Analysis method for predicting UHPC shear resistance based on neural network

Abstract

The invention relates to the technical field of UHPC (Ultra High Performance Concrete) shear resistance research, and discloses an analysis method for predicting UHPC shear resistance based on a neural network, which is characterized by comprising the following steps: data acquisition and preprocessing, selection of a UHPC test piece for a shear failure experiment, acquisition of damaged image and sound wave data by using an ultrasonic detector and a camera, and calculation of the UHPC shear resistance based on the neural network. Graying processing is carried out on the image data; building and training a model, building a prediction model based on a depth separable convolutional neural network, taking the preprocessed image and sound wave data as input, building a sample data set, and carrying out model training; model verification and optimization: verifying the model by using the test set, evaluating the prediction performance, and optimizing the model according to the verification result; and deploying and applying the model. The analysis method for predicting the shear resistance of the UHPC based on the neural network aims to realize rapid and accurate prediction of the shear resistance of the UHPC prefabricated member by combining image processing and deep learning technologies.

Description

Technical Field

[0001] The invention relates to the technical field of UHPC shear performance research, and in particular to an analysis method for predicting UHPC shear performance based on a neural network. Background Art

[0002] As an innovative high-performance concrete material, UHPC (ultra-high performance concrete) has played an important role in many engineering structures due to its excellent mechanical properties and outstanding durability. However, despite the increasingly widespread application of UHPC, the accurate prediction of its shear performance has always been a major challenge facing the engineering field. Shear performance, as an important indicator to measure the material's ability to resist damage under shear force, is crucial to ensuring the safety and stability of engineering structures. However, due to the complex microstructure and multi-factor influence of UHPC, the prediction of its shear performance becomes particularly complex and difficult. Therefore, how to develop an efficient and accurate prediction method to scientifically and objectively evaluate the shear performance of UHPC has become a key issue that needs to be urgently solved in the current engineering field.

[0003] The traditional UHPC shear performance prediction method mainly relies on a large amount of experimental data. Although this method can provide relatively reliable results, its inherent defects cannot be ignored. First, the acquisition of experimental data is often time-consuming and labor-intensive, requiring a large amount of specimen preparation, loading testing and data analysis, which not only increases costs but also prolongs the R&D cycle. Secondly, the experimental method is difficult to fully cover various working conditions, especially for the prediction of shear performance under extreme or special conditions. The experimental data is often insufficient or difficult to obtain, so there are certain defects. Summary of the invention

[0004] The purpose of the present invention is to realize the rapid and accurate prediction of the shear resistance of UHPC preforms by combining image processing and deep learning technology, and propose an analysis method for predicting the shear resistance of UHPC based on neural network.

[0005] The technical solution of the present invention to solve the above technical problems is as follows:

[0006] An analysis method for predicting the shear resistance of UHPC based on a neural network, characterized by comprising the following steps:

[0007] S10, data acquisition and preprocessing, select UHPC specimens for shear failure experiments, use ultrasonic detectors and cameras to collect images and sound wave data after damage, and grayscale the image data;

[0008] S20, model building and training, building a prediction model based on a deep separable convolutional neural network, taking the preprocessed image and sound wave data as input, constructing a sample data set, and performing model training;

[0009] S30, model validation and optimization, using the test set to validate the model, evaluate the prediction performance, and optimize the model based on the validation results;

[0010] S40, model deployment and application, deploy the trained model on the server to analyze, count and display the shear performance prediction results.

[0011] Based on the above technical solution, the present invention can also be improved as follows.

[0012] Furthermore, in step S10, the grayscale processing adopts the grayscale weighted average method, and the formula is:

[0013] f(t)=0.30f R (t)+0.59f G (t)+0.11f B (t)

[0014] Wherein, f(t) represents the pixel matrix of the grayscale image f to be identified at the iteration time;

[0015] f R (t), f G (t), f B (t) represents the iterated color, green, and blue channel pixel matrices respectively.

[0016] Furthermore, in step S20, in the deep convolution stage of the deep separable convolutional neural network, the convolution kernel size, step size and padding method of each channel need to be pre-set according to the UHPC image characteristics to capture subtle textures; in the point-by-point convolution stage, a 1x1 convolution kernel is used to fuse information between channels to generate an output feature map with a higher level of abstraction.

[0017] Furthermore, in step S20, the binary cross entropy loss function used in model training needs to be combined with a weight decay term to prevent the model from falling into a local optimal solution during training. The formula is modified as follows:

[0018]

[0019] Among them, w i,j is the category weight, which is used to balance the imbalance between categories; λ is the weight decay coefficient, and θ is the model parameter.

[0020] Furthermore, in step S30, the cross-validation method adopts K-fold cross-validation, and the K value needs to be reasonably selected according to the size of the data set and the computing resources to ensure the stability and reliability of the validation results. In each cross-validation, the prediction accuracy, recall rate and F1 score indicators of the model need to be recorded to comprehensively evaluate the model performance.

[0021] Furthermore, in step S30, the model optimization also includes introducing an early stopping method, which terminates the training in advance when the validation set loss no longer decreases to avoid model overfitting; at the same time, the learning rate adjustment strategy adopts cosine annealing decay to fine-tune the model parameters in the later stage of training to further improve the model performance.

