virtual material data object, the system performs displacement accumulation calculation until the accumulated displacement covers the physical length of the conveyor belt. The mathematical conditions for determining whether the material has reached the end of the belt are as follows:

[0147] ;

[0148] in, This represents the physical anchor timestamp used to generate the material data object. This indicates the calculated arrival time of the belt end. Indicates the integral variable The constantly collected feedback value of the conveyor belt running speed. This represents the belt slippage correction factor, typically set between 0.95 and 0.99. This factor is an empirical calibration value used to compensate for speed losses caused by belt tension fluctuations, roller wear, or microscopic material slippage. The system dynamically updates this factor by comparing the long-term deviation between the belt scale's cumulative flow rate and the theoretical volumetric flow rate. This represents the time integral infinitesimal element (unit: second), corresponding to the sampling time step in the discrete computation of the system. Multiplying it by the velocity unit (m / s) yields the physical displacement (meters). This indicates the effective physical length of the conveyor belt from the material drop point to the unloading point.

[0149] In step S430, the system performs a silo delay correction based on the volume replacement principle. After the material leaves the conveyor belt and falls into the intermediate buffer silo 140, its residence time depends on the current material level and the downstream discharge rate. In this embodiment, the intermediate buffer silo 140 is designed as a large-cone angle integral flow silo to ensure that the material follows a first-in, first-out (FIFO) flow order, thus making the residence time calculation based on volume integral physically valid. It is assumed that the material flow within the silo conforms to a plug flow model, i.e., the material layer flows first-in, first-out, and extreme axial mixing effects are ignored. The system acquires real-time data from ultrasonic or radar sensors installed on the top of the silo to obtain the real-time material level. Combined with the volume discharge rate of the downstream feeder, the residence time is dynamically calculated using the law of volume conservation. The specific calculation formula is as follows:

[0150] ;

[0151] in, Indicates the current calculation time Estimated buffer silo residence time;

[0152] This represents the cross-sectional area function of the silo, which describes the cross-sectional area of ​​the silo as a function of its height. The function represents a changing geometric relationship. For common frustum-shaped or conical silos, this function is a quadratic polynomial of height, which the system calls through a preset CAD geometric parameter table.

[0153] Indicates time Real-time material level height collected;

[0154] This represents a height-integral infinitesimal element, used for micro-layer accumulation of the silo cross-section in the vertical direction;

[0155] This indicates the downstream feeder (such as a plate feeder) at time [time]. The system uses either a set volume discharge rate or an average discharge rate predicted based on historical trends. It provides an instantaneous estimate based on the current state and updates it dynamically over time. Considering that the discharge rate may be dynamically adjusted during actual operation, the system employs a rolling prediction correction mechanism: when a change in the downstream feeder speed setpoint is detected, the remaining residence time of the material still in the bin is immediately recalculated based on the new flow rate, and its arrival timestamp is updated.

[0156] In response to the detection of downstream feed rate If the output falls below the critical threshold (i.e., in a stopped or blocked state), the system pauses the integral accumulation for that time step and marks the corresponding virtual material object as suspended until discharge resumes, in order to reflect the actual physical blockage process.

[0157] This represents the minimum flow rate reference constant (0.1% of the rated flow rate). This constant is introduced to prevent the denominator from approaching zero when the downstream feeder stops or is in a very low-speed creeping state, which could lead to overflow or singular values ​​in the calculation results.

[0158] In step S440, the system synthesizes the precise arrival timestamp of the downstream device ( The system sequentially superimposes the integral time delay of the conveyor belt stage with the volume replacement time delay of the hopper stage, and introduces the free fall compensation time of the chute section. The final target arrival time is calculated using the following formula:

[0159] ;

[0160] As a preferred embodiment, since The system is inherently affected by the material discharge rate at future moments, and employs a prediction and correction mechanism: after preliminary calculation... Subsequently, if a change in the set value of the downstream feeder speed is detected, the system will immediately trigger the recalculation logic to refresh the arrival time of the virtual material data object until the material actually enters the downstream equipment.

