Building construction safety monitoring system based on Internet of Things technology

Abstract

The invention discloses a building construction safety monitoring system based on the Internet of Things technology, and the system comprises the following steps: collecting various types of data of construction personnel, equipment and environment, constructing a unified data sequence, extracting feature information through employing an incremental principal component analysis method, and further employing an elliptical envelope line algorithm by the system, the method comprises the following steps: establishing an anomaly discrimination model in a plurality of local feature spaces, realizing the recognition of abnormal behaviors, integrating noise suppression processing and a dynamic parameter updating mechanism by a system, improving the monitoring accuracy and adaptive ability, combining the states of personnel, equipment and environment on a construction site, generating a corresponding risk level and an early warning control instruction by the system, and improving the early warning efficiency. Real-time risk identification and response linkage is realized, and the safety guarantee level of a construction site is improved. The method has the key characteristics of multi-source data fusion, boundary adaptive updating, risk response linkage and the like, and intelligent monitoring and dynamic intervention of construction site risks are realized.

Description

Technical Field

[0001] This invention relates to the field of construction safety management technology, and in particular to a construction safety monitoring system based on Internet of Things (IoT) technology. Background Technology

[0002] With the continuous advancement of smart construction sites and building information technology, real-time monitoring of construction site conditions based on the Internet of Things (IoT) has become a key application area for ensuring construction safety and improving management efficiency. Currently, construction safety monitoring systems typically rely on sensor networks or video acquisition devices to collect data on construction personnel behavior, equipment operation, and environmental parameters, and then perform anomaly detection and early warning generation through centralized processing.

[0003] However, existing construction safety monitoring methods still have significant shortcomings in processing dynamic multi-source data and in anomaly detection and risk response. On the one hand, traditional feature extraction methods have limited modeling capabilities for multi-dimensional heterogeneous data, lack an incremental online feature update mechanism, and are difficult to adapt to the non-stationarity and time-varying nature of construction site data flows, easily leading to model degradation and increased anomaly misjudgment rates. On the other hand, commonly used anomaly detection boundary construction methods fail to incorporate local structural differences in the sample space and use a single global discrimination model for risk classification, making it unable to effectively address the accurate identification of multimodal risk features in complex construction scenarios. Furthermore, existing systems largely rely on central servers for risk assessment and response strategy push, lacking edge deployment capabilities and failing to meet the needs of rapid on-site cascading responses.

[0004] Therefore, how to provide a construction safety monitoring system based on Internet of Things (IoT) technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a construction safety monitoring system based on Internet of Things (IoT) technology. This invention combines multi-source data acquisition, incremental principal component analysis, improved elliptical envelope modeling, and edge intelligent discrimination mechanism to describe in detail the entire process of dynamic perception, feature extraction, anomaly detection, model adaptive updating, and risk control response of construction personnel, equipment, and environmental status at the construction site. It has the advantages of high recognition accuracy, fast response time, and strong deployment flexibility.

[0006] A construction safety monitoring system based on Internet of Things (IoT) technology according to an embodiment of the present invention includes the following steps: The data acquisition module is used to collect data on the behavior of construction workers, the operation of construction equipment, and environmental monitoring data, and to preprocess the data to generate a multi-dimensional raw data sequence. The feature extraction module is used to receive multidimensional raw data sequences, extract principal component features through incremental principal component analysis, and dynamically output the low-dimensional feature vector and feature space state of the current period. The boundary modeling module receives low-dimensional feature vectors and feature space states, constructs anomaly discrimination boundary models based on the elliptical envelope algorithm, and outputs a set of boundary parameters. The two-stage discrimination module is used to receive the boundary parameter set and low-dimensional feature vector, perform noise suppression processing on the input feature vector, and perform abnormal state discrimination based on the anomaly discrimination boundary model, and output preliminary risk identification results. The model adaptation module receives the preliminary risk identification results and feature space status, and dynamically updates the core parameters of the feature extraction module and the boundary modeling module according to the discrimination deviation. The edge deployment module is used to perform feature extraction, boundary modeling, discrimination calculation and parameter update processes in edge computing nodes deployed at the construction site, and generate edge discrimination output for the current period; The results output module is used to receive the edge discrimination output and the preliminary risk identification results, and combine them with the construction personnel, equipment and environmental status information to generate the construction risk level and the corresponding early warning and control instructions.

[0007] Optionally, modules can be integrated using the following methods: Collect construction worker behavior data, construction equipment operation data, and environmental monitoring data, perform preprocessing, and generate multidimensional raw data sequences; Incremental principal component analysis is used to extract features from the multidimensional original data sequence to obtain the low-dimensional feature vector and feature space state of the current period; Based on the low-dimensional feature vectors and feature space states, an anomaly discrimination boundary model is constructed using the elliptical envelope algorithm, and a boundary parameter set is output. The system receives the low-dimensional feature vector and boundary parameter set, performs noise suppression processing, and performs anomaly discrimination based on the anomaly discrimination boundary model, outputting preliminary risk identification results. Based on the preliminary risk identification results and feature space status, the parameters of the feature extraction module and the boundary modeling module are dynamically updated according to the deviation. The edge computing nodes deployed at the construction site perform the processes of feature extraction, boundary modeling, discrimination calculation and parameter update, and generate the edge discrimination output for the current period. The system receives the edge discrimination output and preliminary risk identification results, and combines them with information on the status of construction personnel, equipment, and the environment to generate construction risk levels and corresponding early warning and control instructions.

[0008] Optional preprocessing includes data time alignment, missing value imputation, sliding window segmentation, data format normalization, and feature value standardization.