[0022] Furthermore, before step S10, distributed storage technology should be used to back up the original image and acoustic wave data to multiple storage nodes in real time to ensure the security and integrity of the data during the collection and cleaning process. During the data cleaning process, incremental backup should be performed for each new or modified data to reduce backup overhead and improve backup efficiency. In the event of data loss or damage, backup data can be used for rapid recovery to ensure the continuity and stability of the analysis process. Backup data should be verified regularly to ensure the integrity and consistency of the backup data.

[0023] Furthermore, the image denoising in the data cleaning step adopts an adaptive median filtering algorithm, which automatically adjusts the filter window size according to the local noise situation of the image, removes noise while retaining the image edge details. Image enhancement adopts a method that combines histogram equalization and contrast stretching. The acoustic signal filtering adopts a bandpass filter. According to the frequency characteristics of the acoustic signal, an appropriate filtering range is set to remove low-frequency noise and high-frequency interference and retain useful acoustic information.

[0024] Furthermore, the analysis method further includes a model evaluation and optimization iteration step, which is performed after step S30 and includes:

[0025] Performance evaluation: Use the test set to comprehensively evaluate the model, including accuracy, recall, F1 score indicators, and the stability and robustness of the model under different conditions;

[0026] Error analysis: Conduct in-depth analysis on samples with model prediction errors to identify the causes of the errors, data noise and feature extraction, and provide direction for subsequent optimization;

[0027] Optimization iteration: Based on the results of performance evaluation and error analysis, the model is optimized, including adjusting the network structure, increasing training data, and improving feature extraction methods. The model is then retrained and verified until the model performance meets the requirements.

[0028] Version management: Perform version management on each optimized model, record model performance changes and optimization strategies, and facilitate subsequent tracking and backtracking.

[0029] Furthermore, in the result feedback step, the feedback content includes not only the prediction result itself, but also the confidence of the prediction result, the prediction basis and the model performance evaluation report. The specific operation steps are as follows:

[0030] Visual display: Use charts and images to visualize the prediction results and their confidence levels, making it easier for engineers or construction personnel to understand quickly.

[0031] Interactive feedback: Provides an interactive interface that allows engineers or construction personnel to adjust prediction parameters according to actual needs, such as adjusting confidence thresholds and selecting specific features for prediction, thereby improving the flexibility and accuracy of feedback;

[0032] Report generation: Automatically generate detailed documents including prediction results, confidence, prediction basis and model performance evaluation report for engineers or construction personnel to refer to and make decisions;

[0033] Continuous tracking: Establish a database that associates the prediction results with the actual engineering performance, continuously track the accuracy and reliability of the prediction results, and provide data support for subsequent optimization of the model.

[0034] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0035] In the data acquisition and preprocessing stage, the present invention not only collects the image data of the UHPC specimen after shear failure, but also innovatively introduces the acoustic wave data. Through the joint use of the ultrasonic detector and the camera, the physical information of the specimen during the failure process can be more comprehensively captured, providing richer data support for subsequent model training. At the same time, the image data is grayed, which effectively reduces the data dimension and improves the data processing efficiency. Secondly, in the model building and training stage, the deep separable convolutional neural network is used as the prediction model, which can efficiently extract key features from the image and acoustic wave data, and establish a complex relationship model between the shear resistance of UHPC and various influencing factors. Compared with traditional experimental methods, this method not only greatly shortens the prediction cycle, but also improves the accuracy and generalization ability of the prediction. Thirdly, in the model verification and optimization stage, the test set verifies the model. By evaluating the prediction performance, the problems in the model can be discovered in time and optimized. This iterative optimization process enables the model to continuously approach the real situation and improve the reliability of the prediction. Finally, in the model deployment and application stage, the trained model is deployed on the server to realize the analysis, statistics and display of the shear performance prediction results. This online application method not only facilitates engineers' query and use, but also provides convenience for subsequent data analysis and scientific research. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a method flow chart of an analysis method for predicting the shear resistance of UHPC based on a neural network. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Combination Figure 1 As shown, an analysis method for predicting the shear resistance of UHPC based on a neural network of the present invention comprises the following steps:

[0039] S10, data acquisition and preprocessing, select UHPC specimens for shear failure experiments, use ultrasonic detectors and cameras to collect images and sound wave data after damage, and grayscale the image data;

[0040] S20, model building and training, building a prediction model based on a deep separable convolutional neural network, taking the preprocessed image and sound wave data as input, constructing a sample data set, and performing model training;

[0041] S30, model validation and optimization, using the test set to validate the model, evaluate the prediction performance, and optimize the model based on the validation results;

[0042] S40, model deployment and application, deploy the trained model on the server to analyze, count and display the shear performance prediction results.

