Strain sensor arrangement method based on information entropy and neural network input

Abstract

The application provides a strain sensor arrangement method based on information entropy and neural network input, relates to the technical field of physical load strain detection, and solves a series of limitation problems of current sensor arrangement schemes due to lack of scientific optimization mechanism suitable for input requirements of intelligent identification system. The method constructs a finite element model of a target structure, extracts strain response data under a benchmark load condition, calculates strain information entropy of candidate positions, and obtains an optimized candidate measuring point set after screening through engineering constraint conditions. Then, a plurality of load samples are generated, finite element analysis is performed, the optimized candidate measuring points are obtained, and the strain response data of the measuring points at this time are concentrated to construct a feature matrix. The minimum singular value of the feature matrix of the measuring point subset is maximized as an optimization target, a preset number of target measuring points are selected, and a strain sensor arrangement scheme is obtained. The application considers engineering constraints and information theory optimization, and can achieve target precision with fewer strain sensors.

Description

Technical Field

[0001] This invention relates to the field of physical load strain detection technology, and specifically to a method for arranging strain sensors based on information entropy and neural network input. Background Technology

[0002] In the field of structural health monitoring and load identification, strain sensors, as key sensing units, directly affect the reliability and intelligence level of subsequent data analysis through their placement. Traditional placement strategies often rely on engineers' experience, selecting points at symmetrical locations or based on uniform grids. While this offers some engineering intuition, it fails to systematically quantify the ability of each measuring point to represent the load state. With the deepening application of data-driven technologies, especially the widespread deployment of neural networks in load inversion, the quality of input features has become a crucial factor determining model accuracy and generalization. Existing experience-based point placement methods often result in redundant sensor numbers, uneven information density, or insufficient capture of key load features, failing to meet the urgent need of intelligent identification systems for high-quality, high-information-volume input data.

[0003] Sensor optimization placement technology has evolved over many years, resulting in various technical approaches. Some methods originate from modal testing, selecting measurement points by analyzing structural vibration modes, kinetic energy distribution, or modal confidence criteria, focusing on improving the accuracy of dynamic parameter identification. Other technologies target damage-sensitive areas, optimizing point placement based on strain energy changes or probabilistic models to enhance damage detection capabilities. From an information theory perspective, strategies such as mutual information and Bayesian experimental design attempt to improve the information gain of parameter estimation from a statistical standpoint. Regularized placement based on mechanical principles is also common in engineering practice. These methods have accumulated rich experience in their respective application scenarios, demonstrating a multi-dimensional exploration from physical mechanisms to statistical inference.

[0004] However, when the goal shifts to constructing high-quality input features for neural network load recognition systems, existing technologies exhibit significant limitations. Their optimization objectives often deviate from the global representation requirements of static or quasi-static loads, failing to fully meet the inherent requirements of neural networks for input data, such as low redundancy among features, numerical range adaptability, and balanced information distribution. Optimization results under single-condition or simplified load assumptions are insufficient to cover the complex and varied load combinations in reality, easily leading to information blind spots in certain conditions. Simultaneously, theoretical optimization often neglects engineering implementation constraints, such as the installability of measuring points and whether strain signals are within the effective range of sensors, making implementation difficult. Furthermore, existing processes often separate physical simulation from actual load data, lacking a framework for deep integration of finite element analysis and multi-condition sampling, and lacking quantitative evaluation and iterative optimization mechanisms for input feature quality, making it difficult to achieve synergistic improvements in information effectiveness, model adaptability, and engineering feasibility of the deployment scheme.

