Urban road disease prediction and preventive maintenance method based on big data analysis

Abstract

The invention relates to an urban road disease prediction and preventive maintenance method based on big data analysis. The urban road disease prediction method based on big data analysis comprises the steps of obtaining road crowdsourcing data; the data types of the road crowdsourcing data comprise numerical data, image data and text data; processing the road crowdsourcing data based on a big data processing analysis mechanism to obtain a multi-modal database; generating a road disease predictive score by using the multi-modal database according to a preset analysis time unit based on the mixed attention neural network model; the road disease predictive score comprises health indexes of a plurality of road segments. By adopting the method, the disease development trend can be effectively identified; and data-driven support is provided for preventive maintenance of urban roads, the maintenance cost is reduced, and the service life of the roads is prolonged.

Description

Technical Field

[0001] The present invention belongs to the field of data processing, and particularly relates to a method for predicting urban road diseases and preventive maintenance using big data analysis. Background Art

[0002] With the development of big data technology, there has emerged a technology for prediction using data-driven methods. By processing a large amount of complex data and mining the potential laws of the data, accurate prediction and intelligent decision-making can be achieved.

[0003] In traditional technologies, the detection of urban road diseases mainly relies on manual inspections. Maintenance personnel regularly patrol the roads, and rely on visual observation and simple tool measurements to detect road diseases such as cracks and potholes. For the prediction of diseases, it is usually based on the service life of the road, past experience, and some simple statistical data to roughly judge the time and location where diseases may occur. When formulating a maintenance plan, it is also mainly based on manual judgment and fixed cycle arrangements. For example, the road is comprehensively inspected and repaired at regular intervals.

[0004] However, in the above methods, the manual inspection method is inefficient, requires a large amount of manpower and time, and it is difficult to achieve real-time and comprehensive detection, and it is easy to miss some hidden diseases. The disease prediction based on experience and simple statistical data is not accurate enough, and it is impossible to detect potential disease risks in time, resulting in diseases may be discovered only when they develop to a more serious level, increasing the maintenance cost and the impact on traffic. When formulating a maintenance plan, the fixed cycle arrangement may not be able to be flexibly adjusted according to the actual condition of the road, and there may be over-maintenance or under-maintenance, resulting in waste of resources or deterioration of the road condition. Summary of the Invention

[0005] Based on this, in order to solve the above technical problems, it is necessary to provide a method for predicting urban road diseases using big data analysis that can perform disease prediction based on big data analysis.

[0006] In a first aspect, the present application provides a method for predicting urban road diseases using big data analysis, including:

[0007] Obtain road crowdsourcing data; the data types of road crowdsourcing data include numerical data, image data, and text data;

[0008] Process the road crowdsourcing data based on a big data processing and analysis mechanism to obtain a multi-modal database;

[0009] Based on a hybrid attention neural network model, use the multi-modal database to generate a road disease predictive score according to a preset analysis time unit; the road disease predictive score includes the health index of multiple road segments.

[0010] In one embodiment, the road crowdsourcing data is processed based on a big data processing and analysis mechanism to obtain a multi-modal database, including:

[0011] Pre-clean and standardize the road crowdsourcing data to obtain standardized road crowdsourcing data;

[0012] Preprocess the standardized road crowdsourcing data according to the corresponding data type of the standardized road crowdsourcing data to obtain multiple sub-modal data;

[0013] Perform spatial mapping on multiple sub-modal data according to the location information to obtain a road segment grid data set;

[0014] Align the multiple sub-modal data of the same road segment in the road segment grid data set in time to obtain a road segment data sequence;

[0015] Multiple road segment data sequences constitute a multi-modal database.

[0016] In one embodiment, the standardized road crowdsourcing data is preprocessed according to the corresponding data type of the standardized road crowdsourcing data to obtain multiple sub-modal data, including:

[0017] If the data type is numerical data, cut the standardized road crowdsourcing data according to a preset road segment unit, and mark the time stamps of the data of each road segment unit to obtain sub-modal data;

[0018] If the data type is image data, extract image features from the standardized road crowdsourcing data and retain the original metadata to obtain sub-modal data; the metadata includes location information and time;

[0019] If the data type is text data, extract semantic vectors from the standardized road crowdsourcing data using natural language to obtain sub-modal data.

[0020] In one embodiment, based on a hybrid attention neural network model, a road disease predictive score is generated using the multi-modal database according to a preset analysis time unit, including:

[0021] Perform multi-modal feature extraction on the road segment data sequence based on a hybrid structure of CNN and MLP to obtain feature vectors; the road segment data sequence includes numerical data, image feature vectors, and text semantic vectors; the numerical data includes acceleration; the text semantic vectors include potholes; the feature vectors include the abnormal mean index of acceleration and the pothole report density;

[0022] Identify the key feature channels of the feature vectors based on the attention weighting layer to obtain weighted feature representations;

[0023] Based on the prediction layer, a predictive score for road diseases is obtained according to the weighted feature representation and the historical traffic flow of the corresponding road segment.

[0024] In one embodiment, the mean index of acceleration anomaly is obtained through the following formula:

[0025]

[0026] Wherein, AAM i is the mean index of acceleration anomaly; a ij is the longitudinal acceleration collected for the j-th time on the i-th road segment; is the mean acceleration of the i-th road segment; n i is the total number of acceleration records of the i-th road segment;

[0027] The pothole report density is obtained through the following formula:

[0028]

[0029] Wherein, PRD i is the pothole report density; is the number of pothole reports on the i-th road segment within the preset analysis time unit; L i is the length of the road segment; T i is the preset analysis time unit; V i is the average vehicle traffic volume within the preset analysis time unit.

