Soundscape prediction model training method and prediction method integrating multi-source urban data

By integrating multi-source urban data to train a soundscape prediction model, and using a neural network model to train the soundscape prediction model, the geographical limitations of soundscape measurement instruments are solved, and comprehensive and accurate soundscape prediction without the need for on-site data collection is achieved.

CN117743931BActive Publication Date: 2024-11-12SHENZHEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311763103.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-19
Publication Date
2024-11-12
Estimated Expiration
2043-12-19

AI Technical Summary

Technical Problem

Existing soundscape measurement instruments have geographical limitations and cannot perform soundscape measurements in all geographical areas.

Method used

A soundscape prediction model training method that integrates multi-source urban data is adopted. By acquiring environmental data of the training geographic area, performing statistical analysis, and using a neural network model to train the soundscape prediction model, including a feature selection layer, a shared hidden layer, and a sound source hidden layer, the loss function is calculated to train the model.

Benefits of technology

It enables comprehensive and accurate prediction of soundscapes without the need for on-site data collection, expanding the application scenarios of soundscape prediction, making up for the shortcomings of a single data source, and improving the overall efficiency of soundscape prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117743931B_ABST
    Figure CN117743931B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of soundscape prediction technology, and specifically to a soundscape prediction model training method and a prediction method that integrate multi-source urban data. The present invention performs statistical analysis on environmental data to obtain features related to the soundscape, that is, to obtain soundscape training features, and trains a neural network model based on the soundscape training features to obtain a soundscape prediction model. Environmental data in a geographic area to be predicted where soundscape prediction is required is collected, and a soundscape prediction model is applied to the environmental data, and the prediction model outputs the soundscape information of the geographic area to be predicted. The environmental data of the present invention is existing data, and there is no need to collect it on site, that is, the soundscape information can be predicted by using existing environmental data, thereby expanding the application scenario of the soundscape prediction method of the present invention. Moreover, the present invention uses a soundscape prediction model to predict soundscape information, and can comprehensively and accurately predict the soundscape.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of soundscape prediction, and in particular to a soundscape prediction model training method and a prediction method integrating multi-source urban data. Background Art

[0002] Soundscape is the sound environment perceived, experienced or understood by an individual or group. Soundscape covers all sounds in the environment. Accurately knowing the soundscape in the environment can provide technical support for improving the quality of urban soundscape and protecting the health of residents. Existing technologies often use sound measuring instruments to collect soundscapes in the environment, but sound measuring instruments have geographical limitations, that is, it is impossible to set up sound measuring instruments in all geographical areas to measure the soundscape of the area.

[0003] In summary, existing technologies for measuring soundscapes have limitations.

[0004] Therefore, the prior art still needs to be improved and enhanced. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a soundscape prediction model training method and a prediction method integrating multi-source urban data, which solves the problem that the prior art of measuring soundscape has limitations.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for training a soundscape prediction model integrating multi-source urban data, which includes:

[0008] Acquiring environmental data stored in a training geographical area, and performing statistical analysis on the environmental data to obtain soundscape training features of the training geographical area;

[0009] Applying a neural network model to the soundscape training features to obtain a soundscape training label for the training geographic area output by the neural network model;

[0010] Acquire the measured soundscape labels of the training geographical area, calculate the loss function of the neural network model based on the measured soundscape labels and the soundscape training labels, and train the neural network model based on the loss function to obtain a soundscape prediction model.

[0011] In one implementation, acquiring the environmental data stored in the training geographical area and performing statistical analysis on the environmental data to obtain the soundscape training features of the training geographical area includes:

[0012] Obtaining point of interest data information, road data, building data, street view images and green data from the environmental data stored in the training geographic area;

[0013] Performing statistical analysis on the interest point data information, the road data, the building data, the street view image and the greening data to obtain the number of interest points, interest point density, road density, the shortest distance between the center position and the road, building density, building volume ratio, green space coverage, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image in the training geographical area;

[0014] The number of points of interest, the density of points of interest, the road density, the shortest distance, the building density, the building volume ratio, the green space coverage rate, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image are used as soundscape training features of the training geographical area.

