Intelligent real-time regulation and control method and system for construction site environment quality

Abstract

The invention discloses a construction site environment quality intelligent real-time regulation and control method and system, and belongs to the technical field of environment intelligent monitoring and construction management. Particulate matter concentration, gas pollutant concentration, noise level and meteorological parameters of a construction area are collected; obtaining an operation type label and a microenvironment disturbance index, and constructing a multi-dimensional input feature set containing environment data, operation information and disturbance factors; inputting the characteristics into a pollution trend analysis model, and outputting a pollutant concentration trend and pollution source type judgment result; pollution source distinguishing accuracy is evaluated based on an identification result, and a model structure or a training weight is dynamically optimized; differential regulation and control measures are implemented according to the types of pollution sources, wherein the measures comprise linkage of a spraying system and an air draft system, vehicle restriction and operation adjustment suggestions; according to the method, accurate recognition and intelligent response to the construction environment are achieved, and the real-time performance and scientificity of pollution treatment on the construction site are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental intelligent monitoring and construction management, and particularly relates to a construction site environmental quality intelligent real-time regulation method and system. BACKGROUND

[0002] In a complex construction environment, multiple sources of pollution (such as dust, welding fumes, cutting particles, and exhaust emissions) often exist simultaneously. Existing trend analysis methods are usually based on single sensor data modeling, lacking the ability to identify and separate pollution sources, resulting in an inability to accurately determine the dominant pollution source when the concentration of pollutants rises. For example, when the PM2.5 concentration rises, the system may mistakenly judge that it is caused by dust and automatically start the sprinkler system, but in fact the pollution mainly comes from steel structure welding operations. Spraying not only has no effect, but also causes the construction equipment to be damp, posing a safety hazard. Therefore, the lack of pollution source identification causes the control measures to be "symptomatic but not causal", seriously affecting the accuracy of environmental regulation and construction efficiency.

[0003] Especially in the context of highly dynamic changes in power grid operation, how to achieve multi-dimensional feature extraction, multi-scenario fusion analysis, and accurate risk perception based on big data has become an important technical challenge for the safe and stable operation of the power grid. Therefore, it is urgent to build a risk perception method for power grid operation driven by multi-dimensional heterogeneous data, which can extract key feature information from massive data and realize the transition from local anomaly detection to global risk evolution trend prediction, thereby improving the autonomous prediction and fine response capabilities of power grid operation. SUMMARY

[0004] The purpose of the present application is to provide a construction site environmental quality intelligent real-time regulation method and system to solve the problems in the background art.

[0005] To achieve the above purpose, the present application provides the following technical solution: a construction site environmental quality intelligent real-time regulation method, comprising:

[0006] Collecting real-time environmental data of the construction site, including particulate matter concentration, gas pollutant concentration, noise level, and meteorological parameters;

[0007] Obtaining an operation type label related to the construction activity, which is used to identify the current construction phase or specific operation type; the operation type label includes earthwork, steel structure welding, cutting operation, and mechanical operation;

[0008] Calculating a micro-environmental disturbance index related to construction disturbance, which is comprehensively evaluated according to local wind speed change, construction equipment operation frequency, and vehicle entry and exit frequency, reflecting the pollution diffusion and superposition trend in the construction area;

[0009] input the environmental data, the job type label, and the micro-environment disturbance index into the improved pollution trend analysis model, construct a time series model integrating multi-dimensional input features, and output a pollution concentration trend prediction result and a pollution source type discrimination result;

[0010] Based on the pollution source type discrimination result, the current pollution source discrimination accuracy is evaluated, and the pollution trend analysis model structure is adjusted or the training weight is optimized according to the evaluation result;

[0011] According to the pollution source identification result, a control measure is implemented, which includes starting a spraying system, starting a local exhaust system, limiting vehicle operation, or adjusting a job.

[0012] Preferably, the job type label related to the construction activity is obtained, including:

[0013] Based on the construction task plan data, the current job content is extracted by time period and mapped to a preset job type label set;

[0014] The system collects image streams of the construction area, uses a convolutional neural network model to recognize job behavior from the images, and cross- validates the recognition result with the plan data;

[0015] Based on the construction equipment operation state data, including equipment ID, operation time length, and job frequency indicators, through a set job equipment and label mapping relationship, a device-driven job label is automatically generated;

[0016] The job type label results generated by the plan data, video recognition results, and equipment operation data are comprehensively judged, and a weighted voting mechanism is used to output the final label.

[0017] Preferably, the calculation of the micro-environment disturbance index includes:

[0018] Real-time wind speed data of multiple monitoring points in the construction area are obtained, the standard deviation of wind speed in a unit time is calculated based on a time sliding window, and the standard deviation is taken as a wind speed disturbance factor to reflect the intensity of local airflow fluctuation;

[0019] The operation state data of the construction equipment are collected, including the number of equipment start-stop times, continuous operation time length, and operation intensity indicators, and through a correspondence table of equipment types and disturbance factors, the behaviors of various types of equipment are mapped to disturbance scores;

[0020] The frequency of vehicle entry and exit of the construction area entrance and exit in a set time period is counted, and the disturbance weight is set in combination with the vehicle type to calculate the vehicle disturbance factor;

[0021] The wind speed disturbance factor, the disturbance score, and the vehicle disturbance factor are normalized and superimposed according to a weighted formula to form the micro-environment disturbance index.

[0022] Preferably, the environmental data, job type label and micro-environment disturbance index are input into an improved pollution trend analysis model to construct a time series model integrating multi-dimensional input features, including:

[0023] The environmental data is time series reconstructed based on a set time step, including particulate matter concentration, gas pollutant concentration, noise level and meteorological parameters. A sliding window method is used to construct a historical data sequence as an input feature of the model.