[0030] In one embodiment, the hybrid attention neural network model is constructed through the following method:

[0031] Using the crowdsourced data of the road segment and the true health index of the road segment to construct a training set;

[0032] Using the true health index as a supervision signal, pre-training the hybrid attention neural network model with the training set, and optimizing the parameters using the gradient descent algorithm to obtain the model architecture parameters;

[0033] Determine the hybrid attention neural network model using the model architecture parameters.

[0034] In one embodiment, it further includes: obtaining real maintenance feedback information, and updating the multi-modal database and adjusting the model parameters of the hybrid attention neural network model according to the real maintenance feedback information.

[0035] In a second aspect, the present application also provides an urban road preventive maintenance method using big data analysis, including:

[0036] Obtain the health index, traffic flow, and maintenance cost of a road segment; the maintenance cost includes the time cost and distance cost from the road segment to the maintenance resource center; the health index of the road segment is obtained by using the urban road disease prediction method based on big data analysis;

[0037] Generate a maintenance plan according to the health index of the road segment in a set interval mode;

[0038] Generate a maintenance strategy according to the maintenance plan, traffic flow, and maintenance cost; the maintenance strategy includes the execution priority of the maintenance plan;

[0039] Among them, the set interval mode corresponds to the following steps:

[0040] If the health index is 0.00 - 0.30 points, the maintenance plan is emergency repair;

[0041] If the health index is 0.31 - 0.60 points, the maintenance plan includes regular inspection and partial repair;

[0042] If the health index is 0.61 - 0.85 points, the maintenance plan is minor maintenance;

[0043] If the health index is 0.86 - 1.00 points, the maintenance plan is no intervention required.

[0044] In a third aspect, the present application also provides an urban road disease prediction device using big data analysis, including:

[0045] A data acquisition module for acquiring road crowdsourcing data;

[0046] A data processing module for processing the road crowdsourcing data based on a big data processing and analysis mechanism to obtain a multi-modal database;

[0047] A model algorithm module for generating a road disease predictive score based on a hybrid attention neural network model using the multi-modal database according to a preset analysis time unit.

[0048] In a fourth aspect, the present application also provides an urban road preventive maintenance device using big data analysis, including:

[0049] A data preparation module for obtaining the health index, traffic flow, and maintenance cost of a road segment;

[0050] A preventive maintenance module for generating a maintenance plan according to the health index of the road segment in a set interval mode;

[0051] A priority algorithm module for generating a maintenance strategy according to the maintenance plan, traffic flow, and maintenance cost.

[0052] The above-mentioned method for predicting urban road diseases and preventive maintenance using big data analysis makes full use of the multi-source heterogeneous data provided by crowdsourcing data to achieve large-scale and low-cost data collection; constructs a unified standard multi-modal database to make the model have good scalability and real-time performance; uses deep learning and attention mechanism to improve the prediction accuracy and effectively identify the development trend of diseases; provides data-driven support for the preventive maintenance of urban roads, reduces maintenance costs, and extends the service life of roads. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0054] Figure 1 It is a schematic flowchart of the method for predicting urban road diseases using big data analysis according to the present invention;

[0055] Figure 2 It is a schematic flowchart of the method for preventive maintenance of urban roads using big data analysis according to the present invention;

[0056] Figure 3 It is a composition structure diagram of the device for predicting urban road diseases using big data analysis according to the present invention;

[0057] Figure 4 It is a composition structure diagram of the device for preventive maintenance of urban roads using big data analysis according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will further describe the present application in detail with reference to the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0059] In one embodiment, as Figure 1 shown, a method for predicting urban road diseases using big data analysis is provided. In this embodiment, the method is exemplified by being applied to a terminal. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is realized through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] S101. Obtain road crowdsourcing data; the data types of road crowdsourcing data include numerical data, image data, and text data.

[0061] Crowdsourced data is a form of data collected through the Internet with the participation of a large number of users. In the scenario of urban road disease prediction, road crowdsourced data comes from a large number of road users. That is, with the help of their own devices such as smartphones and in-vehicle sensors, they actively or automatically upload information related to road conditions, so as to be able to reflect the actual road conditions in real time and comprehensively, greatly increasing the data coverage and update frequency. Similarly, the data collected by different users and different devices at different times and locations contains various complex road scenarios and environmental factors, providing multiple information dimensions for big data analysis and helping to discover more comprehensive and accurate road disease characteristics and laws. Further, the collection cost of road crowdsourced data is relatively low, without the need for a large investment in professional detection equipment and manpower, reducing the data acquisition cost and conforming to the principles of high efficiency and economy.

[0062] Schematically, road crowdsourced data covers three types of data: numerical data, image data, and text data. Among them, numerical data includes acceleration, gyroscope data, GPS coordinates, vehicle driving speed, etc. automatically uploaded by users' smartphones; image data includes road photos, video frames, etc. measured by users' in-vehicle sensors; text data includes potholes, cracks, abnormal noises, etc. actively reported by users with location markings. Further, it also includes the traffic flow statistics of the road.

[0063] S102. Process the road crowdsourced data based on the big data processing and analysis mechanism to obtain a multi-modal database.

[0064] Convert the obtained multi-source, heterogeneous, and complex road crowdsourced data into a structured and high-quality data set that can support subsequent model analysis, that is, a multi-modal database. Schematically, pre-clean the road crowdsourced data to remove error data such as outliers, noise, duplicates, and invalid coordinates, unify the units of different sensor data and standardize their value ranges. Further, according to the GPS longitude and latitude information in the data, map each user's data to the road segment grid through GIS technology to achieve spatial splicing and integration. Optionally, perform time period aggregation according to different time granularities such as daily, hourly, or weekly to generate an expression that can reflect the road characteristics of a specific time period. For data with missing features, use methods such as historical mean, spatial interpolation, or prediction to complete it, and finally construct a unified multi-modal database that adapts to the density of heterogeneous data.

