Methods, apparatus, and articles of manufacture to physically steer enhanced multiple unknown dynamic load positioning

Abstract

This disclosure provides a physically guided enhanced method, device, and product for locating multiple unknown dynamic loads, relating to the field of computer technology. The physically guided enhanced method for locating multiple unknown dynamic loads includes: randomly generating load positions where loads are applied; binary encoding candidate positions on a target structure to determine an original encoding vector, the original encoding vector including encoded values ​​for load application and encoded values ​​for no load application; performing probability prediction on the encoded values ​​for load application and their adjacent encoded values ​​for no load application to determine the position exchange probability of the original encoding vector; mutating the original encoding vector to determine a predicted encoding vector; performing error prediction on the original and predicted encoding vectors to determine the error values ​​of the original and predicted encoding vectors; and performing threshold judgment based on the error values ​​of the original and predicted encoding vectors to determine a target encoding vector, the target encoding vector being used to determine the actual position of the load application.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a method, device and product for physically guided enhanced positioning of multiple unknown dynamic loads. Background Technology

[0002] Existing dynamic load location technologies mainly rely on empirical models or purely data-driven methods, which make it difficult to ensure both accuracy and efficiency when the load application location is unknown and the time history is unpredictable. Although the traditional transfer ratio method can eliminate the influence of the time history, it requires exhaustively listing all possible locations, and the amount of computation increases dramatically with the structural degrees of freedom. Pure deep learning methods are limited by the impact load assumption and the need for massive samples. Summary of the Invention

[0003] This disclosure provides a method, device, and product for physically guided and enhanced positioning of multiple unknown dynamic loads.

[0004] According to one aspect of this disclosure, a method for locating multiple unknown dynamic loads with physical guidance enhancement is provided, comprising: randomly generating load positions where loads are applied based on candidate positions on a target structure and the number of target loads, wherein the candidate positions are determined by spatial discretization of the target structure; binary encoding the candidate positions on the target structure to determine an original encoding vector, wherein the original encoding vector includes an encoding value for load application and an encoding value for no load application; performing probability prediction on the original encoding vector for the encoding value of load application and its adjacent encoding values ​​for no load application to determine the position exchange probability of the original encoding vector; mutating the original encoding vector based on the position exchange probability to determine a predicted encoding vector; performing error prediction on the original encoding vector and the predicted encoding vector based on a response transfer ratio matrix determined from acceleration responses collected at measurement points, respectively, to determine the error values ​​of the original encoding vector and the predicted encoding vector; and performing threshold judgment based on the error values ​​of the original encoding vector and the predicted encoding vector to determine a target encoding vector, wherein the target encoding vector is used to determine the actual position where the loads are applied.

[0005] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, a target encoding vector is determined by threshold judgment based on the error value of the original encoding vector and the error value of the predicted encoding vector. The method includes: when the error value of the predicted encoding vector is greater than or equal to the error threshold and the error value of the predicted encoding vector is less than the error value of the original encoding vector, the original encoding vector and the predicted encoding vector are regenerated iteratively, and the error values ​​of the original encoding vector and the predicted encoding vector are re-predicted until the error value of the predicted encoding vector is less than the error threshold, thereby determining the target encoding vector.

[0006] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, the original encoding vector is regenerated iteratively, including: constructing sample data based on the original encoding vector and the position exchange probability; and regenerating the original encoding vector based on the sample data.

[0007] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, the method re-iterates and generates a predictive coding vector, comprising: using the original predictive coding vector as sample data; and generating a new predictive coding vector based on the sample data and the regenerated original coding vector.

[0008] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, a target encoding vector is determined by threshold judgment based on the error value of the original encoding vector and the error value of the predicted encoding vector, including: when the error value of the predicted encoding vector is greater than or equal to the error threshold, and the error value of the predicted encoding vector is greater than the error value of the original encoding vector, the position exchange probability is re-predicted based on the original encoding vector; a new predicted encoding vector is determined based on the re-predicted position exchange probability, until the error value determined based on the new predicted encoding vector is less than the error value of the original encoding vector.

[0009] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, a threshold judgment is performed based on the error value of the original coding vector and the error value of the predicted coding vector to determine the target coding vector, including: when the error value of the predicted coding vector is less than the error threshold, the predicted coding vector is used as the target coding vector.

[0010] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, for the original coding vector, the method performs probability prediction on the coded values ​​of the load action and their adjacent coded values ​​without load action to determine the position exchange probability of the original coding vector, including: traversing the coded values ​​of the load action in the original coding vector to identify the coded values ​​without load action adjacent to the coded values ​​of the load action; and determining the position exchange probability of the coded values ​​of the load action in the original coding vector based on the historical exchange probabilities of the coded values ​​of the load action and their adjacent coded values ​​without load action.

[0011] According to at least one embodiment of the present disclosure, a physical guidance-enhanced method for locating multiple unknown dynamic loads, based on acceleration responses acquired from measuring points, includes determining a response transfer ratio matrix. The method comprises: acquiring vibration response data collected from measuring points on a target structure, wherein the vibration response data is a sequence of acceleration responses of each measuring point on the target structure under load over time; spatially grouping the measuring points corresponding to the vibration response data to determine a first measuring point group and a second measuring point group, wherein the number of measuring points in the first measuring point group is greater than the number of target loads; and determining a response transfer ratio matrix based on the mapping relationship between the vibration response data acquired from the first measuring point group and the vibration response data acquired from the second measuring point group.

[0012] According to at least one embodiment of the physical guidance-enhanced method for locating multiple unknown dynamic loads, an error prediction is performed on the original encoding vector and the predicted encoding vector using a deep learning network, wherein the deep learning network is trained based on a loss function constructed from the response transfer ratio matrix.

[0013] According to another aspect of this disclosure, an electronic device is provided, comprising: a memory storing execution instructions; and a processor executing the execution instructions stored in the memory, such that the processor performs a physically guided enhanced method for locating multiple unknown dynamic loads according to any embodiment of this disclosure.

[0014] According to another aspect of this disclosure, a readable storage medium is provided that stores execution instructions, which, when executed by a processor, are used to implement a physically guided enhanced method for locating multiple unknown dynamic loads according to any embodiment of this disclosure.

