Gymnasium steel structure performance evaluation method and system based on multi-modal data fusion

Abstract

This invention provides a method and system for evaluating the performance of a stadium steel structure based on multimodal data fusion, belonging to the field of steel structure analysis technology. This invention acquires strain response time-series data and temperature time-series data of key components of the stadium steel structure over multiple historical time periods. Based on the strain response time-series data and temperature time-series data, a two-layer long short-term memory network is constructed and trained to predict the strain response data of key components at future times. The invention also acquires strain response time-series data and temperature time-series data of key components of the stadium steel structure in the current time period, and uses the two-layer long short-term memory network to predict the strain response data of key components at future times. The strain response data of each key component at the next sampling time is compared with a preset threshold. Based on the comparison results, the steel structure performance is classified as excellent, qualified, or unqualified.

Description

Technical Field

[0001] This invention relates to the field of steel structure analysis technology, specifically to a method and system for evaluating the performance of steel structures in gymnasiums based on multimodal data fusion. Background Technology

[0002] Over long-term use, the steel structure of a gymnasium is affected by various factors such as changes in ambient temperature and load, leading to gradual degradation of structural performance and even safety hazards. Traditional steel structure performance assessment methods often rely on periodic manual inspections or single-type data monitoring, which are insufficient to comprehensively and in real-time reflect the dynamic response of the structure in complex real-world environments. This is especially true for buildings like gymnasiums with large spatial spans and complex structures, where the stress-strain response of key components exhibits significant temporal and multimodal characteristics. Accurate state prediction and performance classification cannot be achieved using only static or single-source data. While existing monitoring methods based on data fusion have emerged, they often focus on single environmental parameters or lack detailed modeling of temperature cumulative effects and temporal dependencies, resulting in insufficient prediction accuracy and delayed assessment results. These methods fail to meet the real-time and accuracy requirements of modern large-scale steel structure health management.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for evaluating the performance of steel structures in gymnasiums based on multimodal data fusion, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for evaluating the performance of steel structures in gymnasiums based on multimodal data fusion, comprising the following steps:

[0007] Step 1: Obtain strain response time-series data and temperature time-series data of key components of the stadium's steel structure over multiple historical time periods;

[0008] Step 2: Construct a strain column vector from the strain response data at the last sampling time in the time series data, construct a first feature matrix from the temperature data at the remaining sampling times, and construct a second feature matrix from the strain response data at the remaining sampling times.

[0009] Step 3: Using the first feature matrix and the second feature matrix as input and the strain column vector as output, construct and train a two-layer long short-term memory network. The two-layer long short-term memory network includes a strain response long short-term memory network and a temperature long short-term memory network. In the temperature long short-term memory network, temperature accumulation is determined based on temperature trend, and a forgetting gate is constructed based on temperature accumulation.

[0010] Step 4: Obtain the strain response time series data and temperature time series data of key parts of the stadium steel structure in the current time period, and construct the first feature matrix and the second feature matrix. Input them into the trained two-layer long short-term memory network to predict the strain response data at the next sampling time.

[0011] Step 5: Compare the strain response data of each key part at the next sampling time with the preset threshold, and classify the steel structure performance into excellent, qualified and unqualified based on the comparison results.

[0012] Furthermore, the key part is the center of the densely represented component and its connection position with other components, and the time period is a time period consisting of N+1 consecutive sampling moments.

[0013] Furthermore, the logic for determining dense representative components is as follows: a preset component quantity threshold is set, starting from the leftmost and bottommost component of the stadium, the number of components in a one cubic decimeter space centered on that component is counted. If the number of components is greater than the component quantity threshold and there are no other dense representative components, then that component is called a dense representative component. This process is repeated for all components in the stadium.

[0014] Furthermore, the rows of the first feature matrix and the second feature matrix represent sampling times, the columns represent key parts, and the element values ​​are the temperature and strain responses at the key parts at the corresponding sampling times.

[0015] The first feature matrix is ​​input into the temperature long short-term memory network to output a strain column vector, and the second feature matrix is ​​input into the strain response long short-term memory network to output a strain column vector. The two strain column vectors are weighted and fused to obtain the final strain column vector.

