A dynamic safety early warning method and system for aerial building construction machines based on multi-source information fusion
By constructing a dynamic safety early warning system for aerial building construction machines that integrates multi-source information and combining graph neural networks and temporal convolutional networks, the problem of controlling the platform's levelness during the lifting process of aerial building construction machines was solved, achieving high-precision safety early warning and construction risk management.
Patent Information
- Application Number
- CN202311431607.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-31
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-10-31
AI Technical Summary
Existing aerial building construction machines have problems such as difficulty in controlling the platform's levelness and asynchronous lifting of hydraulic cylinders during the jacking process, which lead to construction safety hazards. There is an urgent need for a dynamic safety early warning method to identify potential risk factors and prevent accidents.
A multi-source sensor graph network is constructed by using a multi-source information fusion method, combining graph neural networks, temporal convolutional networks, and attention mechanisms. The spatiotemporal correlation of sensors is captured through graph convolutional layers and temporal convolutional layers, enabling accurate prediction and early warning of platform level.
It achieves high-precision platform level prediction, reduces construction risks, improves construction safety and prediction stability, and reduces the probability of safety accidents.
Smart Images

Figure CN117610916B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of structural safety and time-series prediction, and more specifically, relates to a dynamic safety early warning method and system for aerial building machines based on multi-source information fusion. Background Technology
[0002] With the rapid urbanization of the past two decades, high-rise buildings have gained widespread popularity and use worldwide due to their aesthetic design, comfortable environment, and land-saving advantages. However, high-rise building construction presents a series of challenges, including high difficulty, high risk, and complex organization. Traditional climbing formwork construction carries operational risks such as formwork collapse, falling objects, and collisions. Therefore, there is an urgent need for advanced construction equipment to improve the on-site working environment and enhance construction efficiency. Against this backdrop, the aerial building machine (construction equipment integration platform), with its high degree of equipment integration, factory-like operation, fast construction speed, and guaranteed construction safety, has been rapidly deployed in the construction of numerous landmark high-rise buildings in China. This has greatly improved the mechanization and intelligence level of high-rise building construction, propelling my country's high-rise building construction technology to new heights.
[0003] Sky-building machines primarily utilize multiple lifting cylinders simultaneously controlled on an existing concrete structure to elevate the construction platform in an alternating climbing motion, transforming the construction environment from high-altitude, challenging work to safer and more comfortable ground-based construction. However, due to factors such as unbalanced loads on each hydraulic cylinder and the harsh and variable construction environment, asynchronous cylinder strokes can occur during the lifting process. This can lead to platform leveling exceeding limits and localized structural deformation, posing significant safety hazards to high-rise building construction. To improve the safety of high-rise building construction and reduce the probability of accidents, it is necessary to implement dynamic risk prediction and early warning for the lifting process of the sky-building machine. In the context of the digitalization of construction equipment, how to utilize advanced intelligent algorithms to achieve monitoring data processing, real-time risk prediction, and identification of risk factors is a crucial issue. Summary of the Invention
[0004] To address the shortcomings or improvement needs of existing technologies, such as difficulties in controlling platform level and asynchronous lifting by hydraulic cylinders, this invention provides a dynamic safety early warning method for aerial building construction machines based on multi-source information fusion. By combining graph neural networks, temporal convolutional networks, and attention mechanisms, it achieves accurate prediction and early warning of platform level during the lifting process of the building construction machine, thereby quickly identifying potential risk factors and effectively avoiding the occurrence of safety accidents.
[0005] To achieve the above objectives, this invention provides a dynamic safety early warning method for aerial building machines based on multi-source information fusion, comprising the following steps:
[0006] S1: Select the features of the building-building machine's lifting process as model variables, use sensors to acquire the feature information of the model variables, and obtain initial data; use the Z-score method to normalize the initial data; construct a three-dimensional matrix of sample number, time step and features as the training set and test set of the model;
[0007] S2: Constructing the graph convolution module: A multi-source sensor graph network is constructed using the fast DTW algorithm and the Direct-LiNGAM algorithm, respectively; the network is then input into GraphSAGE for node embedding and information aggregation.