[0015] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements a physically guided enhanced method for locating multiple unknown dynamic loads according to any embodiment of this disclosure.

[0016] The physically-guided enhanced method for locating multiple unknown dynamic loads disclosed herein discretizes the structural space into candidate locations and uses binary encoding to represent the load distribution. It then combines the response transfer ratio matrix to perform intelligent mutation and error evaluation of the encoded vectors under physical constraints. A position exchange probability mechanism guides the search direction, avoiding the blindness of traditional random searches and improving convergence efficiency. By using the response transfer ratio matrix as a physical constraint, it ensures that the generated load location solutions conform to the structural dynamics, effectively suppressing the generation of spurious solutions. Ultimately, this method achieves high-precision, robust, and low-computational-cost dynamic load location in complex engineering scenarios. Attached Figure Description

[0017] The accompanying drawings illustrate exemplary embodiments of the present disclosure and, together with the description thereof, serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0018] Figure 1 This is a schematic diagram of the overall process of a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0019] Figure 2 This is a flowchart illustrating the determination of position exchange probability in a physical guidance-enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0020] Figure 3 This is a flowchart illustrating the process of determining the response transfer ratio matrix in a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0021] Figure 4 This is a schematic diagram of the structure of the evaluation network in a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0022] Figure 5 This is a schematic diagram of the structure of the generated network in a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0023] Figure 6 This is a schematic diagram of the structure of an adversarial network in a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0024] Figure 7 This is a schematic structural block diagram of a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0025] Figure 8 This is a schematic diagram of a cantilever beam model of a physically guided enhanced method for positioning multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0026] Figure 9 This is a schematic diagram of the training loss and validation loss of the evaluation network for a physical guidance-enhanced method for locating multiple unknown dynamic loads according to one embodiment of the present disclosure.

[0027] Figure 10 This is a schematic structural block diagram of a dynamic load positioning device according to one embodiment of the present disclosure.

[0028] Figure 11 This is a schematic structural block diagram of an electronic device according to one embodiment of the present disclosure. Detailed Implementation

[0029] The present disclosure will now be described in further detail with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the accompanying drawings.

[0030] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] In structural health monitoring scenarios for large bridges or high-rise buildings, only a few sensors can often be deployed at critical structural sections. Sudden loads such as vehicle collisions, wind vibrations, or people jumping are difficult to predict at their points of application, and it is also difficult to install force sensors for direct measurement. If traditional point-by-point trial calculations or large-scale deep learning solutions are used, either computation timeouts will occur due to the large degree of freedom, or inaccurate positioning will result due to insufficient samples, thus delaying damage assessment and maintenance decisions.

[0032] To address this, this disclosure proposes a physically-guided enhanced method for locating multiple unknown dynamic loads. By discretizing the structural space into candidate locations and using binary encoding to represent the load distribution, and combining the response transfer ratio matrix to perform efficient variation and error assessment of the encoded vector under physical constraints, the precise location of sudden dynamic loads can be quickly pinpointed using only the acceleration response of a limited number of measurement points. This avoids the computational explosion problem of traditional point-by-point calculations and overcomes the positioning inaccuracies caused by insufficient samples in purely data-driven methods. Therefore, it improves the real-time performance and engineering applicability of load identification, providing high-precision, low-cost key inputs for structural damage early warning and operation and maintenance decisions.

[0033] To facilitate description and make the technical solutions of this disclosure easier to understand, the terminology of this disclosure will be explained before describing the technical solutions of this disclosure.

[0034] Load refers to various external forces or factors acting on an engineering structure or component, including but not limited to gravity, wind force, seismic force, vehicle wheel pressure, trampling by crowds, equipment vibration, or thermal stress caused by temperature changes. It can be static (such as the self-weight of a building) or dynamic (such as a moving vehicle or a sudden impact). In the fields of structural mechanics and health monitoring, load is the fundamental cause of structural deformation, internal forces, and vibration response; its magnitude, direction, point of application, and changes over time directly affect the safety and durability of the structure.

[0035] This disclosed physical-guided enhanced method for locating multiple unknown dynamic loads can be deployed across various computing architectures and application scenarios. It can achieve large-scale parallel optimization on cloud server clusters, making it suitable for centralized health monitoring systems of large infrastructures such as bridges, dams, or wind turbine towers. By receiving multi-point sensor data, it can invert the location of dynamic loads in real time, providing decision support for remote operation and maintenance platforms. Alternatively, it can be lightweightly deployed on edge computing terminal devices (such as industrial gateways, embedded controllers, or smart sensor nodes) to achieve localized, low-latency load location in construction sites, old buildings, or temporary structures, enabling immediate response to sudden impacts (such as vehicle collisions or equipment falls) without relying on network backhaul.

[0036] Figure 1 A schematic flowchart illustrating the overall process of a physically guided enhanced method for locating multiple unknown dynamic loads according to one embodiment of this disclosure is shown. Figure 1 The method M100 shown includes steps S110 to S140. This method can be executed by a mobile phone, tablet computer, or server computing node.

[0037] In step S110, based on the candidate locations on the target structure and the number of target loads, the load locations where the loads act are randomly generated. The candidate locations are determined by spatial discretization of the target structure.

[0038] The continuous physical space of the structure is transformed into discrete candidate locations, and several candidate locations are randomly selected from the candidate locations according to the number of target loads as the initial load application points. This ensures that the entire load positioning process starts from a feasible domain that conforms to the actual engineering constraints, avoiding blind search or the generation of invalid solutions.

[0039] For example, based on the geometric shape and mechanical properties of the target structure, its key areas (such as beam spans, nodes, or support points) are divided into a finite number of discrete elements or grid nodes to form a set of candidate positions; several candidate positions with the same number of target loads are randomly selected from the candidate positions as the initial load application point distribution.

[0040] Preferably, the target structure is a physical engineering entity that requires dynamic load identification and location. It is a load-bearing structure with a defined geometric shape and mechanical properties, such as bridges, high-rise buildings, large-span roofs, transmission towers, wind turbine towers, industrial plant frames, or large machinery equipment bases. Its core characteristic is the ability to generate observable vibration responses (such as acceleration) under external loads, and to collect this vibration response data through a sensor network for load inversion.

