A smart city noise source tracing monitoring system

Abstract

The invention discloses an intelligent city noise source tracing monitoring system, which relates to the field of speech sound processing technology and aims to solve the technical problems of insufficient precision and low efficiency in the current positioning and identification of speech noise in high-rise buildings. The system comprises a building layout diagram generating unit for generating a building layout diagram in combination with building data and a sensor array deployment unit arranged at various places of the building for collecting speech noise data. The sensor array deployment unit is electrically connected to a speech noise data acquisition unit for improving data quality, and the building layout diagram generating unit and the speech noise data acquisition unit are electrically connected to a speech noise distribution dynamic image generating module. The invention improves the precision and efficiency of positioning and identifying speech noise in high-rise buildings at night by monitoring speech noise in high-rise buildings and performing speech noise judgment and speech noise source positioning in conjunction with a day-night signal-to-noise ratio threshold.

Description

Technical Field

[0001] The present invention relates to the technical field of speech sound processing, and more specifically to an intelligent city noise source tracing monitoring system. Background Art

[0002] At present, with the acceleration of urbanization, the number of high-rise buildings is increasing, and the resulting voice noise problem is becoming increasingly prominent. Voice noise has a serious impact on residents' lives due to its strong penetration, long propagation distance and high flexibility.

[0003] At present, the voice noise acting on high-rise buildings is mainly divided into two sources: one is the voice noise generated by the flow of people outside the building; the other is the voice noise generated by the activities and communication of people inside the building;

[0004] In the existing technology, it is relatively convenient to control the external voice noise of buildings. It is mainly concentrated during the day. The external voice noise is significantly reduced at night, and it is easy to trace the source and locate. The noise can be significantly reduced by arranging soundproof doors and windows.

[0005] The speech noise concentrated in high-rise buildings at night is easier to be detected by human ears due to the different signal-to-noise ratios during the day and at night. In addition, the speech noise source in high-rise buildings has the characteristics of high flexibility, intermittency, and high propagation. Therefore, the positioning and identification of speech noise in high-rise buildings with dense residents has problems of insufficient accuracy and low efficiency. Therefore, it is particularly important to develop an efficient and accurate speech noise source tracing monitoring system. In view of this, we propose a smart city noise source tracing monitoring system. Summary of the invention

[0006] The purpose of the present invention is to provide an intelligent city noise source tracing monitoring system to solve the technical problems of insufficient accuracy and low efficiency in the positioning and identification of speech noise in current high-rise buildings.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a smart city noise source tracing monitoring system, comprising a building layout diagram generation unit for generating a building layout diagram in combination with building data, and a sensor array deployment unit arranged at various locations of the building for collecting voice noise data;

[0008] The sensor array deployment unit is electrically connected to a voice noise data acquisition unit for improving data quality, and the building layout generation unit and the voice noise data acquisition unit are electrically connected to a voice noise distribution dynamic map generation module for generating a floor voice noise distribution dynamic map to distinguish the source and type of voice noise;

[0009] The speech noise distribution dynamic image generation module is electrically connected to a speech noise source localization unit for generating a sound field model and localizing the speech noise source in combination with a day and night signal-to-noise ratio threshold, and the speech noise source localization unit is electrically connected to a deep learning unit for optimizing the sound field model in real time, and the deep learning unit is electrically connected to a user optimization unit for performing system optimization based on users.

[0010] The present invention collects external and internal voice noises respectively through an external wall array sensor arranged at the external wall of a high-rise building and an internal wall array sensor arranged at the internal wall, and improves the voice noise data quality by relying on a voice noise data acquisition unit, and then generates a building layout map by combining the building data through a building layout map generation unit, so as to generate a floor voice noise distribution dynamic map by relying on the internal and external voice noise data of the high-rise building and the building layout map through a voice noise distribution dynamic map generation module, realize the distinction of the source and type of the voice noise, and generate a sound field model by a voice noise source positioning unit to cooperate with the day and night signal-to-noise ratio threshold to perform voice noise judgment and voice noise source positioning, thereby realizing the judgment and sound source positioning of the voice noise in the high-rise building, and improving the accuracy and efficiency of positioning and identifying the voice noise in the high-rise building at night.

[0011] Preferably, the building layout drawing generation unit includes a building layout data entry unit for classifying and entering building layout data and a layout data modeling module for generating a building layout drawing based on the building layout data, and the building layout data entry unit is electrically connected to the layout data modeling module.

