Expressway construction area risk situation perception method based on environment perception

Abstract

The invention relates to a highway construction area risk situation perception method based on environmental perception, and the method specifically comprises the following steps: deploying a multi-source sensing equipment network along a highway construction area and at a key structure position to collect data, and marking the data; constructing a highway construction area risk situation awareness model, training the model, and jointly optimizing all trainable parameters of the model through back propagation to obtain a trained model; and deploying the trained model to a construction area edge computing platform, outputting a final prediction result through the trained model according to data acquired in real time, judging whether to perform early warning or not according to the prediction result, calibrating a risk position and a risk level, and continuously screening a high-confidence sample through a pseudo-tag self-enhancement module to perform fine adjustment on the model. According to the method, the model is constructed and trained, so that the construction risk can be predicted more quickly and more accurately, workers are timely guided to carry out emergency disposal, and the safety index in the construction process is improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and road construction safety technology, and in particular to a method for risk situation perception in highway construction areas based on environmental perception. Background Technology

[0002] Highway construction zones are an unavoidable part of transportation infrastructure construction. These zones are characterized by complex working environments, dense populations, numerous equipment, and are often accompanied by multiple uncertainties such as traffic flow and the natural environment, making them highly susceptible to safety risks. For example, heavy machinery construction may cause strong vibrations, insufficient slope support may lead to abnormal displacement, sudden heavy rainfall may trigger landslides, and concentrated activity of construction workers may cause traffic congestion or safety accidents. With the development of intelligent sensing and Internet of Things (IoT) technologies, an increasing number of sensing devices are being deployed in construction zones to collect multi-dimensional data such as vibration, displacement, weather, and pedestrian flow.

[0003] Existing technologies have the following shortcomings in practical applications: Existing technologies often rely on simple interpolation or static alignment methods, which easily disrupt spatiotemporal correlations, leading to feature distortion and decreased risk identification accuracy; existing technologies mostly use fixed topology graph modeling, which is difficult to reflect the dynamic relationship between devices, easily causing invalid connections or missing key risk propagation paths; existing graph convolution or temporal modeling methods are not sensitive enough to subtle anomalies and do not explicitly decouple risk factors, resulting in feature confusion and high false positive rates; existing technologies typically use static loss functions or single online learning strategies, which cannot simultaneously consider spatiotemporal sparsity and the scarcity of high-risk samples, easily leading to model performance degradation and missed risk detection.

[0004] Therefore, this invention proposes a risk situation perception method for highway construction areas based on environmental perception to solve the above problems. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by developing a risk situation perception method for highway construction areas based on environmental perception. By constructing and training a model, this invention enables faster and more accurate prediction of construction risks, and timely guidance for workers to take emergency measures, thereby improving the safety index during the construction process.

[0006] The technical solution of this invention to solve the technical problem is a risk situation perception method for highway construction areas based on environmental perception, comprising the following steps: S1. Deploy a network of multi-source sensors along the highway construction area and at key structural locations to collect data throughout the construction cycle, construct a training set, and label the data. S2. Construct and train a risk situation awareness model for highway construction areas. Sequentially construct a spatiotemporal alignment and feature enhancement module, a dynamic adjacency matrix construction module, a gated spatiotemporal convolution module, a risk feature decoupling module, an adaptive dynamic loss weight calculation module, a multi-granularity risk classification head module, and a pseudo-label self-enhancement module. Optimize all trainable parameters of the model through backpropagation to obtain the trained model. S3. Deploy the trained model to the edge computing platform of the construction area. The data collected in real time is processed by the trained model to output the final prediction results. Based on the prediction results, it is determined whether to issue an early warning, as well as to mark the risk location and level. At the same time, the model is fine-tuned by continuously screening high-confidence samples through the pseudo-label self-enhancement module. S1 is as follows: A network of multi-source sensors was deployed along the highway construction area and at key structural locations to collect data from the entire construction cycle, build a training set, and label the data. The multi-source sensor network includes vibration sensors, displacement sensors, weather sensors, and people flow monitoring equipment; Vibration sensors collect vibration amplitudes of the roadbed and structure at a high frequency sampling rate; displacement sensors monitor the displacement of slopes and retaining walls at a medium frequency sampling rate; meteorological sensors record temperature, humidity, wind speed, and rainfall at a low frequency sampling rate; and crowd monitoring equipment captures the activity density of construction workers asynchronously through infrared recognition. The training dataset covers the entire construction cycle. It is stored as a multi-source raw data vector in fixed time window slices. Each sample contains a timestamp, equipment spatial coordinates and sensor values. The collected data is labeled based on construction logs, video surveillance, and expert evaluation. Equipment-level local risk labels and regional-level global risk labels are labeled for each time window. Equipment-level local risk labels include normal, displacement exceeding limits, abnormal vibration, and excessive pedestrian flow. Regional-level global risk labels include normal, equipment-level risk, environmental-level risk, and comprehensive risk.

[0007] The training process of the risk situation awareness model for the construction area of ​​the S2 expressway is as follows: The spatiotemporal alignment and feature enhancement module performs feature-level alignment and local physical correlation feature enhancement on multi-source data in the training set to obtain the spatiotemporal alignment features of each sensing device in a specific time window. A dynamic adjacency matrix is ​​constructed using a dynamic adjacency matrix construction module, and the connection relationship between sensing devices is determined based on spatiotemporal alignment characteristics. The spatiotemporal aligned features are input into the gated spatiotemporal convolution module to generate gated signals and spatiotemporal features layer by layer; The risk feature decoupling module guides the model to decompose spatiotemporal features into independent factors according to risk categories through the feature decoupling loss function, and projects the spatiotemporal features onto the decoupling vectors of each independent factor to generate a decoupling feature matrix; The adaptive dynamic loss weight calculation module calculates feature entropy by decoupling the feature matrix, and then combines the maximum predicted probability of the model, the number of occurrences of historical high-risk events and the total number of samples to adjust the weight of the collected data in the loss function calculation, thereby generating dynamic loss weights in each time window. The multi-granularity risk classification head module processes device-level local risks and region-level global risks respectively, and obtains the local risk probability and region-level global risk probability of each sensing device. The pseudo-label self-enhancement module filters reliable samples by resetting the reliability threshold twice, generates pseudo-labels and assigns them dynamic loss weights for online learning, and then calculates the total loss of the model. The model updates its parameters using the total loss as the basis for parameter updates, and performs multiple rounds of iterative updates until the maximum number of training rounds is reached, resulting in a well-trained model.

[0008] The spatiotemporal alignment and feature enhancement module is as follows: For each sensing device within a specific time window, the data collected by all sensing devices in its spatial neighborhood are aggregated, spatial distance information is fused through location encoding, and the spatiotemporal alignment features of each sensing device within the specific time window are output based on one-dimensional convolution and timestamp offset.