[0043] In a preferred embodiment, the present invention can be further configured as follows: In step S10, the grayscale processing adopts the grayscale weighted average method, and the formula is:

[0044] f(t)=0.30f R (t)+0.59f G (t)+0.11f B (t)

[0045] Wherein, f(t) represents the pixel matrix of the grayscale image f to be identified at the iteration time;

[0046] f R (t), f G (t), f B(t) represents the pixel matrix of the red, green, and blue channels of the iteration, respectively. The grayscale weighted average method is used, and a detailed formula is given. The choice of this method is not accidental, but is based on the careful consideration of the characteristics of UHPC images. After shear failure, the image of UHPC materials often contains rich color information and texture details, which are crucial for the subsequent training of neural network models. However, directly processing color images not only increases the computational complexity, but also may introduce unnecessary noise. Therefore, the grayscale weighted average method is selected. This method can more reasonably retain the key information in the image and remove redundant color information by assigning different weights (0.30, 0.59, and 0.11) to the red, green, and blue channels. In addition, the iterative process of grayscale processing is further supplemented. In practical applications, due to the influence of factors such as image acquisition equipment and lighting conditions, the original image may have certain noise and interference. In order to obtain a higher quality grayscale image, the concept of iterative processing is introduced. In each iteration, the image is grayscaled, and the weight is adjusted or other optimization operations are performed according to the processing results. Through multiple iterations, the ideal grayscale image can be gradually approached, providing more accurate data support for subsequent neural network model training.

[0047] The grayscale weighted average method and iterative processing have significant beneficial effects on the UHPC shear performance prediction method in the present invention. First, the grayscale processing can remove the color information in the image and reduce the data dimension, thereby reducing the computational complexity of the neural network. At the same time, by assigning different weights to different color channels, the key information in the image can be more reasonably retained and the accuracy of feature extraction can be improved. Secondly, the iterative processing process can gradually remove the noise and interference in the image and improve the quality of the grayscale image. This is crucial for subsequent neural network model training, because high-quality input data can improve the training effect of the model, so that it can better learn the complex relationship between the shear performance of UHPC and image features. In addition, it also has good versatility and scalability. As a commonly used image processing technology, the grayscale weighted average method can be widely used in various image processing tasks. The concept of iterative processing provides a flexible and adjustable optimization strategy that can be customized according to actual needs.

[0048] In a preferred embodiment, the present invention can be further configured as follows: in step S20, in the deep convolution stage of the deep separable convolutional neural network, the convolution kernel size, step size and padding method of each channel need to be pre-set according to the UHPC image characteristics to capture subtle textures; in the point-by-point convolution stage, a 1x1 convolution kernel is used to fuse information between channels to generate an output feature map with a higher level of abstraction. In step S20, the construction and application of the deep separable convolutional neural network are further refined and optimized. Specifically, in the deep convolution stage, it not only relies on the general convolution kernel setting, but is personalized and pre-set according to the UHPC image characteristics. This includes the careful selection of the convolution kernel size, step size and padding method of each channel, aiming to more accurately capture the subtle textures and key features in the UHPC image. This customized convolution kernel setting can more effectively extract image information closely related to the shear resistance of UHPC, laying a solid foundation for subsequent analysis and prediction. In the point-by-point convolution stage, a 1x1 convolution kernel is used to fuse information between channels. The purpose of this step is to combine the feature maps of multiple channels extracted in the deep convolution stage to generate an output feature map with a higher level of abstraction. Through the point-by-point convolution operation of the 1x1 convolution kernel, cross-channel information exchange and integration can be achieved, thereby extracting more comprehensive and in-depth image features. This design not only improves the expressiveness of the model, but also helps to improve the accuracy of prediction.

[0049] The personalized convolution kernel setting enables the model to more accurately capture the subtle textures and key features in the UHPC images, thereby improving the accuracy and efficiency of feature extraction. Secondly, the 1x1 convolution kernel design in the point-by-point convolution stage realizes cross-channel information fusion and integration, further improving the model's expressive power and prediction accuracy. In addition, this refined and optimized deep separable convolutional neural network also has good generalization ability. Since the model fully learns the characteristics and rules of UHPC images during the training process, it can still give relatively accurate shear performance prediction results when facing new and unknown UHPC specimens. This not only improves the reliability of the prediction, but also provides strong technical support for the wide application of UHPC in engineering structures.

[0050] In a preferred embodiment, the present invention can be further configured as follows: In step S20, the binary cross entropy loss function used in model training needs to be combined with a weight decay term to prevent the model from falling into a local optimal solution during the training process. The formula is modified as follows:

[0051]

[0052] Among them, w i,jis the category weight, which is used to balance the imbalance between categories; λ is the weight decay coefficient, and θ is the model parameter. The setting of the category weight wi,j does not rely solely on simple empirical values ​​or default values, but is dynamically adjusted according to the actual situation of the UHPC shear performance experimental data. Specifically, the number and distribution of samples in each category in the experimental data and their importance in predicting shear performance are analyzed, and then a suitable weight is assigned to each category based on this information. This method of dynamically adjusting category weights can more effectively balance the imbalance problem between categories and improve the prediction accuracy of the model. In addition, a weight decay term λ∥θ∥2 is introduced into the loss function to prevent the model from overfitting during training and falling into a local optimal solution. The weight decay term limits the complexity of the model parameters by regularizing the model parameters θ, thereby improving the generalization ability of the model. In practical applications, the weight decay coefficient λ is carefully adjusted according to the model training situation and data characteristics to find the best balance point, which not only avoids overfitting but also ensures the prediction performance of the model.