[0015] In one implementation, the neural network model includes a feature screening layer, a shared hidden layer, a sound intensity hidden layer, and a sound source hidden layer, wherein the feature screening layer is cascaded with the shared hidden layer, the sound intensity hidden layer, and the sound source hidden layer, respectively, and the shared hidden layer is cascaded with the sound intensity hidden layer and the sound source hidden layer, respectively. The neural network model is applied to the soundscape training features to obtain a soundscape training label for the training geographical area output by the neural network model, including:

[0016] Inputting the soundscape training features into a feature screening layer, obtaining shared features, sound intensity features, and sound source features screened by the feature screening layer from the soundscape training features, wherein the shared features are soundscape features related to both sound intensity and sound source, the sound intensity features are soundscape features only related to sound intensity, and the sound source features are soundscape features only related to sound source;

[0017] After the shared features are input into the shared hidden layer, the sound intensity features are input into the sound intensity hidden layer, and the sound source features are input into the sound source hidden layer, so as to obtain the sound intensity training labels output by the sound intensity hidden layer and the sound source training labels output by the sound source hidden layer, and use the sound intensity training labels and the sound source training labels as soundscape training labels.

[0018] In one implementation, obtaining the measured soundscape labels of the training geographical area, calculating the loss function of the neural network model based on the measured soundscape labels and the soundscape training labels, and training the neural network model based on the loss function to obtain the soundscape prediction model includes:

[0019] Obtaining a sound intensity measured label and a sound source measured label from the sound scene measured label;

[0020] Calculating a sound intensity loss function according to the difference between the sound intensity training label and the sound intensity measured label;

[0021] Calculating a sound source loss function based on the sound source training label and the sound source measured label;

[0022] According to the sound intensity loss function and the sound source loss function, the parameters of the neural network model are adjusted until the sound intensity loss function and the sound source loss function are both less than set values, so as to obtain a soundscape prediction model.

[0023] In one implementation, the calculating of the sound intensity loss function according to the difference between the sound intensity training label and the sound intensity measured label includes:

[0024] The sound intensity mean square error formed by the sound intensity training label and the sound intensity measured label is calculated, and the sound intensity mean square error is used as the sound intensity loss function.

[0025] In one implementation, the calculating the sound source loss function according to the sound source training label and the sound source measured label includes:

[0026] Calculating each logarithmic function of each of the sound source training labels;

[0027] Multiplying each of the logarithmic functions by each corresponding sound source measured label to obtain each intermediate result;

[0028] A weighted calculation is performed on each of the intermediate results to obtain a sound source loss function.

[0029] In a second aspect, an embodiment of the present invention further provides a soundscape prediction method, applying the above-mentioned soundscape prediction model, the soundscape prediction method comprises:

[0030] Acquiring environmental data in a geographic area to be measured, and performing statistical analysis on the environmental data to obtain soundscape characteristics in the geographic area to be measured;

[0031] A soundscape prediction model is applied to the soundscape feature to obtain a sound source prediction label and a sound intensity prediction label output by the soundscape prediction model.

[0032] In a third aspect, an embodiment of the present invention further provides a soundscape prediction model training device integrating multi-source urban data, wherein the training device includes the following components:

[0033] A feature statistics module is used to obtain the environmental data stored in the training geographical area and perform statistical analysis on the environmental data to obtain the soundscape training features of the training geographical area;

[0034] A prediction module, configured to apply a neural network model to the soundscape training features to obtain a soundscape training label for the training geographical area output by the neural network model;

[0035] The training module is used to obtain the measured soundscape labels of the training geographical area, calculate the loss function of the neural network model according to the measured soundscape labels and the soundscape training labels, and train the neural network model according to the loss function to obtain a soundscape prediction model.

