AutoML-based railway station building typical disease intelligent identification method

Abstract

The embodiment of the invention provides a railway station building typical disease intelligent identification method based on AutoML, and the method comprises the following steps: collecting video image data of railway station building typical diseases, and carrying out the preprocessing of contrast enhancement and size standardization of the video image data, the typical diseases comprise structural connecting piece defects, metal structure surface defects and decorative plate defects; constructing a training data set based on the preprocessed image data, and performing automatic structure search and hyper-parameter tuning on the image recognition model by using an AutoML automatic machine learning framework to obtain an optimized image recognition model; inputting a to-be-recognized railway station building image into the optimized image recognition model for disease recognition, and outputting a classification recognition result; and matching the classification identification result with historical data in a railway station building disease database, and outputting final disease type information according to a matching result.

Description

Technical Field

[0001] This application relates to artificial intelligence.

[0002] Automated Machine Learning

[0003] Image recognition

[0004] Railway Engineering

[0005] Structural health monitoring

[0006] Pattern recognition

[0007] Intelligent Diagnosis

[0008] Computer vision, specifically involving an intelligent identification method for typical defects in railway station buildings based on AutoML. Background Technology

[0009] This paper proposes an intelligent identification method for typical defects in railway station buildings based on AutoML, aiming to improve the accuracy and efficiency of identifying defects such as cracks in railway station buildings through automated machine learning technology. This method faces three major challenges in practical applications: First, how to dynamically adjust the model's classification threshold based on the color, shape, and texture features of the crack image to reduce the false recognition rate; second, how to automatically optimize the enhancement parameters of the image preprocessing module based on changes in illumination in the image acquisition environment to improve image quality stability; and finally, how to combine the time-series features of historical defect data to regulate candidate operators in the AutoML search space to reduce model redundancy and improve generalization ability. Summary of the Invention

[0010] In view of this, the present disclosure provides a method for intelligent identification of typical defects in railway station buildings based on AutoML, which at least partially solves the problems existing in the prior art.

[0011] A method for intelligent identification of typical defects in railway station buildings based on AutoML includes the following steps:

[0012] Video image data of typical defects in railway station buildings are collected, and the video image data is preprocessed including contrast enhancement and size standardization. The typical defects include defects in structural connectors, defects on the surface of metal structures, and defects in decorative panels.

[0013] A training dataset is constructed based on the preprocessed image data, and the AutoML automated machine learning framework is used to perform automated structure search and hyperparameter tuning on the image recognition model to obtain an optimized image recognition model.

[0014] The image of the railway station building to be identified is input into the optimized image recognition model for defect identification, and the classification and recognition results are output.

[0015] The classification and identification results are matched with historical data in the railway station building defect database, and the final defect type information is output based on the matching results.

[0016] In one specific implementation, the model tuning using the AutoML framework includes:

[0017] The number of image samples for each type of disease is counted, and the sample balance weight is calculated based on the number to generate a training dataset with a balanced class distribution.

[0018] Using one of Bayesian optimization, evolutionary algorithms, or reinforcement learning, the optimal network structure units and hyperparameter combinations of an image recognition model are automatically searched within a predefined search space.

[0019] In one specific implementation, the contrast enhancement in the preprocessing of video image data is achieved using a histogram equalization method, which is accomplished by calculating the cumulative distribution function of the pixel grayscale values ​​of the image and mapping it.

[0020] In one specific implementation, the step of disease identification includes:

[0021] After standardizing the size of the input image to be recognized, a sliding window technique is used to scan its local area;

[0022] The optimized image recognition model is used to classify each window region, and a confidence threshold is set to output only the recognition results with a classification confidence higher than the threshold.

[0023] In one specific implementation, the confidence threshold ranges from 0.8 to 0.95.

[0024] In one specific implementation, the matching process includes:

[0025] Analyze the classification and identification results, and extract the disease category labels and confidence scores contained therein;

[0026] The similarity between the identification result and the features recorded in the disease database is calculated. When the similarity is higher than a preset matching threshold, it is determined that the match is successful.

[0027] In one specific implementation, the matching threshold ranges from 0.7 to 0.9.

[0028] In one specific implementation, a data augmentation step is included before constructing the training dataset, which may be one of the following or a combination thereof:

[0029] Calculate the gradient magnitude of pixels in the image, and assign sample weights to different regions in the image based on the gradient magnitude. Use a weighted loss function in model training to focus on high-weight regions.