[0012] Preferably, the building layout data entry unit includes a wall structure type entry module for entering data on the structure types of each wall of the building, a wall material entry module for entering data on each wall material, and a wall thickness entry module for entering data on each wall thickness. The wall structure type entry module, the wall material entry module, and the wall thickness entry module are all electrically connected to the speech noise distribution animation generation module.

[0013] Preferably, the sensor array deployment unit includes an exterior wall array sensor for collecting external voice noise data at the exterior wall and an interior wall array sensor for collecting external voice noise data at the interior wall, the exterior wall array sensor is composed of several microphone arrays, and the interior wall array sensor is composed of several directional microphone arrays.

[0014] Preferably, the speech noise data acquisition unit includes a speech noise data acquisition module for acquiring speech noise data inside and outside the building and a speech noise data processing module for processing the speech noise data, the speech noise data acquisition module is electrically connected to the speech noise data processing module, and the speech noise data acquisition module is electrically connected to the exterior wall array sensor and the interior wall array sensor.

[0015] Preferably, the speech noise source locating unit includes a sound field model construction module that generates an artificial sound field model, an environmental sound field model and an equipment sound field model based on the characteristics of the speech noise data inside and outside the building, and a sound field model analysis module that analyzes the speech noise environment standard and the speech noise source location in combination with the day and night signal-to-noise ratio threshold. The sound field model construction module is connected to the sound field model analysis module.

[0016] Preferably, the deep learning unit includes a deep learning model designed and trained for speech noise characteristics and a model calibration module for simulating the speech noise propagation path and attenuation characteristics in combination with the speech noise data characteristics inside and outside the building, and the deep learning model is connected to the model calibration module.

[0017] Preferably, the user optimization unit includes a user interaction module for outputting a voice noise source tracing monitoring report to the user, a user feedback module for supporting the user to feedback abnormal conditions, a privacy protection module for encrypting voice noise data to protect user privacy, and a privacy compliance module for providing compliance support, the user interaction module is connected to the user feedback module, and the privacy compliance module is connected to the privacy protection module.

[0018] A noise source tracing monitoring method comprises the following steps:

[0019] S1, data entry and modeling;

[0020] S101, building data entry;

[0021] The staff will input the data of each wall structure type, each wall material data, and each wall thickness data for the monitored building, and generate a building layout data table through summarization;

[0022] S102, generating a building layout diagram;

[0023] Using BIM building information modeling technology and combining it with building layout data tables, a three-dimensional building layout diagram of the monitored building is generated to display the internal structure and external environment of the building, providing an accurate basis for voice noise propagation simulation;

[0024] S2, sensor array deployment;

[0025] Relying on the three-dimensional building layout diagram, according to the building layout and voice noise propagation characteristics, the layout position and quantity of the sensor array are calculated, and the layout of the exterior wall array sensor and the interior wall array sensor are completed to fully cover and capture the voice noise in the key areas;

[0026] S3, data preprocessing and feature extraction;

[0027] S301, data cleaning;

[0028] S301a, remove outliers: use Z-score test to identify and remove abnormal speech noise data, set the check value With threshold :

[0029] ;

[0030] when Threshold When , the observed value is considered to be an outlier. Otherwise, when Threshold When , the observed value is considered to be normal;

[0031] in, is the observed value;

[0032] is the mean;

[0033] is the standard deviation;

[0034] S301b, speech noise smoothing: applying sliding average or median filtering and other methods to smooth the original speech noise data to reduce random fluctuations;

[0035] S302, feature extraction;

[0036] S302a, time domain features: extracting statistical features such as mean, variance, kurtosis, and skewness of the speech noise signal;

[0037] S302b, frequency domain features: spectrum obtained by fast Fourier transform , the energy proportion of low frequency band is , calculation expression:

[0038] ;

[0039] in, and is the boundary frequency of the low frequency band;

[0040] and is the boundary frequency of the entire analysis frequency band;

[0041] S302c, time-frequency features: extracting the time-frequency features of the speech noise signal using short-time Fourier transform to capture the characteristics of speech noise changing over time;

[0042] S4, deep learning model optimization;

[0043] Design and train deep learning models, and use deep learning models and sound field models to locate and classify speech noise sources;

[0044] S5, Combination of sound field model and signal processing technology;

[0045] S501, sound field model calibration;

[0046] Use ray tracing to build an accurate sound field model, simulate the effects of different floors and wall materials on the propagation and attenuation of speech noise, and then calibrate the sound field model based on actual measurement data to ensure the accuracy and reliability of the model in different environments;

[0047] S502, signal enhancement and denoising;