[0024] The job type label is vectorized and converted into a fixed-dimensional sparse vector using a one-hot encoding method, and is aligned with the environmental data of the corresponding time period to realize synchronous mapping of the label and time series data.

[0025] The micro-environment disturbance index is normalized to convert it into a continuous variable with the same scale and embedded into the time series feature matrix to form an input feature set.

[0026] Based on the constructed multi-dimensional input features, a long short-term memory neural network is used as the pollution trend analysis model structure to predict the pollutant concentration trend in a preset future time range, and a multi-classification recognition node of pollution source type is set in the output layer to realize joint output of pollution trend prediction and pollution source type identification.

[0027] Preferably, based on the pollution source type identification result, the current pollution source discrimination accuracy is evaluated, including:

[0028] The pollution source type identification result is compared with the artificially labeled real job type data, and the accuracy, recall rate and F1 value indicators are used to measure the recognition performance of the pollution trend analysis model in different pollution source classification tasks.

[0029] Based on a preset performance threshold, it is judged whether the accuracy of the current pollution trend analysis model in identifying a certain type of pollution source is lower than the expected standard. If the F1 value of any category is lower than the set threshold, the category is marked as a weak identification category.

[0030] Preferably, the pollution trend analysis model structure or training weight is adjusted according to the evaluation result, including increasing the weight of the sample of the weak identification category in the training sample or increasing the sample quantity of the weak identification category through oversampling technology.

[0031] According to the overall performance change of the pollution trend analysis model, the network structure parameters of the pollution trend analysis model are dynamically adjusted, including increasing the number of hidden layers, adjusting the learning rate or replacing the activation function.

[0032] Preferably, the pollution source identification result is used to implement control measures, including:

[0033] Based on the pollution source type identification results output by the pollution trend analysis model, the category of the main pollution cause is identified, including earthwork dust, welding fumes, exhaust emissions, or particulate matter accumulation.

[0034] Match the pollution source type with the control rules in the preset control strategy library. The control rules include response threshold, control method, duration and equipment priority to form an executable control instruction set.

[0035] The control commands are sent to the corresponding control equipment, including automatically activating the area sprinkler system for pollution sources identified as earthwork dust, and controlling the spray angle and frequency according to the wind speed and direction; activating the local exhaust or high-level ventilation system for pollution sources identified as welding fumes, prioritizing the closed or semi-closed work areas; and dispatching low-emission vehicles to replace pollution sources identified as exhaust gas.

[0036] This invention also provides an intelligent real-time control system for construction site environmental quality, comprising:

[0037] Environmental sensing and acquisition module: Collects real-time environmental data of the construction site, including particulate matter concentration, gaseous pollutant concentration, noise level and meteorological parameters;

[0038] Construction Operation Information Acquisition Module: Acquires operation type tags related to construction activities to identify the current construction stage or specific operation type; the operation type tags include earthwork operations, steel structure welding, cutting operations, and machinery operation.

[0039] Micro-disturbance sensing and calculation module: calculates the micro-environmental disturbance index related to construction disturbance. The micro-environmental disturbance index is comprehensively evaluated based on local wind speed changes, construction equipment operating frequency and vehicle entry and exit frequency to reflect the trend of pollutant diffusion and accumulation in the construction area.

[0040] The discrimination module inputs the environmental data, operation type labels and microenvironmental disturbance index into the improved pollution trend analysis model, constructs a time series model that integrates multi-dimensional input features, and outputs pollutant concentration trend prediction results and pollution source type discrimination results.

[0041] Optimization module: Based on the pollution source type discrimination results, evaluate the accuracy of current pollution source differentiation, and adjust the pollution trend analysis model structure or optimize training weights according to the evaluation results;

[0042] Control Response Module: Implements control measures based on pollution source identification results. These control measures include turning on the sprinkler system, starting the local ventilation system, and restricting or adjusting vehicle operations.

[0043] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0044] 1. This invention integrates multi-source environmental data from construction sites, operation type labels, and micro-environmental disturbance indices to construct a time-series analysis model for pollution trend prediction and pollution source type identification, significantly improving the accuracy of multi-source complex pollution identification and trend response capability. Compared with traditional analysis methods that rely solely on sensor data, this invention can dynamically capture the causal characteristics of pollution changes, realizing a shift from "data monitoring" to "intelligent judgment."

[0045] 2. Based on the pollution source type identification results, this invention designs a differentiated control and response mechanism that can link multiple control devices such as sprinklers, ventilation systems, and traffic restrictions, or output operational adjustment suggestions, thus constructing a closed-loop system of "sensing-identification-feedback-control". This solution has the technical advantages of strong predictability, accurate response, and high adaptability, significantly improving the level of intelligent management and control of the construction environment and meeting the actual needs of green construction and smart construction sites. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0047] Figure 1 This is a flowchart of the method of the present invention.

[0048] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1, please refer to Figure 1 As shown in this embodiment, a method for intelligent real-time control of construction site environmental quality includes:

[0051] Collect real-time environmental data from construction sites, including particulate matter concentration, gaseous pollutant concentration, noise level, and meteorological parameters;

[0052] Obtain operation type tags related to construction activities to identify the current construction stage or specific operation type; the operation type tags include earthwork, steel structure welding, cutting, and machinery operation.