[0030] or,

[0031] The image's brightness statistical characteristics are calculated, and nonlinear brightness correction is performed based on the difference between the image and the standard brightness to eliminate the influence of different lighting conditions.

[0032] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1-8.

[0033] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any one of claims 1-8.

[0034] This disclosure provides an intelligent identification method for typical defects in railway station buildings based on AutoML, comprising the following steps: collecting video image data of typical defects in railway station buildings, and performing preprocessing on the video image data including contrast enhancement and size standardization; the typical defects include structural connector defects, metal structure surface defects, and decorative panel defects; constructing a training dataset based on the preprocessed image data, and using the AutoML automated machine learning framework to perform automated structure search and hyperparameter tuning on the image recognition model to obtain an optimized image recognition model; inputting the railway station building image to be identified into the optimized image recognition model for defect identification, and outputting the classification identification result; matching the classification identification result with historical data in the railway station building defect database, and outputting the final defect type information based on the matching result. The solution of this disclosure can solve the problem of high false recognition rate by adjusting the classification threshold of the defect identification model based on the color, shape, and texture features of railway station building crack images.

[0035] 2. How to adjust the enhancement parameters of the image preprocessing module according to the changes in illumination in the image acquisition environment to solve the problem of unstable image quality.

[0036] 3. How to adjust the candidate operators in the AutoML search space based on the time series characteristics of historical disease data to solve the problem of high model redundancy. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the exemplary embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this disclosure and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a method for intelligent identification of typical defects in railway station buildings based on AutoML;

[0039] Figure 2 This is a flowchart of model tuning using the AutoML framework;

[0040] Figure 3 This is a flowchart illustrating the process of enhancing contrast in video image data preprocessing using histogram equalization, which is accomplished by calculating and mapping the cumulative distribution function of the pixel gray values ​​of the image.

[0041] Figure 4 This is a flowchart of the steps involved in disease identification;

[0042] Figure 5 The flowchart shows the confidence threshold range of 0.8 to 0.95.

[0043] Figure 6 This is a flowchart of the matching process;

[0044] Figure 7 This is a flowchart showing that the matching threshold ranges from 0.7 to 0.9;

[0045] Figure 8 Before constructing the training dataset, a data augmentation step is also included, which is a flowchart of one or a combination of the following methods;

[0046] Figure 9 It is a flowchart of the method implemented by the processor when executing the program as described in any one of claims 1-8;

[0047] Figure 10 It is a flowchart of the method implemented as claimed in any one of claims 1-8 when the program is executed by the processor. Detailed Implementation

[0048] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0049] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0050] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0051] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0052] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0053] Next, referring to the accompanying drawings, a method for intelligent identification of typical defects in railway station buildings based on AutoML according to the present invention is described, comprising the following steps: acquiring video image data of typical defects in railway station buildings, and performing preprocessing on the video image data including contrast enhancement and size standardization. The typical defects include structural connector defects, surface defects in metal structures, and defects in decorative panels. Specifically, in the image acquisition stage, high-definition camera equipment is used to acquire structural images inside and outside the railway station building at different time periods to ensure coverage of various typical defect scenarios. For example, in one embodiment, high-precision image data is captured using multi-angle cameras, and the images are scaled to a uniform standard size according to the size ratio of the railway station building, thereby ensuring the consistency of model input. Simultaneously, by adjusting the contrast enhancement algorithm, crack features in the image are made more obvious, avoiding color distortion caused by changes in lighting. Furthermore, image denoising processing is performed to improve clarity, providing high-quality data for subsequent modeling.

[0054] Next, a training dataset is constructed based on the preprocessed image data, and the AutoML automated machine learning framework is used to automatically search the structure and tune the hyperparameters of the image recognition model, resulting in an optimized image recognition model. Specifically, the acquired images are first divided into multiple category labels according to type, and the training set, validation set, and test set are split using cross-validation for subsequent model evaluation. Subsequently, a neural network architecture search mechanism is introduced into the AutoML framework, dynamically adjusting the number of layers, nodes, and connections in the network structure to achieve optimal model performance. For example, in one embodiment, the AutoML algorithm explores a hybrid network structure containing residual connections and multi-scale convolution operations in the candidate structure space, effectively improving the ability to identify different types of cracks. Simultaneously, regarding the classification threshold problem, the confidence threshold in the loss function is adaptively adjusted, enabling the model to dynamically distinguish between defects and non-defects based on color, shape, and texture features, thereby significantly reducing the false recognition rate. For example, when detecting surface cracks in metal structures, the model combines color distribution maps and local texture statistics to dynamically set judgment criteria, reducing the possibility of misjudgment.

