AI-based building construction whole-process energy-saving and environment-friendly monitoring optimization system and method

Abstract

The application relates to the technical field of intelligent management of building construction, and discloses an AI-based building construction whole-process energy-saving and environment-friendly monitoring optimization system and method, which comprises a perception layer arranged at a construction site, a data layer comprising an edge computing node and a standardized data engine, the standardized data engine comprising a construction behavior recognition module and a unified data format generation module, the unified data format generation module fusing heterogeneous sensing data and structured construction behavior records according to a preset unified space-time reference and data format to generate standard construction monitoring data flow, an AI analysis layer comprising a construction environment-friendly correlation model, a high-energy-consumption and high-pollution identification module and a dynamic optimization decision module, the dynamic optimization decision module generating a dynamic optimization scheme based on a genetic algorithm, and an application layer comprising a visual interactive platform and an instruction pushing module; the system realizes real-time unified collection, intelligent analysis and dynamic optimization decision closed-loop control of multi-dimensional environment-friendly data in the whole process of building construction.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for building construction, specifically to an AI-based energy-saving and environmental protection monitoring and optimization system and method for the entire building construction process. Background Technology

[0002] With the deepening of the "dual-carbon" strategy and the continuous improvement of green building standards, energy conservation and environmental protection requirements in the construction field are becoming increasingly stringent. Currently, monitoring of environmental indicators such as dust, noise, energy consumption, and building material consumption during the construction process has become increasingly widespread, and various sensors and monitoring equipment are widely used on construction sites.

[0003] However, existing technologies have the following significant drawbacks: First, monitoring data is severely fragmented and heterogeneous. The types of sensors deployed at construction sites are diverse, including dust monitors (optical scattering principle), noise sensors (acoustic acquisition), electricity and gas meters (pulse / Modbus protocol), and material weighing systems (industrial bus). These devices generate data with different formats, sampling frequencies, and spatiotemporal references, creating "data silos." Traditional technologies often employ an independent operation mode for each subsystem, lacking an effective data fusion mechanism, resulting in environmental monitoring data failing to form a unified basis for decision-making.

[0004] Second, real-time performance and accuracy are difficult to balance. To achieve end-to-end monitoring, multimodal data, including video, audio, and time-series sensor data, needs to be processed, resulting in a massive data volume. Uploading all data to the cloud for processing creates network bandwidth and latency bottlenecks; processing only at the edge, however, lacks the computing power to support complex AI analysis. Existing technologies either sacrifice real-time performance for batch processing or sacrifice accuracy for simplified models, failing to simultaneously meet the needs of dynamic construction optimization.

[0005] Third, optimization decisions are lagging and lack proactive control capabilities. The typical operating mode of existing monitoring systems is alarm triggering followed by manual intervention; that is, when environmental indicators exceed thresholds, an alarm is triggered, and managers manually assess the situation and issue adjustment instructions. This mode suffers from a delayed response (usually taking several hours) and relies on personal experience, failing to achieve dynamic optimization of construction parameters. While some systems incorporate BIM models for visualization, this remains merely a digital mapping, failing to form a closed-loop control system with construction equipment.

[0006] Fourth, there is a disconnect between energy conservation and environmental protection and construction efficiency. Current technologies often treat energy conservation and environmental protection as construction constraints, adopting a "passive compliance" strategy. For example, blindly increasing sprinkler systems to reduce dust or halting work to reduce noise often comes at the expense of construction efficiency. There is a lack of systematic technical solutions for optimizing energy conservation and environmental protection while ensuring construction progress and quality. Summary of the Invention

[0007] The present invention aims to provide an AI-based energy-saving and environmental protection monitoring and optimization system and method for the entire construction process, in order to solve the technical problems of fragmented energy-saving and environmental protection monitoring in existing construction projects, difficulty in real-time integration of heterogeneous data, and optimization decisions lagging behind construction dynamics.

[0008] To solve the above problems, the present invention adopts the following technical solution: Option 1: An AI-based energy-saving and environmental protection monitoring and optimization system for the entire building construction process, including: The perception layer, deployed at the construction site, includes a multi-dimensional sensor group, monitoring camera devices, and audio acquisition devices; the multi-dimensional sensor group includes dust sensors, noise sensors, energy consumption metering devices, and building material consumption monitoring devices; the monitoring camera devices are used to collect video images of construction activities; and the audio acquisition devices are used to collect the sounds of construction machinery in operation. The data layer includes edge computing nodes and a standardized data engine. The edge computing nodes are communicatively connected to the perception layer and are used to preprocess heterogeneous sensor data collected by multi-dimensional sensor groups. The standardized data engine includes a construction behavior recognition module and a unified data format generation module. The construction behavior recognition module automatically identifies the type of construction behavior, the type of construction machinery, and the operating status based on video images collected by the monitoring camera and mechanical operation sounds collected by the audio acquisition device, using computer vision algorithms and sound recognition algorithms, and generates structured construction behavior records. The unified data format generation module fuses the heterogeneous sensor data and the structured construction behavior records according to a preset unified spatiotemporal benchmark and data format to generate a standard construction monitoring data stream. The AI ​​analysis layer includes a construction environmental protection correlation model, a high-energy-consumption and high-pollution identification module, and a dynamic optimization decision-making module. The construction environmental protection correlation model stores national environmental protection standard data, construction specification data, and a historical construction environmental protection correlation knowledge base. The high-energy-consumption and high-pollution identification module uses a random forest algorithm to perform real-time analysis on the standard construction monitoring data stream to identify high-energy-consumption and high-pollution stages in the current construction process. The dynamic optimization decision-making module uses a genetic algorithm, taking the high-energy-consumption and high-pollution stages as inputs and combining them with construction progress constraints, to generate a dynamic optimization scheme. The application layer includes a visual interaction platform and an instruction push module. The visual interaction platform displays environmental indicators, energy consumption curves, and optimization suggestions in real time. The instruction push module automatically converts the dynamic optimization scheme into equipment control instructions and pushes them to the construction equipment terminal for execution, forming a closed loop of monitoring, analysis, optimization, and execution.

[0009] Beneficial Effects: This invention automatically identifies construction behaviors based on video images and sound through a construction behavior recognition module, transforming heterogeneous multimodal data into structured records in a unified format, thus solving the data heterogeneity problem without adding additional sensors. The collaboration between edge computing nodes and a standardized data engine enables real-time data processing and fusion, meeting real-time requirements. The linkage between the AI ​​analysis layer and the application layer achieves an automatic closed loop from recognition to execution, significantly improving response speed. This invention realizes closed-loop control of real-time unified collection, intelligent analysis, and dynamic optimization decision-making of multi-dimensional environmental data throughout the entire construction process, achieving a synergistic improvement in energy conservation, environmental protection, and construction efficiency.

[0010] Preferably, the construction behavior recognition module includes: The video analysis submodule uses a convolutional neural network to identify construction behaviors in the video images. The identified content includes the type of construction machinery, work actions, material handling trajectories, and personnel activity areas. The video analysis submodule processes the video stream at a sampling frequency of 5 frames per second to generate a structured video record containing timestamps, spatial coordinates, and behavior types. The audio analysis submodule identifies the mechanical operating sound based on Mel frequency cepstral coefficient feature extraction and support vector machine classifier. The identification content includes the type of machinery, operating load rate, and abnormal operating conditions. The audio analysis submodule performs short-time Fourier transform analysis at a frequency of 10 times per second to generate a structured audio record containing timestamps, sound source location, and mechanical status. The data fusion submodule performs time alignment and spatial matching between the structured video recording and the structured audio recording. When the video recognition result is consistent with the audio recognition result, a structured construction behavior record with a confidence level higher than 0.85 is generated.

[0011] Beneficial effects: The accuracy and robustness of construction behavior recognition are improved through dual-modal recognition and fusion verification of video and audio; the specific sampling frequency setting controls the computational load while ensuring real-time performance; and the confidence threshold setting ensures the quality of data entering downstream analysis.