[0161] In step S450, the edge computing node 410 is also configured to perform instruction fusion processing when generating synchronization control instructions, that is, the system performs control instruction fusion arbitration based on spatiotemporal conflicts. In industrial settings, due to material mixing at transfer points or non-ideal flow within silos, the system maintains a timestamp based on the precise arrival of downstream equipment. The pending control queue is sorted. When the arrival time interval between two adjacent virtual material data objects in the queue is less than the response adjustment dead time of the downstream equipment (…), the queue is processed. When the time required for the hydraulic system to complete one full stroke is typically set to 1.2 times, the system triggers conflict arbitration logic. The arbitrator is based on the confidence level of the physical anchor points of the two data objects (…). ) and normalized hardness coefficient ( Calculate the weighted fusion control objective (where the confidence levels involved in the calculation have all passed the threshold screening in step S380 above to ensure that the denominator is not zero):

[0162] ;

[0163] in, This represents the fusion control hardness coefficient, which will serve as the final setting basis for downstream equipment. and These represent the normalized hardness coefficients of two adjacent data objects in the conflict queue; and These represent the physical anchor confidence levels of two adjacent data objects in the conflict queue. The system then uses this fused control hardness coefficient. Generate a single, compromise control command, prioritizing the mechanical safety and operational stability of the equipment, rather than blindly pursuing immediate responses to every transient change.

[0164] At the physical site of waste household appliance dismantling operations, downstream main equipment such as industrial crushing devices 120 with hydraulic adjustment functions (such as adjustable hammer crushers or large twin-shaft shredders) typically rely on hydraulic or servo mechanisms to drive the movable cutter box or adjust the position of the grid to adjust the crushing gap. However, due to the compressibility of hydraulic fluid, the flow resistance along long pipelines, and the inertial mass of heavy mechanical components, there is an objective and unavoidable electro-hydraulic response lag between the electrical signal command issued by the control system and the actual physical displacement of the actuator. If the control system only triggers actions synchronously based on the arrival time of the material, the mechanical response will inevitably lag behind changes in operating conditions (for example, a high-hardness component has entered the crushing chamber, but the crushing gap has not yet completed its retraction action). This timing misalignment is the core reason for the instantaneous overload shutdown of the crusher or the failure to meet the required degree of dissociation of the crushed products. In view of this, this embodiment, based on obtaining a high-precision material arrival timestamp, constructs a reverse time compensation mechanism based on a hydraulic dynamics model. By pre-simulating the dynamic behavior of the actuator, the lead time is accurately calculated on the time axis to ensure that when the leading edge of the material flow arrives at the geometric boundary of the crushing chamber, the equipment is precisely in the preset optimal geometric state.

[0165] In step S460, the system performs target mechanical state mapping based on material toughness characteristics. The system reads the fused control hardness coefficient ( ) carried in the virtual material data object to be processed. This process, combined with the visual particle size distribution characteristics of the batch of materials, queries a pre-set expert database linking material properties and equipment parameters. The mapping process aims to match the optimal shearing and crushing conditions for the current material. Specifically, for materials with a high hardness coefficient, the system matches a larger target crushing gap value. To prevent clogging; for low-hardness fine materials, match a smaller target crushing gap to increase the crushing ratio.

[0166] In step S470, the system performs a response time delay estimation of the hydraulic actuator. This embodiment recognizes that the response speed of a hydraulic system is a variable that dynamically drifts with operating conditions, influenced by the coupling effect of hydraulic oil temperature, adjustment stroke, and load direction. The system collects the hydraulic oil temperature of the hydraulic station in real time (…). ) and system main pressure ( ), combined with the current breaking gap fed back by the displacement sensor ( The expected motion time is calculated using a dynamic model. To ensure the stability of the mathematical calculations, all physical coefficients involved in the model are constrained to positive definite values. The calculation formula is as follows:

[0167] ;

[0168] in, This indicates the estimated time required for the actuator to complete the movement from its current position to the target position;

[0169] This represents the inherent dead time of the system, encompassing the PLC instruction communication delay, the establishment of the magnetic field by the electromagnetic directional valve coil, and the physical delay of the valve core opening. This value is an inherent physical property of the equipment (typically 100ms-200ms), obtained from the step response test calibration at the time of equipment delivery.

[0170] This indicates the target crushing gap determined based on material properties;

[0171] This indicates the current breaking gap, as fed back by the sensor in real time.

[0172] This represents the reference push rod speed of the hydraulic cylinder under rated system pressure and standard oil temperature, with the unit being mm / s. This parameter is a non-zero system constant.

