Multi-city noise automatic monitoring data analysis method and system based on time period classification

Abstract

The invention discloses a multi-city noise automatic monitoring data analysis method and system based on time period classification, and particularly relates to the field of city noise automatic monitoring, which comprises the steps of collecting sound pressure data, and performing time slot slicing, thereby separating city generality and specific features through feature weighted classification and feature decoupling migration, and obtaining a multi-city noise automatic monitoring data analysis result. And dynamic analysis and evolution of multi-city noise monitoring data are realized. According to the multi-city noise automatic monitoring data analysis method and system based on time period classification, a correlation framework of time period features and noise modes is constructed, the dynamic evolution rule of a noise time domain and a noise frequency domain is accurately captured, and the regional misjudgment rate of cross-time-period monitoring is effectively reduced; through a dual adaptive mechanism, in combination with feature decoupling migration and incremental learning, cross-time and cross-city feature migration and dynamic evolution are realized, and occurrence of time-period misinformation is significantly reduced; by forming a self-optimization closed loop, passive correction is converted into active learning, and the pain point caused by frequent manual intervention in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of urban noise automatic monitoring. More specifically, the present invention relates to a multi-city noise automatic monitoring data analysis method and system based on time period classification. Background Art

[0002] With the acceleration of the intelligentization process of urban environmental monitoring, the real-time identification and dynamic control requirements of large-scale noise pollution sources in multiple cities are particularly important. Traditional noise monitoring consists of distributed sound pressure sensors, cellular network transmission modules, and cloud data processing platforms. Soundprint data is collected through a periodic polling mechanism, and after extracting spectral features through Fourier transform, it is uploaded to the central server for noise classification. However, in practical applications, there is a data transmission delay of up to minutes, and the node energy consumption is concentrated in the spectrum calculation link, making it difficult to meet the real-time response requirements of large-scale data in multiple cities and multiple measurement points.

[0003] To solve the data transmission delay and energy consumption bottlenecks in the traditional architecture, existing improved technologies propose to use a big data training model combined with multi-level wavelet transform to achieve hierarchical compressed expression of soundprint features, and finally achieve automatic noise analysis, intelligent identification, and manual supervision.

[0004] However, in actual use, there are still some shortcomings. For example, the current monitoring system faces the bottleneck of dynamic environment adaptability. Although the general model trained based on static big data improves the single-point recognition efficiency, it does not consider the dynamic evolution law of urban noise sources in the time dimension in heterogeneous scenarios of multiple cities. Since the correlation framework between time period features and noise patterns is not constructed, in typical time periods such as day-night alternation and commuting rush hours, the noise classification threshold and feature weight cannot be dynamically adjusted, resulting in a sharp increase in the regional misjudgment rate during cross-time period monitoring and an increase in the frequency of manual supervision to correct time period false alarms. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-city noise automatic monitoring data analysis method and system based on time period classification, and through the following solutions, to solve the problems raised in the above-mentioned background art.

[0006] To achieve the above object, the present invention provides the following technical solutions: A multi-city noise automatic monitoring data analysis method based on time period classification, including: S1: Obtain noise monitoring data through a multi-city sensor node network, where the noise monitoring data includes a sound pressure signal obtained through dual-mode positioning and high-precision clock synchronization, a millisecond-level time stamp synchronized with the sound pressure signal, and geographical location information; S2: Perform spatio-temporal feature extraction operations based on the noise monitoring data. The spatio-temporal feature extraction operations slice the sound pressure signal by time slots and use an attention mechanism to generate a dynamic time slot fingerprint map with time-sensitive weights and spatial correlation weights; S3: Perform noise iterative analysis on the dynamic time slot fingerprint map, including feature weighted classification, feature decoupling migration, incremental learning, and anomaly feedback; S4: Continuously perform clock alignment between cross-city nodes and use the clock signal as a basic timing constraint condition for feedback adjustment.

[0007] Preferably, in S1, in the multi-city sensor node network, by judging the spatial relationship between the coordinate points corresponding to the sensor nodes and the preset geographical fence, the scene switching event is automatically triggered, specifically including: Load the benchmark noise feature library associated with the target geographical fence and send an update signal of the dynamic time slot fingerprint map to the monitoring data central processing module; Reconfigure the determination threshold of the noise according to the scene type and target time period of the target geographical fence; Dynamically adjust the data upload frequency according to the scene sensitivity of the target geographical fence.