[0065] S103. Based on the hybrid attention neural network model, generate a road disease predictive score using the multi-modal database according to the preset analysis time unit; the road disease predictive score includes the health indexes of multiple road segments.

[0066] The hybrid attention neural network model is used to analyze the data in the multimodal database to obtain the health index of multiple road sections, and then realize the predictive scoring of road diseases. Schematically, the hybrid attention neural network model consists of a feature extraction layer, an attention weighting layer, and a prediction layer. In the feature extraction layer, numerical data, image feature vectors, and text semantic vectors are processed to fuse multimodal data into a unified feature vector. The attention weighting layer adaptively calculates the importance weights of each sub-feature, highlights the data features that have a greater impact on road diseases, reduces noise interference, and forms a weighted feature representation. The prediction layer inputs the weighted features into the fully connected network, combines historical data and traffic flow data, and outputs a health status score between 0 and 1. The score not only reflects the current real-time status of the road, but also takes into account the prediction of future disease development trends by factors such as traffic flow. In this way, a quantitative basis is provided for road maintenance decisions to judge the degree of disease risk of each road section.

[0067] In the above-mentioned urban road disease prediction method using big data analysis, by obtaining multi-type road crowdsourcing data and uploading information with the help of a large number of users' devices, the data coverage and collection frequency are greatly increased. The data are processed using the big data processing and analysis mechanism to build a multimodal database that can fully reflect the various conditions of the road. Based on this, the hybrid attention neural network model can generate a road disease predictive score based on the preset analysis time unit, including the health index of multiple road sections, to achieve efficient and comprehensive detection of road diseases, making up for the shortcomings of manual inspections.

[0068] In one embodiment, the road crowdsourcing data is processed based on a big data processing and analysis mechanism to obtain a multimodal database, including:

[0069] S21. Pre-clean and standardize the road crowdsourcing data to obtain standardized road crowdsourcing data.

[0070] Road crowdsourcing data comes from different users, different devices and different scenarios. The data quality is uneven, and there are a lot of noise, outliers and missing values. The scales and ranges of different types of data are also different. Schematically, the noise and outliers in the data are removed by pre-cleaning. Specifically, the abnormal data can be identified and eliminated by setting a reasonable threshold range. Exemplarily, for acceleration data, if the acceleration value at a certain moment exceeds the reasonable range under normal driving conditions, the data point is regarded as an outlier and eliminated. At the same time, check the missing values in the data. For a small number of missing values, the corresponding records can be deleted; for a large number of missing values, linear interpolation, spline interpolation and other interpolation methods can be used to fill them.

[0071] Since different types of numerical data have different dimensions and value ranges, schematically, the Z-score normalization method is used to eliminate the differences in the data, and its formula is where X is the original data, μ is the data mean, and σ is the standard deviation. After the normalization process, the mean of the data is 0 and the standard deviation is 1, so that different types of data are comparable.

[0072] S22. Preprocess the standardized road crowdsourcing data according to the corresponding data types of the standardized road crowdsourcing data to obtain multiple sub-modal data.

[0073] According to the differences in data types, perform targeted preprocessing on the standardized road crowdsourcing data to extract valuable information for road disease prediction and obtain multiple sub-modal data. Schematically, for numerical data, perform data slicing according to a preset road segment unit, and divide the continuous numerical data into segments corresponding to the road segments. For image data, extract image features from the images in the standardized road crowdsourcing data. For text data, use natural language processing technology to extract semantic vectors from the text in the standardized road crowdsourcing data.

[0074] S23. Perform spatial mapping on the multiple sub-modal data according to the location information to obtain a road segment grid data set.

[0075] Since different types of sub-modal data may have different spatial representation methods, for unified analysis, spatial mapping needs to be performed according to the location information. Specifically, the urban road is divided into several grid units, and each grid unit corresponds to a specific road segment. According to the location information such as GPS coordinates in the sub-modal data, each data point is mapped to the corresponding road segment grid. Exemplarily, if the shooting location of a certain image data is within a specific road segment grid, then the image data and its corresponding feature vector are assigned to this grid. In this way, multiple sub-modal data are integrated into a unified road segment grid structure to form a road segment grid data set, which is convenient for subsequent comprehensive analysis of different types of data within the same road segment.

[0076] S24. Align the multiple sub-modal data of the same road segment in the road segment grid data set in time to obtain a road segment data sequence.

[0077] The different sub-modal data of the same road segment may be collected at different time points and may have different collection time intervals. In order to match and correlate the data in the time dimension for subsequent time series prediction, time alignment is required. Schematically, the method of timestamp matching can be adopted. Based on a unified time scale, the different sub-modal data within the same road segment are sorted in chronological order to construct a road segment data sequence that can reflect the changes in the road conditions. Optionally, for data points with a large time interval, interpolation or smoothing processing methods can be used for filling to ensure the continuity of the data. Exemplarily, in an actual scenario, the multi-modal road crowdsourcing data of the same road segment are uploaded by different users at different times. The data types are diverse and the upload situations are complex. There may be only one data type or multiple data types within a certain time period. Specifically, all the sub-modal data within the same road segment are sorted by timestamp, and a unified time scale is determined as the benchmark. If a user uploads numerical, image, and text data of road segment 1 at different times on different days, these data are first sorted according to the specific time and then unified to the time scale in hours. Further, the different types of data at the same time point or close time points are matched and correlated. At time 1, the user uploads the numerical data of road segment 1, and at a close time 2, the user uploads the text data of road segment 1. These two pieces of data are correlated to mutually corroborate the problems existing on the road.

