Construction site dangerous behavior identification method based on edge calculation

Abstract

The invention discloses a construction site dangerous behavior identification method based on edge calculation, and the method comprises the following steps: collecting construction site multi-source safety perception data, and carrying out the preprocessing of the data to generate a construction site multi-source time series data set; extracting personnel motion features, equipment operation features, operation area constraint features and interaction features of personnel and equipment, constructing behavior state vectors, and generating a behavior state sequence; calculating a security distribution interval, and generating a security state set, a security domain boundary parameter and a security domain discrimination rule; performing forward prediction and reverse reconstruction to generate an irreversibility index; constructing a state association diagram, and detecting mutation points of the structural feature parameters; and performing fusion judgment based on the security state set, the security domain judgment rule, the irreversibility index and the abrupt change point, and outputting a corresponding risk subject, a position and a time interval. The method is based on edge calculation and behavior evolution judgment, construction site dangerous behavior formation precursor recognition is achieved, and the method has the advantages of early warning in advance, weak network adaptation and low label dependence.

Description

Technical Field

[0001] This invention relates to the field of construction site hazardous behavior identification, and more particularly to a construction site hazardous behavior identification method based on edge computing. Background Technology

[0002] Existing construction site hazardous behavior identification technologies mainly rely on video surveillance, sensor data, or wearable devices to detect personnel violations, abnormal equipment conditions, or environmental exceedances. Common methods include safety monitoring based on rule thresholds, behavior classification models based on labeled samples, and anomaly detection methods for single objects or single indicators. These technologies typically perform calculations in the cloud or on central servers, are highly dependent on communication networks and computing resources, and often focus on "danger has occurred" as the identification target, emphasizing post-event identification or immediate alarm.

[0003] Under the actual conditions of weak or disconnected networks and limited edge computing power at construction sites, the above-mentioned technologies are difficult to operate stably. At the same time, their reliance on a large number of labeled hazardous samples or pre-enumerated rules of hazardous actions makes them unsuitable for complex and ever-changing construction scenarios. More importantly, most existing technologies only judge safety from a static state or a single point in time, lacking a holistic model of the evolution of personnel-equipment-work environment behavior. This makes it difficult to identify the precursors of hazardous behavior evolving from a normal state to an unstable state, resulting in delayed warnings and a high false alarm rate, which fails to meet the actual needs of construction sites for early and reliable risk warnings. Summary of the Invention

[0004] One objective of this invention is to propose a method for identifying hazardous behaviors at construction sites based on edge computing. This invention, based on edge computing and behavior evolution determination, enables the identification of precursors to hazardous behaviors at construction sites, and has the advantages of early warning, weak network adaptation, and low labeling dependence.

[0005] A method for identifying hazardous behaviors at construction sites based on edge computing, according to an embodiment of the present invention, includes the following steps: Collect multi-source safety perception data at the construction site and preprocess it to generate a multi-source time-series dataset of the construction site. On edge computing devices, based on multi-source time-series datasets from construction sites, personnel movement features, equipment operation features, work area constraint features, and personnel-equipment interaction features are extracted according to sliding time windows to construct behavioral state vectors and generate behavioral state sequences in chronological order. Select the behavior state sequence of the normal construction period as the normal construction period behavior state sequence from the behavior state sequence, calculate the safety distribution interval based on the sequence, generate the safety state set and safety domain boundary parameters, and generate safety domain discrimination rules based on the safety domain boundary parameters; The behavior state sequence for the real-time running period is obtained from the behavior state sequence as the real-time behavior state sequence. Under the constraint of the security domain discrimination rule, based on the security state set, forward prediction processing and reverse reconstruction processing are performed on the real-time behavior state sequence to generate an irreversible index. Based on real-time behavioral state sequences and irreversibility indicators, a state association graph is constructed within each sliding time window, the corresponding structural feature parameters are calculated, and abrupt changes in the structural feature parameters are detected. Based on the fusion judgment of the set of safe states, the rules for judging safe domains, the irreversibility index and the mutation point, the risk level and early warning result of the precursor of dangerous behavior are generated, and the corresponding risk subject, location and time interval are output.

[0006] Optionally, the multi-source safety perception data at the construction site specifically includes: location information, movement trajectory information, and movement status information of construction personnel; operating status information, movement status information, and working parameter information of construction equipment; spatial constraint information and regional status information of the work area; and environmental parameter information of the construction site. The preprocessing specifically includes: time alignment, abnormal data removal, missing data filling, noise suppression, unit unification, numerical standardization, and time-based combination.