[0055] Next, the image of the railway station to be identified is input into the optimized image recognition model for defect identification, and the classification result is output. Specifically, in practical applications, the system can be deployed on a track inspection vehicle to collect video streams of the area to be inspected in real time and process them synchronously. For example, in one embodiment, when the inspection vehicle travels to the area under a railway bridge, the AI ​​terminal quickly processes the real-time transmitted image and completes the classification task through a pre-trained model. During this process, the model adjusts its internal algorithm parameters according to changes in ambient light. For example, it enhances image saturation under strong light conditions to avoid overexposure, while increasing edge detection sensitivity under cloudy conditions to ensure that recognition accuracy is not limited by external conditions. This automatic adaptation mechanism significantly reduces the risk of recognition failure caused by changes in lighting in traditional fixed algorithms, improving the stability and practicality of the system.

[0056] Next, the classification and identification results are matched with historical data in the railway station building defect database, and the final defect type information is output based on the matching results. Specifically, the system extracts time-series features by comparing the current detection results with past defect records to analyze the regularity of defect development. For example, in one embodiment, when a new peeling phenomenon is found on the edge of a decorative panel, the system compares this situation with similar cases in the past three months, finds that the defect shows a gradual expansion trend, and pushes maintenance warning information to relevant personnel accordingly. This method not only enables real-time diagnosis, but also optimizes the training strategy of subsequent defect prediction models through historical data analysis, thereby further reducing model redundancy. At the same time, combined with the evolution trend of defect types, AutoML can introduce weight coefficient adjustment functions in the search space, such as giving higher search priority to certain frequently occurring or highly harmful defects, to reduce the probability of generating irrelevant or inefficient structures, thereby optimizing model performance.

[0057] This technical solution, by integrating AutoML algorithms, multi-dimensional image processing, and historical data analysis, significantly improves the accuracy and adaptability of railway station building defect identification. It not only achieves efficient identification of multiple defect types but also overcomes key challenges such as lighting fluctuations and model complexity, significantly improving detection efficiency and decision support quality, and providing reliable technical support for the safe maintenance of railway facilities.

[0058] Next, the model tuning using the AutoML framework of this invention is described. First, the number of image samples for each type of defect is counted, and sample balancing weights are calculated based on this number to generate a training dataset with a balanced class distribution. This step identifies class imbalance problems in the dataset by analyzing the number of defect samples of different categories, thereby calculating the weight coefficient of each category to improve the model's ability to identify minority classes. For example, in an AutoML-based intelligent identification method for typical defects in railway station buildings, if there are few crack samples, the weight of that class should be higher than other classes, for example, using the formula: Where count l is the number of samples in class l, and b is a scaling factor between 0.5 and 2 used to adjust the weight strength and make the distribution between classes more balanced.

[0059] Next, the optimal network structure units and hyperparameter combinations for the image recognition model are automatically searched within a pre-defined search space. This step combines strategies such as Bayesian optimization, evolutionary algorithms, or reinforcement learning to efficiently explore different network configurations and hyperparameter combinations to find the best model performance. For example, in one embodiment, Bayesian optimization is used to continuously adjust parameters such as the number of layers, activation function type, and learning rate of the neural network based on historical training results, with the objective function being to minimize the loss value on the validation set. This process can significantly improve the model's recognition accuracy while reducing the workload of manual parameter tuning. This automated approach effectively improves the model's generalization ability and enhances its robustness and reliability in real-world applications.

[0060] Next, the contrast enhancement in the preprocessing of video image data according to the present invention is achieved using a histogram equalization method, which is accomplished by calculating and mapping the cumulative distribution function of the pixel gray values ​​of the image. First, the frequency of occurrence of each gray value in the image is calculated, that is, the number of pixels corresponding to each gray value. The frequency value ranges from 0 to N (where N is the total number of pixels in the image), with the aim of quantifying the distribution of each gray level.