[0005] In summary, current sensor deployment technologies still face multiple challenges when applied to data-driven intelligent load identification tasks, including a disconnect between optimization guidance and application scenarios, insufficient assurance of input feature quality, inadequate load spatial coverage, and shallow integration of physical and data aspects. The industry urgently needs a new deployment approach that can scientifically quantify the value of measurement point information, conform to the input characteristics of machine learning models, and balance theoretical rigor with engineering practicality, to provide reliable support for the intelligent upgrade of structural sensing systems. Summary of the Invention

[0006] The purpose of this invention is to address the technical problems of existing strain sensor placement methods in structural load identification applications, which suffer from redundant sensor configurations, insufficient representation of key load information, and difficulty in coordinating the effectiveness of measurement point data with engineering applicability due to the lack of a scientific optimization mechanism adapted to the input requirements of intelligent identification systems. This invention proposes a strain sensor placement method based on information entropy and neural network input. It transforms the traditional sensor placement problem into a machine learning feature selection problem. Through a two-stage fusion framework of physical selection and data-driven optimization, it balances engineering constraints with information theory optimization, achieving the target accuracy with fewer strain sensors and reducing hardware and data processing costs.

[0007] The present invention employs the following technical solutions to achieve its objective: A method for arranging strain sensors based on information entropy and neural network input includes the following steps: S1. Construct a finite element model of the target structure; apply multiple preset reference load conditions to the finite element model, and extract the first strain response data of each candidate position on the surface of the target structure under the reference load conditions; S2. Based on the first strain response data, calculate the strain information entropy of each candidate location; and filter the candidate locations in conjunction with engineering constraints to form an optimized candidate measurement point set. S3. Generate multiple load samples within a preset load space; analyze the finite element model based on the load samples, obtain the second strain response data of each measuring point in the optimized candidate measuring point set under the action of the load samples, and construct a feature matrix, wherein the rows of the feature matrix correspond to the measuring points, and the columns correspond to the combination of load samples and strain components. S4. With the optimization objective of maximizing the minimum singular value of the feature matrix corresponding to the selected subset of measurement points, a preset number of target measurement points are selected from the optimized candidate measurement point set, and the spatial position of the target measurement points and the corresponding strain measurement direction are output as the arrangement scheme of the strain sensor.

[0008] Specifically, in step S1, several preset reference load conditions include: along the structural coordinate system , and Unit force load applied in three orthogonal directions , and and around , and Unit torque load applied along the three orthogonal axes , and ; The first strain response data consists of the strain tensor components corresponding to each candidate location under all reference load conditions, wherein the strain tensor components include those in the local coordinate system. Directional strain components , Directional strain components and in-plane shear strain components .

[0009] Furthermore, in step S2, the strain information entropy The calculation expression is as follows:

[0010] In the formula, For candidate position index, This is an index for the combination of the reference load condition and strain components. It is the product of the total number of reference load cases and the number of strain components under a single reference load case; The normalized strain response weights are as follows:

[0011] In the formula, Representing the The candidate position is in the Strain response under various load combinations Representing the The candidate position is in the Strain response under various load combinations This is the index number for this location.

[0012] Preferably, in step S2, the engineering constraints include strain amplitude constraints and installation feasibility constraints; The strain amplitude constraint includes: based on the analysis results of the preset maximum load condition, eliminating candidate positions where the strain value exceeds the upper limit of the sensor range or the linear elastic strain threshold of the material, and eliminating candidate positions where the absolute strain value is lower than the preset noise threshold under all preset load conditions. The installation feasibility constraints include: based on the target structure geometric model, identifying infeasible areas including rounded corners, chamfers, hole edges, weld areas, or surface curvature radii smaller than a preset value, and eliminating candidate positions located in the infeasible areas.

[0013] Preferably, in step S3, the load samples are generated using a super Latin cube sampling method, which includes: dividing the value range of each load component in the load space into equally probable sub-intervals with a number equal to the total number of samples; randomly generating a numerical point in each sub-interval; and randomly arranging and combining the numerical points of each load component to form a load sample set covering the load space.