[0009] The dynamic adjacency matrix construction module is as follows: Establish effective connections between physically adjacent and behaviorally related sensing devices. Specifically, for any two sensing devices, calculate the normalized product of spatial distance decay similarity and temporal feature similarity based on spatiotemporal alignment features, combine spatial distance constraints to generate binary adjacency matrix elements, and then output a dynamic adjacency matrix.

[0010] The gated spatiotemporal convolution module is as follows: The gated spatiotemporal convolution module employs multi-scale gated convolutional units. Spatiotemporal aligned features are used as the spatiotemporal features input to the first layer of the multi-scale gated convolutional unit. In each layer of the multi-scale gated convolutional unit, the spatiotemporal features of the current moment are concatenated with the features of the historical moments, and convolution is performed through a gated convolutional kernel. Then, a gated signal is generated by the Sigmoid activation function. The modulated gated signal of the current layer is added to the spatiotemporal features convolved with the spatiotemporal graph, and the updated spatiotemporal features are output through layer normalization.

[0011] The risk feature decoupling module is as follows: The feature decoupling loss function calculates the difference between the expectation and conditional expectation of the inner product of the decoupling vector and the spatiotemporal feature, and calculates the sum of the absolute values ​​of the inner products of each pair of decoupling vectors as an orthogonal constraint term, and then guides the model to decompose the risk feature into independent factors.

[0012] The adaptive dynamic loss weight calculation module is as follows: The adaptive dynamic loss weight calculation module calculates the feature entropy by decoupling the feature matrix, and then combines the maximum predicted probability of the model, the number of occurrences of historical high-risk events and the total number of samples to calculate the classification difficulty coefficient, risk density term and feature uncertainty term, and then weighted and fused to output the dynamic loss weight; The feature entropy of the decoupled feature matrix is ​​obtained by averaging the information entropy of the feature vectors of each sensing device.

[0013] The multi-granularity risk classification header module is as follows: The decoupled feature vector of each sensing device is passed through a shared projection layer and nonlinear activation, and then input into a local classifier to output a device-level local risk probability vector. It aggregates the decoupled feature vectors of all sensing devices, captures cross-device temporal dependencies through a long short-term memory network, and outputs a regional-level global risk probability vector after extracting key features through max pooling.

[0014] The pseudo-tag self-enhancing module is as follows: Based on the global prediction confidence threshold and the KL divergence threshold between the local risk probability distribution and the uniform distribution, high-confidence samples are screened, and a batch set of high-confidence samples is output. Pseudo-labels are generated for high-confidence samples, and online loss is calculated by combining dynamic loss weights. The total loss function is then updated to optimize the model's performance on the high-confidence pseudo-labels.

[0015] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention employs a spatiotemporal alignment and feature enhancement module, utilizing spatial neighborhood aggregation and location encoding fusion instead of conventional interpolation methods to effectively address the issues of asynchronous sampling frequencies and missing data from multi-source sensors, preserving physical resonance characteristics and enhancing the sensitivity of risk identification. This invention proposes a dynamic adjacency matrix construction mechanism, combining spatial distance attenuation and temporal feature similarity to dynamically update the equipment topology, enabling it to adapt to changes in equipment additions, malfunctions, or relocations at the construction site. This invention uses a gated spatiotemporal convolution and risk feature decoupling module, amplifying risk pulse signals through a gating mechanism and explicitly separating independent factors such as displacement, vibration, weather, and pedestrian flow, improving the ability to identify composite risks. This invention employs an adaptive dynamic loss weight and pseudo-label self-enhancement mechanism, enabling continuous online learning and enhanced focus on high-risk samples even under conditions of sample imbalance and dynamic changes in the construction environment. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 This is a comparison chart of the risk identification accuracy of the method of the present invention and existing methods.

[0019] Figure 3 This is a schematic diagram of the original vibration signals from multiple devices.

[0020] Figure 4 This is a schematic diagram of the results of conventional difference alignment.

[0021] Figure 5 This is a schematic diagram of the spatiotemporal alignment result of the method of the present invention.

[0022] Figure 6 This is a schematic diagram showing the layout and connection relationships of the sensing devices.

[0023] Figure 7 This is a schematic diagram of the device behavior similarity matrix.

[0024] Figure 8 This is a schematic diagram of the adjacency matrix using the conventional method.

[0025] Figure 9 This is a schematic diagram of the dynamic adjacency matrix of the method of the present invention.

[0026] Figure 10 This is a schematic diagram showing the change of F1 score between the method of the present invention and existing methods over training time.

[0027] Figure 11This is a risk situation map generated based on the method of the present invention. Detailed Implementation

[0028] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0029] Example 1 A risk situation perception method for highway construction zones based on environmental perception includes the following steps: S1. Deploy a network of multi-source sensors along the highway construction area and at key structural locations to collect data throughout the construction cycle, construct a training set, and label the data. S2. Construct and train a risk situation awareness model for highway construction areas. Sequentially construct a spatiotemporal alignment and feature enhancement module, a dynamic adjacency matrix construction module, a gated spatiotemporal convolution module, a risk feature decoupling module, an adaptive dynamic loss weight calculation module, a multi-granularity risk classification head module, and a pseudo-label self-enhancement module. Optimize all trainable parameters of the model through backpropagation to obtain the trained model. S3. Deploy the trained model to the edge computing platform of the construction area. The data collected in real time is processed by the trained model to output the final prediction results. Based on the prediction results, it is determined whether to issue an early warning, as well as to mark the risk location and level. At the same time, the model is fine-tuned by continuously screening high-confidence samples through the pseudo-label self-enhancement module. In a specific implementation, S1 is as follows: A network of multi-source sensing devices, including vibration sensors, displacement sensors, meteorological sensors, and pedestrian monitoring equipment, will be deployed along the highway construction area and at key structural locations. Vibration sensors collect vibration amplitudes of the roadbed and structure at a high frequency sampling rate; displacement sensors monitor the displacement of slopes and retaining walls at a medium frequency sampling rate; meteorological sensors record temperature, humidity, wind speed, and rainfall at a low frequency sampling rate; and crowd monitoring equipment captures the activity density of construction workers asynchronously through infrared recognition. All devices transmit raw data to the edge computing gateway in real time via IoT nodes.

[0030] The training dataset covers the entire construction cycle and stores multi-source raw data vectors in fixed time window slices. Each sample contains timestamps, equipment spatial coordinates, and sensor values.

[0031] Data annotation is based on construction logs, video surveillance, and expert evaluation. Equipment-level local risk labels (e.g., including four categories: normal, displacement exceeding limits, vibration abnormality, and pedestrian flow exceeding limits) and regional-level global risk labels (e.g., including four categories: normal, equipment-level risk, environmental-level risk, and comprehensive risk) are labeled for each time window, forming a multi-dimensional training sample library with spatiotemporal coordinates, which provides aligned input data and supervision signals for subsequent models.