[0012] Preferably, the standard construction monitoring data stream generated by the unified data format generation module includes: The spatiotemporal reference field includes a timestamp that uniformly adopts Beijing time and spatial coordinates based on the construction plane coordinate system, with time accuracy at the second level and spatial accuracy at the meter level. The data type identifier field uses a four-digit code to identify the data source type. The first digit represents the data category, where 1 represents sensor data, 2 represents video recognition data, and 3 represents audio recognition data. The last three digits represent the specific subtype. The numerical content field uses a floating-point array to store the standardized monitoring values. The array length is fixed at 16 elements. If the length is insufficient, it is filled with null values. If the length is excessive, it is divided into multiple records using time slices. The associated behavior field establishes a foreign key association with the structured construction behavior record, which is used to trace the construction context in which the data was generated.

[0013] Beneficial effects: A unified spatiotemporal benchmark ensures the alignment of multi-source data; a four-bit encoding system enables rapid parsing and classification of data types; a fixed-length numerical array design simplifies the database structure and stream processing logic; and the related behavior field preserves the semantic association between data and construction behavior, providing contextual support for subsequent analysis.

[0014] Preferably, the edge computing node includes: The sensor data preprocessing unit performs outlier removal, moving average filtering, and sampling rate normalization on the heterogeneous sensor data, and converts the different original sampling frequencies into standard time-series data with a 1-minute interval. The cache queue management unit uses a circular buffer structure to cache the preprocessed data of the most recent 72 hours. When the network is interrupted, it is stored locally and re-transmitted in batches after the network is restored. The priority scheduling unit assigns priority to the standard construction monitoring data stream, with dust and noise data assigned the highest priority and a transmission delay of less than 5 seconds; energy consumption and building material data assigned the normal priority and a transmission delay of less than 60 seconds.

[0015] Beneficial effects: The preprocessing function of edge computing nodes reduces the computing pressure on the cloud; 72-hour local caching ensures data integrity; and the priority scheduling mechanism ensures the real-time transmission of key environmental indicators, meeting the differentiated real-time requirements of different data types.

[0016] Preferably, the dynamic optimization decision module includes: A multi-objective optimization objective function construction unit is used to construct a three-dimensional Pareto frontier search space with energy consumption reduction rate, pollution emission reduction rate and construction period delay risk as optimization objectives. The constraint processing unit incorporates construction quality specification constraints, safe operating procedure constraints, and equipment performance boundary constraints, excluding infeasible solutions from the search space. The genetic algorithm solver is set with a population size of 100, an iteration number of 50, a crossover probability of 0.8, and a mutation probability of 0.1, and obtains a non-dominated solution set. The scheme selection unit selects the scheme with the highest overall satisfaction from the non-dominated solution set based on the analytic hierarchy process (AHP) as the dynamic optimization scheme.

[0017] Beneficial effects: The clear multi-objective optimization architecture realizes the synergistic optimization of energy conservation, environmental protection and construction efficiency; the constraint processing unit ensures the executability and safety of the optimization scheme; the specific genetic algorithm parameter settings control the computation time while ensuring the solution quality; the introduction of the analytic hierarchy process enables quantitative decision-making on multi-objective trade-offs.

[0018] Option 2: An AI-based optimization method for energy conservation and environmental protection monitoring throughout the entire building construction process, implemented using the aforementioned system, includes the following steps: S1. Multi-source heterogeneous data acquisition steps: Heterogeneous sensor data is acquired by deploying a multi-dimensional sensor group at the construction site, while video images of construction behavior are acquired by monitoring camera devices, and the sound of construction machinery operation is acquired by audio acquisition devices. S2. Edge preprocessing steps: Use edge computing nodes to preprocess the heterogeneous sensing data, including outlier removal, filtering and sampling rate normalization; S3. Intelligent Recognition Steps for Construction Behavior: Using the construction behavior recognition module in the standardized data engine, based on the video images and mechanical operation sounds, computer vision algorithms and sound recognition algorithms are used to automatically identify the type of construction behavior, the type of construction machinery and its operating status, and generate a structured construction behavior record. S4. Data standardization and fusion steps: Using a unified data format generation module, the preprocessed heterogeneous sensor data and the structured construction behavior records are fused according to a preset unified spatiotemporal benchmark and data format to generate a standard construction monitoring data stream. S5. High energy consumption and high pollution identification steps: The high energy consumption and high pollution identification module uses the random forest algorithm to perform real-time analysis on the standard construction monitoring data stream to identify the high energy consumption and high pollution links in the current construction process. S6. Dynamic optimization decision-making steps: Using the dynamic optimization decision-making module based on the genetic algorithm, the identified high-energy-consuming and high-polluting links are used as inputs, combined with the construction progress constraints, to generate a dynamic optimization scheme. S7. Closed-loop execution steps: The dynamic optimization scheme is automatically converted into equipment control commands and pushed to the construction equipment terminal for execution, forming a closed loop of monitoring, analysis, optimization and execution.

[0019] Beneficial effects: The method of this invention fully covers the entire process from data acquisition to execution closed loop; by recognizing construction behavior through video and sound and generating records in a unified format, it innovatively solves the technical problem of multi-source heterogeneous data fusion; the organic connection of each step realizes the intelligent and automated management of construction environmental protection.

[0020] Preferably, the intelligent recognition step of construction behavior in step S3 includes: S3.1. Extract frames from the video stream at a sampling frequency of 5 frames per second, input the data into a convolutional neural network to identify the type of construction machinery, work actions, material handling trajectories, and personnel activity areas, and generate a structured video record containing timestamps, spatial coordinates, and behavior types. S3.2 Perform a short-time Fourier transform on the audio signal at a frequency of 10 times per second, extract the Mel frequency cepstral coefficient features, input the data into a support vector machine classifier to identify the type of machinery, operating load rate and abnormal working conditions, and generate a structured audio record containing timestamps, sound source location and machinery status. S3.3. Perform time alignment and spatial matching between the structured video recording and the structured audio recording. When the recognition results of the two are consistent and the confidence level is higher than 0.85, output the final structured construction behavior record.

[0021] Beneficial effects: The specific sampling frequency setting balances recognition accuracy and computational overhead; the dual-modal fusion verification mechanism significantly reduces the false judgment rate of a single recognition method; and confidence threshold filtering ensures the reliability of data in downstream analysis.

[0022] Preferably, the standard construction monitoring data stream generated by the data standardization and fusion step in step S4 includes: The timestamps are uniformly adopted using Beijing time and the spatiotemporal reference field is based on the spatial coordinates of the construction plane coordinate system; The data type identifier field uses a four-digit encoding; The numerical content field uses a fixed-length 16-element floating-point array. If the length is insufficient, it is filled with null values, and if the length is excessive, it is divided by time slices. Establish a related behavior field that is associated with the foreign key of the structured construction behavior record.

[0023] Beneficial effects: The unified data format standardizes the fusion output of multi-source heterogeneous data, providing standardized input for subsequent AI analysis; the fixed-length design simplifies stream processing and database storage; and foreign key associations preserve the semantic information of the data.

[0024] Preferably, the dynamic optimization decision-making step in step S6 includes: S6.1 Construct a three-dimensional Pareto frontier search space with energy consumption reduction rate, pollution emission reduction rate and construction delay risk as optimization objectives; S6.2. Incorporate construction quality specifications, safe operating procedures, and equipment performance boundary constraints, and exclude infeasible solutions; S6.3. Using genetic algorithm parameters of population size 100, number of iterations 50, crossover probability 0.8, and mutation probability 0.1, solve for the non-dominated solution set; S6.4. Based on the analytic hierarchy process (AHP), select the solution with the highest overall satisfaction from the non-dominated solution set as the dynamic optimization solution.

[0025] Beneficial effects: The multi-objective optimization framework achieves synergy between energy conservation, environmental protection, and construction efficiency; constraint handling ensures the feasibility of the solution; specific algorithm parameters control the solution time while ensuring optimization quality; and the analytic hierarchy process (AHP) provides a quantitative decision-making basis for multi-objective trade-offs.

[0026] Preferably, the closed-loop execution step in step S7 includes: The dynamic optimization scheme is analyzed into specific equipment control parameters, including concrete mixer speed adjustment value, spraying system start-stop sequence and flow rate setting value, and transport vehicle scheduling route and load limit; Perform safety boundary verification on the generated equipment control parameters; The verified control commands are pushed to the PLC controller of the construction equipment or the vehicle terminal using the MQTT over TLS encryption protocol. A 3-second transmission confirmation timeout is set, and a retransmission is triggered if the timeout occurs.

