A Deep Learning-Based Sensor Optimization Layout Method for Space Frame Construction
By optimizing sensor placement using a deep learning-based approach, combining deep neural networks and the HSIC method, the problem of sensor redundancy is solved, achieving efficient and economical sensor optimization suitable for damage detection in large spatial structures.
Patent Information
- Application Number
- CN202410628525.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-21
- Publication Date
- 2025-10-28
- Estimated Expiration
- 2044-05-21
AI Technical Summary
Existing sensor optimization methods are difficult to achieve comprehensive optimization in large spatial structures, resulting in redundant sensor placement and low computational efficiency, which cannot effectively meet the needs of damage detection.
A deep learning-based approach was adopted, combining the structural morphology and stress information of the steel space frame. The sensor positions were trained through a deep neural network, and redundant sensors were removed by combining the HSIC method to optimize the sensor layout.
It improves the design efficiency and economy of sensor placement, reduces computing costs, and enables rapid and effective sensor optimization, making it suitable for damage detection in large spatial structures.
Smart Images
Figure CN118410714B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of sensor optimization layout technology, and specifically relates to a method for optimizing sensor layout in space frame construction based on deep learning. Background Technology
[0002] In engineering project monitoring systems, sensor systems, as the most reliable monitoring system for the entire project, can quickly reflect the overall health status of the structure. Sensor systems allow for real-time monitoring of the structure, laying the foundation for subsequent work. During the jacking process of a steel space frame, stress concentration is easily generated, leading to deformation of the space frame members. Therefore, it is necessary to understand the deformation and stress conditions of the lifting unit and the installed structure, to issue warnings for potential hazards and accidents, and to ensure the safety of the lifting unit and the installed structure. The arrangement of sensors in the health monitoring system should meet the following two objectives: ① The distribution of sensors can comprehensively reflect the health information of the spatial structure; ② Sensors at the measuring points can quickly detect abnormal changes in the spatial structure. Existing sensor optimization methods mostly aim to minimize the error of the identification parameters. Common optimization methods include: Effective Independence Method (EFI), which ranks sensor positions according to the contribution of each candidate sensor point to the linear independence of the target modal components; Energy Method, which starts from the modal kinetic energy or element strain energy of a certain degree of freedom of the structure and ranks them according to the magnitude of energy to select the sensor position; and Model Reduction Criterion Method, which retains the degrees of freedom where the modal response plays a major role as the measuring point position.
[0003] Existing sensor optimization methods struggle to address overall spatial issues. Different algorithms are based on different theories and have different optimization objectives, leading to varying conclusions under the same structural conditions. For large spatial structures, various uncertainties exist, necessitating the integration of multiple methods to determine the optimal sensor placement based on specific structural characteristics and measurement conditions. Sensor optimization is a multi-objective optimization process. Current sensor placement methods largely focus on maximizing system controllability or objectivity, rather than damage detection. Sensor optimization aimed at damage identification urgently needs to be addressed. Furthermore, most existing sensor optimization methods are based on physical models, resulting in low computational efficiency and a high risk of sensor redundancy. To address these shortcomings, this invention proposes a deep learning-based sensor optimization placement method for space frame construction. Summary of the Invention
[0004] The purpose of this invention is to provide a method for optimizing the arrangement of sensors in space frame construction based on deep learning.
[0005] This invention is achieved through the following measures: a deep learning-based method for optimizing the arrangement of sensors in space frame construction, characterized by comprising,
[0006] A database is constructed using the structural morphology and stress information of the steel space frame as output parameters, and the sensor placement information as output parameters.
[0007] The structural information of the steel space frame and the stress information of the steel space frame are connected as input layer parameters, and the optimized placement information of traditional sensors is used as output layer parameters to build a deep neural network for data training.
[0008] The node information and stress information of the target steel space frame are input into the trained deep neural network to obtain specific sensor spatial position optimization information.
[0009] Furthermore, since the determination of sensor location references three different traditional methods, the results inevitably produce redundancy. Therefore, the method also includes the use of HSIC nonlinear correlation identification method, which uses numerical simulation software to simulate the sensor spatial location optimization information acquisition situation, and uses the HSIC method to remove redundant sensors with highly similar information.
[0010] Furthermore, the HSIC method removes redundant sensors with highly similar information, specifically:
[0011] Once two highly correlated sensor messages are identified, one sensor is immediately and randomly deleted.