[0032] In a specific implementation, S2 is as follows: S2.1 Spatiotemporal Alignment and Feature Enhancement Module: Because the data from multiple sensing devices such as vibration, displacement, and meteorological sensors in highway construction areas are asynchronous in time due to different sampling frequencies, and equipment failures can cause local data gaps, conventional interpolation alignment methods easily disrupt the inherent spatiotemporal correlation of the data when filling in missing data or aligning time, leading to distortion of extracted risk features. Therefore, for each sensing device within a specific time window, the raw data from all sensing devices in its spatial neighborhood are aggregated, spatial distance information is fused through location encoding, and spatiotemporal alignment features are output based on one-dimensional convolution and timestamp offset. This achieves feature-level alignment of multi-source data and enhances local physical correlation features, as expressed below: In the formula, Indicates the first The spatiotemporal alignment features of a sensing device in the τ-th time window, with dimension . It includes aligned multi-source features, such as fused features of vibration, displacement, and meteorological features; Let represent the spatial neighborhood set of the i-th sensor device, containing physically adjacent sensor devices. For example, assuming the sensor devices are deployed on a highway, the i-th sensor device... A sensing device is located at coordinates ,but This may include all equipment within a 50-meter radius, for example, a device located in... Another device is located wait; This represents the timestamp offset between the i-th and j-th sensing devices within the τ-th time window. It is achieved by comparing the timestamps of the i-th and j-th sensing devices. The timestamps of the data from each sensing device within the τ-th time window are obtained. For example, the timestamp of the data from the i-th sensing device is... The data timestamp of the j-th sensor is ,but Because the data is asynchronous, the timestamps may differ; this value represents the degree of data asynchrony. To represent a very small constant, to prevent the denominator from being zero, e.g., ; For indexing sensing devices; j is the index of the sensing device that is different from i; Indicates the use of parameters The one-dimensional convolution operation performed; This represents a one-dimensional convolution kernel parameter matrix with dimension 1. , are trainable parameters, typically initialized randomly and then optimized through backpropagation. They are used to extract local features from time series data and fuse spatial information. The input dimension is . The output dimension depends on the feature dimension after concatenation. , is a hyperparameter; The input feature dimension of a one-dimensional convolution is the sum of the dimension of the original data vector and the dimension of the positional encoding. The output feature dimension of a one-dimensional convolution; This represents the raw data vector of the j-th sensor within the τ-th time window, containing multi-source features such as vibration amplitude, displacement, temperature, and humidity. The dimension depends on the type of sensor; for example, if the sensor measures vibration, displacement, temperature, and humidity, then... It is a 4-dimensional vector, representing A vector formed by combining elements; This represents the vector concatenation operator; This represents the position encoding function, and the input of the position encoding function is defined as... ,Right now The implementation of the position encoding function is then expressed as follows: , is a sinusoidal position code, similar to the position code in Transformer, used to represent spatial distance information; The dimension used for position encoding is typically set to an even number, such as, ; The frequency parameter of the first dimension is calculated as follows: ; The frequency parameter is the second dimension, and its calculation method is expressed as follows: ; For the first The frequency parameter of dimension is calculated as follows: ; This represents the spatial coordinate vector of the i-th sensing device; This represents the spatial coordinate vector of the j-th sensor. This represents the L2 norm, which is equivalent to the Euclidean distance calculation method.

[0033] It should be noted that, unlike conventional interpolation alignment methods, the calculation of spatiotemporal alignment features adopts a joint modeling strategy that fuses spatial neighborhood aggregation and location coding. This comprehensively considers the inherent physical relationships between sensing devices and utilizes a location coding function... Spatial distance information is converted into high-frequency feature vectors, and then compared with the original data vectors. After concatenation, the data is processed by one-dimensional convolution. This coupling operation simultaneously achieves spatiotemporal alignment and physical association enhancement in the feature space. When a device malfunctions and data is missing, the spatial distance encoding of neighboring devices can guide the convolution kernel to adaptively reconstruct the spatiotemporal alignment features of that device. This not only avoids the distortion caused by manual interpolation, but also... Item weighting mechanism strengthens timestamp offset Small device contributions allow the aligned features to retain the micro-vibration propagation patterns of the device cluster. Therefore, this invention transforms physical spatial distance into a learnable feature encoding operation, which can amplify the resonance effect features between devices in vibration signal analysis, providing more sensitive input features for subsequent risk identification.

[0034] S2.2, Dynamic Adjacency Matrix Construction Module: The topology of sensing equipment in highway construction areas changes dynamically as the project progresses, such as when equipment fails or is added. A fixed graph structure cannot adapt to these real-time changes, leading to inaccurate relationships between devices and ultimately affecting the accuracy of risk situational awareness. Therefore, for any two sensing devices, the normalized product of spatial distance attenuation similarity and temporal feature similarity is calculated. Combined with spatial distance constraints, a binary adjacency matrix is ​​generated to ensure that effective connections are established only between physically adjacent and behaviorally related devices. A dynamic adjacency matrix is ​​output to represent the device connection status. The specific operations are as follows: definition Let i be the element in the i-th row and j-th column of the dynamic adjacency matrix at time window τ, and let j be the value of τ. Characterized in the first The first dynamic adjacency matrix within the time window The first sensing device and the first The connection status between two sensing devices: if the value is 1, it means that the two devices are spatially adjacent and their behaviors are related; if the value is 0, it means that there is no connection. Then the element in the i-th row and j-th column of the dynamic adjacency matrix in the τ-th time window is... It is 1 if it is true, otherwise it is 0; in, This represents the spatial decay coefficient, which controls the decay rate of distance similarity; a value of 0.1 is preferred. It is a natural exponential function; Indicates the first The spatiotemporal alignment characteristics of a sensing device in the τ-th time window; This represents the normalized dynamic time warp distance value, since the range of the dynamic time warp distance value is... By normalizing and matching the magnitude of the natural exponential function, the calculation method is expressed as follows: ; This represents the L2 norm, which is equivalent to the Euclidean distance calculation method.

[0035] Represents a dynamic time warping function, which measures the similarity of the timing characteristics of two sensing devices; Let represent the variance of the eigenvector of the i-th sensing device within the τ-th time window. The variance of the eigenvalues ​​of each sensing device within the spatiotemporal alignment feature of the time window τ is calculated. Let represent the variance of the eigenvector of the j-th sensing device within the τ-th time window. The variance of the eigenvalues ​​of each sensing device within the spatiotemporal alignment feature of the time window τ is calculated. This is a function to find the maximum value. This represents the similarity threshold, a critical value for determining behavioral relevance, such as... ; This represents a spatial distance threshold, measured in meters, which limits the maximum distance between physical adjacencies. .