[0014] Specifically, in step S3, when constructing the feature matrix, the second strain response data consists of the strain tensor components corresponding to each measuring point in the optimized candidate measuring point set under all load samples. The type of the strain tensor components is the same as that of the strain tensor components in the first strain response data, both including the strain tensor components in the local coordinate system. Directional strain components , Directional strain components and in-plane shear strain components ; Each measuring point is arranged sequentially according to the sample order under all load samples. , and Formation dimension is eigenvectors, The total number of load samples; the number of rows in the feature matrix is ​​equal to the number of measurement points in the optimized candidate measurement point set, and the number of columns is equal to 3 and the total number of load samples. The product of.

[0015] Furthermore, in step S4, a greedy iterative algorithm is used to select a preset number of target measurement points from the optimized candidate measurement point set. When the greedy iterative algorithm is initialized, the measurement point with the highest strain information entropy value is added to the selected measurement point set, and an initial feature matrix is ​​constructed. In a preset number of iterations, all unselected measurement points are traversed, and the minimum singular value of the temporary feature matrix formed after adding them to the current selected measurement point set is calculated. The unselected measurement point that maximizes the increment of the minimum singular value is selected and added to the selected measurement point set, and the corresponding feature matrix is ​​updated.

[0016] Specifically, in the greedy iterative algorithm, the increment of the minimum singular value is defined as the minimum singular value of the temporary feature matrix. The minimum singular value of the current characteristic matrix The difference; the preset number of iterations is the number of strain sensors corresponding to the feature dimensions required by the neural network input layer, and the preset number of iterations is less than the total number of measurement points in the optimized candidate measurement point set; the number of rows of the feature matrix updated after each iteration corresponds to the number of selected measurement points, and the number of columns corresponds to the product of the total number of load samples and the total number of strain components.

[0017] Specifically, in step S4, the output strain measurement direction includes three preset angular directions, namely... , , The three preset angular directions correspond to the grid pasting orientation of the triaxial strain rosette; the third strain response data collected by each target measuring point during actual measurement is composed of the strain values ​​measured along the three preset angular directions.

[0018] Specifically, the number of measurement points included in the optimized candidate measurement point set formed in step S2, and the preset number of target measurement points in step S4, are determined based on the geometric complexity of the target structure, the load space dimension, and the configuration requirements of the neural network input layer nodes.

[0019] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention significantly improves the quality and effectiveness of input data for neural network load identification tasks. The determined sensor configuration can comprehensively capture the key strain characteristics of structures under complex load combinations, effectively enhancing the prediction accuracy and environmental adaptability of the load identification model. The optimized input features possess good numerical distribution characteristics and a low-redundancy structure, providing a stable and high-quality data foundation for neural network training, significantly improving model convergence efficiency and reducing the risk of overfitting, making the intelligent identification system more reliable in actual working conditions.

[0020] This invention combines theoretical analysis with on-site implementation requirements, automatically avoiding areas with signal distortion risks and physical installation obstacles, ensuring that the selected measuring points possess both measurement effectiveness and engineering feasibility. The layout scheme conforms to the sensor's operating range and the structural material response characteristics, effectively guaranteeing the signal-to-noise ratio and long-term stability of the acquired data, significantly reducing on-site debugging and subsequent maintenance burdens, and improving the operational reliability and service life of the entire monitoring system.

[0021] This invention constructs a systematic design process that covers the load space and integrates multi-source constraints, eliminating excessive reliance on manual experience and achieving the scientific and standardized generation of sensor placement schemes. Its technical framework possesses good structural adaptability and load generalization capabilities, and can be flexibly transferred to various engineering structures and monitoring scenarios. By maximizing information acquisition efficiency with a streamlined and reasonable number of sensors, it significantly reduces hardware deployment costs and data processing complexity while ensuring recognition performance, providing a practical and feasible technical path for the intelligent, economical, and large-scale application of structural health monitoring systems. Attached Figure Description