[0012] Recalculate the HSIC using global sensor information;
[0013] The process continues iteratively until no high correlation exists among all sensors. Finally, the iteration stops, completing the sensor placement optimization work that considers both information sampling effectiveness and cost-effectiveness.
[0014] Furthermore, a deep neural network is built for data training, including...
[0015] The input and output parameters are stretched into a one-dimensional matrix and concatenated with the one-dimensional matrix of force information to form the overall feature.
[0016] Considering the large dimensionality of the original input and output parameters, a stacked denoising autoencoder is used to reduce the dimensionality of the concatenated data, unifying the data size across different samples and facilitating neural network training. Furthermore, the dimensionality-reduced data better reflects its essential characteristics, leading to improved training performance.
[0017] The reduced steel space frame structure information and the steel space frame stress information are connected as input layer parameters, and the reduced traditional sensor optimized placement information is used as output layer parameters to build a deep neural network for data training.
[0018] Furthermore, in deep neural networks:
[0019] The input layer has 36 nodes, and the output layer has 18 nodes.
[0020] Construct a neural network model with 10 hidden layers, where the number of units in each layer is [25, 20, 18, 15, 15, 15, 15, 20, 15, 8].
[0021] The activation function chosen is the ReLU function, whose range is from 0 to infinity, encompassing the range of input and output data.
[0022] The loss function measures the difference between the predicted value and the true value, and the Euclidean distance between matrices of the same shape is chosen as the loss variable;
[0023] The optimizer uses stochastic gradient descent, with a learning rate of 0.001 and 1600 iterations. The prediction results of the neural network input validation set data are rounded to the nearest whole number to obtain the final prediction value.
[0024] Furthermore, different types and sizes of steel space frame solid structures are constructed in the modeling software and exported to obtain the relative position coordinates of the nodes in the model as the structural morphology information of the steel space frame.
[0025] Furthermore, the main stress points during the lifting process of the steel space frame are obtained through Abaqus numerical simulation, and these main stress points are used as stress information.
[0026] Furthermore, by using three traditional sensor optimization methods—the improved effective independence method, the genetic algorithm, and the Fisher information matrix—the placement of all sensors is integrated, and the placement locations are used as the sensor placement information.
[0027] In one embodiment of this application: a storage medium is provided, characterized in that the storage medium stores one or more programs, which can be executed by one or more processors to implement the method for optimizing the arrangement of space frame construction sensors.
[0028] In one embodiment of this application, an electronic device is provided, characterized in that it includes a processor and a memory, wherein the processor is used to execute a program stored in the memory for a method of optimizing the arrangement of space frame construction sensors, so as to realize the method of optimizing the arrangement of space frame construction sensors.
[0029] The beneficial effects of the technical solution provided by this invention are as follows: It integrates existing advanced physical model-based sensor optimization placement methods, establishes a database of sensor optimization placement positions during the steel space frame lifting process, employs a deep learning model to learn the placement method, and uses a nonlinear correlation method to remove redundant sensors. This overcomes the problems of high computational cost and neglect of economic efficiency in traditional physical model-based sensor optimization design. It uses a data-driven deep learning method for sensor optimization placement design and employs a nonlinear correlation determination method to remove redundant sensors, thereby maximizing design efficiency and reducing economic costs. Furthermore, it has advantages over traditional optimization placement methods in terms of speed, portability, scalability, and effectiveness. Attached Figure Description
[0030] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings listed below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0031] Figure 1 This is a flowchart illustrating the steps of the sensor optimization arrangement method for space frame construction in this embodiment of the invention.
[0032] Figure 2 This is a flowchart of the sensor optimization arrangement method for space frame construction in an embodiment of the present invention;
[0033] Figure 3 This is a sensor arrangement diagram of the steel space frame lifting process in an embodiment of the present invention (blue dots are redundant arrangement points, and red dots are the final determined points);
[0034] Figure 4 This is the overall network architecture of the deep neural network in the embodiments of the present invention;
[0035] Figure 5 This is a diagram showing the training results of the deep neural network in an embodiment of the present invention. Detailed Implementation
[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0037] Example 1:
[0038] See also Figure 1-Figure 5 A method for optimizing the arrangement of sensors in space frame construction based on deep learning, characterized by comprising:
[0039] S1. Using the structural morphology and stress information of the steel space frame as output parameters, and the sensor placement information as output parameters, construct a database;
[0040] Input / output parameter determination: This invention uses a data-driven depth method for sensor optimization instead of the traditional physical model-driven sensor optimization method. The most crucial step in this data-driven approach is building a large-capacity, high-quality database. This study uses 3D3S and Rhino instead of finite element modeling to quickly construct steel space frame structures of different types and sizes. The relative spatial coordinates (x, y, z direction coordinates) of the model nodes are input as a three-row, multi-column matrix. ) is used as one of the input parameters.
