Visual intelligent troubleshooting method for hidden danger of internal structure of dam based on machine learning

Abstract

The invention relates to a visual intelligent troubleshooting method for hidden dangers of an internal structure of a dam based on machine learning, and belongs to the technical field of image data processing, and the method comprises the following steps: data collection: employing an array sensor to obtain original monitoring data; data preprocessing: performing noise reduction and normalization processing on the original monitoring data, and dividing the processed data into a training set and a test set; model training: constructing a deep learning neural network model until the deep learning neural network model converges, and obtaining a trained hidden danger prediction model; and hidden danger prediction: inputting the test set into the trained hidden danger prediction model, and outputting a hidden danger prediction result of the internal structure of the dam. Visual display: performing visual display according to the hidden danger prediction result; the method has the beneficial effects that the linear analysis defects of traditional threshold judgment and empirical formulas are overcome, and the hidden danger recognition precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image data processing, and in particular relates to a method for visual intelligent inspection of hidden dangers in the internal structure of a dam based on machine learning. Background Art

[0002] As a key infrastructure of water conservancy projects, the safe operation of dams is crucial to flood control, power generation and irrigation. As the service life of dams increases, affected by geological conditions, water erosion and temperature stress factors, the internal structure is prone to cracks, abnormal seepage and stress concentration. If not discovered and handled in time, it may cause catastrophic accidents such as dam failure.

[0003] At present, the inspection of hidden dangers in the internal structure of the dam mainly relies on a combination of manual inspection and traditional sensor monitoring. Manual inspection is inefficient, highly subjective, and difficult to penetrate into the interior of the dam, making it difficult to discover hidden hidden dangers; although traditional sensor monitoring can obtain some physical parameters, data analysis mostly uses threshold judgment and empirical formula methods, which cannot effectively handle complex nonlinear data relationships and is difficult to accurately identify early minor hidden dangers. In addition, traditional monitoring data are mostly presented in the form of tables and curves, lacking intuitive visual expression, which is not conducive to technical personnel quickly grasping the overall safety status of the dam.

[0004] With the development of artificial intelligence technology, machine learning has shown strong advantages in data processing and pattern recognition. Applying machine learning technology to the inspection of hidden dangers in the internal structure of dams can mine the potential laws in massive monitoring data, realize intelligent prediction of hidden dangers, and combine with visualization technology to intuitively present information on the location and severity of hidden dangers. Summary of the invention

[0005] The present invention provides a method for visual intelligent inspection of hidden dangers in the internal structure of a dam based on machine learning, which is used to solve the technical problem of how to realize the visualization of hidden dangers in the internal structure of a dam. By modeling complex nonlinear relationships through deep learning neural networks, the linear analysis defects of traditional threshold judgments and empirical formulas are overcome, and the accuracy of hidden danger identification is significantly improved.

[0006] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0007] The machine learning-based intelligent visualization method for detecting hidden dangers in the internal structure of a dam includes the following steps:

[0008] Data acquisition: Array sensors are used to collect the physical parameters of stress, strain, seepage and temperature inside the dam in real time to obtain original monitoring data;

[0009] Data preprocessing: perform noise reduction and normalization on the original monitoring data, and divide the processed data into training set and test set;

[0010] Model training: Construct a deep learning neural network model, input the training set into the deep learning neural network model for training, and continuously adjust the parameters of the deep learning neural network model through the backpropagation algorithm until the deep learning neural network model converges, obtaining a trained hidden danger prediction model;

[0011] Hidden danger prediction: Input the test set into the trained hidden danger prediction model, and output the hidden danger prediction results of the dam's internal structure;

[0012] Visualization display: According to the hidden danger prediction results, use 3D modeling technology to construct a 3D model of the dam, and overlay the hidden danger information on the 3D model of the dam in different colors, shapes, and transparencies for visualization display.

[0013] Optionally, the array sensors include fiber optic sensors, strain gauge sensors, and piezometers, which are respectively used to monitor the temperature, strain, and seepage data inside the dam.

[0014] Optionally, a wireless sensor network is used to achieve real-time transmission of the original monitoring data, and the transmitted data is encrypted.

[0015] Optionally, in data preprocessing, wavelet transform is used to denoise the original monitoring data.

[0016] Optionally, the deep learning neural network model is a convolutional neural network or a recurrent neural network.

[0017] Optionally, it also includes deep learning neural network model update: Regularly collect new monitoring data, merge the new monitoring data with historical data, re-divide the training set and test set, and retrain and optimize the hidden danger prediction model.

[0018] Optionally, in model training, transfer learning technology is used to use the model parameters pre-trained on other similar dam monitoring data as initialization parameters to accelerate the model training speed.

