An intelligent sorting and circular regeneration system for construction waste

Abstract

The present application relates to the technical field of fault prediction, and discloses an intelligent sorting and recycling system for construction waste. The system includes: a collection subsystem, which is used to install a sensor array on the crusher working surface and the vibrating screen mesh edge of the construction waste sorting device and collect a feature data set; a simulation subsystem, which is used to establish a simulation model of the construction waste sorting device and generate a simulation data set; a measured data subsystem, which is used to collect a labeled measured wear data set under actual construction waste sorting conditions; a training subsystem, which is used to input the simulation data set and the labeled measured wear data set into an adaptive domain adversarial neural network for training to obtain an equipment wear degree prediction model; an output subsystem, which is used to identify wear fault features through the equipment wear degree prediction model and output a fault closed-loop response plan, thereby realizing the dynamic adjustment of the feeding rate, crushing speed, screening frequency, and screening amplitude of the sorting device.

Description

Technical Field

[0001] This application relates to the technical field of fault prediction, and particularly to an intelligent sorting and recycling system for construction waste. Background Art

[0002] The effective sorting and recycling of construction waste can not only reduce environmental pollution but also create significant economic value, which has become a hot field of global concern. However, existing construction waste treatment equipment faces severe challenges during long-term operation. In particular, key parts such as the working surface of the crusher and the edge of the vibrating screen mesh are prone to "excessive wear" during long-term contact with irregular construction waste materials, resulting in a decline in equipment stability, an increase in maintenance costs, a decrease in processing efficiency, and seriously affecting the sorting quality and the quality of recycled materials.

[0003] At present, there are obvious deficiencies in the fault prediction of construction waste sorting and recycling systems. Traditional mechanical equipment fault prediction mainly relies on empirical models or pure data-driven methods. The former overly simplifies the complexity of material-equipment interaction and cannot accurately reflect the irregular shapes and diverse characteristics of construction waste materials; the latter requires a large amount of historical fault data as support, but it is difficult to obtain sufficient fault samples in actual engineering applications, resulting in poor generalization ability under small sample conditions. In addition, existing fault prediction methods often process data from different sources in isolation, lacking an effective multi-source data fusion mechanism, and it is difficult to comprehensively capture the multi-dimensional characteristics of equipment wear, thus affecting the prediction accuracy. Summary of the Invention

[0004] This application provides an intelligent sorting and recycling system for construction waste, thereby realizing the dynamic adjustment of the feeding rate, crushing speed, screening frequency, and screening amplitude of the sorting device, forming a complete fault closed-loop response scheme, and effectively extending the service life of the equipment while ensuring the sorting efficiency.

[0005] In the first aspect of this application, an intelligent sorting and recycling system for construction waste is provided. The intelligent sorting and recycling system for construction waste includes:

[0006] An acquisition subsystem, configured to install a sensor array on the working surface of the crusher and the edge of the vibrating screen mesh of the construction waste sorting device, and acquire a characteristic data set reflecting the equipment wear state;

[0007] A simulation subsystem, configured to establish a simulation model of the construction waste sorting device according to the characteristic data set, and generate a simulation data set including multiple working condition parameter combinations based on the simulation model;

[0008] An actual measurement subsystem, configured to acquire small-sample actual measurement data and calibrate the wear grade under actual construction waste sorting working conditions to obtain a labeled actual measurement wear data set;

[0009] A training subsystem, configured to input the simulation data set and the labeled measured wear data set into an adaptive domain adversarial neural network for training to obtain a device wear degree prediction model;

[0010] An output subsystem, configured to identify wear fault features through the device wear degree prediction model, and optimize the operating parameters of the construction waste sorting device based on the wear fault features, and output a fault closed-loop response scheme.

[0011] Compared with the prior art, the present application has the following beneficial effects: By configuring a multi-dimensional sensor array including pressure sensors, triaxial acceleration sensors, displacement sensors and depth vision camera modules at the crusher working surface and the edge of the vibrating screen mesh, and combining wavelet transform noise reduction, time synchronization calibration and Kalman filter algorithm for multi-source data fusion, high-precision monitoring of the device wear state is realized, and complex wear features that cannot be recognized by traditional single sensors are captured. Based on the simulation model constructed by finite element analysis, Archard wear theory and discrete element method, combined with small sample measured data, a mechanism-data hybrid-driven fault prediction method is formed, effectively solving the problem of poor generalization ability of pure data-driven methods for small sample data, and at the same time avoiding the prediction deviation caused by the over-simplification of pure mechanism models. By designing an adaptive domain adversarial neural network including a feature extractor, a domain classifier and a wear predictor, effective knowledge transfer from the simulation data domain to the measured data domain is realized, making full use of a large amount of simulation data to improve the training effect of the model, while maintaining the adaptability to actual working conditions, and significantly improving the generalization ability of the fault prediction model. Using the depth vision camera module to collect point cloud data, combining three-dimensional convolutional neural network and variational autoencoder for feature extraction and dimensionality reduction, generating low-dimensional wear fault features with high discrimination through orthogonal constraint and KL divergence optimization, using Takagi-Sugeno fuzzy model to describe the nonlinear dynamic characteristics of the sorting system, combining multi-rate sampling framework to group the sensor acquisition frequency, and determining the feedback control gain by solving the linear matrix inequality optimization problem, effectively coping with the uncertainties caused by sensor faults and time delays, and improving the stability and reliability of the system under complex working conditions. Based on the wear fault risk assessment results and the basic strategy of parameter optimization, combined with the Bayesian optimization system to solve the optimal operating parameter combination, realizing the dynamic adjustment of the feeding rate, crushing speed, screening frequency and screening amplitude of the sorting device, forming a complete fault closed-loop response scheme, effectively extending the service life of the device while ensuring the sorting efficiency. Description of the Drawings

[0012] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0013] The structures, ratios, sizes, etc. shown in the accompanying drawings of this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the limiting conditions under which the present application can be implemented. Therefore, they do not have a substantial technical meaning. Any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects that the present application can produce and the purposes that can be achieved, should still fall within the scope that can be covered by the technical content disclosed in the present application.

[0014] Figure 1 It is a schematic flow chart of the intelligent sorting and circular regeneration system for construction waste provided by the embodiment of the present application. Specific embodiments

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of them. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope protected by the present application.

[0016] The flow chart shown in the accompanying drawings is only an example, and does not necessarily include all contents and operations / steps, nor does it necessarily execute in the described order. For example, some operations / steps can also be decomposed, combined, or partially merged. Therefore, the actual execution order may change according to the actual situation.

[0017] It should also be understood that the terms used in this specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in this specification of the present application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0018] It should be further understood that the term "and / or" used in this specification of the present application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Please refer to Figure 1 , an embodiment of the intelligent sorting and circular regeneration system 10 for construction waste in the embodiment of the present application includes:

[0019] The acquisition subsystem 11 is used to install a sensor array on the crusher working surface and the edge of the vibrating screen mesh of the construction waste separation device, and acquire a characteristic data set reflecting the wear state of the device;

[0020] In this embodiment, a variety of sensing devices are reasonably arranged on the inner wall of the crusher working surface and the edge of the vibrating screen of the construction waste sorting device, including pressure sensors, triaxial acceleration sensors, displacement sensors, and depth vision camera modules, to obtain a multi-dimensional sensor array and acquire the changes in different physical quantities during the operation of the device. The pressure sensor is used to monitor the force change on the inner wall of the crusher, so as to evaluate the working state and wear degree of the crusher; the triaxial acceleration sensor captures the vibration information of the device in all directions, which helps to analyze the smoothness of the device operation and the abnormal vibration caused by wear; the displacement sensor is used to measure the relative displacement of key components, which helps to detect problems such as loosening and offset of mechanical components caused by wear; the depth vision camera module monitors the surface topography of the device by obtaining three-dimensional point cloud data. Especially at the edge of the vibrating screen, the change of mesh holes and structural damage caused by wear can be more intuitively observed through visual information. Different sampling frequencies are set for the multi-dimensional sensor array to ensure the acquisition efficiency and accuracy of various sensing data. The acquisition subsystem sets the sampling frequency according to the characteristics of various signals, making the sampling frequency of the pressure sensor moderate to capture the trend of the force change of the device, while the triaxial acceleration sensor uses a higher sampling frequency to accurately record the high-frequency vibration characteristics of the device. The sampling frequency of the displacement sensor is set according to the speed and change range of the device movement, and the depth vision camera module adjusts the frame rate according to the movement state of the device to obtain the original data set of device wear, including pressure data, vibration data, displacement data, and three-dimensional point cloud data. The original data of device wear is processed by wavelet transform for noise reduction. Wavelet transform is an effective signal processing method. By decomposing the signal at multiple scales, the signal is decomposed into components of different frequencies, and the noise components are filtered out by setting appropriate thresholds, while the effective signal characteristics are retained. When performing wavelet transform for noise reduction, appropriate wavelet basis functions are selected according to the characteristics of different sensing data. For example, for vibration data, wavelets that can finely resolve high-frequency signals are selected, while for pressure and displacement data, wavelets with better smoothness are selected to obtain the denoised data segment set. Time synchronization calibration is performed on the denoised data segment set. The acquisition subsystem aligns the data from different sources to the same time axis through time synchronization calibration technology, realizing the precise correspondence of each sensing data in the time dimension and generating a standardized data set. The features that effectively reflect the device wear state are extracted from the standardized data set, including time-domain features, frequency-domain features, and time-frequency domain features. Time-domain features include the mean, variance, peak value, peak-to-peak value, and signal energy of the signal, reflecting the change of physical quantities during the operation of the device. For example, the change of the peak-to-peak value of the vibration signal indicates the loosening or damage of mechanical components. Frequency-domain features are obtained by performing Fourier transform on the signal to obtain the energy distribution of the signal at different frequency components. For example, by analyzing the spectral characteristics of the vibration signal, the resonance phenomenon at specific frequencies caused by wear can be identified.Time-frequency domain features are realized through short-time Fourier transform or wavelet transform to jointly analyze the signal in the two dimensions of time and frequency, which is suitable for processing non-stationary signals. For example, the vibration signal of a device during load change can capture the change of frequency characteristics over time during the device wear process through time-frequency analysis. The Kalman filtering algorithm is used to perform multi-source data fusion on the multi-dimensional feature vector. Kalman filtering is a recursive estimation method that effectively eliminates noise during the process of multi-dimensional data fusion and predicts the wear state of the current device through a dynamic model. Through Kalman filtering, the data from different sensors are effectively fused in the spatial and temporal dimensions to generate a more stable and accurate feature dataset.