[0036] In a fourth aspect, an embodiment of the present invention further provides a terminal device, wherein the terminal device includes a memory, a processor, and a soundscape prediction model training program for fusing multi-source city data stored in the memory and executable on the processor, and when the processor executes the soundscape prediction model training program for fusing multi-source city data, the steps of the soundscape prediction model training method for fusing multi-source city data described above are implemented.

[0037] In a fifth aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which is stored a soundscape prediction model training program for fusing multi-source city data. When the soundscape prediction model training program for fusing multi-source city data is executed by a processor, the steps of the soundscape prediction model training method for fusing multi-source city data described above are implemented.

[0038] Beneficial effects: The present invention performs statistical analysis on environmental data to obtain features related to the soundscape, that is, to obtain soundscape training features, and trains a neural network model based on the soundscape training features to obtain a soundscape prediction model. Environmental data in a geographic area to be predicted where soundscape prediction is required is collected, and a soundscape prediction model is applied to the environmental data, and the prediction model outputs the soundscape information of the geographic area to be predicted. The environmental data of the present invention is existing data, and there is no need to collect it on-site, that is, the soundscape information can be predicted by using existing environmental data, thereby expanding the application scenarios of the soundscape prediction method of the present invention. Moreover, the present invention uses a soundscape prediction model to predict soundscape information, and can comprehensively and accurately predict the soundscape. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the overall flow chart of the present invention;

[0040] Figure 2 A schematic diagram of model training in an embodiment of the present invention;

[0041] Figure 3 is a simulation diagram of sound intensity prediction in an embodiment of the present invention;

[0042] Figure 4 A structural diagram of a soundscape prediction model training device integrating multi-source urban data provided by the present invention;

[0043] Figure 5 This is a block diagram of the internal structure principle of the terminal device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following is a clear and complete description of the technical solution of the present invention in combination with the embodiments and the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Research has found that soundscape is the sound environment perceived, experienced or understood by an individual or group. Soundscape covers all sounds in the environment. Accurately knowing the soundscape in the environment can provide technical support for improving the quality of urban soundscape and protecting the health of residents. Existing technologies often use sound measuring instruments to collect soundscapes in the environment, but sound measuring instruments have geographical limitations, that is, sound measuring instruments cannot be set up in all geographical areas to measure the soundscape of the area.

[0046] In order to solve the above technical problems, the present invention provides a soundscape prediction model training method and prediction method that integrates multi-source urban data, which solves the problem that the existing technology for measuring soundscape has limitations. In specific implementation, firstly, the environmental data stored in the training geographical area is obtained, and the environmental data is statistically analyzed to obtain the soundscape training features of the training geographical area; then, a neural network model is applied to the soundscape training features to obtain the soundscape training labels for the training geographical area output by the neural network model; finally, the measured soundscape labels of the training geographical area are obtained, and the loss function of the neural network model is calculated based on the measured soundscape labels and the soundscape training labels, and the neural network model is trained based on the loss function to obtain a soundscape prediction model.

[0047] The soundscape prediction model training method for integrating multi-source urban data in this embodiment can be applied to a terminal device, which can be a terminal product with a data collection function, such as a computer. Figure 1 As shown in , the soundscape prediction model training method integrating multi-source urban data specifically includes the following steps:

[0048] S100, obtaining environmental data stored in a training geographical area, and performing statistical analysis on the environmental data to obtain soundscape training features of the training geographical area.

[0049] S200, applying a neural network model to the soundscape training features to obtain a soundscape training label for the training geographical area output by the neural network model.

[0050] S300, obtaining the measured soundscape labels of the training geographical area, calculating the loss function of the neural network model according to the measured soundscape labels and the soundscape training labels, and training the neural network model according to the loss function to obtain a soundscape prediction model.

[0051] In one embodiment, step S100 includes the following specific steps S101 and S102:

[0052] S101, obtaining point of interest data information, road data, building data, street view images and green data from the environmental data stored in the training geographic area.

[0053] Figure 2 The point of interest data information, road data, building data, street view images and green data in the network constitute the multi-source urban data.