[0053] The microenvironmental disturbance index related to construction disturbance is calculated. The microenvironmental disturbance index is comprehensively evaluated based on local wind speed changes, construction equipment operating frequency and vehicle entry and exit frequency to reflect the trend of pollutant diffusion and accumulation in the construction area.

[0054] The environmental data, operation type labels, and microenvironmental disturbance index are input into the improved pollution trend analysis model to construct a time series model that integrates multi-dimensional input features, and output pollutant concentration trend prediction results and pollution source type identification results.

[0055] Based on the pollution source type identification results, assess the accuracy of current pollution source differentiation, and adjust the pollution trend analysis model structure or optimize training weights according to the assessment results;

[0056] Control measures are implemented based on the pollution source identification results. These control measures include turning on the sprinkler system, starting the local ventilation system, and restricting or adjusting vehicle operations.

[0057] To achieve intelligent real-time control of environmental quality at construction sites, this invention first deploys an environmental monitoring subsystem to continuously and dynamically collect data on key environmental elements in the construction area. This subsystem is the foundational module of the entire intelligent control system, and its accuracy and response speed directly affect the accuracy of subsequent pollution source identification and control decisions.

[0058] The environmental data acquisition module in this invention includes multiple distributed sensing nodes, each integrating a particulate matter monitoring sensor, a gas detection sensor, a noise sensor, and a meteorological sensor. The specific data types acquired include, but are not limited to:

[0059] Particulate matter concentration data, including PM2.5 and PM10, is measured using laser scattering or beta-ray methods, with a measurement accuracy of no less than ±10 micrograms per cubic meter and a sampling period of once every 30 seconds. The spatial distribution changes of particulate matter are continuously recorded using a sensor array installed at representative locations within the construction area, forming time-series data.

[0060] Gaseous pollutant concentration data, including carbon monoxide (CO), nitrogen dioxide (NO2), and volatile organic compounds (VOCs), are collected using electrochemical or infrared spectroscopy detection modules. Each monitoring node supports real-time sensing and data caching mechanisms to ensure that critical data is retained even during network outages or signal fluctuations.

[0061] Noise level data: Real-time noise intensity sampling was performed using an A-weighted sound level meter conforming to IEC 61672, with the unit being decibels (dB), and a sampling frequency of once every 10 seconds. Each sensor must support dynamic threshold switching between nighttime and daytime to ensure compliance with environmental standards for comparison at different time periods.

[0062] Meteorological parameters include wind speed, wind direction, temperature, humidity, and atmospheric pressure. Data is collected synchronously using an ultrasonic anemometer and a digital temperature and humidity sensor. The wind speed measurement range is 0 to 30 meters per second, with a measurement accuracy of ±0.1 meters per second.

[0063] To ensure unified management and analysis of the aforementioned diverse environmental data, this invention employs an edge computing gateway to preprocess various types of raw data, including but not limited to unit normalization, noise filtering, missing data imputation, and outlier removal. Outlier detection utilizes a sliding window method based on the three-standard-deviation rule. This involves calculating the mean and standard deviation within any consecutive time window; when a data point exceeds the mean ± three standard deviations, it is marked as an outlier and removed.

[0064] After data cleaning, all data was converted to timestamp format and a time series data structure was constructed as input for subsequent pollution trend analysis models. The time series data was padded and aligned using equal-interval resampling, with a standard time step set at 1 minute.

[0065] In the intelligent real-time control method for construction site environmental quality proposed in this invention, to achieve accurate identification and classification control of pollutant sources, it is necessary to introduce semantic features of construction behavior based on environmental data analysis to assist in the construction of pollution source discrimination models. Among them, the operation type label, as a high-order parameter reflecting the semantics of construction activities, is one of the important features used in this invention for pollution source association modeling and trend judgment.

[0066] The so-called work type label refers to a discrete label that classifies and identifies the construction work being carried out at a construction site within a specific time period, and is used to represent different construction stages or types of work. The label set can be preset according to the type of construction project, and typically includes, but is not limited to, the following categories: earthwork, steel structure welding, rebar cutting, concrete pouring, large machinery operation, material stacking and transportation, etc.

[0067] Retrieve construction task scheduling information from the construction site's progress management system or BIM platform to obtain the task plan corresponding to the current time period. Each construction task includes fields such as task number, start time, end time, construction area, and participating trades.

[0068] By constructing a mapping table between task types and tag names—for example, mapping "foundation pit excavation" to "earthwork" and "steel structure hoisting" to "steel structure welding"—the tag conversion of planned tasks can be achieved. This conversion process can be performed using string matching combined with keyword classification algorithms to achieve a preliminary inference between task and tag.

[0069] The job label output in this step is a "plan label," which has prior characteristics but has certain delays and errors and requires subsequent verification.

[0070] Image sources covering key construction areas were selected from video surveillance points deployed at the construction site, and the acquired video images were analyzed and processed in real time. A trained convolutional neural network (CNN) model was used to perform object detection and behavior recognition on the images.

[0071] This CNN model is built on a ResNet50 backbone network. The input is a time slice consisting of three consecutive frames, and the output is a probability distribution vector of the job type. Training samples are sourced from construction site video datasets, with labels manually reviewed and annotated. The training set contains at least 10,000 samples. The loss function is cross-entropy loss, and the optimization algorithm uses the Adam optimizer with an initial learning rate of 0.0001.

[0072] In the model output, if the confidence level of a certain label exceeds 0.6 (i.e., the recognition probability exceeds 60%), then the label is determined to be the current valid job type and is called "image label".

[0073] To prevent misjudgment, the model also has a built-in region constraint mechanism: that is, based on the correspondence between the work location in the image and the camera installation area, labels with unreasonable spatial locations are filtered out.