[0021] The simulation subsystem 12 is configured to establish a simulation model of the construction waste sorting device according to the feature dataset and generate a simulation dataset including a combination of multi-condition parameters based on the simulation model.

[0022] In this embodiment, based on the feature dataset, the simulation subsystem uses the finite element analysis method to establish a digital model of the equipment structure of the crusher and the vibrating screen. During the model establishment process, the key components of the modeling equipment are included, such as the rotor, hammer head, and lining plate of the crusher, as well as the screen mesh and support structure of the vibrating screen. To improve the accuracy of the model, the equipment structure model considers the material properties, geometric shape, connection method, and boundary conditions of the actual equipment, and discretizes the complex structure into finite elements through mesh generation. These elements reflect the stress and strain distribution inside the equipment during the force and deformation process. After the establishment of the digital model of the equipment structure is completed, the simulation subsystem constructs an equipment wear progress model according to the Archard wear theory and the cumulative damage theory. The Archard wear theory points out that the wear amount is proportional to the contact pressure, relative movement distance, and the reciprocal of the material hardness. When establishing the wear model, parameters such as the wear coefficient (K), contact pressure (P), sliding distance (L), and material hardness (H) are introduced, and the wear amount (W) is expressed as 。On this basis, combined with the cumulative damage theory, considering the cumulative effects of material fatigue and microcracks during the long-term operation of the equipment, a prediction model for equipment wear progress is established, making the model applicable not only to short-term wear prediction but also to evaluating the life attenuation trend of the equipment under long-term use. At the same time, in order to accurately simulate the interaction between construction waste and the equipment in the simulation, a discrete element model of construction waste materials is built. In the simulation subsystem, the physical properties of construction waste such as density distribution, particle size distribution, shape factor, elastic modulus, and friction coefficient are parameterized to form a discrete element model of construction waste materials. The discrete element method regards construction waste as a large number of interacting particle systems. By defining the contact mechanics model and collision recovery coefficient between particles, it simulates the movement, collision, crushing, and screening processes of materials in crushers and vibrating screens. And by introducing particle shape factors (such as spherical, flaky, prismatic, etc.), it more realistically reproduces the stress and movement behaviors of different types of construction waste inside the equipment, ensuring the credibility of the simulation results. The discrete element method is used to simulate the dynamic contact process of the discrete element model of construction waste materials and the digital model of the equipment structure. During the simulation process, through the local stress distribution and contact time data generated by the contact between particles and the equipment surface, the wear degree of the equipment surface material caused by the impact and friction of construction waste is analyzed. The simulation subsystem generates stress distribution nephograms, contact wear heat maps, and material damage accumulation curves of the equipment under different working conditions through multi-step iterative calculations. In order to achieve the full-process automation of simulation calculations and multi-condition evaluation, the simulation subsystem constructs a complex interactive process simulation system, which includes functional modules such as a material input module, a crushing dynamics module, a screening dynamics module, a wear calculation module, an equipment response module, a parameter mapping module, and a data output module. The material input module imports the parameters of construction waste materials with different characteristics into the simulation environment; the crushing dynamics module simulates the collision and crushing behaviors of materials in the crusher; the screening dynamics module is used to calculate the screening efficiency and particle distribution of materials in the vibrating screen; the wear calculation module dynamically evaluates the wear amount of key components of the equipment based on the Archard wear model and the cumulative damage model; the equipment response module provides real-time feedback on the deformation and stress response of the equipment during the force application process through finite element analysis; the parameter mapping module corresponds the equipment operation parameters with the wear characteristics to achieve quantitative analysis of the equipment wear state; the data output module outputs the equipment wear data, material distribution data, and equipment stress response data obtained from the simulation calculations as a standardized simulation data set. The simulation subsystem designs multiple groups of different working condition parameter combinations by the method of controlling variables. In different simulation experiments, the operation parameters of the equipment (such as crusher speed, vibrating screen frequency, feeding speed, etc.) and material characteristic parameters (such as density, particle size, shape factor of construction waste, etc.) are adjusted respectively. By comparing and analyzing the wear degree of the equipment under different working conditions, a simulation data set including equipment operation parameters, material characteristic parameters, and the corresponding equipment wear degree is generated.This dataset contains wear data of the equipment under normal operating conditions, and also includes wear prediction results under extreme conditions.

[0023] The measured subsystem 13 is used to collect small-sample measured data and calibrate the wear grade under the actual construction waste sorting conditions, and obtain a labeled measured wear dataset.

[0024] In this embodiment, multiple target devices are selected as monitoring objects on the actual operating construction waste sorting line. These target devices include key devices such as crushers, vibrating screens, and conveying devices. The measured subsystem configures sensors according to the installation scheme of the sensor array, including installing pressure sensors, triaxial acceleration sensors, displacement sensors, and depth vision camera modules at key parts of the devices, such as the inner wall of the crusher working surface and the edge of the vibrating screen mesh, to obtain the measured monitoring system. After completing the sensor configuration, the measured subsystem sets multiple groups of different typical working condition parameters for the measured monitoring system, such as the operating speed, feed rate, material type, and device load of the device, to cover various working states encountered by the construction waste sorting device in actual production. The measured subsystem continuously monitors the device under different working conditions and real-time collects the surface topography data of the device and various sensor data, including pressure data, vibration data, displacement data, and three-dimensional point cloud data, to ensure the comprehensiveness and multi-dimensional characteristics of data collection. To ensure the continuity and accuracy of the data, the system sets a regular recording mechanism during the operation of the device, such as automatically recording data through periodic triggering or event-driven methods, to obtain the measured original data under multiple working condition conditions. Time domain features are extracted from the measured original data to obtain the time series characteristics of signal changes during the operation of the device. Time domain feature extraction includes the mean, variance, peak value, root mean square value, skewness, kurtosis of the signal, and the envelope of the signal, etc. These time domain features reflect the dynamic changes during the device wear process. At the same time, to capture the frequency characteristics and multi-scale change information of the device wear state, the measured subsystem performs fast Fourier transform and wavelet packet decomposition on the original data to obtain the frequency domain features and time-frequency domain feature sets. The fast Fourier transform converts the time domain signal into a frequency domain signal and analyzes the energy distribution of the signal at different frequencies. Wavelet packet decomposition provides the characteristic information of the signal in both the time and frequency dimensions through multi-scale decomposition, which helps to identify the change characteristics of non-stationary signals during the device wear process. The three-dimensional point cloud data obtained by the depth vision camera module is used to perform refined analysis on the surface topography of the device. By calculating the surface roughness index in the point cloud data, the roughness change caused by wear on the device surface is quantified; by analyzing the principal curvature distribution of the point cloud model, the local details of the surface topography change of the device are detected, such as pits and scratches caused by wear; by calculating the area and depth distribution of the wear area, the material loss situation of the device is quantified. For example, for the crusher hammer head or vibrating screen mesh, the distribution of the wear depth is used to evaluate the service life of the device and determine when the device needs to be replaced or repaired. These features extracted from the three-dimensional point cloud data are classified as the depth feature set. The most representative features are selected from the time domain feature set, frequency domain and time-frequency domain feature sets, and depth feature sets. During the feature selection process, the measured subsystem uses statistical analysis, principal component analysis, or information gain-based methods to ensure that the finally selected features provide the most effective input for the device wear prediction model.To convert these feature data into the actual equipment wear degree, the measurement subsystem calibrates the equipment wear degree by combining visual inspection and laser scanning data. During the visual inspection, the wear level of the equipment is initially evaluated based on the wear conditions on the equipment surface (such as color change, cracks, deformation, etc.). Meanwhile, the laser scanning equipment obtains the changes in the geometric shape of the equipment surface through high-precision scanning data, compares it with the three-dimensional point cloud data, and provides a quantitative basis for the evaluation of the wear level. For example, by analyzing the profile lines of the equipment surface obtained by scanning, measuring the depth changes in the worn area, and comparing these data with the standard wear levels, the wear degree of the equipment is finally divided into different levels (such as mild, moderate, severe wear, etc.).