[0008] S3: Construct a temporal feature module based on a temporal convolutional network (TCN). The temporal feature module includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements.
[0009] S4: Construct a spatiotemporal prediction model based on graph convolution module and temporal feature module; divide the data in the training set into multiple batches to train the model in multiple periods, and use mean squared error (MSE) as the loss function; input the test set data into the trained model to evaluate the model performance;
[0010] S5: Perform single-step prediction, multi-step prediction, and feature importance analysis based on the Saliency method for platform level.
[0011] Furthermore, in step S1, the characteristics of the aerial building machine during the lifting process include, but are not limited to, cylinder load, cylinder stroke, and platform levelness.
[0012] Furthermore, in step S2, the construction of the multi-source sensor graph network using the fast DTW algorithm is based on time similarity and the search length T is limited. The specific steps include:
[0013] S211: Calculate the similarity between the target sensor and the candidate relation;
[0014] S212: Select the k candidate sensors with the highest similarity index and establish a connection with the target sensor;
[0015] S213: Input the temporal similarity adjacency matrix A formed after establishing the connection into the spatiotemporal prediction model.
[0016] Furthermore, in step S2, the step of constructing a multi-source sensor graph network using the Direct-LiNGAM algorithm based on the causal discovery perspective involves constructing another adjacency matrix A′, which is input into the spatiotemporal prediction model along with the temporal similarity adjacency matrix A.
[0017] Furthermore, the graph convolution module has two parallel graph convolution layers with Elu as the activation function.
[0018] Furthermore, in step S3, the temporal convolutional network (TCN) introduces dilated convolution to enable the convolutional kernel to achieve a wider coverage range and skip input samples with a fixed stride; for the time series X = (x0, x1, ..., x... t ,…,x T ) and convolution kernel G = (g0, g1, ..., g f-1 The extended causal convolution result at time t is defined as:
[0019]
[0020] Where f is the kernel size, x t -(f-1-j) is used to prevent the convolution kernel G from acting on x. t+1 ,x t+2 ,…,x T d is the inflation ratio that controls the jump step size; when the length of the past time series is shorter than the designed memory size, zero-padding is usually applied to the left side of the time series to achieve causal convolution.
[0021] Furthermore, in step S3, the temporal convolutional network (TCN) uses deep residual connections. The staggered blocks in the stacked layers contain additional residual connections to guide the input x to be directly added to the output y, as expressed in the following expression:
[0022] y = f(x; W) + x
[0023] Where f(x; W) is the actual computational output of the stacked layers.
[0024] Furthermore, the number of channels in the temporal convolutional network (TCN) is set to [512, 1024], ReLU is used as the activation function, and several fully connected layers are also provided.
[0025] Furthermore, the performance evaluation metrics for the model include, but are not limited to, mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²). 2 :
[0026]
[0027]
[0028]
[0029]
[0030] Where N represents the number of samples, f i (x) represents the prediction result, y i Represents the true value. This represents the average value of the observed data.
[0031] According to a second aspect of the present invention, a dynamic safety early warning system for aerial building machines based on multi-source information fusion is provided, comprising:
[0032] The first main module is used for data acquisition and preprocessing: acquiring feature information of model variables in the sensor, obtaining initial data, and normalizing it; constructing a three-dimensional matrix of sample number, time step and features;
[0033] The second main module, namely the graph convolution module, is used to construct a multi-source sensor graph network based on the fast DTW algorithm and the Direct-LiNGAM algorithm, and to perform node embedding and information aggregation in GraphSAGE;
[0034] The third main module, namely the temporal feature module, includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements.
[0035] The fourth main module is used for single-step prediction and multi-step prediction of platform level, as well as feature importance analysis based on the Saliency method.