[0074] Data acquisition modules are deployed on key equipment at the construction site to record the equipment's operating status, runtime, start / stop frequency, and equipment ID. A mapping table is established between equipment type and operation tag. For example, if a concrete pump truck runs continuously for more than 10 minutes and is located in the pouring area, it can be inferred that it is a "concrete pouring" operation; if an electric welding machine is in a high-power continuous power-on state for more than 5 minutes, it can be identified as "steel structure welding".

[0075] By setting feature thresholds and constructing judgment rules, "equipment labels" with confidence values ​​can be automatically generated. For example, if the equipment type is welding equipment, the power-on time exceeds 5 minutes, and the operating current exceeds the set value of 50 amps, then the label "steel structure welding" is generated with a confidence value of 0.8.

[0076] The three source tags (plan tag, image tag, and equipment tag) are input into the fusion judgment module, and a weighted voting mechanism is used to determine the final job type tag. The default weights are set as follows: plan tag weight is 0.3, image tag weight is 0.5, and equipment tag weight is 0.2.

[0077] Within each decision cycle, the label outputs at the same point in time are summarized and scored according to their respective weights. For example, if "steel structure welding" scores 0.9 in the image label, 0.6 in the planning label, and 0.4 in the equipment label within the same time period, then the overall score is: Overall score = (0.5 × 0.9) + (0.3 × 0.6) + (0.2 × 0.4) = 0.72;

[0078] If the overall score exceeds the set threshold of 0.6, the tag is confirmed as the current operation type tag and participates in pollution source modeling and trend analysis. This tag is refreshed every 5 minutes and can be automatically updated based on actual changes to ensure its timeliness and stability.

[0079] To achieve dynamic modeling and real-time control of the diffusion and accumulation process of pollutants at construction sites, this invention introduces the Microenvironmental Disturbance Index (MEDI) on the basis of traditional environmental parameters. This index is used to quantify the degree of influence of local disturbance behavior at construction sites on the migration, diffusion and superposition trends of pollutants. It is a dynamic indicator that comprehensively reflects the intensity of intervention in construction behavior.

[0080] The calculation of the microenvironmental disturbance index comprehensively considers multiple disturbance source factors such as local wind speed fluctuations, equipment operating intensity, and vehicle traffic frequency, exhibiting good real-time performance, sensitivity, and interpretability. As an auxiliary input feature for environmental trend analysis models and pollution source identification models, this index helps enhance the system's ability to trace pollution events and its adaptability in response.

[0081] Multiple wind speed monitoring points were set up within the construction area (preferably one point per 200 square meters), and real-time wind speed values ​​were collected at each monitoring point every 10 seconds. The disturbance intensity was calculated using the time sliding window method for the wind speed data at each monitoring point.

[0082] Suppose a monitoring point collected 30 wind speed data points in the last 5 minutes, denoted as . Wind speed disturbance factor The calculation method is as follows:

[0083] First, calculate the average wind speed within the window;

[0084] Secondly, calculate the standard deviation of the wind speed data within this time window, denoted as σ (sigma).

[0085] The standard deviation σ represents the degree of wind speed disturbance at the monitoring point during the current time period, and the unit is meters per second.

[0086] To enhance the stability of the indicators, the standard deviation of wind speed disturbance at all monitoring points was averaged to obtain the overall wind speed disturbance factor W for the region.

[0087] The larger the value of the wind speed disturbance factor W, the more intense the local air flow fluctuations in the area, and the more unstable the pollutant diffusion path.

[0088] Large machinery and equipment at construction sites (such as excavators, loaders, welding equipment, and concrete pump trucks) are a significant source of localized disturbances. By deploying status acquisition modules on key equipment, the start / stop status, runtime, operating frequency, and equipment type number of the equipment can be recorded in real time.

[0089] This invention establishes a mapping table between equipment types and disturbance factors, as shown in the following example:

[0090] Table 1 Mapping Relationship Table

[0091] Device type Start score Continuous operation score Intensity factor Excavator 1.5 2.0 1.0 Welding device 1.0 1.5 0.8 Concrete pump truck 2.0 2.5 1.2 Cutting device 1.2 1.8 0.9

[0092] As shown in Table 1, statistics are compiled for each device within a set time window (e.g., 5 minutes):

[0093] Startup count ;

[0094] Continuous runtime T (in minutes);

[0095] Average operating load or power index P (if the equipment has an energy consumption acquisition module);

[0096] The equipment disturbance factor E is calculated as follows: E = ( (×starting score) + (T × continuous running score × intensity factor);

[0097] The E values ​​of multiple devices are summed and normalized to obtain the total disturbance factor E_total of the equipment in the construction area during that time period.

[0098] The frequent passage of construction vehicles (especially heavy transport vehicles such as dump trucks and concrete mixer trucks) is a significant source of local dust disturbances and pollutant movement.

[0099] This invention installs a vehicle access identification module at the construction site entrance and exit, recording the entry and exit records of each construction vehicle through video recognition or RFID, and extracting the following data:

[0100] Total number of vehicles passing through within a set time period ;

[0101] Number of different types of vehicles (e.g., N_a is the number of dump trucks, N_b is the number of ordinary transport vehicles);

[0102] Disturbance weight for each type of vehicle (e.g., 2.0 for dump trucks and 1.0 for ordinary transport vehicles).

[0103] The formula for calculating the vehicle disturbance factor V is: V = ∑ (number of vehicles × corresponding disturbance weight);

[0104] For example, if 6 dump trucks and 4 regular transport trucks pass through within 5 minutes, then: V = (6 × 2.0) + (4 × 1.0) = 16.0; Finally, the V value is normalized by the area of ​​the region to avoid data imbalance between large and small construction sites.