[0007] Optionally, the generation of the behavioral state sequence specifically includes: The multi-source time-series dataset from the construction site is segmented and processed according to a preset sliding time window on the edge computing device; For the data within each sliding time window, the motion characteristics of the construction workers are extracted. These characteristics include changes in the workers' displacement, speed, and state. For data within the same sliding time window, the operating characteristics of the construction equipment are extracted. These operating characteristics include the operating status of the equipment, changes in operating parameters, and the motion state of the equipment. For data within the same sliding time window, extract work area constraint features, which include the spatial relationship between personnel or equipment and the boundary of the work area, as well as the work area constraint state generated based on the spatial relationship. For data within the same sliding time window, the interaction features between people and equipment are extracted. These interaction features include changes in the spatial distance between people and equipment, as well as changes in the interaction state. The characteristics of personnel movement, equipment operation, work area constraints, and personnel-equipment interaction are combined to generate the behavior state vector of the sliding time window. The behavior state vectors corresponding to each sliding time window are arranged in chronological order to generate a behavior state sequence.

[0008] Optionally, the generation of the security domain discrimination rule specifically includes: Select behavioral state vectors from the behavioral state sequence according to the normal construction time period, and arrange them in chronological order to generate the normal construction period behavioral state sequence. Based on the behavioral state sequence during normal construction, the values ​​of each behavioral state vector in the sequence are statistically processed in each feature dimension to generate the safe distribution range of each feature dimension under normal construction conditions. From the normal construction period behavior state sequence, select behavior state vectors whose feature dimensions all fall within the safe distribution range, and aggregate them to generate a safe state set; The range of values ​​for each behavioral state vector in the set of security states on each feature dimension is summarized and processed. The boundary values ​​within the preset confidence interval of each feature dimension are calculated. The boundary values ​​are combined to generate the security domain boundary parameters corresponding to each feature dimension. Based on the safety domain boundary parameters, the values ​​of the behavior state vector in each feature dimension are compared one by one to construct a set of discrimination conditions for determining whether the behavior state vector falls within the range of the safety domain boundary parameters, and thus generate safety domain discrimination rules.

[0009] Optionally, the generation of the irreversibility index specifically includes: Based on the behavior state sequence, a real-time behavior state sequence is generated by filtering according to the time range corresponding to the real-time running time period. Under the constraints of the security domain discrimination rules, based on the security state set, the real-time behavior state sequence is processed time-by-time to generate a processing behavior state sequence that meets the unified security domain discrimination conditions; Based on the temporally adjacent behavioral state vectors in the processing behavioral state sequence, forward prediction processing is performed to generate a predicted behavioral state vector corresponding to the next time step. Based on the temporally adjacent behavioral state vectors in the processing behavioral state sequence, reverse reconstruction processing is performed to generate a reconstructed behavioral state vector corresponding to the previous moment. Based on the differences between the predicted behavior state vector and the corresponding actual behavior state vector, as well as the differences between the reconstructed behavior state vector and the corresponding actual behavior state vector, the changes in the processing behavior state sequence in the direction of temporal evolution are comprehensively calculated to generate an irreversibility index corresponding to the current sliding time window.

[0010] Optionally, the generation of the mutation point specifically includes: Within each sliding time window, based on the real-time behavior state sequence and the irreversibility index, each behavior state vector in the real-time behavior state sequence is indexed and identified to generate a set of state nodes within the sliding time window. Based on the temporal order relationship and irreversibility index between the behavioral state vectors in the state node set, the association relationship between state nodes is generated, and a state association graph corresponding to the sliding time window is constructed. Based on the distribution of state nodes and relationships in the state association graph, the state association graph is traversed and calculated to generate structural feature parameters. Within a continuous sliding time window, the structural feature parameters generated for each sliding time window are arranged in chronological order to generate a sequence of structural feature parameters; Based on the changes in structural feature parameters corresponding to adjacent sliding time windows, mutation detection processing is performed on the structural feature parameter sequence to generate mutation points corresponding to the sliding time windows.

[0011] Optionally, the generation of the risk subject, location, and time interval specifically includes: Under the constraint of the set of safe states, a safe domain discrimination rule is applied to the behavior state vector corresponding to the current sliding time window to generate a set of safe domain discrimination results reflecting whether the behavior state falls into the safe domain. The irreversibility index is subjected to threshold mapping to generate the irreversibility judgment result corresponding to the current sliding time window; Within the current sliding time window, mutation points are recorded and marked, and mutation discrimination results corresponding to the current sliding time window are generated; The results of the security domain discrimination, irreversibility discrimination, and mutation discrimination are combined and fusion judgment processing is performed to generate the risk level and warning result corresponding to the current sliding time window; Based on the risk level and early warning results, relevant information is extracted from the corresponding behavioral state vector in the real-time behavioral state sequence to generate the risk subject, risk location, and risk time interval corresponding to the early warning results.

[0012] The beneficial effects of this invention are: This invention constructs a mechanism for identifying precursors of hazardous behaviors, centered on behavioral state sequences, by continuously modeling and analyzing multi-source safety perception data from construction sites on edge computing devices. This transforms hazard identification from traditional "outcome judgment" to "process judgment." Compared to existing methods that rely on rule thresholds or labeled hazardous samples, this invention uses normal construction behavior as a reference, characterizes the boundaries of normal behavior through safety domain discrimination rules, introduces an irreversibility index combining forward prediction and reverse reconstruction to describe the direction of behavioral evolution, and further utilizes structural feature mutations in the state association graph to characterize the instability process of the personnel-equipment-work environment relationship. Thus, it identifies precursors of hazardous behaviors before they become apparent, achieving true early warning.