[0173] This represents the oil temperature viscosity correction factor, which reflects the effect of temperature on the fluidity of hydraulic oil. This factor is based on the standard operating temperature (e.g., 40°C) (at which point the factor is 1.0); when the hydraulic oil temperature... Below the standard temperature, the oil viscosity increases, leading to a decrease in flow rate; the coefficient ranges from [0.6, 1.0] to [1.0]. Conversely, above the standard temperature, the coefficient ranges from (1.0, 1.1). The system obtains this coefficient in real time through interpolation using the built-in viscosity-temperature characteristic curve.

[0174] This represents the load coefficient in the direction of motion, used to compensate for the asymmetrical influence of gravity and crushing force on the cylinder's movement. When performing the retraction gap action, the cylinder needs to overcome a huge crushing reaction force or accumulator resistance, so this coefficient is relatively small (e.g., 0.85); when performing the yielding action, the self-weight of the movable cutter box assists in oil discharge, so this coefficient is relatively large (e.g., 1.15).

[0175] In step S480, the system calculates the look-ahead control trigger timestamp ( The system calculates the precise arrival timestamp of the downstream device based on the steps S440 above. This involves tracing back to the moment the control command was issued. This step realizes the paradigm shift from event-driven to time-driven control. The calculation logic is as follows:

[0176] ;

[0177] in, This indicates the look-ahead control trigger timestamp when the system sends control commands to the underlying PLC.

[0178] This indicates the precise arrival timestamp of the virtual material data object at the downstream equipment expected to reach the crusher feed inlet;

[0179] This indicates a synchronization safety margin (preferably set to 50ms). The purpose of introducing this margin is to absorb model calculation errors and network communication jitter, ensuring that the action is completed before the material arrives. Even if it causes a brief (millisecond-level) idle wait of the equipment, it avoids the situation where the material has entered the cavity but the equipment is not ready, thereby reducing the risk of spindle damage.

[0180] In step S490, the system performs a timeline scan and issues commands. The system runs a high-frequency clock interrupt service routine (scan period ≤ 10ms) to compare the current system time in real time. The lookahead control trigger timestamps for each pending task in the queue. In response to... The system first calls the parameter conflict arbitration logic from S491 to S493, and calculates the final execution break gap based on the current overall load utilization rate of the equipment. Then, a control message containing the final execution parameters is sent to the actuator controller, and the corresponding virtual material data object is simultaneously marked as a control-triggered state to prevent duplicate execution. As a fault-tolerant mechanism for extreme abnormal operating conditions, if the calculated look-ahead control trigger timestamp is less than the current system time (i.e., ...), ... This indicates that the system has missed the optimal pre-adjustment opportunity due to a sudden increase in upstream material flow rate or computation thread blockage. At this point, the system triggers an emergency safety response mode: instead of performing fine-grained crushing gap adjustment, it directly locks the current crushing gap and controls the hardness coefficient based on the current material's fusion properties. The feeding speed is dynamically reduced, and a warning log for timing alignment loss is generated on the human-machine interface. Simultaneously, the system automatically increases the belt slippage correction coefficient used in step S420 to calculate the arrival time. The weights are used to correct for the predicted arrival time deviations of subsequent materials, thus achieving adaptive calibration of the system. This indicates the timing over-limit threshold, which is usually set to twice the system control cycle (e.g., 200ms) to define whether look-ahead control has completely failed.

[0181] In the closed-loop control of fine crushing and sorting processes, there are often multiple inherent conflicts between control objectives: the process department pursues the ultimate degree of material liberation to improve the metal recovery rate of downstream sorting equipment (such as eddy current separators), which usually requires an extremely small crushing gap width; while the production department pursues the maximum hourly throughput to meet capacity targets, and the equipment maintenance department requires strict limits on vibration and power to extend the life of liners and cutter shafts. The aforementioned step S460 is based on the target crushing gap mapped from the material properties ( While theoretically consistent with material crushing characteristics, under certain extreme transient conditions (such as motor thermal overload, entry of non-crushable materials, or uneven liner wear), this theoretical optimal value may exceed the safety boundaries of the physical equipment. Therefore, this embodiment constructs a value-priority envelope model based on a multi-dimensional state space. This model no longer seeks rigid tracking of the set value but defines a dynamic feasible region that includes mechanical safety constraints and economic value gradients. When the preset target parameter deviates from this feasible region, the system automatically calculates the optimal compromise operating point based on the equipment survival priority principle, achieving a smooth transition from fault shutdown to degraded operation. The edge computing node 410 is configured to perform parameter conflict adjustment.