[0008] Preferably, in S2, generating the dynamic time slot fingerprint map specifically includes: Group the power spectrum values corresponding to the continuous PCM waveform data in the noise monitoring data into a power spectrum matrix in frame order and input it into a time slice convolutional network to obtain time domain features, frequency domain features, and compressed feature vectors; Use the time slot index , and generate a time slot embedding vector through a trainable look-up table; Integrate the attention mechanism and construct a query vector , a key vector , and a value vector based on the time slot embedding vector ; Based on the query vector and the key vector , calculate the attention weight , specifically expressed as: , where represents the dimension of the key vector , represents the matrix product of the query vector and the transpose of the key vector ; Weighted sum the value vector through the attention weight to obtain a time slot sensitive feature vector , specifically expressed as: ; Package the compressed feature vector, the time slot index, the period-sensitive feature vector, and the power spectral matrix corresponding to each city to form a dynamic period fingerprint map.

[0009] Preferably, in step S2, calculate the power spectral values corresponding to the continuous PCM waveform data in the noise monitoring data , specifically expressed as: , wherein, represents the power spectral value of the th frame, represents the number of sampling points corresponding to the Hamming window, represents the index of the sampling points corresponding to the Hamming window, represents the signal value of the th sampling point, represents the Hamming window function value, represents the basis function of the discrete Fourier transform.

[0010] Preferably, in step S3, perform feature weighted classification, specifically including: Form a dual adaptive mechanism by constructing an LSTM period prediction unit and a noise pattern discriminator driven by a gradient reversal network; Perform adversarial training on the dynamic period fingerprint map to generate the probability distribution of each noise type.

[0011] Preferably, in step S3, perform feature decoupling migration, specifically including: By constructing a feature decoupling distillation framework, decompose the feature space of the multi-city noise analysis model into a city-independent common subspace and a city-specific private subspace to achieve cross-domain transfer of common features.

[0012] Preferably, in step S3, perform incremental learning, specifically including: Adopt the elastic weight consolidation algorithm to realize the dynamic evolution of the multi-city noise analysis model; If periodic false alarms are detected, freeze the common feature parameters and dynamically adjust the specific feature parameters, and realize incremental update through a dual-channel backpropagation mechanism.

[0013] Preferably, in step S3, perform abnormal feedback construction, specifically including: A spatio-temporal joint anomaly detection mechanism based on Mahalanobis distance; When the classification result deviates from the preset threshold of the historical period pattern, automatically trigger the manual review interface and generate a training sample stream with corrected labels.

[0014] To achieve the above object, the present invention provides the following technical solutions: A multi-city noise automatic monitoring data analysis system based on time period classification, including a multi-city noise monitoring database and a monitoring data central processing module. Implementing the above multi-city noise automatic monitoring data analysis method based on time period classification includes: Multi-source spatio-temporal data acquisition module: Through a multi-city sensor node network, noise monitoring data is obtained. The noise monitoring data includes a sound pressure signal obtained through dual-mode positioning and high-precision clock synchronization, a millisecond-level timestamp synchronized with the sound pressure signal, and geographical location information; Time period feature encoding module: Based on the noise monitoring data, a spatio-temporal feature extraction operation is performed. The spatio-temporal feature extraction operation slices the sound pressure signal by time slots and uses an attention mechanism to generate a dynamic time period fingerprint map containing time-sensitive weights and spatial correlation weights; Dynamic evolution cooperation module: Perform noise iterative analysis on the dynamic time period fingerprint map, including feature weighted classification, feature decoupling migration, incremental learning, and anomaly feedback; Clock synchronization module: Continuously perform clock alignment between cross-city nodes and use the clock signal as a basic timing constraint condition for feedback adjustment; The multi-city noise monitoring database includes all data texts of the multi-city noise automatic monitoring data analysis system based on time period classification, and real-time collects the information data output by each module; the monitoring data central processing module is used for the information data instructions output by each module in the central control system.

[0015] Preferably, the clock synchronization module continuously performs clock alignment between cross-city nodes, specifically including: The master clock node sends a synchronization message every 100 ms; Adopt a linear regression compensation algorithm to eliminate transmission delay jitter; Maintain the clock deviation of the entire network < 1 μs.

[0016] The technical effects and advantages of the present invention: 1. Through time slot slicing and the attention mechanism, the present invention converts the sound pressure signal into a dynamic time period fingerprint map containing time-sensitive weights and spatial correlation weights, constructs an association framework between time period features and noise patterns, accurately captures the dynamic evolution laws of the time domain and frequency domain of noise during periods such as day-night alternation and commuting peaks, enables the model to adaptively adjust the classification strategy according to the feature weights of different periods, and effectively reduces the regional misjudgment rate of cross-period monitoring; 2. Through the dual adaptive mechanisms of the LSTM time period prediction unit and the gradient reversal network, combined with feature decoupling migration and incremental learning, the present invention realizes cross-time-period and cross-city feature migration of noise patterns and dynamic evolution of the model. While freezing common features, it dynamically optimizes city-specific parameters, retains the classification experience of historical time periods, and can quickly adapt to the noise changes in new time periods, significantly reducing the occurrence of time-period false alarms. 3. Through the spatio-temporal joint detection mechanism based on Mahalanobis distance, when classifying anomalies, it automatically triggers manual review and generates a corrected sample stream, forming a self-optimizing closed loop of the monitoring system. By reducing the frequency of manual active supervision and converting passive correction into active learning, it effectively reduces the manual supervision cost and solves the pain point of frequent manual intervention in the prior art. Description of the Drawings

[0017] Figure 1 It is a flowchart of the implementation steps of the multi-city noise automatic monitoring data analysis method based on time period classification provided by an embodiment of the present application.