[0027] Beneficial effects: The optimized scheme is transformed into directly executable equipment control parameters, realizing automation from decision-making to execution; security boundary verification prevents the issuance of dangerous commands; encrypted transmission and timeout retransmission mechanisms ensure industrial control security and command delivery reliability.

[0028] (a) Advantages of the present invention This invention, through an innovative four-layer architecture encompassing perception, data, AI, and applications, achieves a leapfrog upgrade in building construction energy conservation and environmental protection management, moving from passive compliance to proactive optimization, from manual experience to intelligent decision-making, and from single-point monitoring to a closed-loop process. Specific advantages include: First, the originality of the data fusion mechanism. It innovatively introduces automatic construction behavior recognition technology based on video image and sound recognition, which integrates heterogeneous multimodal sensor data with semantic construction behavior records. Without relying on complex technologies such as digital twins, it solves the problem of real-time fusion of multi-source heterogeneous data in a low-cost software manner.

[0029] Second, optimize the synergy of objectives. Breaking away from the traditional mindset that energy conservation and environmental protection are opposed to construction efficiency, the optimal trade-off is sought on the Pareto frontier through a multi-objective genetic algorithm, achieving a synergistic optimization effect of "reducing consumption without reducing efficiency and reducing emissions without delaying progress".

[0030] Third, the real-time nature of the response mechanism. The complete closed-loop control from data acquisition to command execution is completed within seconds to minutes, which is two orders of magnitude faster than the response cycle of several hours for traditional over-limit alarms and manual handling.

[0031] Fourth, the flexibility of application deployment. The knowledge base switching mechanism ensures compatibility with both traditional civil engineering and prefabricated building scenarios, avoiding hardware modifications during scenario switching and significantly reducing the system's total lifecycle cost.

[0032] (ii) Unexpected solutions to technical problems When faced with the technical contradiction between data heterogeneity and real-time performance, existing technologies typically adopt two extreme strategies: one is to use complex technologies such as digital twins for full-element modeling, which can solve heterogeneity but makes it difficult to guarantee real-time performance and is costly; the other is to abandon data fusion, with each subsystem operating independently, which ensures real-time performance but fails to form a unified decision-making mechanism. This invention unexpectedly discovers that existing monitoring cameras and audio acquisition equipment at construction sites, enhanced with appropriate AI algorithms, can become a unified data fusion hub: video and sound, as universal sensors, can identify various construction behaviors and generate standardized semantic records, thereby unifying heterogeneous physical sensor data to the behavioral semantic level. This approach breaks away from the traditional thinking of adding hardware to unify protocols or complex software middleware, and solves the technical contradiction at low cost through software-defined perception.

[0033] Furthermore, existing technologies generally consider the automatic optimization instructions pushed to construction equipment to pose safety risks, thus limiting their application to the level of visualization. This invention unexpectedly discovers that by introducing mature constraint handling mechanisms from the process industry and a hierarchical human-machine permission design, automatic control within a specific parameter range can be achieved while ensuring safety, transforming the monitoring, analysis, optimization, and execution closed loop from a theoretical concept into an engineering reality.

[0034] (iii) Unexpected Technological Means Technical Approach 1: Dual-Modal Construction Behavior Recognition. Introducing computer vision and sound recognition technologies into the field of construction environmental monitoring is not immediately obvious, as traditional thinking considers construction behavior recognition a production management requirement, unrelated to environmental monitoring. This invention unexpectedly combines the two, discovering that construction behavior is a key context for interpreting variations in environmental data. For example, the same dust emission value can have drastically different meanings and require different handling strategies depending on whether it's related to earthwork excavation or concrete pouring. By unifying data semantics through behavior recognition, the interpretability and actionability of environmental data are achieved.

[0035] Technical Approach Two: Cascaded Architecture of Genetic Algorithm and Random Forest. Random forest is used for identification (classification problems), and genetic algorithm is used for optimization (combinatorial optimization problems). The cascaded use of these two is not standard practice in the AI ​​field. This invention unexpectedly found that this combination is particularly suitable for construction environmental optimization scenarios: the high robustness of random forest adapts to the high noise level of construction site data, and the global search capability of genetic algorithm adapts to the complex optimization needs of multiple objectives and constraints. The combination of the two achieves a synergistic effect of accurate identification and global optimization.

[0036] Technical Means 3: Standard Data Format for Fixed-Length Arrays. In the field of data engineering, fixed-length designs are often seen as lacking in flexibility. This invention unexpectedly adopts a 16-element fixed-length floating-point array as the standard format. It has been found that this design greatly simplifies the development of stream processing operators and the schema design of databases. In scenarios where edge computing resources are limited, the simple data structure actually results in higher processing throughput and lower latency, representing a successful practice of trading structural rigidity for processing efficiency. Attached Figure Description

[0037] Figure 1 This is a system architecture diagram of the present invention.

[0038] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0039] The following detailed description illustrates the specific implementation method: This invention provides an AI-based energy-saving and environmental protection monitoring and optimization system and method for the entire construction process. It aims to solve the technical problems of fragmented energy-saving and environmental protection monitoring, delayed optimization decisions, and the disconnect between environmental protection and efficiency in the construction field through innovative data fusion mechanisms, intelligent analysis algorithms, and automatic execution closed loops.

[0040] like Figure 1 As shown, the overall architecture of the AI-based building construction process energy conservation and environmental protection monitoring and optimization system of the present invention includes four layers: perception layer, data layer, AI analysis layer and application layer.

[0041] The perception layer deploys multi-dimensional sensor groups, monitoring cameras, and audio acquisition devices to achieve comprehensive data collection at the construction site.

[0042] The perception layer, deployed at the construction site, includes a multi-dimensional sensor group, a monitoring camera device, and an audio acquisition device; the multi-dimensional sensor group includes a dust sensor, a noise sensor, an energy consumption metering device, and a building material consumption monitoring device; the monitoring camera device is used to collect video images of construction activities; and the audio acquisition device is used to collect the operating sounds of construction machinery.

[0043] The data layer uses edge computing nodes and a standardized data engine to preprocess and unify the format of heterogeneous data. In particular, the construction behavior recognition module automatically identifies construction behavior based on video images and sounds, generates structured records, and resolves the contradiction between data heterogeneity and real-time performance.

[0044] The data layer includes edge computing nodes and a standardized data engine. The edge computing nodes are communicatively connected to the perception layer and are used to preprocess heterogeneous sensor data collected by multi-dimensional sensor groups. The standardized data engine includes a construction behavior recognition module and a unified data format generation module. The construction behavior recognition module automatically identifies the type of construction behavior, the type of construction machinery, and the operating status based on video images collected by the monitoring camera and mechanical operation sounds collected by the audio acquisition device, using computer vision algorithms and sound recognition algorithms, and generates structured construction behavior records. The unified data format generation module fuses the heterogeneous sensor data and the structured construction behavior records according to a preset unified spatiotemporal benchmark and data format to generate a standard construction monitoring data stream.

[0045] The edge computing nodes include: The sensor data preprocessing unit performs outlier removal, moving average filtering, and sampling rate normalization on the heterogeneous sensor data, and converts the different original sampling frequencies into standard time-series data with a 1-minute interval. The cache queue management unit uses a circular buffer structure to cache the preprocessed data of the most recent 72 hours. When the network is interrupted, it is stored locally and re-transmitted in batches after the network is restored. The priority scheduling unit assigns priority to the standard construction monitoring data stream, with dust and noise data assigned the highest priority and a transmission delay of less than 5 seconds; energy consumption and building material data assigned the normal priority and a transmission delay of less than 60 seconds.

[0046] In one specific implementation, the capacity of the circular buffer is calculated as: 72 hours × 3600 seconds / hour × 100 records / second = 25,920,000 records. Assuming an average record size of 500 bytes, the total capacity is approximately 12.96 GB. A dual-buffer alternating write strategy is employed: when Buffer-A is full, the system switches to Buffer-B, while Buffer-A is asynchronously flushed to disk, ensuring zero data loss.

[0047] The construction behavior recognition module includes a video analysis submodule, an audio analysis submodule, and a data fusion submodule.