[0013] The technical solution of this invention fully considers the actual working conditions of weak network, network outage, and limited edge computing power at construction sites. The entire identification and judgment process can be completed on edge computing devices, avoiding dependence on cloud computing and continuous network connection, and improving the stability and real-time performance of the system in complex construction site environments. Since this invention does not rely on a large number of labeled hazardous samples, nor does it require exhaustive enumeration of specific hazardous action rules in advance, it can adapt to different construction scenarios, work processes, and differences in personnel behavior, significantly reducing system deployment and maintenance costs, and improving generalization ability and practical value. By fusing and judging multi-dimensional information such as safety domain state, irreversible behavioral evolution, and structural mutation, this invention effectively reduces the false alarm rate while ensuring timely early warning, providing a more forward-looking, reliable, and engineering-feasible technical means for construction site safety management. Attached Figure Description

[0014] 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:

[0015] Figure 1 This is a flowchart of a construction site hazardous behavior identification method based on edge computing proposed in this invention; Figure 2 This is a schematic diagram illustrating the forward prediction and reverse reconstruction to generate irreversibility indicators in a construction site hazardous behavior identification method based on edge computing proposed in this invention. Figure 3 This is a schematic diagram illustrating the construction state association diagram and generation of structural feature parameters for a construction site hazardous behavior identification method based on edge computing proposed 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 1-3 A method for identifying hazardous behaviors at construction sites based on edge computing includes the following steps: Collect multi-source safety perception data at the construction site and preprocess it to generate a multi-source time-series dataset of the construction site. On edge computing devices, based on multi-source time-series datasets from construction sites, personnel movement features, equipment operation features, work area constraint features, and personnel-equipment interaction features are extracted according to sliding time windows to construct behavioral state vectors and generate behavioral state sequences in chronological order. Select the behavior state sequence of the normal construction period as the normal construction period behavior state sequence from the behavior state sequence, calculate the safety distribution interval based on the sequence, generate the safety state set and safety domain boundary parameters, and generate safety domain discrimination rules based on the safety domain boundary parameters; The behavior state sequence for the real-time running period is obtained from the behavior state sequence as the real-time behavior state sequence. Under the constraint of the security domain discrimination rule, based on the security state set, forward prediction processing and reverse reconstruction processing are performed on the real-time behavior state sequence to generate an irreversible index. Based on real-time behavioral state sequences and irreversibility indicators, a state association graph is constructed within each sliding time window, the corresponding structural feature parameters are calculated, and abrupt changes in the structural feature parameters are detected. Based on the fusion judgment of the set of safe states, the rules for judging safe domains, the irreversibility index and the mutation point, the risk level and early warning result of the precursor of dangerous behavior are generated, and the corresponding risk subject, location and time interval are output.

[0018] Optionally, the multi-source safety perception data at the construction site specifically includes: location information, movement trajectory information, and movement status information of construction personnel; operating status information, movement status information, and working parameter information of construction equipment; spatial constraint information and regional status information of the work area; and environmental parameter information of the construction site. The preprocessing specifically includes: time alignment, abnormal data removal, missing data filling, noise suppression, unit unification, numerical standardization, and time-based combination.