[0025] The training subsystem 14 is used to input the simulation data set and the labeled measured wear data set into the adaptive domain adversarial neural network for training to obtain a prediction model for the equipment wear degree;

[0026] In this embodiment, in the design subsystem, a neural network overall architecture including a feature extractor, a domain classifier, and a wear predictor is constructed. The feature extractor, as the core part of the network, maps the input multi-dimensional feature data to a high-dimensional feature space. Its internal structure consists of multiple fully connected layers, batch normalization layers, and non-linear activation functions (such as ReLU or Leaky ReLU). The fully connected layers extract deep features of the data through linear transformation. The batch normalization layers are used to stabilize the training process and avoid problems such as gradient vanishing or explosion. The activation function introduces non-linearity, enabling the model to learn complex feature representations. The multi-level feature extractor can effectively extract common features from simulation data and measured data and retain the unique information of the equipment wear state. After completing the network structure design, the adversarial unit combines the wear prediction task and the domain adversarial task by defining a multi-task loss function to achieve multi-objective optimization of the model. The wear prediction loss calculates the difference between the predicted value of the equipment wear degree and the actual labeled value through the mean squared error or cross-entropy loss function. The goal of this loss term is to improve the prediction accuracy of the model for the labeled measured data. The domain adversarial loss predicts the source of the input features (simulation data or measured data) through the domain classifier and uses the gradient reversal layer to implement adversarial training. The gradient reversal layer does not change the data flow during forward propagation but reverses the gradient sign during backpropagation, making the feature extractor tend to generate feature representations that "confuse" the domain classifier during training, thereby reducing the difference in the feature distributions between simulation data and measured data and achieving the effect of domain adaptation. To ensure the stability and generalization ability of model training, the partitioning unit partitions the dataset. The simulation dataset is divided into a first training set and a validation set to ensure that the network conducts sufficient feature learning on a large dataset and evaluates the generalization ability of the model on the validation set. At the same time, the labeled measured wear dataset is divided into a second training set and a test set. The second training set is used for the joint training of the model, and the test set is used to evaluate the wear prediction ability of the model under real working conditions. During the partitioning process, ensure the random allocation of data and the consistency of data distribution to avoid the risks of data leakage and overfitting. After completing the dataset partitioning, the joint unit pre-trains the neural network using the first training set. The pre-training stage mainly focuses on the basic feature learning ability of the feature extractor. Through training with a large amount of simulation data, the feature extractor learns to extract general features related to the equipment wear state. When the pre-training is completed, the joint unit extends the training of the model to the labeled measured wear data. By inputting the second training set into the network, the joint training process is started. During the joint training, the model continues to optimize the wear prediction task and gradually introduces the domain adversarial task. Through a dynamic weight adjustment strategy, the weight of the domain adversarial loss is gradually increased, enabling the model to maintain adaptability to the features of simulation data while learning the features of measured data.The dynamic weight adjustment strategy adopts the method of linear increase or exponential increase. In this way, the domain adversarial task gradually occupies a larger proportion in the later stage of training, ensuring that the model achieves the domain adversarial effect while maintaining the prediction accuracy. During the entire training process, the optimization objective of the model is to simultaneously minimize the wear prediction loss and maximize the error rate of the domain classifier. This adversarial training mechanism enables the feature extractor to continuously adjust the output features to "confuse" the domain classifier, ultimately achieving the seamless fusion of the feature distributions of the simulation data and the measured data. After the training process is completed, the model validates its prediction accuracy and generalization ability through the test set, obtaining a device wear degree prediction model with good domain adaptability and high-precision wear prediction ability.

[0027] The joint unit initializes the parameters of the adaptive domain adversarial neural network. This process involves initializing all the weight and bias parameters of the feature extractor, domain classifier, and wear predictor. The Xavier initialization or He initialization method is adopted to keep the initial parameters of the model within a reasonable distribution range, thus avoiding the initial state of the network deviating too much from the optimal solution and affecting the convergence speed of training and the performance of the model. During the initialization process, an initial inversion coefficient is set for the gradient reversal layer to minimize the influence of the domain adversarial loss in the initial stage of training, so that the model can preferentially learn the effective features in the wear prediction task. After completing the initialization of the model parameters, the joint unit inputs the first training set into the adaptive domain adversarial neural network in batches, and only updates the parameters of the feature extractor and wear predictor. The parameters of the domain classifier remain frozen at this stage. The core objective of this pre-training stage is to enable the feature extractor to learn the effective features related to the equipment wear state by minimizing the wear prediction loss without domain adversarial interference. By using mini-batch stochastic gradient descent or the Adam optimizer, the model updates the parameters of the feature extractor and wear predictor in each batch of training, gradually reducing the validation loss. When the validation loss of the model shows stability on the validation set during the training process, it means that the feature extractor already has the preliminary ability to extract wear features, and at this time, the pre-training model is considered to be completed. After completing the pre-training model, the joint unit constructs a data loader, marks the first training set as source domain data, marks the second training set as target domain data, and mixes the source domain and target domain data to generate a batch data stream by setting the same sampling ratio. The mixed batches in this process ensure that each batch contains samples from simulation data (source domain data) and measured data (target domain data), enabling the model to learn the features of both data distributions simultaneously in the subsequent adversarial training. The data loader uses random shuffling and batch balancing techniques to make the data distributions of the source domain and target domain in each batch as uniform as possible, thus preventing the model from being biased towards a certain data domain during the training process. In the domain adversarial training stage, the joint unit inputs each batch of data in the domain adversarial training data stream into the adaptive domain adversarial neural network. The input data is converted into high-dimensional feature vectors by the feature extractor, and these feature vectors are then simultaneously input into the domain classifier and wear predictor to obtain the domain prediction result and wear prediction result respectively. The goal of the domain classifier is to distinguish whether the features of the input data come from the source domain or the target domain, while the wear predictor predicts the wear degree of the equipment. In this process, the feature extractor plays an "adversarial role". Under the action of the gradient reversal layer, the feature extractor has to ensure the accuracy of wear prediction and "confuse" the domain classifier so that the domain classifier cannot accurately determine the source domain of the data, thereby achieving the domain adaptation effect. When calculating the loss function, the joint unit combines the wear prediction loss and the domain adversarial loss to calculate the total loss.The wear prediction loss uses the mean squared error or cross-entropy loss function to measure the deviation between the wear degree predicted by the model and the actual label, while the domain adversarial loss is calculated based on the cross-entropy loss between the output of the domain classifier and the true domain label. In the loss function, the weight coefficient (λ) of the domain adversarial loss is not fixed but dynamically adjusted according to the current training progress. In the initial stage of model training, the value of λ is small, enabling the model to prioritize optimizing the wear prediction task. As the training progresses, the value of λ gradually increases through a linear or exponential growth strategy, causing the model to gradually increase its attention to the domain adversarial task. During the training process, every 50 training rounds, the joint unit decays the learning rate to 0.1 of the original value. By adopting the learning rate decay strategy, it effectively prevents the model from oscillating when approaching convergence and helps the model find a better local minimum. At the same time, after each learning rate decay, the model calculates the wear prediction error and the domain classification confusion matrix on the test set, and judges the training status of the model through evaluation metrics. When the wear prediction error is less than the set threshold and the domain classification accuracy is close to the random guessing level (i.e., close to 50%), it means that the feature extractor has successfully "confused" the domain classifier, achieving an ideal domain adaptation effect. When the model shows a stable low wear prediction error on the test set, and at the same time the domain classifier cannot effectively distinguish source domain and target domain data, the joint unit obtains a cross-domain transfer model after dynamic adjustment. This model can accurately predict the wear degree of the equipment under actual working conditions and also has strong generalization ability, capable of migrating the features learned from simulation data to measured data, thus realizing the intelligence, automation, and high efficiency of equipment wear monitoring and prediction in the intelligent sorting and recycling system of construction waste.

[0028] The output subsystem 15 is used to identify wear fault features through the equipment wear degree prediction model, and based on the wear fault features, optimize the operating parameters of the construction waste sorting device and output a fault closed-loop response plan.

[0029] In this embodiment, the filtering unit collects the point cloud data of the working surface of the construction waste sorting device through the depth vision camera module, which reflects the three-dimensional topography of the equipment surface, including the geometric characteristics of the crusher working surface, the vibrating screen mesh, and other key components. The filtering unit performs adaptive voxel filtering on the original three-dimensional point cloud data. By dividing the space into a grid of cubes (voxels) of the same size, multiple points within each voxel are simplified into a representative point, effectively reducing the amount of point cloud data and being able to eliminate noise and redundant points, improving the efficiency and stability of data processing. After the voxel filtering is completed, the filtering unit estimates the normal vectors of the point cloud data. By calculating the normal vector directions of each point on the point cloud surface, local geometric features (such as curvature, flatness, edge characteristics, etc.) are obtained, and a regularized voxel grid and local geometric features are obtained. The feature extraction unit uses a feature extraction algorithm to convert the spatial geometric information of the point cloud into a high-dimensional feature vector, which contains the geometric shape information of the point cloud and integrates the spatial relationship and surface characteristics between the point clouds. The dimensionality reduction processing unit uses a variational autoencoder to perform dimensionality reduction on the high-dimensional feature vector of the point cloud. The variational autoencoder compresses the high-dimensional input features into a low-dimensional latent space through an encoder-decoder structure. In this process, by introducing a KL divergence regularization term, the model retains the main distribution information of the feature data during dimensionality reduction, while avoiding the distribution in the latent space from being too concentrated or sparse. After the dimensionality reduction by the variational autoencoder, the low-dimensional feature vector retains the main information of the wear fault features. The clustering unit combines the prediction results of the equipment wear degree prediction model and performs a wear mode clustering analysis on the wear fault features. The clustering analysis uses unsupervised learning algorithms such as K-means, DBSCAN, or Gaussian mixture model to divide the wear features on the equipment surface into different wear modes, such as mild wear, uniform wear, local severe wear, abnormal wear, and other types. During the clustering process, the model assigns data points to the closest cluster center according to the similarity of the wear fault features, generating a wear fault risk assessment result. By analyzing the distribution of various wear modes, the current wear state and potential fault risks of the equipment are evaluated, and the stability and safety of the equipment during future operation are predicted. The construction unit further constructs a Takagi-Sugeno (T-S) fuzzy model to describe the nonlinear dynamic characteristics of the construction waste sorting device. The T-S fuzzy model is a hybrid model that can decompose a complex nonlinear system into a set of linear sub-models and realizes the accurate modeling of the equipment operating state through a fuzzy logic inference mechanism. When constructing the model, a multi-rate sampling framework is designed to reasonably configure the sampling frequencies of the grouped sensors, enabling different types of sensors to dynamically adjust the sampling rate in different working modes, thereby ensuring the timeliness and effectiveness of the sampled data. The construction unit determines the feedback control gain by solving a linear matrix inequality optimization problem, and this process can minimize the control error and ensure the stability of the equipment control system.After the feedback control gain is determined, the construction unit obtains the parameter optimization basic strategy. Based on the wear failure risk assessment result and the parameter optimization basic strategy, the solution unit combines with the Bayesian optimization system to solve the optimal operation parameter combination of the construction waste sorting device. During the Bayesian optimization process, the model constructs a surrogate model (such as a Gaussian process model) based on the current device state and historical data, predicts the operation effect of the device under different parameter combinations, and calculates the corresponding objective function values (such as device wear, sorting efficiency, energy consumption, etc.). Through iterative updates, the Bayesian optimization continuously selects the parameter combination that can maximize the acquisition function, so as to find the optimal operation parameters with the least number of experiments. After obtaining the optimal parameter combination, the solution unit applies these parameters to the construction waste sorting device, dynamically adjusts the feeding rate, crushing speed, screening frequency and screening amplitude, and realizes the adaptive operation optimization of the device. The solution unit feeds back the optimal parameter setting to the device control system through the controller to form a closed-loop control path, so that the device automatically adjusts the operation parameters according to the real-time monitored wear state and failure risk, ensures that the device operates in the best state, thereby prolonging the service life of the device, improving the sorting efficiency of construction waste, reducing the maintenance cost, and realizing the efficient, safe and stable operation of the intelligent sorting and recycling system.