[0048] S502a, signal processing: using spectral subtraction and Wiener filtering as signal processing techniques to remove background speech noise and improve the signal-to-noise ratio of the target speech noise signal;

[0049] Spectral Subtraction: Estimating the Speech Noise Spectrum , and then from the noisy signal spectrum Subtract it from the original signal to get the enhanced signal spectrum:

[0050] ;

[0051] in, is the oversubtraction factor;

[0052] Wiener filtering: In the frequency domain, the Wiener filter is estimated as:

[0053] ;

[0054] in, , are the power spectral densities of speech noise and signal, respectively;

[0055] S502b, interference suppression: using beamforming technology, through the phase difference of the signal received by the array sensor, the voice noise signal from a specific direction is enhanced and interference from other directions is suppressed;

[0056] S6, speech noise source localization algorithm;

[0057] S601, Positioning based on sound source imaging:

[0058] S601a, preliminary positioning: using the delay-sum beamforming in the sound source imaging technology, the signal received by the array sensor is converted into a sound source distribution image to achieve preliminary positioning of the speech noise source. , calculate the delay And apply, then sum:

[0059] ;

[0060] in, It is The signal of a microphone;

[0061] is the weight;

[0062] The sound wave travels from the source to the The delay of each microphone;

[0063] is the azimuth of the sound source;

[0064] S601b, positioning accuracy optimization: Combined with the sound field model, the sound source image is corrected and optimized to improve the positioning accuracy;

[0065] S7, disturbing voice noise recognition;

[0066] S701, setting of standards for nuisance to the public;

[0067] Set the threshold of speech noise nuisance according to local speech noise standards , and dynamically adjust the nuisance threshold by considering the changes in signal-to-noise ratio at different time periods during the day and night To adapt to different environments;

[0068] S702, determining disturbing voice noise;

[0069] S702a, speech noise recognition: compare the detected speech noise with the set nuisance threshold Compare and identify disturbing speech noise :

[0070] ;

[0071] in, is the equivalent continuous sound level;

[0072] when When , it indicates nuisance speech noise;

[0073] On the contrary, when When , it indicates non-nuisance speech noise;

[0074] S702b, voice noise source location: combining the disturbing voice noise determination and voice noise source location results, determining the specific source and location of the disturbing voice noise;

[0075] S8, user interaction and privacy protection;

[0076] The speech noise determination and speech noise source location results are transmitted to users through the user interaction module, so that users can view the speech noise distribution map, speech noise source location and type. During the monitoring process, the collected speech noise data is encrypted and stored in accordance with laws and regulations, and it is prohibited for anyone other than users and officials to access it illegally.

[0077] Preferably, the S4 deep learning model optimization further includes the following steps:

[0078] S401, model architecture selection;

[0079] A hybrid architecture of convolutional neural network combined with recurrent neural network is used to simultaneously capture the spatial and temporal features of speech noise signals. The attention mechanism is then introduced to enable the model to automatically focus on features that are more critical to speech noise classification and localization.

[0080] S402, loss function design;

[0081] S402a, joint optimization: design a multi-task loss function, consider the cross entropy loss of speech noise classification and the mean square error loss of speech noise source localization, and achieve joint optimization through weighted summation;

[0082] The multi-task loss function can be expressed as:

[0083] ;

[0084] Categorical cross entropy loss:

[0085] ;

[0086] Positioning mean square error loss:

[0087] ;

[0088] in, is the categorical cross entropy loss, is the positioning mean square error loss;

[0089] and is the weight coefficient;

[0090] is the true label, is the predicted probability;

[0091] is the number of categories;

[0092] It is the real location. is the predicted position;

[0093] is the number of anchor points;

[0094] S402b, Hard Example Mining: Introduce a hard example mining strategy to give higher weights to samples that are misclassified or inaccurately located, thereby improving the model's ability to learn hard-to-classify samples.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] 1. The present invention collects external and internal voice noises respectively through an outer wall array sensor arranged at the outer wall of a high-rise building and an inner wall array sensor arranged at the inner wall, and relies on a voice noise data acquisition unit to improve the quality of voice noise data, and then generates a building layout map through a building layout map generation unit in combination with the building data, thereby generating a floor voice noise distribution dynamic map based on the voice noise data inside and outside the high-rise building and the building layout map through a voice noise distribution dynamic map generation module, thereby achieving the distinction between the source and type of voice noise, and generating a sound field model through a voice noise source positioning unit to cooperate with the day and night signal-to-noise ratio threshold to perform voice noise judgment and voice noise source positioning, thereby achieving the judgment and sound source positioning of voice noise in high-rise buildings, and improving the accuracy and efficiency of positioning and identifying voice noise in high-rise buildings at night.