[0053] By dynamically adjusting the category weights, the imbalance problem between categories can be more effectively balanced, thereby improving the prediction accuracy of the model. This is particularly important for materials such as UHPC, which have complex microstructures and are affected by multiple factors, because data of different categories often have different distributions and importance. Secondly, the introduction of the weight decay term effectively prevents the model from overfitting during the training process and improves the generalization ability of the model. This means that when the model faces new and unknown UHPC specimens, it can still give relatively accurate shear performance prediction results. This is of great significance to the engineering field, because in practical applications, it is often necessary to predict the shear performance of UHPC specimens that have not yet been tested. In addition, the optimization of the loss function calculation process further improves the training efficiency and prediction speed of the model. This makes the method of the present invention more efficient and convenient in practical applications, providing strong technical support for engineers.

[0054] In a preferred embodiment, the present invention can be further configured as follows: In step S30, the cross-validation method adopts K-fold cross-validation, and the K value needs to be reasonably selected according to the size of the data set and the computing resources to ensure the stability and reliability of the verification result. In each cross-validation, the prediction accuracy, recall rate and F1 score indicators of the model need to be recorded to comprehensively evaluate the model performance. The selection of the K value of the K-fold cross-validation is a key step, which is directly related to the stability and reliability of the verification result. The K value is not fixed to a specific number, but is flexibly selected according to the size of the data set and the actual situation of the computing resources. Specifically, when the data set is large and the computing resources are sufficient, a larger K value (such as K=10) can be selected to obtain a more stable and reliable verification result; and when the data set is small or the computing resources are limited, a smaller K value (such as K=5 or K=3) can be selected to reduce the amount of calculation and speed up the verification process. In addition, in each cross-validation, the three key indicators of the prediction accuracy, recall rate and F1 score of the model are recorded to comprehensively evaluate the performance of the model. The prediction accuracy reflects the proportion of samples that the model predicts correctly and is a basic indicator for evaluating model performance; the recall rate reflects the model's ability to identify positive samples, that is, how many positive samples the model can correctly identify; and the F1 score is the harmonic mean of the prediction accuracy and recall rate, which takes these two indicators into consideration and can more comprehensively reflect the performance of the model. By recording these three indicators at the same time, we can have a more comprehensive understanding of the advantages and disadvantages of the model and provide strong support for subsequent optimization work.

[0055] The K-fold cross-validation method is used in combination with the prediction accuracy, recall rate and F1 score indicators to comprehensively evaluate the model performance, which has a significant beneficial effect on the UHPC shear performance prediction method in the present invention. First, the K-fold cross-validation method can effectively avoid overfitting and underfitting problems and improve the stability and reliability of the verification results. Through multiple cross-validations, more accurate and reliable model performance evaluation results can be obtained, providing strong support for subsequent optimization work. Secondly, by simultaneously recording the three key indicators of prediction accuracy, recall rate and F1 score, the performance characteristics of the model can be more comprehensively understood. These three indicators reflect the prediction ability and recognition ability of the model from different angles. By comprehensively analyzing these three indicators, the advantages and disadvantages of the model can be more accurately judged, providing targeted guidance for subsequent optimization work. In addition, the actual situation of the data set size and computing resources is also considered, and the K value is flexibly selected for cross-validation, which improves the efficiency and accuracy of the verification process. At the same time, by introducing technical means such as parallel computing and data enhancement, the effectiveness and practicality of the method are further improved.

[0056] In a preferred embodiment, the present invention can be further configured as follows: in step S30, the model optimization also includes the introduction of an early stopping method, when the validation set loss no longer decreases, the training is terminated in advance to avoid model overfitting; at the same time, the learning rate adjustment strategy adopts cosine annealing decay to finely adjust the model parameters in the later stage of training to further improve the model performance. Regarding the introduction of the early stopping method, its implementation details are clarified. During the model training process, the loss change on the validation set is continuously monitored. When the validation set loss no longer decreases significantly within a certain number of consecutive iterations (such as 10 consecutive iterations), it is considered that the model is close to the optimal state, and the training is terminated in advance to avoid model overfitting. This strategy can effectively save computing resources and improve the generalization ability of the model. Secondly, regarding the learning rate adjustment strategy, the cosine annealing decay method is selected. Cosine annealing decay is a method for periodically adjusting the learning rate, which can adjust the model parameters more finely in the later stage of training to further improve the model performance. In the specific implementation, an initial learning rate and a minimum learning rate are set, and then the learning rate is decayed according to the shape of the cosine function according to the training rounds and the preset period. This adjustment method can make the model converge quickly in the early stage of training, while maintaining a certain update ability in the later stage of training to avoid falling into the local optimal solution. In addition, in order to further improve the effect of model optimization, other optimization techniques can be combined, such as weight decay (L2 regularization) and gradient clipping. Weight decay can prevent the model parameters from being too large to a certain extent and improve the generalization ability of the model; while gradient clipping can prevent the gradient explosion problem and ensure the stability of model training. The introduction of these optimization techniques will further improve the effectiveness and practicality of.