[0030] The dimensionality reduction processing unit constructs a variational autoencoder model composed of an encoder and a decoder. The encoder consists of three fully connected layers, and the number of nodes is 256, 128, and 64 in sequence. This decreasing node design can gradually compress the dimension of the input data, enabling the feature vector to extract more compact and key feature information in each layer. After each fully connected layer, a batch normalization layer and a ReLU activation function are connected. The batch normalization layer can stabilize the model training process and prevent gradient vanishing or explosion, while the ReLU activation function enables the model to have the ability to learn complex features by introducing a non-linear transformation. After completing the encoder structure design, the high-dimensional feature vector of the point cloud is input into the variational autoencoder for feature processing. The encoder maps the input high-dimensional feature data layer by layer to a lower-dimensional space, and the feature vector output by the last fully connected layer is called the encoded output feature. The key difference between the variational autoencoder and the traditional autoencoder is that the variational autoencoder does not directly output a deterministic latent representation, but maps the encoded output feature to a mean vector (μ) and a variance vector (σ²) respectively. These vectors define a multi-dimensional normal distribution in the latent space. In this process, the variational autoencoder calculates the mean and variance of the features through the fully connected layer, making the latent representation not only compact but also able to capture the uncertainty of the data distribution. To perform effective sampling in the latent space, the dimensionality reduction processing unit introduces the reparameterization technique by calculating , where ε is the random noise sampled from the standard normal distribution. This reparameterization technique enables the backpropagation of gradients and keeps the latent representation continuous and smooth during model training, resulting in the latent representation after dimensionality reduction. The dimensionality reduction processing unit imposes an orthogonality constraint on the latent representation after dimensionality reduction, making the correlation between latent features as close to zero as possible, which helps to eliminate feature redundancy and improve the generalization ability of the model. The specific implementation method is to calculate the difference between the covariance matrix Σ of the latent variables and the identity matrix I and minimize its norm (such as the Frobenius norm). By adding the orthogonality constraint loss term to the total loss function, the model will automatically adjust the parameters during training to make the latent features satisfy the orthogonality requirement as much as possible, obtaining orthogonal latent features. The dimensionality reduction processing unit maps the orthogonal latent features back to the original feature space through the decoder to obtain the reconstructed feature vector. The decoder structure is similar to the encoder, consisting of multiple fully connected layers and activation functions. By gradually amplifying the dimension of the latent features, the low-dimensional latent representation is restored to a high-dimensional feature vector with the same dimension as the original input features. To evaluate the reconstruction effect, the dimensionality reduction processing unit calculates the mean square error between the reconstructed feature vector output by the decoder network and the original input, which is used as the reconstruction loss. The goal of the reconstruction loss is to make the reconstructed features as close to the original features as possible, thus ensuring that as much original information as possible is retained during dimensionality reduction. At the same time, to maintain the distribution characteristics of the latent representation, the variational autoencoder calculates the KL divergence between the latent variable distribution and the standard normal distribution, which is used as the regularization loss. The KL divergence can quantify the difference between two probability distributions. By minimizing the KL divergence, the variational autoencoder makes the distribution of the latent variables close to the standard normal distribution. This regularization process prevents the latent representation from being too concentrated or too sparse and enhances the generalization ability of the model on new data. The reconstruction loss and the regularization loss are combined to form the total loss function. During the model training process, the dimensionality reduction processing unit performs iterative optimization based on the total loss function. Through backpropagation and gradient descent methods, the parameters of the model are continuously updated to minimize the total loss function. After multiple rounds of iterative optimization, the model gradually converges and outputs the wear fault features.

[0031] Optionally, the acquisition subsystem 11 is specifically configured to:

[0032] Install pressure sensors, triaxial acceleration sensors, displacement sensors, and depth vision camera modules on the inner wall of the crusher working surface and the edge of the vibrating screen mesh of the construction waste sorting device to obtain a multi-dimensional sensor array;

[0033] Set different sampling frequencies for the multi-dimensional sensor array to collect pressure data, vibration data, displacement data, and three-dimensional point cloud data respectively to obtain the original equipment wear data;

[0034] Perform wavelet transform denoising on the original equipment wear data to obtain a denoised data segment set, and perform time synchronization calibration on the denoised data segment set to obtain a standardized data set;

[0035] Extract time-domain features, frequency-domain features, and time-frequency domain features from the standardized data set to obtain a multi-dimensional feature vector, and use the Kalman filter algorithm to perform multi-source data fusion on the multi-dimensional feature vector to obtain a feature data set reflecting the current wear state of the equipment.

[0036] In this embodiment, various types of sensors are installed on the inner wall of the crusher working surface and the edge of the vibrating screen mesh, including pressure sensors, triaxial acceleration sensors, displacement sensors, and depth vision camera modules, to construct a multi-dimensional sensor array to achieve comprehensive acquisition of different physical quantities during the operation of the equipment. The pressure sensor is used to measure the change in the force on the inner wall when the equipment crushes materials, and the obtained pressure data is denoted as , where represents pressure, is time; the triaxial acceleration sensor obtains the vibration data of the equipment in , , three directions, and the vibration data is denoted as , where represents acceleration; the displacement sensor is used to monitor the displacement change of the key components of the equipment, and the displacement data is denoted as , where represents displacement; the depth vision camera module collects the three-dimensional point cloud data on the surface of the equipment, and the data is represented as , where represents three-dimensional vision data. During the data acquisition process, a differential sampling frequency is set for the multi-dimensional sensor array to ensure accurate acquisition of various signals within their respective specific frequency ranges, and an original data set of equipment wear is obtained, including , , and . Perform wavelet transform denoising on the original equipment wear data. Wavelet transform effectively separates the noise components in the signal by decomposing the signal into sub-bands of different frequencies. For example, for the vibration signal perform wavelet decomposition, and represent the signal as coefficients of different scales and frequencies:

[0037]

[0038] Among them, is the th wavelet coefficient, is the wavelet basis function, is the number of decomposition scales. In the denoising process, the low-amplitude noise coefficients are set by setting a threshold Set it to zero, then perform wavelet reconstruction to obtain the vibration data after noise reduction , where represents the acceleration data after noise reduction. Other data such as 、 and are denoised using the same method to obtain 、 and represent the pressure, displacement, and three-dimensional visual data after noise reduction respectively. Time synchronization calibration is performed on the denoised data segment set, and interpolation methods (such as linear interpolation or spline interpolation) are used to align the sampled data at different frequencies to a unified time axis , obtaining a standardized data set. For example, the pressure data after interpolation is , the vibration data is , the displacement data is , and the three-dimensional point cloud data is , where is the unified time point after standardization. Time-domain, frequency-domain, and time-frequency domain features are extracted from the standardized data set to form a multi-dimensional feature vector. Time-domain features include the mean , standard deviation , peak value , and root mean square value of the signal, and these features are calculated by the following formulas:

[0039]

[0040]

[0041] where represents the number of sampling points, is the signal value at time . Frequency-domain features are obtained through fast Fourier transform and mainly include the central frequency and bandwidth of the spectrum:

[0042]

[0043] where is the spectrum of the signal, is the frequency. Time-frequency domain features are obtained through short-time Fourier transform or continuous wavelet transform and can describe the dynamic change characteristics of the signal in time and frequency. The Kalman filtering algorithm is used to perform multi-source data fusion on the multi-dimensional feature vector. Assuming the multi-dimensional feature vector is , the current feature state of the device is predicted through the state equation:

[0044]

[0045] Among them, is the state transition matrix, is the control input matrix, is the control input. In the observation equation, the actually measured eigenvector is combined with the predicted eigenvector, and the fused eigenvector is updated through the Kalman gain :

[0046]

[0047] Among them, is the observation matrix, is the Kalman gain, which can dynamically adjust the weight of feature fusion, and finally obtain a feature dataset reflecting the current wear state of the equipment.