[0041] The main stress points during the lifting process of the steel space frame are obtained by numerical simulation using Abaqus. The coordinates of the stress points, the magnitude of the force, and the direction of the force are used as another input feature parameter and set as a one-dimensional matrix.
[0042] Three conventional sensor optimization methods were used: the improved effective independence method, the genetic algorithm, and the Fisher information matrix. The strain sensor position information obtained by the three methods was integrated as the output parameters of the database, and the output was also a matrix with three rows and multiple columns.
[0043] S2. Connect the steel space frame structure information and the steel space frame stress information as input layer parameters, and use the optimized arrangement information of traditional sensors as output layer parameters to build a deep neural network for data training.
[0044] In this invention, the input and output parameters of the database may have the same number of columns but different numbers of rows due to variations in the steel space frame structure, resulting in different input and output parameter sizes and hindering training. This invention stretches the input and output parameters into a one-dimensional matrix (arranged horizontally in x, y, z order) and concatenates it with a one-dimensional matrix of force information as a unified feature. A stacked noise-reducing autoencoder (SDA) is used to process the input and output parameters, which not only uncovers the essential features of the parameters but also reduces the dimensionality of the data, accelerating training efficiency while standardizing data size. Based on actual usage (mainly the number of steel space frame nodes and the number of sensors), the stacked noise-reducing autoencoder encodes the input parameters into a 36-column one-dimensional array and the output parameters into an 18-column one-dimensional array (this can be adjusted according to the specific project requirements).
[0045] A deep neural network was chosen as the main training model, with the following special features: the input layer had 36 nodes and the output layer had 18 nodes; a neural network model with 10 hidden layers was constructed, with the number of units in each layer being [25, 20, 18, 15, 15, 15, 15, 20, 15, 8]; the ReLU activation function was chosen, with a range from 0 to infinity, encompassing the range of input and output data in this study; the loss function measures the difference between the predicted and true values, and the Euclidean distance between matrices of the same shape was chosen as the loss variable; the optimizer was chosen as stochastic gradient descent, with a learning rate of 0.001 and 1600 iterations, and the prediction results of the neural network input validation set data were rounded to the nearest whole number to obtain the final prediction value.
[0046] S3. Input the node information and force information of the target steel space frame into the trained deep neural network to obtain specific sensor spatial position optimization information. The trained deep neural network can obtain the sensor placement information based on the space frame node coordinate information and force information. The output matrix is decoded by a stacked noise-reducing autoencoder to output the specific spatial position of the sensor.
[0047] S4. The number of sensors output by the trained network is based on the integration of three traditional sensor optimization methods. Therefore, there will be a certain range of sensor duplication and sensor information redundancy. This invention explores the nonlinear correlation between the received information of the deployed sensors using the HSIC method. Employing the nonlinear correlation identification method of HSIC, and using numerical simulation software to simulate the sensor spatial location optimization information acquisition, two highly correlated sensors will be randomly retained, while the redundant one will be removed. To prevent excessive redundancy removal, this process only considers the correlation between two sensors, removing only one redundant sensor at a time, iterating continuously until there is no high correlation among all sensors globally. This completes the sensor optimization deployment considering information sampling effectiveness and economy.
[0048] This invention provides a solution to the problem of optimizing the placement of structural sensors using deep learning methods based on a large dataset. The database considers different grid structures and various traditional sensor optimization objectives. Therefore, the sensor placement results obtained during training also take into account the comprehensiveness and accuracy of sensor monitoring of structural deformation. In designing the deep neural network, input and output parameter design issues were considered, and a series of operations such as stretching and dimensionality reduction were performed to improve the efficiency and rationality of network training. Furthermore, the output results of the neural network, i.e., the placement and number of sensors, undergo further redundancy removal, significantly reducing economic costs. After training in its neural network, the deep learning-based sensor placement optimization method calculates the sensor placement for new structures much faster than traditional methods (no more than 1 second). This speedup effect is even more pronounced when dealing with large structures.