[0057] The early stopping method and the cosine annealing decay learning rate adjustment strategy are used to optimize the model, which has a significant beneficial effect on the UHPC shear performance prediction method in the present invention. First, the introduction of the early stopping method can effectively avoid model overfitting and improve the generalization ability of the model. By terminating the training early, a large amount of computing resources can be saved, and more stable and reliable prediction results can be obtained. Secondly, the cosine annealing decay learning rate adjustment strategy can adjust the model parameters more finely in the later stage of training, further improving the model performance. This adjustment method can enable the model to maintain a certain update ability during the training process, avoid falling into the local optimal solution, and thus obtain more accurate prediction results. In addition, combined with the introduction of other optimization techniques (such as weight decay and gradient clipping), the effect of model optimization can be further improved, and the stability and robustness of the model can be enhanced. The comprehensive application of these optimization techniques will make the UHPC shear performance prediction method in the present invention more efficient, accurate and reliable in practical applications.

[0058] In a preferred embodiment, the present invention can be further configured as follows: before step S10, distributed storage technology should be used to back up the original image and acoustic wave data to multiple storage nodes in real time to ensure the security and integrity of the data during the acquisition and cleaning process. During the data cleaning process, incremental backup is performed for each newly added or modified data to reduce backup overhead and improve backup efficiency. In the case of data loss or damage, the backup data is used to achieve rapid recovery to ensure the continuity and stability of the analysis process. The backup data is regularly verified to ensure the integrity and consistency of the backup data. Before step S10, distributed storage technology is used to back up the original image and acoustic wave data to multiple storage nodes in real time. The implementation of this strategy depends on an efficient distributed storage system that can automatically divide the data into multiple small blocks and store these small blocks in different storage nodes. Such a design not only improves the storage efficiency of the data, but also enhances the fault tolerance of the data, because even if a storage node fails, the data on other nodes can still remain intact. During the data cleaning process, an incremental backup strategy is adopted to reduce backup overhead and improve backup efficiency. Specifically, each time new data is collected or existing data is modified, only the newly added or modified data parts are backed up instead of re-backing up the entire data set. This strategy can significantly reduce the time and storage space required for backup, while ensuring the real-time and consistency of data. In addition, in order to ensure the validity and reliability of the backup data, the backup data is verified regularly. The verification process includes data integrity check and consistency verification to ensure that the backup data is completely consistent with the original data and there is no damage or loss. Once data anomalies are found, the data recovery process will be started immediately, and the backup data will be used for rapid recovery to ensure the continuity and stability of the analysis process. In order to further improve the security and reliability of data, technical means such as encrypted storage and access control can also be considered. Encrypted storage can ensure the security of data during transmission and storage, preventing data from being illegally accessed or tampered with; while access control can limit access rights to data to ensure that only authorized personnel can access and operate data.

[0059] The use of distributed storage technology and incremental backup strategy to back up and verify the original image and acoustic wave data in real time has a significant beneficial effect on the UHPC shear performance prediction method in the present invention. First, the implementation of distributed storage technology improves the storage efficiency and fault tolerance of data, and ensures the security and integrity of data during the acquisition and cleaning process. Even if a storage node fails, the data on other nodes can still remain intact, thereby avoiding the risk of data loss or damage. Secondly, the implementation of the incremental backup strategy significantly reduces the time and storage space required for backup and improves backup efficiency. This strategy can ensure the real-time and consistency of data, while reducing backup overhead, making the data backup process more efficient and economical. In addition, the backup data is verified regularly to ensure the integrity and consistency of the backup data. Once a data anomaly is found, the backup data can be used immediately to achieve rapid recovery to ensure the continuity and stability of the analysis process. This data recovery capability is particularly important in the engineering field, because in practical applications, the loss or damage of data may lead to serious consequences.

[0060] In a preferred embodiment, the present invention can be further configured as follows: the image denoising in the data cleaning step adopts an adaptive median filtering algorithm, automatically adjusts the filter window size according to the local noise situation of the image, removes the noise while retaining the image edge details, and the image enhancement adopts a method combining histogram equalization and contrast stretching. The acoustic signal filtering adopts a bandpass filter, sets a suitable filtering range according to the frequency characteristics of the acoustic signal, removes low-frequency noise and high-frequency interference, and retains useful acoustic information. Regarding image denoising, an adaptive median filtering algorithm is adopted. The core of this algorithm is that it can automatically adjust the size of the filter window according to the local noise situation of the image. Specifically, when the noise in a certain area of ​​the image is more serious, the algorithm will automatically increase the filter window to more effectively remove the noise; and when the edge details in a certain area of ​​the image are rich, the algorithm will reduce the filter window to avoid the loss of edge details. This adaptive adjustment method makes the image denoising effect more superior, and can retain the edge details of the image to the maximum extent while removing the noise. Secondly, in terms of image enhancement, a method combining histogram equalization and contrast stretching is adopted. Histogram equalization can adjust the grayscale distribution of the image, making the grayscale value of the image more uniform, thereby enhancing the contrast of the image. Contrast stretching can further enlarge the difference between different grayscale values ​​in the image, making the details of the image clearer. The combination of these two methods can significantly improve the overall visual effect of the image and provide more valuable information for subsequent analysis and prediction. Finally, in terms of acoustic signal filtering, a bandpass filter is used. The bandpass filter can set a suitable filtering range according to the frequency characteristics of the acoustic signal, thereby removing low-frequency noise and high-frequency interference and retaining useful acoustic information. In practical applications, appropriate filtering parameters will be selected according to the specific characteristics of the acoustic signal to ensure the optimization of the filtering effect. In addition, a multi-stage filtering method can be used to filter the acoustic signal multiple times to further improve the accuracy and effect of the filtering.