[0105] The three disturbance factors mentioned above—wind speed disturbance factor W, equipment disturbance factor E_total, and vehicle disturbance factor V—are input into the disturbance index fusion model. The fusion method uses a weighted superposition method, as shown in the formula: MEDI = α × W + β × E_total + γ × V; where α, β, and γ are the weighting coefficients of the disturbance factors. Based on the influence of each factor on pollutant concentration changes in historical data, coefficients are trained using a linear regression model. The initial recommended values ​​are: wind speed disturbance factor weight α = 0.3; equipment disturbance factor weight β = 0.4; vehicle disturbance factor weight γ = 0.3.

[0106] The MEDI value is calculated every 5 minutes and normalized to fall between 0 and 1. The higher the value, the greater the disturbance intensity in the current construction area, the more complex the pollutant diffusion path, and the more significant the impact on environmental quality.

[0107] To enhance the ability of intelligent environmental quality control systems at construction sites to predict changes in pollutant concentrations and identify pollution sources, this invention proposes a multi-dimensional feature input method that integrates environmental data, work type labels, and micro-environmental disturbance indices, building upon traditional pollution trend modeling. This constructs an improved pollution trend analysis model suitable for multi-source pollution scenarios at construction sites.

[0108] First, time-series data collected by various environmental sensors deployed at the construction site are processed. This environmental data includes, but is not limited to:

[0109] Particulate matter concentration (PM2.5, PM10);

[0110] Concentration of gaseous pollutants (such as CO, NO2, VOC);

[0111] Noise level (in decibels);

[0112] Meteorological parameters (wind speed, wind direction, temperature, humidity, atmospheric pressure, etc.).

[0113] The time step is set to 1 minute, and the past 30 minutes are selected as a sliding window to construct a historical sequence. That is, each input sample contains multidimensional environmental data at 30 time points, forming a three-dimensional tensor structure with the shape [30, N], where N is the environmental feature dimension at each time point.

[0114] The sliding window mechanism slides once per minute to generate new training samples, enabling continuous modeling of contamination trends.

[0115] To ensure data quality, the following preprocessing is performed on the raw data within each sliding window:

[0116] Null value filling: Use the previous value filling method;

[0117] Outlier removal: Use the 3σ rule to filter data points that exceed 3 times the standard deviation;

[0118] Unit normalization: Scaling all numerical features to the range of 0 to 1, using the maximum and minimum value normalization method.

[0119] The job type label is used to describe the main types of jobs carried out at the construction site in different time periods, including earthwork, steel structure welding, cutting, concrete pouring, and machinery operation.

[0120] To use them for model training, they need to be vectorized. This invention uses one-hot encoding to convert classification labels into sparse vectors. For example, if there are 5 job types, each label will be converted into a vector of length 5, where only the position corresponding to the job type is 1, and the rest are 0.

[0121] The label vectors are aligned with the environmental data using minute-level timestamps. The label vector for each time point is appended to the end of the corresponding environmental feature vector, forming an expanded time-series input matrix. The final input feature dimension for each time point is updated to N+5.

[0122] To ensure a sensitive response to label changes, the system sets a minimum label change cycle of 5 minutes. This means that if a job type label changes, it must remain continuously for at least 5 minutes before being adopted by the model, effectively filtering out short-term misidentifications.

[0123] The Microenvironmental Disturbance Index (MEDI), a dynamic indicator reflecting the intensity of disturbance at a construction site, is calculated based on the following factors:

[0124] Wind speed fluctuation intensity;

[0125] Frequency of start-up and shutdown of construction equipment and intensity of operation;

[0126] Disturbance weights for the frequency and type of construction vehicles entering and exiting.

[0127] The index is a continuous variable between 0 and 1. The higher the value, the stronger the local disturbance and the more obvious the pollution diffusion trend.

[0128] When constructing the time series input, the MEDI value is aligned with the environmental data and job labels at a frequency of once per minute, and embedded as a continuous variable into the feature matrix. The final input dimension of each time point is expanded to N+5+1, which is the sum of the original environmental features, job label vector and disturbance index.

[0129] This invention employs a Long Short-Term Memory (LSTM) neural network as the core structure of a pollution trend prediction model. LSTM is a recurrent neural network with memory capabilities, suitable for processing time series data, and can effectively model the time dependence and nonlinear trends of pollutant concentrations.

[0130] The specific model structure is as follows:

[0131] Input layer: Receives time series tensors of shape [30, N+6];

[0132] LSTM layer: Set the number of hidden units to 128, and use bidirectional LSTM to enhance context modeling capabilities;

[0133] Dropout layer: Set the dropout rate to 0.3 to prevent overfitting;

[0134] Fully connected layer 1: Connected to the pollutant concentration trend prediction output, outputting a single or multi-value sequence (such as predicting the PM2.5 concentration for the next 10 minutes).

[0135] Fully connected layer 2: Connected to the pollution source type discrimination output, using the Softmax function for multi-classification, outputting the probability distribution of pollution sources (such as dust, welding fumes, exhaust gas, etc.).

[0136] To achieve multi-task joint training for pollution trend prediction and pollution source identification, this invention designs a joint loss function L_total, which consists of two parts:

[0137] The pollutant concentration prediction error is calculated using the mean squared error (MSE) as the loss function.

[0138] The error in identifying pollution source types is assessed using the cross-entropy loss function.