[0078] The road segment data sequence not only records the information of a single data type at different time points, but more importantly, it can present the correlation and changes between different types of data, reflecting the evolution of the road conditions over time. Exemplarily, through the change trend of numerical data combined with subsequent text and image data, the process of road diseases from potential to appearance can be clearly seen.

[0079] S25. The data sequences of multiple road segments constitute a multi-modal database.

[0080] Schematically, the road segment data sequences of all road segments in the city are collected, classified and stored according to the location information of the road segments. Since road crowdsourcing data contains a large amount of unstructured and semi-structured data, a NoSQL or distributed database that supports such data can be used, including MongoDB, ElasticSearch, or a data lake in the Hadoop / Spark ecosystem. Among them, the data sequence of each road segment is used as a record in the database, that is, documents or entries are divided according to road segments, which include fields such as numerical values, image features, and text semantics. At the same time, the data source is recorded and arranged in chronological order. To facilitate data query and analysis, a data association and indexing mechanism is established in the database. Exemplarily, according to the road segment index, all information arranged in data order within a specific road segment can be quickly retrieved. Further, through the time index, the information of the road segment within a certain time period can be quickly retrieved; specific types of data can also be filtered out through the data type index. Further, an association relationship between different data types is established, so that when querying a certain type of data, other types of data related to it can be conveniently obtained, and collaborative analysis of multi-modal data can be realized.

[0081] Since road crowdsourcing data is continuously uploaded, the multi-modal database needs to be updated in real time to reflect the latest changes in road conditions. When new data is uploaded, it is added to the corresponding road segment data sequence according to the above time alignment and integration method. Optionally, the database is maintained regularly to clean up invalid data and update data indexes to ensure the efficient operation of the database and the accuracy of the data.

[0082] In one embodiment, the standardized road crowdsourcing data is preprocessed according to the corresponding data type of the standardized road crowdsourcing data to obtain a plurality of sub-modal data, including:

[0083] S31. If the data type is numerical data, the standardized road crowdsourcing data is cut according to a preset road segment unit, and a time stamp is marked on the data of each road segment unit to obtain sub-modal data.

[0084] In road crowdsourcing data, numerical data such as acceleration, vehicle speed, and traffic flow can reflect road conditions from a quantitative perspective. Schematically, the road is divided into road segment units of a fixed length. According to the GPS positioning information, the continuous numerical data is divided into each road segment unit. Exemplarily, each road segment unit is divided into 100 meters. During the vehicle driving process, the acceleration data is continuously recorded. The road position where the vehicle is located at different times is determined by GPS, and the acceleration data at the corresponding time is assigned to the corresponding road segment unit. Further, a timestamp mark is added to the data of each road segment unit, that is, the specific time of data collection is recorded. The timestamp can provide information in the time dimension and help analyze the change of road conditions over time. Exemplarily, by analyzing the acceleration change of the same road segment in different time periods, the flatness change of the road segment at different times can be understood.

[0085] S32. If the data type is image data, image feature extraction is performed on the standardized road crowdsourcing data, and the original metadata is retained to obtain multi-modal data; the metadata includes location information and time.

[0086] Image data can visually present the conditions of the road surface, such as diseases like cracks and potholes. The original image data contains a large amount of redundant information, and it is less efficient to directly use it for analysis. Therefore, image features are extracted and the original metadata is retained. Schematically, computer vision technology is used to extract representative features from the standardized image data. Specifically, pre-trained convolutional neural networks (CNNs) such as VGG and ResNet are used to automatically learn features such as textures, edges, and shapes in the image, and convert the high-dimensional image data into low-dimensional feature vectors. Exemplarily, for an image containing a road pothole, the CNN can extract the size, shape, edge features, etc. of the pothole and represent them with a feature vector to reduce the amount of data and improve the analysis efficiency. Further, in addition to extracting image features, the original metadata including location information and time should also be retained. The location information can determine the specific road position where the image is taken, and the time information can reflect the moment when the image is taken.

[0087] S33. If the data type is text data, semantic vector extraction is performed on the standardized road crowdsourcing data using natural language to obtain multi-modal data.

[0088] The text data contains subjective descriptions of the road conditions by users. Exemplarily, "The road is bumpy" and "Saw a big hole" etc. The text information is very valuable for understanding the road conditions and the text data can be converted into semantic vectors for direct processing by a computer. Exemplarily, natural language processing techniques are used to extract semantic vectors from the standardized text data. Specifically, pre-trained word vector models such as Word2Vec and GloVe are used to convert each word in the text into a vector representation, and the word vectors are combined into semantic vectors of sentences or text segments through methods such as average pooling and max pooling. Exemplarily, for the text "The road is bumpy", first convert words such as "road", "is", and "bumpy" into vectors, and then obtain the semantic vector of the whole sentence through average pooling. The text data is converted into a numerical vector form that can be processed by a computer, facilitating subsequent fusion analysis with other types of data. Through semantic vector extraction, the semantic information in the text data can be mined, providing a more comprehensive basis for road disease prediction.

[0089] In one embodiment, based on a hybrid attention neural network model, a road disease predictive score is generated using a multi-modal database according to a preset analysis time unit, including:

[0090] S41. Perform multi-modal feature extraction on the road segment data sequence based on a hybrid structure of CNN and MLP to obtain feature vectors; the road segment data sequence includes numerical data, image feature vectors, and text semantic vectors; the numerical data includes acceleration; the text semantic vectors include potholes; the feature vectors include the abnormal mean index of acceleration and the pothole report density.