[0022] The present invention is described in detail with reference to the following figures, which include six figures as follows: Figure 1 This is a schematic diagram illustrating the overall process of the strain sensor arrangement method of the present invention; Figure 2 This is a schematic diagram of the strain information entropy sorting of the initially selected measurement points in this invention example; Figure 3 This is a schematic diagram illustrating the change of the minimum singular value during the execution of the greedy iterative algorithm in this invention example; Figure 4 This is a schematic diagram illustrating the selection change from the original node to the final measurement point in this invention. Figure 5 This is a schematic diagram illustrating the prediction effect of the original node features in an example of the present invention. Figure 6 This is a schematic diagram illustrating the prediction effect of the final measurement point features in an example of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] A method for arranging strain sensors based on information entropy and neural network input. Figure 1 The overall process of this method is briefly described below and can be viewed concurrently; the key steps of this method can be summarized as follows: S1. Construct a finite element model of the target structure; apply multiple preset reference load conditions to the finite element model, and extract the first strain response data of each candidate position on the surface of the target structure under the reference load conditions; S2. Based on the first strain response data, calculate the strain information entropy of each candidate location; combine the engineering constraints to screen the candidate locations and form an optimized candidate measurement point set; S3. Generate multiple load samples within the preset load space; analyze the finite element model based on the load samples, obtain the second strain response data of each measuring point in the optimized candidate measuring point set under the action of the load samples, and construct a feature matrix. The rows of the feature matrix correspond to the measuring points, and the columns correspond to the combination of load samples and strain components. S4. With the optimization objective of maximizing the minimum singular value of the feature matrix corresponding to the selected subset of measurement points, select a preset number of target measurement points from the optimized candidate measurement point set, and output the spatial position of the target measurement points and the corresponding strain measurement direction as the arrangement scheme of the strain sensor.

[0026] The method described in this embodiment involves a two-stage fusion approach, specifically designed for high-quality input feature acquisition in neural network-based load identification systems, i.e., how to arrange strain sensors. The two stages are: steps S1 to S2, which involve constructing and initially screening the physical information database; and steps S3 to S4, which involve final optimization and selection based on the neural network. Additionally, a third application verification stage may be included to evaluate and verify the strain sensor arrangement scheme obtained in step S4, allowing for further optimization. This embodiment will provide a detailed and optimized description of the above method process.

[0027] In step S1 of this embodiment, a parametric finite element model of the target structure is first established, which involves multiple reference load conditions including: along the structural coordinate system. , and Unit force load applied in three orthogonal directions , and and around , and Unit torque load applied along the three orthogonal axes , and The six loads here are the basic unit loads. After performing static analysis on them, the data related to the unit load response can be obtained.

[0028] Simultaneously, the first strain response data of each candidate location on the target structure surface under the reference load condition are the strain tensor components corresponding to each candidate location under all reference load conditions. The strain tensor components include those in the local coordinate system. Directional strain components , Directional strain components and in-plane shear strain components After the strain tensor components at all candidate locations are extracted, they are combined with the aforementioned six basic unit loads, for example... Directional unit force load Combining these terms with the three strain tensor components yields three combined terms, which together form a sensitivity matrix on the target structure surface encompassing all candidate locations. , It represents the set of real numbers; the total number of combinatorial terms is 18, which serve as columns of the matrix; The total number of candidate positions is represented by rows in the matrix. Essentially, a candidate position is equivalent to an original node that could potentially be used as a test point.

[0029] In step S2 of this embodiment, for the nodes at the above candidate positions... Its unit load response vector is denoted as This can describe the node. Sensitive modes to base loads. This implementation calculates each candidate location (i.e., node) here. Strain information entropy The calculation expression is as follows:

[0030] In the formula, For candidate position index, This is an index for the combination of the reference load condition and strain components. In this embodiment, the product of the total number of reference load cases and the number of strain components under a single reference load case is used. That is, the total number of combination items, which is 18; The normalized strain response weights are as follows:

[0031] In the formula, Representing the The candidate position is in the Strain response under various load combinations Representing the The candidate position is in the Strain response under various load combinations This is the index number at this location. This implementation calculates the candidate location nodes under the original condition. The strain information entropy, nodes with high entropy values ​​have significant and differentiated responses to various loads, thus having great information potential and are the preferred measurement points for subsequent strain sensor deployment.

