Building waste intelligent prediction, classification and recycling treatment system based on multi-mode AI

Abstract

The invention relates to a building waste intelligent prediction, classification and resource disposal system based on multi-modal AI. The building waste intelligent prediction, classification and resource disposal system comprises a data capture module, a pre-judgment module, a dynamic accounting module, a resource disposal module, an intelligent market distribution module and a carbon accounting module. Precise prediction, dynamic accounting, intelligent classification, recycling treatment and carbon footprint management of the construction waste are achieved through multi-modal data perception and AI intelligent decision, a set of complete intelligent construction waste treatment system is constructed, and full-chain automatic and digital management of the construction waste from generation to recycling can be achieved; and an intelligent solution is provided for environmental protection industry, engineering construction and supervision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of environmental protection technology application of artificial intelligence, in particular to a construction waste intelligent prediction, classification and resource disposal system based on multi-modal AI. BACKGROUND

[0002] With the acceleration of urbanization, the rapid development of the construction industry leads to a sharp increase in the amount of construction waste. Currently, the traditional construction waste management method has the following problems: (1) Construction waste is mostly treated by landfill or stacking, occupying a large amount of land, and the recyclable resources (such as concrete, steel bars, wood) cannot be efficiently sorted and reused; (2) Manual classification is low in efficiency and high in error rate, resulting in insufficient recycling rate of recycled aggregates, metals, etc.; (3) Relying on manual recording and regular inspection, it is impossible to real-time grasp the waste generation, transportation track and treatment status; (4) The data of various supervision systems are not connected, making it difficult to realize whole-process tracing and collaborative management; (5) No linkage mechanism with the carbon trading market is established, and the emission reduction potential is not fully tapped; (6) Garbage transfer relies on experience scheduling, often resulting in idle or overloading of vehicles, and high transportation cost proportion; (7) Unreasonable distribution of production lines leads to mixed treatment of high-value resources (such as steel bars) and low-value materials (such as slag), resulting in poor economic benefits.

[0003] At the same time, the existing technology almost presents the characteristics of "fragmentation" and "isolation", and only optimizes one link in the huge construction waste governance system, failing to build an end-to-end intelligent closed-loop system. The existing technology has the shortcoming of "disconnection between perception and decision". For example, Chinese patent application numbers 2024111009605 and 2021107234679, whether based on meta deep learning for fine classification or combining YOLOv10 and SAM for component estimation, the technical focus is limited to "recognition" itself. The classification data eventually becomes the end of information, rather than the starting point of decision, and fails to form linkage with the resource disposal and market absorption in the back end. This leads to the system only replacing the human eye, but not the human brain for scheduling and optimization.

[0004] Secondly, in terms of data management, the existing technology (Chinese patent application number 2023105496322) exposes the limitation of "emphasis on storage and neglect of intelligence". The management system based on blockchain does solve the problem of data authenticity and traceability, but its technical essence is to digitize the management process, and it has not realized intelligence. The system lacks deep mining and real-time analysis capability of data, and cannot actively predict risks, optimize disposal paths or dynamically adjust strategies based on whole life cycle data, so that the valuable data value cannot be fully released.

[0005] Again, in the field of carbon accounting, the existing method (such as Chinese patent application number 202411036470.3) has the defects of "isolated measurement and lagging results". Although the special carbon emission calculation system can relatively accurately calculate the carbon footprint of the transportation link through a complex mathematical model, it is equivalent to an ex post facto accounting tool. The accounting process is independent of the main business flow, and the results cannot be fed back in real time and guide the front-end transportation scheduling, the middle-end disposal method selection, or the back-end market transaction. This lag and isolation make carbon management a reporting behavior rather than a dynamic optimization-driven emission reduction force.

[0006] In order to effectively solve the above problems, an intelligent system that integrates multi-source data perception, AI dynamic decision-making, and full-chain carbon management is urgently needed to achieve accurate prediction, real-time accounting, adaptive resourceization, and carbon footprint closure. SUMMARY

[0007] The purpose of the present application is to propose an intelligent prediction, classification and resourceization disposal system for construction waste based on multi-modal AI, to build a complete intelligent processing system for construction waste, to achieve accurate prediction, dynamic accounting, intelligent classification, resourceization disposal and carbon footprint management of construction waste through multi-modal data perception and AI intelligent decision-making, and to build a full-chain intelligent closed-loop system covering "prediction-monitoring-disposal-market-carbon management" through cross-disciplinary technology integration.

[0008] To achieve the above purpose, the present application provides an intelligent prediction, classification and resourceization disposal system for construction waste based on multi-modal AI, comprising:

[0009] Data grabbing module: obtain planning files and statistical files of regulatory departments through API interface to grab basic engineering information and design parameters;

[0010] Prediction module: predict the amount of construction waste based on historical data in planning files and statistical files and machine learning model;

[0011] Dynamic accounting module: multi-angle on-site image acquisition by unmanned aerial vehicle or camera and preprocessing, data labeling, identification of construction waste volume and waste classification by deep learning model, generation of three-dimensional model with classification label, calculation of actual volume and prediction difference, assistance in supervision, customization of photo shooting frequency by reinforcement learning model, and dynamic adjustment of pulling strategy according to prediction results of prediction module and output of construction waste classification results by deep learning model and production line load;