[0097] 2. The present invention adopts the classification and collection of internal and external speech noise, and relies on the sound field model to simulate the influence of different floors and different wall materials on the propagation and attenuation of speech noise, and then calibrates the sound field model in combination with the actual measurement data to ensure the accuracy and reliability of the model in different environments. The sound source distribution image generated by the speech noise source localization algorithm is used to achieve preliminary positioning of the speech noise, and then combined with the sound field model, the sound source image is corrected and optimized, which can significantly improve the positioning accuracy.

[0098] 3. The present invention can also capture the spatial and temporal characteristics of speech noise signals at the same time through the arrangement of deep learning models and the hybrid architecture of convolutional neural networks combined with recurrent neural networks. Combined with the introduced attention mechanism, it can automatically focus on the features that are more critical to speech noise classification and positioning, and at the same time supplement the multi-task loss function, while considering the cross entropy loss of speech noise classification and the mean square error loss of speech noise source positioning. Then, through difficult example mining, higher weights are given to samples that are misclassified or inaccurately positioned, thereby improving the model's learning ability for difficult samples, thereby further improving the accuracy and reliability of the sound field model, and further improving the recognition and positioning accuracy of subsequent nighttime speech noise in high-rise buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 is a schematic block diagram of the system of the present invention;

[0100] Figure 2 It is a system schematic block diagram of a building layout diagram generating unit in the present invention;

[0101] Figure 3 It is a system schematic block diagram of the building layout data entry unit in the present invention;

[0102] Figure 4 It is a system schematic block diagram of the sensor array deployment unit in the present invention;

[0103] Figure 5 It is a system schematic block diagram of the speech noise data acquisition unit in the present invention;

[0104] Figure 6 It is a system schematic block diagram of the speech noise source localization unit in the present invention;

[0105] Figure 7 It is a system schematic block diagram of the deep learning module in the present invention;

[0106] Figure 8 It is a system schematic block diagram of the user optimization unit in the present invention. DETAILED DESCRIPTION

[0107] Embodiment 1:

[0108] like Figures 1 to 8 As shown, the present invention relates to a smart city noise source tracing monitoring system, comprising a building layout diagram generation unit for generating a building layout diagram in combination with building data, and a sensor array deployment unit arranged at various locations of the building for collecting voice noise data;

[0109] The sensor array deployment unit is electrically connected to a voice noise data acquisition unit for improving data quality, and the building layout generation unit and the voice noise data acquisition unit are electrically connected to a voice noise distribution dynamic map generation module for generating a floor voice noise distribution dynamic map to distinguish the source and type of voice noise;

[0110] The speech noise distribution dynamic image generation module is electrically connected to a speech noise source localization unit for generating a sound field model and localizing the speech noise source in combination with a day and night signal-to-noise ratio threshold, and the speech noise source localization unit is electrically connected to a deep learning unit for optimizing the sound field model in real time, and the deep learning unit is electrically connected to a user optimization unit for performing system optimization based on users.

[0111] The present invention collects external and internal voice noises respectively through an external wall array sensor arranged at the external wall of a high-rise building and an internal wall array sensor arranged at the internal wall, and improves the voice noise data quality by relying on a voice noise data acquisition unit, and then generates a building layout map by combining the building data through a building layout map generation unit, so as to generate a floor voice noise distribution dynamic map by relying on the internal and external voice noise data of the high-rise building and the building layout map through a voice noise distribution dynamic map generation module, realize the distinction of the source and type of the voice noise, and generate a sound field model by a voice noise source positioning unit to cooperate with the day and night signal-to-noise ratio threshold to perform voice noise judgment and voice noise source positioning, thereby realizing the judgment and sound source positioning of the voice noise in the high-rise building, and improving the accuracy and efficiency of positioning and identifying the voice noise in the high-rise building at night.

[0112] In an embodiment of the present invention, the building layout drawing generation unit includes a building layout data entry unit for classifying and entering building layout data and a layout data modeling module for generating a building layout drawing based on the building layout data, and the building layout data entry unit is electrically connected to the layout data modeling module.

[0113] In an embodiment of the present invention, the building layout data entry unit includes a wall structure type entry module for entering data on the structure types of each wall of the building, a wall material entry module for entering data on each wall material, and a wall thickness entry module for entering data on each wall thickness. The wall structure type entry module, the wall material entry module, and the wall thickness entry module are all electrically connected to the speech noise distribution animation generation module.