[0061] The use of an adaptive median filter algorithm, a method combining histogram equalization and contrast stretching, and a bandpass filter for data cleaning has significant beneficial effects on the UHPC shear performance prediction method in the present invention. First, the application of the adaptive median filter algorithm makes the image denoising effect more superior, and can retain the edge details of the image to the maximum extent while removing the noise, providing more accurate and valuable information for subsequent analysis and prediction. Secondly, the method combining histogram equalization and contrast stretching can significantly improve the overall visual effect of the image, making the details of the image clearer, and helping to more accurately identify and extract key information in the image. This is of great significance for improving the accuracy and reliability of UHPC shear performance prediction. Finally, the application of the bandpass filter can effectively remove low-frequency noise and high-frequency interference in the acoustic wave signal and retain useful acoustic wave information. This has a significant effect on improving the quality and analysis accuracy of the acoustic wave signal. At the same time, the multi-stage filtering method can further improve the accuracy and effect of filtering, making the analysis results of the acoustic wave signal more accurate and reliable.

[0062] In a preferred embodiment, the present invention may be further configured as follows: the analysis method further includes a model evaluation and optimization iteration step, which is executed after step S30 and includes:

[0063] Performance evaluation: Use the test set to comprehensively evaluate the model, including accuracy, recall, F1 score indicators, and the stability and robustness of the model under different conditions;

[0064] Error analysis: Conduct in-depth analysis on samples with model prediction errors to identify the causes of the errors, data noise and feature extraction, and provide direction for subsequent optimization;

[0065] Optimization iteration: Based on the results of performance evaluation and error analysis, the model is optimized, including adjusting the network structure, increasing training data, and improving feature extraction methods. The model is then retrained and verified until the model performance meets the requirements.

[0066] Version management: Perform version management on the model after each optimization, record the model performance changes and optimization strategies, and facilitate subsequent tracking and backtracking. In the performance evaluation phase, in addition to the conventional accuracy, recall and F1 score indicators, the stability and robustness evaluation of the model under different conditions is also introduced. This includes but is not limited to the performance of the model under different data sets, different noise levels, different feature subsets, etc. By comprehensively evaluating the performance of the model under various conditions, we can more accurately understand the generalization ability and adaptability of the model, and provide more valuable reference for subsequent optimization. In the error analysis phase, not only the samples with model prediction errors are deeply analyzed, but also the causes of the errors are tried to be found. This includes data noise, insufficient or redundant feature extraction, unreasonable model structure, etc. Through meticulous error analysis, the bottleneck of model performance improvement can be located, providing a clear direction for subsequent optimization. In the optimization iteration phase, the model is optimized according to the results of performance evaluation and error analysis. This includes adjusting the network structure (such as increasing or decreasing the number of layers, changing the activation function, etc.), increasing training data (such as collecting more samples, performing data enhancement, etc.), and improving feature extraction methods (such as extracting more useful features, removing redundant features, etc.). After each optimization, the model will be retrained and verified to ensure the effectiveness of the optimization strategy. At the same time, the model performance changes and optimization strategies after each optimization will be recorded for subsequent tracking and backtracking. In addition, in the version management link, the version control system is used to manage the model after each optimization. This not only records the evolution history of the model, but also makes it easy to compare the performance differences between different versions, so as to quickly locate the best model. At the same time, version management also facilitates subsequent research and development work, allowing team members to work together and share model resources.

[0067] By comprehensively evaluating the performance and stability of the model under different conditions, we can more accurately understand the generalization ability and adaptability of the model, and provide more valuable reference for subsequent optimization. This helps to quickly locate the bottleneck of model performance improvement, so as to develop a more effective optimization strategy. Secondly, through in-depth error analysis, we can find out the root cause of model prediction errors and provide a clear direction for subsequent optimization. This can not only improve the efficiency of model optimization, but also avoid blind attempts and waste of resources. Thirdly, by continuously optimizing the model structure, increasing training data, and improving feature extraction methods, the performance of the model can be continuously improved. This can not only improve the accuracy of prediction, but also enhance the robustness and stability of the model, so that it can better adapt to actual application scenarios. Finally, by managing the version of each optimized model through the version management system, it is easy to track and trace back the evolution history of the model, so as to quickly locate the best model. This helps to better manage and utilize model resources and improve the efficiency of research and development work.