[0091] The road segment data sequence contains multi-modal information such as numerical data, image feature vectors, and text semantic vectors. Different types of data reflect the road conditions from multiple perspectives and can corroborate each other, providing rich information for accurately extracting road disease-related features. Performing multi-modal feature extraction based on a hybrid structure of CNN and MLP can make full use of the characteristics of each modal data and mine more representative feature vectors. Among them, different types of data are complementary when reflecting road conditions. Exemplarily, in the same time unit of the same road segment, an abnormal acceleration in the numerical data may imply unevenness on the road surface, while the image data can visually show the disease conditions on the road surface, such as cracks and potholes. The text data may contain subjective descriptions of the road conditions by users, such as "There are obvious bumps" and "Saw a big hole", further determining the road disease information. In the same road segment, in different time units, it may not be possible to obtain various types of information exactly, but if there is abnormal acceleration information in the first time unit and text data is reflected in the second time unit, they can also corroborate each other to determine the road disease situation.

[0092] Schematically, CNN can be used as a front-end module of the model to process image data and store image features in a database for direct model calls. Specifically, CNN can automatically extract local and global features in images through operations such as convolutional layers and pooling layers. For example, the convolutional layer can identify features such as edges and textures in an image, and the pooling layer can reduce the dimensionality of the features to reduce the amount of calculation. Through multi-layer convolution and pooling operations, CNN can learn important features related to road hazards in images, such as the shape and size of potholes.

[0093] Schematically, MLP (Multi-layer Perceptron) is a general neural network structure suitable for processing numerical data and text semantic vectors. For numerical data, MLP can perform nonlinear transformations on it through multiple fully connected layers to mine the potential patterns in the data. For text semantic vectors, MLP can further extract and transform them to convert text information into more representative numerical features. Exemplarily, the text semantic vector containing "holes" is mapped to a higher-dimensional feature space for better fusion with other modal data.

[0094] After processing the image feature vector, numerical data and text semantic vector respectively, their outputs are fused. The features of different modalities can be combined by splicing, weighted summation, etc. to form a comprehensive feature vector, which can reflect the disease evidence information in the road section to improve the accuracy of prediction.

[0095] S42. Identify key feature channels of the feature vector based on the attention weighted layer to obtain a weighted feature representation.

[0096] In the feature vector, different feature channels have different importance for road disease prediction. The role of the attention weighting layer is to automatically identify these key feature channels and assign different weights to them, so as to highlight important features, suppress low-correlation features, and obtain a more representative weighted feature representation. Specifically, the attention mechanism can dynamically adjust the weights of each feature channel according to the content of the feature vector. In the attention weighting layer, the importance score of each feature channel is calculated according to the attention function. Exemplarily, the dot product attention mechanism is used to calculate the dot product of the feature vector and a learnable query vector, and then the scores are normalized to the interval [0,1] through the softmax function to obtain the attention weights of each feature channel. According to the calculated attention weights, the key feature channels are identified. The feature channels with larger weights indicate that they are more important for road disease prediction. Exemplarily, if the weight of the acceleration anomaly mean index in the feature vector is large, it means that this index plays a key role in road disease prediction within a certain road section. Specifically, within a certain road section, there is acceleration numerical data covering continuous time and provided by multiple different users, then the confidence of this acceleration numerical data is relatively high, and the weight is relatively high. For the description of potholes in text data, only one user responds, then the confidence of this text data is relatively low, and the weight is relatively low.

[0097] Multiply the feature vector by the corresponding attention weights to obtain a weighted feature representation, enabling the subsequent prediction layer to focus more on key features and improve the accuracy of road disease prediction.

[0098] S43. Based on the prediction layer, obtain the road disease predictive score according to the weighted feature representation and the historical traffic volume of the corresponding road section.

[0099] The main task of the prediction layer is to generate a road disease predictive score by comprehensively considering the actual condition of the road and traffic pressure according to the weighted feature representation and the historical traffic volume of the corresponding road section. Exemplarily, use the weighted feature representation and the historical traffic volume of the corresponding road section as the input of the prediction layer. The historical traffic volume reflects the usage frequency and traffic pressure of the road and has an important impact on the development of road diseases. Exemplarily, roads with larger traffic volumes are more prone to wear and damage, and the speed of disease development may be faster. Fusing the weighted feature representation and the historical traffic volume can more comprehensively consider the factors affecting road diseases. The prediction layer can be constructed using a fully connected neural network. Through learning historical data, the model can establish a mapping relationship between the weighted feature representation, the historical traffic volume, and the road disease predictive score. Specifically, through multiple fully connected layers, non-linear transformations are performed on the input data, and finally a road disease predictive score is output. The prediction layer calculates the road disease predictive score according to the input data and the learned mapping relationship. Exemplarily, the score is between 0 and 1, and the larger the value, the higher the possibility of road diseases.

[0100] In one of the embodiments, the mean index of acceleration anomaly is obtained through the following formula:

[0101]

[0102] where AAM i is the mean index of acceleration anomaly; a ij is the longitudinal acceleration collected for the j-th time in the i-th road segment; is the mean acceleration of the i-th road segment; n i is the total number of acceleration records of the i-th road segment;

[0103] The pothole report density is obtained through the following formula:

[0104]

[0105] where PRD i is the pothole report density; is the number of pothole reports in the i-th road segment within the preset analysis time unit; L i is the length of the road segment; T i is the preset analysis time unit; V i is the average vehicle traffic volume within the preset analysis time unit.

[0106] In one of the embodiments, the hybrid attention neural network model is constructed through the following method:

[0107] S51. Use the crowdsourcing data of the road segment and the true health index of the road segment to construct a training set.

[0108] Collect the crowdsourcing data of the road segment from multiple sources, including numerical data, image data, and text data. At the same time, obtain the true health index of the corresponding road segment, which is usually determined based on professional road detection reports, historical maintenance records, and expert evaluations, etc., and can accurately reflect the actual health status of the road. Integrate the crowdsourcing data and the true health index to ensure that each data sample contains various feature information of the road segment and the corresponding health index label, and form a training set.