[0114] In an embodiment of the present invention, the sensor array deployment unit includes an exterior wall array sensor for collecting external voice noise data at the exterior wall and an interior wall array sensor for collecting external voice noise data at the interior wall, the exterior wall array sensor is composed of a plurality of microphone arrays, and the interior wall array sensor is composed of a plurality of directional microphone arrays.

[0115] In an embodiment of the present invention, the speech noise data acquisition unit includes a speech noise data acquisition module for acquiring speech noise data inside and outside the building and a speech noise data processing module for processing the speech noise data. The speech noise data acquisition module is electrically connected to the speech noise data processing module, and the speech noise data acquisition module is electrically connected to the exterior wall array sensor and the interior wall array sensor.

[0116] In an embodiment of the present invention, the speech noise source locating unit includes a sound field model construction module that generates an artificial sound field model, an environmental sound field model and an equipment sound field model based on the characteristics of the speech noise data inside and outside the building, and a sound field model analysis module that analyzes the speech noise environment standard and the speech noise source location in combination with the day and night signal-to-noise ratio threshold. The sound field model construction module is connected to the sound field model analysis module.

[0117] In an embodiment of the present invention, the deep learning unit includes a deep learning model designed and trained for speech noise characteristics and a model calibration module for simulating the speech noise propagation path and attenuation characteristics in combination with the speech noise data characteristics inside and outside the building. The deep learning model is connected to the model calibration module.

[0118] In an embodiment of the present invention, the user optimization unit includes a user interaction module for outputting a voice noise source tracing monitoring report to the user, a user feedback module for supporting the user to feedback abnormal conditions, a privacy protection module for encrypting voice noise data to protect user privacy, and a privacy compliance module for providing compliance support. The user interaction module is connected to the user feedback module, and the privacy compliance module is connected to the privacy protection module.

[0119] Embodiment 2:

[0120] A noise source tracing monitoring method comprises the following steps:

[0121] S1, data entry and modeling;

[0122] S101, building data entry;

[0123] The staff will input the data of each wall structure type, each wall material data, and each wall thickness data for the monitored building, and generate a building layout data table through summarization;

[0124] S102, generating a building layout diagram;

[0125] Using BIM building information modeling technology and combining it with building layout data tables, a three-dimensional building layout diagram of the monitored building is generated to display the internal structure and external environment of the building, providing an accurate basis for voice noise propagation simulation;

[0126] S2, sensor array deployment;

[0127] Relying on the three-dimensional building layout diagram, according to the building layout and voice noise propagation characteristics, the layout position and quantity of the sensor array are calculated, and the layout of the exterior wall array sensor and the interior wall array sensor are completed to fully cover and capture the voice noise in the key areas;

[0128] S3, data preprocessing and feature extraction;

[0129] S301, data cleaning;

[0130] S301a, remove outliers: use Z-score test to identify and remove abnormal speech noise data, set the check value With threshold :

[0131] ;

[0132] when Threshold When , the observed value is considered to be an outlier. Otherwise, when Threshold When , the observed value is considered to be normal;

[0133] in, is the observed value;

[0134] is the mean;

[0135] is the standard deviation;

[0136] S301b, speech noise smoothing: applying sliding average or median filtering and other methods to smooth the original speech noise data to reduce random fluctuations;

[0137] S302, feature extraction;

[0138] S302a, time domain features: extracting statistical features such as mean, variance, kurtosis, and skewness of the speech noise signal;

[0139] S302b, frequency domain features: spectrum obtained by fast Fourier transform , the energy proportion of low frequency band is , calculation expression:

[0140] ;

[0141] in, and is the boundary frequency of the low frequency band;

[0142] and is the boundary frequency of the entire analysis frequency band;

[0143] S302c, time-frequency features: extracting the time-frequency features of the speech noise signal using short-time Fourier transform to capture the characteristics of speech noise changing over time;

[0144] S4, deep learning model optimization;

[0145] Design and train deep learning models, and use deep learning models and sound field models to locate and classify speech noise sources;

[0146] S5, Combination of sound field model and signal processing technology;

[0147] S501, sound field model calibration;

[0148] Use ray tracing to build an accurate sound field model, simulate the effects of different floors and wall materials on the propagation and attenuation of speech noise, and then calibrate the sound field model based on actual measurement data to ensure the accuracy and reliability of the model in different environments;

[0149] S502, signal enhancement and denoising;