[0032] Before proceeding to the final optimization selection for the neural network, this implementation method first needs to screen the candidate positions in conjunction with engineering constraints to form an optimized candidate measurement point set; in step S2, the engineering constraints include strain amplitude constraints and installation feasibility constraints.

[0033] The strain amplitude constraint is determined by applying the expected maximum load combination. The specific constraint includes: based on the analysis results of the preset maximum load condition, eliminating candidate positions where the strain value exceeds the upper limit of the sensor range or the linear elastic strain threshold of the material, and eliminating candidate positions where the absolute strain value is lower than the preset noise threshold under all preset load conditions.

[0034] The installation feasibility constraint mainly considers whether the final triaxial strain rose can be pasted at specific measurement points on the target structure. Therefore, the corresponding pastability constraint is determined based on the CAD model. The constraint includes: based on the geometric model of the target structure, identifying infeasible areas including fillets, chamfers, hole edges, weld areas, or surface curvature radii smaller than preset values, and eliminating candidate positions located in infeasible areas.

[0035] Then, based on the previously calculated candidate location nodes... Strain information entropy The nodes are sorted in descending order, and only the top-ranked nodes corresponding to the required number of candidate positions are retained, thus forming the optimized candidate measurement point set. In this embodiment, this is equivalent to the first measurement point screening. The number of measurement points to be retained can be determined based on the geometric complexity of the target structure, the load space dimension, and the configuration requirements of the neural network input layer nodes.

[0036] In the second stage, step S3 of this embodiment generates load samples using a super Latin cube sampling method, including: dividing the value range of each load component in the load space into equally probable sub-intervals with a number equal to the total number of samples; randomly generating a numerical point in each sub-interval; and randomly arranging and combining the numerical points of each load component to form a load sample set covering the load space.

[0037] The load components here are of the same type as the reference load condition in step S1, and are also... , , , , and These six load components are obtained from the load space and a random generation method, with the value range and specific numerical points of each component determined by the load space. Furthermore, a load sample is formed by combining the numerical points of these six load components. After generating a sufficient number of load samples, this serves as the basis for finite element analysis. Following finite element analysis, all measurement points in the optimized candidate measurement point set yield corresponding second strain response data for each load sample, which can be used to construct a characteristic matrix.

[0038] In this embodiment, the second strain response data consists of the strain tensor components corresponding to each measuring point in the optimized candidate measuring point set under all load samples. The type of strain tensor components is the same as that in the first strain response data, both including the strain tensor components in the local coordinate system. Directional strain components , Directional strain components and in-plane shear strain components .

[0039] Each measuring point is arranged sequentially according to the sample order under all load samples. , and Formation dimension is eigenvectors, This is the total number of payload samples. Each load sample should be representative in order to optimize the number of measurement points in the candidate measurement point set. For example, the corresponding constructed feature vector (It can also be viewed as a matrix with 1 row) and can be recorded as follows:

[0040] The feature matrix can be formed by optimizing the feature vectors of all measurement points in the candidate measurement point set. , among them To optimize the number of measurement points in the candidate measurement point set, the number of rows in the feature matrix is ​​equal to the number of measurement points in the optimized candidate measurement point set, and the number of columns is equal to 3 and the total number of load samples. The product of.

[0041] In step S4 of this embodiment, a greedy iterative algorithm is used to select a preset number of target measurement points from the optimized candidate measurement point set. The specific objective function is: to maximize the measurement point matrix selected from the optimized candidate measurement point set. The amount of information, that is, maximizing the matrix of the selected measurement points. Minimum singular value This is equivalent to minimizing the condition number. .