[0016] Furthermore, the temperature long short-term memory network includes:

[0017] The input layer is used to receive the first feature matrix;

[0018] The LSTM layer is used to process the first feature matrix, and it includes a gating control module, a memory update module, and a memory output module.

[0019] The gating control module initializes the hidden state as a vector in the same row and column as the strain column vector, but with all elements set to 0. For any time within the previous N time steps, the hidden vector from the previous time step and the temperature column vectors of each key location in the first feature matrix at that time step are summed using the input weight vector at that time step. After adding the input bias at that time step, the input is passed to the Sigmoid activation function to form the input gate. The temperature difference between each key location at that time step and the previous time step is calculated. Based on the temperature difference, the time period of each key location is divided into several consecutive time steps with the same temperature trend. For any time step within the same consecutive time step, the temperature difference between that time step and all previous time steps within that consecutive time step is summed to form the temperature accumulation at that time step. The temperature accumulation of all key locations at the same time step is used to form the temperature accumulation vector. For any time step within the N time steps, the temperature accumulation vector at that time step and the temperature column vectors of each key location at that time step are summed using the forgetting weight vector. After adding the forgetting bias at that time step, the sum is negative and passed to the Sigmoid activation function to form the forgetting gate.

[0020] It should be noted that since a vector is being activated, when the activation function is applied, each element in the vector is activated and the activated element replaces the original element.

[0021] In the health monitoring of steel structures in gymnasiums, the effect of temperature on structural strain is not instantaneous, but rather a physical process with a significant time-cumulative effect. Traditional LSTM forget gates rely solely on the current input and hidden state, making it difficult to model the long-term cumulative characteristics of temperature changes.

[0022] The memory update module, for any given time, sums the temperature column vectors at each key position in the first feature matrix at that time and the hidden vectors from the previous time through the state weight vector at that time, and then, after adding the state bias at that time, it connects to the tanh activation function to form a candidate memory state.

[0023] For the first time step, its candidate memory states are taken as memory states.

[0024] For all time points except the first time point, calculate the Hadamard product of the forget gate and the memory state of the previous time point, and the Hadamard product of the candidate memory state of the current time point and the input gate. The sum of the two Hadamard products constitutes the memory state of the current time point.

[0025] Fully connected layers are used to map the output of the LSTM layer onto the strain column vector;

[0026] The output layer is used to output the strain column vector.

[0027] Furthermore, the logic for dividing several consecutive moments is as follows: if the temperature difference between a moment and the previous moment is greater than 0, then the state at that moment is called a positive state; if the temperature difference between that moment and the previous moment is not greater than 0, then the state at that moment is called a negative state.

[0028] Sort the moments within the same time period in chronological order, take the first moment as the starting point of the first consecutive moment, traverse the subsequent moments until the state of a moment is inconsistent with the starting point for the first time, take this moment as the end point of the first consecutive moment, and take the next moment as the starting point of the second consecutive moment, traverse the subsequent moments again, and repeat the iteration to divide several consecutive moments.

[0029] Furthermore, a first component strain response threshold, a second component strain response threshold, and a component abnormal response number threshold are preset, wherein the first component strain response threshold is less than the second component strain response threshold;

[0030] The number of components whose strain response is greater than the strain response threshold of the first component but not greater than the strain response threshold of the second component is called the first response number. The number of components whose strain response in key parts is greater than the strain response threshold of the second component is called the second response number.

[0031] If the number of second responses is greater than 0, the steel structure performance is unqualified.

[0032] If the number of second responses is 0 and the number of first responses is not less than the threshold for the number of abnormal responses of a component, then the performance of the steel structure is unqualified.

[0033] If the number of second responses is 0, the number of first responses is less than the threshold for the number of abnormal responses of a component, and the number of first responses is not 0, then the steel structure performance is qualified.

[0034] If the number of first and second responses are both equal to 0, then the steel structure has excellent performance.