[0048] The video analysis submodule performs construction behavior recognition on the video images based on a convolutional neural network. The recognition content includes the type of construction machinery, operation actions, material handling trajectory, and personnel activity area. The video analysis submodule processes the video stream at a sampling frequency of 5 frames per second to generate a structured video record containing timestamps, spatial coordinates, and behavior types.

[0049] The video analysis submodule employs a computer vision algorithm that, in one specific implementation, uses a YOLOv5 object detection network to identify construction machinery and operational behaviors. This network is a specific implementation of a convolutional neural network. Network structure configuration: The input is a 640×640 pixel RGB image. Features are extracted through the CSPDarknet53 backbone network, and multi-scale feature fusion is performed using the PANet structure to output detection heads at three scales (13×13, 26×26, and 52×52). Detection categories include: six types of construction machinery such as tower cranes, construction elevators, concrete pump trucks, excavators, loaders, and rebar processing machinery; and five types of work actions such as hoisting, excavation, pouring, processing, and transportation. Simultaneously, the spatial coordinates of the machinery and personnel (pixel coordinates of the bounding box center point, converted to construction plane coordinates after camera calibration) and material handling trajectories (temporal trajectories after cross-frame association) are output.

[0050] Model Training and Performance: The network was pre-trained on a self-built construction scenario dataset containing 150,000 labeled images, covering diverse working conditions such as sunny days, cloudy days, nighttime, and dusty conditions. Considering the comprehensive system overhead of video decoding, image preprocessing, network transmission, and post-processing, the actual end-to-end processing latency is approximately 180-200ms. Therefore, the system sampling frequency was set to 5 frames per second (processing one frame every 200ms) to control the computational load and energy consumption of edge computing nodes while meeting real-time requirements. This sampling frequency was found to be sufficient to capture typical movements of construction machinery (e.g., the lifting speed of a tower crane hook is approximately 0.5m / s, and 5 frames / second can resolve displacements of up to 10cm).

[0051] As an alternative implementation, the Faster R-CNN two-stage detector can be used, which has higher accuracy but is slower and is suitable for scenarios where real-time requirements are not high; or the lightweight model MobileNetV3-YOLO can be used, which can reduce the single-frame inference time to less than 5ms and is suitable for scenarios where edge computing resources are limited.

[0052] The audio analysis submodule identifies the mechanical operating sound based on Mel frequency cepstral coefficient feature extraction and support vector machine classifier. The identification content includes the type of machinery, operating load rate, and abnormal operating conditions. The audio analysis submodule performs short-time Fourier transform analysis at a frequency of 10 times per second to generate a structured audio record containing timestamps, sound source location, and machinery status.

[0053] The audio analysis submodule employs a sound recognition algorithm, which, in one specific implementation, is achieved through the following steps: Step 1: Preprocessing. The original audio signal is pre-emphasized (coefficient 0.97), framed (frame length 25ms, frame shift 10ms), and Hamming windowed.

[0054] Step 2: Feature Extraction. Perform Fast Fourier Transform (FFT, 512 points) on each frame of signal to calculate the power spectrum; filter through 26 triangular filter banks (Mel scale, frequency range 20Hz-8000Hz), take the logarithm and perform Discrete Cosine Transform (DCT), retain the first 13 dimensions of coefficients to obtain the MFCC (Mel frequency cepstral coefficients) feature vector; add first-order and second-order differences to finally form a 39-dimensional feature vector.

[0055] Step 3: Classification and Recognition. A Radial Basis Function (RBF) support vector machine was used for classification, with kernel parameter γ=0.01 and penalty coefficient C=10. The classification categories included: tower crane operation (light load, medium load, heavy load, abnormal), concrete pump truck operation (idling, pumping, cleaning), construction elevator operation (ascending, descending, braking), and environmental background noise, totaling 8 categories. The classification accuracy reached 91.5% on the test set, with a processing time of 8ms per audio clip (1 second).

[0056] The audio analysis submodule performs short-time Fourier transform analysis at a frequency of 10 times per second (corresponding to 100 frames per second of audio, with a frame shift of 10ms) to generate a structured audio record containing timestamps, sound source locations, and mechanical states.

[0057] As an alternative implementation, deep learning methods, such as end-to-end recognition based on one-dimensional convolutional neural networks (1D-CNN) or recurrent neural networks (LSTM), can also be used, which are suitable for scenarios with sufficient training data.

[0058] The data fusion submodule performs time alignment and spatial matching between the structured video recording and the structured audio recording. When the video recognition result is consistent with the audio recognition result, a structured construction behavior record with a confidence level higher than 0.85 is generated.

[0059] The data fusion submodule performs time alignment and spatial matching between the structured video recordings and the structured audio recordings. Time alignment uses a nearest neighbor matching strategy: for the timestamp tv of the video recording, it searches for the record with the smallest time difference |ta - tv| in the audio recording time series; when the time difference is less than 100ms, it is considered a matching record from the same moment. Spatial matching uses region overlap determination: when the distance between the mechanical bounding box recognized by the video and the sound source location (located by the microphone array) recognized by the audio is less than 5 meters, it is considered a spatial match. When the video recognition result and the audio recognition result are consistent (same mechanical type) and the confidence level is higher than 0.85, the final structured construction behavior record is generated.

[0060] The standard construction monitoring data stream generated by the unified data format generation module includes: The spatiotemporal reference field includes a timestamp that uniformly adopts Beijing time and spatial coordinates based on the construction plane coordinate system, with time accuracy at the second level and spatial accuracy at the meter level. The data type identifier field uses a four-digit code to identify the data source type. The first digit represents the data category, where 1 represents sensor data, 2 represents video recognition data, and 3 represents audio recognition data. The last three digits represent the specific subtype. The numerical content field uses a floating-point array to store the standardized monitoring values. The array length is fixed at 16 elements. If the length is insufficient, it is filled with null values. If the length is excessive, it is divided into multiple records using time slices. The associated behavior field establishes a foreign key association with the structured construction behavior record, which is used to trace the construction context in which the data was generated.

[0061] In one specific implementation, the standard construction monitoring data stream generated by the unified data format generation module can be defined in JSON format.

[0062] The four-digit coding system is defined as follows: the first digit is 1 = sensor data, 2 = video recognition data, and 3 = audio recognition data; the second digit is 0 = environmental type, 1 = machinery type, 2 = personnel type, and 3 = material type; the third and fourth digits are the specific subtype serial numbers. This coding system supports the expansion of up to 4×4×100=1600 data types.

[0063] The behavior field is linked to establish a foreign key relationship with the structured construction behavior record: In relational databases (such as MySQL) or time-series databases (such as InfluxDB), the standard construction monitoring data flow table establishes a foreign key relationship with the structured construction behavior record table through the behavior_ref field, supporting the tracing of the construction context generated by the data through SQL queries or time series query languages. For example, when analyzing the cause of excessive dust pollution during a certain period, the specific construction behavior at that time (such as "tower crane hoisting - heavy load") can be quickly located through the foreign key relationship, and then video clips and audio records can be retrieved for verification.

[0064] The AI ​​analysis layer constructs a construction and environmental protection correlation model, uses the random forest algorithm to identify high-energy-consuming and high-polluting links, and uses the genetic algorithm to generate dynamic optimization schemes to achieve multi-objective collaborative optimization.

[0065] The AI ​​analysis layer includes a construction environmental protection correlation model, a high-energy-consumption and high-pollution identification module, and a dynamic optimization decision-making module. The construction environmental protection correlation model stores national environmental protection standard data, construction specification data, and a historical construction environmental protection correlation knowledge base. The high-energy-consumption and high-pollution identification module uses a random forest algorithm to perform real-time analysis on the standard construction monitoring data stream to identify high-energy-consumption and high-pollution stages in the current construction process. The dynamic optimization decision-making module uses a genetic algorithm, taking the high-energy-consumption and high-pollution stages as inputs and combining them with construction progress constraints, to generate a dynamic optimization scheme.

[0066] In one specific implementation, the construction environmental protection association model is constructed using knowledge graph technology, including: Entity layer: Defines construction activity entities (such as concrete pouring, tower crane hoisting), environmental indicator entities (such as PM2.5, noise decibels, energy consumption kWh), equipment entities (such as tower crane A01, concrete pump truck B02), and standard entities (such as GB12523-2011 noise limits).