[0182] In step S491, the system performs a normalized mapping of the overall equipment load utilization rate. Since the constraints of the crusher involve multiple physical dimensions such as current (reflecting torque), hydraulic pressure (reflecting crushing force), and frame vibration (reflecting mechanical stability), and the dimensions, orders of magnitude, and response characteristics of each dimension are completely different, direct algebraic comparisons are not possible. Therefore, this embodiment introduces a dimensionless overall load utilization rate (… The concept maps multi-source, heterogeneous sensor data to a unified [0, 1] normalized state space. The computational logic of this step follows the maximum membership principle (i.e., the weakest link principle), where the overall safety of the system is determined by the physical quantity closest to the tripping limit. The specific calculation formula is as follows:

[0183] ;

[0184] in, This represents the overall load utilization rate, which is a floating-point number between 0.0 and 1.0. The closer the value is to 1.0, the closer the device is to the physical tripping boundary. This indicates the real-time power of the main motor, as collected by the sensor. This represents the rated power of the main motor, which is preset as a non-zero system constant. This represents the environmental derating factor, used to compensate for the reduced heat dissipation capacity of motors in high-altitude or high-temperature environments. For example, in mining areas above 3000 meters altitude, the thin air leads to reduced cooling efficiency; this factor is set at 0.85-0.95 to ensure that the power assessment matches the actual insulation withstand capability. This indicates the main system pressure collected in the preceding steps; This indicates the hydraulic relief valve value of the hydraulic system, which is the pressure setting value for the safety valve to open. This indicates the frame vibration intensity measured by an acceleration sensor mounted on the frame beam (usually expressed as the effective value of vibration velocity in RMS). This indicates the vibration trip limit set in the equipment protection logic.

[0185] In step S492, the system constructs a dynamic value-priority envelope correction function. When the control commands given by the material property model cause the load to exceed the limit, the system needs a clear mathematical criterion to determine the concession range. This embodiment employs an arbitration algorithm based on a nonlinear penalty mechanism, where the load utilization rate is within the safe zone ( ) Fully respond to feedforward control commands to obtain maximum process value; when entering the buffer ( When the crushing gap is forcibly increased exponentially, the ability to operate continuously without shutting down is achieved at the expense of some product particle size. The final crushing gap after correction is ( )calculate

[0186] ;

[0187] in, This indicates the final execution break gap sent to the underlying PLC after conflict arbitration; This represents the target crushing gap obtained in step S460 based on material property mapping; It represents the full range of adjustment of the equipment, that is, the difference between the maximum allowable crushing gap and the minimum allowable crushing gap, and is used as a normalization coefficient to assign physical dimensions to the arbitration item. This represents the arbitration response strength gain, and its value is typically set between 0.5 and 1.5. For scenarios involving high-value copper / gold-containing scrap, a smaller value can be set. The value is set to allow for short-term overload operation of the equipment; for slave equipment, a larger value is set to ensure absolute safety. This represents the safe load baseline, serving as the trigger threshold for intervention control. To prevent the denominator from being zero, this value is strictly constrained during system initialization. (Preferred value: 0.85); The envelope penalty exponent (usually 2.0 or 3.0) determines the steepness of the correction curve. A higher power means a small correction is applied initially in the buffer zone to maintain granularity; however, as the load approaches its limit... The correction force increases exponentially, forming a flexible yet rigid barrier to prevent equipment from hitting the hard shutdown threshold.

[0188] In step S493, the system performs execution smoothing based on the rate of change limit. Taking into account the mechanical inertia and oil circuit shock of the hydraulic system, even the calculated... While meeting static safety requirements, excessively rapid changes in command rates can lead to hydraulic shock (water hammer effect) or system oscillation. This is particularly true for high-rigidity metal components, where rapid shrinkage of the breakage gap can trigger extremely high transient pressure spikes. The system is subject to... Dynamic slope limiter using first-order gradient:

[0189] ;

[0190] in, Indicates the rate of change of control commands over time; This indicates the maximum permissible rate of physical motion in the hydraulic system (e.g., 5 mm / s). This indicates the minimum guaranteed operating speed (e.g., 0.5 mm / s). This term is introduced to prevent issues when the material has extremely high hardness. When the calculated allowable rate approaches zero, it can cause the actuator to deadlock or respond slowly, ensuring that the system always has basic adjustment capabilities. This represents the hardness damping scaling factor (range 0.5-0.9), used to adjust the degree to which hardness inhibits the rate of motion. This represents the fusion control hardness coefficient. This formula implements a nonlinear damping control strategy that adjusts quickly for soft materials and slowly for hard materials: when handling high-hardness metals, it actively limits the action rate to smooth pressure fluctuations; when handling low-hardness materials, it removes the rate limit to improve the system's dynamic tracking performance to flow fluctuations.