[0035] This invention also provides a performance evaluation system for gymnasium steel structures based on multimodal data fusion. The system is used to implement the aforementioned performance evaluation method for gymnasium steel structures based on multimodal data fusion, specifically including:

[0036] The data acquisition module is used to acquire strain response time-series data and temperature time-series data of key parts of the stadium's steel structure over multiple historical time periods;

[0037] The data analysis module is used to construct a strain column vector from the strain response data at the last sampling time in the time series data, construct a first feature matrix from the temperature data at the remaining sampling times, and construct a second feature matrix from the strain response data at the remaining sampling times.

[0038] The model building module is used to construct and train a two-layer long short-term memory network with the first feature matrix and the second feature matrix as input and the strain column vector as output. The two-layer long short-term memory network includes a strain response long short-term memory network and a temperature long short-term memory network. In the temperature long short-term memory network, the temperature accumulation is determined based on the temperature trend, and the forgetting gate is constructed based on the temperature accumulation.

[0039] The model prediction module is used to acquire the strain response time series data and temperature time series data of key parts of the steel structure of the stadium in the current time period, and to form the first feature matrix and the second feature matrix, which are then input into the trained two-layer long short-term memory network to predict the strain response data at the next sampling time.

[0040] The performance evaluation module is used to compare the strain response data of each key part at the next sampling time with the preset threshold, and classify the steel structure performance into excellent, qualified and unqualified based on the comparison results.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention achieves dynamic prediction and performance evaluation of the strain response of a gymnasium's steel structure by fusing multimodal data such as ambient temperature, load data, and structural temperature, combined with finite element simulation and a two-layer long short-term memory network (LSTM) for time-series modeling. This method accurately captures changes in structural behavior under the combined effects of temperature accumulation and load, improving prediction accuracy and timeliness. Through recursive prediction and threshold grading mechanisms, it enables intelligent grading of structural performance (excellent, qualified, unqualified), providing a reliable basis for structural health monitoring and maintenance decisions. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0044] Figure 2 The graph shows the changes in the actual and predicted strain response of a key component as a function of training rounds.

[0045] Figure 3 The graph shows the absolute difference between the actual and predicted strain responses of a key component as a function of training rounds.

[0046] Figure 4 A graph showing the Euclidean distance between the actual and predicted strain column vectors as a function of training epochs;

[0047] Figure 5 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0049] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0050] Example:

[0051] Please see Figure 1-5 The present invention provides a technical solution:

[0052] A method for evaluating the performance of steel structures in gymnasiums based on multimodal data fusion, comprising the following steps:

[0053] Step 1: Obtain strain response time-series data and temperature time-series data of key components of the stadium's steel structure over multiple historical time periods;

[0054] The stadium has a large number of steel structural components, making it impossible to monitor them all. Selecting densely packed areas as "representatives" can reflect the dynamic response trend of the overall structure and has good representativeness. It can cover high-stress and high-risk areas with fewer monitoring points, thereby improving monitoring efficiency.

[0055] The key part is the center of the densely represented component and its connection position with other components, and the time period is a time period consisting of N+1 consecutive sampling moments.

[0056] The center of densely packed components often represents a region with high local stiffness or complex load transfer in the structure, and its strain response can reflect the overall stress state of that region.

[0057] Connection points are stress concentration areas in steel structures, and are prone to large strains due to load changes, temperature fluctuations, etc., making them sensitive areas for structural health.

[0058] The logic for determining dense representative components is as follows: a preset component quantity threshold is set. Starting from the leftmost and bottommost component of the stadium, the number of components in a one-cubic-decimeter space centered on that component is counted. If the number of components is greater than the component quantity threshold and there are no other dense representative components, then that component is called a dense representative component. This process is repeated for all components in the stadium.

[0059] For a given area, one dense representative component is sufficient to represent it. Therefore, considering cost and computational simplification, the condition "no other dense representative components exist" is set.

[0060] The cross-sectional dimensions of most individual steel structural components in a gymnasium (such as steel pipes and structural steel sections) are typically a few centimeters (larger components usually do not present strain response issues). One cubic decimeter can cover the cross-section of a typical small individual component under study and include its surrounding connection area, making it a reasonable spatial unit reflecting the local density. It should be noted that in this embodiment, if a component only partially appears in the corresponding one cubic decimeter space, it is still included in the component count within that one cubic decimeter space.