[0009] Optional, the generation of low-dimensional feature vectors and feature space states includes: Extract key feature fields from the multidimensional raw data sequence that correspond to construction worker behavior data, construction equipment operation data, and environmental monitoring data, and construct a data sample set with a unified structure. Perform time series partitioning on the data sample set, segmenting it according to the set time window size and sliding step size, and generating a multidimensional original data subsequence for the current time window; For the multidimensional raw data subsequence of the current time window, initialize the cumulative sample count parameter, the mean vector of the feature field and the set of principal component direction vectors in the feature extraction module to construct the initial feature space state; After receiving the multidimensional raw data subsequence of the current time window, update the cumulative sample count parameter, and calculate the offset and correct the feature field mean vector based on the difference between the multidimensional raw data subsequence and the current feature field mean vector. By utilizing the relationship between the offset and the current set of principal component direction vectors, the incremental principal component analysis method is applied to recursively update the set of principal component direction vectors, and orthogonality and unitity preservation operations are performed to form the principal component space of the current period. The multidimensional original data subsequence of the current time window is dimensionality-reduced and mapped in the principal component space composed of the principal component direction vector set to generate the low-dimensional feature vector of the current period. Output the low-dimensional feature vector of the current period along with the feature space state, which includes the set of principal component direction vectors, the mean vector of the feature fields, and the cumulative sample count parameter.

[0010] Optionally, the output of the boundary parameter set includes: Collect the set of low-dimensional feature vectors for the current period to form a feature sample set, and receive the data points that have been identified as abnormal in the previous period to form a historical abnormal sample set. Based on the feature sample set and the historical abnormal sample set, the low-dimensional feature space is divided by spatial clustering method to obtain multiple local feature subspaces. Each local feature subspace contains a sample set with similar structural distribution. For each local feature subspace, calculate the sample mean vector of that local feature subspace. Covariance matrix and anomaly detection threshold Construct a local anomaly detection boundary model; Upon receiving a new low-dimensional feature vector x to be judged, the Mahalanobis distance function value calculated in each local subspace is used. The subspace that minimizes the Mahalanobis distance function is selected as the home subspace, and the corresponding local anomaly discrimination boundary model is called within the home subspace for discrimination. At that time, preliminary anomaly detection results are generated; When a local feature subspace has a dense distribution of abnormal samples or frequent out-of-bounds discrimination results in multiple consecutive periods, the anomaly feedback expansion mechanism is triggered to automatically divide a new local feature subspace and construct a new anomaly discrimination boundary model based on the local feature subspace, thereby achieving adaptive expansion of the boundary model structure. The output contains the boundary parameter set of the boundary model for all local anomaly detection methods. The boundary parameter set includes the sample mean vector corresponding to each subspace. Covariance matrix and anomaly detection threshold This set of boundary parameters will then be used in the subsequent two-stage discrimination module to perform anomaly identification operations.

[0011] Optionally, the output of the preliminary risk identification results includes: Collect the low-dimensional feature vector of the current period and construct a feature time buffer by combining it with the historical low-dimensional feature vectors within the previous fixed number of periods, and arrange them in chronological order to form a feature sliding window sequence; In the feature sliding window sequence, a moving average filtering process is performed on each low-dimensional feature vector to be judged. The low-dimensional feature vector to be judged is smoothed by taking values ​​symmetrically before and after, suppressing short-period fluctuations and abnormal spike interference, and outputting a smooth feature vector after noise suppression. Call the boundary parameter set, which includes multiple local anomaly detection boundary models. Each local anomaly detection boundary model contains a local sample mean vector, a covariance matrix, and an anomaly detection threshold. For the current smooth feature vector, the statistical deviation in each local feature subspace is calculated in all local anomaly discrimination boundary models, and the local feature subspace with the smallest deviation is determined as the belonging subspace; Within the subspace of the origin, based on the corresponding local anomaly discrimination boundary model, the relationship between the current smooth feature vector and the anomaly discrimination threshold is compared. If the deviation is greater than the corresponding anomaly discrimination threshold, the smooth feature vector is marked as anomaly; otherwise, it is marked as normal. The labeled status of all low-dimensional feature vectors to be judged is associated with the corresponding timestamp, and the abnormal status set is output periodically to form the preliminary risk identification result for the current period.

[0012] Optionally, updating the parameters of the feature extraction module and the boundary modeling module includes: Receive the preliminary risk identification results and corresponding low-dimensional feature vectors of the current period, extract the low-dimensional feature vectors that are judged to be abnormal, and combine them with the data of the previous period to form an abnormal sample sequence. Determine whether there are multiple consecutive periods in the abnormal sample sequence that belong to the same local feature subspace and whose deviation exceeds the current anomaly detection threshold. If the condition is met, record the parameter update identifier that triggers the local feature subspace. For the local feature subspace that triggers the update flag, collect all low-dimensional feature vector samples that are labeled as normal, and calculate the updated sample mean vector and covariance matrix based on the low-dimensional feature vector sample set, while constructing new sample variance statistics. Based on the parameters of the normal sample set and the cumulative sample count, the mean vector of the feature fields and the set of principal component direction vectors in the feature extraction module are updated, and orthogonal normalization is performed on the updated set of principal component direction vectors to generate the corrected feature space state; Based on the corrected sample mean vector and covariance matrix, the boundary function of the corresponding local anomaly discrimination boundary model is updated, while keeping the original anomaly discrimination threshold unchanged, thus completing the synchronous update of the boundary parameter set. The updated parameters of the feature extraction module and the boundary modeling module are written into the system buffer and synchronously distributed to the feature extraction module and the boundary modeling module in the edge computing node before the start of the next cycle, ensuring that the anomaly detection in subsequent cycles is performed based on the updated parameters.