[0049] Example 2:
[0050] This application also provides an electronic device, characterized in that it includes: a processor and a memory, wherein the processor is used to execute a program stored in the memory for a method of optimizing the arrangement of space frame construction sensors, so as to realize the method of optimizing the arrangement of space frame construction sensors.
[0051] An electronic device includes at least one processor, memory, at least one network interface, and other user interfaces. The various components of the electronic device are coupled together via a bus system. It is understood that the bus system is used to enable communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.
[0052] The user interface may include a display, keyboard, or clicking device (e.g., mouse, trackball, touchpad, or touchscreen). It is understood that the memory in this embodiment may be volatile memory or non-volatile memory, or may include both.
[0053] In this embodiment of the invention, the processor executes the method steps provided in each method embodiment by calling a program or instruction stored in the memory, specifically a program or instruction stored in an application program.
[0054] In some implementations, the memory stores elements such as executable units or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.
[0055] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. The program implementing the method of this invention can be included in the application programs.
[0056] Example 3:
[0057] This application also provides a storage medium, characterized in that the storage medium stores one or more programs, which can be executed by one or more processors to implement a method for optimizing the arrangement of sensors in space frame construction.
[0058] The method steps described in Embodiment 1 disclosed herein can be implemented using hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for optimizing the arrangement of sensors in space frame construction based on deep learning, characterized in that, include, A database is constructed using the structural morphology and stress information of the steel space frame as output parameters, and the sensor placement information as output parameters. The structural information of the steel space frame and the stress information of the steel space frame are connected as input layer parameters, and the optimized placement information of traditional sensors is used as output layer parameters to build a deep neural network for data training. The node information and stress information of the target steel space frame are input into the trained deep neural network to obtain specific sensor spatial position optimization information; It also includes a nonlinear correlation identification method using HSIC, which uses numerical simulation software to simulate the sensor spatial location optimization information acquisition, and uses the HSIC method to remove redundant sensors with highly similar information; The HSIC method removes redundant sensors with highly similar information, specifically: Once two highly correlated sensor messages are identified, one sensor is immediately and randomly deleted. Recalculate the HSIC using global sensor information; Iterate continuously until there is no high correlation among all sensors; Building a deep neural network for data training includes, The input and output parameters are stretched into a one-dimensional matrix and concatenated with the one-dimensional matrix of force information to form the overall feature. Stacked noise reduction autoencoders are used to reduce the dimensionality of the spliced data, thus unifying the data size of different samples; The reduced steel space frame structure information and the steel space frame stress information are connected as input layer parameters, and the reduced traditional sensor optimized placement information is used as output layer parameters to build a deep neural network for data training. In deep neural networks: The input layer has 36 nodes, and the output layer has 18 nodes. Construct a neural network model with 10 hidden layers, where the number of units in each layer is [25, 20, 18, 15, 15, 15, 15, 20, 15, 8]. The activation function chosen is the ReLU function, whose range is from 0 to infinity, encompassing the range of input and output data. The loss function measures the difference between the predicted value and the true value, and the Euclidean distance between matrices of the same shape is chosen as the loss variable; The optimizer is stochastic gradient descent, with a learning rate of 0.001 and 1600 iterations. The prediction results of the neural network input validation set data are rounded to the nearest whole number to obtain the final prediction value. Construct and export steel space frame solid structures of different types and sizes in modeling software, and obtain the relative position coordinates of the nodes in the model as the structural morphology information of the steel space frame; The main stress points during the lifting process of the steel space frame are obtained through Abaqus numerical simulation, and these main stress points are used as stress information.
2. The method for optimizing the arrangement of sensors in space frame construction according to claim 1, characterized in that, This paper uses three traditional sensor optimization methods—the improved effective independence method, the genetic algorithm, and the Fisher information matrix—to integrate all sensor placement locations and use these locations as sensor placement information.
3. A storage medium, characterized in that, The storage medium stores one or more programs, which can be executed by one or more processors to implement the sensor optimization arrangement method for space frame construction as described in any one of claims 1-2.
4. An electronic device, characterized in that, include: A processor and a memory, the processor being configured to execute a program stored in the memory for a method of optimizing the arrangement of sensors for space frame construction, to implement the method of optimizing the arrangement of sensors for space frame construction as described in any one of claims 1-2.