[0061] Setting a component quantity threshold can be achieved using existing technology. First, the total number of components in the stadium can be counted to obtain the stadium's volume (cubic decimeters). The total number of components can then be divided by the stadium's volume to obtain the number of components per unit space. This number can then be multiplied by 1.5 to 2.0 as the component quantity threshold. Alternatively, experts in the field can be invited to analyze and evaluate the design drawings to directly determine the component quantity threshold.

[0062] The strain response time series data and temperature time series data can be obtained by installing temperature sensors and strain sensors at the corresponding locations, or by installing load sensors at the corresponding locations based on strain data. The steel structure parameters can be obtained, and a strain finite element model corresponding to the finite element method can be constructed. The collected load is input into the constructed finite element model to obtain the corresponding strain response. This is a conventional technical method in this field and will not be elaborated here.

[0063] Step 2: Construct a strain column vector from the strain response data at the last sampling time in the time series data, construct a first feature matrix from the temperature data at the remaining sampling times, and construct a second feature matrix from the strain response data at the remaining sampling times.

[0064] The rows of the first and second characteristic matrices represent sampling times, the columns represent key locations, and the element values ​​are the temperature and strain responses at the key locations at the corresponding sampling times.

[0065] Step 3: Using the first feature matrix and the second feature matrix as input and the strain column vector as output, construct and train a two-layer long short-term memory network. The two-layer long short-term memory network includes a strain response long short-term memory network and a temperature long short-term memory network. In the temperature long short-term memory network, temperature accumulation is determined based on temperature trend, and a forgetting gate is constructed based on temperature accumulation.

[0066] The first feature matrix is ​​input into the temperature long short-term memory network to output a strain column vector, and the second feature matrix is ​​input into the strain response long short-term memory network to output a strain column vector. The two strain column vectors are weighted and fused to obtain the final strain column vector.

[0067] The formula is:

[0068]

[0069] in, For the final strain column vector, Output column vectors for the temperature long short-term memory network. To provide a column vector for the output of the long short-term memory network in response to strain, , The weight parameters are determined by the least squares method, which is an existing technique and will not be elaborated here.

[0070] Temperature long short-term memory networks include:

[0071] The input layer is used to receive the first feature matrix;

[0072] The LSTM layer is used to process the first feature matrix, and it includes a gating control module, a memory update module, and a memory output module.

[0073] The gating control module initializes the hidden state as a vector in the same row and column as the strain column vector, but with all elements set to 0. For any time within the previous N time steps, the hidden vector from the previous time step and the temperature column vectors of each key location in the first feature matrix at that time step are summed using the input weight vector at that time step. After adding the input bias at that time step, the input is passed to the Sigmoid activation function to form the input gate. The temperature difference between each key location at that time step and the previous time step is calculated. Based on the temperature difference, the time period of each key location is divided into several consecutive time steps with the same temperature trend. For any time step within the same consecutive time step, the temperature difference between that time step and all previous time steps within that consecutive time step is summed to form the temperature accumulation at that time step. The temperature accumulation of all key locations at the same time step is used to form the temperature accumulation vector. For any time step within the N time steps, the temperature accumulation vector at that time step and the temperature column vectors of each key location at that time step are summed using the forgetting weight vector. After adding the forgetting bias at that time step, the sum is negative and passed to the Sigmoid activation function to form the forgetting gate.

[0074] In the health monitoring of steel structures in gymnasiums, the effect of temperature on structural strain is not instantaneous, but rather a physical process with a significant time-cumulative effect. Traditional LSTM forget gates rely solely on the current input and hidden state, making it difficult to display the long-term cumulative characteristics of temperature changes. The temperature accumulation vector is obtained by accumulating the temperature difference over a continuous heating or cooling period. Its magnitude characterizes the degree of net temperature influence on the structure during that period. The greater the temperature accumulation, the more effective it is. Using the temperature accumulation vector as the input signal for the forget gate allows the model to dynamically adjust its memory retention strategy based on the accumulated temperature.