[0036] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:
[0037] 1. This invention selects multi-source monitoring data, such as cylinder load, cylinder stroke, platform level, etc., and constructs a multi-dimensional graph relationship structure between sensors based on time similarity and causal discovery. Through graph convolutional layers, temporal convolutional layers and attention mechanisms, it achieves high-precision prediction of platform level during the lifting process.
[0038] 2. The spatiotemporal prediction model of this invention has good prediction robustness, and the R-value between multi-step predictions is high. 2 The highest rate of change was only 1.370%, indicating that the spatiotemporal prediction model proposed in this invention can accurately and stably predict the trend of platform level changes, thereby improving the dynamic safety of construction during the jacking process and reducing the probability of risk occurrence. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of an MGTCN network architecture according to an embodiment of the present invention;
[0040] Figure 2 This is a schematic diagram of a sensor relationship based on causal discovery in one embodiment of the present invention;
[0041] Figure 3 This is a sensor relationship diagram based on the fast-DTW algorithm in one embodiment of the present invention;
[0042] Figure 4 The following is a statistical diagram of a multi-step prediction index in one embodiment of the present invention: (a) y1; (b) y2; (c) y3; (d) y4;
[0043] Figure 5 The following is a schematic diagram of the multi-step prediction results of y1 in one embodiment of the present invention: (a) prediction step 1; (b) prediction step 10;
[0044] Figure 6 The following is a schematic diagram of the multi-step prediction results of y2 in one embodiment of the present invention: (a) prediction step 1; (b) prediction step 10;
[0045] Figure 7 The following is a schematic diagram of the multi-step prediction results of y3 in one embodiment of the present invention: (a) prediction step 1; (b) prediction step 10;
[0046] Figure 8 The following is a schematic diagram of the multi-step prediction results of y4 in one embodiment of the present invention: (a) prediction step 1; (b) prediction step 10;
[0047] Figure 9 Here are heatmaps showing the model interpretability in one embodiment of the present invention: (a) y1; (b) y2; (c) y3; (d) y4;
[0048] Figure 10 This is a schematic diagram of a dynamic safety early warning system for an aerial building machine based on multi-source information fusion, according to one embodiment of the present invention. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0050] This invention provides a dynamic safety early warning method for aerial building machines based on multi-source information fusion, namely, the construction of a multi-source sensor graph network, a graph convolution module, a temporal convolution module, and the establishment of a spatiotemporal prediction model. Specifically, it includes the following steps:
[0051] S1: Select the features of the building-building machine's lifting process as model variables, use sensors to acquire the feature information of the model variables, and obtain initial data; use the Z-score method to normalize the initial data; construct a three-dimensional matrix of sample number, time step and features as the training set and test set of the model;
[0052] S2: Constructing the graph convolution module: A multi-source sensor graph network is constructed using the fast DTW algorithm and the Direct-LiNGAM algorithm, respectively; the network is then input into GraphSAGE for node embedding and information aggregation.
[0053] The graph edge establishment method based on the fast DTW algorithm identifies the time series of two sensor nodes from the perspective of temporal similarity, establishes connections between the nodes, and then inputs them into the network in the form of an adjacency matrix. To reduce the complexity of conventional DTW, this invention limits the search length T. The pseudocode for applying the fast DTW algorithm to arbitrary two-end time series is shown below:
[0054]
[0055]
[0056] The adjacency matrix is represented using a directed graph, where nodes represent sensors and edges represent dependencies between sensors. The adjacency matrix A represents a directed graph, where A... ij This indicates the existence of a directed edge from node i to node j. Each sensor node i has a set of candidate relations. In the absence of prior information, candidate sensors for target sensor i are all sensors other than itself. The fast DTW algorithm can be used to calculate the temporal similarity between the target sensor and candidate sensor i, thereby establishing a relationship connection.
[0057] Specifically, the steps for constructing the relationship connection by calculating the temporal similarity between the target sensor and candidate sensor i using the fast DTW algorithm include:
[0058] S21: Calculate sensor i and candidate relationships The similarity between them;
[0059] S22: Select the top-k candidate sensors, i.e. the k sensors with the highest similarity index, and establish a connection with the target sensor;
[0060] S23 uses the adjacency matrix A in a time series prediction model.