[0042] The specific process of the greedy iterative algorithm is as follows: During the initialization of the greedy iterative algorithm, the measurement point with the highest strain information entropy value is selected. Add the selected measurement points to the set and construct the initial feature matrix, which is the feature matrix of the selected measurement point set at this point. At this point, the target number of strain sensors can be preset, typically with a minimum value of 2. Assuming the preset target number of strain sensors is 20, meaning the final solution requires 20 strain sensors, and the selected measurement points... We need to select 19 measurement points to set up, which means we need to perform 19 more iterations.

[0043] In these 19 iterations, all unselected measurement points at each iteration are traversed. Calculate the temporary feature matrix formed after adding it to the currently selected measurement point set. Minimum singular value Finally, the unselected test point that maximizes the increment of the minimum singular value is added to the set of selected test points, and the corresponding feature matrix is ​​updated. .

[0044] In the greedy iterative algorithm, the increment of the minimum singular value is defined as the minimum singular value of the temporary characteristic matrix. The minimum singular value of the current characteristic matrix The difference; the preset number of iterations is the number of strain sensors corresponding to the feature dimensions required by the neural network input layer, i.e., 20 iterations in the example above, including the first initialization, and the preset number of iterations should be less than the total number of measurement points in the optimized candidate measurement point set; this is the second measurement point selection based on the optimized candidate measurement point set, which can also be determined according to the geometric complexity of the target structure, the load space dimension, and the node configuration requirements of the neural network input layer. The number of rows in the updated feature matrix after each iteration corresponds to the number of selected measurement points, and the number of columns corresponds to the product of the total number of load samples and the total number of strain components.

[0045] The mathematical principle applied in this embodiment of the greedy iterative algorithm is as follows: maximizing the minimum singular value can ensure that the feature matrix is ​​far from the singular, and make the directions of each feature vector as orthogonal as possible, thereby providing the neural network with complementary information and numerically stable input.

[0046] In step S4 of this embodiment, the output strain measurement direction includes three preset angular directions, namely... , , The three preset angular directions correspond to the grid placement orientation of the triaxial strain rosette; the triaxial strain rosette is a mature device in this field, which has three strain measurement units, each corresponding to an independent measurement component, which in this embodiment correspond to... , , These three local coordinate angles; in this embodiment, the grid pasting orientation refers to the sensitive extension direction of each strain measurement unit in the triaxial strain rosette, that is, the sensing direction of each strain measurement unit. The third strain response data collected by each target measuring point during actual measurement is composed of the strain values ​​measured along three preset angular directions.

[0047] Finally, in the third stage, application verification is conducted. Based on the optimized strain sensor arrangement, triaxial strain rosettes are attached to corresponding positions on the target structure or test specimen. Strain loads are applied, strain data is collected, and a dataset for neural network training is constructed. The neural network can be a typical multilayer perceptron (MLP) neural network, trained using the optimized dataset. When evaluating the arrangement scheme of this embodiment, it can be compared with conventional arrangements such as random and uniform arrangements, and the evaluation is based on the root mean square error (RMSE) and coefficient of determination. The evaluation indicators such as training convergence speed and robustness to noise were analyzed.

[0048] This embodiment will now use more specific example data to illustrate the application of the method.

[0049] Taking a cantilever beam with a circular cross-section as an example, its length is 0.5m, its cross-section is 0.1m × 0.1m, and its material is steel. The design load space is as follows: ;

[0050] Using hypercubic Latin sampling, 3000 load samples were obtained. Then, following the traditional point placement method that relies on expert experience, three cross sections were selected, and eight original nodes were arranged on each cross section. Each node required strain data in three directions to be measured using a triaxial strain rosette, resulting in a total of 3×8×3=72 features. The neural network that requires strain data was modeled using an adaptive multilayer perceptron.