[0139] The total loss function is expressed as: ; where λ is the weight coefficient, initially set to 0.7, and can be dynamically adjusted during training based on model performance.

[0140] The training dataset was constructed using a sliding window, with each sample containing a 30-minute input sequence and a predicted target for the next 10 minutes. The training epochs were 100, the batch size was 64, the optimizer was Adam, the learning rate was set to 0.001, and an early stopping mechanism was enabled to control overfitting.

[0141] This invention allows for performance evaluation of pollution source type identification results during both model training and operation. To ensure the accuracy and traceability of the evaluation results, the model identification results must be compared item by item with manually labeled real operation type data. The manually labeled data can originate from on-site video review records, operation logs exported from the construction progress management platform, or supervision records provided by third-party supervision agencies.

[0142] To comprehensively evaluate the performance of pollution source type identification, this invention adopts the following three classic classification performance indicators:

[0143] Accuracy: refers to the proportion of pollution sources correctly identified by the model as a certain type of pollution source.

[0144] Recall: The proportion of samples that are actually sources of pollution that are correctly identified by the model.

[0145] F1 score (F1-score): is the harmonic mean of precision and recall. The formula is: F1 = 2 × (precision × recall) ÷ (precision + recall). The above indicators are calculated for each type of pollution source to evaluate the model's ability to distinguish different pollution source scenarios such as earthwork dust, welding fumes, and exhaust emissions.

[0146] After the assessment is completed, the calculated F1 value for each pollution source category is compared with a set performance threshold. In this invention, an F1 value of 0.7 (i.e., 70%) is used as the minimum acceptable standard by default. If the F1 value of any category is lower than this threshold, the category is considered to be a weak identification category.

[0147] The system automatically marks this category as an optimization target and records the following information:

[0148] Error types (missed detection, incorrect detection);

[0149] Time period and working condition context (e.g., high disturbance index or frequent job label switching);

[0150] Current model parameter configuration (such as the number of hidden layers, learning rate, etc.).

[0151] These labeling results will be fed into the model optimization module as input to drive subsequent adjustments to the architecture or training strategy.

[0152] To address the identification of weak categories, this invention provides two training data augmentation strategies to improve their participation in the model and learning efficiency:

[0153] Sample weighting mechanism: During model training, the system automatically increases the weight of weakly classified samples in the loss function. Specifically, the standard cross-entropy loss function is modified to a class-weighted version, where the loss of weakly classified samples is multiplied by a weight factor greater than 1 (such as 1.5 or 2.0), thus making them more influential on the overall gradient.

[0154] Sample oversampling mechanism: When the number of weak class samples in the original data is insufficient, the system uses oversampling (such as the SMOTE algorithm) to generate approximate samples, thereby balancing the class distribution of the training set.

[0155] The two methods described above can be used in combination depending on the specific scenario. The system can dynamically determine the enhancement method and intensity through an automatic learning mechanism.

[0156] In addition to data-level optimizations, if the overall model performance still falls short of expectations, the system will dynamically adjust the model structure itself. The pollution trend analysis model in this invention is built upon a Long Short-Term Memory (LSTM) neural network and supports the following dynamic parameter adjustment strategies:

[0157] Increase the number of hidden layers or units: If the model is not good at time-dependent modeling, the number of LSTM layers or the number of neurons per layer can be increased, for example, from a single-layer LSTM to a double-layer stacked LSTM, or the number of units per layer can be increased from 64 to 128.

[0158] Adjusting the learning rate: If the model experiences oscillations or slow convergence during training, the learning rate can be adjusted from the initial 0.001 to a more suitable value, such as 0.0005, to improve stability and training efficiency.

[0159] Replace the activation function: Replace the nonlinear transformation activation function of the fully connected layer, for example, from ReLU to LeakyReLU, to alleviate gradient vanishing or activation saturation problems.

[0160] The system can automatically trigger the above-mentioned structural optimization strategy based on changes in model training logs and validation set metrics. The optimized model will be redeployed in the next training cycle, and the latest performance metrics will be used as feedback input into the next evaluation process, forming a continuous performance evolution loop.

[0161] The pollution trend analysis model outputs not only future pollutant concentration trends but also a classification of pollution source types. This classification is generated by the Softmax output layer in the multi-task learning architecture, representing the probability distribution of each pollution source type within the current time period. For example, the model outputs the following probability vector:

[0162] Dust from earthwork: 0.65;

[0163] Welding fumes: 0.20;

[0164] Exhaust emissions: 0.10;

[0165] Other unknown sources: 0.05.

[0166] This invention adopts the principle of maximum probability, taking the category with the highest probability value as the current dominant pollution source type, recording its pollution source type identifier and occurrence time interval, and using it as a control trigger condition.

[0167] The system sets a confidence threshold for pollution source identification, with a default value of 0.6. This means that the system will only trigger the subsequent control response module when the predicted probability of a pollution source exceeds 0.6 and the pollutant concentration shows an upward trend, thus avoiding invalid responses caused by misidentification.

[0168] This invention includes a pollution source control strategy library, which pre-sets control schemes for various types of pollution sources. Each strategy consists of the following fields:

[0169] Types of pollution sources (e.g., dust from earthwork).

[0170] Response measure type (e.g., sprinkler system);

[0171] Response trigger threshold (e.g., PM10 > 200 micrograms per cubic meter);

[0172] Execution parameters (e.g., spray frequency, nozzle angle, duration);

[0173] Linkage device number and control command template.