[0048] Optionally, the simulation subsystem 12 is specifically used for:

[0049] Based on the feature dataset, a digital model of the equipment structure of the crusher and the vibrating screen is established using the finite element analysis system;

[0050] According to the Archard wear theory and the cumulative damage theory, an equipment wear progress model of the wear coefficient and the equipment material hardness is established;

[0051] The density distribution, particle size distribution, shape factor, elastic modulus and friction coefficient of different types of construction waste are parameterized to obtain a discrete element model of construction waste materials;

[0052] The dynamic contact process of the discrete element model of construction waste materials and the digital model of the equipment structure is simulated using the discrete element method to obtain local stress distribution and contact time data;

[0053] A joint simulation environment including a material input module, a crushing dynamics module, a screening dynamics module, a wear calculation module, an equipment response module, a parameter mapping module and a data output module is constructed to obtain an interactive process simulation system;

[0054] By using the control variable method, multiple groups of different combinations of working condition parameters are designed, and the interactive process simulation system is simulated to obtain a simulation dataset including equipment operation parameters, material characteristic parameters and corresponding equipment wear degree data.

[0055] In this embodiment, a finite element analysis system is used to establish digital models of crushers and vibrating screens. By analyzing the structural characteristics of the equipment, including the rotor, hammers, liners of the crusher, and the screen mesh and support structure of the vibrating screen, a 3D model of the equipment is drawn using CAD software, and then the model is imported into the finite element analysis system for mesh division processing. The equipment structure is discretized into a finite number of elements, and a mathematical model of the equipment during the stress and deformation processes is established through the connection relationships of nodes and elements. The finer the mesh, the higher the accuracy of the simulation results, but the computational cost will also increase accordingly. For example, in the area of the crusher hammer head, due to high impact force and friction, high-precision hexahedral mesh division is used to make the stress and strain distribution in this area more accurate. After establishing the digital model of the equipment structure, an equipment wear progress model is constructed based on Archard's wear theory and cumulative damage theory. Archard's wear theory states that the wear volume is proportional to the contact pressure , the sliding distance , and the reciprocal of the material hardness . Its calculation formula is:

[0056]

[0057] where is the wear coefficient, indicating the wear sensitivity of the material under specific working conditions; is the stress condition on the equipment surface; is the relative sliding distance; is the hardness of the material; is the wear volume. In practical applications, the wear coefficient is obtained through laboratory wear tests or material databases, and combined with the local pressure and sliding distance calculated in the finite element analysis, the wear degree of the key components of the equipment under different operating conditions is predicted. At the same time, the cumulative damage theory is used to evaluate the fatigue damage of the equipment material under multiple stress cycles, and its damage accumulation model is expressed as:

[0058]

[0059] where represents the degree of cumulative damage, is the actual number of cycles under the th stress amplitude, is the failure life of the material under this stress condition. When When it reaches 1, the equipment component is considered to have suffered fatigue failure. By combining the wear model with the cumulative damage model, the wear progress prediction of the equipment during long-term operation is realized. At the same time, in order to accurately simulate the interaction between construction waste and the equipment, a discrete element model of construction waste materials is constructed. The material properties of construction waste include density distribution , particle size distribution , shape factor , elastic modulus and friction coefficient . These parameters are obtained through experimental tests or sample analyses. For example, the cumulative distribution function of the particle size distribution is obtained through a screening test , the elastic modulus is determined through a triaxial compression test , and the static friction coefficient of the material is obtained through a contact friction test . Based on these parameters, in the discrete element software, a discrete element model of the material is constructed through a particle generation tool. Each particle is given specific physical properties and mechanical behavior models, such as the restitution coefficient and rolling friction coefficient during particle collision. These models can simulate the real movement state of the material inside the equipment. After the construction of the digital model of the equipment structure and the discrete element model of the construction waste materials is completed, the coupling technology of the discrete element method and the finite element method is used to simulate the dynamic contact process between the construction waste materials and the equipment. During the simulation, when each particle contacts the surface of the equipment, the system calculates the local stress distribution and contact time at the contact point. The local stress calculation formula is:

[0060]

[0061] where is the contact force is the contact area. Through dynamic simulation, the stress distribution map and contact time data on the surface of the equipment at different times and positions are obtained, and a combined simulation environment containing multiple functional modules is constructed, including a material input module, a crushing dynamics module, a screening dynamics module, a wear calculation module, an equipment response module, a parameter mapping module, and a data output module. For example, the material input module automatically generates an input particle flow according to the material property parameters of the construction waste. The crushing dynamics module calculates the crushing efficiency by analyzing the impact process between the material and the crusher hammer head. The screening dynamics module simulates the movement trajectory and screening efficiency of the material in the vibrating screen. The wear calculation module calculates the wear amount of the equipment by combining the stress distribution and sliding distance data. The equipment response module analyzes the force and deformation conditions of the equipment under load changes. The parameter mapping module maps the equipment operation state to the wear characteristics. The data output module outputs the simulation results as a standardized data set, including equipment operation parameters , material characteristic parameters and the corresponding degree of equipment wear :

[0062]

[0063] Among them, is an equipment wear prediction model. By using the control variable method, multiple groups of different working condition parameter combinations are designed, such as adjusting the feeding rate of the equipment , the rotation speed of the crusher , the screening frequency and the screening amplitude , and the simulation calculation is carried out on the interactive process simulation system, and the degree of equipment wear under each working condition is evaluated one by one , and finally a simulation data set including equipment operation parameters , material characteristic parameters and the corresponding equipment wear degree data is formed.

[0064] Optionally, the measured subsystem 13 is specifically used for:

[0065] Select multiple target devices on the actual operating construction waste separation line as monitoring objects, configure sensors according to the same installation scheme as the sensor array, and obtain a measured monitoring system;

[0066] Set multiple groups of different typical working condition parameters for the measured monitoring system, and continuously operate under different working conditions and regularly record the surface topography data and sensor data of the equipment to obtain measured original data;

[0067] Extract time-domain features from the measured original data to obtain a time-domain feature set, and perform fast Fourier transform and wavelet packet decomposition on the measured original data to obtain frequency-domain and time-frequency domain feature sets;

[0068] Calculate surface roughness indexes, principal curvature distributions, areas and depth distributions of worn areas using the three-dimensional point cloud data obtained by the depth vision camera module to obtain a depth feature set;

[0069] Select the most representative features from the time-domain feature set, frequency-domain and time-frequency domain feature sets and depth feature sets, and combine visual inspection and laser scanning data to calibrate the equipment wear degree level to obtain a labeled measured wear data set.

[0070] In this embodiment, multiple target devices are selected as monitoring objects on the actual construction waste sorting line, including key devices such as crushers, vibrating screens, and conveyor belts. These devices are prone to the impact and friction of materials and mechanical fatigue caused by long-term operation during the sorting process. Therefore, the wear problem is particularly prominent. Sensors are configured on these target devices according to the same installation scheme as the sensor array to form a measured monitoring system. The sensors include pressure sensors, triaxial acceleration sensors, displacement sensors, and depth vision camera modules. The pressure sensor is installed on the inner wall of the crusher working surface to measure the change in the force on the inner wall when the device crushes materials and obtain pressure data , where represents pressure, is time; the triaxial acceleration sensor is installed at the key connection points of the device to capture vibration data during the operation of the device , and the components of the acceleration in , , directions are respectively , , ; the displacement sensor is mainly installed at the edge of the vibrating screen mesh to monitor the movement trajectory of the screen in real time and obtain displacement data ; the depth vision camera module is fixed at a suitable position outside the device to obtain three-dimensional point cloud data of the device surface through non-contact laser scanning or structured light technology , which is used to analyze the change in the surface topography of the device. After the measured monitoring system is built, by setting multiple groups of different typical working condition parameters, the device is continuously operated under various actual operating conditions to comprehensively evaluate the wear condition of the device under different working states. The typical working condition parameters include the feeding rate , the crusher speed , the screening frequency , and the screening amplitude . Through the combined changes of these parameters, the operating states of the device under low load, high load, and extreme working conditions are simulated. Under these different working conditions, the measured monitoring system continuously collects data at the preset sampling frequency and regularly records the device surface topography data and sensor data to obtain the measured raw data, including pressure , acceleration , displacement and three-dimensional point cloud data. Time domain features are extracted from the measured raw data to obtain a time domain feature set. Time domain features are characteristic parameters of a signal in the time domain, including the mean value , the standard deviation , the peak value and the root mean square value . These features are calculated by the following formulas:

[0071]

[0072]

[0073] Among them, represents the number of data sampling points, is the signal value at time . For the acceleration signal , the time-domain characteristics of , , are calculated respectively. These characteristics reflect the vibration intensity and stability of the device in different directions during operation. At the same time, the measured raw data is subjected to fast Fourier transform and wavelet packet decomposition to obtain the frequency-domain and time-frequency-domain feature sets. The frequency-domain features mainly include the center frequency of the signal spectrum, the main frequency component, and the bandwidth :

[0074]

[0075] Among them, is the spectrum of the signal, is the frequency. In frequency-domain analysis, for example, if a peak at a specific frequency in the spectrum of the vibration signal is significantly higher than other frequencies, it indicates that resonance has occurred in some components of the device, which is a typical feature of mechanical wear or looseness. Through wavelet packet decomposition, the signal is decomposed into sub-signals with different frequency bandwidths. By analyzing the distribution of these sub-signals on the time axis, time-frequency-domain features are obtained, which helps to capture the frequency response characteristics of the device under dynamic loads. The three-dimensional point cloud data obtained by using the depth vision camera module is used to calculate the surface roughness index , the main curvature distribution , the area of the worn area, and the depth distribution to obtain the depth feature set. For example, the surface roughness is expressed as:

[0076]

[0077] Among them, is the height value of points in the point cloud data, is the average surface height, is the number of data points. The area of the worn area is obtained by calculating the projected area of the damaged surface in the three-dimensional model, while the depth distribution Then, by analyzing the depression depth of the worn area and combining the fine topological data obtained by laser scanning, the quantitative analysis of the surface topography change of the device is realized. The most representative features are selected from the time-domain feature set, frequency-domain and time-frequency domain feature sets, and depth feature set through feature selection methods (such as principal component analysis or mutual information method). Combining visual inspection and laser scanning data, the wear degree of the device is calibrated to obtain a labeled measured wear data set.