[0013] Optionally, the generation of construction risk levels and corresponding early warning and control instructions includes: Collect construction worker behavior data, construction equipment operation data and environmental monitoring data within the current period, extract corresponding personnel activity intensity indicators, equipment operation stability indicators and environmental abnormal change indicators, and construct a multi-source status information set; The preliminary risk identification results output in the current cycle are correlated and matched with the multi-source state information set. Based on the low-dimensional feature vector corresponding to the abnormal event, the local feature subspace to which it belongs, and the joint situation of the current personnel, equipment, and environmental status, the risk impact factor is calculated. Based on the risk impact factors and combined with the risk level classification rules, each abnormal event is assigned to a corresponding construction risk level label. The risk level classification rules are determined based on the level mapping relationship constructed in the historical risk event database. Based on the construction risk level label and the corresponding subspace identifier of the abnormal event, the early warning control template that meets the risk level, event type and subspace constraints is matched from the preset early warning response rule set, and the control template information containing response action, response range and trigger condition parameters is extracted. Based on the current construction scheduling plan and the availability status of construction equipment, structured early warning control instructions are generated according to the extracted control template information. The structured early warning control instructions include response action definitions, target equipment identifiers, and trigger condition parameters. Output the construction risk level label and structured early warning control instructions corresponding to each abnormal event in the current period, as the risk control output result of the construction site safety response system.

[0014] The beneficial effects of this invention are: (1) This invention constructs a construction safety monitoring system based on the Internet of Things, integrating multi-source data acquisition, incremental principal component analysis, improved elliptical envelope boundary modeling, and a two-stage anomaly discrimination mechanism. This enables collaborative perception and modeling of construction personnel behavior, equipment operating status, and environmental parameters. It can dynamically extract low-dimensional features, identify potential risks, and perform refined boundary division, thereby improving the accuracy of anomaly discrimination and the adaptability of boundary modeling. (2) This invention deploys feature extraction and model evolution processes at edge computing nodes, enabling localized computation and rapid risk identification and response at the construction site. This reduces the system's dependence on the central server and ensures stable operation even under network fluctuations or latency, thus significantly improving the overall response efficiency and practicality of the system. (3) This invention introduces a risk level classification and early warning response strategy library, combined with construction scheduling plans and equipment availability status, to generate structured early warning control instructions. This enables coordinated intervention and control at the construction site, providing high risk warning accuracy, flexible response strategies, and traceable control execution. This provides intelligent auxiliary decision support for construction safety management. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a construction safety monitoring system based on Internet of Things (IoT) technology proposed in this invention; Figure 2 This is a flowchart illustrating the construction of the anomaly discrimination boundary model based on the improved elliptical envelope algorithm in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-2 A construction safety monitoring system based on Internet of Things (IoT) technology includes the following steps: The data acquisition module is used to collect data on the behavior of construction workers, the operation of construction equipment, and environmental monitoring data, and to preprocess the data to generate a multi-dimensional raw data sequence. The feature extraction module is used to receive multidimensional raw data sequences, extract principal component features through incremental principal component analysis, and dynamically output the low-dimensional feature vector and feature space state of the current period. The boundary modeling module receives low-dimensional feature vectors and feature space states, constructs anomaly discrimination boundary models based on the elliptical envelope algorithm, and outputs a set of boundary parameters. The two-stage discrimination module is used to receive the boundary parameter set and low-dimensional feature vector, perform noise suppression processing on the input feature vector, and perform abnormal state discrimination based on the anomaly discrimination boundary model, and output preliminary risk identification results. The model adaptation module receives the preliminary risk identification results and feature space status, and dynamically updates the core parameters of the feature extraction module and the boundary modeling module according to the discrimination deviation. The edge deployment module is used to perform feature extraction, boundary modeling, discrimination calculation and parameter update processes in edge computing nodes deployed at the construction site, and generate edge discrimination output for the current period; The results output module is used to receive the edge discrimination output and the preliminary risk identification results, and combine them with the construction personnel, equipment and environmental status information to generate the construction risk level and the corresponding early warning and control instructions.

[0018] In this embodiment, the modules are interconnected using the following method: Collect construction worker behavior data, construction equipment operation data, and environmental monitoring data, perform preprocessing, and generate multidimensional raw data sequences; Incremental principal component analysis is used to extract features from the multidimensional original data sequence to obtain the low-dimensional feature vector and feature space state of the current period; Based on the low-dimensional feature vectors and feature space states, an anomaly discrimination boundary model is constructed using the elliptical envelope algorithm, and a boundary parameter set is output. The system receives the low-dimensional feature vector and boundary parameter set, performs noise suppression processing, and performs anomaly discrimination based on the anomaly discrimination boundary model, outputting preliminary risk identification results. Based on the preliminary risk identification results and feature space status, the parameters of the feature extraction module and the boundary modeling module are dynamically updated according to the deviation. The edge computing nodes deployed at the construction site perform the processes of feature extraction, boundary modeling, discrimination calculation and parameter update, and generate the edge discrimination output for the current period. The system receives the edge discrimination output and preliminary risk identification results, and combines them with information on the status of construction personnel, equipment, and the environment to generate construction risk levels and corresponding early warning and control instructions.

[0019] In this embodiment, preprocessing includes data time alignment, missing value imputation, sliding window segmentation, data format normalization, and feature value standardization.