[0012] Resource disposal module: According to the classification results of construction waste output by the dynamic accounting module and the real-time capacity data, the reinforcement learning model is used to dynamically allocate to the optimal resource production line; wherein the process of closed-loop optimization and feedback of the reinforcement learning model in the dynamic accounting module includes: the dynamic accounting module pushes the construction waste related data to the reinforcement learning model, generates allocation instructions and issues them to the resource disposal module to control the production line; after each batch is processed, the actual resource utilization rate, energy consumption and other data are collected, the reward is calculated and the strategy is updated; the actual resource utilization rate, processing time and energy consumption data, and garbage composition deviation of various types of garbage are fed back to the dynamic accounting module;

[0013] Intelligent market allocation module: train the reinforcement learning model to automatically match market demand according to the characteristics of recycled products to optimize logistics distribution; wherein the working process of the reinforcement learning model in the intelligent market allocation module includes:

[0014] Step S61, environment construction: build a virtual market environment through Agent-Based Modeling, simulate the dynamic interaction of customers, suppliers and logistics merchants, and generate extreme cases by combining historical data enhancement technology;

[0015] Step S62, reinforcement learning model training:

[0016] State space: real-time inventory of various recycled products, specifications, quality grades, order types, demand, priority, regional distribution, transportation vehicle availability, route congestion, policy subsidy changes, and raw material price fluctuations;

[0017] Action space: allocate products to designated orders; transportation vehicle scheduling, path planning; dynamically adjust product prices to balance supply and demand;

[0018] Reward function: increase order satisfaction rate, reduce logistics cost, and improve inventory turnover rate as positive reward; order default and resource waste as negative reward;

[0019] Model training: based on historical orders, inventory and logistics data, build a simulation environment, pre-train using reinforcement learning, and optimize the initial strategy; after deployment, fine-tune model parameters through real-time feedback, and manually review high-risk decisions;

[0020] Step S63, closed-loop optimization and feedback: the resource disposal module pushes the data of recycled products to the intelligent market allocation module, and the external data source synchronizes in real time through the API interface, transmits market feedback to the resource disposal module, and dynamically adjusts the priority of the production line;

[0021] Carbon accounting module: based on the life cycle assessment model LCA, use machine learning model to calculate carbon emissions in transportation and processing links and emission reduction of recycled products.

[0022] Preferably, the basic engineering information includes engineering address, engineering type, land use type, engineering area, construction period; and the design parameters include structure type, material list in design drawings, and demolition and alteration plan.

[0023] Preferably, the step of the data grabbing module grabbing the basic engineering information and the design parameters comprises:

[0024] In step S11, an API interface call application is submitted to a regulatory department, and data is grabbed by using Python;

[0025] In step S12, fields returned by the API are mapped into internal standard fields of the system, and then data cleaning is performed, invalid data is filtered, and data is standardized, missing items are filled with default values or marked as “to be completed”;

[0026] In step S13, the cleaned data is stored in a relational database, and the data is synchronized at a regular time.

[0027] Preferably, in the pre-judgment module, the training model of the machine learning model comprises:

[0028] Feature encoding: select engineering type, land use type, engineering area, structure type, material consumption, construction period, etc. as key features, and use model feature importance to screen high-contribution features; then standardize and normalize numerical variables, and use one-hot encoding for classification variables;

[0029] Model training: select random forest algorithm to predict garbage quantity and classification; divide the data into training set: validation set: test set = 6:2:2, use 5-fold cross-validation to optimize hyperparameters, and select MAE, RMSE, R² to evaluate the continuous quantity of the regression model to predict construction waste, and select F1-score to evaluate the construction waste type identified by the classification model.

[0030] Preferably, the step of collecting multi-angle field images by using a drone or a camera and performing preprocessing and data labeling comprises:

[0031] Image collection: use a drone and an RGB camera to collect multi-angle images and videos of the construction waste pile; when the construction period is long, install a fixed camera to cover the main area of the waste pile, ensure that the image heading overlap reaches 70-80% and the lateral overlap reaches 30-50%, and use LIDAR to collect depth data in the face of complex occlusion scenes;

[0032] Preprocessing: use bilateral filtering to denoise, and combine CLAHE to enhance contrast; identify feature points of the same object in different images through SIFT feature matching, establish the correspondence between images, and use RANSAC to remove false matching points;

[0033] Data labeling: Label the pre-processed images and videos, including garbage categories, bounding boxes, and pixel-level masks, for training deep learning models.

[0034] Preferably, the reinforcement learning model in the dynamic accounting module includes:

[0035] State space: construction site dynamic data, including the remaining days of the construction period, the garbage accumulation rate, and the weather in the next month; system real-time data of the prediction module, including garbage quantity prediction value, deep learning classification result, and production line load;

[0036] Action space: scheduling vehicle priority, transportation route optimization, and image shooting frequency and coverage;

[0037] Reward function: Set shooting coverage of key accumulation areas and discovery of classification abnormalities as positive rewards for shooting frequency decision; Set repeated shooting of invalid areas as negative rewards for shooting frequency decision; Set production line load matching and transportation delay reduction as positive rewards for pulling strategy; Set production line overload and inefficient transportation route as negative rewards for pulling strategy;

[0038] Model training: The model training uses entropy regularization SAC to continuously balance exploration and utilization in the closed loop of "input state - policy network generates action - environment feedback reward - update policy". After each round of training, the system evaluates the shooting coverage, pulling efficiency, and total cost saving three core indicators in parallel, real-time tracks the strategy optimization effect and guides the next round of training.