[0019] Optionally, in hidden danger prediction, it also includes early warning: According to the hidden danger prediction results, when serious hidden dangers are predicted, trigger the alarm device to send out early warning information.

[0020] Optionally, in the visualization display, it supports immersive viewing of the 3D model of the dam and hidden danger information through virtual reality devices, and provides a dynamic playback function for hidden danger information to display the development and change of hidden dangers in time series.

[0021] Optionally, in the visualization display, it also includes setting hidden danger level identifiers for different areas of the dam, and obtaining detailed hidden danger information of the corresponding area by clicking on the identifier.

[0022] Advantages of the present invention:

[0023] Through noise reduction and normalization operations, the data preprocessing of the present invention eliminates environmental noise interference and unifies the data scale, ensuring the effectiveness of subsequent model training. The model training uses a deep learning neural network to model complex non-linear relationships, overcomes the linear analysis defects of traditional threshold judgment and empirical formulas, and significantly improves the accuracy of hidden danger identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0025] Figure 1 It is a schematic diagram of the working process of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] The following will describe the embodiments of the present application in detail with reference to the drawings.

[0027] Embodiment 1;

[0028] As Figure 1 shown, this embodiment provides a visual intelligent inspection method for hidden dangers in the internal structure of a dam based on machine learning, including the following steps:

[0029] Data acquisition: Using an array of sensors, the physical parameters of stress, strain, seepage, and temperature inside the dam are collected in real time to obtain the original monitoring data;

[0030] Data preprocessing: Perform noise reduction and normalization on the original monitoring data, and divide the processed data into a training set and a test set;

[0031] Model training: Construct a deep learning neural network model, input the training set into the deep learning neural network model for training, and continuously adjust the parameters of the deep learning neural network model through the backpropagation algorithm until the deep learning neural network model converges to obtain a trained hidden danger prediction model;

[0032] Hidden danger prediction: Input the test set into the trained hidden danger prediction model to output the hidden danger prediction results of the internal structure of the dam;

[0033] Visualization display: According to the hidden danger prediction results, use 3D modeling technology to construct a 3D model of the dam, and overlay the hidden danger information on the 3D model of the dam in different colors, shapes, and transparencies for visualization display.

[0034] Through the combination of deep learning of the machine and three-dimensional visualization technology, the intelligent detection and intuitive expression of dam hidden dangers are realized.

[0035] Specifically, in data acquisition, array sensors are used to monitor multi-dimensional physical parameters in real time, breaking through the limitation of insufficient data coverage of traditional single sensors and providing a comprehensive data basis for hidden danger identification; in the data preprocessing step, through noise reduction and normalization operations, environmental noise interference is eliminated and the data scale is unified to ensure the effectiveness of subsequent model training;

[0036] In model training, a deep learning neural network is used to model complex non-linear relationships, overcoming the linear analysis defects of traditional threshold judgment and empirical formulas, and significantly improving the accuracy of hidden danger identification; in the hidden danger prediction step, the trained model is used to infer the test data to realize the automatic positioning and grading of hidden dangers; in the visualization display step, the abstract prediction results are mapped into the visualization elements of a three-dimensional model, and the spatial distribution and severity of hidden dangers are intuitively presented through differences in color, shape and transparency, solving the problem of unintuitive information expression of traditional tables and curves.

[0037] The array sensors include fiber optic sensors, strain gauge sensors and piezometers, which are used to monitor the temperature, strain and seepage data inside the dam respectively.

[0038] Accurate acquisition of multi-dimensional physical parameters is achieved through sensor combination. Fiber optic sensors have high sensitivity and anti-electromagnetic interference characteristics, and can accurately capture the temperature field distribution inside the dam, providing data support for temperature stress analysis; strain gauge sensors are directly attached to the surface of the structure to measure the deformation of the material in real time, reflecting the change trend of the stress state of the dam structure; piezometers dynamically monitor seepage parameters and can identify abnormal seepage paths or sudden changes in pore water pressure. The combined use of the three sensors not only covers the physical quantities for the safety assessment of the dam structure, but also makes up for the monitoring blind areas of single sensors through different sensing principles, forming multi-source data complementarity, providing complete and heterogeneous input features for subsequent machine learning models, thereby enhancing the reliability of hidden danger prediction.