[0020] In this embodiment, the generation of low-dimensional feature vectors and feature space states includes: Extract key feature fields from the multidimensional raw data sequence that correspond to construction worker behavior data, construction equipment operation data, and environmental monitoring data, and construct a data sample set with a unified structure. Perform time series partitioning on the data sample set, segmenting it according to the set time window size and sliding step size, and generating a multidimensional original data subsequence for the current time window; For the multidimensional raw data subsequence of the current time window, initialize the cumulative sample count parameter, the mean vector of the feature field and the set of principal component direction vectors in the feature extraction module to construct the initial feature space state; After receiving the multidimensional raw data subsequence of the current time window, update the cumulative sample count parameter, and calculate the offset and correct the feature field mean vector based on the difference between the multidimensional raw data subsequence and the current feature field mean vector. By utilizing the relationship between the offset and the current set of principal component direction vectors, the incremental principal component analysis method is applied to recursively update the set of principal component direction vectors, and orthogonality and unitity preservation operations are performed to form the principal component space of the current period. The multidimensional original data subsequence of the current time window is dimensionality-reduced and mapped in the principal component space composed of the principal component direction vector set to generate the low-dimensional feature vector of the current period. Output the low-dimensional feature vector of the current period and the feature space state containing the set of principal component direction vectors, the mean vector of feature fields, and the cumulative sample number parameter; The time window size and sliding step size are configured based on the sampling frequency of sensor data, feature evolution rate, and actual response requirements at the construction site. The time window size is defined as the continuous time period covered by a single feature extraction, and the sliding step size is defined as the time offset between two adjacent windows. The time window size is typically selected between 10 and 300 seconds to ensure sufficient coverage of multidimensional raw data to reflect the current construction status. The sliding step size is set to 10% to 50% of the window length to balance feature update frequency and computational overhead, enabling continuous evolution of the feature space and online modeling capabilities. All of the above parameters can be flexibly configured according to the on-site construction rhythm, sensor node deployment density, and anomaly response time requirements. Performing orthogonality and unitity preservation operations refers to the process of orthogonalizing the set of principal component direction vectors one by one using Gram-Schmidt orthogonalization or its equivalent numerical method during the update of the principal component direction vector set, ensuring that different principal component directions are mutually orthogonal; and performing unit norm normalization on each orthogonal vector to keep its Euclidean norm at 1, thereby maintaining the stability and invertibility of the feature space formed by the principal component directions in terms of geometric structure. The above operations can be performed in real time in each incremental update cycle, ensuring that the set of principal component directions does not undergo linear correlation degradation or scale drift in long-term operation, and guaranteeing the stability of feature projection and anomaly boundary discrimination.

[0021] In this embodiment, the output of the boundary parameter set includes: Collect the set of low-dimensional feature vectors for the current period to form a feature sample set, and receive the data points that have been identified as abnormal in the previous period to form a historical abnormal sample set. Based on the feature sample set and the historical abnormal sample set, the low-dimensional feature space is divided by spatial clustering method to obtain multiple local feature subspaces. Each local feature subspace contains a sample set with similar structural distribution. For each local feature subspace, calculate the sample mean vector of that local feature subspace. Covariance matrix and anomaly detection threshold A local anomaly detection boundary model is constructed, where the boundary function is defined as follows: ; in, This represents the low-dimensional feature vector to be judged in the current period. For the first The sample mean vector of each local feature subspace Let covariance be the local eigenspace. It is the inverse of the covariance matrix. The result of the Mahalanobis distance function is used to determine the degree of deviation of the sample from the center of the local feature subspace. This formula originates from the squared Mahalanobis distance form in classical statistics, used to measure the deviation of a sample from the population mean in the covariance-weighted space. Based on this mathematical foundation, this application introduces a local feature subspace partitioning mechanism, improving the original global modeling method into a multi-subspace parallel modeling mode, and independently calculating the mean vector for each subspace. With covariance matrix An adaptive boundary function is constructed. Furthermore, by incorporating a historical anomaly feedback mechanism, model expansion and dynamic boundary updates are achieved for newly added anomaly patterns. In the above formula... This represents the low-dimensional feature vector of the current period. For the first The mean vector of each subspace. Its covariance matrix is The function value obtained is the inverse of the matrix. As a dimensionless anomaly deviation index, in terms of dimensional analysis, vector difference The unit is the characteristic quantity. The inverse unit of the covariance matrix is The calculation results are scalar units with consistent dimensions, that is: This ensures that the physical meaning of the formula is rigorous and that its dimensions are closed and reasonable. Upon receiving a new low-dimensional feature vector x to be judged, the Mahalanobis distance function value calculated in each local subspace is used. The subspace that minimizes the Mahalanobis distance function is selected as the home subspace, and the corresponding local anomaly discrimination boundary model is called within the home subspace for discrimination. At that time, preliminary anomaly detection results are generated; When a local feature subspace has a dense distribution of abnormal samples or frequent out-of-bounds discrimination results in multiple consecutive periods, the anomaly feedback expansion mechanism is triggered to automatically divide a new local feature subspace and construct a new anomaly discrimination boundary model based on the local feature subspace, thereby achieving adaptive expansion of the boundary model structure. The output contains the boundary parameter set of the boundary model for all local anomaly detection methods. The boundary parameter set includes the sample mean vector corresponding to each subspace. Covariance matrix and anomaly detection threshold This set of boundary parameters is then used in the subsequent two-stage discrimination module to perform anomaly identification operations. The data points identified as anomalous in the previous cycle refer to the set of sample points that were identified as anomalous in the previous cycle of the construction safety monitoring system. These points were identified by executing the anomalous state discrimination process in the two-stage discrimination module, comparing the input low-dimensional feature vector with the currently deployed anomalous discrimination boundary model, and generating anomalous state identification results. Each data point in this set corresponds to the original multi-dimensional sensor data collected at the actual construction site. After being processed by the feature extraction module, it is mapped to a low-dimensional representation in the feature space. If the data point exceeds the anomalous threshold of the corresponding subspace in the Mahalanobis distance function result, it is identified as high-risk or anomalous. The system records it as a historical anomalous sample and uses it in subsequent cycles to update the boundary structure, assist model expansion, or optimize the judgment threshold to enhance the adaptability of the anomalous detection model to marginally distributed samples or new anomalous patterns. Spatial clustering methods partition low-dimensional feature spaces by obtaining a set of feature samples composed of all low-dimensional feature vectors within the current period, and then using unsupervised clustering algorithms to perform pattern partitioning on the sample distribution in the feature space. Feature points with similar structures, continuous distribution density, or close aggregation are grouped into the same cluster subset. The clustering method can employ density-based clustering, distance-based partitioning, or hierarchical clustering algorithms, and the final partitioning result is determined based on parameters such as the set cluster radius, minimum number of samples within a cluster, or distance threshold. Each cluster subset is defined as a local feature subspace, which serves as the basic input region for constructing the corresponding local anomaly discrimination boundary model, ensuring the adaptability of the anomaly boundary and the consistency of regional features.