[0039] Preferably, in the dynamic accounting module, the assistance supervision content includes:

[0040] Threshold setting rule: dynamically adjust the threshold based on historical data and engineering files using statistical methods;

[0041] Attribution analysis: When the predicted quantity is less than the actual quantity and the difference exceeds the threshold, the project progress is accelerated and unregistered garbage is mixed; When the predicted quantity is greater than the actual quantity and the difference exceeds the threshold, the project is delayed and the classification equipment misjudges;

[0042] Closed-loop optimization and feedback:

[0043] Optimize the prediction module using difference data: Record high difference cases, label the real reasons, and feed the corrected data back to the machine learning model of the prediction module to adjust the feature weights;

[0044] Optimize the dynamic accounting module using difference data: Strengthen data auditing for high-frequency difference construction sites, adjust image collection frequency, and statistically train garbage categories with high misjudgment frequency.

[0045] Preferably, the life cycle assessment model LCA construction comprises:

[0046] System boundary: covering the whole life cycle of construction waste from transportation, classification, treatment to recycled product utilization;

[0047] Emission factor library: integrating emission factors from authoritative databases, including: transportation emission factors, energy consumption emission factors, carbon offset factors of recycled products;

[0048] Computational design:

[0049] ;

[0050] ;

[0051] .

[0052] Preferably, in the carbon accounting module, the working steps of the machine learning model comprise:

[0053] Step S71, extracting garbage types, treatment processes, transportation paths, and energy types as key features, and performing one-hot encoding on the classification variables and standardizing the numerical values;

[0054] Step S72, using random forest to predict carbon emissions at each link, and using cross-validation to optimize hyperparameters, and selecting RMSE and R² as evaluation indexes;

[0055] Step S73, visualizing the output results according to regional distribution to generate a carbon emission heat map, and outputting a report at the same time.

[0056] Preferably, in the carbon accounting module, the closed-loop optimization and feedback process of the machine learning model comprises:

[0057] The carbon emission data is fed back to the dynamic accounting module and the intelligent market allocation module to optimize the transportation route or the treatment priority; in cooperation with the intelligent market allocation module, low-carbon recycled product orders are preferentially recommended; the difference between the carbon emission and the predicted value is recorded, and the model parameters and the emission factor library are updated regularly.

[0058] Based on the above technical solutions, the advantages of the present application are:

[0059] The construction waste intelligent prediction, classification and resource disposal system of the present application integrates environmental engineering, computer vision, Internet of Things, machine learning and smart city management, and is mainly applied to intelligent prediction, dynamic monitoring, accurate classification, resource disposal and carbon footprint management of construction waste, and can realize automatic and digital management of the whole chain of construction waste from generation to recycling, and provides an intelligent solution for environmental protection industry, engineering construction and supervision. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0061] Figure 1 A general diagram of a building waste intelligent prediction, classification and resource disposal system;

[0062] Figure 2 A step diagram for the data grabbing module to grab basic engineering information and design parameters;

[0063] Figure 3 A workflow diagram of a reinforcement learning model in the intelligent market allocation module;

[0064] Figure 4 A working step diagram of a machine learning model in the carbon accounting module. DETAILED DESCRIPTION

[0065] The technical solutions of the application will be further described in detail below with the aid of the accompanying drawings and embodiments.

[0066] The application proposes a building waste intelligent prediction, classification and resource disposal system based on multi-modal AI, and constructs a complete building waste intelligent processing system. Through multi-modal data sensing and AI intelligent decision-making, the system realizes accurate prediction, dynamic accounting, intelligent classification, resource disposal and carbon footprint management of building waste.

[0067] As shown in Figure 1 The building waste intelligent prediction, classification and resource disposal system mainly includes a data grabbing module, a prediction module, a dynamic accounting module, a resource disposal module, an intelligent market allocation module and a carbon accounting module, as follows:

[0068] (1) Data grabbing module: Obtain planning files of law enforcement bureaus and housing bureaus through API interfaces, so that the platform grabs information (address, engineering quantity, land area type, use nature, consumption use quantity, etc.).

[0069] (2) Prediction module: Predict the amount of building waste based on historical data in planning files and statistical files and machine learning models.

[0070] (3) Dynamic accounting module: Collect multi-angle field photos through drones or cameras, identify building waste volume and waste classification through deep learning, generate a three-dimensional model with classification labels, calculate the difference between actual volume and prediction, and assist in supervision. Utilize reinforcement learning models to customize photo shooting frequency according to factors such as construction period, and dynamically adjust the pulling strategy according to the prediction results of the prediction module, real-time classification results of the deep learning model and production line load.

[0071] (4) Resource recycling module: Based on the construction waste classification results (such as concrete, steel bars, wood, etc.) output by the dynamic accounting module and the real-time capacity data, the waste is dynamically allocated to the optimal resource recycling production line (such as crushing, sorting, and recycled aggregate processing) using a reinforcement learning model to maximize resource utilization and minimize processing costs.