[0075] The stronger the thermal accumulation effect, the more significant the persistence of the structure's historical state's influence on the present; therefore, more relevant memory should be retained. Conversely, weaker temperature accumulation indicates a change in temperature trend, with temperature beginning to accumulate again, resulting in a completely new memory model. Previous memory patterns need to be forgotten, hence the choice of negative values ​​for reactivation. During diurnal or seasonal temperature cycles, structural responses exhibit hysteresis and memory effects. The forgetting gate based on temperature accumulation enables the model to retain memory of previous states during periods of sustained temperature trends, more accurately simulating the hysteresis characteristics of structural responses under cyclic loading. By sensing the continuous temperature change trend through the temperature accumulation vector, the model actively suppresses the forgetting of past states, making the network memory more closely match the continuous characteristics of thermal stress accumulation. This reduces predictive abrupt changes caused by premature memory discarding, improving the smoothness and reliability of strain time-series predictions.

[0076] The logic for dividing a series of consecutive moments is as follows: if the temperature difference between a given moment and the previous moment is greater than 0, then the state at that moment is called a positive state; if the temperature difference between a given moment and the previous moment is not greater than 0, then the state at that moment is called a negative state.

[0077] Sort the moments within the same time period in chronological order, take the first moment as the starting point of the first consecutive moment, traverse the subsequent moments until the state of a moment is inconsistent with the starting point for the first time, take this moment as the end point of the first consecutive moment, and take the next moment as the starting point of the second consecutive moment, traverse the subsequent moments again, and repeat the iteration to divide several consecutive moments.

[0078] In the monitoring of the steel structure of the gymnasium, temperature changes are not isolated fluctuations, but rather exhibit a clear phased trend of rising and falling temperatures, influenced by diurnal variation, season, and usage conditions. By determining the positive or negative value of the temperature difference between adjacent moments to define time periods with consistent trends, the model can fundamentally distinguish the different mechanisms by which heating and cooling processes affect the thermal stress of the structure. This trend is also crucial for the analysis of temperature accumulation. If the temperature trend changes frequently within a continuous time period (i.e., the continuous time periods defined in this example are constantly changing), the temperatures cancel each other out during the rises and falls, making accumulation difficult. If the trends are consistent, the temperature accumulation within the corresponding time period can be considered a linear process, effectively quantifying the temperature accumulation and providing a theoretical basis for the construction of the forgetting gate.

[0079] The memory update module, for any given time, sums the temperature column vectors at each key position in the first feature matrix at that time and the hidden vectors from the previous time through the state weight vector at that time, and then, after adding the state bias at that time, it connects to the tanh activation function to form a candidate memory state.

[0080] For the first time step, its candidate memory states are taken as memory states.

[0081] For all time points except the first time point, calculate the Hadamard product of the forget gate and the memory state of the previous time point, and the Hadamard product of the candidate memory state of the current time point and the input gate. The sum of the two Hadamard products constitutes the memory state of the current time point.

[0082] Fully connected layers are used to map the output of the LSTM layer onto the strain column vector;

[0083] The output layer is used to output the strain column vector.

[0084] Strain response long short-term memory networks can be constructed using existing technologies, including:

[0085] The input layer is used to connect to the second feature matrix;

[0086] The LSTM layer is used to perform feature processing on the data received from the input layer in order to capture the temporal dependencies in the time series data. The number of LSTM units in the LSTM layer is set between 50 and 100.

[0087] Fully connected layers are used to map the output of the LSTM layer onto the strain column vector;

[0088] The output layer is used to output the strain column vector.

[0089] The training process for the prediction model is as follows:

[0090] Strain response time-series data and temperature time-series data from multiple historical time periods are used to construct corresponding first feature matrices, second feature matrices, and corresponding column vectors, forming a sample dataset. The first and second feature matrices from multiple historical time periods in the sample dataset are used as input, and the corresponding column vectors are used as output to train a two-layer long short-term memory network.