[0036] It should be noted that by nonlinearly coupling spatial proximity constraints with behavioral correlations, the limitations of conventional graph structures relying on fixed topologies are overcome. The term is passed through the spatial attenuation coefficient. Controlling physical proximity while normalizing dynamic time-warped distance values It captures the similarity of dynamic behaviors between devices, and combines them with... This mechanism enhances robustness by automatically weakening the connection strength between a device and its neighboring devices when a device generates high variance noise due to a malfunction. This prevents abnormal devices from contaminating the graph structure. For example, transient vibrations caused by the movement of construction machinery can interfere with device correlation judgments, but variance normalization effectively suppresses such transient interference, ensuring the stability of the constructed dynamic adjacency matrix. To more stably represent the real risk transmission path.

[0037] S2.3, Gated Spatiotemporal Convolution Module: Because conventional graph convolutional networks have difficulty distinguishing the subtle differences between normal vibration signals and risk precursors in construction areas, they are not sensitive enough to sudden risk signals and cannot effectively fuse spatiotemporal features, making it difficult to capture subtle signal changes and sudden risks, resulting in poor spatiotemporal feature fusion effects.

[0038] Therefore, this invention employs multi-scale gated convolutional units, which adaptively fuse spatiotemporal features through a gating mechanism to enhance sensitivity to sudden risk signals, effectively integrate spatial and temporal information, and improve risk perception capabilities. The specific steps are as follows: 1) Gating signal generation The spatiotemporal features of the current moment are concatenated with features from historical moments, convolution is performed using a gated convolution kernel, and then a gated signal is generated using a sigmoid activation function. Feature importance weights are learned to modulate subsequent features, as shown below: In the formula, Represents the multi-scale gated convolutional unit. The output gated signal of the layer in the τ-th time window, dimensionality and spatiotemporal characteristics same; This represents the parameter matrix of the gated convolution kernel, used to learn the feature importance weights. It is a trainable parameter that extracts the dependencies between features through convolution operations and outputs a gated signal to modulate the features. This indicates a convolution operator, meaning that a convolution operation is performed on the input. For example, for a one-dimensional convolution, the input is a sequence of features, and the convolution kernel uses a sliding window to calculate the dot product. Represents the multi-scale gated convolutional unit. The spatiotemporal characteristics of the layer's output in the τ-th time window, when hour, Equivalent to the spatiotemporal alignment feature of the sensing device in the τ-th time window ; Let be the spatiotemporal alignment feature of the sensing device in the τ-th time window, which is the spatiotemporal alignment feature of all sensing devices in the τ-th time window. The spatiotemporal alignment feature set of time windows, spatiotemporal alignment features The i-th element is ; Represents the multi-scale gated convolutional unit. The layer in the first The time window outputs spatiotemporal characteristics; The time step represents the offset of the historical time window, and is specifically set according to the sequence length and task requirements, such as... Indicates the previous time window; This represents the channel concatenation operator, which joins two feature matrices along the feature dimension. This indicates the Sigmoid activation function, which compresses the output to... Interval.

[0039] 2) Spatiotemporal feature fusion The spatial graph convolution result is added to the gated and modulated temporal features, and the updated spatiotemporal features are output through layer normalization, fusing spatial and temporal dimensional information, as follows: In the formula, Represents the multi-scale gated convolutional unit. The layer in the first The time window outputs spatiotemporal characteristics; Let the dynamic adjacency matrix of the τ-th time window have dimension 1. , It is a dynamic adjacency matrix The element in the i-th row and j-th column; The number of sensing devices; Represents a spatial convolution kernel, which slides across the spatial dimension during convolution. It is a trainable parameter that extracts spatial features between devices. The size of the spatial convolution kernel, such as, ; This represents the temporal convolution kernel, which slides across the convolutional plane along the time dimension. It consists of trainable parameters and extracts time-series features. For the temporal convolution kernel size, e.g., ; This represents the Hadamard product, which is an element-wise multiplication used for modulating features with gated signals. The representation layer normalization operation stabilizes the feature distribution.

[0040] In its implementation, the multi-scale gated convolutional unit is a neural network module that combines spatial graph convolution and temporal convolution, and adaptively fuses spatiotemporal features through a gating mechanism. It contains multiple layers, such as three layers, each processing features representing the current characteristics. and characterizing historical features Generate gate signal Then, spatial and temporal features are fused to output spatiotemporal features. .

[0041] It should be noted that the gated convolution kernel parameter matrix It can directly learn the dependency patterns between spatiotemporal features. When a strong instantaneous vibration occurs in the construction area, such as when a heavy-load vehicle passes by, it outputs a gating signal. It will significantly enhance the spatiotemporal characteristics of the output. The weighting of mid-to-high frequency components, while suppressing stationary background noise in historical features, makes... This can amplify risk pulse signals, and at the same time, Spatial convolution of terms representation and The temporal convolution of the terms forms a complementary relationship. The former integrates the spatial correlation of the equipment cluster, while the latter strengthens the local temporal anomalies. In slope micro-displacement monitoring, it can simultaneously capture spatial chain reactions and temporal precursors of sudden changes.

[0042] S2.4 Risk Feature Decoupling Module: This invention employs a feature decoupling loss function to guide the model in separating potential risk factors and outputting decoupled feature representations, thereby achieving independent decomposition of risk features and improving the ability to identify complex risks. The specific steps are as follows: 1) Calculation of Feature Decoupling Loss Define a feature decoupling loss function, calculate the difference between the expected and conditional expected values ​​of the inner product of the decoupling vector and the spatiotemporal feature, and calculate the sum of the absolute values ​​of the pairwise inner products of the decoupling vectors as an orthogonality constraint term to guide the model to decompose the risk feature into independent factors, expressed as: In the formula, This represents the feature decoupling loss, which guides the model to decompose risk features into independent factors, such as displacement, vibration, weather, and pedestrian flow, thereby improving the ability to identify complex risks. The number of risk factor categories is determined by domain knowledge; for example, in construction area risks, it is set as follows: These correspond to four types of factors: displacement, vibration, weather, and human flow. Index for risk factor categories; Indicates the first The decoupling vector of the risk factor class, used to extract the corresponding risk features, is a trainable parameter. for transpose; Indicates the first Decoupling vectors of risk factors Indicates the first Decoupling vectors of risk factors To distinguish it from the first risk factor category index of k, To distinguish it from the second risk factor category index of k, , The term represents the sum of the absolute values ​​of the inner products between any two decoupled vectors, and enforces vector orthogonality; Indicates the first The spatiotemporal feature matrix of the time window is the output of the last layer of the multi-scale gated convolutional unit, i.e. ; Represents the multi-scale gated convolutional unit. The layer in the first The time window outputs spatiotemporal characteristics; It is the total number of multi-scale gated convolutional units, such as... ; This represents the expectation operator, which calculates the mean in the training batch; express Expectations of the item Indicates when risk label hour The conditional expectation of an item is calculated by comparing the conditional expectation for each category with the global expectation. The risk label indicates the true risk category of the sample. ; It represents the total number of risk categories, determined by domain knowledge. For example, in a highway construction zone, it might be possible to set up... The categories include normal conditions, equipment-level risks, environmental risks, and comprehensive risks. m is the risk category index, with values ​​ranging from 1 to... The real risk labels in the training data Therefore, when calculating the conditional expectation, the filtering is performed. ; This represents the absolute value operation; This represents the decoupling strength coefficient, which controls the weight of the orthogonal constraint term, such as... .