[0109] S52. Using the true health index as a supervision signal, pre-train the hybrid attention neural network model with the training set, and use the gradient descent algorithm to optimize the parameters to obtain the model architecture parameters.

[0110] Schematically, initialize the hybrid attention neural network model, and input the crowdsourced data in the training set into the initialized model. The data undergoes forward propagation through the feature extraction layer, attention weighting layer, prediction layer, etc. In the feature extraction layer, the model uses CNN and MLP to extract features from multi-modal data to obtain feature vectors; the attention weighting layer assigns weights according to the importance of the feature vectors to highlight key features; the prediction layer outputs a road disease prediction score based on the weighted feature representation and information such as historical traffic flow.

[0111] The predictive score is the prediction result of the model on the road health condition under the current parameters. Compare the prediction result of the model, that is, the predictive score, with the true health index, and calculate the difference between the two through the loss function. Schematically, use the weighted mean square error (Weighted MSE) as the loss function, and the formula is where y i is the labeled true health index; is the predictive score; ω i is the sample weight, which is used to emphasize the influence of high-confidence data. The loss value reflects the deviation degree between the model prediction and the real situation. The smaller the loss value, the more accurate the model prediction.

[0112] Furthermore, use the gradient descent algorithm to optimize the parameters of the model. The gradient descent algorithm determines the update direction of the parameters by calculating the gradient of the loss function with respect to the model parameters, so that the value of the loss function gradually decreases in each iteration. Specifically, adjust the weights and biases of the model according to the opposite direction of the gradient. Exemplarily, the update formula for the weights is where α is the learning rate, which controls the step size of parameter update. Through multiple iterations, continuously adjust the model parameters to make the prediction result of the model gradually approach the real health index, thereby improving the accuracy of the model. During the training process, techniques such as cross-validation can also be used to select appropriate hyperparameters such as the learning rate and batch size to prevent the model from overfitting and improve the generalization ability of the model.

[0113] S53. Determine the hybrid attention neural network model using the model architecture parameters.

[0114] Apply the optimized model architecture parameters to the corresponding structures of the hybrid attention neural network model to formally determine the specific form of the model. The parameters of each layer of the model, such as the feature extraction layer, attention weighting layer, and prediction layer, have been determined, and the model has the ability to accurately predict road diseases based on the input crowdsourced data of road segments.

[0115] In one embodiment, it further includes: obtaining real maintenance feedback information, and updating the multi-modal database and adjusting the model parameters of the hybrid attention neural network model according to the real maintenance feedback information.

[0116] The real maintenance feedback information mainly comes from the actual implementation of road maintenance work. After the maintenance personnel complete the road maintenance tasks, they will record the specific content of the maintenance, including the disease conditions before maintenance, filling potholes and repairing cracks, etc. They will also evaluate the road conditions after maintenance, including the flatness of the road and whether the diseases have been effectively treated. Further, the real maintenance feedback information is integrated into the multi-modal database to hide the road segment data within a period of time and continuously monitor the road segment conditions after maintenance.

[0117] The real maintenance feedback information provides an important basis for adjusting the model parameters. When there is a deviation between the prediction result of the model and the actual maintenance situation, it is necessary to adjust the model parameters. Exemplarily, if the model predicts that the disease of this road segment is relatively light and does not require immediate maintenance, but it is found that the disease is relatively serious during actual maintenance, it indicates that there is a deviation in the prediction of this road segment by the model. At this time, according to the feedback information, the backpropagation algorithm and the gradient descent method are used to fine-tune the parameters of the model. Specifically, adjust the weights of each feature in the attention weighting layer to make the model pay more attention to the features related to the actual disease in subsequent predictions; or adjust the parameters of the prediction layer to make the output of the model closer to the actual situation. By continuously adjusting the model parameters according to the real maintenance feedback information, the prediction accuracy of the model will be continuously improved, so as to better serve the road disease prediction and preventive maintenance work.

[0118] In the above method, the maintenance record is used as the real maintenance feedback information to update the multi-modal database. At the same time, compare the prediction result with the actual maintenance situation, and adjust the parameters of the hybrid attention neural network model according to the difference. Implement a closed-loop process of "prediction - implementation - feedback - retraining" to enable the model to continuously adapt to the changes of the road and continuously improve the prediction accuracy.

[0119] In one embodiment, as Figure 2 shown, a method for preventive maintenance of urban roads using big data analysis is provided. In this embodiment, taking the same application environment as the method for predicting urban road diseases using big data analysis, the following steps are included:

[0120] S201. Obtain the health index, traffic flow, and maintenance cost of the road segment; the maintenance cost includes the time cost and distance cost from the road segment to the maintenance resource center; the health index of the road segment is obtained by using the method for predicting urban road diseases using big data analysis.

[0121] Traffic flow reflects the usage frequency and busyness of roads, which not only affects the wear speed of roads, but also affects the implementation difficulty and influence scope of maintenance operations. The maintenance cost includes the time cost and distance cost from the road section to the maintenance resource center.

[0122] S202. Generate a maintenance plan according to the health index of the road section in a set interval mode.

[0123] Among them, the set interval mode corresponds to the following steps:

[0124] If the health index is between 0.00 - 0.30 points, the maintenance plan is emergency repair;

[0125] If the health index is between 0.31 - 0.60 points, the maintenance plan includes regular inspections and partial repairs;

[0126] If the health index is between 0.61 - 0.85 points, the maintenance plan is minor maintenance;

[0127] If the health index is between 0.86 - 1.00 points, the maintenance plan is no intervention required.