[0019] Optionally, the generation of the behavioral state sequence specifically includes: The multi-source time-series dataset from the construction site is segmented and processed according to a preset sliding time window on the edge computing device; For the data within each sliding time window, the motion characteristics of the construction workers are extracted. These characteristics include changes in the workers' displacement, speed, and state. For data within the same sliding time window, the operating characteristics of the construction equipment are extracted. These operating characteristics include the operating status of the equipment, changes in operating parameters, and the motion state of the equipment. For data within the same sliding time window, extract work area constraint features, which include the spatial relationship between personnel or equipment and the boundary of the work area, as well as the work area constraint state generated based on the spatial relationship. The generation of the work area constraint features specifically includes: Within the current sliding time window, the predefined work area boundary in the construction site is acquired, and the spatial location data of construction personnel and equipment within the corresponding time window are obtained, forming a correspondence between the work area boundary information and the spatial locations of construction personnel and equipment. Based on the correspondence, the spatial relationship between construction personnel or equipment and the work area boundary is calculated, wherein the spatial relationship includes at least the states of being inside or outside the work area and the relative distance from the work area boundary. According to the spatial relationship, within the current sliding time window, it is determined whether the construction personnel or equipment meet the work area constraints, generating a work area constraint determination result. The work area constraint determination result is then represented in a state-based manner within the current sliding time window, generating work area constraint features corresponding to the construction personnel or equipment. For data within the same sliding time window, the interaction features between people and equipment are extracted. These interaction features include changes in the spatial distance between people and equipment, as well as changes in the interaction state. The specific features of the interaction between people and equipment include: Within the current sliding time window, acquire spatial location data of construction personnel and equipment, associate personnel and equipment, and generate a personnel-equipment spatial correspondence corresponding to the sliding time window; based on the personnel-equipment spatial correspondence, calculate the changes in spatial distance between personnel and equipment, and generate a personnel-equipment distance change relationship corresponding to the current sliding time window; based on the personnel-equipment distance change relationship, determine whether personnel and equipment are in an interactive state, and generate a personnel-equipment interaction determination result corresponding to the current sliding time window; represent the personnel-equipment interaction determination result in a stateful manner within the current sliding time window, and generate an interaction state corresponding to personnel and equipment. The characteristics of personnel movement, equipment operation, work area constraints, and personnel-equipment interaction are combined to generate the behavior state vector of the sliding time window. The generation of the behavior state vector specifically includes: Within the current sliding time window, feature alignment processing is performed on the extracted personnel movement features, equipment operation features, work area constraint features, and personnel-equipment interaction features to ensure consistency in time dimension and sampling granularity, generating a multi-source feature set corresponding to the sliding time window. The features in the multi-source feature set are arranged according to a preset feature order, and scale consistency processing is performed on different types of features to generate a feature representation set with consistent feature dimensions. The feature representations in the feature representation set are concatenated according to feature dimensions to form a joint feature representation corresponding to the sliding time window. The joint feature representation is then vectorized to generate the behavior state vector corresponding to the sliding time window. The behavior state vectors corresponding to each sliding time window are arranged in chronological order to generate a behavior state sequence.

[0020] Optionally, the generation of the security domain discrimination rule specifically includes: Select behavioral state vectors from the behavioral state sequence according to the normal construction time period, and arrange them in chronological order to generate the normal construction period behavioral state sequence. Based on the behavioral state sequence during normal construction, the values ​​of each behavioral state vector in the sequence are statistically processed in each feature dimension to generate the safe distribution range of each feature dimension under normal construction conditions. From the normal construction period behavior state sequence, select behavior state vectors whose feature dimensions all fall within the safe distribution range, and aggregate them to generate a safe state set; The range of values ​​for each behavioral state vector in the set of security states on each feature dimension is summarized and processed. The boundary values ​​within the preset confidence interval of each feature dimension are calculated. The boundary values ​​are combined to generate the security domain boundary parameters corresponding to each feature dimension. The generation of the security domain boundary parameters specifically includes: Based on the set of safe states, the value sequences of each behavioral state vector in the set are obtained on each feature dimension to form a set of safe values ​​corresponding to each feature dimension. For each feature dimension's set of safe values, the value distribution is limited within a preset confidence interval to determine the effective value range of that feature dimension within the confidence interval. Within the effective value range, the corresponding boundary values ​​are selected as the safe domain boundary values ​​of that feature dimension. The safe domain boundary values ​​obtained for each feature dimension are aggregated to generate a set of safe domain boundary parameters composed of multiple feature dimension boundary values. Based on the safety domain boundary parameters, the values ​​of the behavior state vector in each feature dimension are compared one by one to construct a set of discrimination conditions for determining whether the behavior state vector falls within the range of the safety domain boundary parameters, and generate safety domain discrimination rules. The generation of the security domain discrimination rules specifically includes: Based on the safety domain boundary parameters, corresponding safety value constraints are determined for each feature dimension. These safety value constraints are used to limit the value range of the behavior state vector in that feature dimension. The safety value constraints are then organized dimension by dimension to form a set of single-dimensional discrimination conditions that correspond one-to-one with the feature dimensions of the behavior state vector. The single-dimensional discrimination condition sets are then combined to generate a set of multi-dimensional discrimination conditions that describe the joint constraint relationship of the behavior state vector in multiple feature dimensions, which is then solidified as a safety domain discrimination rule.