[0072] (5) Intelligent market allocation module: The reinforcement learning model is trained to automatically match market demand based on the characteristics of recycled products (such as strength and specifications) and optimize logistics allocation.

[0073] (6) Carbon accounting module: Based on the LCA (Life Cycle Assessment) method, a machine learning model is used to calculate carbon emissions in transportation, processing and other links and the emission reduction of recycled products.

[0074] In the data capture module, the module mainly captures basic engineering information and design parameters from planning documents and statistical documents, specifically including:

[0075] (1) Basic engineering information

[0076] 1) Project address: Used to locate the project's geographical location, and to associate it with regional waste disposal policies or transportation costs.

[0077] 2) Project type: new construction, demolition, renovation, municipal engineering, etc. There are significant differences in the waste generation coefficient among different project types.

[0078] 3) Land use type: residential, commercial, industrial, etc. Different land use types affect the types of building materials and waste composition.

[0079] 4) Project area: Building area, land area, etc., which are directly related to the amount of waste generated.

[0080] 5) Construction period: The amount of waste generated will vary depending on the duration of the project (construction period).

[0081] (2) Design parameters

[0082] 1) Structural type: The material composition of different structures directly affects the composition of waste, such as steel structure, concrete structure, and brick-concrete structure.

[0083] 2) Material list in the design drawings: quantity data of cement, steel bars, bricks, etc.

[0084] 3) Demolition and renovation plan: The capacity of construction waste is determined by two methods: overall demolition and partial renovation.

[0085] like Figure 2 As shown, the steps of the data capture module in capturing basic engineering information and design parameters include:

[0086] Step S11, submit API interface call application to the regulatory department, and use Python for data scraping;

[0087] Step S12, map the fields returned by the API to the system internal standard fields, then perform data cleaning, filter invalid data and standardize the data, fill in the missing items for the mandatory fields with default values or mark them as "to be completed";

[0088] Step S13, use a relational database to store the cleaned data, and synchronize the newly added or modified data at regular intervals.

[0089] In the pre-judgment module, the machine learning model training operation is as follows:

[0090] (1) Data preparation

[0091] Data source integration: integrate historical engineering data (such as engineering type, area, material list, construction period, etc.) and corresponding construction waste generation data, and supplement external data such as regional climate, policies and regulations that may affect waste generation during the construction period.

[0092] Data cleaning: according to professional knowledge, use clustering algorithm in statistical method to report error for nominal data and numerical data; fill in default values or mark "to be completed" for missing items in mandatory fields, and unify units and formats for integrated data.

[0093] (2) Model training

[0094] Feature encoding: select engineering type, land use type, engineering area, structure type, material consumption, construction period, etc. as key features, and use model feature importance (such as random forest) to screen high contribution features. Standardize and normalize numerical variables, and use one-hot encoding for categorical variables.

[0095] Model training: select random forest algorithm to predict waste quantity and classification. Divide the data into training set: validation set: test set = 6:2:2, use 5-fold cross-validation to optimize hyperparameters, and select MAE, RMSE, R² to evaluate the regression model's prediction of continuous construction waste, and F1-score to evaluate the classification model's recognition of construction waste types.

[0096] In the dynamic accounting module, the deep learning model technology implementation process is as follows:

[0097] (1) Data preparation

[0098] Image collection: Use drones and RGB cameras to collect multi-angle images and videos of construction waste piles; install fixed cameras when the construction period is long to cover the main area of the waste pile, ensuring that the image heading overlap reaches 70-80% and the lateral overlap reaches 30-50%; use LIDAR to collect depth data in complex occlusion scenarios.

[0099] Preprocessing: Use bilateral filtering for denoising and combine CLAHE to enhance contrast to solve uneven lighting and overexposure problems; identify feature points of the same object in different images through SIFT feature matching, establish the correspondence between images, and use RANSAC to remove false matching points to ensure accurate matching points.

[0100] Data labeling: Label the preprocessed images and videos, including waste categories (such as concrete, steel, wood, etc.), bounding boxes, and pixel-level masks, for training deep learning models.

[0101] (2) Deep learning model training

[0102] Model selection: Use YOLOv8 to identify and classify construction waste in images; use SfM technology to generate a three-dimensional model with classification labels from multi-angle images; align the point cloud generated by SfM with the classification labels of YOLOv8; based on the three-dimensional model, calculate the actual volume of waste through point cloud analysis.

[0103] Model training: Use the labeled dataset to train the model, optimize parameters to improve recognition and classification accuracy; select mAP (mean average precision) for target detection and volume calculation error rate for three-dimensional models.

[0104] Further, the reinforcement learning model technology implementation process in the dynamic accounting module is as follows:

[0105] (1) Environment initialization

[0106] Construct a virtual construction site environment, define state space, action space, and reward function, and randomly generate dynamic rules such as waste accumulation rate, weather changes, and production line status.

[0107] (2) Reinforcement learning model training

[0108] State space: remaining days of construction period, waste accumulation rate, future one-month weather, and other construction site dynamic data; waste quantity prediction value of prediction module, deep learning classification result, and production line load and other system real-time data.

[0109] Action space: vehicle priority scheduling (such as concrete priority transportation), transportation route optimization (such as shortest path or low-carbon path), and image shooting frequency and coverage range.