[0091] During training, mean squared error is selected as the loss function, and the Adam optimization algorithm is selected to update the model weights. After training, the prediction model is evaluated using a test set. Specifically, mean squared error can be selected as the evaluation index. If the evaluation index reaches the expected value, the training is considered complete. Otherwise, the various weights and biases of the temperature long short-term memory network, as well as the parameters of the strain response long short-term memory network such as the number of LSTM units, the number of layers, the learning rate, and the batch size are adjusted based on the evaluation results and the model is trained again. The trainable weights in the attention mechanism are also adjusted. This is existing technology and will not be elaborated here.

[0092] The model parameters include batch size, feedback training period, and total training period. The feedback training period is set between one-tenth and one-fifth of the total training period. The batch size is the number of samples used in each training iteration, typically between 32 and 256. The total training period is the total number of training iterations, typically between 30 and 100. The feedback training period is used to determine the time point for obtaining model training feedback results. If this time point is too early, the model training feedback results will lack reference value, while if it is too late, the total model training time will be too long. Therefore, the feedback training period is set between one-tenth and one-fifth of the total training period to balance the reference value of model training feedback results and the efficiency of model training.

[0093] Please see Figure 2 , Figure 2 The graph shows the changes in the actual and predicted strain response of a key component as a function of training rounds.

[0094] Please see Figure 3 , Figure 3 The graph shows the absolute difference between the actual and predicted strain responses of a key component as a function of training rounds.

[0095] Please see Figure 4 , Figure 4 The graph shows the Euclidean distance between the actual and predicted strain column vectors as a function of training epochs.

[0096] Step 4: Obtain the strain response time series data and temperature time series data of key parts of the stadium steel structure in the current time period, and construct the first feature matrix and the second feature matrix. Input them into the trained two-layer long short-term memory network to predict the strain response data at the next sampling time.

[0097] The strain response time series data and temperature time series data from the 2nd to the N+1th time in the current time period are used to construct the first feature matrix and the second feature matrix.

[0098] Step 5: Compare the strain response data of each key part at the next sampling time with the preset threshold, and classify the steel structure performance into excellent, qualified and unqualified based on the comparison results.

[0099] A first component strain response threshold, a second component strain response threshold, and a component abnormal response number threshold are preset, wherein the first component strain response threshold is less than the second component strain response threshold;

[0100] The preset first component strain response threshold, second component strain response threshold, and component abnormal response number threshold can be achieved using existing technology. Historical data from multiple inspections of steel structure components assessed as abnormal, along with their corresponding strain responses, are obtained and sorted from smallest to largest. The average of the top 10% is used as the second component strain response threshold. While 80% of the second component strain response thresholds show no abnormality, a trend towards abnormality is observed. This 80% of the second component strain response threshold is then used as the first component strain response threshold. The threshold for the number of abnormal responses in critical components is set at 20%. If more than 20% of critical components simultaneously exhibit an abnormal trend (even if not exceeding the danger threshold), exceeding the range of random fluctuations, it indicates that the structure may be in a systematic abnormal state.

[0101] The number of components whose strain response is greater than the strain response threshold of the first component but not greater than the strain response threshold of the second component is called the first response number. The number of components whose strain response in key parts is greater than the strain response threshold of the second component is called the second response number.

[0102] If the number of second responses is greater than 0, an abnormal component will appear at the next sampling time, and the steel structure performance will be unqualified.

[0103] If the number of second responses is 0 and the number of first responses is not less than the threshold for the number of abnormal responses of components, even if no abnormal components appear at the next sampling time, many components show an abnormal trend and the danger is still very high, then the performance of the steel structure is unqualified.

[0104] If the number of second responses is equal to 0, the number of first responses is less than the threshold for the number of abnormal responses of components, and the number of first responses is not equal to 0, although no abnormal components appear at the next sampling time, there is a trend that all components are abnormal, but the number is small. After the risk is explained, the steel structure performance is qualified.

[0105] If the number of first responses and the number of second responses are both equal to 0, there is no danger, and the steel structure has excellent performance.

[0106] Please see Figure 5 The present invention further provides a performance evaluation system for steel structures of gymnasiums based on multimodal data fusion. This system is used to implement the aforementioned performance evaluation method for steel structures of gymnasiums based on multimodal data fusion, specifically including:

[0107] The data acquisition module is used to acquire strain response time-series data and temperature time-series data of key parts of the stadium's steel structure over multiple historical time periods;

[0108] The data analysis module is used to construct a strain column vector from the strain response data at the last sampling time in the time series data, construct a first feature matrix from the temperature data at the remaining sampling times, and construct a second feature matrix from the strain response data at the remaining sampling times.