[0191] Specific application example: a scenario involving the mixed disassembly of a drum washing machine and an air conditioner outdoor unit.

[0192] Scene Background: A waste appliance shredding line with a capacity of 15 tons per hour. Materials are continuously being transported on a frequency converter conveyor belt. Current Time The material input vision acquisition device 210 detected a waste drum washing machine (including cement counterweight and motor) (referred to as material A), and 1.5 meters behind it was a thin single-door refrigerator shell (referred to as material B).

[0193] Step 1: Feed Sensing and Complexity Calculation:

[0194] On-site operating conditions: The feed vision acquisition device 210 captures a depth image of material A. The edge computing node 410 extracts three core features:

[0195] Volumetric characteristics (Normalized value, medium size).

[0196] High resistance metal density (Motor and counterweight area detected).

[0197] Structural integrity (The entire machine was not disassembled, and it has strong resistance to breakage.) According to the formula... The complexity index was calculated. .

[0198] Step 2: Adjusting the anti-aliasing spacing:

[0199] On-site conditions: System calculations show that if the current belt speed of 1.0 m / s is maintained, material B will enter the crusher in 1.5 seconds. At this time, material A has not yet been crushed (requiring 5.2 seconds), posing a very high risk of load overlap. Action: The edge computing node generates a deceleration command. The variable frequency conveyor belt performs an S-shaped deceleration, reducing the belt speed to 0.2 m / s.

[0200] See attached document Figure 3 , Figure 3 The dashed line represents the risk of aliasing, which increases with the predicted clearing time. As the time increases (moving to the right to 5.2s in this example), the aliasing probability rises sharply; the solid line represents the control algorithm based on... The generated target velocity curve. Due to the circumstances in this example... The system forces the belt speed down to the bottom range (0.2m / s) along the solid line, thus increasing the time difference of material B entering the machine in physical space, ensuring that the crusher shaft current returns to the no-load reference before processing B.

[0201] Step 3: Physical Anchoring and Truth Verification:

[0202] After material A enters the crushing chamber, the load characteristic monitoring device 220 detects a violent vibration and a high-frequency current pulse. At this point, the system needs to confirm whether the washing machine previously seen visually is the object causing the vibration. Action: The system extracts the physical load waveform and performs cross-correlation analysis with the visually predicted load sequence. The location with the highest correlation is locked.

[0203] See attached document Figure 4 , Figure 4 The dashed line represents the theoretical impact moment and intensity estimated by the vision system based on the image; the solid line represents the actual current / vibration data collected by the sensor, which includes noise and physical lag. The system calculates the lag time accurately at the peak by determining the cross-correlation function between the two. The time stamp was 2.1s. This timestamp became the unique physical fingerprint of the material, used to correct the tracking timing of all subsequent materials, eliminating the cumulative error caused by belt slippage.

[0204] Step 4: Conflict Arbitration and Downstream Adaptation:

[0205] When the motor is crushing material A, the instantaneous load rate of the spindle is... The value spiked to 0.95 (close to the overload trip threshold of 1.0). At this point, the system faced a conflict between maintaining a small gap for thorough crushing and increasing the gap to prevent shutdown. Action: The arbitration module intervened, temporarily and dynamically widening the discharge port gap from 50mm to 85mm, allowing the difficult-to-crush motor stator to pass through quickly, followed by a rapid reset. Simultaneously, based on flow tracking, the cloud notified the downstream eddy current separator 200ms in advance to increase its speed to 2800rpm.

[0206] See attached document Figure 5 , Figure 5 The diagram illustrates the value-priority arbitration envelope of the system. The left side of the diagram represents the safe zone when the load rate... 0.85 (safety threshold) When the load spikes to 0.95, the system prioritizes efficiency, maintaining a target clearance of 50mm. The gray shaded area on the right is the arbitration zone; in this example, when the load surges to 0.95, it falls into this zone. According to the formula... The curve rises exponentially and steeply, which means that the system prioritizes survival (preventing shutdown) over efficiency and forces a larger gap correction value to be output. This flexible avoidance mechanism effectively eliminates the risk of hard tripping in traditional PID control.