[0110] Reward function: Set the shooting coverage of key accumulation areas and the discovery of classification abnormalities as positive rewards for shooting frequency decision-making; Set repeated shooting of invalid areas as negative rewards for shooting frequency decision-making; Set production line load matching and transportation delay reduction as positive rewards for pulling and transporting strategies; Set production line overload and inefficient transportation routes as negative rewards for pulling and transporting strategies.

[0111] Model training: The model training uses the entropy regularization (SAC) algorithm to continuously balance exploration and utilization in the closed loop of "input state - policy network generates action - environmental feedback reward - update policy". After each round of training, the system evaluates three core indicators in parallel: shooting coverage, pulling efficiency and total cost savings, to track the effect of policy optimization in real time and guide the next round of training.

[0112] Preferably, in the dynamic accounting module, the assistance supervision content includes:

[0113] (1) Set threshold rules

[0114] Based on historical data and engineering files, the threshold is dynamically adjusted by statistical methods to avoid fixed thresholds that are not suitable for all engineering types.

[0115] (2) Attributive analysis

[0116] When the predicted amount is less than the actual amount and the difference exceeds the threshold: project progress is accelerated, unregistered garbage is mixed, etc.; When the predicted amount is greater than the actual amount and the difference exceeds the threshold: project delay, misjudgment of classification equipment, etc.

[0117] (3) Closed-loop optimization and feedback

[0118] Optimize the prediction module using difference data: record high difference cases, label the real cause, and feed back the corrected data to the machine learning model of the prediction module to adjust the feature weights.

[0119] Optimize the dynamic accounting module using difference data: strengthen data auditing for high-frequency difference construction sites and adjust image acquisition frequency; Statistically train garbage categories with high misjudgment frequency.

[0120] Further, in the resource disposal module, the reinforcement learning model technology implementation process is as follows:

[0121] (1) Environment construction

[0122] Based on historical data, a virtual construction site scenario is generated to simulate garbage arrival rate, composition changes and equipment failure events; Random disturbances (such as sudden increase in demolition garbage) are introduced to enhance model robustness.

[0123] (2) Reinforcement learning model training

[0124] State space: construction waste characteristics, production line status, external factors (time-of-day fluctuations in electricity prices, priority of urgent orders).

[0125] Action space: production line allocation decisions, production line processing parameter adjustments, task priority scheduling.

[0126] Reward function: set the resource utilization rate, energy cost reduction, equipment utilization rate as positive reward; set the processing delay caused by accumulation, equipment overload triggered maintenance as negative reward.

[0127] Model training: use simulation data to train the initial strategy, explore high return action sequences through Monte Carlo Tree Search (MCTS); after deployment, fine-tune with real data, and use imitation learning to incorporate expert rules.

[0128] (3) Closed-loop optimization and feedback

[0129] The dynamic accounting module pushes construction waste-related data to the reinforcement learning model, generates allocation instructions and issues them to the resource utilization module to control the production line; after each batch is processed, actual resource utilization rate, energy consumption, etc. Data is collected, rewards are calculated and strategies are updated; the actual resource utilization rate of various types of waste, processing time and energy consumption data, and waste composition deviation are fed back to the dynamic accounting module.

[0130] Further, as shown in Figure 3 , the workflow of the reinforcement learning model in the intelligent market allocation module includes:

[0131] Step S61, Environment Construction: Build a virtual market environment through Agent-Based Modeling (ABM) to simulate the dynamic interaction of customers, suppliers and logistics companies; combine historical data enhancement techniques to generate extreme cases (such as supply chain disruption, demand surge) to ensure that the training environment covers the uncertainties of the real market.

[0132] Step S62, Reinforcement Learning Model Training:

[0133] State space: real-time inventory of various types of recycled products, specifications, quality grades; order type (such as municipal engineering, commercial real estate), demand, priority (urgent orders), regional distribution; availability of transport vehicles, road congestion; policy subsidy changes, raw material price fluctuations.

[0134] Action space: allocate products to designated orders; transport vehicle scheduling, path planning; dynamically adjust product prices to balance supply and demand.

[0135] Reward function: set order satisfaction rate, logistics cost reduction, inventory turnover rate improvement as positive reward; set order default, resource waste as negative reward.

[0136] Model training: Based on historical orders, inventory and logistics data, a simulation environment is constructed, pre-trained using reinforcement learning algorithms, and the initial strategy is optimized. After deployment, the model parameters are fine-tuned through real-time feedback. High-risk decisions are manually reviewed.

[0137] Step S63, closed-loop optimization and feedback: The resource disposal module pushes the data of the recycled products to the intelligent market distribution module; external data sources (such as changes in market demand, logistics status, and policy subsidy adjustments) are synchronized in real time through API interfaces to ensure the timeliness of the model input data; market feedback is transmitted to the resource disposal module to dynamically adjust the production line priority.

[0138] Furthermore, in the carbon accounting module, based on the life cycle assessment model LCA, a machine learning model is used to calculate the carbon emissions of transportation and processing links and the emission reduction of recycled products. Specifically, the technical implementation process is as follows:

[0139] (1) Data preparation

[0140] Collect carbon emission related data in the whole process of construction waste treatment, including transportation distance, vehicle type and load, energy consumption of processing link, carbon emission reduction potential of recycled products, etc.; obtain GPS data of transportation vehicles, production line energy consumption monitoring data, etc. in real time through API interface or Internet of Things devices.