[0109] The model building module is used to construct and train a two-layer long short-term memory network with the first feature matrix and the second feature matrix as input and the strain column vector as output. The two-layer long short-term memory network includes a strain response long short-term memory network and a temperature long short-term memory network. In the temperature long short-term memory network, the temperature accumulation is determined based on the temperature trend, and the forgetting gate is constructed based on the temperature accumulation.

[0110] The model prediction module is used to acquire the strain response time series data and temperature time series data of key parts of the steel structure of the stadium in the current time period, and to form the first feature matrix and the second feature matrix, which are then input into the trained two-layer long short-term memory network to predict the strain response data at the next sampling time.

[0111] The performance evaluation module is used to compare the strain response data of each key part at the next sampling time with the preset threshold, and classify the steel structure performance into excellent, qualified and unqualified based on the comparison results.

[0112] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that cannot be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for evaluating the performance of steel structures in gymnasiums based on multimodal data fusion, characterized in that, The specific steps include: Step 1: Obtain strain response time-series data and temperature time-series data of key components of the stadium's steel structure over multiple historical time periods; Step 2: Construct a strain column vector from the strain response data at the last sampling time in the time series data, construct a first feature matrix from the temperature data at the remaining sampling times, and construct a second feature matrix from the strain response data at the remaining sampling times. Step 3: Using the first feature matrix and the second feature matrix as input and the strain column vector as output, construct and train a two-layer long short-term memory network. The two-layer long short-term memory network includes a strain response long short-term memory network and a temperature long short-term memory network. In the temperature long short-term memory network, temperature accumulation is determined based on temperature trend, and a forgetting gate is constructed based on temperature accumulation. Step 4: Obtain the strain response time series data and temperature time series data of key parts of the stadium steel structure in the current time period, and construct the first feature matrix and the second feature matrix. Input them into the trained two-layer long short-term memory network to predict the strain response data at the next sampling time. Step 5: Compare the strain response data of each key part at the next sampling time with the preset threshold, and classify the steel structure performance into excellent, qualified and unqualified based on the comparison results; The key part is the center of the densely represented component and its connection position with other components, and the time period is a time period consisting of N+1 consecutive sampling moments; The rows of the first and second feature matrices represent sampling times, the columns represent key locations, and the element values ​​are the temperature and strain responses at the key locations at the corresponding sampling times. The first feature matrix is ​​input into the temperature long short-term memory network to output a strain column vector, and the second feature matrix is ​​input into the strain response long short-term memory network to output a strain column vector. The two strain column vectors are weighted and fused to obtain the final strain column vector.

2. The method for evaluating the performance of a gymnasium steel structure based on multimodal data fusion according to claim 1, characterized in that: The logic for determining dense representative components is as follows: a preset component quantity threshold is set. Starting from the leftmost and bottommost component of the stadium, the number of components in a one-cubic-decimeter space centered on that component is counted. If the number of components is greater than the component quantity threshold and there are no other dense representative components, then that component is called a dense representative component. This process is repeated for all components in the stadium.