[0054] Obtain POI data information through Amap Open Platform and Baidu Map Open Platform. POI data information includes Figure 2 The point of interest POI data and area of ​​interest AOI data in the data, where the point of interest POI includes shopping malls, schools, and stations, and the area of ​​interest AOI data is used to characterize the density or area of ​​the point of interest POI.

[0055] Obtain OSM road and building data by downloading from the Open Street Map website. Obtain street view images through the Baidu Map Panorama Platform. Extract urban green space areas by classifying the Sentinel-2 high-resolution remote sensing images.

[0056] S102, statistically analyzing the interest point data information, the road data, the building data, the street view image and the green data to obtain the number of interest points, interest point density, road density, the shortest distance between the center position and the road, building density, building volume ratio, green space coverage, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image in the training geographical area. The number of interest points, the interest point density, the road density, the shortest distance, the building density, the building volume ratio, the green space coverage, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image are used as the soundscape training features of the training geographical area.

[0057] Number of points of interest

[0058]

[0059] P i is the number of points of interest in the i-th grid in the training geographic area, such as P i The number of schools or shopping malls in the training geographical area.

[0060] Point of interest density

[0061]

[0062] A i is the area of ​​the i-th interest point in the training geographic area, and S is the total area of ​​the training geographic area.

[0063] Road density D R :

[0064]

[0065] L is the total length of roads within the training geographic area.

[0066] Shortest distance

[0067]

[0068] r i,j For the jth road of level i, Dist(c,r i,j ) is the center point c to all r in the training geographic area i,j distance.

[0069] Building density D B :

[0070]

[0071] n is the total number of buildings in the training geographic area, A b,i is the base area of ​​the ith building.

[0072] Building volume ratio FAR:

[0073]

[0074] n is the total number of buildings in the training geographic area, A f,i is the surface area of ​​the i-th building.

[0075] Green space coverage FVC:

[0076]

[0077] NVDI soil is the NVDI value of the pure bare soil covered pixel, NVDI veg is the NVDI value of the pixel with pure vegetation cover.

[0078] The nearest green space is Dist G :

[0079]

[0080] g i is the i-th green space outside the training geographic area, Dist(c,g i ) are the center points c and g in the training geographic area i The distance between.

[0081] Recent green area A g :

[0082]

[0083] The distance between the center point c in the training geographic area and the nearest green space distance Dist G The corresponding green area.

[0084] Proportion of each element of street view image

[0085]

[0086] is the area value of the i-th type pixel, that is, the area occupied by the i-th type pixel on the street view image, T n is the total area of ​​the street view image, that is, T n is the sum of all pixels in the Street View imagery.

[0087] In one embodiment, the neural network model in S200 is as follows: Figure 2 As shown, it includes a feature screening layer, a shared hidden layer, a sound intensity hidden layer, and a sound source hidden layer, wherein the feature screening layer is cascaded with the shared hidden layer, the sound intensity hidden layer, and the sound source hidden layer, respectively, and the shared hidden layer is cascaded with the sound intensity hidden layer and the sound source hidden layer, respectively. Figure 2 The feature engineering layer in step S200 includes the following specific steps S201 and S202:

[0088] S201, inputting the soundscape training features into a feature screening layer, obtaining shared features, sound intensity features, and sound source features screened by the feature screening layer from the soundscape training features, wherein the shared features are soundscape features related to both sound intensity and sound source, the sound intensity features are soundscape features only related to sound intensity, and the sound source features are soundscape features only related to sound source.

[0089] The feature screening layer screens and classifies the soundscape features in step S102, and divides them into sound source features, sound intensity features, and shared features. The shared features include the road density D in step S102. R , the shortest distance to the secondary road Green space coverage FVC, traffic station density, and landscape density; sound intensity characteristics include the nearest green space area A g and the proportion of each element of street view image The sound source characteristics include the shortest distance to the primary road Shortest distance to third-level roads Shortest distance to Class 4 road Building density D B , building volume ratio FAR, point of interest density Number of points of interest

[0090] S202, after inputting the shared features into the shared hidden layer, inputting the sound intensity features into the sound intensity hidden layer, and inputting the sound source features into the sound source hidden layer, to obtain a sound intensity training label output by the sound intensity hidden layer and a sound source training label output by the sound source hidden layer, and using the sound intensity training label and the sound source training label as soundscape training labels.