[0051] Corresponding to step S2 in the method, a feasibility screening is performed on a total of 24 original nodes to remove nodes with stress concentration, geometric abrupt changes, or strain absolute values ​​that are outside the preset range. An example upper limit threshold is 0.00095. Seven nodes are removed as an example: Node 4 has an absolute strain value of 0.000959 > 0.00095, so it is discarded. Node 7 has an absolute strain value of 0.000951 > 0.00095, therefore it is discarded. Node 16 has an absolute strain value of 0.000960 > 0.00095, therefore it is discarded. Node 19 has an absolute strain value of 0.000951 > 0.00095, therefore it is discarded. Nodes 9, 11, and 17 are located in the area where pasting is not allowed, so they are removed. Thus, a total of 17 candidate location nodes remain as measurement points awaiting a second screening, forming an optimized candidate measurement point set.

[0052] Based on the strain information entropy calculation in step S2 of the method, the 17 measurement points in the optimized candidate measurement point set are arranged in descending order, and the arrangement result is as follows. Figure 2 As shown. Figure 2 In the diagram, the node numbers on the horizontal axis are the original sequence numbers of the original 24 nodes.

[0053] After performing steps S3 and S4, the change in the minimum singular value of the feature matrix is ​​calculated using a greedy iterative algorithm, as follows: Figure 3 As shown. By Figure 3 As can be seen, the minimum singular value decreases rapidly as the number of selected measurement points increases. After the fifth measurement point is selected, the minimum singular value has dropped to near zero. When the eighth measurement point is selected, the change in the minimum singular value has basically leveled off. Therefore, these eight measurement points are sufficient, and these eight measurement points are the optimized strain sensor arrangement scheme.

[0054] Figure 4 The process from the initial 24 nodes to the final 8 selected measurement points is illustrated. The original indices of the 8 remaining measurement points are 2, 6, 8, 10, 13, 15, 21, and 22. Based on the established neural network, models are built for all features involved in the initial 24 nodes and the features of the final 8 selected measurement points, with corresponding feature counts of 72 and 24, respectively. For the complete set of features, such as... Figure 5 As shown, the determination coefficient of its true Y load The features selected through the method, such as Figure 6 As shown, its true Y-load determination coefficient A comparison of the prediction results of the two methods shows that this implementation method, through two rounds of selection, eliminates features of low information entropy measurement points, which can make the network structure smaller and maintain or even slightly improve the prediction accuracy.

Claims

1. A method for arranging strain sensors based on information entropy and neural network input, characterized in that, Includes the following steps: S1. Construct a finite element model of the target structure; apply multiple preset reference load conditions to the finite element model, including loads along the structural coordinate system. , and Unit force load applied in three orthogonal directions , and and around , and Unit torque load applied along the three orthogonal axes , and ; The first strain response data of each candidate location on the surface of the target structure under the reference load condition is extracted. The first strain response data is the strain tensor component corresponding to each candidate location under all reference load conditions. The strain tensor component includes the strain tensor components in the local coordinate system. Directional strain components , Directional strain components and in-plane shear strain components ; S2. Based on the first strain response data, calculate the strain information entropy at each candidate location. The calculation expression is as follows: In the formula, For candidate position index, This is an index for the combination of the reference load condition and strain components. It is the product of the total number of reference load cases and the number of strain components under a single reference load case; The normalized strain response weights are as follows: In the formula, Representing the The candidate position is in the Strain response under various load combinations Representing the The candidate position is in the Strain response under various load combinations This is the index number at this location; The candidate locations are screened in combination with engineering constraints. During the screening, the nodes corresponding to each candidate location are sorted in descending order based on the calculated strain information entropy of each node. Only the required number of nodes corresponding to the top-ranked candidate locations are retained to form an optimized set of candidate measurement points. S3. Generate multiple load samples within the preset load space; The finite element model is analyzed based on the load sample to obtain the second strain response data of each measuring point in the optimization candidate measuring point set under the action of the load sample and to construct a feature matrix. The rows of the feature matrix correspond to the measuring points, and the columns correspond to the combination of load sample and strain components. S4. With the optimization objective of maximizing the minimum singular value of the feature matrix corresponding to the selected subset of measurement points, a greedy iterative algorithm is adopted to select a preset number of target measurement points from the optimized candidate measurement point set. When the greedy iterative algorithm is initialized, the measurement point with the highest strain information entropy value is added to the selected measurement point set to construct an initial feature matrix. Finally, the spatial location of the target measuring point and the corresponding strain measurement direction are output as the arrangement scheme of the strain sensor; The number of measurement points included in the optimized candidate measurement point set formed in step S2, and the preset number of target measurement points in step S4, are determined based on the geometric complexity of the target structure, the load space dimension, and the configuration requirements of the neural network input layer nodes, respectively; the preset number of iterations of the greedy iteration algorithm in step S4 is the number of strain sensors corresponding to the feature dimension required by the neural network input layer.