[0021] Optionally, the generation of the irreversibility index specifically includes: Based on the behavior state sequence, a real-time behavior state sequence is generated by filtering according to the time range corresponding to the real-time running time period. Under the constraints of the security domain discrimination rules, based on the security state set, the real-time behavior state sequence is processed time-by-time to generate a processing behavior state sequence that meets the unified security domain discrimination conditions; Based on the temporally adjacent behavioral state vectors in the processing behavioral state sequence, forward prediction processing is performed to generate a predicted behavioral state vector corresponding to the next time step. The generation of the predicted behavior state vector specifically includes: In processing the behavior state sequence, adjacent behavior state vector pairs are selected in chronological order to form a time-series sample pair consisting of the behavior state vector at the current moment and the behavior state vector at the next moment. Using the behavior state vector at the current moment in the time-series sample pair as input, the value of the behavior state vector at the next moment is forward-calculated based on the established time evolution relationship in the processing behavior state sequence to generate the predicted behavior state vector corresponding to the next moment. Based on the temporally adjacent behavioral state vectors in the processing behavioral state sequence, reverse reconstruction processing is performed to generate a reconstructed behavioral state vector corresponding to the previous moment. The generation of the reconstructed behavior state vector specifically includes: In processing the behavior state sequence, the behavior state vectors that are adjacent in time are selected in chronological order to form the state input sequence for reverse reconstruction; based on the state input sequence, the value of the behavior state vector at the previous moment is reversed to generate the reconstructed behavior state vector corresponding to the previous moment. Based on the difference between the predicted behavior state vector and the corresponding actual behavior state vector, as well as the difference between the reconstructed behavior state vector and the corresponding actual behavior state vector, the changes in the processing behavior state sequence in the direction of time evolution are comprehensively calculated to generate an irreversibility index corresponding to the current sliding time window. The generation of the irreversibility index specifically includes: Based on the predicted behavior state vector and its corresponding actual behavior state vector in time sequence, the differences between the two in each feature dimension are calculated to generate the predicted difference result corresponding to the current sliding time window. Based on the reconstructed behavior state vector and its corresponding actual behavior state vector in time sequence, the differences between the two in each feature dimension are calculated to generate the reconstructed difference result corresponding to the current sliding time window. The predicted difference result and the reconstructed difference result are combined in time sequence to form a difference combination result used to characterize the difference in the forward and reverse evolution of the behavior state in time. Based on the difference combination result, the changes in the processing behavior state sequence in the direction of time evolution are comprehensively calculated to generate an irreversibility index corresponding to the current sliding time window.

[0022] Optionally, the generation of the mutation point specifically includes: Within each sliding time window, based on the real-time behavior state sequence and the irreversibility index, each behavior state vector in the real-time behavior state sequence is indexed and identified to generate a set of state nodes within the sliding time window. Based on the temporal order relationship and irreversibility index between the behavioral state vectors in the state node set, the association relationship between state nodes is generated, and a state association graph corresponding to the sliding time window is constructed. Based on the distribution of state nodes and relationships in the state association graph, the state association graph is traversed and calculated to generate structural feature parameters. The generation of the structural feature parameters specifically includes: Within the current sliding time window, all state nodes and their corresponding relationships in the state association graph are acquired, forming a node set and a relationship set for structural analysis. Based on the node set and relationship set, a traversal operation is performed on the state association graph, and the distribution of the number of associations of each state node during the traversal process is statistically analyzed to generate node association distribution parameters. Based on the node association distribution parameters, the degree of concentration of the distribution of relationships among nodes in the state association graph is statistically analyzed to generate connection distribution parameters that characterize the overall connectivity of the state association graph. Based on the node association distribution parameters and connection distribution parameters, the overall structural form of the state association graph is summarized and calculated to generate structural feature parameters corresponding to the current sliding time window. Within a continuous sliding time window, the structural feature parameters generated for each sliding time window are arranged in chronological order to generate a sequence of structural feature parameters; Based on the changes in structural feature parameters corresponding to adjacent sliding time windows, a mutation detection process is performed on the structural feature parameter sequence to generate mutation points corresponding to the sliding time windows; The generation of the mutation point specifically includes: Within a continuous sliding time window, structural feature parameters corresponding to adjacent sliding time windows are selected in chronological order to form adjacent structural feature parameter pairs. Based on the adjacent structural feature parameter pairs, the changes in structural feature parameters between adjacent sliding time windows are calculated to generate structural change results corresponding to adjacent sliding time windows. The structural change results are aggregated in chronological order to form a structural change sequence. Based on the structural change sequence, it is identified whether the changes in structural feature parameters between adjacent sliding time windows meet preset change discrimination conditions, generating abrupt change positions corresponding to the sliding time windows. The abrupt change positions are mapped to the corresponding sliding time windows to generate abrupt change points corresponding to the sliding time windows.

[0023] Optionally, the generation of the risk subject, location, and time interval specifically includes: Under the constraint of the set of safe states, a safe domain discrimination rule is applied to the behavior state vector corresponding to the current sliding time window to generate a set of safe domain discrimination results reflecting whether the behavior state falls into the safe domain. The irreversibility index is subjected to threshold mapping to generate the irreversibility judgment result corresponding to the current sliding time window; Within the current sliding time window, mutation points are recorded and marked, and mutation discrimination results corresponding to the current sliding time window are generated; The results of the security domain discrimination, irreversibility discrimination, and mutation discrimination are combined and fusion judgment processing is performed to generate the risk level and warning result corresponding to the current sliding time window; The generation of the risk level and early warning results specifically includes: For the current sliding time window, the corresponding set of security domain discrimination results, irreversibility discrimination results, and mutation discrimination results are obtained to form a multi-criteria input set for fusion judgment. Based on the multi-criteria input set, consistency analysis is performed on the security domain discrimination results, irreversibility discrimination results, and mutation discrimination results to generate an intermediate fusion judgment result that characterizes the comprehensive risk features of the behavior state within the current sliding time window. According to the intermediate fusion judgment result, the behavior state within the current sliding time window is graded to generate a risk level and warning result corresponding to the sliding time window. Based on the risk level and early warning results, relevant information is extracted from the corresponding behavioral state vector in the real-time behavioral state sequence to generate the risk subject, risk location, and risk time interval corresponding to the early warning results.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a construction site of a large urban complex. This site presented complex conditions including high-altitude operations, hoisting and lifting, edge work, and multiple trades operating simultaneously. The number of workers was large, the types of equipment were diverse, and the spatial relationships between personnel, equipment, and the work area constantly changed over time. Furthermore, the on-site network conditions were unstable, with localized weak networks and even brief outages. Previous safety monitoring methods relying on video recognition or fixed rules were prone to problems in this environment, such as recognition delays, frequent false alarms, or inability to operate continuously during network anomalies. These methods struggled to detect early signs of dangerous behavior evolving from a normal state to an unstable one.