[0067] Relationship layer: Defines the generation relationship (construction activity → environmental protection indicators), the constraint relationship (construction activity → specifications and standards), the influence relationship (equipment parameters → environmental protection indicators), and the belonging relationship (specific equipment → equipment type).

[0068] Attribute layer: Stores the attribute values ​​of each entity, such as typical_duration=2h and typical_dust_emission=150μg / m³ for "concrete pouring"; and the daytime noise limit of 70dB and the nighttime noise limit of 55dB in "GB12523-2011".

[0069] The knowledge graph is stored in the Neo4j graph database and uses the Cypher query language to achieve related retrieval.

[0070] The high-energy-consumption and high-pollution identification module performs real-time analysis of the standard construction monitoring data stream based on the random forest algorithm. In one specific embodiment, the input features of the random forest algorithm include: Current construction behavior type encoding (16-dimensional one-hot vector, corresponding to 16 typical construction behaviors); Energy consumption time series characteristics over the last 15 minutes (including mean, variance, peak value, and trend slope, a total of 4 dimensions); Current environmental parameters (dust concentration, noise level, temperature, humidity, 4 dimensions in total); Construction schedule deviation rate (the ratio of actual progress to planned progress, expressed as a percentage, 1D).

[0071] The total input feature dimension is 25. The random forest algorithm is set to 200 decision trees, a maximum depth of 15, and a minimum number of leaf samples of 5. Gini impurity is used as the splitting criterion. The algorithm outputs the probability value of whether the current construction stage belongs to a high-energy-consuming stage or a high-polluting stage. If the probability value exceeds 0.7, it is judged as positive (i.e., identified as a high-energy-consuming or high-polluting stage).

[0072] The model uses historical project data during the training phase and optimizes parameters through 10-fold cross-validation. The final model achieves an AUC of 0.89, precision of 0.85, and recall of 0.82 on the test set, meeting the accuracy requirements for real-time identification at construction sites. The inference time per record is approximately 5ms, and it supports a data throughput of over 200 records per second.

[0073] In one specific implementation, the 16 typical construction behaviors are extracted from the common processes of civil construction and prefabricated buildings, covering the main high-energy-consuming and high-polluting links, as shown in Table 1: Table 1

[0074] The dynamic optimization decision module includes: A multi-objective optimization objective function construction unit is used to construct a three-dimensional Pareto frontier search space with energy consumption reduction rate, pollution emission reduction rate and construction period delay risk as optimization objectives. The constraint processing unit incorporates construction quality specification constraints, safe operating procedure constraints, and equipment performance boundary constraints, excluding infeasible solutions from the search space. The genetic algorithm solver is set with a population size of 100, an iteration number of 50, a crossover probability of 0.8, and a mutation probability of 0.1, and obtains a non-dominated solution set. The scheme selection unit selects the scheme with the highest overall satisfaction from the non-dominated solution set based on the analytic hierarchy process (AHP) as the dynamic optimization scheme.

[0075] The dynamic optimization decision-making module generates a dynamic optimization scheme based on a genetic algorithm. In a specific embodiment, multi-objective optimization is implemented according to the following steps: Step 1: Coding. A real-number coding method is adopted, and the chromosome length is equal to the number n of decision variables (in this embodiment, n=12, including the speed adjustment coefficients of 3 tower cranes, the flow setting values of 2 sets of sprinkler systems, the scheduling priorities of 5 transport vehicles, and the timing adjustment amounts of 2 working faces). Each gene is randomly initialized within its value range: the value range of the speed adjustment coefficient is [0.85, 1.15] (i.e., ±15%), the value range of the sprinkler flow setting value is [0, 20] liters per minute, the vehicle scheduling priority is an integer within [1, 5], and the working face timing adjustment amount is [-2, +2] hours.

[0076] Step 2: Fitness evaluation, constructing a three-dimensional Pareto front search space. For each individual in the population, it is decoded into specific construction parameters, which are substituted into the construction simulation model to calculate the values of the three objective functions: f1 (energy consumption reduction rate, %) = (base energy consumption - energy consumption after optimization) / base energy consumption × 100; f2 (pollutant emission reduction rate, %) = (base emission - emission after optimization) / base emission × 100, where the pollution emission comprehensively calculates multiple indicators including dust, noise and wastewater; f3 (construction delay risk index, 0-1) = 1 - exp(-λ·Δt), where Δt is the estimated construction delay in hours, and λ is the risk coefficient (taken as 0.1 in this embodiment).

[0077] Non-dominated sorting (NSGA-II) is used for Pareto stratification: for a minimization problem, individual p dominates individual q if and only if for all i∈{1,2,3}, fi(p)≤fi(q) and there exists j∈{1,2,3} such that fj(p)<fj(q). Individuals in the same non-dominated layer are sorted according to crowding distance, and the crowding distance is calculated as the sum of normalized distances of adjacent individuals on each objective function.

[0078] Step 3: Selection. Binary tournament selection is adopted: 2 individuals are randomly selected each time, and the individual with a higher Pareto level (or a larger crowding distance) is selected to enter the mating pool. The size of the mating pool is equal to the population size (100).

[0079] Step 4: Crossover. Simulated binary crossover (SBX) is adopted, with a crossover probability of 0.8 and a distribution index ηc=20. For parent individuals x1 and x2, the formula for generating offspring y1 and y2 is: y1= 0.5[(x1+x2) - β(x2-x1)] y2 = 0.5[(x1+x2) + β(x2-x1)] Where β is the diffusion factor, calculated from the random number u∈[0,1]: .

[0080] Step 5: Mutation. Multinomial mutation is used, with a mutation probability of 0.1 and a distribution exponent η. m =20. For the i-th gene of individual x, the mutated gene is:

[0081] Where δ is calculated from the random number r∈[0,1]: , Wherein, the normalized distance to the lower boundary Normalized distance to the lower boundary .

[0082] Step 6: Elite Preservation. Combine the parent and offspring generations (200 individuals in total), and select the top 100 individuals based on Pareto hierarchy and crowding to form the new generation population.

[0083] Step 7: Termination Judgment. The iteration terminates when the rate of change of the Pareto front hypervolume index is less than 1% after 50 generations or 10 consecutive generations, and outputs the non-dominated solution set (usually containing 8-15 solutions).

[0084] Step 8: Scheme Selection – Based on the Analytic Hierarchy Process (AHP). A judgment matrix is ​​constructed to compare the relative importance of the three objectives. In this embodiment, the importance ratio of energy consumption, pollution, and construction period is 3:3:2. The judgment matrix is ​​shown in Table 2. Table 2

[0085] The weight vector w = [0.375, 0.375, 0.25] is calculated, and the consistency ratio CR = 0 < 0.1, passing the consistency test. The comprehensive satisfaction score S = w1·f1 + w2·f2 + w3·(1-f3) is calculated for each non-dominated solution (the lower the project schedule risk, the better, so 1-f3 is taken), and the solution with the highest score is selected as the dynamic optimization scheme.

[0086] As an alternative implementation, the Particle Swarm Optimization (PSO) algorithm can be used, which has a faster convergence speed but slightly weaker global search capability; or the Multi-Objective Evolutionary Algorithm (MOEA / D) can be used, which is suitable for scenarios where the objective function has extremely high computational cost.

[0087] The application layer displays key indicators through a visualization platform and automatically converts optimization solutions into equipment control commands through the command push module, forming a complete closed loop.

[0088] The application layer includes a visual interaction platform and an instruction push module. The visual interaction platform displays environmental indicators, energy consumption curves, and optimization suggestions in real time. The instruction push module automatically converts the dynamic optimization scheme into equipment control instructions and pushes them to the construction equipment terminal for execution, forming a closed loop of monitoring, analysis, optimization, and execution.