[0018] Figure 2 It is a flowchart of the implementation steps of noise iterative analysis in the multi-city noise automatic monitoring data analysis method based on time period classification provided by an embodiment of the present application.

[0019] Figure 3 It is a block diagram of the architecture of the multi-city noise automatic monitoring data analysis system based on time period classification provided by an embodiment of the present application. Detailed Embodiments

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. As used in the specification of the present application, the singular forms "a", "an", "the", "above", "said", "this" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term " / and / " used in the present application refers to and includes any or all possible combinations of one or more of the listed items. In the description of the embodiments of the present application, unless otherwise stated, the meaning of "a plurality" is two or more.

[0022] As shown in the attached Figure 1The multi-city noise automatic monitoring data analysis method based on time period classification shown in the figure realizes the dynamic analysis evolution of multi-city noise monitoring data by collecting timestamped sound pressure data and performing time slot slicing, so as to separate the common and specific features of cities through feature weighted classification and feature decoupling migration, and specifically includes the following steps: S1: Obtain noise monitoring data through a multi-city sensor node network. The noise monitoring data includes sound pressure signals obtained through dual-mode positioning and high-precision clock synchronization, millisecond-level timestamps synchronized with the sound pressure signals, and geographical location information; S2: Perform spatio-temporal feature extraction operations based on the noise monitoring data. The spatio-temporal feature extraction operations slice the sound pressure signals by time slots and use an attention mechanism to generate a dynamic time period fingerprint map containing time-sensitive weights and spatial correlation weights; S3: Perform noise iterative analysis on the dynamic time period fingerprint map, including feature weighted classification, feature decoupling migration, incremental learning, and anomaly feedback; S4: Continuously perform clock alignment between cross-city nodes and use the clock signal as a basic timing constraint condition for feedback adjustment.

[0023] Specifically, in S1, by deploying multiple sensor nodes in a multi-city area, the sensor nodes integrate a positioning function with GPS and Beidou dual-mode positioning capabilities, a high-precision clock synchronization function, and an acoustic signal acquisition function, and complete necessary network configuration and initialization; encapsulate a data packet containing a microsecond-level timestamp provided by a synchronous clock, longitude and latitude obtained by fusion positioning, instantaneous sound pressure level, PCM waveform data of a preset sampling window, and a dynamically associated geofence identifier as noise monitoring data.

[0024] In this embodiment, the acoustic signal acquisition function captures the original PCM waveform data through the ALSA audio driver layer at a sampling rate of 48 kHz and calculates the instantaneous sound pressure level using the real-time noise dose calculation method.

[0025] Furthermore, the high-precision clock synchronization function is based on an improved IEEE 1588 PTPv2.1 protocol stack. The master clock node periodically sends Sync messages containing the transmission timestamp and the slave node records the arrival time after receiving the message , obtains the slave node response time through message interaction and the master node reception time , calculates the delay compensation , and is specifically expressed as: , This embodiment uses a Kalman filter to eliminate network jitter in the synchronization result of the delay compensation, so that the clock deviation between nodes is less than 1 μs.

[0026] Further, call the underlying APIs of the devices corresponding to each sensor node, and integrate the GPS and Beidou dual-mode positioning engines to obtain the GPS positioning data in the WGS-84 coordinate system in real time and the Beidou positioning data and their corresponding horizontal dilution of precision (HDOP) ; then execute the multi-source positioning data fusion algorithm, and obtain the coordinate points corresponding to each sensor node according to the dynamic weighting formula , specifically expressed as: , wherein and are respectively expressed as and the corresponding weights during the fusion process; in this embodiment is set to , which represents the horizontal dilution of precision of GPS; is set to , which represents the horizontal dilution of precision of the Beidou system, to ensure that the positioning error of the weighted fusion is less than 3 meters, and obtain the fused longitude and latitude data at intervals of 50 ms.

[0027] It should be noted that by preloading the multi-city GIS geographical fence data stored in the multi-city noise monitoring database, the ray method point-in-polygon determination algorithm is executed in real time to determine the spatial relationship between the coordinate points corresponding to the sensor nodes and the preset geographical fences, and the acoustic scene tags are dynamically associated. The acoustic scene tags include but are not limited to residential quiet areas, industrial operation areas, traffic artery areas, commercial gathering areas, cultural and medical areas, park green spaces, etc.; when the mobile node crosses the fence boundary, a triple-scene switching event is automatically triggered.