[0025] At the construction site, edge computing devices are deployed near key work areas and connected to multi-source safety sensing terminals to continuously collect the location, movement trajectory, and movement status information of construction personnel, the operating status, movement status, and working parameters of construction equipment, as well as the spatial constraints and environmental parameters of the work area. The collected multi-source safety sensing data is first processed on the edge computing devices through time alignment, abnormal data removal, missing data imputation, noise suppression, and numerical normalization to form a unified multi-source time-series dataset for the construction site. Using a sliding time window approach, this time-series dataset is continuously segmented. Within each time window, the movement changes of construction personnel, the changes in equipment operating status, the spatial relationship between personnel or equipment and the boundary of the work area, and the changes in the interaction status between personnel and equipment are extracted. These features are then combined to construct a behavioral state vector, forming a continuous behavioral state sequence in chronological order.

[0026] In the initial stable operation phase of the construction site, the corresponding normal construction time period is selected from the behavior state sequence as the normal construction period behavior state sequence. The feature distribution of each behavior state vector in the sequence is statistically analyzed to generate a safe distribution interval reflecting the range of behavior fluctuations under normal construction conditions. Based on this, a safe state set and safe domain boundary parameters are formed. By comparing the safe domain boundary parameters dimension by dimension, a safe domain discrimination rule is constructed, enabling the edge computing device to determine whether the real-time behavior state is still within the range of normal construction behavior during subsequent operation. As construction continues, the edge computing device selects the behavior state vector corresponding to the current operating time period from the behavior state sequence in real time to form a real-time behavior state sequence. Under the constraint of the safe domain discrimination rule, the real-time behavior state sequence is subjected to forward prediction processing and backward reconstruction processing based on the safe state set to obtain an irreversibility index describing the characteristics of the temporal evolution direction of the behavior state.

[0027] As construction intensity increases and working conditions change, edge computing devices map each behavior state vector to a state node within each sliding time window based on real-time behavior state sequences and irreversibility indicators. This constructs a state association graph reflecting the temporal sequence and evolutionary relationship of behavior states. The system then traverses and calculates this state association graph to obtain structural feature parameters that characterize the relationship between personnel, equipment, and the working environment. By analyzing the changes in structural feature parameters within continuous time windows, the system can identify the locations where structural features undergo abrupt changes during temporal evolution, thereby determining the time nodes when the behavior relationship transitions from a stable state to an unstable state.

[0028] Based on this, the edge computing device simultaneously acquires the security domain discrimination result, irreversibility discrimination result, and mutation discrimination result within each sliding time window, and fuses the three types of discrimination results to generate corresponding risk levels and warning results. When the system detects that the behavior state has not yet obviously violated the safety rules, but has shown an irreversible evolution trend accompanied by changes in structural relationships, it can be determined that it has entered the pre-dangerous behavior formation stage. Subsequently, the system extracts the behavior state vector association information corresponding to the warning result from the real-time behavior state sequence, and outputs clear risk subjects, risk locations, and risk time intervals, providing intuitive and operable warning information for on-site management personnel.

[0029] In practical applications, this embodiment can operate continuously even under unstable network conditions at construction sites. The edge computing device completes data processing, judgment, and early warning output locally, without relying on cloud computing support. Continuous operation observation across multiple work periods reveals that this invention can provide early warnings before dangerous behaviors develop into obvious violations or accidents, giving on-site management personnel sufficient time to take intervention measures, such as adjusting work processes, reminding relevant personnel, or suspending the operation of relevant equipment. This embodiment demonstrates that this invention can effectively solve the problems of existing technologies that rely heavily on a large number of dangerous samples and fixed rules, have difficulty identifying early warning signs of dangerous behaviors, and struggle to operate stably in weak network environments. It possesses good engineering applicability and promotional value.

[0030] Table 1. Comparison of the overall performance of different construction site hazard identification methods in actual construction scenarios.