[0089] The instruction push module includes: The instruction generation unit parses the dynamic optimization scheme into specific equipment control parameters, including the speed adjustment value of the concrete mixer (unit: revolutions per minute, adjustment range not exceeding ±15% of the rated speed), the start and stop sequence and flow rate setting value of the spraying system (unit: liters per minute), and the scheduling route and load limit of the transport vehicles (unit: tons). The instruction verification unit performs safety boundary verification on the generated device control parameters, and the instructions can only be issued after the verification is passed. The command transmission unit uses the MQTT over TLS encryption protocol to push the verified control commands to the PLC controller of the construction equipment or the vehicle terminal. The transmission confirmation timeout is 3 seconds, and a retransmission mechanism is triggered if the timeout occurs.

[0090] The instruction push module transforms the optimization scheme into specific equipment control parameters, achieving automation from decision-making to execution; the percentage limit of the speed adjustment range ensures safe equipment operation; TLS encrypted transmission safeguards industrial control security; and the timeout retransmission mechanism improves the reliability of instruction delivery.

[0091] This invention's system is compatible with both traditional civil engineering construction scenarios and prefabricated building construction scenarios. In traditional civil engineering construction scenarios, the perception layer primarily deploys concrete pouring monitoring sensors, formwork support stress sensors, and on-site dust and noise monitoring points. In prefabricated building construction scenarios, the perception layer primarily deploys component hoisting posture sensors, grouting fullness detection devices, and component transport vehicle positioning devices. The AI ​​analysis layer automatically adapts to the current construction scenario's optimization strategy by loading different scenario knowledge bases. This system's scenario-adaptive architecture expands its applicability; scenario adaptation is achieved through knowledge base switching rather than hardware replacement, reducing deployment costs and maintenance complexity.

[0092] like Figure 2As shown, the overall process of the present invention includes seven core steps: multi-source heterogeneous data acquisition, edge preprocessing, intelligent recognition of construction behavior, standardized data fusion, high-energy-consumption and high-pollution identification, dynamic optimization decision-making, and closed-loop execution. The method generates a unified format of construction behavior records through dual-modal recognition of video and audio, fusing heterogeneous sensor data with semantic behavior records to provide standardized input for AI analysis; it achieves intelligent decision-making from identification to optimization through the cascading of random forests and genetic algorithms; and it realizes an automatic closed loop from optimization scheme to equipment execution through automatic instruction generation, security verification, and encrypted transmission.

[0093] The AI-based energy conservation and environmental protection monitoring and optimization method for the entire building construction process, implemented using the aforementioned system, includes the following steps: S1. Multi-source heterogeneous data acquisition steps: Heterogeneous sensor data is acquired by deploying a multi-dimensional sensor group at the construction site, while video images of construction behavior are acquired by monitoring camera devices, and the sound of construction machinery operation is acquired by audio acquisition devices. S2. Edge preprocessing steps: Use edge computing nodes to preprocess the heterogeneous sensing data, including outlier removal, filtering and sampling rate normalization; In one specific implementation, the preprocessing of the edge computing nodes is as follows: Outlier removal: The 3σ criterion is used in conjunction with domain knowledge. For dust data, if |x-μ|>3σ and x>500μg / m³ (exceeding the severe pollution threshold), it is determined to be a sensor malfunction rather than a true value and is marked as missing; for energy consumption data, if the power of a single device is greater than the rated power × 1.2, it is determined to be data anomaly.

[0094] Moving average filtering: Employs exponentially weighted moving average (EWMA), the formula is as follows:

[0095] Where α=0.3, it balances response speed and noise suppression. It is suitable for indicators with large fluctuations such as dust and noise.

[0096] Sampling rate normalization: The original sampling frequencies varied (1 minute for dust, 5 minutes for electricity meters, and triggering events on weighbridges), and were uniformly converted into standard time-series data with a 1-minute interval. Linear interpolation was used to fill missing data, with the following formula:

[0097] Where t1 and t2 are adjacent sampling times. For event-type data (such as weighbridges), aggregation (summation or counting) is performed within a 1-minute window.

[0098] S3. Intelligent Recognition Steps for Construction Behavior: Using the construction behavior recognition module in the standardized data engine, based on the video images and mechanical operation sounds, computer vision algorithms and sound recognition algorithms are used to automatically identify the type of construction behavior, the type of construction machinery and its operating status, and generate a structured construction behavior record. The intelligent recognition steps for construction behavior described in step S3 include: S3.1. Extract frames from the video stream at a sampling frequency of 5 frames per second, input the data into a convolutional neural network to identify the type of construction machinery, work actions, material handling trajectories, and personnel activity areas, and generate a structured video record containing timestamps, spatial coordinates, and behavior types. S3.2 Perform a short-time Fourier transform on the audio signal at a frequency of 10 times per second, extract the Mel frequency cepstral coefficient features, input the data into a support vector machine classifier to identify the type of machinery, operating load rate and abnormal working conditions, and generate a structured audio record containing timestamps, sound source location and machinery status. S3.3. Perform time alignment and spatial matching between the structured video recording and the structured audio recording. When the recognition results of the two are consistent and the confidence level is higher than 0.85, output the final structured construction behavior record.

[0099] S4. Data standardization and fusion steps: Using a unified data format generation module, the preprocessed heterogeneous sensor data and the structured construction behavior records are fused according to a preset unified spatiotemporal benchmark and data format to generate a standard construction monitoring data stream. The standard construction monitoring data stream generated by the data standardization and fusion step described in step S4 includes: The time stamp (accuracy down to the second) based on Beijing time and the spatiotemporal reference field based on the spatial coordinates of the construction plane coordinate system (accuracy down to the meter) are used uniformly. The data type identifier field uses a four-digit code (the first digit represents the major data category, and the last three digits represent the specific subtype); The numerical content field uses a fixed-length 16-element floating-point array. If the length is insufficient, it is filled with null values, and if the length is excessive, it is divided by time slices. Establish a foreign key association field with the structured construction behavior record. The database foreign key association is achieved by establishing a foreign key association with the structured construction behavior record through the associated behavior field. In one specific implementation, the database table structure is designed as follows: Structured Construction Behavior Records: behavior_id (VARCHAR(32), primary key): A unique identifier, formatted as "bhv_YYYYMMDDhhmmss_serial number". timestamp(DATETIME): The time when the action occurred. zone(VARCHAR(10)): Construction area code behavior_type(VARCHAR(50)): Behavior type, such as "Tower crane hoisting - heavy load". confidence (FLOAT): Identifies confidence level video_clip_path(VARCHAR(255)): The associated video clip storage path audio_clip_path(VARCHAR(255)): The associated audio clip storage path Standard construction monitoring data flow table (monitoring_data): data_id (VARCHAR(32), primary key): unique identifier timestamp(DATETIME): Data collection time type_code(CHAR(4)): Four-bit code values ​​(JSON): A 16-element numeric array behavior_ref(VARCHAR(32), foreign key): associated with behavior_records.behavior_id quality_flag (TINYINT): Data quality flag Foreign key relationships via behavior_ref enable efficient multi-table join queries.

[0100] S5. High energy consumption and high pollution identification steps: The high energy consumption and high pollution identification module uses the random forest algorithm to perform real-time analysis on the standard construction monitoring data stream to identify the high energy consumption and high pollution links in the current construction process. In step S5, the high-energy-consumption and high-pollution identification step, the input features of the random forest algorithm include: Current construction behavior type encoding (16-dimensional one-hot vector); Energy consumption time series characteristics over the last 15 minutes (including mean, variance, peak value, and trend slope); Current environmental parameters (dust concentration, noise level, temperature, humidity); Construction schedule deviation rate (the ratio of actual progress to planned progress, expressed as a percentage); In one specific implementation, the real-time analysis of the random forest algorithm is based on an Apache Kafka plus Flink stream processing architecture: Data stream access: Standard construction monitoring data streams are accessed through a Kafka message queue. Topics are partitioned according to data priority (priority-high, priority-normal), and the number of partitions is consistent with the number of edge computing nodes (2 in this embodiment).

[0101] Stream processing operator: Flink DataStream API implements sliding window analysis, with a window size of 15 minutes (corresponding to historical feature extraction) and a sliding step size of 30 seconds. Execution within the window function: Feature engineering: Aggregate the data within the window into a 25-dimensional feature vector (as described above); Model inference: Load a pre-trained random forest model (PMML or ONNX format), single inference time <5ms; Output result: If the probability is greater than 0.7, output a high-energy-consumption, high-pollution label to trigger downstream genetic algorithm optimization.