2. The strain sensor arrangement method according to claim 1, characterized in that: In step S2, the engineering constraints include strain amplitude constraints and installation feasibility constraints; The strain amplitude constraint includes: based on the analysis results of the preset maximum load condition, eliminating candidate positions where the strain value exceeds the upper limit of the sensor range or the linear elastic strain threshold of the material, and eliminating candidate positions where the absolute strain value is lower than the preset noise threshold under all preset load conditions. The installation feasibility constraints include: based on the target structure geometric model, identifying infeasible areas including rounded corners, chamfers, hole edges, weld areas, or surface curvature radii smaller than a preset value, and eliminating candidate positions located in the infeasible areas.

3. The strain sensor arrangement method according to claim 1, characterized in that, In step S3, the load samples are generated using the super Latin cube sampling method, which includes: dividing the value range of each load component in the load space into equal probability sub-intervals with a number equal to the total number of samples; randomly generating a numerical point in each sub-interval; and randomly arranging and combining the numerical points of each load component to form a load sample set covering the load space.

4. The strain sensor arrangement method according to claim 3, characterized in that: In step S3, when constructing the feature matrix, the second strain response data consists of the strain tensor components corresponding to each measuring point in the optimized candidate measuring point set under all load samples. The type of the strain tensor components is the same as that of the strain tensor components in the first strain response data, both including the strain tensor components in the local coordinate system. Directional strain components , Directional strain components and in-plane shear strain components ; Each measuring point is arranged sequentially according to the sample order under all load samples. , and Formation dimension is eigenvectors, This represents the total number of payload samples. The number of rows in the feature matrix is ​​equal to the number of measurement points in the optimized candidate measurement point set, and the number of columns is equal to 3 and the total number of load samples. The product of.

5. The strain sensor arrangement method according to claim 1, characterized in that: In step S4, during a preset number of iterations, all unselected test points are traversed, and the minimum singular value of the temporary feature matrix formed after adding them to the current set of selected test points is calculated. The unselected test point that maximizes the increment of the minimum singular value is selected and added to the set of selected test points, and the corresponding feature matrix is ​​updated.

6. The strain sensor arrangement method according to claim 5, characterized in that: In the greedy iterative algorithm, the increment of the minimum singular value is defined as the minimum singular value of the temporary characteristic matrix. The minimum singular value of the current characteristic matrix The difference; the preset number of iterations is the number of strain sensors corresponding to the feature dimensions required by the neural network input layer, and the preset number of iterations is less than the total number of measurement points in the optimized candidate measurement point set; The number of rows in the updated feature matrix after each iteration corresponds to the number of selected measurement points, and the number of columns corresponds to the product of the total number of load samples and the total number of strain components.

7. The strain sensor arrangement method according to claim 1, characterized in that: In step S4, the output strain measurement direction includes three preset angular directions, namely... , , The three preset angular directions correspond to the grid pasting orientation of the triaxial strain rosette; the third strain response data collected by each target measuring point during actual measurement is composed of the strain values ​​measured along the three preset angular directions.

Images