[0031] As shown in Table 1, traditional rule-based hazardous behavior identification methods still have a certain degree of effectiveness in construction site applications, with an accuracy rate of over 80%. However, since these methods mainly rely on pre-set thresholds and rules, the rules are difficult to adapt in a timely manner when the operating methods of construction personnel, the operating rhythm of equipment, or the conditions of the work area change, which limits the identification accuracy and early warning capabilities. Identification methods based on labeled samples have improved accuracy compared to rule-based methods, but their performance is highly dependent on the completeness and representativeness of the training samples. When the sample coverage is insufficient, it is still difficult to stably identify newly emerging hazardous evolution processes.

[0032] The method of this invention achieves an accuracy of 90.1% in identifying dangerous behaviors, which is a steady improvement compared to the two traditional methods. This improvement does not come from more complex models or larger training data, but from the continuous modeling of behavioral state sequences, which enables the system to understand the comprehensive state of personnel, equipment and working environment in a more complete time dimension, thereby reducing misjudgments caused by abnormalities at a single moment or short-term fluctuations.

[0033] In terms of early warning capability, the advantages of the method of this invention are more obvious. Traditional rule-based methods, which focus on whether a threshold is reached, have an early warning success rate of less than half and a short average warning lead time. Although sample-based methods can identify some dangerous trends, their identification targets are still biased towards known dangerous patterns, and the lead time is limited. This invention introduces irreversibility indicators and structural feature change analysis, enabling the system to identify its evolution trend before the behavior state has obviously crossed the boundary. As a result, it has achieved significant improvements in both early warning success rate and early warning lead time, giving on-site managers more time to respond.

[0034] In terms of false alarm rate, the method of this invention is significantly lower than the two traditional methods. Rule-based methods are prone to triggering alarms frequently due to local state fluctuations in complex operating conditions, while sample-based methods also have the risk of false alarms when facing uncovered scenarios. This invention only triggers an alarm when the safety domain state, behavioral evolution direction, and structural changes all show abnormal trends simultaneously, effectively suppressing false alarms caused by abnormal single indicators and making the overall alarm results more stable and reliable.

[0035] From an engineering application perspective, the present invention demonstrates higher continuous operational stability and real-time processing success rate of single edge devices under weak network conditions compared to traditional methods. This indicates that deploying the main analysis process on edge computing devices reduces reliance on network connectivity and centralized computing power, making it more suitable for long-term operational needs at construction sites. Overall, the present invention achieves earlier warnings, lower false alarms, and higher operational stability while ensuring identification accuracy. Its performance improvement stems from the systematic analysis of behavioral evolution and the fusion of multi-dimensional criteria, rather than a simple aggregation of the number of rules or sample size, demonstrating significant engineering practical value.

[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 method for identifying hazardous behaviors on construction sites based on edge computing, characterized in that, Includes the following steps: Collect multi-source safety perception data at the construction site and preprocess it to generate a multi-source time-series dataset of the construction site. On edge computing devices, based on multi-source time-series datasets from construction sites, personnel movement features, equipment operation features, work area constraint features, and personnel-equipment interaction features are extracted according to sliding time windows to construct behavioral state vectors and generate behavioral state sequences in chronological order. Select the behavior state sequence of the normal construction period as the normal construction period behavior state sequence from the behavior state sequence, calculate the safety distribution interval based on the sequence, generate the safety state set and safety domain boundary parameters, and generate safety domain discrimination rules based on the safety domain boundary parameters; The behavior state sequence for the real-time running period is obtained from the behavior state sequence as the real-time behavior state sequence. Under the constraint of the security domain discrimination rule, based on the security state set, forward prediction processing and reverse reconstruction processing are performed on the real-time behavior state sequence to generate an irreversible index. Based on real-time behavioral state sequences and irreversibility indicators, a state association graph is constructed within each sliding time window, the corresponding structural feature parameters are calculated, and abrupt changes in the structural feature parameters are detected. Based on the fusion judgment of the set of safe states, the rules for judging safe domains, the irreversibility index and the mutation point, the risk level and early warning result of the precursor of dangerous behavior are generated, and the corresponding risk subject, location and time interval are output.

2. The method for identifying hazardous behaviors at construction sites based on edge computing according to claim 1, characterized in that, The multi-source safety perception data at the construction site specifically includes: location information, movement trajectory information, and movement status information of construction personnel; operating status information, movement status information, and working parameter information of construction equipment; spatial constraint information and regional status information of the work area; and environmental parameter information of the construction site. The preprocessing specifically includes: time alignment, abnormal data removal, missing data filling, noise suppression, unit unification, numerical standardization, and time-based combination.