[0091] That is, the shared features are input first, and when the shared features reach the shared hidden layer, the sound source features and the sound intensity features are input into the sound source hidden layer and the sound intensity hidden layer respectively, so that the above three features are input into the neural network model in batches.

[0092] The three hidden layers, shared hidden layer, sound intensity hidden layer, and sound source hidden layer, are all composed of convolutional layers. Among them, the shared hidden layer receives the features after feature selection and screening. For the features required by both sound source prediction and sound intensity prediction, a shared hidden layer (fully connected layer) is designed to extract the features of the similar parts of the two tasks. The sound intensity hidden layer and the sound source hidden layer belong to the unique feature layer. For the unique features of the two tasks, the unique feature layer adopts the approach of designing two hidden layers (fully connected layers) after the shared hidden layer for the unique features of the two tasks, so as to train the unique features of the two tasks and learn the different features of the two tasks.

[0093] Due to the small size of training samples and the risk of overfitting caused by multiple fully connected layers, a Drop-out layer is added after the sound source hidden layer and the sound intensity hidden layer respectively to enhance the generalization of the model. A softmax activation function is set on the output layer cascaded with the sound source hidden layer, and a linear activation function is set on the output layer cascaded with the sound intensity hidden layer.

[0094] The sound intensity training labels output by the neural network model are as follows: Figure 3 As shown, the neural network model predicts the sound intensity and uses the sound intensity as the sound intensity training label. At the same time, the neural network model also outputs the sound source training label, that is, the sound source category is represented by a digital label. The sound source categories are shown in Table 1.

[0095] Table 1

[0096] Sound source category Detailed categories Natural Sound Birds singing, waves crashing, insects chirping, wind blowing through leaves Human Voice Conversation, activity, broadcast, and sound Traffic Sound Motor vehicle driving sound, horn sound Mechanical sound Construction noise, loading and unloading noise

[0097] In one embodiment, step S300 includes the following specific steps S301 to S306:

[0098] S301, obtaining a sound intensity measured label and a sound source measured label in the sound scene measured label.

[0099] The measured sound intensity labels and sound source labels are the sound intensity and sound source categories of the training geographical area mentioned above. The sound intensity and sound source categories are obtained in the following way:

[0100] Point the sound level meter toward the sound source, record the latitude and longitude of the sampling point, the number of the sampling point and the location, and use the application on the mobile device where the sound level meter is located to record and present the data synchronously.

[0101] The measured sound intensity label is the average sound intensity of each sampling point.

[0102] S302, calculating the sound intensity training label and the sound intensity measured label y i The sound intensity mean square error RMSE is constructed, and the sound intensity mean square error is used as the sound intensity loss function.

[0103]

[0104] Where n is the total number of sampling points.

[0105] S303, calculating each of the sound source training labels p i The logarithmic functions log(p i ).

[0106] S304, multiplying each of the logarithmic functions by the corresponding sound source measured labels y i , and obtain the intermediate results y i log(p i ).

[0107] S305, performing weighted calculation on each of the intermediate results to obtain a sound source loss function H(y,p):

[0108]

[0109] S306: According to the sound intensity loss function and the sound source loss function, adjust the parameters of the neural network model until the sound intensity loss function and the sound source loss function are both less than set values, so as to obtain a soundscape prediction model.

[0110] That is, in model training, the loss functions of the two tasks of sound intensity prediction and sound source prediction are separated, and their minimum values ​​are deduced respectively using back propagation. When both loss functions reach the minimum value, the model training is completed.