[0061] Furthermore, this invention employs the Direct-LiNGAM algorithm to construct another adjacency matrix A′ from a causal discovery perspective. This matrix, along with the temporal similarity adjacency matrix A, is then input into the prediction model to improve its generalization ability. Direct-LiNGAM is a method for exploring causal relationships between features, based on the widely used LiNGAM causal discovery algorithm. The Direct-LiNGAM method does not require prior knowledge to obtain causal relationships between variables and has stable computational complexity with easy convergence.
[0062] During construction, multiple sensors collected a large amount of monitoring data. Choosing a suitable GNN (Graph Neural Networks) model for data feature extraction is essential. GraphSAGE is an inductive learning framework that can quickly generate information embeddings for unknown vertices based on their known properties. It samples the neighboring nodes of each root node and updates nodes by hierarchically aggregating hidden nodes. This process is called node embedding. The basic idea of node embedding is to use dimensionality reduction methods to extract the high-dimensional information of the root node's neighbors into dense vector embeddings. These node embeddings are then fed into downstream machine learning systems to help complete tasks such as node classification, clustering, and prediction. Compared to conventional GNNs, GraphSAGE offers greater flexibility and generalization ability.
[0063] S3: Construct a temporal feature module based on a temporal convolutional network (TCN). The temporal feature module includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements.
[0064] This invention uses TCN (Temporal convolution network) to capture the temporal relationships between multiple sensors. TCN not only includes features of traditional CNNs, such as local connections and weight sharing, but also includes some special structural forms, such as causal convolution, dilated connections, and residual connections. Therefore, TCN has more flexibility and a longer receptive field size, and has a stronger historical memory ability than RNN and LSTM.
[0065] TCN utilizes causal convolution, where the output at time t is convolved only with neurons at time t and previous time steps. Causal convolution ensures that future information is not leaked into the current prediction and considers history within the receptive field. Furthermore, dilated convolution is introduced without increasing the model parameters to exponentially increase the receptive field. Unlike normal convolution where each kernel covers a range equal to its length, the kernels in dilated convolution can achieve a wider coverage, skipping input samples with a fixed stride. Formally, for a time series X = (x0, x1, ..., x... t ,…,x T ) and convolution kernel G = (g0, g1, ..., g f-1 The extended causal convolution result at time t is defined as:
[0066]
[0067] Where f is the kernel size, x t -(f-1-j) is used to prevent the convolution kernel G from acting on x. t+1 ,x t+2 ,…,x T d is the inflation ratio that controls the jump step size; when the length of the past time series is shorter than the designed memory size, zero-padding is usually applied to the left side of the time series to achieve causal convolution.
[0068] Furthermore, to ensure the stability of deeper network layers, deep residual connections are also used. For several stacked layers, a residual block contains an additional residual connection to guide the input x directly to the output y, expressed as:
[0069] y = f(x; W) + x
[0070] Where f(x; W) is the actual computational output of the stacked layers.
[0071] S4: Construct a spatiotemporal prediction model based on graph convolution module and temporal feature module; divide the data in the training set into multiple batches to train the model in multiple periods, and use mean squared error (MSE) as the loss function; input the test set data into the trained model to evaluate the model performance;
[0072] This invention proposes a Multi-graph Temporal Convolution Network (MGTCN) for constructing a spatiotemporal prediction model. The network framework is as follows: Figure 1As shown, the main advantage of MGTCN lies in its ability to simultaneously capture multi-dimensional spatiotemporal latent features of monitoring data from both the node relationships and temporal relationships of sensors. This spatiotemporal prediction model mainly consists of two modules: a graph convolution module and a temporal feature module. In the graph convolution module, monitoring data from the time window is reconstructed into a multi-dimensional graph structure, which is then input into GraphSAGE for information aggregation. In the temporal module, TCN is used to capture spatiotemporal correlations from the spatiotemporally aggregated vectors. Furthermore, an attention mechanism is introduced into the model, which helps the prediction model focus on the input information to improve prediction accuracy. Finally, the output layer, composed of fully connected networks, reconstructs the prediction data from the latent vectors processed by the attention mechanism, based on the length and dimension of the prediction requirements.