[0041] Preferably, the number of target loads refers to the number of unknown dynamic loads that are simultaneously acting on the target structure in this event and have been given prior to the dynamic load positioning.

[0042] Preferably, the load locations for the load action are randomly generated by a generative network in a generative adversarial neural network (GAN) model, which uses a deep learning network structure.

[0043] In step S120, the candidate positions on the target structure are binary encoded to determine the original encoding vector, which includes the encoding value under load and the encoding value without load.

[0044] By mapping continuous or discrete candidate locations in physical space to structured binary encoded vectors, the complex load location search problem is transformed into a machine-processable combinatorial optimization problem.

[0045] For example, during the binary encoding process at candidate positions on the target structure based on the load location, the encoded value of the load (e.g., 1) is inserted, and the remaining candidate positions are filled with encoded values ​​of no load (e.g., 0), thus determining the original encoded vector. This ensures that the original encoded vector strictly satisfies the constraint of a fixed number of loads, while maintaining the compactness and operability of the encoding space, facilitating subsequent mutation operators and error assessment predictions.

[0046] Preferably, the candidate positions on the target structure are binary encoded by the generative network in the generative adversarial neural network model.

[0047] In step S130, for the original coding vector, the probability prediction is performed on the coding value of the load effect and its adjacent coding value without load effect to determine the position exchange probability of the original coding vector.

[0048] By analyzing the spatial relationship between the load position in the original encoding vector and its neighboring unloaded candidate positions, the probability weight of potential position swaps is predicted, thus providing directional guidance for subsequent mutation operations.

[0049] For example, the code values ​​of all load actions in the original code vector are traversed, and their spatially adjacent code values ​​without load actions are identified. Based on the historical exchange probabilities of these adjacent code values ​​without load actions, the exchange probability of moving from a code value with load actions to that adjacent code value without load actions is predicted. A normalized probability value is assigned to each exchangeable direction to obtain the position exchange probability. This position exchange probability serves as a control parameter for the mutation operator, enabling physics- or data-driven directional searches.

[0050] Preferably, the position swap probability of the original encoded vector is determined by an adversarial network in the generative adversarial neural network model, wherein the adversarial network is a deep learning network structure.

[0051] In step S140, the original coding vector is mutated based on the position swap probability to determine the predicted coding vector.

[0052] The original coding vector is perturbed in a controlled manner based on the position exchange probability to generate an exploratory predictive coding vector. This introduces local diversity while preserving the good features of the parent, and promotes the optimization process to evolve towards a better region. The probabilistic directional information is transformed into specific structural modification actions.

[0053] For example, based on the original encoded vector, and combined with the position exchange probability distribution, one or more positions with encoded values ​​that correspond to load effects are selected. These positions are then mutated according to the exchange probability of their adjacent encoded values ​​that do not correspond to load effects; that is, the mutated position is changed from a encoded value with load effects to a encoded value without load effects, while the corresponding adjacent encoded values ​​without load effects are changed to encoded values ​​with load effects. Step S140 is a constrained single-point or multi-point position migration operation, ensuring that the mutated predicted encoded vector still satisfies the physical premise that the target load quantity remains unchanged.

[0054] Preferably, the predictive encoding vector is determined by generating the adversarial network in the adversarial neural network model.

[0055] In step S150, based on the response transfer ratio matrix determined by the acceleration response collected from the measurement points, error prediction is performed on the original coding vector and the predicted coding vector respectively to determine the error value of the original coding vector and the error value of the predicted coding vector.

[0056] Preferably, the original encoded vector and the predicted encoded vector are input into the evaluation network of the generator adversarial neural network model to perform error prediction on the original encoded vector and the predicted encoded vector respectively, thereby obtaining the error value of the original encoded vector and the error value of the predicted encoded vector. The evaluation network is a convolutional neural network, and the loss function is constructed based on the response transfer ratio matrix.

[0057] In one specific embodiment, a deep learning network is used to predict the error between the original encoded vector and the predicted encoded vector. The deep learning network is trained based on a loss function constructed from the response transfer ratio matrix. This loss function measures the prediction error of the deep learning network model and serves as the basis for updating model parameters during training, ultimately improving the model's stability and generalization ability. The loss function is:

[0058]

[0059] in, Represents the loss function. Indicates data loss items, Represents the physical loss term. This represents the error between the estimated acceleration response and the measured acceleration response based on the response transfer ratio matrix estimation. This represents the error between the predicted acceleration response and the measured acceleration response, as indicated by the evaluation network. Represents the root mean square. The response transfer ratio matrix is ​​determined only by the location of the load and the quantity of the target load, and is independent of the time history of the unknown load. This represents the acceleration response vector of the predetermined first set of measuring points. This represents the acceleration response vector of the predetermined second set of measuring points.

[0060] Optionally, for either the original coding vector or the predicted coding vector, the acceleration response collected by the first set of measurement points on the target structure is obtained; using the response transfer ratio matrix, the acceleration response of the predetermined second set of measurement points is predicted based on the acceleration response collected by the predetermined first set of measurement points, resulting in a predicted acceleration response; the actual measured acceleration response collected by the second set of measurement points within the same time period is extracted; the residual between the predicted acceleration response and the measured acceleration response is predicted and converted into an error value through normalization. This process is performed on both the original coding vector and the predicted coding vector generated by mutation, respectively, to obtain the corresponding error values ​​of the original coding vector and the predicted coding vector.

[0061] In step S160, a threshold judgment is performed based on the error values ​​of the original encoding vector and the predicted encoding vector to determine the target encoding vector, which is used to determine the actual location of the load.

[0062] Preferably, when the error value of the predicted encoding vector is less than the error threshold, the predicted encoding vector is used as the target encoding vector. This achieves high-precision convergent identification of the actual application location of dynamic loads, and can adaptively complete the stable evolution from the initial assumption to the optimal solution under complex noise environments and limited measurement points, ultimately outputting load distribution results that meet the accuracy requirements, thus improving the accuracy, reliability, and engineering practicality of dynamic load identification.