[0207] Experimental verification and effect comparison:

[0208] Experimental setup:

[0209] Baseline group: Traditional constant gap + PID load feedback (passive deceleration after overload).

[0210] Experimental group: The flow tracking and look-ahead compensation control system of the present invention is used.

[0211] Sample: 500 mixed waste household appliances (including washing machines, refrigerators, and air conditioner outdoor units), with a total weight of approximately 20 tons.

[0212] Experimental Results Statistical Table: Comparative experimental data show that the flow tracking and look-ahead control system proposed in this invention has significant technical advantages in handling complex mixed waste household appliances scenarios. First, through visual-machine collaborative anti-overlapping control (step 2), the system successfully eliminated ineffective grinding caused by feed congestion, increasing the throughput by 18.4% while reducing unit energy consumption by 14.6%, verifying the effectiveness of adjusting belt speed based on predicted emptying time; Secondly, the value-first conflict arbitration mechanism (step 4) solves the hysteresis problem of traditional PID control under hard object impact, reduces the number of unexpected equipment shutdowns to zero, and significantly extends the service life of core toolbox components (reducing wear by 28.0%). Based on the time axis calibration function of physical anchoring (step 3), the downstream sorting equipment can achieve microsecond-level precision, breaking the situation of each link in the traditional production line operating independently and increasing the metal recovery rate to 94.2%. This embodiment confirms that the system has successfully achieved a technological leap from traditional passive overload protection to active flow planning, taking into account the high efficiency, high recovery rate and safe operation of the dismantling line.

Claims

1. A material flow tracking and control system for dismantling waste household appliances, characterized in that, It includes a variable frequency feeding conveyor (110), an industrial crushing device (120) and a multi-stage sorting execution device (300) connected in sequence according to the process steps, and also includes a feed vision acquisition device (210) set directly above the variable frequency feeding conveyor (110), a load characteristic monitoring device (220) connected to the power circuit of the industrial crushing device (120), and an edge computing node (410). The edge computing node (410) is communicatively connected to the above devices, and in response to the data from the feed vision acquisition device (210), it constructs a virtual material data object containing a unique index ID and extracts the visual prediction load sequence. The edge computing node (410) collects real-time load current waveform data from the load characteristic monitoring device (220) to construct a physical response load sequence. By matching the physical response load sequence with the visual prediction load sequence, the physical anchor time of the virtual material data object in the industrial crushing device (120) is locked. The edge computing node (410) extrapolates the expected arrival time of the virtual material data object to the multi-level sorting execution device (300) based on the physical anchor time, and outputs an operating parameter adjustment command to adjust the multi-level sorting execution device (300) before the expected arrival time.

2. The waste household appliance dismantling material flow tracking and control system according to claim 1, characterized in that, The system also includes a discharge conveying device (130) located downstream of the industrial crushing device (120), a discharge visual monitoring device (131) located directly above the discharge conveying device (130), and an intermediate buffer silo (140) connected to the end of the discharge conveying device (130). The intermediate buffer silo (140) is equipped with a silo level monitoring component (141) and a quantitative feeding device (142). The edge computing node (410) performs integral calculation on the operating speed feedback value of the discharge conveyor (130) to obtain the integral time delay of the conveyor belt link; The edge computing node (410) obtains the real-time material level height of the silo monitoring component (141) and the volume discharge rate of the quantitative feeding device (142), and calculates the buffer silo residence time based on the volume replacement principle. The edge computing node (410) superimposes the physical anchor point time, the integral time delay of the conveyor belt link, the residence time of the buffer hopper, and the preset free fall compensation time to generate the expected arrival time.

3. The waste household appliance dismantling material flow tracking and control system according to claim 1, characterized in that, The edge computing node (410) runs adaptive discretization feed control logic and performs the following operations after generating the virtual material data object: The crushing resistance complexity index of the current waste household appliance unit is calculated based on the image data, and the predicted crushing and emptying time is generated by mapping it. Monitor the physical distance between two adjacent waste household appliances on the variable frequency feeding conveyor (110); In response to the fact that the crushing resistance complexity index of two adjacent waste household appliances exceeds a preset threshold and the physical distance is less than the safe signal separation distance calculated based on the predicted crushing and emptying time, a deceleration control command is generated to adjust the variable frequency feeding conveyor (110).