[0073] Model training is a crucial step in the entire process, as it determines the model's final performance. The training process is repeated over multiple epochs by further dividing the training data into multiple batches. The model uses mean squared error (MSE) as the loss function. MSE can be calculated using the following equation:
[0074]
[0075] Where N represents the number of samples, f i (x) represents the prediction result, y i Represents the actual value.
[0076] The performance of the trained model is evaluated by applying it to test data. This invention selects four commonly used metrics for performance evaluation: mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²). 2 These indicators can be calculated using the following equations:
[0077]
[0078]
[0079]
[0080]
[0081] Where N represents the number of samples, f i (x) represents the prediction result, y i Represents the true value. This represents the average value of the observed data.
[0082] One embodiment of the present invention provides a dynamic safety early warning method for aerial building construction machines based on multi-source information fusion, specifically applied to the construction of a 356m high-rise building using an aerial building construction machine. The main steps include:
[0083] S1: Selection and Processing of Model Variables. To accurately predict the development of the steel platform's levelness, all 26 sets of cylinder load data were selected. Since the stroke of each cylinder is similar, only one set of cylinder stroke data was chosen. Furthermore, a total of 16 static level sensors were installed on the steel platform to measure levelness. Platform levelness is affected by various factors, and each sensor exhibits a different degree of fluctuation. Therefore, the four sensors with the most significant levelness fluctuations during the jacking process were selected. In summary, the model variables include the 26 cylinder loads (X1 to X...). 26 ), Cylinder stroke (X) 27 The data includes 4 platform levels (y1 to y4). Z-score normalization is then used to prevent feature unbiasedness due to excessive data variance. After standardization, the data is a two-dimensional matrix of time steps and features, with each time step representing one sample. To use past monitoring data as training data for predicting platform levels, the two-dimensional data needs to be redefined as the number of samples S. n A three-dimensional matrix of time step and features.
[0084] Table 1 Description of Monitoring Data Characteristics
[0085]
[0086] Note: X1-X 26 The unit of the variable is t; X 27 The unit for variables y1-y4 is mm.
[0087] S2: Construct the graph convolution module. After completing the reconstruction process, a graph can be constructed for each sample. This represents the sensors deployed in the building machine (i.e., the features in the reconstructed data), and its value is the reconstructed time subsequence of each feature. The edges connecting two nodes are represented by the direct-LiNGAM and fast DTW algorithms used to construct feature maps within the sensor, such as... Figure 2 and Figure 3 As shown. The size of a node represents the number of connected nodes. Figure 2 The data analysis reveals the causal relationships between different nodes. As shown in the graph, the hydraulic cylinder stroke is the node with the most causal relationships. Figure 3The graph structure with Top-k=5 is plotted, representing the temporal similarity relationship of multiple source sensors. This graph shows a clear temporal similarity between platform levelness, cylinder stroke, and partial cylinder load, thus allowing the establishment of connections between multiple nodes.
[0088] S3: Construct a temporal feature module based on a temporal convolutional network (TCN). The temporal feature module includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements.
[0089] S4: Construct a spatiotemporal prediction model based on graph convolution module and temporal feature module; divide the data in the training set into multiple batches to train the model in multiple periods, and use mean squared error (MSE) as the loss function; input the test set data into the trained model to evaluate the model performance;
[0090] Specifically, the spatiotemporal prediction model has two parallel graph convolutional layers, each with 256 neurons and ELU activation function. The number of channels in the TCN layer is set to [512, 1024], with ReLU as the activation function and a kernel size of 2. Additionally, there are two fully connected layers with 64 and 10 neurons respectively. During training, the batch size is set to 128, and the number of heads for the attention mechanism is set to 4. The initial learning rate of the model is 0.001. The model is trained using the AdamW optimization algorithm with 400 training epochs. The input is 32 steps of historical data, and the model output is a 10-step prediction sequence. Monitoring data from a complete jacking process is selected, with 80% of the monitoring data used for model training and the remainder for model testing. The mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²) are used to measure the mean absolute error. 2 .