[0102] Latency guarantee: End-to-end latency = Kafka transmission (<100ms) + Flink window (30s sliding) + inference (5ms) + network transmission (<50ms) < 40 seconds, meeting real-time requirements (relative to minute-level changes in the construction phase).

[0103] Model updates: The model is retrained offline monthly, and the model parameters are updated hot through Flink's Broadcast State mechanism without restarting the stream processing job.

[0104] The random forest algorithm is set to have 200 decision trees, a maximum depth of 15, and a minimum number of leaf samples of 5. It outputs the probability value of whether the current construction stage belongs to a high-energy-consuming stage or a high-polluting stage. If the probability value exceeds 0.7, it is judged as positive.

[0105] Multidimensional feature engineering fully considers the dynamics of construction behavior and environmental sensitivity; specific algorithm parameter settings balance model complexity and generalization ability; probability threshold settings provide adjustable recognition sensitivity.

[0106] S6. Dynamic optimization decision-making steps: Using the dynamic optimization decision-making module based on the genetic algorithm, the identified high-energy-consuming and high-polluting links are used as inputs, combined with the construction progress constraints, to generate a dynamic optimization scheme. In one specific implementation, the construction schedule constraint is quantitatively calculated as follows: Schedule Deviation Rate = (Actual Completed Work - Planned Completed Work) / Planned Completed Work × 100% The workload is measured in working hours or output value and is obtained in real time from the project management system (such as Glodon, Mingyuan Cloud) via API.

[0107] Hard boundary constraints: When the schedule deviation rate is >+10% (too much ahead of schedule, potentially leading to quality risks) or <-15% (too much behind schedule, leading to project time risks), the weight of the task option in the optimization objective of the genetic algorithm is automatically increased (from 0.25 to 0.4), prioritizing the generation of solutions that catch up with the schedule.

[0108] The dynamic optimization decision-making step described in step S6 includes: S6.1 Construct a three-dimensional Pareto frontier search space with energy consumption reduction rate, pollution emission reduction rate and construction delay risk as optimization objectives; S6.2. Incorporate construction quality specifications, safe operating procedures, and equipment performance boundary constraints, and exclude infeasible solutions; S6.3. Using genetic algorithm parameters of population size 100, number of iterations 50, crossover probability 0.8, and mutation probability 0.1, solve for the non-dominated solution set; S6.4. Based on the analytic hierarchy process (AHP), select the solution with the highest overall satisfaction from the non-dominated solution set as the dynamic optimization solution.

[0109] S7. Closed-loop execution steps: The dynamic optimization scheme is automatically converted into equipment control commands and pushed to the construction equipment terminal for execution, forming a closed loop of monitoring, analysis, optimization and execution.

[0110] In one specific implementation, the automatic conversion of the instruction push module is achieved through a rule engine plus a protocol adapter.

[0111] The closed-loop execution steps described in step S7 include: The dynamic optimization scheme is analyzed into specific equipment control parameters, including the concrete mixer speed adjustment value (not exceeding the rated speed ±15%), the start and stop sequence and flow rate setting value of the spraying system, and the dispatching route and load limit of the transport vehicles. Perform safety boundary verification on the generated equipment control parameters; The verified control commands are pushed to the PLC controller of the construction equipment or the vehicle terminal using the MQTT over TLS encryption protocol. A 3-second transmission confirmation timeout is set, and a retransmission is triggered if the timeout occurs.

[0112] Furthermore, in other embodiments, the method of the present invention further includes a scene adaptive configuration step: Identify whether the current construction scenario is a traditional civil engineering construction scenario or a prefabricated building construction scenario; In traditional civil engineering construction scenarios, the focus is on collecting monitoring data on concrete pouring, stress data on formwork support, and data on on-site dust and noise. In the context of prefabricated building construction, the key data to be collected include component hoisting posture data, grouting fullness data, and component transportation positioning data. Load the knowledge base corresponding to the current construction scenario and adjust the optimization strategy of the AI ​​analysis layer.

[0113] This invention expands the applicability of the method through a scenario-adaptive approach, achieving scenario switching through software configuration rather than hardware replacement, thereby improving deployment flexibility and economy.

[0114] In other embodiments, the method of the present invention further includes a priority adaptive transmission step: The standard construction monitoring data stream is prioritized. When the dust concentration exceeds 150 μg / m³ or the noise exceeds 70 dB, the relevant data is marked as the highest priority, and the transmission delay is required to be less than 5 seconds. Energy consumption and building material data are marked as normal priority, with a transmission delay requirement of less than 60 seconds; When network bandwidth is insufficient, high-priority data is transmitted first, while ordinary-priority data is placed in the local cache queue and retransmitted after bandwidth is restored.

[0115] The priority-adaptive transmission mechanism ensures real-time reporting of critical environmental pollution incidents, while also guaranteeing data integrity when network resources are limited, achieving a dynamic balance between reliability and real-time performance.

[0116] Compared with conventional methods in the field, the present invention also has significant differences: Conventional Approach 1: Digital Twin Technology. Faced with the problem of multi-source heterogeneous data fusion, the mainstream approach in this field is to use digital twin technology to construct a virtual mapping of the construction site and achieve data fusion through unified modeling. This invention explicitly abandons this conventional approach for the following reasons: (1) Digital twins require high-precision 3D modeling, have long construction cycles, and are difficult to adapt to the dynamic changes of the construction site; (2) The twin model has a large data volume, which is difficult to handle at the edge, and cloud processing introduces latency; (3) The twin model's interface with the AI ​​analysis layer is complex and has high development costs. This invention uses video plus sound as a universal perception method, achieving semantic-level data unification in a software manner, reducing costs by an order of magnitude and improving real-time performance by an order of magnitude.

[0117] Conventional Approach Two: Single-Objective Optimization. In construction optimization problems, single-objective optimization (such as optimizing only cost or only construction period) is commonly used, or multiple objectives are transformed into a single objective (such as weighted summation). This invention explicitly adopts multi-objective Pareto optimization, directly addressing the three conflicting objectives of energy consumption, pollution, and construction period, generating a non-dominated solution set for decision-making, avoiding the subjectivity of weight setting, and providing richer decision options.

[0118] Common Method 3: Threshold Alarm Mechanism. The conventional design of environmental monitoring systems is to set fixed thresholds; exceeding these thresholds triggers an alarm. This invention employs a machine learning-based dynamic identification mechanism. The random forest model can learn complex patterns in historical data, identifying potential high-energy-consuming and high-polluting processes that, while not exceeding the limits, exhibit anomalies, thus achieving a shift from passive response to proactive prevention.

[0119] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An AI-based energy-saving and environmental protection monitoring and optimization system for the entire building construction process, characterized in that: include: The perception layer, deployed at the construction site, includes multi-dimensional sensor arrays, monitoring cameras, and audio acquisition devices; The multi-dimensional sensor group includes a dust sensor, a noise sensor, an energy consumption metering device, and a building material consumption monitoring device; the monitoring camera device is used to collect video images of construction activities; the audio acquisition device is used to collect the operating sounds of construction machinery. The data layer includes edge computing nodes and a standardized data engine. The edge computing nodes are communicatively connected to the perception layer and are used to preprocess heterogeneous sensor data collected by multi-dimensional sensor groups. The standardized data engine includes a construction behavior recognition module and a unified data format generation module. The construction behavior recognition module automatically identifies the type of construction behavior, the type of construction machinery, and the operating status based on video images collected by the monitoring camera and mechanical operation sounds collected by the audio acquisition device, using computer vision algorithms and sound recognition algorithms, and generates structured construction behavior records. The unified data format generation module fuses the heterogeneous sensor data and the structured construction behavior records according to a preset unified spatiotemporal benchmark and data format to generate a standard construction monitoring data stream. The AI ​​analysis layer includes a construction environmental protection correlation model, a high-energy-consumption and high-pollution identification module, and a dynamic optimization decision-making module. The construction environmental protection correlation model stores national environmental protection standard data, construction specification data, and a historical construction environmental protection correlation knowledge base. The high-energy-consumption and high-pollution identification module uses a random forest algorithm to perform real-time analysis on the standard construction monitoring data stream to identify high-energy-consumption and high-pollution stages in the current construction process. The dynamic optimization decision-making module uses a genetic algorithm, taking the high-energy-consumption and high-pollution stages as inputs and combining them with construction progress constraints, to generate a dynamic optimization scheme. The application layer includes a visual interaction platform and a command push module; the visual interaction platform displays environmental protection indicators, energy consumption curves, and optimization suggestions in real time. The instruction push module automatically converts the dynamic optimization scheme into equipment control instructions and pushes them to the construction equipment terminal for execution, forming a closed loop of monitoring, analysis, optimization and execution.