[0174] Once a pollution source type is identified as the primary cause, the system searches its policy library and performs rule matching based on pollution level and operating parameters. For example:

[0175] If the pollution source is "dust from earthwork" and the current PM10 concentration exceeds the set threshold of 200 micrograms per cubic meter, then a spraying strategy will be used.

[0176] If the pollution source is "welding fumes", and the disturbance index is greater than 0.6, and the operation is carried out in a closed structural area, then the exhaust system control strategy should be matched.

[0177] After matching is completed, the system generates a corresponding set of control instructions, which includes specific execution parameters such as the device number to be turned on, execution duration, and operating power level. The instruction set will be sent to the device control module through the communication interface.

[0178] Based on the generated control instruction set, the system sends control signals to the interface of intelligent control equipment deployed at the construction site. The types of equipment include, but are not limited to:

[0179] Sprinkler system linkage control: When the pollution source type is "earthwork dust", the system will control the sprinkler equipment to start, and dynamically adjust the spray angle and nozzle spray range based on wind speed and direction data. If the wind speed exceeds 6 meters per second, the system will reduce the spray intensity to avoid secondary dust diffusion.

[0180] Localized exhaust and ventilation system: When the pollution source is "welding fumes," the system activates the exhaust equipment in the welding area, automatically selecting either high-level exhaust or fixed-point exhaust mode. The airflow is adjustable in three levels based on the pollutant concentration. The system prioritizes processing the semi-enclosed work space to control the accumulation of harmful gases.

[0181] Construction vehicle operation restrictions: When the pollution source type is "exhaust emissions," the system implements access control for construction vehicles, restricting unnecessary vehicles from entering or leaving high-pollution areas. The dispatching system automatically adjusts vehicle routes, prioritizing the use of electric or low-emission vehicles.

[0182] All control commands have an execution confirmation mechanism, meaning that after receiving a control command, the device must return a status signal to the control platform to confirm successful control execution and ensure safe and effective closed-loop feedback.

[0183] When the pollution source identified by the system cannot be directly intervened through physical control means, or when the control equipment is in a faulty, maintenance, or offline state, the system will automatically switch to the management suggestion output mode.

[0184] In this mode, the system generates optimization suggestions based on the current pollution source type and operation type, including but not limited to:

[0185] It is recommended to adjust the operation time to nighttime or a period of low disturbance;

[0186] It is recommended to suspend high-pollution processes and prioritize the completion of low-pollution tasks;

[0187] It is recommended to increase protective facilities, such as portable dustproof cloths and temporary ventilation facilities;

[0188] Relevant personnel are reminded to strengthen protective measures.

[0189] The above suggestions will be automatically pushed to the construction scheduling platform and the management mobile application, and can be manually reviewed and confirmed.

[0190] Example 2, please refer toFigure 2 As shown in this embodiment, a construction site environmental quality intelligent real-time control system includes:

[0191] Environmental sensing and acquisition module: Collects real-time environmental data of the construction site, including particulate matter concentration, gaseous pollutant concentration, noise level and meteorological parameters;

[0192] Construction Operation Information Acquisition Module: Acquires operation type tags related to construction activities to identify the current construction stage or specific operation type; the operation type tags include earthwork operations, steel structure welding, cutting operations, and machinery operation.

[0193] Micro-disturbance sensing and calculation module: calculates the micro-environmental disturbance index related to construction disturbance. The micro-environmental disturbance index is comprehensively evaluated based on local wind speed changes, construction equipment operating frequency and vehicle entry and exit frequency to reflect the trend of pollutant diffusion and accumulation in the construction area.

[0194] The discrimination module inputs the environmental data, operation type labels and microenvironmental disturbance index into the improved pollution trend analysis model, constructs a time series model that integrates multi-dimensional input features, and outputs pollutant concentration trend prediction results and pollution source type discrimination results.

[0195] Optimization module: Based on the pollution source type discrimination results, evaluate the accuracy of current pollution source differentiation, and adjust the pollution trend analysis model structure or optimize training weights according to the evaluation results;

[0196] Control Response Module: Implements control measures based on pollution source identification results. These control measures include turning on the sprinkler system, starting the local ventilation system, and restricting or adjusting vehicle operations.

[0197] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent real-time control of environmental quality at construction sites, characterized in that: The method comprises the following steps: Collecting real-time environmental data of the construction site, including particulate matter concentration, gas pollutant concentration, noise level and meteorological parameters; Obtaining an operation type label related to the construction activity, which is used to identify the current construction stage or specific operation type; the operation type label includes earthwork operation, steel structure welding, cutting operation and mechanical operation; Calculating a micro-environmental disturbance index related to the construction disturbance, which is comprehensively evaluated according to the local wind speed change, construction equipment operation frequency and vehicle access frequency, and reflects the pollution diffusion and superposition trend in the construction area; Inputting the environmental data, operation type label and micro-environmental disturbance index into the improved pollution trend analysis model to construct a time series model integrating multi-dimensional input features, and outputting pollution concentration trend prediction results and pollution source type discrimination results; Based on the pollution source type discrimination results, evaluating the current pollution source discrimination accuracy, and adjusting the pollution trend analysis model structure or optimizing the training weight according to the evaluation results; Implementing control measures according to the pollution source identification results, the control measures including starting the spraying system, starting the local exhaust system, limiting vehicle operation or adjusting the operation. 2.The intelligent real-time regulation and control method for construction site environment quality according to claim 1, characterized in that: The operation type label related to the construction activity is obtained, including: Based on the construction task plan data, the current operation content is extracted by time period and mapped to the preset operation type label set; The system collects image stream of the construction area, uses convolutional neural network model to identify the operation behavior of the image, and cross- validates the identification results with the plan data; Based on the construction equipment operation state data, including equipment ID, operation time and operation frequency indicators, the device-driven operation label is automatically generated through the set operation equipment and label mapping relationship; The operation type label results generated by the plan data, video identification results and equipment operation data are comprehensively judged, and the final label is output by using weighted voting mechanism. 3.The intelligent real-time regulation method for construction site environment quality according to claim 2, characterized in that: The calculation of the micro-environmental disturbance index includes: Obtain real-time wind speed data of multiple monitoring points in the construction area, calculate the standard deviation of wind speed in unit time based on time sliding window, and take the standard deviation as the wind speed disturbance factor to reflect the local airflow fluctuation intensity; Collect the operation state data of the construction equipment, including the number of equipment start-stop times, continuous operation time and operation intensity indicators, and map each type of equipment behavior to the disturbance score through the corresponding relationship table of equipment type and disturbance factor; Statistical vehicle access frequency of the construction area entrance and exit within a set time period, and set the disturbance weight combined with the vehicle type to calculate the vehicle disturbance factor; The wind speed disturbance factor, disturbance score and vehicle disturbance factor are normalized and superimposed according to the weighted formula to form the micro-environmental disturbance index.