[0022] In this embodiment, the output of the preliminary risk identification results includes: Collect the low-dimensional feature vector of the current period and construct a feature time buffer by combining it with the historical low-dimensional feature vectors within the previous fixed number of periods, and arrange them in chronological order to form a feature sliding window sequence; In the feature sliding window sequence, a moving average filtering process is performed on each low-dimensional feature vector to be judged. The low-dimensional feature vector to be judged is smoothed by taking values ​​symmetrically before and after, suppressing short-period fluctuations and abnormal spike interference, and outputting a smooth feature vector after noise suppression. Call the boundary parameter set, which includes multiple local anomaly detection boundary models. Each local anomaly detection boundary model contains a local sample mean vector, a covariance matrix, and an anomaly detection threshold. For the current smooth feature vector, the statistical deviation in each local feature subspace is calculated in all local anomaly discrimination boundary models, and the local feature subspace with the smallest deviation is determined as the belonging subspace; Within the subspace of the origin, based on the corresponding local anomaly discrimination boundary model, the relationship between the current smooth feature vector and the anomaly discrimination threshold is compared. If the deviation is greater than the corresponding anomaly discrimination threshold, the smooth feature vector is marked as anomaly; otherwise, it is marked as normal. The labeled states of all low-dimensional feature vectors to be judged are associated with their corresponding timestamps, and an abnormal state set is output periodically to form the preliminary risk identification result for the current period; Performing a moving average filtering process involves, within the feature time-series buffer, backtracking a fixed number of historical period data points chronologically backward from the current period's low-dimensional feature vector. Based on this set of historical feature vectors, the average value for each feature dimension is calculated. This average value replaces the original feature value for the current period, resulting in a smoothed feature result. Specifically, with a sliding window length of w, the value of each feature dimension in the current period is obtained by averaging its value with the corresponding dimensions from the previous w-1 periods. This reduces disturbances caused by random noise or transient anomalies, allowing the feature sequence to retain its trend while removing high-frequency interference. This process is performed independently for each low-dimensional feature dimension, yielding a fully smoothed low-dimensional feature vector.

[0023] In this embodiment, updating the parameters of the feature extraction module and the boundary modeling module includes: Receive the preliminary risk identification results and corresponding low-dimensional feature vectors of the current period, extract the low-dimensional feature vectors that are judged to be abnormal, and combine them with the data of the previous period to form an abnormal sample sequence. Determine whether there are multiple consecutive periods in the abnormal sample sequence that belong to the same local feature subspace and whose deviation exceeds the current anomaly detection threshold. If the condition is met, record the parameter update identifier that triggers the local feature subspace. For the local feature subspace that triggers the update flag, collect all low-dimensional feature vector samples that are labeled as normal, and calculate the updated sample mean vector and covariance matrix based on the low-dimensional feature vector sample set, while constructing new sample variance statistics. Based on the parameters of the normal sample set and the cumulative sample count, the mean vector of the feature fields and the set of principal component direction vectors in the feature extraction module are updated, and orthogonal normalization is performed on the updated set of principal component direction vectors to generate the corrected feature space state; Based on the corrected sample mean vector and covariance matrix, the boundary function of the corresponding local anomaly discrimination boundary model is updated, while keeping the original anomaly discrimination threshold unchanged, thus completing the synchronous update of the boundary parameter set. The updated parameters of the feature extraction module and the boundary modeling module are written into the system buffer and synchronously distributed to the feature extraction module and the boundary modeling module in the edge computing node before the start of the next cycle, ensuring that the anomaly detection in subsequent cycles is performed based on the updated parameters.

[0024] In this embodiment, the generation of construction risk levels and corresponding early warning and control instructions includes: Collect construction worker behavior data, construction equipment operation data and environmental monitoring data within the current period, extract corresponding personnel activity intensity indicators, equipment operation stability indicators and environmental abnormal change indicators, and construct a multi-source status information set; The preliminary risk identification results output in the current cycle are correlated and matched with the multi-source state information set. Based on the low-dimensional feature vector corresponding to the abnormal event, the local feature subspace to which it belongs, and the joint situation of the current personnel, equipment, and environmental status, the risk impact factor is calculated. Based on the risk impact factors and combined with the risk level classification rules, each abnormal event is assigned to a corresponding construction risk level label. The risk level classification rules are determined based on the level mapping relationship constructed in the historical risk event database. Based on the construction risk level label and the corresponding subspace identifier of the abnormal event, the early warning control template that meets the risk level, event type and subspace constraints is matched from the preset early warning response rule set, and the control template information containing response action, response range and trigger condition parameters is extracted. Based on the current construction scheduling plan and the availability status of construction equipment, structured early warning control instructions are generated according to the extracted control template information. The structured early warning control instructions include response action definitions, target equipment identifiers, and trigger condition parameters. Output the construction risk level label and structured early warning control instructions corresponding to each abnormal event in the current period, as the risk control output result of the construction site safety response system; The risk level classification rule is a hierarchical standard system built upon statistical analysis of historical construction anomalies, quantifying events according to their risk impact factors. These factors consist of multiple dimensions: the degree of deviation of the anomaly from the feature space within the current period, the frequency of anomalies, the number of construction personnel involved, equipment operational stability indicators, and the magnitude of environmental parameter changes. These dimensions are then comprehensively calculated using a weighted scoring model to obtain an event's risk score. Based on the score's position within a preset threshold range, different risk level labels are assigned. Each level corresponds to a specific early warning strategy and response level, enabling quantitative assessment and level allocation of event risk. The pre-defined early warning response rule set refers to a set of mapping rules pre-constructed during the system deployment phase, based on different anomaly types, risk levels, and their associated characteristic subspaces in the construction scenario, combined with industry safety standards, construction management requirements, and on-site control mechanisms. Each rule in this set includes the following fields: risk level label, anomaly event category, associated subspace identifier, corresponding response action definition (e.g., equipment shutdown, personnel evacuation, information reporting), response scope (e.g., specified construction area or specific personnel group), and trigger condition parameters (e.g., risk duration threshold, risk density, number of consecutive triggers). This rule set is invoked at runtime to select a matching response strategy based on the currently detected risk situation, thereby achieving automated intervention and control of the construction site. The current construction scheduling plan and construction equipment availability status refer to a dynamic set of information describing on-site resource allocation and equipment operating capacity during the construction process. The construction scheduling plan reflects the arrangement of various construction tasks at different time periods, including task number, scheduled start time, estimated duration, responsible personnel, and assigned construction area. The construction equipment availability status includes the current operating status of each piece of equipment, remaining available time, number of assigned tasks, and equipment maintenance plan. This information set is typically generated or updated in real time by the construction site's scheduling and equipment management systems. It is used to determine the feasibility of target equipment, whether it is idle or under low load, and accordingly allocate intervention resources and scheduling response priorities when generating early warning control commands.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to a construction site. Due to frequent personnel changes, high-load operation of machinery and equipment, and drastic environmental changes, the identification and early warning of safety risks at construction sites has become an urgent problem to be solved. Traditional construction safety monitoring methods mostly rely on static rule settings and manual inspections, which have limited ability to detect abnormal behavior and cannot achieve real-time feedback. They also suffer from problems such as delayed response, high false alarm rate, and inability to adapt to deployment in multiple scenarios.