[0150] S502a, signal processing: using spectral subtraction and Wiener filtering as signal processing techniques to remove background speech noise and improve the signal-to-noise ratio of the target speech noise signal;

[0151] Spectral Subtraction: Estimating the Speech Noise Spectrum , and then from the noisy signal spectrum Subtract it from the original signal to get the enhanced signal spectrum:

[0152] ;

[0153] in, is the oversubtraction factor;

[0154] Wiener filtering: In the frequency domain, the Wiener filter is estimated as:

[0155] ;

[0156] in, , are the power spectral densities of speech noise and signal, respectively;

[0157] S502b, interference suppression: using beamforming technology, through the phase difference of the signal received by the array sensor, the voice noise signal from a specific direction is enhanced and interference from other directions is suppressed;

[0158] S6, speech noise source localization algorithm;

[0159] S601, Positioning based on sound source imaging:

[0160] S601a, preliminary positioning: using the delay-sum beamforming in the sound source imaging technology, the signal received by the array sensor is converted into a sound source distribution image to achieve preliminary positioning of the speech noise source. , calculate the delay And apply, then sum:

[0161] ;

[0162] in, It is The signal of a microphone;

[0163] is the weight;

[0164] The sound wave travels from the source to the The delay of each microphone;

[0165] is the azimuth of the sound source;

[0166] S601b, positioning accuracy optimization: Combined with the sound field model, the sound source image is corrected and optimized to improve the positioning accuracy;

[0167] S7, disturbing voice noise recognition;

[0168] S701, setting of standards for nuisance to the public;

[0169] Set the threshold of speech noise nuisance according to local speech noise standards , and dynamically adjust the nuisance threshold by considering the changes in signal-to-noise ratio at different time periods during the day and night To adapt to different environments;

[0170] S702, determining disturbing voice noise;

[0171] S702a, speech noise recognition: compare the detected speech noise with the set nuisance threshold Compare and identify disturbing speech noise :

[0172] ;

[0173] in, is the equivalent continuous sound level;

[0174] when When , it indicates nuisance speech noise;

[0175] On the contrary, when When , it indicates non-nuisance speech noise;

[0176] S702b, voice noise source location: combining the disturbing voice noise determination and voice noise source location results, determining the specific source and location of the disturbing voice noise;

[0177] S8, user interaction and privacy protection;

[0178] The speech noise determination and speech noise source location results are transmitted to users through the user interaction module, so that users can view the speech noise distribution map, speech noise source location and type. During the monitoring process, the collected speech noise data is encrypted and stored in accordance with laws and regulations, and it is prohibited for anyone other than users and officials to access it illegally.

[0179] As another embodiment of the present invention, the S4 deep learning model optimization further includes the following steps:

[0180] S401, model architecture selection;

[0181] A hybrid architecture of convolutional neural network combined with recurrent neural network is used to simultaneously capture the spatial and temporal features of speech noise signals. The attention mechanism is then introduced to enable the model to automatically focus on features that are more critical to speech noise classification and localization.

[0182] S402, loss function design;

[0183] S402a, joint optimization: design a multi-task loss function, consider the cross entropy loss of speech noise classification and the mean square error loss of speech noise source localization, and achieve joint optimization through weighted summation;

[0184] The multi-task loss function can be expressed as:

[0185] ;

[0186] Categorical cross entropy loss:

[0187] ;

[0188] Positioning mean square error loss:

[0189] ;

[0190] in, is the categorical cross entropy loss, is the positioning mean square error loss;

[0191] and is the weight coefficient;

[0192] is the true label, is the predicted probability;

[0193] is the number of categories;

[0194] It is the real location. is the predicted position;

[0195] is the number of anchor points;

[0196] S402b, Hard Example Mining: Introduce a hard example mining strategy to give higher weights to samples that are misclassified or inaccurately located, thereby improving the model's ability to learn hard-to-classify samples.

[0197] Through the arrangement of deep learning models and the hybrid architecture of convolutional neural networks combined with recurrent neural networks, the spatial and temporal characteristics of speech noise signals can be captured at the same time. Combined with the introduced attention mechanism, it can automatically focus on the features that are more critical to speech noise classification and positioning. At the same time, a multi-task loss function is added, and the cross entropy loss of speech noise classification and the mean square error loss of speech noise source positioning are considered at the same time. Through difficult example mining, higher weights are given to samples that are misclassified or inaccurately positioned, which improves the model's learning ability for difficult samples, thereby further improving the accuracy and reliability of the sound field model, and further improving the recognition and positioning accuracy of subsequent nighttime speech noise in high-rise buildings.