[0028] In a possible implementation manner, the scene switching event includes: loading the reference noise feature library associated with the target geographical fence, and sending an update signal of the dynamic time fingerprint map to the monitoring data central processing module; reconfiguring the determination threshold of the noise according to the scene type and target time period of the target geographical fence; dynamically adjusting the data upload frequency according to the scene sensitivity of the target geographical fence.

[0029] ​​​​​In this embodiment, the residential quiet zone is the night time period from 22:00 to 6:00, and the equivalent sound level limit is ≤ 45 dB(A); the industrial operation zone is dominated by continuous mechanical noise spectrum, and the energy proportion in the frequency band of 63 Hz - 4 kHz is > 70%; the traffic artery zone has the spectral characteristics of typical motor vehicle noise, and the peak frequency point is 500 Hz ± 50 Hz; the commercial gathering zone has mixed human voice noise, and the variance of sound pressure fluctuation in the frequency band of 1 - 2 kHz is > 15 dB; the cultural and educational medical zone has the threshold of sudden loud events ≤ 65 dB(C) and the duration < 200 ms; the park and green space has intermittent pulses of natural soundscapes such as bird songs and wind sounds in the frequency band of 2 - 8 kHz.

[0030] Specifically, in S2, the noise monitoring data is received, and slicing and preprocessing based on time slots are performed. The 24-hour cycle is divided into minute-level time slots, and a time-sensitive feature vector is extracted through a time-slice convolutional network combined with an attention mechanism to generate a dynamic time slot fingerprint map including time-sensitive weights and spatial correlation weights. The dynamic time slot fingerprint map provides a spatio-temporal coupling feature representation for noise classification.

[0031] It should be noted that the slicing and preprocessing of the time slots include: using a sliding window mechanism to divide the continuous PCM waveform data in the noise monitoring data into 1440 independent 1-minute time slots according to the 24-hour cycle; based on the microsecond-level timestamps corresponding to the continuous PCM waveform data in the noise monitoring data, determining the time slot index corresponding to the noise data through a periodic slicing algorithm , specifically expressed as: , Among them, , represents the microsecond-level timestamp corresponding to the continuous PCM waveform data in the noise monitoring data; preprocessing is performed on the sound pressure signal data in each time slot, and the preprocessing includes but is not limited to filtering the sound pressure signal using an A-weighting filter, performing frame division on the filtered signal using a Hamming window by the frame addition window technique, and performing Fourier transform on each frame to calculate the corresponding power spectrum , specifically expressed as: , Among them, represents the power spectrum value of the th frame. In this embodiment , represents the number of sampling points corresponding to the Hamming window, represents the index of the sampling points corresponding to the Hamming window, represents the signal value of the th sampling point, Expressed as Hamming window function values, Expressed as the basis functions of the discrete Fourier transform, used to transform the signal from the time domain to the frequency domain.

[0032] In a possible implementation, generating the dynamic time segment fingerprint map includes: grouping the power spectral values corresponding to the continuous PCM waveform data in the noise monitoring data into a power spectral matrix in the frame order of the Hamming window, and inputting it into the time slice convolutional network.

[0033] It should be noted that the time slice convolutional network includes multiple levels; among them, the input layer receives the power spectral matrix; the time domain convolutional layer uses multiple 1D convolutional kernels and the ReLU activation function to extract time domain features; the frequency domain convolutional layer includes multiple 2D convolutional kernels and uses the LeakyReLU activation function to extract frequency domain features; the attention pooling layer performs multi-scale feature compression through multiple scaling factors to obtain a compressed feature vector.

[0034] Furthermore, the time domain features are preliminary features extracted from the power spectral matrix, reflecting the energy distribution pattern or change trend within a specific frequency range, including but not limited to the time domain envelope features of short-time energy mutations of transient events, the time domain periodic features of continuous noise, the zero crossing rate, the short-time energy integral value autocorrelation function, and the peak interval, etc. The transient events include but are not limited to impact sounds, blasting sounds, etc., and the continuous noise includes but is not limited to engine roars, fireworks blasting sounds, etc.; the frequency domain features are more complex and discriminative spectral patterns further extracted based on the time domain features, including but not limited to octave band energy distribution, the ratio of fundamental frequency to harmonic components, spectral centroid, special frequency identification, etc., and the special frequencies include but are not limited to traffic noise, animal calls, etc.; the compressed feature vector is a compact representation that comprehensively compresses the features extracted through time domain and frequency domain convolutions.