[0128] Evaluate the obtained health index of the road section to determine the interval range it is in. When the health index is between 0.00 - 0.30 points, it indicates that the road has serious diseases, such as large - area potholes, structural cracks, etc., which pose a greater threat to driving safety. Therefore, emergency repair is required immediately to eliminate potential safety hazards. When the health index is between 0.31 - 0.60 points, it shows that there are early damage signs on the road, such as slight cracks, local unevenness, etc. At this time, adopting the strategy of regular inspections and partial repairs can timely detect the development trend of diseases and repair them before the diseases further deteriorate, avoiding small problems from evolving into major safety hazards, and at the same time effectively controlling the maintenance cost. If the health index is between 0.61 - 0.85 points, it means that the road condition is relatively good, but there are still some minor problems, such as slightly uneven or slightly abnormal vibrations. At this time, arranging minor maintenance, through daily maintenance and upkeep of the road, such as cleaning, filling small potholes, etc., can maintain the good state of the road and extend the service life of the road. When the health index reaches 0.86 - 1.00 points, it indicates that the road is in a normal traffic state and no additional maintenance operations are required. Continue to monitor the road, collect data, and provide a basis for subsequent evaluation and decision - making.

[0129] S203. Generate a maintenance strategy according to the maintenance plan, traffic flow, and maintenance cost; the maintenance strategy includes the execution priority of the maintenance plan.

[0130] Schematically, for road sections requiring emergency repair, due to their serious impact on traffic safety, maintenance operations should be prioritized. When determining the execution priority, traffic flow factors also need to be considered, and it is advisable to choose a time period with relatively low traffic flow for operations to minimize interference with traffic. For road sections undergoing regular inspections, local repairs, and minor maintenance, the priority is determined comprehensively based on their health index, traffic flow, and maintenance cost. Exemplarily, for road sections with a relatively low health index and high traffic flow, maintenance should be prioritized because the diseases of such road sections may deteriorate rapidly with the increase in traffic flow, and at the same time, they have a greater impact on traffic. Further, according to the traffic flow, when formulating a maintenance strategy, it is necessary to avoid carrying out maintenance operations during peak traffic hours to prevent traffic congestion. For road sections with high traffic flow, the maintenance operation time should be minimized as much as possible, and efficient maintenance technologies and equipment should be adopted, or operations should be carried out during time periods with low traffic flow such as at night. At the same time, according to the changing pattern of traffic flow, the maintenance sequence can be reasonably arranged, and road sections with less impact on traffic should be given priority. Even further, on the premise of meeting the maintenance requirements, a maintenance plan with relatively low cost should be selected as much as possible. Exemplarily, for adjacent road sections to be repaired, they can be arranged in the same maintenance strategy to reduce the maintenance cost and achieve efficient, economical, and safe road maintenance.

[0131] Based on the accurate prediction results, the above method can arrange the maintenance plan and resource allocation more reasonably in preventive maintenance decision-making. For example, the maintenance measures are determined according to the health index of the road section, and the maintenance strategy is optimized by combining traffic flow and maintenance cost, improving the efficiency and economic benefits of road maintenance, thereby solving the problem of lack of dynamic adjustment in decision-making.

[0132] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown sequentially in the direction of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential either, but can be executed alternately or in turn with at least some of the steps or stages in other steps or other steps.

[0133] Based on the same inventive concept, an embodiment of the present application further provides a big data analysis-based urban road disease prediction device for implementing the above-mentioned urban road disease prediction method using big data analysis. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the big data analysis-based urban road disease prediction device provided below can refer to the limitations on the big data analysis-based urban road disease prediction method in the above text, and will not be repeated here.

[0134] In an exemplary embodiment, as Figure 3 shown, a big data analysis-based urban road disease prediction device is provided, including:

[0135] A data acquisition module, configured to acquire road crowdsourcing data;

[0136] A data processing module, configured to process the road crowdsourcing data based on a big data processing and analysis mechanism to obtain a multi-modal database;

[0137] A model algorithm module, configured to generate a road disease predictive score using the multi-modal database based on a hybrid attention neural network model according to a preset analysis time unit.

[0138] In one of the embodiments, it further includes:

[0139] A standardization module, configured to perform pre-cleaning and standardization processing on the road crowdsourcing data to obtain standardized road crowdsourcing data;

[0140] A preprocessing module, configured to perform preprocessing on the standardized road crowdsourcing data according to the corresponding data type of the standardized road crowdsourcing data to obtain multiple sub-modal data;

[0141] A spatial mapping module, configured to perform spatial mapping on the multiple sub-modal data according to location information to obtain a road segment grid data set;

[0142] A time series module, configured to perform time alignment on the multiple sub-modal data of the same road segment in the road segment grid data set to obtain a road segment data sequence.

[0143] In one of the embodiments, it further includes:

[0144] A feature extraction module, configured to perform multi-modal feature extraction on the road segment data sequence based on a hybrid structure of CNN and MLP to obtain a feature vector;

[0145] An attention weighting module, configured to identify key feature channels of the feature vector based on an attention weighting layer to obtain a weighted feature representation;

[0146] A prediction module, configured to obtain a road disease predictive score based on a prediction layer according to a weighted feature representation and the historical traffic volume of the corresponding road segment.

[0147] Similarly, as Figure 4 shown, an embodiment of the present application further provides a big data analysis-based urban road preventive maintenance device for implementing the above-mentioned big data analysis-based urban road preventive maintenance method, including:

[0148] A data preparation module, configured to obtain the health index, traffic flow, and maintenance cost of a road segment;

[0149] A preventive maintenance module, configured to generate a maintenance plan according to the health index of the road segment in a set interval mode;

[0150] A priority algorithm module, configured to generate a maintenance strategy according to the maintenance plan, traffic flow, and maintenance cost.