[0111] In one embodiment, the above model training is completed, and the soundscape prediction model obtained by training is evaluated. Specifically, the accuracy Acc, precision PREC, recall REC, and F1 function are used to evaluate the model sound source prediction task, and R 2 Function to evaluate the model for the sound intensity prediction task.

[0112]

[0113]

[0114]

[0115]

[0116] Where TP is the number of sound source prediction labels output by the model that are consistent with the sound source measured labels, FP is the number of sound source prediction labels output by the model that are not equal to the sound source measured labels, FN is the number of sound source prediction labels output by the model that are equal to the non-sound source measured labels, and TN is the number of non-sound source prediction labels output by the model that are equal to the sound source measured labels.

[0117]

[0118] In the formula, is the mean of the measured labels of sound intensity.

[0119] Use accuracy Acc, precision PREC, recall REC, F1 function and R 2 Function and RMSE function to construct the comprehensive indicator ε:

[0120] ε=w1·ACC norm +w2·PEC norm +w2 REC norm +w3·F1 norm +w4·RMSE norm +w5·R 2 norm

[0121] In the formula, ACC norm is the normalized value of Acc, PEC norm is the normalization of the precision PREC, REC norm is the normalization of REC, F1 norm is the normalization of F1, RMSE normis the normalized RMSE, R 2 norm For R 2 Normalization of .

[0122] The normalized formula is as follows:

[0123]

[0124] X is any one of the above six indicators. min is the minimum value of any indicator, X max is the maximum value of any indicator.

[0125] By weighing the contributions of different metrics, the comprehensive performance metric ε provides a more comprehensive performance evaluation in soundscape prediction.

[0126] In summary, the soundscape prediction module constructed by the present invention can simultaneously handle the sound source classification and sound intensity prediction tasks, that is, simultaneously predict the sound source category and sound intensity. The model is used to predict the urban soundscape in unknown areas, achieving a comprehensive and accurate prediction of the urban soundscape.

[0127] The present invention makes full use of information of different dimensions of the city, including data on buildings, road networks, street scenes, etc., to make up for the shortcomings of a single data source, thereby capturing the characteristic information of urban soundscape more comprehensively and making more efficient and accurate predictions of urban soundscape.

[0128] The present invention uses the characteristic connection between sound source category and sound intensity to establish a shared feature layer and a unique feature layer, thereby achieving the task of simultaneously predicting the sound source category and sound intensity. After training with a small sample, the model can predict a large range of urban soundscapes only through urban soundscape features without the need for sound data, thereby improving the overall efficiency of soundscape prediction.

[0129] This embodiment also provides a soundscape prediction model training device that integrates multi-source urban data. Figure 4 As shown, the training device includes the following components:

[0130] The feature statistics module 01 is used to obtain the environmental data stored in the training geographical area and perform statistical analysis on the environmental data to obtain the soundscape training features of the training geographical area;

[0131] Prediction module 02, used for applying a neural network model to the soundscape training features to obtain a soundscape training label for the training geographical area output by the neural network model;

[0132] The training module 03 is used to obtain the measured soundscape labels of the training geographical area, calculate the loss function of the neural network model based on the measured soundscape labels and the soundscape training labels, and train the neural network model based on the loss function to obtain a soundscape prediction model.

[0133] Based on the above embodiments, the present invention further provides a terminal device, whose principle block diagram can be shown as follows: Figure 5 As shown. The terminal device includes a processor, a memory, a network interface, and a display screen connected via a system bus. Among them, the processor of the terminal device is used to provide computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a soundscape prediction model training method that integrates multi-source urban data is implemented. The display screen of the terminal device can be a liquid crystal display screen or an electronic ink display screen.