[0026] The system has been successfully applied in a large-scale municipal project with a building area exceeding 100,000 square meters. The project involved over 150 construction workers daily and a wide variety of construction equipment, including tower cranes, concrete pump trucks, and steel reinforcement processing machinery. More than 50 environmental sensor nodes, over 20 video acquisition units, and several edge computing devices were deployed on-site. The system's deployment goal was to automatically generate risk behavior identification and early warning control commands for personnel, equipment, and the environment.

[0027] In this scenario, the system first receives data from sensors and video nodes via a data acquisition module. This data includes construction worker location information, acceleration sensing data, equipment operating current, voltage, vibration frequency, ambient temperature and humidity, and PM2.5 concentration. All data undergoes standardized structure at the acquisition end, including data field filtering, timestamp alignment, and vector normalization, forming a multidimensional raw data sequence. Subsequently, the system segments the data into time series segments based on a set time window and sliding step size (e.g., a 60-second window and a 30-second step size). Incremental Principal Component Analysis (IPCA) is then used to dynamically update the mean vector of the feature fields and the set of principal component directions, yielding the low-dimensional feature vector and feature space state for the current period.

[0028] The generated low-dimensional feature vectors are then input into the boundary modeling module. This module combines historical anomaly data (such as behavioral data of individuals previously marked as anomalous) and, using an improved elliptic envelope algorithm, constructs multiple discriminative boundary function models in the local feature subspace. This model introduces dynamic clustering and adaptive subspace expansion mechanisms, which not only improves its adaptability to anomaly events but also dynamically adjusts the model structure through a feedback mechanism, ensuring detection accuracy.

[0029] In the feature discrimination process, to reduce noise interference, the system employs a moving average filtering method. Within the constructed feature sliding window, the feature vector to be detected is symmetrically smoothed to remove short-period disturbances such as sudden jitter and signal spikes. The smoothed feature vector is then substituted into the discrimination boundary model to identify whether it exceeds the boundary, thereby outputting preliminary risk identification results, including multiple risk labels such as behavioral boundary violations, excessive equipment vibration, and sudden environmental changes.

[0030] Based on the discrimination results and the degree of feature offset, the system calls the feedback mechanism to update the parameters of the feature extraction module and the boundary modeling module, and promptly corrects the principal component direction vector and the discrimination boundary. The result output module combines the current work plan, personnel schedule and equipment operation status information obtained from the construction scheduling system, integrates the risk label and status data, divides the risk level according to the preset risk classification rules, and automatically matches the response strategy template according to the early warning rule library to generate corresponding early warning control instructions, such as "personnel evacuation suggestion", "equipment emergency shutdown command" or "local area environmental adjustment suggestion", etc.

[0031] To verify the actual effectiveness of the system, we deployed and ran it continuously at the construction site for 60 days and compared it with the traditional safety management system (mainly based on inspection and video surveillance). The results are shown in Tables 1 and 2 below.

[0032] Table 1 Comparative Test Results of Intelligent Monitoring Systems at Construction Sites Project Category Traditional system recognition rate The recognition rate of the system of this invention Increase Identification of personnel working at heights without wearing safety ropes 71.2% 95.6% +24.4% Equipment abnormal vibration identification 64.8% 93.1% +28.3% Identification of sudden changes in toxic gas concentration 57.5% 90.4% +32.9% Response latency (seconds) 87 seconds 11 seconds -76 seconds False alarm rate 14.3% 3.5% -10.8% As can be seen from the table, the system of the present invention exhibits excellent recognition capabilities and response speed in a variety of typical scenarios, especially showing significant advantages in the recognition of high-frequency, low-amplitude, and unstructured risk behaviors.

[0033] Further data analysis showed that in the first month after using the system, the number of minor safety accidents at the construction site caused by failure to detect abnormal behavior in a timely manner decreased by 73.5% year-on-year, and the potential equipment damage rate decreased by 64.2%. Statistical analysis of the execution of control commands generated by the system revealed that the average response time after automatically triggering an early warning was reduced from 6 minutes to 1 minute and 15 seconds, greatly improving the efficiency of on-site emergency response.