[0198] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A smart city noise source tracing monitoring system, characterized in that: It includes a building layout diagram generation unit for generating a building layout diagram in combination with building data, and a sensor array deployment unit arranged at various locations of the building for collecting voice noise data; The sensor array deployment unit is electrically connected to a voice noise data acquisition unit for improving data quality, and the building layout generation unit and the voice noise data acquisition unit are electrically connected to a voice noise distribution dynamic map generation module for generating a floor voice noise distribution dynamic map to distinguish the source and type of voice noise; The speech noise distribution dynamic image generation module is electrically connected to a speech noise source positioning unit for generating a sound field model and positioning the speech noise source in combination with a day and night signal-to-noise ratio threshold, and the speech noise source positioning unit is electrically connected to a deep learning unit for optimizing the sound field model in real time, and the deep learning unit is electrically connected to a user optimization unit for optimizing the system based on the user; The voice noise source positioning unit includes a sound field model construction module that generates an artificial sound field model, an environmental sound field model, and an equipment sound field model based on the characteristics of the voice noise data inside and outside the building, and a sound field model analysis module that analyzes the voice noise environment standard and the voice noise source positioning in combination with the day and night signal-to-noise ratio threshold, and the sound field model construction module is connected to the sound field model analysis module; The deep learning unit includes a deep learning model designed and trained for speech noise characteristics and a model calibration module for simulating speech noise propagation path and attenuation characteristics in combination with speech noise data characteristics inside and outside the building, and the deep learning model is connected to the model calibration module; The user optimization unit includes a user interaction module for outputting a voice noise source tracing monitoring report to the user, a user feedback module for supporting the user to feedback abnormal conditions, a privacy protection module for encrypting voice noise data to protect user privacy, and a privacy compliance module for providing compliance support. The user interaction module is connected to the user feedback module, and the privacy compliance module is connected to the privacy protection module.

2. The intelligent city noise source tracing monitoring system according to claim 1 is characterized in that: The building layout diagram generating unit comprises a building layout data entry unit for classifying and entering building layout data and a layout data modeling module for generating a building layout diagram based on the building layout data. The building layout data entry unit is electrically connected to the layout data modeling module.

3. The intelligent city noise source tracing monitoring system according to claim 2 is characterized in that: The building layout data entry unit includes a wall structure type entry module for entering data on the structure types of each wall of the building, a wall material entry module for entering data on each wall material, and a wall thickness entry module for entering data on each wall thickness. The wall structure type entry module, the wall material entry module, and the wall thickness entry module are all electrically connected to the speech noise distribution dynamic image generation module.

4. The intelligent city noise source tracing monitoring system according to claim 1 is characterized in that: The sensor array deployment unit includes an exterior wall array sensor for collecting external voice noise data at the exterior wall and an interior wall array sensor for collecting external voice noise data at the interior wall. The exterior wall array sensor is composed of a plurality of microphone arrays, and the interior wall array sensor is composed of a plurality of directional microphone arrays.

5. The intelligent city noise source tracing monitoring system according to claim 4 is characterized in that: The speech noise data acquisition unit includes a speech noise data acquisition module for acquiring speech noise data inside and outside the building and a speech noise data processing module for processing the speech noise data. The speech noise data acquisition module is electrically connected to the speech noise data processing module, and the speech noise data acquisition module is electrically connected to the exterior wall array sensor and the interior wall array sensor.