[0063] In one specific embodiment, when the error value of the predicted encoding vector is greater than or equal to the error threshold, and the error value of the predicted encoding vector is greater than the error value of the original encoding vector, the position exchange probability is re-predicted based on the original encoding vector; based on the re-predicted position exchange probability, a new predicted encoding vector is determined, until the error value determined based on the new predicted encoding vector is less than the error value of the original encoding vector. During the optimization process, the mutated predicted encoding vector not only fails to approximate the true load distribution but also leads to an increase in prediction error. Instead of simply reverting to the original encoding vector, this approach dynamically adjusts the probability distribution of the mutation direction by re-evaluating the local structural features of the original encoding vector. This guides subsequent searches to escape the current invalid path and explore more promising neighborhood spaces. This is a key guarantee for improving the model's global optimization capability in complex error surfaces, effectively avoiding the problem of traditional intelligent optimization models easily getting stuck in local stagnation or blindly exploring.

[0064] In one specific embodiment, when the error value of the predicted encoding vector is greater than or equal to the error threshold, and the error value of the predicted encoding vector is less than the error value of the original encoding vector, the original encoding vector and the predicted encoding vector are regenerated iteratively, and the error values ​​of the original encoding vector and the predicted encoding vector are re-predicted until the error value of the predicted encoding vector is less than the error threshold, thus determining the target encoding vector. By setting dynamic judgment conditions, the generative adversarial neural network model is guided to continuously explore better solutions during the search process, avoiding localization failure due to premature convergence or local stagnation. Ensuring the continuous improvement of solution quality while preventing the model from falling into invalid loops is a key technical support for ensuring the high accuracy and strong robustness of the final output results.

[0065] Preferably, the original encoded vector is regenerated iteratively, including: constructing sample data based on the original encoded vector and position exchange probabilities; and regenerating the original encoded vector based on the sample data using the adversarial network of the generative adversarial neural network model. This not only preserves the effective information in the original optimization path (such as the location of high-confidence loads) but also injects appropriate perturbations through the sample data reconstruction mechanism, enhancing the solution exploration capability and preventing the generative adversarial neural network model from getting stuck in local stagnation or repeating ineffective searches. Simultaneously, constructing samples based on historical exchange probabilities ensures that the newly generated original encoded vector remains within the high-potential solution space region, balancing search efficiency and convergence stability. This improves the adaptability of the generative adversarial neural network model to complex load distributions and its global optimization performance while ensuring physical rationality.

[0066] Preferably, the re-iterative generation of the predicted encoding vector includes: using the original predicted encoding vector as sample data; and generating a new predicted encoding vector based on the sample data and the regenerated original encoding vector. This achieves context continuation and direction optimization of the mutation operation. While retaining historical search experience, the updated original encoding vector serves as a benchmark to guide the construction process of the new predicted encoding vector, ensuring that the mutation direction always evolves around the current optimal solution. Simultaneously, using the original predicted encoding vector as a sample allows for the extraction of its local structural features or effective payload patterns, avoiding information loss and improving the quality and stability of the solution. This enhances the adaptive adjustment capability of the generative adversarial neural network model in complex error spaces, maintaining search coherence while avoiding getting trapped in invalid loops, thereby accelerating convergence and improving the accuracy and reliability of the final localization result.

[0067] This disclosed physical-guided enhanced method for locating multiple unknown dynamic loads discretizes the continuous structural space of the target structure into candidate locations and generates the load locations based on the number of target loads. It transforms this into a computable combinatorial optimization problem using binary encoded vectors to obtain the original encoded vectors. An error quantization mechanism based on the response transfer ratio matrix and a deep learning evaluation network is introduced, combined with position exchange probabilities to achieve physical / data-driven directional mutation, shifting the search process from blind exploration to intelligent evolution. This eliminates interference from the time histories of unknown loads and overcomes the limitation of previous load location identification research being only applicable to impact loads. Furthermore, through multi-branch convergence control logic, including deterioration escape, direction reset, continuous refinement, and threshold termination strategies, it effectively avoids the problems of traditional generative adversarial neural network models easily getting trapped in local optima, slow convergence, or misjudgments. It possesses both physical interpretability and data self-learning capabilities, enabling it to stably and quickly approximate the true load distribution in complex noisy environments, nonlinear responses, and practical engineering scenarios with insufficient prior information, thus improving the automation level, positioning accuracy, and engineering applicability of dynamic load identification.

[0068] Regarding step S130, for the original encoded vector, probability prediction is performed on the encoded value of the load effect and its adjacent encoded values ​​without load effects to determine the position exchange probability of the original encoded vector. In some embodiments of this disclosure, this may include, for example... Figure 2 Steps S1301 to S1302 are shown.

[0069] In step S1301, the coded values ​​of the load action of the original coded vector are traversed to identify the coded values ​​of the load action that are adjacent to the coded values ​​of the load action and have no load action.

[0070] By analyzing the spatial structure of the original encoded vectors, potential evolutionary paths are identified, providing fundamental support for the intelligent mutation operations of subsequent generative adversarial neural network models. This represents a technological leap from global random search to localized targeted exploration. By identifying the encoded values ​​of transferable locations, i.e., load effects, it ensures that while maintaining physical continuity and engineering rationality, focused optimization is conducted on high-potential regions, improving search efficiency and convergence stability.

[0071] In step S1302, the position exchange probability of the coded value of the load action in the original coded vector is determined based on the historical exchange probability of the coded value of the load action and its adjacent coded values ​​without load action.

[0072] By integrating multi-dimensional prior information, the rationality weight of the load location migration direction of the encoded values ​​under different load effects is dynamically evaluated, thus providing directional guidance driven by both data and physics for subsequent mutation operations. By introducing multi-source factors such as structural response characteristics and historical optimization experience, the mutation process not only meets the basic constraints of combinatorial optimization, but also possesses engineering interpretability and environmental adaptability, enhancing the system's load location capability under complex noise backgrounds.

[0073] Therefore, by sensing and analyzing the spatial structure of the current load distribution, a physically reasonable and data-driven optimization path was constructed. While maintaining the constraint on the number of loads, the path guides the mutation operation to focus on high-potential neighborhood regions, improving the directionality, efficiency, and stability of the optimization process. Simultaneously, historical optimization experience and structural response priors are introduced to enhance the robustness and adaptability of the generative adversarial neural network model under complex noisy environments, ultimately achieving high-precision and rapid convergence identification of dynamic load locations.