[0061] Next, the cumulative distribution function of the grayscale values ​​is calculated, and its expression is: Where i represents the current gray level, k increases from 0 to i, and frequency(k) is the frequency of gray level k, i.e., the number of pixels in the image whose gray value is exactly k. CDF(i) ranges from 0 to 1 and is used to reflect the proportion of all pixels from the lowest gray level to the current gray level. The optimal value set by this formula is to make the gray value distribution as uniform as possible, thereby improving the overall contrast.

[0062] Then, the CDF is mapped proportionally to a new grayscale range, typically [0, 255], using the formula: new pixel=floor(CDF(i)*255+0.5), where floor is the floor function, and CDF(i)*255 linearly stretches the probability value CDF(i) to the range [0,255]. This is the core step of the mapping. floor(CDF(i)*255+0.5) is used for rounding. This step ensures that the changes in brightness in the image are more significant, enhancing image details. In the identification of typical defects in railway station buildings, for example, images of track areas taken on rainy days are generally dark and have low contrast due to reflection. After applying this method, ground cracks can be clearly distinguished from the background, which helps the subsequent detection model to accurately identify them.

[0063] Finally, the mapping results are used to generate processed images, thereby improving the overall clarity and visual information richness of video frames, which is beneficial to improving the recognition accuracy and generalization ability of the AutoML recognition system.

[0064] Next, the steps for defect identification according to the present invention are described. The input image to be identified is standardized in size. This step ensures that all images are processed at the same scale, avoiding model recognition errors caused by scale differences. Parameters include the height and width of the image, typically set to 512×512 pixels, within which detail preservation and computational efficiency can be balanced. For example, in an intelligent identification method for typical defects in railway station buildings based on AutoML, camera images of different resolutions are standardized to a standard size for subsequent processing. A sliding window technique is used to scan local areas. This technique analyzes the image block by block to capture information at different locations by setting window size and step size parameters. The window size is typically 64×64 or 128×128 pixels, and the step size is set to half the window size, covering more possible defect locations. Specifically, in one embodiment, scanning the crack area of ​​the station building roof can accurately locate multiple potential damages. An optimized image recognition model is used to classify each window area. This model is trained based on the AutoML framework and can adaptively adjust feature extraction and classification strategies, significantly improving accuracy. A confidence threshold is set, and only recognition results with a classification confidence level higher than this threshold are output. This aims to filter out false positives and improve the reliability of the results. This threshold is typically set to 0.7; predictions below this value are considered unreliable, thereby reducing invalid feedback. In one embodiment, setting a threshold of 0.7 effectively reduces false positives for minor deformations and improves diagnostic stability. Through the above steps, high-precision and high-efficiency identification of typical defects in railway station buildings is achieved, improving the scientific and intelligent level of maintenance work.

[0065] Next, the confidence threshold of this invention is described as ranging from 0.8 to 0.95. First, during training, an initial confidence threshold is set based on historical data to filter valid detection results. The threshold range is defined as 0.8 to 0.95 to ensure the model has sufficient sensitivity while avoiding excessive false positives. Specifically, in the task of identifying defects in railway station buildings, the model may output different confidence values ​​for different types of defects, such as wall cracks and tile detachment. If a defect is marked as a crack and its corresponding confidence value is higher than 0.8, it is considered a reliable result. Second, during the model deployment phase, the threshold is optimized according to the actual environment to adapt to different lighting conditions, angle changes, etc. The formula can be expressed as: T = a*C + (1-a)*K, where C is the classifier output confidence, K is the spatial matching coefficient, and a is a balancing factor between 0.5 and 0.7, allowing the model to balance accuracy and stability. For example, in an image captured by a camera inside a train station, the system needs to determine whether there are traces of water seepage. If a suspected area is detected with a confidence level of 0.85, the result is considered valid and triggers further analysis. Finally, setting the confidence threshold to 0.8 to 0.95 has the advantage of ensuring that most real defects are identified while reducing false alarms, thus improving the robustness and applicability of the overall identification system.

[0066] Next, the matching process of this invention is described. First, the classification and identification results are parsed to extract the disease category labels and confidence scores. This step aims to obtain key information from the output of the automated machine learning model to provide a basis for subsequent processing. The disease category label represents the specific type detected, while the confidence score reflects the model's degree of certainty regarding the conclusion. For example, in the identification of typical defects in railway station buildings, if the model determines that a crack exists in a certain area, the label might be "concrete crack," and the confidence score might be 0.92.