[0043] 2) Decoupling Feature Generation The spatiotemporal features are projected onto each decoupling vector, and concatenated to form a decoupling feature matrix. The feature intensity of each device under different risk factors is output as follows: In the formula, This represents the decoupling feature matrix, with dimension 1. Characterizes the strength of each device's features under different risk factors; Indicates the first Decoupling vectors of risk factors for transpose, Indicates the first Decoupling vectors of risk factors for transpose, Indicates the first Decoupling vectors of risk factors for transpose; This represents the vector concatenation operator; This indicates the transpose operation.

[0044] It should be noted that, The expected difference of the terms forces the decoupling vector Specific responses are activated when risk events occur. For example, the decoupling vector of a type 1 risk factor characterizes the displacement risk factor, which is significantly activated only during displacement anomalies. Orthogonal constraints ensure the independence of the feature spaces of different risk factors, and joint optimization generates composite risk analysis capabilities. When meteorological risks such as rainstorms trigger geological risks such as slope displacement, the feature matrix is ​​decoupled. It can clearly separate meteorological characteristic peaks from displacement characteristic trends, avoiding feature confusion. Orthogonal constraint terms enable the model to distinguish signals that appear simultaneously but have different sources in the construction machinery vibration interference scenario, significantly reducing the false alarm rate.

[0045] S2.5 Adaptive dynamic loss weight calculation: High-risk events in highway construction areas exhibit spatiotemporal sparsity, and normal state samples constitute the majority in these areas. This leads to model training bias towards the majority class, resulting in insufficient ability to identify risk samples. Conventional methods employ a focus loss function, but this fails to consider the spatiotemporal risk distribution characteristics, making it difficult to effectively alleviate the sample imbalance problem. This invention calculates the classification difficulty coefficient, risk density term, and feature uncertainty term based on the model's maximum predicted probability, the number of historical high-risk events, the total number of samples, and feature entropy. These are then weighted and fused to output dynamic loss weights, adjusting the importance of samples in the loss function, as expressed below: In the formula, This represents the dynamic loss weight for the τ-th time window, used to adjust the importance of a sample in the loss function; Indicates the first The probability of the maximum value in the model's predicted probability within the time window, i.e. This characterizes the confidence level of the model in predicting regional-level global risks. The classification difficulty coefficient is represented by a larger value for a lower predicted probability, indicating a focus on difficult-to-classify samples. This represents the maximum value of the regional-level global risk probability in the τ-th time window, characterizing the model's confidence level in the regional-level risk. This represents a time-focusing factor, amplifying the weighting effect of difficult samples, such as... ; This represents the number of historical high-risk events that occurred within the τ-th time window, calculated from historical data records. This represents the total number of samples within the τ-th time window, i.e., the number of data points within that time window. It is calculated as the number of sensing devices multiplied by the number of sampling points within the time window. This represents the risk density attenuation coefficient, which controls the strength of the influence of risk density on the weights, such as... ; Represents the characteristic entropy function. Item calculation decoupling characteristic matrix The information entropy method is to calculate the information entropy of the feature vector of each device and then take the average to represent the feature uncertainty. This represents the weighted fusion coefficient, balancing the classification difficulty term and the feature uncertainty term, such as... .

[0046] It should be noted that, The term generates a priori historical risk density, giving higher weight to samples from high-risk periods. The term quantifies uncertainty based on decoupling feature entropy. In the case of imbalanced samples, for the risk of sporadic slope collapse, the number of historical events is considered. When an accident has occurred in the area, even if the current predicted probability is... The risk density term remains high, and it will continue to focus on samples in that region to prevent the model from forgetting sparse risk patterns. Furthermore, Items and their representations Weighted combination of terms helps identify hidden risks with high prediction confidence but chaotic features, such as when the slope is in the creep stage, although the model predicts a high probability. The confidence level may be high due to the lack of obvious displacement, but the decoupling characteristic matrix... An abnormal increase in the entropy of geological and meteorological factors will trigger an increase in their weights, prompting the model to pay attention to such potential threats.

[0047] S2.6, Multi-granularity risk classification header module: Highway construction risks require the simultaneous identification of both equipment-level local risks and regional-level global risks. For example, exceeding displacement limits for a single sensing device or slope collapse. Conventional single classifiers cannot simultaneously handle risk features of different granularities, leading to local anomalies being overwhelmed by global features, failing to balance local and global characteristics, and affecting the comprehensiveness of risk identification. Therefore, this invention employs a dual-branch classifier structure to process equipment-level local risks and regional-level global risks separately, thereby simultaneously capturing equipment-level and regional-level risk features, expressed as: 1) Equipment-level local risk classification For each sensing device, its decoupled feature vector is passed through a shared projection layer and nonlinear activation. The risk probability distribution is then output by a local classifier, representing the device-level local risk, as follows: In the formula, Let represent the local risk probability vector of the i-th sensing device in the τ-th time window, characterizing device-level risks such as displacement exceeding limits; The decoupling feature vector of the i-th sensing device is represented by the decoupling feature matrix. The Row vector, dimension ; Represents the shared feature projection matrix, dimension The decoupled features are mapped to a unified space, and these are trainable parameters. The projection feature dimension is a hyperparameter preset by humans, such as... ; This represents a linear rectified activation function, enhancing the ability to express nonlinearities. This represents the parameter matrix of the local risk classifier, with dimensions... The projected features are mapped to local risk categories, and these are trainable parameters. This represents the number of local risk categories, preset according to actual application, and may include, for example, normal, displacement exceeding limits, abnormal vibration, etc. This represents the normalization exponential function, which transforms the classifier output into a probability distribution.