[0068] In a preferred embodiment, the present invention can be further configured as follows: in the result feedback step, the feedback content includes not only the prediction result itself, but also the confidence of the prediction result, the prediction basis and the model performance evaluation report. The specific operation steps are as follows:

[0069] Visual display: Use charts and images to visualize the prediction results and their confidence levels, making it easier for engineers or construction personnel to understand quickly.

[0070] Interactive feedback: Provides an interactive interface that allows engineers or construction personnel to adjust prediction parameters according to actual needs, such as adjusting confidence thresholds and selecting specific features for prediction, thereby improving the flexibility and accuracy of feedback;

[0071] Report generation: Automatically generate detailed documents including prediction results, confidence, prediction basis and model performance evaluation report for engineers or construction personnel to refer to and make decisions;

[0072] Continuous tracking: Establish a database that associates the prediction results with the actual engineering performance, continuously track the accuracy and reliability of the prediction results, and provide data support for the subsequent optimization of the model. In terms of visualization, not only visualization methods such as charts and images are used to intuitively display the prediction results and their confidence, but also dynamic interactive elements such as dynamic charts and interactive image annotations are introduced to enable engineers or construction personnel to have a deeper understanding of the prediction results and the data distribution and feature relationships behind them. In the interactive feedback link, a highly flexible interactive interface is designed to allow engineers or construction personnel to freely adjust the prediction parameters according to actual needs. This includes but is not limited to adjusting the confidence threshold to filter prediction results with higher or lower confidence, selecting specific features or feature combinations for prediction to focus on specific engineering problems, and adjusting the parameters of the prediction model to optimize the prediction performance. In addition, a real-time feedback mechanism is provided, allowing engineers or construction personnel to see the changes in the prediction results immediately after adjusting the parameters, so as to more accurately evaluate and adjust the prediction parameters. In terms of report generation, a detailed document containing the prediction results, confidence, prediction basis, and model performance evaluation report is automatically generated. This report not only lists the prediction results and confidence levels in detail, but also provides a detailed explanation of the basis for the prediction, including information on key links such as model selection, feature extraction, and data processing. At the same time, the model performance evaluation report section records in detail the key performance indicators of the model, such as accuracy, recall, F1 score, and the stability and robustness evaluation results of the model under different conditions, providing engineers or construction personnel with a comprehensive model performance reference. Finally, in the continuous tracking link, a database associating the prediction results with the actual engineering performance is established to continuously track and record the accuracy and reliability of the prediction results. This database not only records the comparative data between the prediction results and the actual engineering performance, but also contains feedback and improvement suggestions from engineers or construction personnel, providing valuable data support and experience reference for the subsequent optimization of the model.

[0073] Through visual display and interactive feedback mechanism, engineers or construction personnel can understand the prediction results and the data distribution and characteristic relationship behind them more intuitively and deeply, so as to more accurately evaluate the accuracy and reliability of the prediction results. This helps to enhance the confidence and accuracy of engineers or construction personnel in the decision-making process. Secondly, by providing a flexible interactive interface and real-time feedback mechanism, engineers or construction personnel can freely adjust the prediction parameters according to actual needs, so as to more accurately locate and solve specific engineering problems. This not only improves the practicality and pertinence of the prediction results, but also enhances the initiative and creativity of engineers or construction personnel in the problem-solving process. Thirdly, by automatically generating detailed documents containing prediction results, confidence, prediction basis and model performance evaluation reports, engineers or construction personnel can fully and systematically understand the prediction results and model performance, providing strong data support and theoretical basis for decision-making. This helps to improve the scientificity and accuracy of decision-making. Finally, by establishing a database that associates the prediction results with the actual engineering performance, and continuously tracking and recording the accuracy and reliability of the prediction results, it provides valuable data support and experience reference for the subsequent optimization of the model.

[0074] In the data collection and preprocessing stage, the UHPC specimens were selected for shear failure experiments, and ultrasonic detectors and cameras were used to capture the images and acoustic wave data after the failure. Subsequently, the collected image data were grayed for subsequent processing and analysis. Next, the model building and training stage was entered. Based on the deep separable convolutional neural network, a prediction model was built. The model takes the preprocessed image and acoustic wave data as input. In order to train this model, a sample data set containing these input data was constructed, and the model was allowed to learn the features and laws in the data through the training process. In the model verification and optimization stage, the trained model was verified using the test set to evaluate its prediction performance. According to the verification results, the model was optimized to improve its prediction accuracy and generalization ability. Finally, in the model deployment and application stage, the trained and optimized model was deployed on the server. In this way, the model can be used to predict the shear performance of UHPC, and the prediction results can be analyzed, counted and displayed. This step enables this method to be applied in actual engineering and provides engineers and researchers with a powerful tool to evaluate and optimize the shear performance of UHPC.

[0075] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0076] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An analytical method for predicting the shear resistance of UHPC based on neural network, characterized in that: The following steps are involved: S10, data acquisition and preprocessing, select UHPC specimens for shear failure experiments, use ultrasonic detectors and cameras to collect images and sound wave data after damage, and grayscale the image data; S20, model building and training, building a prediction model based on a deep separable convolutional neural network, taking the preprocessed image and sound wave data as input, constructing a sample data set, and performing model training; S30, model validation and optimization, using the test set to validate the model, evaluate the prediction performance, and optimize the model based on the validation results; S40, model deployment and application, deploy the trained model on the server to analyze, count and display the shear performance prediction results.