[0067] Secondly, the similarity between the identified disease characteristics and the features recorded in the disease database is calculated. This process assesses the degree of match by comparing the identified disease features with historical data in the database. Specifically, the cosine similarity formula can be used: A and B are two types of vectors, representing the identification results and disease features in the disease database, respectively. A·B is the dot product of vectors A and B, and |A|*|B| is the product of the moduli of vectors A and B. The value range is [-1,1]. Usually, the optimal value is set to 0.8 or above to ensure reliability.

[0068] Finally, when the similarity score is higher than a preset matching threshold, a successful match is determined. This judgment mechanism helps to filter out highly reliable results and reduce the probability of false alarms. In one embodiment, if the similarity calculation result reaches 0.85, a successful match is considered, and the system can automatically associate corresponding repair suggestions. This process enhances the accuracy and practicality of intelligent recognition.

[0069] Next, the matching threshold of this invention is described as ranging from 0.7 to 0.9. First, a numerical parameter is determined when the system matches the target with candidate objects. This parameter measures the similarity between the two and is typically a floating-point number between 0 and 1, where 0 indicates no match and 1 indicates a perfect match. In this invention, the value range of this parameter is limited to 0.7 to 0.9. Second, based on feedback analysis of historical test data and actual application scenarios, the threshold range that best balances recognition accuracy and false alarm rate is selected, thereby reducing the risk of false recognition while ensuring detection accuracy. Specifically, for an intelligent identification method for typical defects in railway station buildings based on AutoML, the matching threshold is used to evaluate the similarity between the model's detection results for typical defects such as structural cracks and exposed rebar and the standard image. For example, when identifying cracks in the exterior wall of a station building, if the probability value output by the model is 0.85, a successful match is considered. Furthermore, the main purpose of setting this range is to ensure recognition accuracy while avoiding false detections due to excessive sensitivity. The formula for calculating the matching threshold can be expressed as: Among them, S i The similarity score for a single detection region ranges from [0,1], and m represents the number of detection regions. The optimal value is set to 0.8 to achieve the best recognition efficiency and accuracy. This design helps improve the stability and practicality of the railway station health monitoring system.

[0070] Next, the invention further includes a data augmentation step before constructing the training dataset. This step can be one or a combination of the following: calculating the gradient magnitude of pixels in the image and assigning sample weights to different regions in the image based on the gradient magnitude, and using a weighted loss function in model training to focus on high-weight regions; or calculating the brightness statistics of the image and performing nonlinear brightness correction based on its difference from the standard brightness to eliminate the influence of different lighting conditions.

[0071] The purpose of this step is to improve the model's ability to focus on key regions of an image while reducing the impact of environmental changes on model performance. For example, in an AutoML-based intelligent identification method for typical defects in railway station buildings, the uneven contrast in railway images due to weather or lighting changes makes it difficult to effectively detect defective areas.

[0072] Specifically, the gradient magnitude of pixels in the image is calculated using the formula... Among them G x and G y These represent the gradients in the horizontal and vertical directions, respectively, with values ​​ranging from 0 to 255. G is the gradient magnitude, a scalar representing the overall rate of change in grayscale at that point (i.e., edge intensity). A larger value indicates a more pronounced edge at that point. This formula is used to detect edge information in images; high gradient regions typically indicate significant structural changes, such as cracks or peeling areas. Based on this, sample weights are assigned. Among them G max The maximum gradient value is set to 255. This setting gives higher weight to image edge regions, thereby improving the model's learning performance on those regions.

[0073] Another step is to calculate the brightness statistical characteristics of the image, using a nonlinear transformation formula I. corrected =α×I β Where I is the original brightness value, I corrected The corrected brightness values ​​are α∈[0.8,1.2] and β∈[0.5,1.0]. This transformation can adapt to dynamic changes in images under different lighting conditions. For example, if the lighting in an image of a railway station is too dim, adjusting the values ​​of α and β will bring the image brightness closer to the standard value, enhancing the visual contrast of the damaged areas and helping the model accurately identify problems such as cracks or erosion.

[0074] The two enhancement methods described above not only improved image quality but also enhanced the model's ability to identify railway defects in complex environments, effectively reducing the impact of factors such as lighting differences and edge blurring on the model's inference results.