[0134] Those skilled in the art will understand that Figure 5 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal device to which the scheme of the present invention is applied. The specific terminal device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0135] In one embodiment, a terminal device is provided. The terminal device includes a memory, a processor, and a soundscape prediction model training program for integrating multi-source city data stored in the memory and executable on the processor. When the processor executes the soundscape prediction model training program for integrating multi-source city data, the following operation instructions are implemented:

[0136] Acquiring environmental data stored in a training geographical area, and performing statistical analysis on the environmental data to obtain soundscape training features of the training geographical area;

[0137] Applying a neural network model to the soundscape training features to obtain a soundscape training label for the training geographic area output by the neural network model;

[0138] Acquire the measured soundscape labels of the training geographical area, calculate the loss function of the neural network model based on the measured soundscape labels and the soundscape training labels, and train the neural network model based on the loss function to obtain a soundscape prediction model.

[0139] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A soundscape prediction model training method integrating multi-source urban data, characterized in that: include: Acquiring environmental data stored in a training geographical area, and performing statistical analysis on the environmental data to obtain soundscape training features of the training geographical area; Applying a neural network model to the soundscape training features to obtain a soundscape training label for the training geographic area output by the neural network model; Acquire the measured soundscape labels of the training geographical area, calculate the loss function of the neural network model based on the measured soundscape labels and the soundscape training labels, and train the neural network model based on the loss function to obtain a soundscape prediction model, wherein the soundscape prediction model simultaneously predicts the sound source category and the sound intensity; The step of acquiring the environmental data stored in the training geographical area and performing statistical analysis on the environmental data to obtain the soundscape training features of the training geographical area includes: Obtaining point of interest data information, road data, building data, street view images and green data from the environmental data stored in the training geographic area; Performing statistical analysis on the interest point data information, the road data, the building data, the street view image and the greening data to obtain the number of interest points, interest point density, road density, the shortest distance between the center position and the road, building density, building volume ratio, green space coverage, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image in the training geographical area; The number of points of interest, the density of points of interest, the road density, the shortest distance, the building density, the building volume ratio, the green space coverage ratio, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image are used as the soundscape training features of the training geographical area; The neural network model includes a feature screening layer, a shared hidden layer, a sound intensity hidden layer, and a sound source hidden layer, wherein the feature screening layer is cascaded with the shared hidden layer, the sound intensity hidden layer, and the sound source hidden layer, respectively, and the shared hidden layer is cascaded with the sound intensity hidden layer and the sound source hidden layer, respectively. The neural network model is applied to the soundscape training features to obtain the soundscape training label for the training geographical area output by the neural network model, including: Inputting the soundscape training features into a feature screening layer, obtaining shared features, sound intensity features, and sound source features screened by the feature screening layer from the soundscape training features, wherein the shared features are soundscape features related to both sound intensity and sound source, the sound intensity features are soundscape features only related to sound intensity, and the sound source features are soundscape features only related to sound source; After the shared features are input into the shared hidden layer, the sound intensity features are input into the sound intensity hidden layer, and the sound source features are input into the sound source hidden layer, so as to obtain the sound intensity training labels output by the sound intensity hidden layer and the sound source training labels output by the sound source hidden layer, and use the sound intensity training labels and the sound source training labels as soundscape training labels.

2. The soundscape prediction model training method integrating multi-source urban data according to claim 1, characterized in that: The step of obtaining the measured soundscape labels of the training geographical area, calculating the loss function of the neural network model based on the measured soundscape labels and the soundscape training labels, and training the neural network model based on the loss function to obtain a soundscape prediction model comprises: Obtaining a sound intensity measured label and a sound source measured label from the sound scene measured label; Calculating a sound intensity loss function according to the difference between the sound intensity training label and the sound intensity measured label; Calculating a sound source loss function based on the sound source training label and the sound source measured label; According to the sound intensity loss function and the sound source loss function, the parameters of the neural network model are adjusted until the sound intensity loss function and the sound source loss function are both less than set values, so as to obtain a soundscape prediction model.

3. The soundscape prediction model training method for integrating multi-source urban data according to claim 2, characterized in that: The calculating of the sound intensity loss function based on the difference between the sound intensity training label and the sound intensity measured label includes: The sound intensity mean square error formed by the sound intensity training label and the sound intensity measured label is calculated, and the sound intensity mean square error is used as the sound intensity loss function.