[0091] S5: Platform Level Multi-Step Prediction. The prediction model of this invention exhibits high performance in multi-step prediction. According to... Figure 4 It can be seen that single-step prediction is the most accurate, and the R² of 4 sensors is... 2 The values were 0.970, 0.949, 0.945, and 0.967, respectively. Among the four sensors, the R value predicted in step 10 was... 2 They reached 0.923, 0.920, 0.915 and 0.912 respectively, by Figures 5 to 8 Multi-step prediction comparisons show that the predicted values from each sensor match the actual values well, the error distribution is relatively uniform, and the prediction error is relatively small when data changes abruptly. According to Table 2, the R-values between every two steps for the four sensors... 2The rates of change were stable, with average values of 0.555%, 0.348%, 0.358%, and 0.645%, respectively. Among all sensors, R... 2 The maximum rate of change is only 1.370%, indicating that the spatiotemporal prediction model has strong robustness. Calculation results show that the proposed prediction model provides reliable estimation performance, which can assist in dynamic risk assessment and decision-making during the jacking process of aerial building machines. Furthermore, the Saliency method is used to further analyze the importance of features. Figure 9 The heatmap illustrates the impact of the previous 16 steps on the prediction results. Colors in the heatmap represent the importance of variables; darker colors contribute more to the target value. Platform level prediction is more sensitive to the most recent time step. The most important features of the four level sensors are their raw values over the past three time steps, followed by the cylinder load x. 17 And the cylinder stroke x 27 The cylinder load x in the past time t-1 17 Sensors y1 and y4 contribute relatively more to the result. The results indicate that the platform's levelness during the lifting process is affected by its previous time step, the load on some cylinders, and the stroke. Therefore, it is recommended to focus on the cylinders with higher importance.
[0092] Table 2 R every two steps 2 rate of change
[0093]
[0094] The implementation of the various embodiments of this invention is based on programmed processing by a device with a central processing unit (CPU). Therefore, in practice, the technical solutions and functions of the various embodiments of this invention can be encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of this invention provide a dynamic safety early warning system for aerial building-building machines based on multi-source information fusion, used to execute a dynamic safety early warning method for aerial building-building machines based on multi-source information fusion from the above method embodiments. For example... Figure 10 As shown, it includes:
[0095] The first main module is used for data acquisition and preprocessing: acquiring feature information of model variables in the sensor, obtaining initial data, and normalizing it; constructing a three-dimensional matrix of sample number, time step and features;
[0096] The second main module, namely the graph convolution module, is used to construct a multi-source sensor graph network based on the fast DTW algorithm and the Direct-LiNGAM algorithm, and to perform node embedding and information aggregation in GraphSAGE;
[0097] The third main module, namely the temporal feature module, includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements.
[0098] The fourth main module is used for single-step prediction and multi-step prediction of platform level, as well as feature importance analysis based on the Saliency method.
[0099] The above embodiment is only one specific embodiment of the system embodiment of the present invention. In specific implementation, specific adjustments can be made to each module of the system according to the present invention.