[0075] Next, the method described herein is implemented when the processor executes the program, as claimed in any one of claims 1-8. After receiving railway station image data, the processor performs preliminary preprocessing to improve image quality. For example, the image is grayscaled and standardized using formulas. Where I represents the original image brightness value, ranging from 0 to 255. norm The normalized brightness value, ranging from 0 to 1, I max I is the maximum value of a pixel. min The minimum value of a pixel is used to improve the training effect of subsequent models.

[0076] Next, the processor invokes the AutoML framework to automatically select a suitable deep learning model for disease identification. For example, in one embodiment, a ResNet-50 architecture is selected based on the training set data. This model has high feature extraction capabilities, and its parameters are optimized during training using a cross-entropy loss function, expressed as follows: N is the total number of pixels in the image, W i Let y be the weight of the i-th pixel.i For real labels, p i The probability (between 0 and 1) that the model predicts for this pixel to be of the positive class is 1-y. i It is the true probability that the pixel belongs to the negative class, 1-p i This is the probability that the model predicts the pixel to be a negative class. This formula helps the model converge quickly and classify accurately.

[0077] Subsequently, the processor uses the trained model to infer the new input image and outputs information on the type and location of the defects. For example, in practical implementation, the model can detect typical defects such as concrete cracks and water seepage, and mark the defective areas. This step effectively improves the efficiency and accuracy of railway station building inspections.

[0078] Finally, the processor outputs the identification results to the user interface for relevant personnel to use in decision-making. This technical solution can reduce the workload of manual identification, improve the level of intelligent detection, and achieve rapid and accurate disease diagnosis.

[0079] Next, the method described herein, when executed by a processor, implements any one of claims 1-8. First, image data of the railway station structure is acquired and input into an AutoML framework. The image data includes high-resolution aerial photographs and structural inspection images for subsequent processing. For example, in one embodiment, RGB images with a resolution higher than 4K are used to ensure a detail sharpness of at least 0.1mm.

[0080] Secondly, the AutoML algorithm is used to automatically optimize the model structure and hyperparameter configuration. Specifically, the optimal model combination is found within a preset range through random search or Bayesian optimization methods. For example, the hyperparameter range is set to a learning rate of 0.0001 to 0.1 and a kernel size of 3×3 to 5×5, and models with an accuracy of over 85% are selected as initial candidates. This step can significantly improve model training efficiency and generalization ability.

[0081] Third, the input images are preprocessed and enhanced to extract disease features. For example, histogram equalization is used to improve contrast, and transformations such as rotation and scaling are used to enhance data diversity and ensure the robustness of the model's recognition. This step helps reduce the risk of misidentification caused by environmental interference.

[0082] Finally, the trained model is deployed on on-site computing equipment to automatically identify and classify newly acquired images. For example, combined with the deep learning object detection model YOLOv5, it can identify multiple images per second with an accuracy rate of no less than 90%, thereby improving detection efficiency. This technical solution effectively improves the automation and accuracy of railway station building defect identification, reducing reliance on manual labor and maintenance costs.

[0083] This invention discloses an intelligent identification method for typical defects in railway station buildings based on AutoML. The method includes: First, acquiring video image data of typical defects in railway station buildings and performing contrast enhancement and size standardization on the images to improve image quality and provide a unified data foundation for subsequent analysis. Next, constructing a training dataset based on the preprocessed image data and using the AutoML automated machine learning framework for model structure search and hyperparameter tuning. The optimized image recognition model achieves high-accuracy defect classification. In practical applications, after inputting the image to be identified into the model, the system outputs preliminary classification results and matches them with historical data in the railway station building defect database to further generate final defect type information.

[0084] To address the first technical problem of how to adjust the classification threshold of the defect identification model based on the color, shape, and texture features of railway station building crack images to solve the problem of high false recognition rate, this invention introduces a feature-weighted dynamic threshold adjustment mechanism. Based on the differences in color, shape, and texture of different types of defects, the probability distribution of the model output is adaptively corrected, thereby effectively reducing the risk of false recognition caused by feature similarity.

[0085] Regarding the second technical problem of how to adjust the enhancement parameters of the image preprocessing module according to the changes in illumination in the image acquisition environment to solve the problem of unstable image quality, this invention designs a dynamic parameter optimization algorithm based on illumination detection. In the preprocessing stage, the algorithm senses the changes in ambient light in real time and then dynamically adjusts enhancement parameters such as contrast, brightness, and sharpening to ensure the quality and consistency of the image under different illumination conditions.