[0035] In a possible implementation, generating the dynamic time segment fingerprint map further includes: using the time slot index to generate a time slot embedding vector through a trainable lookup table ; integrating the attention mechanism, based on the time slot embedding vector to construct a query vector a key vector and a value vector , specifically expressed as; , , , wherein, , and are respectively expressed as the query vector , key vector and value vector The corresponding weight matrix, represented as the power spectral matrix corresponding to the time slot embedding vector, , and are respectively represented as the adjusted query vector , key vector and value vector offsets; based on the query vector and key vector , calculate the attention weight , specifically represented as: , wherein, is represented as the dimension of the key vector , is represented as the matrix product of the query vector and the transpose of the key vector ; through the attention weight weighted sum of the value vector is performed to obtain the time period sensitive feature vector , specifically represented as: ; In this embodiment, during the morning rush hour, that is, 7:00 - 9:00, by applying a gradient amplification factor to the convolution kernel corresponding to the 500 Hz frequency band, the noise feature weight is enhanced; during the late night period, that is, 22:00 - 6:00, by setting the activation threshold of the feature channel, the activation threshold of the 2 - 4 kHz frequency band feature channel is increased.

[0036] In a possible implementation manner, generating the dynamic time period fingerprint map further includes: encapsulating the compressed feature vector, the time slot index, the time period sensitive feature vector and the power spectral matrix corresponding to each city to form a dynamic time period fingerprint map.

[0037] Specifically, in S3, the feature weighted classification, feature decoupling migration, the incremental learning and the anomaly feedback form a technical closed loop through the shared feature space and parameter update chain. The result of the feature decoupling determines the parameter freezing range of the incremental learning. The model update amount generated by the incremental learning reversely restricts the adversarial training process of the feature decoupling. The threshold dynamic adjustment algorithm of the anomaly feedback depends on the common feature distribution of cities output by the feature decoupling.

[0038] In a possible implementation manner, the noise iterative analysis includes: the feature weighted classification forms a dual adaptive mechanism by constructing an LSTM time period prediction unit and a noise pattern discriminator driven by a gradient reversal network to perform adversarial training on the dynamic time period fingerprint map.

[0039] It should be noted that taking the dynamic time period fingerprint map as the input, generating the probability distribution of each noise type through the final classification layer of the noise pattern discriminator, applying the time period weight and the spatial correlation information to the feature vector, and encapsulating the final classification result. The encapsulated content includes: noise type encoding, probability arrays of each category, time dimension weight of the current time slot, spatial dimension weight of the current position, and time slot index.

[0040] Furthermore, the network architecture of the LSTM time period prediction unit includes an input layer, an LSTM hidden layer, an attention fusion layer, and an output layer. Among them, the input layer is used to receive the dynamic time period fingerprint map and perform time series modeling according to the feature vector sequence in the dynamic time period fingerprint map. The LSTM hidden layer contains multiple memory units. The attention fusion layer is used to calculate the attention weights of the historical time period features. The output layer is used to generate a time period feature weight matrix. The noise pattern discriminator constructed by the gradient reversal network includes a feature input layer, a gradient reversal layer, a fully connected layer, and a classification output layer. Among them, the feature input layer is used to receive the fingerprint map features. The gradient reversal layer is used to reverse the gradient during backpropagation. The fully connected layer contains multiple fully connected nodes. The classification output layer is used to output the probabilities of each noise type.

[0041] In this embodiment, the time step is set to 60, and the attention mechanism is applied to the output hidden states of the past 60 time steps to calculate the attention scores of each hidden state , and calculate the attention weights through the Softmax function , which is specifically expressed as: , Among them, represents the attention weight corresponding to the hidden state at the -th time step, represents the attention score corresponding to the hidden state at the -th time step, represents a preset weight vector used to perform weighted summation on the transformation results of the hidden states, represents a preset weight matrix used to map the hidden states, , respectively represent the -th and the The hidden state of the hidden layer of the LSTM for a time step; the weighted sum of the hidden states is passed through a fully connected layer and a Sigmoid activation function to generate a predicted time period feature weight matrix.

[0042] Furthermore, an adversarial training mechanism is adopted to optimize the LSTM time period prediction unit and the noise pattern discriminator driven by the gradient reversal network. In each training cycle: the LSTM unit generates prediction weights according to the input information, performs feature weighting on the original features and the prediction weights to obtain weighted features and inputs them into the noise pattern discriminator to obtain the classification probabilities of each noise type; calculates the classification loss of the noise pattern discriminator and the loss of the LSTM prediction unit, and defines the total loss, where the total loss is to maximize the loss of the noise pattern discriminator while minimizing the prediction error of the prediction unit; performs backpropagation on the total loss, the parameters of the noise pattern discriminator are updated normally according to the classification loss of the noise pattern discriminator, and the parameter update of the LSTM prediction unit is affected by the gradient reversal layer.