3. The method for evaluating the performance of a gymnasium steel structure based on multimodal data fusion according to claim 1, characterized in that: Temperature long short-term memory networks include: The input layer is used to receive the first feature matrix; The LSTM layer is used to process the first feature matrix, and it includes a gating control module, a memory update module, and a memory output module. The gating control module initializes the hidden state as a vector in the same row and column as the strain column vector, but with all elements being 0. For any time within the previous N time steps, the hidden vector from the previous time step and the temperature column vectors of each key position in the first feature matrix at that time step are summed using the input weight vector at that time step. After adding the input bias at that time step, the input is connected to the Sigmoid activation function to form an input gate. The temperature difference between each key part at that time step and the previous time step is calculated. Based on the temperature difference, the time period of each key part is divided into several consecutive time steps with the same temperature trend. For any time step within the same consecutive time step, the temperature difference between that time step and all previous time steps within that consecutive time step is summed to form the temperature accumulation at that time step. The temperature accumulation of all key parts at the same time step is used to form a temperature accumulation vector. For any time step within the N time steps, the temperature accumulation vector at that time step and the temperature column vectors of each key position at that time step are summed using the forgetting weight vector. After adding the forgetting bias at that time step, the sum is negative and connected to the Sigmoid activation function to form a forgetting gate. The memory update module, for any given time, sums the temperature column vectors at each key position in the first feature matrix at that time and the hidden vectors from the previous time through the state weight vector at that time, and then, after adding the state bias at that time, it connects to the tanh activation function to form a candidate memory state. For the first time step, its candidate memory states are taken as memory states. For all time points except the first time point, calculate the Hadamard product of the forget gate and the memory state of the previous time point, and the Hadamard product of the candidate memory state of the current time point and the input gate. The sum of the two Hadamard products constitutes the memory state of the current time point. Fully connected layers are used to map the output of the LSTM layer onto the strain column vector; The output layer is used to output the strain column vector.

4. The method for evaluating the performance of a gymnasium steel structure based on multimodal data fusion according to claim 3, characterized in that: The logic for dividing a series of consecutive moments is as follows: if the temperature difference between a moment and the previous moment is greater than 0, then the state at that moment is called a positive state; otherwise, the state at that moment is called a negative state. Sort the moments within the same time period in chronological order, take the first moment as the starting point of the first consecutive moment, traverse the subsequent moments until the state of a moment is inconsistent with the starting point for the first time, take this moment as the end point of the first consecutive moment, and take the next moment as the starting point of the second consecutive moment, traverse the subsequent moments again, and repeat the iteration to divide several consecutive moments.

5. The method for evaluating the performance of a gymnasium steel structure based on multimodal data fusion according to claim 1, characterized in that: A first component strain response threshold, a second component strain response threshold, and a component abnormal response number threshold are preset, wherein the first component strain response threshold is less than the second component strain response threshold; The number of key components whose strain response is greater than the strain response threshold of the first component but not greater than the strain response threshold of the second component is called the first response number. The number of components whose strain response at key components is greater than the strain response threshold of the second component is called the second response number. If the number of second responses is greater than 0, the steel structure performance is unqualified. If the number of second responses is 0 and the number of first responses is not less than the threshold for the number of abnormal responses of a component, then the performance of the steel structure is unqualified. If the number of second responses is 0, the number of first responses is less than the threshold for the number of abnormal responses of a component, and the number of first responses is not 0, then the steel structure performance is qualified. If the number of first and second responses are both equal to 0, then the steel structure has excellent performance.

6. A performance evaluation system for steel structures of gymnasiums based on multimodal data fusion, characterized in that: The system is used to implement the performance evaluation method for steel structures of gymnasiums based on multimodal data fusion as described in any one of claims 1-5, specifically including: The data acquisition module is used to acquire strain response time-series data and temperature time-series data of key parts of the stadium's steel structure over multiple historical time periods; The data analysis module is used to construct a strain column vector from the strain response data at the last sampling time in the time series data, construct a first feature matrix from the temperature data at the remaining sampling times, and construct a second feature matrix from the strain response data at the remaining sampling times. The model building module is used to construct and train a two-layer long short-term memory network with the first feature matrix and the second feature matrix as input and the strain column vector as output. The two-layer long short-term memory network includes a strain response long short-term memory network and a temperature long short-term memory network. In the temperature long short-term memory network, the temperature accumulation is determined based on the temperature trend, and the forgetting gate is constructed based on the temperature accumulation. The model prediction module is used to acquire the strain response time series data and temperature time series data of key parts of the steel structure of the stadium in the current time period, and to form the first feature matrix and the second feature matrix, which are then input into the trained two-layer long short-term memory network to predict the strain response data at the next sampling time. The performance evaluation module is used to compare the strain response data of each key part at the next sampling time with the preset threshold, and classify the steel structure performance into excellent, qualified and unqualified based on the comparison results.

Images