[0078] Optionally, the training subsystem 14 further includes:

[0079] A design unit for designing an adaptive domain adversarial neural network structure including a feature extractor, a domain classifier, and a wear predictor, where the feature extractor consists of multiple fully connected layers, batch normalization layers, and activation functions;

[0080] An adversarial unit for defining a multi-task loss function including a wear prediction loss and a domain adversarial loss, and implementing adversarial training through a gradient reversal layer to obtain a training optimization target;

[0081] A partitioning unit for partitioning the simulation data set into a first training set and a validation set, and partitioning the labeled measured wear data set into a second training set and a test set;

[0082] A joint unit for performing network pre-training using the first training set, then performing joint training in combination with the second training set, and gradually increasing the domain adversarial loss weight using a dynamic weight adjustment strategy to obtain a device wear degree prediction model.

[0083] In this embodiment, a network structure including a feature extractor, a domain classifier, and a wear predictor is designed. Through a multi-task loss function and an adversarial training method, accurate prediction of the device wear degree and seamless migration between simulation data and measured data are realized. The whole process includes the collaborative work of the design unit, the adversarial unit, the partitioning unit, and the joint unit to ensure that the model can not only accurately predict the device wear state but also adapt to the distribution differences between different data domains. In the design unit, a feature extractor is constructed. The feature extractor consists of multiple fully connected layers, batch normalization layers, and activation functions. Its goal is to convert the input multi-dimensional feature vector into a latent feature representation , where is the dimension of the input feature, is the dimension of the extracted latent feature. The feature extraction process is expressed as:

[0084] ;

[0085] where, is the output feature of the th layer, is the weight matrix of the fully connected layer, is the bias vector, is the activation function (e.g., ReLU function ), represents the output dimension of the th layer. The role of the batch normalization layer is to calculate the mean of the feature distribution at each layer and the standard deviation , and then normalize the features:

[0086] ;

[0087] This processing method can accelerate the model convergence speed and avoid the problems of gradient vanishing or gradient explosion. The latent features extracted by the feature extractor are simultaneously input into the domain classifier and the wear predictor. The role of the domain classifier is to distinguish whether the input features come from simulation data or measured data, which is the key component of domain adversarial training. The domain classifier consists of two to three fully connected layers and a sigmoid activation function, and outputs the domain prediction result , where represents originating from simulation data (source domain), represents originating from measured data (target domain). The domain classification loss is calculated using the binary cross-entropy loss function:

[0088]

[0089] where, is the true domain label (0 or 1), is the prediction result of the domain classifier. Through the gradient reversal layer, the gradient sign is reversed during backpropagation, so that the features learned by the feature extractor are more domain-invariant, achieving the effect of domain adversarial training. The wear predictor is used to predict the wear degree of the device, and the prediction result is expressed as , and the actual wear label is . The wear prediction loss is calculated through the mean square error loss:

[0090] ;

[0091] where, is the number of samples, and are the actual wear value and the predicted wear value of the th sample respectively. The adversarial unit defines the multi-task loss function including the wear prediction loss and the domain adversarial loss , and realizes adversarial training by dynamically adjusting the loss weight :

[0092]

[0093] Among them, is the domain adversarial loss weight coefficient, which is smaller initially to ensure that the model gives priority to learning the wear prediction task and gradually increases in the later stage of training, so that the model pays more attention to the domain adaptation ability. During the training process, the gradient descent method (such as the Adam optimizer) is adopted, and the gradient of the loss function with respect to the model parameters is backpropagated and the parameters are updated:

[0094] ;

[0095] Among them, is the learning rate, which controls the step size of model parameter update. In the partitioning unit, the simulation data set is partitioned into the first data set and the second data set , and the measured wear data set with labels is partitioned into the third data set and the fourth data set . During the data partitioning process, the consistency of the data distribution is ensured. For example, the data set is partitioned according to the ratio of 80%:20% to avoid overfitting and data leakage problems. The joint unit pre-trains the model through the first data set , with the goal of minimizing the wear prediction loss without introducing the domain adversarial loss, in the initial stage . When the loss of the model on the second data set tends to be stable, it means that the model has learned sufficient wear prediction features. The model is extended to the third data set for joint training. At this time, the value of the domain adversarial loss weight is dynamically increased, for example, through a linear growth strategy:

[0096]

[0097] Among them, is the current training progress (the ratio of the number of training epochs to the total number of epochs), is the adjustment coefficient. The dynamic adjustment strategy ensures that the model focuses on the wear prediction task in the early stage and enhances the domain adaptation ability in the later stage. After the joint training is completed, the wear prediction accuracy and the domain classification confusion degree of the model are evaluated on the fourth data set . The calculation formula for the wear prediction accuracy is:

[0098] ;

[0099] wherein, is the number of samples. When the model reaches the domain classification confusion state with a small wear prediction error, it is considered that the model has successfully achieved cross-domain migration of simulation data and measured data, and finally a device wear degree prediction model is obtained.

[0100] Optionally, the joint unit is specifically used for:

[0101] Initialize the parameters of the adaptive domain adversarial neural network to obtain the initial network parameters;

[0102] Input the first training set into the adaptive domain adversarial neural network batch by batch, and only update the parameters of the feature extractor and the wear predictor until the validation loss is stable to obtain a pre-trained model;

[0103] Construct a data loader, mark the first training set as source domain data, mark the second training set as target domain data, and sample them in the same proportion to form a mixed batch to obtain a domain adversarial training data stream;

[0104] For each batch of data in the domain adversarial training data stream, extract features through the feature extractor of the adaptive domain adversarial neural network, and input the extracted features into the domain classifier and the wear predictor at the same time to obtain domain prediction and wear prediction results;

[0105] Calculate the domain adversarial loss weight coefficient according to the current training progress, and calculate the total loss by combining the wear prediction loss and the domain adversarial loss. Update the network parameters through the gradient reversal layer to obtain dynamically adjusted network weights;

[0106] Decay the learning rate to 0.1 of the original value every 50 rounds of training, and calculate the wear prediction error and the domain classification confusion matrix on the test set. When the prediction error is less than the threshold and the domain classification accuracy is close to the random guessing level, obtain a cross-domain migration model.

[0107] In this embodiment, the parameters of the neural network are initialized to obtain the initial network parameters. In the model initialization stage, all the weight parameters and bias parameters in the feature extractor, the domain classifier, and the wear predictor are all initialized by the Xavier initialization method or the He initialization method. These two initialization methods can keep the initial values of the parameters within a reasonable distribution range, effectively avoiding the problems of gradient disappearance or gradient explosion in the early stage of model training. For example, for the weight matrix and of the fully connected layer

[0108] ​

[0109] Among them, represents a uniform distribution on the interval . This initialization method ensures that the variance of the signals between different layers remains stable, which helps the model converge quickly. After completing the initialization of the model parameters, the first training set is input into the adaptive domain adversarial neural network in batches for pre-training. In the pre-training stage, only the parameters of the feature extractor and the wear predictor are updated, while the parameters of the domain classifier are kept frozen, which means that in the multi-task loss function, the domain adversarial loss weight coefficient is set to 0, enabling the model to focus on the device wear prediction task first. The wear prediction loss is calculated by the mean squared error, and the formula is:

[0110]

[0111] Among them, is the number of samples in a batch, is the true wear degree of the th sample, is the wear value predicted by the model. In each batch, the model calculates the predicted value through forward propagation and calculates the gradient of the loss function with respect to the model parameters through backpropagation, and then updates the network parameters of the feature extractor and the wear predictor by the gradient descent method (such as the Adam optimizer):

[0112]

[0113] Among them, is the learning rate, which controls the step size of the model parameter update. During the pre-training process, the loss of the model on the validation set gradually decreases. When the validation loss converges to a stable value, it is considered that the model has learned sufficient device wear prediction features and the pre-trained model is obtained. After the model pre-training is completed, a data loader is constructed. The first training set is marked as the source domain data (simulation data), and the second training set is marked as the target domain data (measured data). By setting the same sampling strategy ratio, the source domain and target domain data are mixed to generate a batch data stream, which is called the domain adversarial training data stream. In the mixed data batch, each batch of data contains both source domain data and target domain data to ensure that the model can learn both the wear prediction task of the source domain and the domain adaptation task of the target domain during the training process. In the domain adversarial training stage, for each mixed batch of data, the high-dimensional feature representation of the input feature , the feature extraction process is expressed as:

[0114]

[0115] Among them, is the weight matrix of the feature extractor, is the bias, is the activation function (such as the ReLU function). The extracted feature is simultaneously input into the domain classifier and the wear predictor. The domain classifier outputs the domain prediction result , where represents the source domain data represents the target domain data. The domain classification loss is calculated by the cross-entropy loss function:

[0116] ;

[0117] Among them, is the actual domain label (0 or 1). Through the gradient reversal layer, the gradient is inverted during backpropagation to achieve the effect of adversarial training. During the training process, according to the current training progress calculate the dynamic change of the domain adversarial loss weight coefficient , and adopt a linear increase strategy, for example:

[0118]

[0119] Among them, , as increases, smoothly transitions from close to 0 to close to 1, enabling the model to focus on the wear prediction task in the early stage and enhancing the domain adaptation ability in the later stage. The total loss function is composed of the wear prediction loss and the domain adversarial loss weighted:

[0120]

[0121] During each update of the model parameters, the total loss function is minimized by the gradient descent method to achieve dynamically adjusted network weights. The learning rate is decayed to 0.1 of the original value every 50 training rounds to avoid oscillations in the later stage of model training and improve the convergence accuracy of the model. During the model training process, the wear prediction error and the domain classification confusion matrix are calculated on the test set. The wear prediction error is expressed as the deviation between the predicted value and the true value:

[0122]

[0123] Among them, is the number of samples in the test set, and are the actual and predicted wear values of the th sample respectively. When the prediction error is less than a preset threshold (e.g., 0.05) and the domain classification accuracy in the domain classification confusion matrix is close to the random guessing level (50%), it indicates that the model has achieved the alignment of the feature distributions of the simulation data and the measured data, achieving the expected effect of domain adversarial training, and finally obtaining a cross-domain transfer model.