6. A noise source tracing monitoring method, according to the smart city noise source tracing monitoring system according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1, data entry and modeling; S101, building data entry; The staff will input the data of each wall structure type, each wall material data, and each wall thickness data for the monitored building, and generate a building layout data table through summarization; S102, generating a building layout diagram; Using BIM building information modeling technology and combining it with building layout data tables, a three-dimensional building layout diagram of the monitored building is generated to display the internal structure and external environment of the building, providing an accurate basis for voice noise propagation simulation; S2, sensor array deployment; Relying on the three-dimensional building layout diagram, according to the building layout and voice noise propagation characteristics, the layout position and quantity of the sensor array are calculated, and the layout of the exterior wall array sensor and the interior wall array sensor are completed to fully cover and capture the voice noise in the key areas; S3, data preprocessing and feature extraction; S301, data cleaning; S301a, remove outliers: use Z-score test to identify and remove abnormal speech noise data, set the check value With threshold : ； when Threshold When , the observed value is considered to be an outlier. Otherwise, when Threshold When , the observed value is considered to be normal; in, is the observed value; is the mean; is the standard deviation; S301b, speech noise smoothing: applying a sliding average or median filtering method to smooth the original speech noise data to reduce random fluctuations; S302, feature extraction; S302a, time domain features: extracting the mean, variance, kurtosis, and skewness statistical features of the speech noise signal; S302b, frequency domain features: spectrum obtained by fast Fourier transform , the energy proportion of low frequency band is , calculation expression: ； in, and is the boundary frequency of the low frequency band; and is the boundary frequency of the entire analysis frequency band; S302c, time-frequency features: extracting the time-frequency features of the speech noise signal using short-time Fourier transform to capture the characteristics of speech noise changing over time; S4, deep learning model optimization; Design and train deep learning models, and use deep learning models and sound field models to locate and classify speech noise sources; S5, Combination of sound field model and signal processing technology; S501, sound field model calibration; Use ray tracing to build an accurate sound field model, simulate the effects of different floors and wall materials on the propagation and attenuation of speech noise, and then calibrate the sound field model based on actual measurement data to ensure the accuracy and reliability of the model in different environments; S502, signal enhancement and denoising; S502a, signal processing: using spectral subtraction and Wiener filtering as signal processing techniques to remove background speech noise and improve the signal-to-noise ratio of the target speech noise signal; Spectral Subtraction: Estimating the Speech Noise Spectrum , and then from the noisy signal spectrum Subtract it from the original signal to get the enhanced signal spectrum: ； in, is the oversubtraction factor; Wiener filtering: In the frequency domain, the Wiener filter is estimated as: ； in, , are the power spectral densities of speech noise and signal, respectively; S502b, interference suppression: using beamforming technology, through the phase difference of the signal received by the array sensor, the voice noise signal from a specific direction is enhanced and interference from other directions is suppressed; S6, speech noise source localization algorithm; S601, Positioning based on sound source imaging: S601a, preliminary positioning: using the delay-sum beamforming in the sound source imaging technology, the signal received by the array sensor is converted into a sound source distribution image to achieve preliminary positioning of the speech noise source. , calculate the delay And apply, then sum: ； in, It is The signal of a microphone; is the weight; The sound wave travels from the source to the The delay of each microphone; is the azimuth of the sound source; S601b, positioning accuracy optimization: Combined with the sound field model, the sound source image is corrected and optimized to improve the positioning accuracy; S7, disturbing voice noise recognition; S701, setting of standards for nuisance to the public; Set the threshold of speech noise nuisance according to local speech noise standards , and dynamically adjust the nuisance threshold by considering the changes in signal-to-noise ratio at different time periods during the day and night To adapt to different environments; S702, determining disturbing voice noise; S702a, speech noise recognition: compare the detected speech noise with the set nuisance threshold Compare and identify disturbing speech noise : ； in, is the equivalent continuous sound level; when When , it indicates nuisance speech noise; On the contrary, when When , it indicates non-nuisance speech noise; S702b, voice noise source location: combining the disturbing voice noise determination and voice noise source location results, determining the specific source and location of the disturbing voice noise; S8, user interaction and privacy protection; The speech noise determination and speech noise source location results are transmitted to users through the user interaction module, so that users can view the speech noise distribution map, speech noise source location and type. During the monitoring process, the collected speech noise data is encrypted and stored in accordance with laws and regulations, and it is prohibited for anyone other than users and officials to access it illegally.

7. A noise source tracing monitoring method according to claim 6, characterized in that: The S4 deep learning model optimization also includes the following steps: S401, model architecture selection; A hybrid architecture of convolutional neural network combined with recurrent neural network is used to simultaneously capture the spatial and temporal features of speech noise signals. The attention mechanism is then introduced to enable the model to automatically focus on features that are more critical to speech noise classification and localization. S402, loss function design; S402a, joint optimization: design a multi-task loss function, consider the cross entropy loss of speech noise classification and the mean square error loss of speech noise source localization, and achieve joint optimization through weighted summation; The multi-task loss function is expressed as: ； Categorical Cross Entropy Loss: ； Positioning mean square error loss: ； in, is the categorical cross entropy loss, is the positioning mean square error loss; and is the weight coefficient; is the true label, is the predicted probability; is the number of categories; It is the real location. is the predicted position; is the number of anchor points; S402b, Hard Example Mining: Introduce a hard example mining strategy to give higher weights to samples that are misclassified or inaccurately located, thereby improving the model's ability to learn hard-to-classify samples.

Images