3. The method for identifying hazardous behaviors at construction sites based on edge computing according to claim 1, characterized in that, The generation of the behavioral state sequence specifically includes: The multi-source time-series dataset from the construction site is segmented and processed according to a preset sliding time window on the edge computing device; For the data within each sliding time window, the motion characteristics of the construction workers are extracted. These characteristics include changes in the workers' displacement, speed, and state. For data within the same sliding time window, the operating characteristics of the construction equipment are extracted. These operating characteristics include the operating status of the equipment, changes in operating parameters, and the motion state of the equipment. For data within the same sliding time window, extract work area constraint features, which include the spatial relationship between personnel or equipment and the boundary of the work area, as well as the work area constraint state generated based on the spatial relationship. For data within the same sliding time window, the interaction features between people and equipment are extracted. These interaction features include changes in the spatial distance between people and equipment, as well as changes in the interaction state. The characteristics of personnel movement, equipment operation, work area constraints, and personnel-equipment interaction are combined to generate the behavior state vector of the sliding time window. The behavior state vectors corresponding to each sliding time window are arranged in chronological order to generate a behavior state sequence.

4. The method for identifying hazardous behaviors at construction sites based on edge computing according to claim 1, characterized in that, The generation of the security domain discrimination rules specifically includes: Select behavioral state vectors from the behavioral state sequence according to the normal construction time period, and arrange them in chronological order to generate the normal construction period behavioral state sequence. Based on the behavioral state sequence during normal construction, the values ​​of each behavioral state vector in the sequence are statistically processed in each feature dimension to generate the safe distribution range of each feature dimension under normal construction conditions. From the normal construction period behavior state sequence, select behavior state vectors whose feature dimensions all fall within the safe distribution range, and aggregate them to generate a safe state set; The range of values ​​for each behavioral state vector in the set of security states on each feature dimension is summarized and processed. The boundary values ​​within the preset confidence interval of each feature dimension are calculated. The boundary values ​​are combined to generate the security domain boundary parameters corresponding to each feature dimension. Based on the safety domain boundary parameters, the values ​​of the behavior state vector in each feature dimension are compared one by one to construct a set of discrimination conditions for determining whether the behavior state vector falls within the range of the safety domain boundary parameters, and thus generate safety domain discrimination rules.

5. The method for identifying hazardous behaviors at construction sites based on edge computing according to claim 1, characterized in that, The generation of the irreversibility index specifically includes: Based on the behavior state sequence, a real-time behavior state sequence is generated by filtering according to the time range corresponding to the real-time running time period. Under the constraints of the security domain discrimination rules, based on the security state set, the real-time behavior state sequence is processed time-by-time to generate a processing behavior state sequence that meets the unified security domain discrimination conditions; Based on the temporally adjacent behavioral state vectors in the processing behavioral state sequence, forward prediction processing is performed to generate a predicted behavioral state vector corresponding to the next time step. Based on the temporally adjacent behavioral state vectors in the processing behavioral state sequence, reverse reconstruction processing is performed to generate a reconstructed behavioral state vector corresponding to the previous moment. Based on the differences between the predicted behavior state vector and the corresponding actual behavior state vector, as well as the differences between the reconstructed behavior state vector and the corresponding actual behavior state vector, the changes in the processing behavior state sequence in the direction of temporal evolution are comprehensively calculated to generate an irreversibility index corresponding to the current sliding time window.

6. The method for identifying hazardous behaviors at construction sites based on edge computing according to claim 1, characterized in that, The generation of the mutation point specifically includes: Within each sliding time window, based on the real-time behavior state sequence and the irreversibility index, each behavior state vector in the real-time behavior state sequence is indexed and identified to generate a set of state nodes within the sliding time window. Based on the temporal order relationship and irreversibility index between the behavioral state vectors in the state node set, the association relationship between state nodes is generated, and a state association graph corresponding to the sliding time window is constructed. Based on the distribution of state nodes and relationships in the state association graph, the state association graph is traversed and calculated to generate structural feature parameters. Within a continuous sliding time window, the structural feature parameters generated for each sliding time window are arranged in chronological order to generate a sequence of structural feature parameters; Based on the changes in structural feature parameters corresponding to adjacent sliding time windows, mutation detection processing is performed on the structural feature parameter sequence to generate mutation points corresponding to the sliding time windows.

7. The method for identifying hazardous behaviors at construction sites based on edge computing according to claim 1, characterized in that, The generation of the risk subject, location, and time interval specifically includes: Under the constraint of the set of safe states, a safe domain discrimination rule is applied to the behavior state vector corresponding to the current sliding time window to generate a set of safe domain discrimination results reflecting whether the behavior state falls into the safe domain. The irreversibility index is subjected to threshold mapping to generate the irreversibility judgment result corresponding to the current sliding time window; Within the current sliding time window, mutation points are recorded and marked, and mutation discrimination results corresponding to the current sliding time window are generated; The results of the security domain discrimination, irreversibility discrimination, and mutation discrimination are combined and fusion judgment processing is performed to generate the risk level and warning result corresponding to the current sliding time window; Based on the risk level and early warning results, relevant information is extracted from the corresponding behavioral state vector in the real-time behavioral state sequence to generate the risk subject, risk location, and risk time interval corresponding to the early warning results.

Images