[0086] To address the third technical problem of how to regulate candidate operators in the AutoML search space based on the time-series characteristics of historical disease data to solve the problem of high model redundancy, this invention analyzes the temporal evolution pattern of disease data, extracts key features as prior information, and guides the AutoML model to prioritize network structures and functional modules with time correlation during structure search, thereby improving the model's relevance and efficiency, reducing unnecessary computational burden, and enhancing the model's generalization ability and performance.

[0087] The methods, programs, systems, apparatuses, etc., in the embodiments of the present invention can be executed or implemented in one or more networked computers, or practiced in a distributed computing environment. In the embodiments of this specification, in these distributed computing environments, tasks can be performed by remote processing devices connected via a communication network.

[0088] Those skilled in the art will understand that the embodiments described in this specification can be provided as methods, systems, or computer program products. Therefore, those skilled in the art will realize that the functional modules / units or controllers and related method steps described in the above embodiments can be implemented in software, hardware, or a combination of both.

[0089] Unless explicitly stated otherwise, the actions or steps of the methods and procedures described in the embodiments of the present invention do not necessarily have to be performed in a specific order and can still achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0090] This document describes several embodiments of the present invention; however, for the sake of brevity, the descriptions of the embodiments are not exhaustive, and identical or similar features or parts between the embodiments may be omitted. In this document, "one embodiment," "some embodiments," "example," "specific example," or "some examples" refers to embodiments applicable to at least one, but not all, of the present invention. The above terms do not necessarily refer to the same embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of the different embodiments or examples.

[0091] The exemplary systems and methods of the present invention have been specifically shown and described with reference to the above embodiments, which are merely examples of the best mode for implementing the systems and methods. Those skilled in the art will understand that various changes can be made to the embodiments of the systems and methods described herein without departing from the spirit and scope of the invention as defined in the appended claims when implementing the systems and / or methods.

Claims

1. A method for intelligent identification of typical defects in railway station buildings based on AutoML, characterized in that, Includes the following steps: Video image data of typical defects in railway station buildings are collected, and the video image data is preprocessed including contrast enhancement and size standardization. The typical defects include defects in structural connectors, defects on the surface of metal structures, and defects in decorative panels. A training dataset is constructed based on the preprocessed image data, and the AutoML automated machine learning framework is used to perform automated structure search and hyperparameter tuning on the image recognition model to obtain an optimized image recognition model. The image of the railway station building to be identified is input into the optimized image recognition model for defect identification, and the classification and recognition results are output. The classification and identification results are matched with historical data in the railway station building defect database, and the final defect type information is output based on the matching results.

2. The method according to claim 1, characterized in that, The model tuning using the AutoML framework includes: The number of image samples for each type of disease is counted, and the sample balance weight is calculated based on the number to generate a training dataset with a balanced class distribution. Using one of Bayesian optimization, evolutionary algorithms, or reinforcement learning, the optimal network structure units and hyperparameter combinations of an image recognition model are automatically searched within a predefined search space.

3. The method according to claim 1, characterized in that, The contrast enhancement in the preprocessing of video image data is achieved using the histogram equalization method, which is accomplished by calculating the cumulative distribution function of the pixel gray values ​​of the image and mapping it.

4. The method according to claim 1, characterized in that, The steps for disease identification include: After standardizing the size of the input image to be recognized, a sliding window technique is used to scan its local area; The optimized image recognition model is used to classify each window region, and a confidence threshold is set to output only the recognition results with a classification confidence higher than the threshold.

5. The method according to claim 4, characterized in that, The confidence threshold ranges from 0.8 to 0.

95.

6. The method according to claim 1, characterized in that, The matching process includes: Analyze the classification and identification results, and extract the disease category labels and confidence scores contained therein; The similarity between the identification result and the features recorded in the disease database is calculated. When the similarity is higher than a preset matching threshold, it is determined that the match is successful.

7. The method according to claim 6, characterized in that, The matching threshold ranges from 0.7 to 0.

9.

8. The method according to claim 1, characterized in that, Before building the training dataset, a data augmentation step is also included, which can be one of the following or a combination thereof: Calculate the gradient magnitude of pixels in the image, and assign sample weights to different regions in the image based on the gradient magnitude. Use a weighted loss function in model training to focus on high-weight regions. Alternatively, the brightness statistics of the image can be calculated, and nonlinear brightness correction can be performed based on the difference between the image and the standard brightness to eliminate the influence of different lighting conditions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

Images