[0048] 2) Regional-level overall risk classification Aggregate the decoupled feature vectors of all sensing devices, capture cross-device temporal dependencies through a long short-term memory network, extract key features through max pooling, and output the regional risk probability distribution to represent the regional-level global risk, as shown below: In the formula, This represents a regional-level global risk probability vector. Indicates that the j-th sensor device starts from... Time window to the The decoupled feature vector sequence of the time window, with dimension . ; This indicates the number of review time windows, which is the length of the input sequence to the Long Short-Term Memory network. It is preset by the user, such as T=10. This represents a Long Short-Term Memory (LSTM) network, which includes an input gate, a forget gate, and an output gate. The size of the hidden layer can be set according to requirements. For example, it can be configured as a single-layer LSM network with 64 hidden units. This indicates traversing all One sensing device; This represents a spatiotemporal max pooling operation that takes the maximum value of the output of the long short-term memory network of all sensing devices to extract the most significant risk features. This represents the parameter matrix of the global risk classifier, with dimensions... The pooled features are mapped to global risk categories, and these are trainable parameters. This represents the number of hidden units in a Long Short-Term Memory (LSTM) network, for example, 64 hidden units. This represents the total number of risk categories, preset according to actual application, and may include, for example, normal risk, slope risk, traffic congestion risk, etc.

[0049] It should be noted that the dual-branch structure shares the feature projection matrix through the collaboration of device-level projection and region-level temporal aggregation. This approach enables all device local classifiers to share underlying feature representations, while the long short-term memory network in the regional branch captures the time sequence of risk propagation across devices, and max pooling extracts the most significant risk features. This helps to accurately identify the critical state where local anomalies have not exceeded the threshold but the regional trend is deteriorating.

[0050] S2.7, Pseudo-tag self-enhancing module: Due to the continuous changes in the highway construction environment, such as the movement of sensing equipment or the addition of monitoring points, the performance of statically trained models degrades. Conventional online learning methods do not consider the confidence characteristics of model predictions, which can easily generate false labels, causing model drift and making it impossible to select high-confidence samples, thus affecting model stability.

[0051] This invention uses a double-reset reliability threshold to screen reliable samples, generates pseudo-labels, and assigns dynamic weights for online learning. This effectively avoids model performance degradation caused by dynamic changes in the construction environment. The dynamic loss weights strengthen the focus on high-risk samples, suppress interference from erroneous pseudo-labels, and significantly improve the model's robustness and continuous perception accuracy in dynamic environments. The specific steps are as follows: 1) Screening of high-confidence samples Based on the global prediction confidence threshold and the KL divergence threshold between the local risk probability distribution and the uniform distribution, high-confidence samples are selected, and the high-confidence sample batch set is output, represented as: In the formula, Denotes the set of high-confidence sample batches in the τ-th time window; Represents the original data vector within the τ-th time window, characterizing the... The raw data set of all sensing devices within the time window, and the raw data vector. The j-th element is ; This represents the global confidence threshold, used to filter high-certainty predictions, such as... ; The device-level local risk probability distribution vector is the set of local risk probability vectors for all devices. The i-th element is ; Represents a uniformly distributed vector, with dimensions equal to the device-level local risk probability distribution vector. same; Indicates the Kullback-Leibler divergence. The term is calculated separately for each device, measuring the device-level local risk probability distribution vector. The difference from a uniform distribution is significant; a large difference indicates a high prediction confidence level. The term is characterized for the current time window only if the average KL divergence between the local risk prediction distributions of all devices and the uniform distribution is greater than the local distribution uncertainty threshold. Only then are they included in the high-confidence sample set; This represents the threshold for local distribution uncertainty, ensuring that local predictions have significant biases, such as... .

[0052] 2) Construction and calculation of loss function Pseudo-labels are generated for high-confidence samples. The online loss is calculated using dynamic loss weights, and the total loss function is updated to optimize the model's performance on the high-confidence pseudo-labels. This is represented as: In the formula, The online learning loss is represented only for the high-confidence sample batch set within the τ-th time window. The sample loss is calculated to optimize the model's performance on high-confidence pseudo-labels and prevent model drift. The calculation method is expressed as follows: ; This represents the dynamic loss weight for the τ-th time window, used to adjust the importance of a sample in the loss function; The classification loss includes equipment-level local risk classification loss and region-level global risk classification loss, and is calculated as follows: ; For the first The local risk real label of each device is a vector in one-hot encoded format; These serve as regional-level global risk pseudo-labels, used to supervise the training of regional-level classifiers during the online learning phase. This represents a regional-level global risk probability vector. Represents the cross-entropy loss function; This represents the total loss of the model, used for overall model optimization.

[0053] In practical implementation, regional-level global risk pseudo-labels These are not actual labeled data, but rather model predictions generated through a high-confidence sample selection mechanism. Specifically, they are obtained by selecting samples that meet certain conditions. For the samples, the current regional-level prediction results of the model are directly used as pseudo-labels, i.e. , The term represents taking the regional-level global risk probability vector. The category with the highest probability is represented by a one-hot encoded vector, with dimensions equal to the number of global risk categories. The same, for example, if the number of global risk categories is... It includes four categories: normal, slope risk, traffic congestion, and comprehensive risk. Characterize slope risk.

[0054] It should be noted that the high-confidence sample selection criteria employ a dual verification mechanism. The term characterizes the global confidence level to ensure the reliability of regional-level predictions, while The requirement is that device-level predictions have significant biases and avoid uniform distribution. When a new sensing device is added, the prediction should only be true if the local risk of that sensing device is accurately labeled. The Kullback-Leibler divergence meets the criteria and the regional prediction probability is consistent with the uniform distribution. Pseudo-labels are only generated when the confidence level is high, and this is combined with dynamic loss weights. This allows high-confidence samples during high-risk periods to receive stronger supervisory signals, effectively preventing model drift caused by erroneous pseudo-labels.

[0055] S2.8 Iterative training and parameter update of trainable parameters: All trainable parameters of the model are jointly optimized through backpropagation, specifically using the Adam optimizer, with an initial learning rate of 0.001 and a batch size of 64. Samples are randomly drawn from the training set for each batch. The training process involves the total loss of the model. As a basis for parameter updates, the gradient clipping threshold is set to 5.0 to prevent gradient explosion; After every 5 epochs of training, the macro F1 score is evaluated on the validation set. If there is no improvement for 3 consecutive times, the learning rate is reduced to 1 / 10 of the original value. The iteration terminates when the maximum training epoch (preset 200) is reached, and the model parameters with the best performance on the validation set are saved.