[0043] In a possible implementation manner, the noise iterative analysis further includes: the feature decoupling migration decomposes the feature space of the multi-city noise analysis model into a city-independent common subspace and a city-specific private subspace by constructing a feature decoupling distillation framework to achieve cross-domain transmission of common features.

[0044] In this embodiment, the dynamic time period fingerprint map is used as the input and decomposed into common features and private features through two independent feature extraction channels. The common features extract the general features shared across cities through the ReLU function, and the private features extract the city-specific features through the tanh function; an orthogonal regularization loss function is introduced to ensure that the common features and the private features are decoupled from each other in the feature space; calculates the element-wise product of the common features and the private features and sums them to obtain a correlation metric; takes the mean square value of the correlation metric as the orthogonal regularization loss term and forces the learning to generate independent common features and private features.

[0045] Further, using the city-independent common features as the knowledge source, through the knowledge distillation technology, the common feature extraction ability is migrated to the model of the target city, and the learned common features are applied to the model of the target city. The common features representing specific city types are loaded from the pre-constructed common feature library, and then the common features are frozen to keep them unchanged during the training process of the target city model, and only the layers responsible for extracting private features in the target model are trained to adapt to the specific noise pattern of the target city while retaining the migrated common feature extraction ability.

[0046] In a possible implementation, the noise iterative analysis further includes: the incremental learning uses the elastic weight consolidation algorithm to implement the dynamic evolution of the multi-city noise analysis model. If periodic false alarms are detected, the common feature parameters are frozen and the specific feature parameters are dynamically adjusted, and the model incremental update is realized through the dual-channel backpropagation mechanism.

[0047] It should be noted that by calculating the diagonal elements of the Fisher information matrix corresponding to the parameters of each multi-city noise analysis model, the sensitivity of the change of the loss function on the labeled time period fingerprint map samples is evaluated, and this is used as a measure of the parameter importance. , specifically expressed as; , where, represents the importance measure of parameter , represents the number of samples in the historical dataset, represents the sample index in the historical dataset, represents the loss function of the model on sample , represents the gradient of the loss function with respect to parameter ; Furthermore, the model parameters are sorted in descending order of importance, and the cumulative importance ratio when the parameters are arranged in this order is calculated. The parameters with a cumulative importance ratio reaching a preset threshold are selected as the core parameter set. The preset threshold in this embodiment is 85%. For the parameters determined to be core parameters, their gradient calculation function is disabled during the subsequent incremental learning process, while the gradient calculation of auxiliary parameters includes the standard loss gradient term and the regularization term based on EWC to balance the new task learning and the retention of old task knowledge.

[0048] In a possible implementation, the noise iterative analysis further includes: the anomaly feedback constructs a spatio-temporal joint anomaly detection mechanism based on the Mahalanobis distance. When the classification result deviates from the preset threshold of the historical time period pattern, the manual review interface is automatically triggered and a training sample stream with corrected labels is generated to realize the self-optimizing closed loop of the noise monitoring system.

[0049] It should be noted that the Mahalanobis distance between the current dynamic time period fingerprint map and the historical pattern of the corresponding time slot is calculated and compared with a dynamically set threshold; the calculation of the Mahalanobis distance is specifically expressed as: , where, Denoted as the inverse of the historical covariance matrix; the anomaly detection threshold is dynamically set based on the statistical distribution of the historical Mahalanobis distance, and the three-standard-deviation principle is adopted; Further, when it is detected that the Mahalanobis distance exceeds the dynamic threshold, an artificial review process is triggered, and a sample stream for model self-optimization is generated according to the review result; in this embodiment, different strategies are adopted to generate and transmit correction samples according to the type and persistence of the anomaly, so as to optimize the feedback efficiency and the model training effect; when there is a single-point sudden anomaly, an independent sample is immediately generated and has the highest transmission priority; when the threshold is adaptively updated and the dynamic threshold is adjusted according to the new historical data, additional metadata is generated to explain the reason and scope of the threshold adjustment, no specific sample is generated, and the transmission priority is the lowest.

[0050] Specifically, in S4, an improved PTPv2.1 protocol is adopted to achieve sub-microsecond time synchronization between cross-city nodes, and a unified spatio-temporal reference system is constructed to provide a benchmark guarantee for the spatio-temporal correlation analysis of noise monitoring data.