[0034] Table 2. Statistics on Response to Early Warning and Control Commands Risk type Trigger count Number of successful responses Response success rate Average response time Working at height was not done in accordance with regulations. 42 41 97.6% 58 seconds Excessive equipment vibration 35 33 94.3% 76 seconds Abnormal ambient gas concentration 23 23 100% 48 seconds Overall average - - 96.9% 60 seconds As can be seen from the numerical results in Tables 1 and 2 of the embodiments, the IoT-based construction safety monitoring method proposed in this invention exhibits significant performance advantages in practical applications. In terms of recognition accuracy, in key risk scenarios such as working at heights without safety harnesses, abnormal equipment vibration, and sudden changes in toxic gas concentration, the system's recognition rate is significantly better than traditional systems, with improvements of 24.4%, 28.3%, and 32.9%, respectively. This indicates that the present invention possesses stronger anomaly detection capabilities in multi-dimensional feature fusion and boundary modeling. Simultaneously, the response latency is reduced from 87 seconds to 11 seconds, significantly improving response efficiency, and the false alarm rate decreases from 14.3% to 3.5%, demonstrating the system's accuracy in implementing noise suppression and multi-layer discrimination strategies. Further analysis of the response statistics in Table 2 shows that the early warning control commands of this invention achieved a response success rate of over 95% in multiple high-risk events, with an average response time maintained within 1 minute. In particular, it achieved a 100% accurate identification and response rate in environmental gas concentration anomalies, fully demonstrating that the method possesses comprehensive advantages of high real-time performance, strong adaptability, and stability, and can effectively support risk identification and early warning control in complex dynamic construction environments.

[0035] In summary, this embodiment fully demonstrates the collaborative capability of the present invention's method in the entire process of perceiving, identifying, modeling, and controlling safety risks in construction scenarios. The system, through a complete set of modular feature processing and discrimination logic, achieves a closed-loop intelligent processing flow from data acquisition to early warning generation. Actual deployment data fully verifies that the technical solution proposed in this invention possesses strong robustness, real-time performance, and adaptability, effectively solving the problems of slow response, low identification accuracy, and poor adaptability in existing construction site safety monitoring systems.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A construction safety monitoring system based on Internet of Things (IoT) technology, characterized in that, Includes the following modules: The data acquisition module is used to collect data on the behavior of construction workers, the operation of construction equipment, and environmental monitoring data, and to preprocess the data to generate a multi-dimensional raw data sequence. The feature extraction module is used to receive multidimensional raw data sequences, extract principal component features through incremental principal component analysis, and dynamically output the low-dimensional feature vector and feature space state of the current period. The boundary modeling module receives low-dimensional feature vectors and feature space states, constructs anomaly discrimination boundary models based on the elliptical envelope algorithm, and outputs a set of boundary parameters. The two-stage discrimination module is used to receive the boundary parameter set and low-dimensional feature vector, perform noise suppression processing on the input feature vector, and perform abnormal state discrimination based on the anomaly discrimination boundary model, and output preliminary risk identification results. The model adaptation module receives the preliminary risk identification results and feature space status, and dynamically updates the core parameters of the feature extraction module and the boundary modeling module according to the discrimination deviation. The edge deployment module is used to perform feature extraction, boundary modeling, discrimination calculation and parameter update processes in edge computing nodes deployed at the construction site, and generate edge discrimination output for the current period; The results output module is used to receive the edge discrimination output and the preliminary risk identification results, and combine them with the construction personnel, equipment and environmental status information to generate the construction risk level and the corresponding early warning and control instructions.

2. A construction safety monitoring system based on Internet of Things (IoT) technology, characterized in that, The modules are connected in the following way: Collect construction worker behavior data, construction equipment operation data, and environmental monitoring data, perform preprocessing, and generate multidimensional raw data sequences; Incremental principal component analysis is used to extract features from the multidimensional original data sequence to obtain the low-dimensional feature vector and feature space state of the current period; Based on the low-dimensional feature vectors and feature space states, an anomaly discrimination boundary model is constructed using the elliptical envelope algorithm, and a boundary parameter set is output. The system receives the low-dimensional feature vector and boundary parameter set, performs noise suppression processing, and performs anomaly discrimination based on the anomaly discrimination boundary model, outputting preliminary risk identification results. Based on the preliminary risk identification results and feature space status, the parameters of the feature extraction module and the boundary modeling module are dynamically updated according to the deviation. The edge computing nodes deployed at the construction site perform the processes of feature extraction, boundary modeling, discrimination calculation and parameter update, and generate the edge discrimination output for the current period. The system receives the edge discrimination output and preliminary risk identification results, and combines them with information on the status of construction personnel, equipment, and the environment to generate construction risk levels and corresponding early warning and control instructions.

3. A construction safety monitoring system based on Internet of Things (IoT) technology according to claim 2, characterized in that, Preprocessing includes data time alignment, missing value imputation, sliding window segmentation, data format normalization, and feature value standardization.

4. A construction safety monitoring system based on Internet of Things (IoT) technology according to claim 2, characterized in that, The generation of low-dimensional feature vectors and feature space states includes: Extract key feature fields from the multidimensional raw data sequence that correspond to construction worker behavior data, construction equipment operation data, and environmental monitoring data, and construct a data sample set with a unified structure. Perform time series partitioning on the data sample set, segmenting it according to the set time window size and sliding step size, and generating a multidimensional original data subsequence for the current time window; For the multidimensional raw data subsequence of the current time window, initialize the cumulative sample count parameter, the mean vector of the feature field and the set of principal component direction vectors in the feature extraction module to construct the initial feature space state; After receiving the multidimensional raw data subsequence of the current time window, update the cumulative sample count parameter, and calculate the offset and correct the feature field mean vector based on the difference between the multidimensional raw data subsequence and the current feature field mean vector. By utilizing the relationship between the offset and the current set of principal component direction vectors, the incremental principal component analysis method is applied to recursively update the set of principal component direction vectors, and orthogonality and unitity preservation operations are performed to form the principal component space of the current period. The multidimensional original data subsequence of the current time window is dimensionality-reduced and mapped in the principal component space composed of the principal component direction vector set to generate the low-dimensional feature vector of the current period. Output the low-dimensional feature vector of the current period along with the feature space state, which includes the set of principal component direction vectors, the mean vector of the feature fields, and the cumulative sample count parameter.