[0074] Regarding step S150, based on the response transfer ratio matrix determined from the acceleration response acquired at the measurement points, error prediction is performed on both the original coding vector and the predicted coding vector to determine the error values ​​of the original coding vector and the predicted coding vector. In some embodiments of this disclosure, the response transfer ratio matrix determined from the acceleration response acquired at the measurement points may include, for example: Figure 3 Steps S1501 to S1503 are shown.

[0075] In step S1501, vibration response data collected at measuring points on the target structure are obtained. The vibration response data is the acceleration response sequence of each measuring point on the target structure under load as time changes.

[0076] By acquiring high-precision synchronized acceleration data, we can ensure that the obtained vibration response data can truly reflect the dynamic behavior of the target structure.

[0077] The above measuring points indicate the installation location of the sensors on the target structure, preferably in areas sensitive to structural dynamic response, at node connections, at mid-span, or near supports.

[0078] In step S1502, the measuring points corresponding to the vibration response data are spatially grouped to determine the first measuring point group and the second measuring point group. The number of measuring points in the first measuring point group is greater than the number of target loads.

[0079] All valid measuring points deployed on the target structure are divided into two non-overlapping functional groups: the first measuring point group (input group) and the second measuring point group (output / verification group). The first measuring point group characterizes the dynamic input features of the structure under external loads, and its collected vibration response data (i.e., acceleration response) serves as contextual information for judging the rationality of the encoding vector. The second measuring point group provides independent observation data for comparison with theoretical predictions or model inferences, thereby quantifying errors. The first measuring point group is required to have more measuring points than the target load to ensure the system has sufficient degrees of freedom to uniquely or approximately uniquely invert the load distribution. This grouping process can be statically set during the initialization phase or dynamically optimized based on the structural modal characteristics to adapt to the identification needs under different working conditions.

[0080] In step S1503, the response transfer ratio matrix is ​​determined based on the mapping relationship between the vibration response data collected by the first measuring point group and the vibration response data collected by the second measuring point group.

[0081] Using multiple sets of synchronous vibration response data collected from the target structure under known excitation or typical working conditions, the dynamic coupling relationship between the first and second measurement point groups is analyzed. Assuming the target structure satisfies linear time-invariant characteristics, the acceleration response of the second measurement point group can be obtained from the response of the first measurement point group through a linear transformation. The response transfer ratio matrix is ​​obtained by methods such as least squares fitting, frequency domain transfer function estimation, or multivariate regression.

[0082] Therefore, it realizes the extraction and modeling of the inherent dynamic laws of the structure from the original observation data, and provides an interpretable and computable physical basis for the rationality evaluation of the subsequent coding vector. It avoids the dependence on complex finite element models, improves the accuracy, stability and engineering applicability of dynamic load identification, and supports efficient and reliable load position inversion and error quantification without the need for full structural perception.

[0083] The technical solution of this disclosure will be further explained below with specific implementation and application examples.

[0084] Based on the time-domain response transfer ratio matrix, a method is proposed to identify the location of unknown dynamic loads using physical guidance and generative adversarial neural networks. The mechanism of physical guidance is as follows:

[0085] In the time domain, if it is a pulse load The generated system response is Then, based on position Load at the location The system caused by the location Response at the location It can be expressed as the integral form of a series of unit impulse response functions:

[0086] (1)

[0087] Equation (1) can be written in matrix form as shown in equation (2):

[0088] (2)

[0089] in The total number of sampling points. For time sampling points, Given the sampling interval, equation (2) can be written as follows:

[0090] (3)

[0091] in,

[0092] (4)

[0093] Formula (4) represents the position. j Load time history at the location and location i Response time at the location y i The relationship between the structures. If the structure is subjected to... One load is applied, and random measurements are assumed. M The structural responses at each location are denoted in ascending order of measurement point number as follows: The relationship between load and structural response can be expressed as:

[0094] (5)

[0095] make Each element in the matrix The specific representation is shown in formula (4), and formula (5) can be further written in the following form:

[0096] (6)

[0097] Furthermore, this Structural response at each measurement point The data is randomly divided into two groups, denoted as the first group response vector. (i.e., the acceleration response vector of the first set of measuring points) and the response vector of the second set. (i.e., the acceleration response vector of the second measuring point group), and and Each contains m and The structural response at each measurement point. The two sets of responses in the time domain can be expressed as Equation (7) according to Equation (6):

[0098] (7)

[0099] When the number of acceleration responses in the first group was observed Greater than the unknown load Number of Then, from the first row of equation (7), we can obtain equation (8):

[0100] (8)

[0101] in represent The pseudo-inverse of the second set of responses can be further expressed as shown in the equation: (9)

[0102] (10)

[0103] T 12 The matrix is ​​the acceleration response transfer ratio matrix in the time domain, which depends only on the location and quantity of the load, and is independent of the time history of the unknown load. Therefore, using... T 12 The matrix eliminates the influence of unknown load time histories, and then estimates the structural acceleration vector based on the response transfer ratio matrix. Compared with the measured structural acceleration vector Y 2. Construct the objective function, as shown in equation (11):

[0104] (11)

[0105] in, The relative error between the estimated second set of acceleration responses and the measured second set of acceleration responses; This refers to the second set of acceleration responses estimated based on the response transfer ratio matrix. i One value; This refers to the second set of measured acceleration responses. i One value; The total number of discrete time history data points for the second set of acceleration responses, where This represents the structural acceleration response at the location of the first measuring point in the second set of responses. This indicates the corresponding number in the second group of responses. Structural acceleration response at each measuring point.

[0106] Therefore, by constructing an objective function based on the response transfer ratio matrix, the influence of unknown load time histories is eliminated. This simplifies the complex problem of coupled load time histories to considering only the application locations of unknown multiple loads, transforming the identification of load application locations into an optimization problem. A mapping between load application locations and acceleration responses is constructed using a physical evaluation network based on the response transfer ratio matrix, overcoming the inefficiency and brute-force nature of traditional techniques. This reduces the number of repetitive calculations required by traditional finite element models, improving computational efficiency.