[0039] The use of a wireless sensor network to realize the real-time transmission of original monitoring data and encrypt the transmitted data. By using a wireless sensor network to replace the traditional wired transmission method, the problems of low transmission efficiency and poor real-time performance caused by complex wiring during the data acquisition process of sensors inside the dam are solved. The wireless network can achieve multi-point synchronous transmission and reduce data latency. At the same time, encrypting the transmitted data can prevent the monitoring data from being illegally intercepted or tampered with during transmission, ensuring data integrity and security, and avoiding misjudgment caused by data leakage or errors. The combination of these two technical means further strengthens the anti-interference ability and security protection level of the monitoring system on the basis of improving data transmission efficiency, providing a reliable data basis for subsequent hidden danger prediction models.

[0040] Example 2;

[0041] Based on Example 1, in data preprocessing, wavelet transform is used to denoise the original monitoring data. To address the problem of noise interference in dam monitoring data, data optimization is achieved by introducing the signal processing method of wavelet transform.

[0042] The wavelet transform for denoising the original monitoring data is as follows:

[0043] Signal decomposition (multi-scale decomposition);

[0044] Decompose the original signal into layers of low-frequency approximation components and high-frequency detail components of each layer. Taking discrete wavelet transform as an example, the decomposition formula for the th layer is:

[0045] ;

[0046] Among them, is the approximation component of the th layer (low-frequency, representing the main body of the signal); is the detail component of the th layer (high-frequency, possibly containing noise);

[0047] Threshold processing (key step for denoising);

[0048] Assume that the noise mainly exists in the high-frequency detail components. The detail coefficients are processed through a threshold function to retain the effective signal components and suppress the noise components.

[0049] Threshold calculation: Universal threshold (VisuShrink): Applicable to white noise scenarios, the threshold , where is the noise standard deviation, is the signal length. Coefficients greater than the threshold are shrunk (subtracting the threshold) to avoid discontinuity points that may cause signal distortion.

[0050] Signal reconstruction;

[0051] Reconstruct the denoised signal through the inverse wavelet transform of the processed approximation coefficients and detail coefficients :

[0052] ;

[0053] Among them, is the inverse discrete wavelet transform. The discrete wavelet transform is a process of decomposing the original signal into different frequency components, while the inverse discrete wavelet transform is a process of recombining the processed (such as: threshold processing) frequency components to restore the signal. Among them, Indicates the -th approximation coefficient obtained after the original signal is decomposed by layers of discrete wavelet transform. The approximation coefficient reflects the low-frequency part of the signal, that is, the main trend and general characteristics of the signal.

[0054] Indicates the detail coefficient of the -th layer after the original signal is decomposed by discrete wavelet transform. The detail coefficient reflects the high-frequency part of the signal, and usually most of the noise is concentrated in the high-frequency region.

[0055] Through the inverse discrete wavelet transform (IDWT), the processed -th layer approximation coefficient and the processed detail coefficients from the 1st layer to the -th layer are combined to reconstruct the denoised or otherwise processed signal , which is the last step in the process of using wavelet transform for signal processing (such as denoising, compression), and synthesizes the information of each frequency component decomposed and processed previously into the finally desired signal. , which is the last step in the process of using wavelet transform for signal processing (such as denoising, compression), and synthesizes the information of each frequency component decomposed and processed previously into the finally desired signal.

[0056] In practical applications, the Mallat algorithm (fast wavelet transform) is usually used to implement the above decomposition and reconstruction processes.

[0057] Specifically, wavelet transform has the ability of multi-resolution analysis and can adaptively match the characteristics of monitoring signals with different frequencies. Compared with traditional filtering methods, its advantages are as follows: it has localization characteristics in both the time domain and the frequency domain, and can accurately identify the hidden danger characteristics corresponding to the signal mutation points; by selecting appropriate wavelet basis functions, it can effectively separate the real signal from the random noise in the physical parameters of the dam; it is particularly suitable for processing non-stationary stress and strain monitoring data, removing high-frequency interference while retaining the detail characteristics of the signal. This processing method provides a higher-quality training data basis for subsequent deep learning models, enables the retention of tiny hidden danger characteristics, and thus improves the accuracy of hidden danger prediction.

[0058] Embodiment 3;

[0059] Based on Embodiment 1, the deep learning neural network model is a convolutional neural network or a recurrent neural network. By defining the specific type of the deep learning neural network model as a convolutional neural network or a recurrent neural network, it provides adaptability for the data characteristics of the hidden danger prediction scenario of the dam internal structure.

[0060] For the obtained trained hidden danger prediction model:

[0061] ;

[0062] Among them, is the probability of potential risk (the value range is usually mapped to [0, 1] through the sigmoid function and is applicable to logistic regression); are feature variables (such as: equipment temperature, operation duration, and maintenance times); is the intercept term (basic risk value); are feature weight coefficients (determined through training and reflecting the influence degree of each factor on potential risks); is the error term (obeying the normal distribution and representing random factors not modeled).