5. A construction safety monitoring system based on Internet of Things (IoT) technology according to claim 2, characterized in that, The output of the boundary parameter set includes: Collect the set of low-dimensional feature vectors for the current period to form a feature sample set, and receive the data points that have been identified as abnormal in the previous period to form a historical abnormal sample set. Based on the feature sample set and the historical abnormal sample set, the low-dimensional feature space is divided by spatial clustering method to obtain multiple local feature subspaces. Each local feature subspace contains a sample set with similar structural distribution. For each local feature subspace, calculate the sample mean vector of that local feature subspace. Covariance matrix and anomaly detection threshold Construct a local anomaly detection boundary model; Upon receiving a new low-dimensional feature vector x to be judged, the Mahalanobis distance function value calculated in each local subspace is used. The subspace that minimizes the Mahalanobis distance function is selected as the home subspace, and the corresponding local anomaly discrimination boundary model is called within the home subspace for discrimination. At that time, preliminary anomaly detection results are generated; When a local feature subspace exhibits a dense distribution of abnormal samples or frequent out-of-bounds discrimination results within multiple consecutive periods, an anomaly feedback expansion mechanism is triggered to automatically divide a new local feature subspace and construct a new anomaly discrimination boundary model based on the local feature subspace. The output contains the boundary parameter set of the boundary model for all local anomaly detection methods. The boundary parameter set includes the sample mean vector corresponding to each subspace. Covariance matrix and anomaly detection threshold .

6. A construction safety monitoring system based on Internet of Things (IoT) technology according to claim 2, characterized in that, The output of the preliminary risk identification results includes: Collect the low-dimensional feature vector of the current period and construct a feature time buffer by combining it with the historical low-dimensional feature vectors within the previous fixed number of periods, and arrange them in chronological order to form a feature sliding window sequence; In the feature sliding window sequence, a moving average filtering process is performed on each low-dimensional feature vector to be judged. The low-dimensional feature vector to be judged is smoothed by taking values ​​symmetrically before and after, suppressing short-period fluctuations and abnormal peak interference, and outputting a smooth feature vector after noise suppression. Call the boundary parameter set, which includes multiple local anomaly detection boundary models. Each local anomaly detection boundary model contains a local sample mean vector, a covariance matrix, and an anomaly detection threshold. For the current smooth feature vector, the statistical deviation in each local feature subspace is calculated in all local anomaly discrimination boundary models, and the local feature subspace with the smallest deviation is determined as the belonging subspace; Within the subspace, based on the corresponding local anomaly discrimination boundary model, the relationship between the current smooth feature vector and the anomaly discrimination threshold is compared. If the deviation is greater than the corresponding anomaly discrimination threshold, the smooth feature vector is marked as anomaly; otherwise, it is marked as normal. The labeled status of all low-dimensional feature vectors to be judged is associated with the corresponding timestamp, and the abnormal status set is output periodically to form the preliminary risk identification result for the current period.

7. A construction safety monitoring system based on Internet of Things (IoT) technology according to claim 2, characterized in that, The parameters for updating the feature extraction module and the boundary modeling module include: Receive the preliminary risk identification results and corresponding low-dimensional feature vectors of the current period, extract the low-dimensional feature vectors that are judged to be abnormal, and combine them with the data of the previous period to form an abnormal sample sequence. Determine whether there are multiple consecutive periods in the abnormal sample sequence that belong to the same local feature subspace and whose deviation exceeds the current anomaly detection threshold. If the condition is met, record the parameter update identifier that triggers the local feature subspace. For the local feature subspace that triggers the update flag, collect all low-dimensional feature vector samples that are labeled as normal, and calculate the updated sample mean vector and covariance matrix based on the low-dimensional feature vector sample set, while constructing new sample variance statistics. Based on the parameters of the normal sample set and the cumulative sample count, the mean vector of the feature fields and the set of principal component direction vectors in the feature extraction module are updated, and orthogonal normalization is performed on the updated set of principal component direction vectors to generate the corrected feature space state; Based on the corrected sample mean vector and covariance matrix, the boundary function of the corresponding local anomaly discrimination boundary model is updated, while keeping the original anomaly discrimination threshold unchanged, thus completing the synchronous update of the boundary parameter set.

8. A construction safety monitoring system based on Internet of Things (IoT) technology according to claim 2, characterized in that, The generation of construction risk levels and corresponding early warning and control instructions includes: Collect construction worker behavior data, construction equipment operation data and environmental monitoring data within the current period, extract corresponding personnel activity intensity indicators, equipment operation stability indicators and environmental abnormal change indicators, and construct a multi-source status information set; The preliminary risk identification results output in the current cycle are correlated and matched with the multi-source state information set. Based on the low-dimensional feature vector corresponding to the abnormal event, the local feature subspace to which it belongs, and the joint situation of the current personnel, equipment, and environmental status, the risk impact factor is calculated. Based on the risk impact factors and the risk level classification rules, each abnormal event is assigned to the corresponding construction risk level label. Based on the construction risk level label and the corresponding subspace identifier of the abnormal event, the early warning control template that meets the risk level, event type and subspace constraints is matched from the preset early warning response rule set, and the control template information containing response action, response range and trigger condition parameters is extracted. Based on the current construction scheduling plan and the availability status of construction equipment, structured early warning control instructions are generated according to the extracted control template information. The structured early warning control instructions include response action definitions, target equipment identifiers, and trigger condition parameters. Output the construction risk level label and structured early warning control instructions corresponding to each abnormal event in the current cycle.

Images