4. The intelligent real-time regulation and control method for construction site environment quality according to claim 3, characterized in that: The environmental data, operation type label and micro-environmental disturbance index are input into the improved pollution trend analysis model to construct a time series model integrating multi-dimensional input features, including: Reconstructing the time series of the environmental data based on the set time step, the environmental data including particulate matter concentration, gas pollutant concentration, noise level and meteorological parameters, and using sliding window method to construct historical data sequence as input features of the model; The job type label is vectorized, the classification label is converted into a fixed dimension sparse vector by using a one-hot encoding method, and is aligned with the environmental data of the corresponding time period, so as to realize the synchronous mapping of the label and the time series data; The microenvironment disturbance index is normalized to convert it into a continuous variable with the same scale and embedded into the time series feature matrix to form the input feature set; Based on the constructed multi-dimensional input features, a long short-term memory neural network is used as the pollution trend analysis model structure to predict the pollution concentration trend in the future preset time range, and a multi-classification recognition node of the pollution source type is set in the output layer to realize the joint output of pollution trend prediction and pollution source type identification.

5. The intelligent real-time regulation and control method for construction site environmental quality according to claim 4, characterized in that: Based on the pollution source type identification result, the current pollution source differentiation accuracy is evaluated, including: The pollution source type identification result is compared with the manually labeled real job type data, and the accuracy, recall rate and F1 value indicators are used to measure the identification performance of the pollution trend analysis model in different pollution source classification tasks; Based on the preset performance threshold, it is judged whether the accuracy of the current pollution trend analysis model in identifying a certain type of pollution source is lower than the expected standard, if the F1 value of any category is lower than the set threshold, the category is marked as a weak identification category.

6. The intelligent real-time regulation and control method for construction site environmental quality according to claim 5, characterized in that: According to the evaluation result, the pollution trend analysis model structure or the training weight is adjusted, including: for the weak identification category, increasing the weight of its sample in the training sample, or increasing its sample quantity by oversampling technology; According to the overall performance change of the pollution trend analysis model, the network structure parameters of the pollution trend analysis model are dynamically adjusted, including increasing the number of hidden layers, adjusting the learning rate or replacing the activation function.

7. The intelligent real-time regulation and control method for construction site environmental quality according to claim 6, characterized in that: According to the pollution source identification result, the control measures are implemented, including: Based on the pollution source type identification result output by the pollution trend analysis model, the category of the pollution main cause is identified, including soil dust, welding dust, exhaust emission or particulate matter superposition; Matching the pollution source type with the control rules in the preset control strategy library, the control rules include response threshold, control mode, duration and device priority, forming an executable control instruction set; The control instruction is sent to the corresponding control device, including automatically starting the regional spraying system for the identified soil dust pollution source, and controlling the spraying angle and frequency according to the wind direction; for the identified welding dust pollution source, start the local exhaust or high-level exhaust system, and preferentially act on the closed or semi-closed operation area; for the identified exhaust pollution pollution source, dispatch low-emission vehicles to replace.

8. An intelligent real-time construction site environment quality control system for implementing the intelligent real-time construction site environment quality control method of any one of claims 1-7, characterized in that: Including: Environmental perception acquisition module: collecting real-time environmental data of the construction site, including particulate matter concentration, gas pollutant concentration, noise level and meteorological parameters; Construction operation information acquisition module: obtaining job type labels related to construction activities, which are used to identify the current construction stage or specific job type; the job type label includes earthwork, steel structure welding, cutting operation and mechanical operation; The perturbation perception calculation module calculates a micro-environmental disturbance index related to the construction disturbance, the micro-environmental disturbance index being comprehensively evaluated according to local wind speed variation, construction equipment operation frequency and vehicle access frequency, and reflecting the pollution diffusion and superposition trend in the construction area. The discrimination module inputs the environmental data, the work type label and the micro-environmental disturbance index into the improved pollution trend analysis model, constructs a time series model fusing multi-dimensional input features, and outputs pollution concentration trend prediction results and pollution source type discrimination results. The optimization module evaluates the current pollution source discrimination accuracy based on the pollution source type discrimination results, and adjusts the pollution trend analysis model structure or optimizes the training weight according to the evaluation results. The control response module implements control measures according to the pollution source identification results, and the control measures include spraying system opening, local exhaust system starting, vehicle operation restriction or work adjustment suggestion.

Images