[0056] In a specific implementation, S3 is as follows: The trained model is deployed to the edge computing platform in the construction area. Real-time data flows through the spatiotemporal alignment and feature enhancement module. First, the original data of each device's neighborhood is aggregated, and after fusing the position encoding, the spatiotemporal alignment features are output through one-dimensional convolution. Then, the dynamic adjacency matrix module generates real-time topological connections based on device spatial distance and behavioral similarity. The aligned features are input into the gated spatiotemporal convolution module, which integrates historical and current spatiotemporal features through a multi-scale gating mechanism to enhance the response to sudden risk signals. The risk feature decoupling module projects the output features onto an independent risk factor space, generating four types of decoupling feature matrices: displacement, vibration, meteorology, and pedestrian flow. Then, the data is processed in parallel by a dual-branch classifier: the equipment-level branch outputs the probability of local risks such as displacement exceeding limits and vibration anomalies of each sensor based on decoupling features; the region-level branch aggregates the time-series features of all equipment through a long short-term memory network, and outputs the probability of global risks such as slope risk and traffic congestion after max pooling. Finally, by integrating the results at the device and regional levels, if the probability of any risk exceeds the threshold, an audible and visual alarm will be triggered, and the risk location and level will be marked on the digital twin platform to guide on-site personnel in emergency response.

[0057] The system updates the risk situation map every 5 minutes and continuously fine-tunes the model by filtering high-confidence samples through a pseudo-label self-reinforcement module to adapt to dynamic changes in the construction environment. The risk situation map is as follows: Figure 11 As shown.

[0058] Example 2 like Figure 2 As shown, to compare the overall performance of different methods in the risk identification task of highway construction areas, the experiment set up four different technical solutions for comparison, including the conventional interpolation alignment combined with fixed graph convolution method, the conventional interpolation alignment combined with dynamic graph convolution method, the spatiotemporal alignment technology of the present invention combined with fixed graph convolution method, and the complete method of the present invention. The experiment evaluated the accuracy from two dimensions: equipment-level risk identification accuracy and area-level risk identification accuracy. The accuracy unit is a percentage value, and the higher the value, the better the performance. like Figure 2The experimental results show that the complete method of this invention significantly outperforms other comparative methods in both device-level and region-level risk identification accuracy. Conventional interpolation alignment methods suffer from the lowest performance because they destroy the inherent spatiotemporal correlation of data during data preprocessing, leading to distorted feature extraction. While simply using dynamic graph convolution can improve performance to some extent, it fails to fundamentally solve the data alignment problem. The spatiotemporal alignment and feature enhancement module of this invention achieves feature-level alignment and local physical correlation feature enhancement by aggregating spatial neighborhood sensor data and fusing position encoding. Therefore, it can achieve significant improvement even when combined with fixed graph convolution. The complete method of this invention combines innovative modules such as spatiotemporal alignment, dynamic adjacency matrix construction, and gated spatiotemporal convolution, fully exploiting the spatiotemporal correlation characteristics of multi-source sensor data. Therefore, it achieves the highest level in both types of accuracy indicators, demonstrating the significant advantages of this invention in overall risk identification performance.

[0059] Example 3 like Figures 3 to 5 As shown, a comparative analysis of spatiotemporal alignment features was conducted to verify the effectiveness of the spatiotemporal alignment module in solving the problem of data asynchrony among multi-source sensing devices. The experiment simulated vibration signals collected by three vibration sensors (device 1, device 2, and device 3) in a highway construction area. Device 1 served as the reference device, device 2 had a 150-millisecond time delay, and device 3 had a 100-millisecond time lead. In the figure, the horizontal axis represents time (in seconds), and the vertical axis represents vibration amplitude (in millimeters). The experimental results show that the original signal... Figure 3 The vibration peaks of devices 2 and 3 are significantly shifted from those of device 1, particularly in the resonance region of 1.8–2.5 seconds and the anomalous region of 6.8–7.5 seconds. (Conventional interpolation alignment results) Figure 4 While time alignment was achieved, waveform distortion occurred in the resonance region (peak value decreased by about 30% at 2 seconds) and pulse smoothing occurred in the abnormal region (amplitude attenuation of about 25% at 7 seconds). The alignment results of this technique... Figure 5 As shown, the resonant waveform features are fully preserved in the resonant region, and the steep rising edge of the pulse is clearly recovered in the abnormal region. Experimental results show that the joint modeling of position encoding and one-dimensional convolution enables this technique to reconstruct continuous vibration waveforms from neighboring devices during the data loss period (3-4 seconds) of device 2, avoiding the problem of physical correlation breakage caused by conventional interpolation.

[0060] Example 4 like Figures 6 to 9 As shown, dynamic adjacency matrix analysis is performed to verify the adaptability of the dynamic adjacency matrix module to changes in device topology. Figure 6As shown, 20 sensors (blue for vibration, orange for displacement, green for weather, and red for pedestrian flow) were deployed in a 100×50 meter construction area. The horizontal and vertical coordinates are in meters, and the lines represent the dynamic connections established by this technology.

[0061] Figure 7 This is the behavioral similarity matrix (0-1 range). Figure 8 This is a standard distance adjacency matrix (30-meter threshold). Figure 9 This is the dynamic adjacency matrix of this technology; Depend on Figures 7 to 9 It is known that conventional methods rely solely on spatial distance, resulting in numerous erroneous connections between physically adjacent but behaviorally unrelated devices, while neglecting distant but behaviorally related devices. This technology, through dual constraints of spatial attenuation coefficient and temporal feature similarity, can effectively reduce erroneous connections while preserving correct connection relationships. Figure 7 The yellow area indicates that this technology can identify behavioral correlations across different device types, providing accurate topological support for gated spatiotemporal convolution.

[0062] Example 5 like Figure 10 As shown, the performance trends of each method under different training stages are analyzed to evaluate the convergence speed and final performance of the model. Experiments recorded the changes in the macro F1 score of the conventional method (using a standard convolutional neural network with 3 layers) and the method of this invention over multiple training cycles. The training cycle unit is the number of model iterations. The macro F1 score is a performance indicator that comprehensively considers precision and recall; a higher value indicates better overall model performance. The curves show that the method of this invention maintains a performance advantage throughout the entire training process, and its convergence speed is significantly faster than that of the conventional method. In the early stages of training, the F1 score of the method of this invention rises rapidly, indicating that the model can quickly learn effective feature representations. As the training cycle increases, the performance of the method of this invention continues to improve and eventually stabilizes at a high level. In contrast, the conventional method has a slower initial convergence speed, and its final stable performance level is significantly lower than that of the method of this invention. Experimental results demonstrate the effectiveness of each module in the method of this invention, especially the pseudo-label self-enhancement module, which uses double-reset confidence thresholds to screen reliable samples, generate pseudo-labels, and assign dynamic weights for online learning, effectively accelerating the model convergence process and improving the final performance. The dynamic loss weight calculation module also enhances the focus on high-risk samples by adjusting the importance of samples in the loss function, thereby further improving model performance.