2. The system according to claim 1, characterized in that, The construction behavior recognition module includes: The video analysis submodule uses a convolutional neural network to identify construction behaviors in the video images. The identified content includes the type of construction machinery, work actions, material handling trajectories, and personnel activity areas. The video analysis submodule processes the video stream at a sampling frequency of 5 frames per second to generate a structured video record containing timestamps, spatial coordinates, and behavior types. The audio analysis submodule identifies the mechanical operating sound based on Mel frequency cepstral coefficient feature extraction and support vector machine classifier. The identification content includes the type of machinery, operating load rate, and abnormal operating conditions. The audio analysis submodule performs short-time Fourier transform analysis at a frequency of 10 times per second to generate a structured audio record containing timestamps, sound source location, and mechanical status. The data fusion submodule performs time alignment and spatial matching between the structured video recording and the structured audio recording. When the video recognition result is consistent with the audio recognition result, a structured construction behavior record with a confidence level higher than 0.85 is generated.

3. The system according to claim 2, characterized in that, The standard construction monitoring data stream generated by the unified data format generation module includes: The spatiotemporal reference field includes a timestamp that uniformly adopts Beijing time and spatial coordinates based on the construction plane coordinate system, with time accuracy at the second level and spatial accuracy at the meter level. The data type identifier field uses a four-digit code to identify the data source type. The first digit represents the data category, where 1 represents sensor data, 2 represents video recognition data, and 3 represents audio recognition data. The last three digits represent the specific subtype. The numerical content field uses a floating-point array to store the standardized monitoring values. The array length is fixed at 16 elements. If the length is insufficient, it is filled with null values. If the length is excessive, it is divided into multiple records using time slices. The associated behavior field establishes a foreign key association with the structured construction behavior record, which is used to trace the construction context in which the data was generated.

4. The system according to claim 1, characterized in that, The edge computing nodes include: The sensor data preprocessing unit performs outlier removal, moving average filtering, and sampling rate normalization on the heterogeneous sensor data, and converts the different original sampling frequencies into standard time-series data with a 1-minute interval. The cache queue management unit uses a circular buffer structure to cache the preprocessed data of the most recent 72 hours. When the network is interrupted, it is stored locally and re-transmitted in batches after the network is restored. The priority scheduling unit assigns priority to the standard construction monitoring data stream, with dust and noise data assigned the highest priority and a transmission delay of less than 5 seconds; energy consumption and building material data assigned the normal priority and a transmission delay of less than 60 seconds.

5. The system according to claim 1, characterized in that, The dynamic optimization decision module includes: A multi-objective optimization objective function construction unit is used to construct a three-dimensional Pareto frontier search space with energy consumption reduction rate, pollution emission reduction rate and construction period delay risk as optimization objectives. The constraint processing unit incorporates construction quality specification constraints, safe operating procedure constraints, and equipment performance boundary constraints, excluding infeasible solutions from the search space. The genetic algorithm solver is set with a population size of 100, an iteration number of 50, a crossover probability of 0.8, and a mutation probability of 0.1, and obtains a non-dominated solution set. The scheme selection unit selects the scheme with the highest overall satisfaction from the non-dominated solution set based on the analytic hierarchy process (AHP) as the dynamic optimization scheme.

6. An AI-based method for energy conservation and environmental protection monitoring and optimization throughout the entire building construction process, characterized in that: The system described in claim 1 is implemented by comprising the following steps: S1. Multi-source heterogeneous data acquisition steps: Heterogeneous sensor data is acquired by deploying a multi-dimensional sensor group at the construction site, while video images of construction behavior are acquired by monitoring camera devices, and the sound of construction machinery operation is acquired by audio acquisition devices. S2. Edge preprocessing steps: Use edge computing nodes to preprocess the heterogeneous sensing data, including outlier removal, filtering and sampling rate normalization; S3. Intelligent Recognition Steps for Construction Behavior: Using the construction behavior recognition module in the standardized data engine, based on the video images and mechanical operation sounds, computer vision algorithms and sound recognition algorithms are used to automatically identify the type of construction behavior, the type of construction machinery and its operating status, and generate a structured construction behavior record. S4. Data standardization and fusion steps: Using a unified data format generation module, the preprocessed heterogeneous sensor data and the structured construction behavior records are fused according to a preset unified spatiotemporal benchmark and data format to generate a standard construction monitoring data stream. S5. High energy consumption and high pollution identification steps: The high energy consumption and high pollution identification module uses the random forest algorithm to perform real-time analysis on the standard construction monitoring data stream to identify the high energy consumption and high pollution links in the current construction process. S6. Dynamic optimization decision-making steps: Using the dynamic optimization decision-making module based on the genetic algorithm, the identified high-energy-consuming and high-polluting links are used as inputs, combined with the construction progress constraints, to generate a dynamic optimization scheme. S7. Closed-loop execution steps: The dynamic optimization scheme is automatically converted into equipment control commands and pushed to the construction equipment terminal for execution, forming a closed loop of monitoring, analysis, optimization and execution.

7. The method according to claim 6, characterized in that, The intelligent recognition step for construction behavior described in step S3 includes: S3.

1. Extract frames from the video stream at a sampling frequency of 5 frames per second, input the data into a convolutional neural network to identify the type of construction machinery, work actions, material handling trajectories, and personnel activity areas, and generate a structured video record containing timestamps, spatial coordinates, and behavior types. S3.2 Perform a short-time Fourier transform on the audio signal at a frequency of 10 times per second, extract the Mel frequency cepstral coefficient features, input the data into a support vector machine classifier to identify the type of machinery, operating load rate and abnormal working conditions, and generate a structured audio record containing timestamps, sound source location and machinery status. S3.

3. Perform time alignment and spatial matching between the structured video recording and the structured audio recording. When the recognition results of the two are consistent and the confidence level is higher than 0.85, output the final structured construction behavior record.

8. The method according to claim 6, characterized in that, The standard construction monitoring data stream generated by the data standardization and fusion step described in step S4 includes: The timestamps are uniformly adopted using Beijing time and the spatiotemporal reference field is based on the spatial coordinates of the construction plane coordinate system; The data type identifier field uses a four-digit encoding; The numerical content field uses a fixed-length 16-element floating-point array. If the length is insufficient, it is filled with null values, and if the length is excessive, it is divided by time slices. Establish a related behavior field that is associated with the foreign key of the structured construction behavior record.

9. The method according to claim 6, characterized in that, The dynamic optimization decision-making steps described in step S6 include: S6.1 Construct a three-dimensional Pareto frontier search space with energy consumption reduction rate, pollution emission reduction rate and construction delay risk as optimization objectives; S6.

2. Incorporate construction quality specifications, safe operating procedures, and equipment performance boundary constraints, and exclude infeasible solutions; S6.

3. Using genetic algorithm parameters of population size 100, number of iterations 50, crossover probability 0.8, and mutation probability 0.1, solve for the non-dominated solution set; S6.

4. Based on the analytic hierarchy process (AHP), select the solution with the highest overall satisfaction from the non-dominated solution set as the dynamic optimization solution.

10. The method according to claim 6, characterized in that, The closed-loop execution steps described in step S7 include: The dynamic optimization scheme is analyzed into specific equipment control parameters, including concrete mixer speed adjustment value, spraying system start-stop sequence and flow rate setting value, and transport vehicle scheduling route and load limit; Perform safety boundary verification on the generated equipment control parameters; The verified control commands are pushed to the PLC controller of the construction equipment or the vehicle terminal using the MQTT over TLS encryption protocol. A 3-second transmission confirmation timeout is set, and a retransmission is triggered if the timeout occurs.

Images