[0124] Optionally, the output subsystem 15 further includes:

[0125] A filtering unit, configured to collect point cloud data of the working surface of the construction waste sorting device through a depth vision camera module, perform adaptive voxel filtering and normal vector estimation on the collected point cloud data, and obtain a regularized voxel grid and local geometric features;

[0126] A feature extraction unit, configured to extract features from the regularized voxel grid and local geometric features to obtain a high-dimensional feature vector of the point cloud;

[0127] A dimensionality reduction processing unit, configured to input the high-dimensional feature vector of the point cloud into a variational autoencoder for dimensionality reduction processing to obtain wear fault features;

[0128] A clustering unit, configured to perform wear mode clustering on the wear fault features based on the prediction results of the device wear degree prediction model to obtain a wear fault risk assessment result;

[0129] A construction unit, configured to construct a Takagi-Sugeno fuzzy model to describe the nonlinear dynamic characteristics of the construction waste sorting device, design a multi-rate sampling framework to group the sensor acquisition frequencies, and determine the feedback control gain by solving a linear matrix inequality optimization problem to obtain a basic parameter optimization strategy;

[0130] A solving unit, configured to dynamically adjust the feeding rate, crushing speed, screening frequency, and screening amplitude of the construction waste sorting device based on the wear fault risk assessment result and the basic parameter optimization strategy, in combination with a Bayesian optimization system to solve the optimal operating parameter combination, and obtain a fault closed-loop response plan.

[0131] In this embodiment, the filtering unit collects point cloud data of the working surface of the construction waste sorting device through a depth vision camera module (such as a lidar or a structured light 3D scanner). These point cloud data are stored in the form of three-dimensional coordinates, and each point contains spatial coordinates , where , , respectively represent the point Position in three-dimensional space. Adaptive voxel filtering is performed on the original point cloud data. By dividing the point cloud into cubes (voxels) of consistent size and reducing multiple points within each voxel to a representative point, this process significantly reduces the amount of point cloud data while retaining spatial structure information. Assume the voxel size is , and the center position of the voxel is , then the point after voxel filtering is calculated as follows:

[0132]

[0133] where is the number of points within the voxel. After filtering, a regularized voxel grid is obtained, improving the efficiency of point cloud processing. After completing voxel filtering, the filtering unit estimates the normal vectors of the point cloud data to obtain the local geometric features of each point. The normal vector reflects the directional characteristics of the point cloud surface in three-dimensional space. The normal vector is obtained by calculating the covariance matrix of the neighborhood (using the k-nearest neighbor algorithm, assume the neighborhood size is ) of each point in the point cloud through eigenvalue decomposition:

[0134]

[0135] where is the centroid of the neighborhood points, and the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is the normal vector . Geometric features such as local curvature, flatness, and edge characteristics are obtained through normal vector estimation. The feature extraction unit performs high-dimensional feature extraction on the regularized voxel grid and local geometric features. Using three-dimensional point cloud feature descriptors (such as FPFH or SHOT features), the spatial geometric information and normal vector information of the point cloud are converted into a high-dimensional feature vector , where is the feature dimension. For example, by calculating the distribution of the normal vector angles and distances between each point in the point cloud and the points in its neighborhood, a multi-dimensional feature representation describing the surface topography and wear characteristics of the device is obtained. After obtaining the high-dimensional feature vector, the dimensionality reduction processing unit uses a variational autoencoder to perform dimensionality reduction processing on the point cloud high-dimensional feature vector. The variational autoencoder compresses the high-dimensional input feature into a low-dimensional latent space , where . The encoder maps the input feature to a mean vector and a standard deviation vector :

[0136]

[0137] Among them, and are weight matrices, and are bias vectors. The latent representation is sampled through the reparameterization technique:

[0138]

[0139] The decoder remaps the latent representation back to the original feature space and calculates the reconstruction loss (mean square error) and the regularization loss (KL divergence) . The total loss function is:

[0140]

[0141] Among them, is the regularization term weight coefficient. Feature dimensionality reduction is achieved by minimizing the total loss function to obtain the wear fault feature z. The clustering unit performs wear mode clustering on the wear fault features based on the prediction results of the device wear degree prediction model. Using the K-means or DBSCAN clustering algorithm, the low-dimensional features are divided into different wear modes, such as mild wear, uniform wear, local wear, and abnormal wear, etc., to obtain the wear fault risk assessment result. After realizing wear mode recognition, the construction unit constructs a Takagi-Sugeno (T-S) fuzzy model to describe the nonlinear dynamic characteristics of the construction waste sorting device. The T-S fuzzy model consists of multiple linear sub-models and performs inference through the fuzzy rules

[0142] . The model output is the weighted average of all sub-models:

[0143]

[0144] Among them, is the fuzzy weight, is the input at the th rule's membership function value. To improve the sampling efficiency, a multi-rate sampling framework is constructed to group the sensor sampling frequencies and dynamically adjust the sampling rate according to the device state. The feedback control gain is determined by solving the linear matrix inequality optimization problem to form the basic strategy for parameter optimization. The solution unit, based on the wear fault risk assessment result and the basic strategy for parameter optimization, combines the Bayesian optimization system to solve the optimal operating parameter combination. Bayesian optimization predicts the device performance under different operating parameters by constructing a surrogate model and selects the operating parameters that maximize the acquisition function Parameter combinations:

[0145]

[0146] By iterating repeatedly, dynamically adjust the feeding rate of the construction waste sorting device , crushing speed , screening frequency and screening amplitude , and finally output a closed-loop fault response scheme to achieve automatic control of the equipment under complex working conditions and improve the stability and safety of the equipment operation.

[0147] Optionally, the dimensionality reduction processing unit is specifically used for:

[0148] Construct a variational autoencoder including an encoder and a decoder, where the encoder consists of three fully connected layers with the number of nodes being 256, 128, and 64 in sequence, and each fully connected layer is followed by a batch normalization layer and a ReLU activation function;

[0149] Input the high-dimensional feature vector of the point cloud into the variational autoencoder for feature processing to obtain the encoded output feature, map the encoded output feature to a mean vector and a variance vector respectively, and sample through the reparameterization technique to obtain the latent representation after dimensionality reduction;

[0150] Apply an orthogonality constraint to the latent representation after dimensionality reduction, and obtain the orthogonal latent feature by calculating the difference between the covariance matrix of the latent variable and the identity matrix and minimizing its norm;

[0151] Map the orthogonal latent feature back to the original feature space to obtain the reconstructed feature vector, calculate the mean square error between the decoder network and the reconstructed feature vector as the reconstruction loss, and calculate the KL divergence between the latent variable distribution and the standard normal distribution as the regularization loss, and combine the reconstruction loss and the regularization loss to form the total loss function;

[0152] Based on the total loss function, perform iterative optimization and output the wear fault feature.

[0153] In this embodiment, a variational autoencoder model composed of an encoder and a decoder is constructed. The variational autoencoder is a generative model that performs dimensionality reduction by learning the latent space representation of the input data and reconstructs the latent representation back to the original data space through the decoder. The encoder part consists of multiple fully connected layers. The designed encoder contains three fully connected layers, consisting of 256, 128, and 64 nodes respectively. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function, which helps to accelerate the training process of the model and prevent overfitting. Assume that the input high-dimensional feature vector is , where is the dimension of the features, and the encoder maps these features to the latent space, generating two vectors: the mean vector and the variance vector , which are used to define the distribution of the latent space. The first fully connected layer maps the input features to an intermediate space with a size of 256, obtaining , where is the weight matrix of the first layer, is the bias term. After processing through the batch normalization layer and the ReLU activation function, BatchNorm is obtained. The second and third layers are processed in a similar manner to obtain the intermediate features and , and the output is mapped to the mean vector and the variance vector through two independent fully connected layers:

[0154]

[0155] After obtaining the mean and the variance , the latent variable is sampled from the distribution of the latent space through the reparameterization trick. By introducing the noise variable , the latent variable becomes differentiable, enabling backpropagation optimization. The latent variable is calculated through the following formula:

[0156]

[0157] This approach brings the control of the mean and variance into the latent space and provides higher flexibility and robustness for the subsequent generation process. To enhance the structure of the latent space and avoid high correlations between latent representations, an orthogonal constraint is imposed on the latent representations after dimensionality reduction to ensure that there is no strong correlation between different dimensions of the latent space. Let be the latent features after dimensionality reduction, and calculate the covariance matrix of the latent variables:

[0158]

[0159] where is the number of samples, is the mean vector of all latent representations. To impose the orthogonal constraint, minimize the difference between the covariance matrix and the identity matrix , and the objective function is expressed as:

[0160]

[0161] Among them, is the Frobenius norm, representing the sum of the squares of the matrix elements. This constraint ensures that each dimension of the latent space is orthogonal to each other, resulting in more structured latent features. In the latent features after dimensionality reduction The decoder maps it back to the original feature space. The structure of the decoder is a mirror image of the encoder and consists of multiple fully connected layers. Assuming the output of the decoder is the reconstructed feature vector , the goal of the decoder is to minimize the reconstruction error, that is, the difference between the original feature and the reconstructed feature . The mean squared error is used as the reconstruction loss :

[0162]

[0163] Calculate the regularization loss of the latent variables by calculating the Kullback-Leibler divergence (KL divergence) between the distribution of the latent variables and the standard normal distribution . The KL divergence measures the difference between two probability distributions, and the formula is:

[0164]

[0165] Among them, and are the elements of the mean vector and the variance vector respectively. The role of the KL divergence is to push the distribution of the latent space closer to the standard normal distribution, prompting the model to learn smooth and meaningful latent features. The final total loss function is the weighted sum of the reconstruction loss and the regularization loss, and is optimized using the gradient descent algorithm during training:

[0166]

[0167] Among them, and are the weight coefficients of the KL divergence and the orthogonal constraint respectively, used to adjust the contribution of each loss to the total loss.