[0051] As shown in the appendix Figure 2 The multi-city noise automatic monitoring data analysis system based on time period classification shown in the figure includes a multi-city noise monitoring database and a monitoring data central processing module, and also includes: Multi-source spatio-temporal data acquisition module: Through the multi-city sensor node network, noise monitoring data is obtained. The noise monitoring data includes sound pressure signals obtained through dual-mode positioning and high-precision clock synchronization, millisecond-level timestamps synchronized with the sound pressure signals, and geographical location information; Time period feature encoding module: Based on the noise monitoring data, spatio-temporal feature extraction operations are performed. The spatio-temporal feature extraction operations slice the sound pressure signals by time slots and use an attention mechanism to generate a dynamic time period fingerprint map containing time-sensitive weights and spatial correlation weights; Dynamic evolution cooperation module: Perform noise iterative analysis on the dynamic time period fingerprint map, including feature weighted classification, feature decoupling migration, incremental learning, and anomaly feedback; Clock synchronization module: Continuously perform clock alignment between cross-city nodes and use the clock signal as a basic timing constraint condition for feedback adjustment; In this embodiment, the clock synchronization module continuously performing clock alignment between cross-city nodes includes: Through the master clock node pre-deployed in the multi-city noise monitoring database, a Sync synchronization message is generated and broadcast every 100 ms. The message structure adopts a streamlined frame header design and uses a multicast address to achieve efficient one-to-many transmission. When the network is congested, the frequency is automatically reduced to 200 ms, and retransmission is triggered if the response from the slave node is not received continuously twice; Furthermore, the node records multiple consecutive sets of time deviation data, and uses a linear regression compensation algorithm to solve the predicted deviation at the current moment, and then adjusts the local clock to eliminate the transmission delay jitter; Furthermore, each node generates time slot indexes based on a unified clock, maintaining the clock deviation of the entire network < 1 μs, providing a unified benchmark for time slot division; The multi-city noise monitoring database includes all data texts of the multi-city noise automatic monitoring data analysis system classified based on time periods, and real-time collects the information data output by each module. The monitoring data central processing module is used for the information data instructions output by each module in the central control system.

[0052] Among them, the multi-city noise monitoring database is stored using a spatio-temporal sub-database architecture, and is used to store a large amount of data related to the multi-city noise automatic monitoring data analysis classified based on time periods, including a cross-city feature decoupling model established based on an adversarial discriminator and orthogonal constraints, historical versions of the LSTM time period prediction weight matrix, and dynamic time period fingerprint maps stored in a two-dimensional manner by city-time period, etc.; when the monitoring data central processing module performs dynamic feature weighted classification, it real-time calls the common feature template data in the model library, and combines the spatio-temporal weight matrix in the feature library to perform adversarial training gradient calculation, parameter importance evaluation, and dual-channel backpropagation operations.

[0053] Among them, the monitoring data central processing module is the core operation and control unit of the entire multi-city noise automatic monitoring data analysis system classified based on time periods; it includes a multi-core heterogeneous computing architecture, which is connected to a distributed clock synchronization module, a multi-source spatio-temporal calibration acquisition module, and a dynamic evolution cooperation module through a high-speed bus and a dedicated interface; by running or executing the convolutional neural network instruction set, incremental learning algorithm code set, and spatio-temporal attention calculation instructions stored in the system operation database, and being able to call the cross-city feature template data stored therein, thereby performing operations such as spatio-temporal joint feature encoding, dynamic weight matrix generation, and feature decoupling distillation calculation, to ensure that the minute-level feature slicing of the time slot is strictly synchronized with the incremental learning process.

[0054] Secondly: In the attached drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved. For other structures, reference can be made to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other; Finally: The above are only the preferred embodiments of the present invention, and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for automatically analyzing multi-city noise monitoring data based on time period classification, characterized in that, It also includes: S1: Obtain noise monitoring data through a multi-city sensing node network. The noise monitoring data includes a sound pressure signal obtained through dual-mode positioning and high-precision clock synchronization, a millisecond-level timestamp synchronized with the sound pressure signal, and geographical location information; S2: Perform spatio-temporal feature extraction operations based on the noise monitoring data. The spatio-temporal feature extraction operations slice the sound pressure signal by time slots and use an attention mechanism to generate a dynamic time-slot fingerprint map containing time-sensitive weights and spatial correlation weights; S3: Perform noise iterative analysis on the dynamic time-slot fingerprint map, including feature weighted classification, feature decoupling migration, incremental learning, and anomaly feedback; S4: Continuously perform clock alignment between cross-city nodes and use the clock signal as a basic timing constraint condition for feedback adjustment.

2. The method for automatically analyzing multi-city noise monitoring data based on time period classification according to claim 1, characterized in that: In S1, in the multi-city sensing node network, by judging the spatial relationship between the coordinate point corresponding to the sensor node and the preset geographical fence, a scene switching event is automatically triggered, specifically including: Loading a reference noise feature library associated with the target geographical fence and sending an update signal of the dynamic time-slot fingerprint map to the monitoring data central processing module; Reconfiguring the determination threshold of the noise according to the scene type and target time period of the target geographical fence; Dynamically adjusting the data upload frequency according to the scene sensitivity of the target geographical fence.