[0063] The linear regression model intuitively reflects the influence of each factor on potential risks through coefficients and is suitable for preliminary risk assessment and interpretive analysis. In practical applications, features need to be selected in combination with the business scenario, and the accuracy and stability of the model need to be verified.

[0064] Through risk assessment, according to the calculation results (such as: 11.82%), a threshold is set to judge the risk level: if the threshold is 10%, then this equipment is "high risk" and immediate maintenance is required.

[0065] Convolutional neural networks can effectively extract spatial features from the monitoring data of sensor arrays. For example, the spatial correlation of stress and temperature distributions collected by fiber optic sensors and strain gauge sensors at different positions of the dam body; recurrent neural networks are suitable for processing monitoring data with time-series evolution characteristics such as seepage and temperature, and capturing the dynamic change laws in the development process of potential risks. The selection of these two network structures targets the spatial and time dimension features hidden in the dam monitoring data respectively, strengthening the model's fitting ability for complex non-linear relationships, thereby improving the accuracy of potential risk prediction.

[0066] Deep learning neural network model update: Regularly collect new monitoring data. After merging the new monitoring data with historical data, re-divide the training set and test set, and retrain and optimize the potential risk prediction model. Facilitating the regular collection of new monitoring data can avoid the problem of data distribution deviation caused by the change of the dam's internal structure over time and ensure the timeliness of the model input data; after merging new data with historical data and re-dividing the training set and test set, both the regular features of historical data are retained and new features under the current working conditions are integrated, enhancing the completeness of the data set; retraining and optimizing the potential risk prediction model can correct the deviation between the original model parameters and the new data, and through iterative update, the model continuously adapts to the change of the dam structure state, thus solving the problem of prediction failure caused by data aging of the static model.

[0067] Example 4;

[0068] Based on Embodiment 1, in model training, transfer learning technology is adopted. The model parameters pre-trained on other similar dam monitoring data are used as the initial parameters to accelerate the model training speed. Through transfer learning technology, the prior knowledge obtained in the existing dam monitoring scenarios is transferred to the current model training. Specifically, the pre-trained model parameters of other similar dams are used as the initial parameters, fully utilizing the data features and pattern recognition capabilities learned in similar engineering scenarios, and avoiding the inefficiency caused by the random initialization of parameters when the neural network model is trained from scratch. This initialization method based on pre-trained parameters can significantly shorten the convergence time of the model on the target dataset. Especially when the sample size of the current dam monitoring data is limited, parameter transfer can effectively improve the stability and efficiency of model training while reducing the consumption demand for computing resources.

[0069] Embodiment 5;

[0070] Hidden danger prediction also includes early warning: According to the hidden danger prediction results, when serious hidden dangers are predicted, the alarm device is triggered to send out early warning information. By establishing a linkage mechanism between hidden danger prediction and early warning, a closed-loop system from data analysis to risk disposal is constructed. In the specific technical features, the judgment basis for early warning according to the hidden danger prediction results realizes the quantitative application of the output of the machine learning model, enabling the early warning to have objective data support; the trigger condition setting when serious hidden dangers are predicted differentiates the hidden danger levels by setting thresholds to ensure special responses to high-risk situations; the execution method of triggering the alarm device to send out early warning information converts digital signals into physical alarm actions, solving the problem of response delay in traditional manual judgment and forming an automated risk disposal process.

[0071] Embodiment 6;

[0072] In the visual display, it supports immersive viewing of the dam three-dimensional model and hidden danger information through virtual reality devices, and provides a dynamic playback function for hidden danger information, showing the development and changes of hidden dangers in time series.

[0073] Enhance the spatio-temporal dimension expressiveness of hidden danger analysis through dual innovation. Through the immersive viewing function of virtual reality devices, break through the limitations of traditional two-dimensional plane displays, enabling technicians to observe the spatial distribution characteristics of hidden dangers from a three-dimensional spatial perspective in multiple dimensions. In particular, achieve three-dimensional perception of complex forms such as crack directions and seepage paths. The introduction of the dynamic playback function establishes a time dimension analysis framework, converting discrete hidden danger detection data into continuous time-series animations, intuitively presenting the evolution process of stress concentration areas and the dynamic characteristics of the expansion trend of seepage paths. The time series display realizes the smooth transition of the hidden danger development process through data frame interpolation technology, facilitating technicians to trace the mechanism of hidden danger generation and predict the trend of structural deterioration. The synergistic effect of these two technical features not only solves the problem of insufficient presentation of spatial dimension information but also makes up for the shortcoming of the lack of time dimension analysis, forming a four-dimensional spatio-temporal analysis ability.