[0141] (2) Life cycle assessment (LCA) model construction

[0142] System boundary: covering the whole life cycle of construction waste from transportation, classification, processing to recycled product utilization;

[0143] Emission factor library: integrate emission factors from authoritative databases, including: transportation emission factors, energy consumption emission factors, carbon offset factors of recycled products;

[0144] Calculation design:

[0145] ;

[0146] ;

[0147] .

[0148] Preferably, as shown in Figure 4 , in the carbon accounting module, the working steps of the machine learning model include:

[0149] Step S71, extract garbage type, processing technology, transportation path, and energy type as key features, and perform one-hot encoding on the classification variables and standardize the numerical values;

[0150] Step S72, each link carbon emission is predicted by using random forest, and hyperparameters are optimized by using cross-validation, and RMSE and R² are selected as evaluation indexes;

[0151] Step S73, the output results are visualized according to the regional distribution to generate a carbon emission heat map, and a report is output at the same time, and the details of the report include the proportion of each sub-link emission, the comparison of emission reduction benefits and the like.

[0152] In the carbon accounting module, the closed-loop optimization and feedback process of the machine learning model includes:

[0153] The carbon emission data is fed back to the dynamic accounting module and the intelligent market allocation module, and the transportation route is optimized or the priority is handled (for example, low-carbon orders are selected, and high-emission-potential waste is optimized); in combination with the intelligent market allocation module, low-carbon recycled product orders are preferentially recommended; the difference between the carbon emission and the predicted value is recorded, and the model parameters and the emission factor library are updated regularly.

[0154] To further illustrate the technical solutions of the present application, the following is an example of the application of the building waste intelligent prediction, classification and resource disposal system based on multi-modal AI in the project of "a 20-year-old commercial complex in a core area of a certain district is demolished and reconstructed as an ecological intelligent building" to illustrate the application of the present application.

[0155] (1) On-site situation

[0156] ① Urgent construction period: only 6 months, the traditional manual waste counting method is seriously lagging behind.

[0157] ② Complex composition: including concrete, steel bars, glass curtain walls, waste equipment, decoration waste and the like, and it is difficult to accurately classify.

[0158] ③ Municipal environmental protection: the municipal government requires that the resource utilization rate of the project be not less than 80%, and a complete carbon emission reduction report needs to be submitted.

[0159] ④ Cost control: located in the city center, the transportation and disposal cost is high, and it needs to be optimized to the extreme.

[0160] (2) System whole-chain processing flow

[0161] The system automatically obtains the planning permit, structural construction drawing and detailed material list (such as recording cement consumption of 12000 tons, steel bars of 3500 tons) of the project from the municipal housing and construction bureau API interface through the data grabbing module. The prediction module is started immediately, based on these high-value data, the trained random forest machine learning model is used for first-round prediction, and the key conclusion is output: it is predicted that 28500 tons of construction waste will be generated (±1500 tons), and the steel component accounts for 12.5%.

[0162] After the project starts, the dynamic accounting module immediately takes over the site supervision. The system deploys 3 drones and 5 fixed high-definition cameras, which dynamically customize the "3 times high-altitude cruising per day, real-time monitoring of key areas" strategy based on the reinforcement learning model to perform multi-angle image acquisition. After preprocessing the raw images collected by bilateral filtering denoising and CLAHE contrast enhancement, they are input into the deep learning model based on the YOLOv8 architecture for recognition and classification. The model successfully identifies about 150 tons of steel bars buried in the concrete blocks, and simultaneously generates a three-dimensional point cloud model with classification labels, calculating the actual garbage volume as 29100 tons. The system automatically compares the predicted amount with the measured amount, with an error of only 2.1%, far below the industry's 5% error threshold, while the traditional experience estimation error is usually 10%-20%, fully verifying the reliability of the prediction model.

[0163] At the same time, the reinforcement learning model in the module dynamically generates the "priority removal of steel bars, and detours the city east morning peak road segment" optimal pulling instruction based on the real-time accumulation rate (150 tons / day), the remaining processing capacity of the production line (80% load), and the real-time traffic congestion index.

[0164] Based on the accurate classification results and real-time capacity data from the dynamic accounting module, the resource disposal module makes optimal allocation decisions through its internal reinforcement learning model: dynamically allocates more than 18000 tons of concrete blocks to the recycled aggregate production line, and directs the allocation of about 3500 tons of identified steel bars to the metal recycling line. After each batch (500 tons as a unit) is processed, the module collects actual data (the steel bar recycling rate of this batch reaches 95%, and the energy consumption is 18 kWh / ton), and calculates the reward value to feedback to the model for strategy updating, forming a closed loop that becomes more intelligent with use. Through the continuous optimization of this closed loop, the overall resource utilization rate of the project is finally increased to 88.7%, the steel bar recycling rate is increased from about 90% to 95%, and the accurate classification avoids the omission and waste of high-value resources.