4. The soundscape prediction model training method for integrating multi-source urban data according to claim 2, characterized in that: The calculating the sound source loss function according to the sound source training label and the sound source measured label includes: Calculating each logarithmic function of each of the sound source training labels; Multiplying each of the logarithmic functions by each corresponding sound source measured label to obtain each intermediate result; A weighted calculation is performed on each of the intermediate results to obtain a sound source loss function.

5. A soundscape prediction method, using the soundscape prediction model according to any one of claims 1 to 4, characterized in that: The soundscape prediction method comprises: Acquiring environmental data in a geographic area to be measured, and performing statistical analysis on the environmental data to obtain soundscape characteristics in the geographic area to be measured; A soundscape prediction model is applied to the soundscape feature to obtain a sound source prediction label and a sound intensity prediction label output by the soundscape prediction model.

6. A soundscape prediction model training device integrating multi-source urban data, characterized in that: The training device comprises the following components: A feature statistics module is used to obtain the environmental data stored in the training geographical area and perform statistical analysis on the environmental data to obtain the soundscape training features of the training geographical area; A prediction module, configured to apply a neural network model to the soundscape training features to obtain a soundscape training label for the training geographical area output by the neural network model; A training module is used to obtain the measured soundscape labels of the training geographical area, calculate the loss function of the neural network model according to the measured soundscape labels and the soundscape training labels, and train the neural network model according to the loss function to obtain a soundscape prediction model, wherein the soundscape prediction model simultaneously predicts the sound source category and the sound intensity; The step of acquiring the environmental data stored in the training geographical area and performing statistical analysis on the environmental data to obtain the soundscape training features of the training geographical area includes: Obtaining point of interest data information, road data, building data, street view images and green data from the environmental data stored in the training geographic area; Performing statistical analysis on the interest point data information, the road data, the building data, the street view image and the greening data to obtain the number of interest points, interest point density, road density, the shortest distance between the center position and the road, building density, building volume ratio, green space coverage, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image in the training geographical area; The number of points of interest, the density of points of interest, the road density, the shortest distance, the building density, the building volume ratio, the green space coverage ratio, the nearest green space distance, the nearest green space area, and the proportion of each element of the street view image are used as the soundscape training features of the training geographical area; The neural network model includes a feature screening layer, a shared hidden layer, a sound intensity hidden layer, and a sound source hidden layer, wherein the feature screening layer is cascaded with the shared hidden layer, the sound intensity hidden layer, and the sound source hidden layer, respectively, and the shared hidden layer is cascaded with the sound intensity hidden layer and the sound source hidden layer, respectively. The neural network model is applied to the soundscape training features to obtain the soundscape training label for the training geographical area output by the neural network model, including: Inputting the soundscape training features into a feature screening layer, obtaining shared features, sound intensity features, and sound source features screened by the feature screening layer from the soundscape training features, wherein the shared features are soundscape features related to both sound intensity and sound source, the sound intensity features are soundscape features only related to sound intensity, and the sound source features are soundscape features only related to sound source; After the shared features are input into the shared hidden layer, the sound intensity features are input into the sound intensity hidden layer, and the sound source features are input into the sound source hidden layer, so as to obtain the sound intensity training labels output by the sound intensity hidden layer and the sound source training labels output by the sound source hidden layer, and use the sound intensity training labels and the sound source training labels as soundscape training labels.

7. A terminal device, characterized in that: The terminal device includes a memory, a processor, and a soundscape prediction model training program for fusing multi-source city data stored in the memory and executable on the processor. When the processor executes the soundscape prediction model training program for fusing multi-source city data, the steps of the soundscape prediction model training method for fusing multi-source city data according to any one of claims 1 to 4 are implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a soundscape prediction model training program that integrates multi-source city data. When the soundscape prediction model training program that integrates multi-source city data is executed by the processor, the steps of the soundscape prediction model training method that integrates multi-source city data according to any one of claims 1 to 4 are implemented.