[0074] In the visual display, it also includes setting hidden danger level identifiers for different areas of the dam and obtaining detailed hidden danger information for the corresponding areas by clicking on the identifiers. It is through the collaborative mechanism of hierarchical identification and interactive query that the efficiency of hidden danger visual analysis is improved. Specifically, by setting hidden danger level identifiers in different spatial areas of the three-dimensional dam model and using differentiated graphic symbols to achieve an intuitive hierarchical display of the severity of hidden dangers, it solves the problem that traditional visualization only shows the location of hidden dangers and cannot quickly judge the priority level for handling. Further, by clicking on the identifier to trigger the data retrieval function and establishing the associated mapping between the spatial position and the background database, technicians can directly obtain detailed information such as the stress change curve and seepage history data of a specific area, overcoming the cumbersome operation of manually retrieving multi-source data in the traditional display method. The combination of these two technical features not only retains the spatial perception advantage of the three-dimensional model but also forms a complete closed loop from macroscopic situation perception to microscopic data analysis through the visual guidance of hierarchical identification and the interactive design of instant data retrieval.

[0075] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope recorded in the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. A method for visually and intelligently detecting potential hazards in the internal structure of a dam based on machine learning, characterized in that, It includes the following steps: Data acquisition: Using an array of sensors, the physical parameters of stress, strain, seepage, and temperature inside the dam are collected in real time to obtain the original monitoring data; Data preprocessing: The original monitoring data is denoised and normalized, and the processed data is divided into a training set and a test set; Model training: A deep learning neural network model is constructed, and the training set is input into the deep learning neural network model for training. The parameters of the deep learning neural network model are continuously adjusted through the backpropagation algorithm until the deep learning neural network model converges, obtaining a trained hidden danger prediction model; Hidden danger prediction: The test set is input into the trained hidden danger prediction model, and the hidden danger prediction results of the internal structure of the dam are output; Visualization display: According to the hidden danger prediction results, a three-dimensional model of the dam is constructed using three-dimensional modeling technology, and the hidden danger information is superimposed on the three-dimensional model of the dam in different colors, shapes, and transparencies for visualization display.

2. The method for visual intelligent investigation of potential hazards in the internal structure of a dam based on machine learning according to claim 1, characterized in that, The array of sensors includes fiber optic sensors, strain gauge sensors, and piezometers, which are used to monitor the temperature, strain, and seepage data inside the dam respectively.

3. The method for visual intelligent investigation of potential hazards in the internal structure of a dam based on machine learning according to claim 2, characterized in that, A wireless sensor network is used to realize the real-time transmission of the original monitoring data, and the transmitted data is encrypted.

4. The method for visual intelligent investigation of potential hazards in the internal structure of a dam based on machine learning according to claim 1, characterized in that In the data preprocessing, wavelet transform is used to denoise the original monitoring data.

5. The method for visual intelligent investigation of potential hazards in the internal structure of a dam based on machine learning according to claim 1, wherein The deep learning neural network model is a convolutional neural network or a recurrent neural network.

6. The visualization intelligent inspection method for hidden dangers of the internal structure of a dam based on machine learning according to claim 5, characterized in that, It also includes the update of the deep learning neural network model: New monitoring data is collected regularly. After merging the new monitoring data with the historical data, the training set and the test set are re-divided, and the hidden danger prediction model is retrained and optimized.

7. The method for visual intelligent inspection of hidden dangers in the internal structure of a dam based on machine learning according to claim 1, characterized in that In the model training, transfer learning technology is adopted, and the model parameters pre-trained on other similar dam monitoring data are used as the initial parameters to accelerate the model training speed.

8. The method for visual intelligent investigation of hidden dangers in the internal structure of a dam based on machine learning according to claim 1, wherein In the hidden danger prediction, it also includes early warning: According to the hidden danger prediction results, when serious hidden dangers are predicted, the alarm device is triggered to send out early warning information.

9. The method for visual intelligent detection of hidden dangers in the internal structure of a dam based on machine learning according to claim 1, wherein In the visualization display, it supports immersive viewing of the three-dimensional model of the dam and the hidden danger information through virtual reality devices, and provides a dynamic playback function for the hidden danger information to display the development and changes of the hidden danger in time series.

10. The method for visual intelligent investigation of potential hazards in the internal structure of a dam based on machine learning according to claim 9, characterized in that, In the visualization display, it also includes setting hidden danger level marks for different areas of the dam, and obtaining detailed hidden danger information of the corresponding area by clicking on the marks.

Images