2. The analysis method for predicting the shear resistance of UHPC based on a neural network according to claim 1 is characterized in that: In step S10, the grayscale processing adopts the grayscale weighted average method, and the formula is: f(t)=0.30f R (t)+0.59f G (t)+0.11f B (t) Wherein, f(t) represents the pixel matrix of the grayscale image f to be identified at the iteration time; f R (t), f G (t), f B (t) represents the iterated color, green, and blue channel pixel matrices respectively.

3. The analysis method for predicting the shear resistance of UHPC based on a neural network according to claim 1 is characterized in that: In step S20, in the deep convolution stage of the deep separable convolutional neural network, the convolution kernel size, step size and padding method of each channel need to be pre-set according to the UHPC image characteristics to capture subtle textures; in the point-by-point convolution stage, a 1x1 convolution kernel is used to fuse information between channels to generate an output feature map with a higher level of abstraction.

4. The analysis method for predicting the shear resistance of UHPC based on a neural network according to claim 3 is characterized in that: In step S20, the binary cross entropy loss function used in model training needs to be combined with a weight decay term to prevent the model from falling into a local optimal solution during the training process. The formula is modified as follows: Among them, w i,j is the category weight, which is used to balance the imbalance between categories; λ is the weight decay coefficient, and θ is the model parameter.

5. The analysis method for predicting the shear resistance of UHPC based on neural network according to claim 1 is characterized in that: In step S30, the cross-validation method adopts K-fold cross-validation, and the K value needs to be reasonably selected according to the size of the data set and the computing resources to ensure the stability and reliability of the validation results. In each cross-validation, the prediction accuracy, recall rate and F1 score indicators of the model need to be recorded to comprehensively evaluate the model performance.

6. The analysis method for predicting the shear resistance of UHPC based on a neural network according to claim 1 is characterized in that: In step S30, the model optimization also includes introducing an early stopping method, which terminates the training early when the validation set loss no longer decreases to avoid model overfitting; at the same time, the learning rate adjustment strategy adopts cosine annealing decay to fine-tune the model parameters in the later stage of training to further improve the model performance.

7. The analysis method for predicting the shear resistance of UHPC based on a neural network according to claim 1, characterized in that: Before step S10, distributed storage technology should be used to back up the original image and acoustic wave data to multiple storage nodes in real time to ensure the security and integrity of the data during the collection and cleaning process. During the data cleaning process, incremental backup should be performed for each new or modified data to reduce backup overhead and improve backup efficiency. In the event of data loss or damage, backup data can be used for rapid recovery to ensure the continuity and stability of the analysis process. Backup data should be verified regularly to ensure the integrity and consistency of the backup data.

8. The method for predicting the shear resistance of UHPC based on neural network according to claim 7, characterized in that: The image denoising in the data cleaning step adopts an adaptive median filtering algorithm, which automatically adjusts the filter window size according to the local noise situation of the image, removes noise while retaining the image edge details. The image enhancement adopts a method combining histogram equalization and contrast stretching. The acoustic signal filtering adopts a bandpass filter, sets a suitable filtering range according to the frequency characteristics of the acoustic signal, removes low-frequency noise and high-frequency interference, and retains useful acoustic information.

9. The method for predicting the shear resistance of UHPC based on a neural network according to claim 1, characterized in that: The analysis method further includes a model evaluation and optimization iteration step, which is performed after step S30 and includes: Performance evaluation: Use the test set to comprehensively evaluate the model, including accuracy, recall, F1 score indicators, and the stability and robustness of the model under different conditions; Error analysis: Conduct in-depth analysis on samples with model prediction errors to identify the causes of the errors, data noise and feature extraction, and provide direction for subsequent optimization; Optimization iteration: Based on the results of performance evaluation and error analysis, the model is optimized, including adjusting the network structure, increasing training data, and improving feature extraction methods. The model is then retrained and verified until the model performance meets the requirements. Version management: Perform version management on each optimized model, record model performance changes and optimization strategies, and facilitate subsequent tracking and backtracking.

10. The method for predicting the shear resistance of UHPC based on neural network according to claim 1, characterized in that: In the result feedback step, the feedback content includes not only the prediction result itself, but also the confidence of the prediction result, the prediction basis and the model performance evaluation report. The specific operation steps are as follows: Visual display: Use charts and images to visualize the prediction results and their confidence levels, making it easier for engineers or construction workers to understand quickly. Interactive feedback: Provides an interactive interface that allows engineers or construction personnel to adjust prediction parameters according to actual needs, such as adjusting confidence thresholds and selecting specific features for prediction, thereby improving the flexibility and accuracy of feedback; Report generation: Automatically generate detailed documents including prediction results, confidence, prediction basis and model performance evaluation report for engineers or construction personnel to refer to and make decisions; Continuous tracking: Establish a database that associates the prediction results with the actual engineering performance, continuously track the accuracy and reliability of the prediction results, and provide data support for subsequent optimization of the model.

Images