[0100] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A dynamic safety early warning method for aerial building machines based on multi-source information fusion, characterized in that, Includes the following steps: S1: Select the features of the building-building machine's lifting process as model variables, use sensors to acquire the feature information of the model variables to obtain initial data; use the Z-score method to normalize the initial data; construct a three-dimensional matrix of sample number, time step and features as the training set and test set of the model; S2: Constructing the graph convolution module: A multi-source sensor graph network is constructed using the fast DTW algorithm and the Direct-LiNGAM algorithm, respectively; the network is then input into GraphSAGE for node embedding and information aggregation. S3: Construct a temporal feature module based on Temporal Convolutional Network (TCN). The temporal feature module includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements. S4: Construct a spatiotemporal prediction model based on graph convolution modules and temporal feature modules; divide the data in the training set into multiple batches to train the model in multiple periods, using mean squared error (MSE) as the loss function; input the test set data into the trained model to evaluate the model performance; S5: Perform single-step prediction, multi-step prediction, and feature importance analysis based on the Saliency method for platform level. In step S1, the characteristics of the aerial building machine during the lifting process include, but are not limited to, cylinder load, cylinder stroke, and platform levelness; In step S2, the construction of the multi-source sensor graph network using the fast DTW algorithm is based on time similarity and the search length T is limited. The specific steps include: S211: Calculate the similarity between the target sensor and the candidate relation; S212: Select the k candidate sensors with the highest similarity index and establish a connection with the target sensor; S213: The time similarity adjacency matrix formed after establishing the connections. Input spatiotemporal prediction model; In step S2, the Direct-LiNGAM algorithm is used to construct a multi-source sensor graph network based on the causal discovery perspective. The steps include constructing another adjacency matrix. The adjacency matrix of time similarity The spatiotemporal prediction model is input together; the graph convolution module has two parallel graph convolution layers with Elu as the activation function.
2. The dynamic safety early warning method for aerial building machines based on multi-source information fusion according to claim 1, characterized in that, In step S3, the temporal convolutional network (TCN) introduces dilated convolution to enable the convolutional kernel to achieve a wider coverage range and skip input samples with a fixed stride; for time series... and convolution kernel ,time The extended causal convolution result is defined as: in, It is the kernel size. Used to prevent convolution kernels Acting on , It controls the expansion ratio of the jump step; when the length of the past time series is shorter than the designed memory size, zero-padding is usually applied to the left side of the time series to achieve causal convolution.
3. The dynamic safety early warning method for aerial building machines based on multi-source information fusion according to claim 1, characterized in that, In step S3, the temporal convolutional network (TCN) uses deep residual connections, and the staggered blocks in the stacked layers contain additional residual connections to guide the input. Add it directly to the output y, the expression is: in, It is the actual computational output of the stacked layers.
4. The dynamic safety early warning method for aerial building machines based on multi-source information fusion according to claim 1, characterized in that, The Temporal Convolutional Network (TCN) has the number of channels set to [512, 1024], ReLU as the activation function, and several fully connected layers.
5. The dynamic safety early warning method for aerial building machines based on multi-source information fusion according to claim 1, characterized in that, The performance evaluation metrics for the model include, but are not limited to, mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²). 2 : in, Indicates the number of samples. Indicates the prediction result. Represents the true value. This represents the average value of the observed data.
6. A dynamic safety early warning system for aerial building construction machines based on multi-source information fusion, used to implement the dynamic safety early warning method for aerial building construction machines based on multi-source information fusion as described in any one of claims 1-5, characterized in that, include: The first main module is used for data acquisition and preprocessing: acquiring feature information of model variables in the sensor, obtaining initial data, and normalizing it; Construct a three-dimensional matrix of sample number, time step, and features; The second main module, namely the graph convolution module, is used to construct a multi-source sensor graph network based on the fast DTW algorithm and the Direct-LiNGAM algorithm, and to perform node embedding and information aggregation in GraphSAGE; The third main module, namely the temporal feature module, includes a temporal convolutional layer, an attention mechanism, and an output layer. The temporal convolutional layer can be used to capture the spatiotemporal correlation of data. The attention mechanism can help the prediction model focus on the input information to improve prediction accuracy. The output layer can reconstruct the prediction data from the latent vectors processed by the attention mechanism according to the length and dimension of the prediction requirements. The fourth main module is used for single-step prediction and multi-step prediction of platform level, as well as feature importance analysis based on the Saliency method.
Citation Information
Patent Citations
DSTGCN-based traffic prediction method
CN115578852A
Large power grid transient power angle stability assessment method based on high-dimensional map attention network
CN116760015A