[0168] In the embodiments of the present application, a multi-dimensional sensor array including pressure sensors, three-axis acceleration sensors, displacement sensors, and depth vision camera modules is configured at the crusher working surface and the edge of the vibrating screen mesh. By combining wavelet transform noise reduction, time synchronization calibration, and Kalman filter algorithms for multi-source data fusion, high-precision monitoring of the equipment wear state is achieved, and complex wear characteristics that cannot be recognized by traditional single sensors are captured. A simulation model constructed based on finite element analysis, Archard wear theory, and discrete element method, combined with small sample measured data, forms a mechanism-data hybrid-driven fault prediction method, effectively solving the problem of poor generalization ability of pure data-driven methods for small sample data, and at the same time avoiding prediction deviations caused by over-simplification of pure mechanism models. By designing an adaptive domain adversarial neural network including a feature extractor, a domain classifier, and a wear predictor, effective knowledge transfer from the simulation data domain to the measured data domain is achieved, making full use of a large amount of simulation data to improve the training effect of the model, while maintaining adaptability to actual working conditions, and significantly improving the generalization ability of the fault prediction model. Using the depth vision camera module to collect point cloud data, combined with three-dimensional convolutional neural network and variational autoencoder for feature extraction and dimensionality reduction, low-dimensional wear fault features with high discrimination are generated through orthogonal constraint and KL divergence optimization. The Takagi-Sugeno fuzzy model is used to describe the nonlinear dynamic characteristics of the sorting system. Combining the multi-rate sampling framework to group the sensor acquisition frequencies, the feedback control gain is determined by solving the linear matrix inequality optimization problem, effectively coping with the uncertainties caused by sensor faults and time delays, and improving the stability and reliability of the system under complex working conditions. Based on the wear fault risk assessment results and the basic strategy of parameter optimization, combined with the Bayesian optimization system to solve the optimal operating parameter combination, the dynamic adjustment of the feeding rate, crushing speed, screening frequency, and screening amplitude of the sorting device is realized, forming a complete fault closed-loop response scheme, effectively extending the service life of the equipment while ensuring the sorting efficiency.

[0169] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing system embodiments and will not be elaborated herein.

[0170] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the system described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0171] As described above, the above embodiments are only used to illustrate the technical solution of this application, rather than to limit it; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of this application.

Claims

1. An intelligent sorting and recycling system for construction waste, characterized in that, The system includes: A data acquisition subsystem, which is used to install a sensor array on the crusher working surface and the edge of the vibrating screen mesh of the construction waste sorting device, and acquire a characteristic data set reflecting the equipment wear state; A simulation subsystem, which is used to establish a simulation model of the construction waste sorting device according to the characteristic data set, and generate a simulation data set containing multiple working condition parameter combinations based on the simulation model; specifically, the simulation subsystem is used to: construct a joint simulation environment including a material input module, a crushing dynamics module, a screening dynamics module, a wear calculation module, an equipment response module, a parameter mapping module and a data output module to obtain an interactive process simulation system; the wear calculation module calculates the wear amount of the equipment by combining stress distribution and sliding distance data, the equipment response module analyzes the force and deformation conditions of the equipment under load changes, the parameter mapping module maps the equipment operation state and wear characteristics, and the data output module outputs the simulation results as the standardized simulation data set, including equipment operation parameters, material characteristic parameters and the corresponding equipment wear degree; A measured data subsystem, which is used to collect small-sample measured data under actual construction waste sorting working conditions and calibrate the wear grade to obtain a labeled measured wear data set; A training subsystem, which is used to input the simulation data set and the labeled measured wear data set into an adaptive domain adversarial neural network for training to obtain an equipment wear degree prediction model; An output subsystem, which is used to identify wear fault characteristics through the equipment wear degree prediction model, and based on the wear fault characteristics, optimize the operation parameters of the construction waste sorting device and output a fault closed-loop response plan; the output subsystem further includes: a filtering unit, which is used to collect point cloud data of the working surface of the construction waste sorting device through a depth vision camera module, perform adaptive voxel filtering and normal vector estimation on the collected point cloud data to obtain a regularized voxel grid and local geometric features; a feature extraction unit, which is used to extract features from the regularized voxel grid and local geometric features to obtain a point cloud high-dimensional feature vector; a dimensionality reduction processing unit, which is used to input the point cloud high-dimensional feature vector into a variational autoencoder for dimensionality reduction processing to obtain wear fault characteristics; a clustering unit, which is used to perform wear mode clustering on the wear fault characteristics based on the prediction result of the equipment wear degree prediction model to obtain a wear fault risk assessment result; a construction unit, which is used to construct a Takagi-Sugeno fuzzy model to describe the nonlinear dynamic characteristics of the construction waste sorting device, design a multi-rate sampling framework to group the sensor acquisition frequencies, and determine the feedback control gain by solving a linear matrix inequality optimization problem to obtain a basic parameter optimization strategy; a solving unit, which is used to dynamically adjust the feeding rate, crushing speed, screening frequency and screening amplitude of the construction waste sorting device based on the wear fault risk assessment result and the basic parameter optimization strategy, combined with a Bayesian optimization system to solve the optimal operation parameter combination, to obtain a fault closed-loop response plan.

2. The intelligent sorting and recycling system for construction waste according to claim 1, wherein Specifically, the data acquisition subsystem is used to: Install pressure sensors, triaxial acceleration sensors, displacement sensors, and depth vision camera modules on the inner wall of the crusher working surface and the edge of the vibrating screen of the construction waste separation device to obtain a multi-dimensional sensor array; Set different sampling frequencies for the multi-dimensional sensor array, respectively collect pressure data, vibration data, displacement data, and three-dimensional point cloud data to obtain the original equipment wear data; Perform wavelet transform noise reduction processing on the original equipment wear data to obtain a denoised data segment set, and perform time synchronization calibration on the denoised data segment set to obtain a standardized data set; Extract time-domain features, frequency-domain features, and time-frequency domain features from the standardized data set to obtain a multi-dimensional feature vector, and use the Kalman filter algorithm to perform multi-source data fusion on the multi-dimensional feature vector to obtain a feature data set reflecting the current wear state of the equipment.

3. The intelligent sorting and recycling system for construction waste according to claim 1, characterized in that The actual measurement subsystem is specifically used for: Select multiple target devices on the actual operating construction waste separation line as monitoring objects, and configure sensors according to the same installation scheme as the sensor array to obtain an actual measurement monitoring system; Set multiple groups of different typical working condition parameters for the actual measurement monitoring system, and continuously operate under different working conditions and regularly record the equipment surface topography data and sensor data to obtain the actual measurement original data; Extract time-domain features from the actual measurement original data to obtain a time-domain feature set, and perform fast Fourier transform and wavelet packet decomposition on the actual measurement original data to obtain a frequency-domain and time-frequency domain feature set; Use the three-dimensional point cloud data obtained by the depth vision camera module to calculate surface roughness indexes, principal curvature distributions, and area and depth distributions of wear areas to obtain a depth feature set; Select the most representative features from the time-domain feature set, the frequency-domain and time-frequency domain feature set, and the depth feature set, and combine visual inspection and laser scanning data to calibrate the equipment wear degree level to obtain a labeled actual measurement wear data set.

4. The intelligent sorting and circular regeneration system for construction waste according to claim 1, wherein, The training subsystem further includes: A design unit for designing an adaptive domain adversarial neural network structure including a feature extractor, a domain classifier, and a wear predictor, where the feature extractor consists of multiple fully connected layers, batch normalization layers, and activation functions; An adversarial unit for defining a multi-task loss function including a wear prediction loss and a domain adversarial loss, and implementing adversarial training through a gradient reversal layer to obtain a training optimization target; A division unit for dividing the simulation data set into a first training set and a validation set, and dividing the labeled actual measurement wear data set into a second training set and a test set; A joint unit for using the first training set for network pre-training, and then combining the second training set for joint training, and gradually increasing the domain adversarial loss weight using a dynamic weight adjustment strategy to obtain an equipment wear degree prediction model.

5. The intelligent sorting and recycling system for construction waste according to claim 1, characterized in that, The dimensionality reduction processing unit is specifically used for: Construct a variational autoencoder including an encoder and a decoder, where the encoder consists of three fully connected layers, and the number of nodes is 256, 128, and 64 in sequence. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function; Input the high-dimensional feature vector of the point cloud into the variational autoencoder for feature processing to obtain the encoded output features, map the encoded output features to a mean vector and a variance vector respectively, and sample through the reparameterization technique to obtain the latent representation after dimensionality reduction; Apply an orthogonality constraint to the latent representation after dimensionality reduction, and obtain the orthogonal latent features by calculating the difference between the latent variable covariance matrix and the identity matrix and minimizing its norm; Map the orthogonal latent features back to the original feature space to obtain the reconstructed feature vector, calculate the mean square error between the decoder network and the reconstructed feature vector as the reconstruction loss, calculate the KL divergence between the latent variable distribution and the standard normal distribution as the regularization loss, and combine the reconstruction loss and the regularization loss to form the total loss function; Based on the total loss function, perform iterative optimization and output the wear fault features.

Images