[0165] The product information of about 15000 tons of recycled aggregate and 3300 tons of recycled steel produced by the resource disposal module is pushed to the intelligent market allocation module in real time. The reinforcement learning model of this module runs in a highly simulated virtual market environment, considering real-time inventory (4200 tons of recycled aggregate inventory), new order demand (a highway project urgently needs 5000 tons of aggregate), and dynamic logistics costs, and within 3 minutes, it automatically matches and locks three optimal potential buyers for recycled aggregate, and plans a cost-optimal path that shortens the transportation distance by 18%, successfully reducing the overall transportation cost of the project to 16.3% of the total disposal cost.

[0166] The carbon accounting module throughout the whole process of the project is based on the system LCA model, automatically collects the whole chain data from transportation (total mileage 12,500 kilometers, average load 15 tons) to production line processing (total energy consumption 185,000 kWh), and accurately calculates the carbon emission of the project in the processing link as 1,150 tons of CO2e, but realizes carbon emission reduction of 1,570 tons of CO2e due to the replacement of renewable materials with primary resources, finally contributes net carbon emission reduction benefit of 420 tons of CO2e to the project. At the same time, these refined carbon data are fed back to the dynamic accounting and market module in real time, used to optimize the transportation route of future vehicles (the system preferentially selects a path that can reduce carbon emission by 5%) and market recommendation of products (a "green priority" label is given to low-carbon renewable products on the trading interface), and finally builds a complete intelligent closed loop from accurate prediction, dynamic supervision, efficient disposal, intelligent transaction to carbon trace management, realizes the maximization and unification of environmental benefits and economic benefits.

[0167] The building waste intelligent prediction, classification and resource disposal system of the application integrates environmental engineering, computer vision, Internet of Things, machine learning and smart city management and is mainly applied to intelligent prediction, dynamic monitoring, accurate classification, resource disposal and carbon footprint management of building waste, can realize automatic and digital management of the whole chain of building waste from generation to recycling, and provides an intelligent solution for environmental protection industry, engineering construction and supervision.

[0168] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit them; although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalents; without departing from the spirit of the technical solutions of the present application, they should be covered in the technical solution range of the present application claimed.

Claims

1. A multi-modal AI-based intelligent prediction, classification and resourceful disposal system for construction waste, characterized in that: Comprise: Data extraction module: Obtain planning files and statistical files of regulatory departments through API interface to extract basic engineering information and design parameters; Prediction module: Predict the amount of construction waste based on historical data in planning files and statistical files and machine learning models; Dynamic accounting module: Collect multi-angle field images through drones or cameras, preprocess and data label, identify construction waste volume and waste classification using deep learning models, generate three-dimensional models with classification labels, calculate actual volume and prediction difference, assist in supervision, and use reinforcement learning models to customize photo shooting frequency. Adjust the pulling strategy according to the prediction module prediction results and the deep learning model output construction waste classification results and production line load; Resource disposal module: According to the construction waste classification results and real-time capacity data output by the dynamic accounting module, use the reinforcement learning model to dynamically allocate to the optimal resource disposal production line; The process of closed-loop optimization and feedback of the reinforcement learning model in the dynamic accounting module includes: the dynamic accounting module pushes the construction waste related data to the reinforcement learning model, generates allocation instructions and issues them to the resource disposal module to control the production line; After each batch is processed, actual resource recovery rate, energy consumption and other data are collected, rewards are calculated and strategies are updated; The actual resource recovery rate, processing time and energy consumption data, and garbage composition deviation of various types of garbage are fed back to the dynamic accounting module; Intelligent market allocation module: Train reinforcement learning model to automatically match market demand according to the characteristics of recycled products to optimize logistics allocation; The content of the reinforcement learning model in the intelligent market allocation module includes: Step S61, environment construction: build a virtual market environment through Agent-Based Modeling, simulate the dynamic interaction of customers, suppliers and logistics companies, and generate extreme cases by combining historical data enhancement technology; Step S62, reinforcement learning model training: State space: Real-time inventory, specifications, quality grades of various recycled products; Order type, demand, priority, regional distribution; Transportation vehicle availability, route congestion; Policy subsidy changes, raw material price fluctuations; Action space: Assign products to designated orders; Transportation vehicle scheduling, path planning; Dynamically adjust product prices to balance supply and demand; Reward function: Increase order satisfaction rate, reduce logistics cost, and improve inventory turnover rate as positive reward; Order default, resource waste as negative reward; Model training: Based on historical orders, inventory and logistics data, build a simulation environment, pre-train using reinforcement learning, and optimize the initial strategy; After deployment, fine-tune model parameters through real-time feedback, and manually review high-risk decisions; Step S63, closed-loop optimization and feedback: Resource disposal module pushes recycled product data to intelligent market allocation module, external data sources synchronize in real time through API interface, and market feedback is transmitted to resource disposal module to dynamically adjust production line priority; Carbon accounting module: Based on the life cycle assessment model LCA, use machine learning models to calculate carbon emissions in transportation and processing and the emission reduction amount of recycled products.

2. The intelligent construction waste prediction, classification and recycling disposal system according to claim 1, characterized in that: The basic engineering information includes engineering address, engineering type, land use type, engineering area, construction period; the design parameters include structure type, material list in design drawings, demolition and reconstruction plan.

3. The intelligent construction waste prediction, classification and recycling disposal system according to claim 2, characterized in that: The step of the data grabbing module grabbing the basic engineering information and the design parameters comprises: Step S11, submitting an API interface call application to a regulatory department, and using Python to perform data grabbing; Step S12, mapping fields returned by the API into internal standard fields of the system, and then performing data cleaning, filtering invalid data, and standardizing the data, and filling in missing items as default values or marking as "to be completed"; Step S13, storing the cleaned data in a relational database, and synchronizing newly added or modified data at a fixed time.