[0063] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A highway construction zone risk situation awareness method based on environmental perception, characterized in that, The method comprises the following steps: S1, deploying a multi-source sensor device network along the highway construction zone and at key structure locations, collecting data throughout the construction cycle to build a training set, and labeling the data; S2, constructing a highway construction zone risk situation awareness model and training it, sequentially constructing a time-space alignment and feature enhancement module, a dynamic adjacency matrix construction module, a gated time-space convolution module, a risk feature decoupling module, an adaptive dynamic loss weight calculation module, a multi-granularity risk classification head module, and a pseudo-label self-enhancement module, jointly optimizing all trainable parameters of the model through back propagation, and obtaining the trained model; S3, deploying the trained model to the edge computing platform in the construction zone, outputting the final prediction result from the trained model for real-time collected data, determining whether to issue a warning based on the prediction result, and calibrating the risk location and level, and continuously screening high-confidence samples through the pseudo-label self-enhancement module to fine-tune the model.

2. The environmental perception-based highway work zone risk situation awareness method according to claim 1, characterized in that, The training process of the highway construction zone risk situation awareness model in S2 is as follows: Through the time-space alignment and feature enhancement module, the multi-source data in the training set is aligned at the feature level and the local physical correlation features are enhanced, and the time-space alignment features of each sensor device in a specific time window are obtained; The dynamic adjacency matrix construction module constructs a dynamic adjacency matrix, and determines the connection relationship between the sensor devices according to the time-space alignment features; The time-space alignment features are input into the gated time-space convolution module to generate gate signals and time-space features layer by layer; The risk feature decoupling module guides the model to decompose the time-space features into independent factors according to the risk categories through the feature decoupling loss function, projects the time-space features onto the decoupling vectors of each independent factor, and generates a decoupling feature matrix; The adaptive dynamic loss weight calculation module calculates the feature entropy through the decoupling feature matrix, and adjusts the weight of the collected data in the loss function calculation by combining the maximum value of the model prediction probability, the number of historical high-risk events, and the total number of samples, to generate dynamic loss weights in each time window; Through the multi-granularity risk classification head module, the device-level local risk and the regional-level global risk are processed respectively, and the local risk probability of each sensor device and the regional-level global risk probability are obtained; The pseudo-label self-enhancement module screens reliable samples through double confidence thresholds, generates pseudo-labels and dynamic loss weights for online learning, and then calculates the total loss of the model; The model updates the parameters through the total loss as the basis for parameter update, and iteratively updates the parameters for multiple rounds until the maximum training round is terminated, and the trained model is obtained. 3.The method according to claim 2, wherein, The time-space alignment and feature enhancement module is as follows: For each sensor device in a specific time window, aggregate the data collected by all sensor devices in its spatial neighborhood, fuse the spatial distance information through position encoding, and output the time-space alignment features of each sensor device in the specific time window based on one-dimensional convolution and timestamp offset.

4. The environment perception-based highway work zone risk situation awareness method according to claim 2, characterized in that, The dynamic adjacency matrix construction module is as follows: An effective connection is established between physically adjacent and behaviorally related sensing devices, specifically, for any two sensing devices, a normalized product of spatial distance decay similarity and time feature similarity is calculated based on spatio-temporal alignment features, a binary adjacency matrix element is generated by combining the spatial distance constraint, and a dynamic adjacency matrix is output. 5.The environment-aware based highway work zone risk situation awareness method according to claim 2, characterized in that, The gated spatio-temporal convolution module is specifically as follows: The gated spatio-temporal convolution module adopts a multi-scale gated convolution unit, the spatio-temporal alignment features are used as the spatio-temporal features of the first layer input of the multi-scale gated convolution unit, in each layer of the multi-scale gated convolution unit, the spatio-temporal features of the current time are spliced with the historical time features, convolution operation is performed through a gating convolution kernel, and then a gating signal is generated through a Sigmoid activation function, the modulated gating signal of the current layer is added to the spatio-temporal features after spatio-temporal graph convolution, and the updated spatio-temporal features are output through layer normalization. 6.The environment-aware based highway work zone risk situation awareness method according to claim 2, characterized in that, The risk feature decoupling module is specifically as follows: The feature decoupling loss function calculates the difference between the expected value and the conditional expected value of the inner product of the decoupling vector and the spatio-temporal features, and calculates the sum of the absolute values of the inner products between the decoupling vectors as the orthogonal constraint term, and then guides the model to decompose the risk features into independent factors.

7. The environment-aware based highway work zone risk situation awareness method according to claim 2, characterized in that adaptive The dynamic loss weight calculation module is specifically as follows: The adaptive dynamic loss weight calculation module calculates the feature entropy of the decoupled feature matrix, and then combines the maximum prediction probability of the model, the number of historical high-risk events and the total number of samples to calculate the classification difficulty coefficient, the risk density term and the feature uncertainty term, and then weightedly fuses to output the dynamic loss weight; The feature entropy of the decoupled feature matrix is obtained by calculating the information entropy of the feature vector of each sensing device and then averaging. 8.The environment-aware based highway work zone risk situation awareness method according to claim 2, characterized in that, The multi-granularity risk classification head module is specifically as follows: The decoupled feature vector of each sensing device is input into a local classifier after passing through a shared projection layer and a nonlinear activation, and a device-level local risk probability vector is output; The decoupled feature vectors of all sensing devices are aggregated, the long short-term memory network is used to capture the cross-device time sequence dependence, the key features are extracted through maximum pooling, and a region-level global risk probability vector is output. 9.The environment-aware based highway work zone risk situation awareness method according to claim 1, characterized in that, The pseudo-label self-enhancement module is specifically as follows: Based on the global prediction confidence threshold and the KL divergence threshold of the local risk probability distribution and the uniform distribution, high-confidence samples are screened, and a high-confidence sample batch set is output; Pseudo-labels are generated for high-confidence samples, online loss is calculated combined with dynamic loss weight, and the total loss function is updated to optimize the model's performance on high-confidence pseudo-labels. 10.The environment-aware based highway work zone risk situation awareness method according to claim 1, characterized in that, S1 is specifically as follows: Deploy a multi-source sensing device network along the highway construction zone and at key structure locations, collect data throughout the construction cycle to build a training set, and label the data; The multi-source sensing device network includes vibration sensors, displacement sensors, weather sensors, and people flow monitoring devices; The vibration sensor collects the vibration amplitude of the roadbed and structure at a high sampling rate; the displacement sensor monitors the displacement of the slope and retaining wall at a medium sampling rate; the weather sensor records temperature, humidity, wind speed, and rainfall at a low sampling rate; the people flow monitoring device captures the activity density of construction personnel in an asynchronous manner through infrared recognition; The construction training data set covers the whole construction cycle, and the multi-source original data vector is stored by slicing according to a fixed time window, each sample contains a time stamp, a device spatial coordinate and a sensing value; The collected data are labeled based on the construction log, video monitoring and expert evaluation, and the device-level local risk label and the regional-level global risk label are labeled for each time window, the device-level local risk label includes normal, displacement overrun, vibration anomaly and human flow overrun, and the regional-level global risk label includes normal, device-level risk, environment-level risk and comprehensive risk.

Images