[0107] Furthermore, a generative adversarial neural network (GAN) is employed to optimize the identified load application locations. This technique uses only a combination of deep learning network models to optimize the identified load application locations without the need for other optimization algorithms. It mainly consists of three network modules with different functions: an evaluation network module, a generative network module, and an adversarial network module. The evaluation network module uses a Convolutional Neural Network (CNN) structure, while the generative and adversarial network modules both use Deep Neural Network (DNN) structures. The main functions of these three modules and the details of the network inputs and outputs will be described in detail below.

[0108] like Figure 4As shown, the deep learning model used in the evaluation network module is a convolutional neural network (CNN). CNNs achieve local connectivity and weight sharing through convolutional layers. Each convolutional kernel performs linear operations with a local region of the input data, and activation functions enhance the network's non-linear expressive power, enabling it to capture local features in the data and solve complex pattern classification and regression problems in practical engineering. A typical CNN consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. Convolutional layers help reduce the number of model parameters and computational complexity while maintaining effective capture of image features. By increasing the number of convolutional layers, CNNs can extract high-level features of images layer by layer, thereby better recognizing and classifying images. Pooling layers in CNNs are used to reduce feature dimensionality, enhance feature invariance, reduce overfitting, extract more abstract features, control computational resource consumption, and improve the model's generalization ability. However, too many convolutional layers may cause the network to over-abstract the input data, resulting in the loss of important information. Furthermore, an excessive number of neurons in fully connected layers may cause the network to learn noise in the data, leading to overfitting. Therefore, hyperparameters of convolutional neural networks, such as kernel size, number of kernels, stride, and padding, are crucial for evaluating network performance and generalization ability. By appropriately selecting these hyperparameters, the performance of convolutional neural networks on specific tasks can be optimized, enabling effective processing and feature extraction of image data.

[0109] Therefore, this disclosure converts the load application location coordinates into binary encoded vectors, where 0 indicates no load is applied at the possible load location (i.e., the encoded value of no load application), and 1 indicates a load is applied at the possible load location (i.e., the encoded value of load application). The evaluation network module's role is to establish the binary encoded vector of the load location. ( This represents the total number of all possible load locations; Indicates the location The binary code value of the load location (a value of 0 or 1), and the acceleration response error. (The mapping relationship between the acceleration response identified based on the transfer ratio matrix between two sets of structural responses and the measured acceleration response) is established. Since the evaluation network is a crucial part of this optimization framework affecting identification, it is necessary to limit the sample size required by the evaluation network while ensuring its prediction accuracy. Therefore, based on the data loss function, the response transfer ratio matrix calculated from the physical model is introduced into the objective function, proposing a data-physical hybrid driven evaluation network to establish the mapping relationship between the binary encoded vector of the load position and the acceleration response error. The loss between the prediction error and the actual error of the evaluation network (data loss) is as follows:

[0110] (12)

[0111] in This represents the true error between the second set of acceleration responses estimated based on the response transfer ratio matrix and the measured second set of acceleration responses. To evaluate the prediction error between the network-predicted second set of acceleration responses and the measured second set of acceleration responses.

[0112] Additionally, the loss (physical loss) is evaluated between the network prediction error and the error between the second set of responses estimated based on the transfer ratio matrix and the measured responses:

[0113] (13)

[0114] Therefore, the proposed data-physical hybrid loss function is as follows:

[0115] (14)

[0116] Therefore, by introducing a data-physics hybrid loss function into the evaluation network, a physics-guided generative adversarial neural network is proposed. This network leverages the prior knowledge of the physical model to enhance the data-driven learning process, enabling the acquisition of high-precision unknown load location identification results with fewer samples.

[0117] like Figure 5 As shown, the network model used in the generative network module is a deep neural network. Unlike traditional perceptrons, deep neural networks have linear operational relationships between each neuron in each layer and all neurons in the next layer, and each neuron contains a built-in non-linear activation function. This allows the network to express non-linear mapping relationships. Deep neural networks typically consist of an input layer, hidden layers, and an output layer. They usually have multiple hidden layers. Increasing the number of hidden layers and neurons can better separate data features and enhance the network's non-linear mapping ability. However, an excessive number of hidden layers and neurons may lead to learning noise in the data, resulting in overfitting. Therefore, the hyperparameters of a deep neural network are also important parameters for evaluating whether the network can effectively represent the relationship between input and output.

[0118] The generator network module generates binary encoded vectors of load locations based on the known number of loads. Therefore, the input to the generator network is the number of loads. The output is a binary encoded vector of the load position. Therefore, a generative network module is used to automatically generate binary codes for the load positions of each possible load.

[0119] like Figure 6As shown, the adversarial network module is also a deep neural network. Its function is to generate the probability of swapping each 1 in the load position binary encoding vector with its left or right encoded value based on the load position binary encoding vector generated by the generator network module. That is, the input of the adversarial network is the load position binary encoding vector output by the generator network module. The output is the probability that each 1 in the load position encoding vector will be swapped with its left or right encoded value. This is used to determine how each load application location should be optimized during the optimization process.

[0120] like Figure 7 As shown, the process of identifying the load application location using the physically enhanced guided generative adversarial neural network optimization framework proposed in this disclosure is as follows:

[0121] The first step is to train and evaluate the network. First, assume the number of unknown loads is... Several sets of unknown load location coordinates are randomly generated and converted into corresponding binary encoded vectors. Loads are then applied to these location combinations on the finite element model to obtain the structural acceleration response at the corresponding measurement points. Next, the error between the identified structural acceleration response and the measured structural acceleration response is calculated based on the first set of structural acceleration responses. Finally, all load location binary encoded vectors are... As input to the network training set, the error between the recognized structural acceleration response and the measured acceleration response is used as the input. The output of the training set is used as the evaluation network; finally, the evaluation network is trained to obtain the trained evaluation network.

[0122] The second step is to randomly initialize the network parameters of the generator network and the adversarial network. The input to the generator network is the number of unknown payloads. The output is a binary encoded vector of the load position. The input to the adversarial network is the binary encoded vector of the payload location. The output is the probability value of swapping each 1 in the load position encoding vector with its left or right encoded value. The probability of swapping each 1 with its left or right encoded value is summed to 1, and the number of iterations is set to 1. K Second-rate.