3. The multi - city noise automatic monitoring data analysis method based on time - period classification according to claim 1, characterized in that: In S2, generating a dynamic time-slot fingerprint map specifically includes: Grouping the power spectral values corresponding to the continuous PCM waveform data in the noise monitoring data into a power spectral matrix in frame order and inputting it into a time slice convolutional network to obtain time-domain features, frequency-domain features, and compressed feature vectors; Using time slot index , generating a time slot embedding vector through a trainable lookup table; Integrated attention mechanism, based on time slot embedding vectors , construct query vectors , key vectors and value vectors ; Based on the query vector and the key vector , calculate the attention weights , which is specifically expressed as: ， Among them, denotes the dimension of the key vector , denotes the matrix product of the query vector and the transpose of the key vector ; Through attention weights weight the value vectors and perform weighted summation to obtain a time period sensitive feature vector , which is specifically expressed as: ； Encapsulating the compressed feature vector, the time slot index, the time period sensitive feature vector, and the power spectral matrix corresponding to each city to form a dynamic time-slot fingerprint map.

4. The multi-city noise automatic monitoring data analysis method based on time period classification according to claim 3, characterized in that: S2 calculates the power spectral values corresponding to consecutive PCM waveform data in the noise monitoring data , which is specifically expressed as: ， Among them, is expressed as the power spectral value of the th frame, is expressed as the number of sampling points corresponding to the Hamming window, is expressed as the index of the sampling points corresponding to the Hamming window, is expressed as the th signal value of the sampling point, is expressed as the Hamming window function value, is expressed as the basis function of the discrete Fourier transform.

5. The method for automatically analyzing multi-city noise monitoring data based on time period classification according to claim 1, characterized in that: In S3, performing feature weighted classification specifically includes: Forming a dual adaptive mechanism by constructing an LSTM time period prediction unit and a noise pattern discriminator driven by a gradient reversal network; Performing adversarial training on the dynamic time-slot fingerprint map to generate the probability distribution of each noise type.

6. The multi-city noise automatic monitoring data analysis method based on time period classification according to claim 1, characterized in that: In S3, performing feature decoupling migration specifically includes: By constructing a feature decoupling distillation framework, decomposing the feature space of the multi-city noise analysis model into a city-independent common subspace and a city-specific private subspace to achieve cross-domain transfer of common features.

7. The method for automatically analyzing multi-city noise monitoring data based on time period classification according to claim 1, characterized in that: In S3, performing incremental learning specifically includes: Using the elastic weight consolidation algorithm to realize the dynamic evolution of the multi-city noise analysis model; If periodic false alarms are detected, freeze the common feature parameters and dynamically adjust the specific feature parameters, and realize incremental update through a two-channel backpropagation mechanism.

8. The multi-city noise automatic monitoring data analysis method based on time period classification according to claim 1, characterized in that: In S3, performing anomaly feedback construction specifically includes: A spatio-temporal joint anomaly detection mechanism based on Mahalanobis distance; When the classification result deviates from the preset threshold of the historical time period pattern, automatically trigger an artificial review interface and generate a training sample stream with corrected labels.

9. An automatic multi-city noise monitoring data analysis system based on time period classification, comprising a multi-city noise monitoring database and a central monitoring data processing module, which is used for the automatic multi-city noise monitoring data analysis method according to any one of the above claims 1-8, characterized in that It also includes: Multi-source spatio-temporal data acquisition module: Obtain noise monitoring data through a multi-city sensor node network. The noise monitoring data includes sound pressure signals obtained through dual-mode positioning and high-precision clock synchronization, millisecond-level timestamps synchronized with the sound pressure signals, and geographical location information. Time period feature encoding module: Perform spatio-temporal feature extraction operations based on the noise monitoring data. The spatio-temporal feature extraction operations slice the sound pressure signals by time slots and use an attention mechanism to generate a dynamic time period fingerprint map containing time-sensitive weights and spatial correlation weights. Dynamic evolution collaboration module: Perform noise iterative analysis on the dynamic time period fingerprint map, including feature weighted classification, feature decoupling migration, incremental learning, and anomaly feedback. Clock synchronization module: Continuously perform clock alignment between cross-city nodes and use the clock signal as a basic timing constraint condition for feedback adjustment. The multi-city noise monitoring database includes all data texts of a multi-city noise automatic monitoring data analysis system based on time period classification, and real-time collects the information data output by each module; the monitoring data central processing module is used for the information data instructions output by each module in the central control system.

10. The multi-city noise automatic monitoring data analysis system based on time period classification according to claim 9, characterized in that: The clock synchronization module, continuously performing clock alignment between cross-city nodes, specifically includes: The master clock node sends synchronization messages every 100 ms. Adopt a linear regression compensation algorithm to eliminate transmission delay jitter. Maintain the clock deviation of the entire network < 1 μs.

Images