[0151] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0152] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0153] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are only illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure solution. Those of ordinary skill in the art can understand and implement it without creative work.

[0154] The above embodiments only represent several implementation manners of the embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the application embodiments. It should be noted that for those of ordinary skill in the art, without departing from the concept of the embodiments of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. A method for predicting urban road diseases using big data analysis, characterized in that, The method includes: Obtaining road crowdsourcing data; the data types of the road crowdsourcing data include numerical data, image data, and text data; Processing the road crowdsourcing data based on a big data processing and analysis mechanism to obtain a multimodal database; Based on a hybrid attention neural network model, using the multimodal database to generate a road disease predictive score according to a preset analysis time unit; the road disease predictive score includes health indexes of multiple road segments.

2. The method according to claim 1, wherein The processing of the road crowdsourcing data based on a big data processing and analysis mechanism to obtain a multimodal database includes: Performing pre-cleaning and standardization processing on the road crowdsourcing data to obtain standardized road crowdsourcing data; Performing preprocessing on the standardized road crowdsourcing data according to the data types corresponding to the standardized road crowdsourcing data to obtain multiple sub-modal data; Performing spatial mapping on the multiple sub-modal data according to location information to obtain a road segment grid data set; Performing time alignment on the multiple sub-modal data of the same road segment in the road segment grid data set to obtain a road segment data sequence; Multiple road segment data sequences constitute a multimodal database.

3. The method according to claim 2, characterized in that, The performing preprocessing on the standardized road crowdsourcing data according to the data types corresponding to the standardized road crowdsourcing data to obtain multiple sub-modal data includes: If the data type is the numerical data, cutting the standardized road crowdsourcing data according to a preset road segment unit and performing timestamp marking on the data of each road segment unit to obtain the sub-modal data; If the data type is the image data, performing image feature extraction on the standardized road crowdsourcing data and retaining the original metadata to obtain the sub-modal data; the metadata includes location information and time; If the data type is the text data, performing semantic vector extraction on the standardized road crowdsourcing data using natural language to obtain the sub-modal data.

4. The method according to claim 2, wherein The generating a road disease predictive score based on a hybrid attention neural network model using the multimodal database according to a preset analysis time unit includes: Performing multimodal feature extraction on the road segment data sequence based on a CNN and MLP hybrid structure to obtain feature vectors; the road segment data sequence includes numerical data, image feature vectors, and text semantic vectors; the numerical data includes acceleration; the text semantic vectors include potholes; the feature vectors include an acceleration anomaly mean index and a pothole report density; Identifying key feature channels of the feature vectors based on an attention weighting layer to obtain a weighted feature representation; Based on a prediction layer, obtaining a road disease predictive score according to the weighted feature representation and the historical traffic flow of the corresponding road segment.

5. The method according to claim 4, wherein: The acceleration anomaly mean index is obtained through the following formula: Among them, AAM i is the acceleration anomaly mean index; a ij is the longitudinal acceleration collected for the j-th time on the i-th road section; is the mean acceleration of the i-th road section; n i is the total number of acceleration records of the i-th road section; The pothole report density is obtained through the following formula: Among them, PRD i is the pothole report density; is the number of pothole reports on the i-th road section within the preset analysis time unit; L i is the road section length; T i is the preset analysis time unit; V i is the average vehicle traffic volume within the preset analysis time unit.

6. The method according to any one of claims 1 to 5, characterized in that The hybrid attention neural network model is constructed through the following method: Constructing a training set using the crowdsourcing data of the road segment and the true health index of the road segment; The true health index is used as a supervision signal to pre-train the hybrid attention neural network model using the training set, and the gradient descent algorithm is used to optimize the parameters to obtain the model architecture parameters; The hybrid attention neural network model is determined using the model architecture parameters.

7. The method according to claim 6, characterized in that, The method further includes: Obtaining real maintenance feedback information, and updating the multi-modal database and adjusting the model parameters of the hybrid attention neural network model according to the real maintenance feedback information.

8. A preventive maintenance method for urban roads using big data analysis, characterized in that, Including Obtaining the health index, traffic flow, and maintenance cost of a road segment; the maintenance cost includes the time cost and distance cost from the road segment to the maintenance resource center; the health index of the road segment is obtained by the urban road disease prediction method using big data analysis according to claim 1; Generating a maintenance plan according to the health index of the road segment in a set interval mode; Generating a maintenance strategy according to the maintenance plan, traffic flow, and the maintenance cost; the maintenance strategy includes the execution priority of the maintenance plan; Wherein, the set interval mode corresponds to the following steps: If the health index is 0.00 - 0.30 points, the maintenance plan is emergency repair; If the health index is 0.31 - 0.60 points, the maintenance plan includes regular inspection and partial repair; If the health index is 0.61 - 0.85 points, the maintenance plan is small-scale maintenance; If the health index is 0.86 - 1.00 points, the maintenance plan is no intervention required.

9. An urban road disease prediction device using big data analysis, characterized in that, The device includes: A data acquisition module for acquiring road crowdsourcing data; A data processing module for processing the road crowdsourcing data based on a big data processing and analysis mechanism to obtain a multi-modal database; A model algorithm module for generating a road disease predictive score using the multi-modal database based on a hybrid attention neural network model according to a preset analysis time unit.

10. An urban road preventive maintenance device using big data analysis, characterized in that, The device includes: A data preparation module for obtaining the health index, traffic flow, and maintenance cost of a road segment; A preventive maintenance module for generating a maintenance plan according to the health index of the road segment in a set interval mode; A priority algorithm module for generating a maintenance strategy according to the maintenance plan, traffic flow, and the maintenance cost.

Images