[0123] The third step is to proceed. K The next iteration. The generator network first randomly obtains the first set of binary encoded vectors for the load positions. This refers to the original encoded vector. When this set of load position encoded vectors is input into the adversarial network, the probability of swapping each 1 in the corresponding load position encoded vector with its left or right encoded value can be obtained. Based on the maximum probability corresponding to each load position encoding vector, it is determined whether the encoding value to be swapped with the encoding value to its left or right is used for each load position, thus obtaining a second set of new binary encoding vectors for load positions. These are the predicted encoding vectors. Inputting these two sets of binary encoded vectors for the load positions into the evaluation network yields two sets of errors. (i.e., the error value of the original encoded vector) and (i.e., the error value of the predicted encoding vector), and determine Is it less than the error threshold? If the value is less than the specified value, the loop will exit and the output will be displayed. The corresponding binary encoding vector of the load position. Otherwise, continue the judgment. Is it less than If so, then construct a new set of binary encoded vector samples of load positions, with the sample input being the error. The corresponding binary encoded vector of the load application location is output as the interchange probability of the corresponding load location. This sample is added to the training dataset of the adversarial network, and the adversarial network is trained. The corresponding binary encoded vector of the load application location is added to the training dataset of the generator network, and the generator network is trained. This completes one iteration. Otherwise, the adversarial network re-predicts the probability of swapping each 1 in the load application location encoded vector with its left or right encoded value, obtaining a new binary encoded vector of the load application location, until... > .

[0124] To illustrate the technical effects of the embodiments of this disclosure, the following verification cases are provided.

[0125] like Figure 8 As shown, taking a cantilever beam under three concentrated white noise loads as an example, the cantilever beam is 1m long and is divided into 50 element nodes, each element being 0.02m long. The Young's modulus of elasticity is... mass density The acceleration response was observed at the lower surface of the cantilever beam connection node with a sampling frequency of 50 Hz and a duration of 30 s.

[0126] In this case, the number of unknown loads is known to be 3. Given that 3 loads are placed on 50 element nodes, if an exhaustive method is used, the total number of unknown loads is... This layout scheme involves a large amount of calculation and has extremely low computational efficiency.

[0127] Since the number of acceleration responses in the first group needs to be greater than the number of loads, the acceleration responses observed on the lower surfaces of the 10th, 15th, 20th, 25th, 28th, 35th, 40th, and 45th element nodes are used as the first group of acceleration responses, and the acceleration responses observed on the lower surfaces of the 9th, 14th, 26th, and 39th element nodes are used as the second group of acceleration responses. Furthermore, to account for the influence of measurement noise, Gaussian white noise with a 5% mean square error is added to the acceleration responses. When the actual load is applied to the right node of element [15 25 35], it is converted into the corresponding binary encoding vector of the load position, verifying the recognition effectiveness of the physical enhancement-guided generative adversarial neural network optimization framework proposed in this disclosure.

[0128] The loss changes of the evaluation network on the training and validation sets during the load application location optimization process are as follows: Figure 9 As shown, the horizontal axis represents the number of training epochs, and the vertical axis represents the mean squared error (MSE) loss value. The solid blue line represents the training loss, and the dashed orange line represents the validation loss. With the increase in the number of training epochs, both the training loss and validation loss decrease rapidly, converging quickly to near zero within the first 10 epochs and remaining stable in subsequent iterations. This indicates that the model effectively learns the mapping relationship between the input data and the target output during training, without significant overfitting, demonstrating good generalization ability.

[0129] Furthermore, the optimized results of the load application location predicted by the network are transformed into the load location in natural coordinates as shown in the table below. The actual load application location is consistent with the load application location identified in this disclosure.

[0130]

[0131] The physically enhanced guided evaluation network in the proposed physically enhanced guided generative adversarial neural network (GAN) optimization framework can learn the relationship between the load application location and the corresponding acceleration response error value. Using the trained network, unknown multiple load application locations can be accurately identified. Furthermore, by introducing a response transfer ratio matrix, a physically enhanced guided GAN is proposed, which can eliminate the influence of load time history, thus breaking the limitation of traditional load location identification methods that are only applicable to impact loads. Compared with traditional methods, the method of this disclosure simplifies the process of repeated calculations of the finite element model, improves computational efficiency, and requires only a small number of samples to achieve high-precision load location identification. Numerical simulations to identify unknown multiple load application locations verify the effectiveness of the invention, and this technology can provide key technical support for the accurate identification of load application locations in two-dimensional and three-dimensional structures.

[0132] Based on any of the above embodiments, this disclosure also provides a dynamic load positioning device.

[0133] Figure 10 This is a schematic block diagram of the structure of a dynamic load positioning device according to one embodiment of the present disclosure.

[0134] like Figure 10 As shown, the dynamic load positioning device includes:

[0135] The load location generation module 1102 randomly generates the load location of the load action based on the candidate locations on the target structure and the number of target loads. The candidate locations are determined by spatial discretization of the target structure.

[0136] The encoding module 1104 performs binary encoding on the candidate positions on the target structure to determine the original encoding vector, which includes the encoding value under load and the encoding value without load.

[0137] The exchange probability determination module 1106 performs probability prediction on the encoded value of the load action and its adjacent encoded values ​​without load action for the original encoded vector, and determines the position exchange probability of the original encoded vector.

[0138] The position mutation module 1108 mutates the original coding vector based on the position exchange probability to determine the predicted coding vector;

[0139] The error prediction module 1110, based on the response transfer ratio matrix determined by the acceleration response collected from the measurement points, performs error prediction on the original coding vector and the predicted coding vector respectively, and determines the error value of the original coding vector and the error value of the predicted coding vector.

[0140] The dynamic load positioning module 1112 determines the target encoding vector by performing a threshold judgment based on the error value of the original encoding vector and the error value of the predicted encoding vector. The target encoding vector is used to determine the actual position of the load.

[0141] The aforementioned dynamic load positioning device can be in the form of computer software, and each module of the aforementioned dynamic load positioning device can be implemented through computer software modules.