4. The intelligent construction waste prediction, classification, and resource disposition system of claim 1, wherein: In the prediction module, the training model of the machine learning model comprises: Feature encoding: selecting engineering type, land use type, engineering area, structure type, material consumption, construction period, etc. as key features, and using model feature importance to screen high-contribution features; then, standardizing and normalizing numerical variables, and using one-hot encoding for classification variables; Model training: selecting a random forest algorithm to predict the amount of garbage and classification; dividing the data into a training set, a validation set, and a test set in a ratio of 6:2:2, using 5-fold cross-validation to optimize hyperparameters, and selecting MAE, RMSE, and R² to evaluate the regression model for predicting the continuous amount of construction waste, and using F1-score to evaluate the construction waste type identified by the classification model.

5. The intelligent construction waste prediction, classification, and resource disposition system of claim 1, wherein: The steps of collecting multi-angle site images by using a drone or a camera and performing preprocessing and data labeling comprise: Image collection: using a drone and an RGB camera to collect multi-angle images and videos of the construction waste pile; when the construction period is long, installing a fixed camera to cover the main area of the waste pile, ensuring that the image heading overlap reaches 70-80% and the lateral overlap reaches 30-50%; using LIDAR to collect depth data in the face of complex occlusion scenarios; Preprocessing: using bilateral filtering to denoise, and combining CLAHE to enhance contrast; identifying feature points of the same object in different images through SIFT feature matching, establishing a correspondence between images, and using RANSAC to remove false matching points; Data labeling: labeling the preprocessed images and videos, and the labeling content includes garbage categories, bounding boxes, and pixel-level masks, which are used to train a deep learning model.

6. The intelligent construction waste prediction, classification, and resource disposition system of claim 1, wherein: The reinforcement learning model in the dynamic calculation module comprises: State space: construction site dynamic data, including the number of days remaining in the construction period, the waste accumulation rate, and the weather in the next month; real-time data of the prediction module, including garbage quantity prediction value, deep learning classification result, and production line load; Action space: scheduling vehicle priority, transportation route optimization, and image shooting frequency and coverage range; Reward function: setting shooting key accumulation areas and finding classification abnormalities as positive rewards for shooting frequency decision; setting repeated shooting of invalid areas as negative rewards for shooting frequency decision; setting production line load matching and transportation delay reduction as positive rewards for pulling strategy; setting production line overload and inefficient transportation route as negative rewards for pulling strategy; Model training: Entropy regularization SAC is used for model training, which continuously balances exploration and utilization in the closed loop of "input state-policy network generates action-environment feedback reward-updates policy". After each round of training, the system evaluates the core indicators of coverage rate, efficiency and total cost saving in parallel, and real-time tracks the optimization effect of the policy to guide the next round of training.

7. The intelligent construction waste prediction, classification, and resource disposition system of claim 1, wherein: In the dynamic accounting module, the assistance supervision content includes: Setting threshold rules: dynamically adjusting thresholds based on historical data and engineering files using statistical methods; Attribution analysis: when the predicted quantity is less than the actual quantity and the difference exceeds the threshold, the project progress is accelerated and unregistered garbage is mixed; when the predicted quantity is greater than the actual quantity and the difference exceeds the threshold, the project is delayed and the classification equipment misjudges; Closed-loop optimization and feedback: Optimizing the prediction module using difference data: record high difference cases, label the real cause, and feed the corrected data back to the machine learning model of the prediction module to adjust the feature weights; Optimizing the dynamic accounting module using difference data: intensify data auditing for high-frequency difference construction sites, adjust image acquisition frequency, and conduct targeted training on garbage categories with high misjudgment frequency.

8. The intelligent construction waste prediction, classification, and resource disposition system of claim 1, wherein: The life cycle assessment model LCA includes: System boundary: covering the whole life cycle of construction waste from transportation, classification, treatment to recycled product utilization; Emission factor library: integrating authoritative database emission factors, including: transportation emission factors, energy consumption emission factors, and carbon offset factors of recycled products; Calculation design: ； ； 。 9. The intelligent construction waste prediction, classification, and resource disposition system of claim 1, wherein: In the carbon accounting module, the working steps of the machine learning model include: Step S71, extract garbage types, treatment processes, transportation paths, and energy types as key features, and perform one-hot encoding on classification variables and standardize numerical values; Step S72, use random forest to predict carbon emissions at each link, and use cross-validation to optimize hyperparameters, selecting RMSE and R² as evaluation indicators; Step S73, visualize the carbon emission heat map according to the regional distribution of the output results, and output the report at the same time.

10. The intelligent construction waste prediction, classification, and resource disposition system of claim 9, wherein: In the carbon accounting module, the closed-loop optimization and feedback process of the machine learning model includes: Feed the carbon emission data back to the dynamic accounting module and the intelligent market allocation module to optimize the transportation route or treatment priority; and link with the intelligent market allocation module to preferentially recommend low-carbon recycled product orders; record the difference between carbon emissions and predicted values, and update model parameters and emission factor library regularly.

Images