4. The waste household appliance dismantling material flow tracking and control system according to claim 1, characterized in that, The edge computing node (410) performs the following operations when locking the physical anchor point time: The visual predicted load sequence is constructed based on the pixel projection area and normalized stiffness coefficient in the visual image, and the physical response load sequence is constructed by extracting the event boundaries of the real-time load current waveform data using the energy operator. Within the physical delay search window, a sliding window normalized cross-correlation analysis is performed on the visual prediction load sequence and the physical response load sequence. The time corresponding to the maximum value of the normalized cross-correlation coefficient is marked as the physical anchor time, and the maximum value of the normalized cross-correlation coefficient is recorded as the physical anchor confidence level.

5. A material flow tracking and control system for dismantling waste household appliances according to claim 4, characterized in that, The edge computing node (410) runs a closed-loop correction mechanism based on physical truth feedback: The visual perception bias index is obtained by calculating the relative difference between the physical response load sequence and the visual prediction load sequence after time alignment. In response to the visual perception deviation index exceeding the preset positive deviation lockout threshold or negative deviation lockout threshold, it is determined that there is a hidden hardness abnormality, and a dynamic control compensation vector is generated to correct the control parameters of the industrial crushing device (120). The physical inversion stiffness coefficient is calculated based on the physical response load sequence and the visual perception bias index, and the model parameters used to construct the visual prediction load sequence are updated online based on the physical inversion stiffness coefficient.

6. The waste household appliance dismantling material flow tracking and control system according to claim 1, characterized in that, The multi-stage sorting execution device (300) includes an adjustable eddy current sorting device (310), a magnetic sorting device (320) and a wind sorting device (330) arranged in sequence. The system also includes a flow feedback detection device (230) located at the end of the sorting process. The edge computing node (410) adjusts the magnetic roller speed and belt running speed of the adjustable eddy current sorting device (310) according to the characteristics of the virtual material data object. The edge computing node (410) generates operating parameter adjustment instructions for the magnetic sorting device (320) and the wind sorting device (330) based on the material composition distribution data in the virtual material data object; The edge computing node (410) receives the actual output data collected by the streaming feedback detection device (230) for closed-loop verification.

7. A material flow tracking and control system for dismantling waste household appliances according to claim 1, characterized in that, The system also includes a cloud server (420). The cloud server (420) calculates the equipment performance degradation index based on the historical cumulative load, instantaneous active power and rotor vibration acceleration effective value of the industrial crushing device (120); The cloud server (420) constructs a joint input vector containing the batch feature vector of the virtual material data object and the equipment performance degradation index, and generates a dynamic state vector; The edge computing node (410) receives the dynamic state vector and corrects the target execution parameters in the running parameter adjustment instruction according to the dynamic state vector.

8. A material flow tracking and control system for dismantling waste household appliances according to claim 1, characterized in that, The edge computing node (410) estimates the response time delay of the actuator when outputting the execution parameter adjustment command: Using a hydraulic dynamics model, the estimated action time required to adjust to the target parameters is calculated based on the real-time status and drive characteristics of the multi-stage sorting actuator (300). Subtract the estimated action time and the preset synchronization safety margin from the estimated arrival time to obtain the look-ahead control trigger timestamp; The operating parameter adjustment command is sent when the system time reaches the forward control trigger timestamp via a high-frequency clock interrupt service routine.

9. A material flow tracking and control system for dismantling waste household appliances according to claim 8, characterized in that, The edge computing node (410) performs parameter conflict arbitration: Calculate the overall load utilization rate of the multi-stage sorting execution device (300); When the target parameter of the operation parameter adjustment instruction causes the overall load utilization rate to enter the preset buffer, the target parameter is corrected using a non-linear penalty mechanism. A dynamic slope limiter is applied to the modified target parameters to limit the action rate of the actuator to prevent hydraulic shock.

10. A material flow tracking and control system for dismantling waste household appliances according to claim 4, characterized in that, The edge computing node (410) performs control command fusion arbitration when generating the operating parameter adjustment instruction: Detect the time interval between the expected arrival times of two adjacent virtual material data objects in the control queue to be processed; In response to the time interval being less than the response adjustment dead time of the multi-stage sorting execution device (300), the normalized hardness coefficient and physical anchor confidence of the two adjacent virtual material data objects are obtained. The normalized hardness coefficient is weighted and calculated using the confidence level of the physical anchor point as a weight to obtain the fusion control hardness coefficient, and a single fusion control command is generated accordingly.

Images