[0142] The specific implementation process of the functions and roles of each module in the above-mentioned dynamic load positioning device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0143] This disclosure also provides an electronic device 1000. Figure 11 A schematic diagram of the hardware implementation using the processing system is shown.

[0144] The hardware structure of electronic device 1000 can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application and overall design constraints of the hardware. Bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules. Bus 1100 can also connect various other circuits 1400 such as peripherals, voltage regulators, power management circuits, external antennas, etc. Bus 1100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one connection line is used in this figure, but this does not indicate that there is only one bus or one type of bus.

[0145] For ease of explanation, certain steps of the above method are described in relation to modules. It should be understood that the corresponding module performing one or more steps of the above method may be one or more hardware modules specifically configured to perform the corresponding step, or implemented by a processor configured to perform the corresponding step, or stored in a computer-readable medium for implementation by a processor, or implemented by some combination thereof.

[0146] This disclosure also provides a readable storage medium storing a computer program that, when executed by a processor, is used to implement the methods described above. A "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples of a readable storage medium include: an electrical connection with one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM), etc.

[0147] This disclosure also provides a computer program product, the methods of which can be implemented wholly or partially through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented wholly or partially as a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, all or part of the processes or functions of this disclosure are performed.

[0148] Computer programs or instructions can be stored in a readable storage medium or transferred from one readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The readable storage medium can be any available medium capable of access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or it can include both volatile and non-volatile types of storage media.

[0149] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0150] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0151] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0152] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0153] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., refer to specific features, structures, or characteristics described in connection with that embodiment / mode or example, which are included in at least one embodiment / mode or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0154] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0155] Those skilled in the art should understand that the above embodiments are merely for illustrating the present disclosure and are not intended to limit the scope of the disclosure. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present disclosure.

Claims

1. A method for locating multiple unknown dynamic loads with enhanced physical guidance, characterized in that, include: Based on the candidate locations on the target structure and the number of target loads, the load locations where the loads act are randomly generated, and the candidate locations are determined by spatial discretization of the target structure. Candidate positions on the target structure are binary encoded to determine the original encoding vector, which includes the encoding value under load and the encoding value without load. For the original encoded vector, the probability of position swapping between the encoded value of the load effect and its adjacent encoded values ​​without load effect is predicted. Based on the position swap probability, the original coding vector is mutated to determine the predicted coding vector; Based on the response transfer ratio matrix determined from the acceleration response collected at the measurement points, a deep learning network is used to predict the errors of the original encoding vector and the predicted encoding vector, respectively, to determine the error values ​​of the original encoding vector and the predicted encoding vector; the deep learning network is trained based on the loss function constructed from the response transfer ratio matrix. as well as A threshold judgment is performed based on the error values ​​of the original encoding vector and the predicted encoding vector to determine the target encoding vector, which is used to determine the actual location of the load. The response transfer ratio matrix determined based on the acceleration response collected from the measurement points includes: Vibration response data collected at measuring points on the target structure is obtained, wherein the vibration response data is the acceleration response sequence of each measuring point on the target structure under load as a function of time. The measuring points corresponding to the vibration response data are spatially grouped to determine a first measuring point group and a second measuring point group, wherein the number of measuring points in the first measuring point group is greater than the number of target loads. Based on the mapping relationship between the vibration response data collected by the first measuring point group and the vibration response data collected by the second measuring point group, the response transfer ratio matrix is ​​determined.

2. The method for locating multiple unknown dynamic loads with enhanced physical guidance as described in claim 1, characterized in that, The target coding vector is determined by threshold judgment based on the error values ​​of the original coding vector and the predicted coding vector, including: When the error value of the predicted coding vector is greater than or equal to the error threshold, and the error value of the predicted coding vector is less than the error value of the original coding vector, the original coding vector and the predicted coding vector are regenerated iteratively, and the error values ​​of the original coding vector and the predicted coding vector are re-predicted until the error value of the predicted coding vector is less than the error threshold, and the target coding vector is determined.

3. The method for locating multiple unknown dynamic loads with enhanced physical guidance as described in claim 2, characterized in that, Regenerate the original encoded vector iteratively, including: Sample data is constructed based on the original original encoding vector and the position swap probability; Based on the sample data, the original encoding vector is regenerated.

4. The method for locating multiple unknown dynamic loads with enhanced physical guidance as described in claim 2, characterized in that, Re-iterate to generate the predicted encoding vector, including: Use the original predicted encoding vector as sample data; Based on the sample data and the regenerated original encoding vector, a new predicted encoding vector is generated.

5. The method for locating multiple unknown dynamic loads with enhanced physical guidance as described in claim 1, characterized in that, The target coding vector is determined by threshold judgment based on the error values ​​of the original coding vector and the predicted coding vector, including: When the error value of the predicted coding vector is greater than or equal to the error threshold, and the error value of the predicted coding vector is greater than the error value of the original coding vector, the position swap probability is re-predicted based on the original coding vector. Based on the re-predicted position swap probability, a new predictive coding vector is determined until the error value determined based on the new predictive coding vector is less than the error value of the original coding vector.

6. The method for locating multiple unknown dynamic loads with enhanced physical guidance as described in claim 1, characterized in that, The target coding vector is determined by threshold judgment based on the error values ​​of the original coding vector and the predicted coding vector, including: When the error value of the predicted coding vector is less than the error threshold, the predicted coding vector is used as the target coding vector.

7. The method for locating multiple unknown dynamic loads with enhanced physical guidance as described in claim 1, characterized in that, For the original encoded vector, probability prediction is performed on the encoded value of the load action and its adjacent encoded values ​​without load action to determine the probability of position swapping of the original encoded vector, including: The original encoded vector is traversed for the encoded values ​​of the load action, and the encoded values ​​that are adjacent to the encoded values ​​of the load action and have no load action are identified. Based on the historical exchange probabilities of the encoded values ​​of the load action and their adjacent encoded values ​​without load action, the position exchange probabilities of the encoded values ​​of the load action in the original encoded vector are determined.

8. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes execution instructions stored in the memory, causing the processor to perform the physical guidance-enhanced method for locating multiple unknown dynamic loads as described in any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the physical guidance-enhanced method for locating multiple unknown dynamic loads as described in any one of claims 1 to 7.

Images