A deep learning-based dust noise monitoring data intelligent analysis method

Abstract

The application discloses a kind of based on deep learning's dust raising noise monitoring data intelligent analysis method, by obtaining initial multi-source monitoring data, utilize wavelet threshold to carry out the noise reduction of dust raising data in data, by adaptive band filter, non-environmental interference in data is carried out the noise reduction, utilize 1D-CNN hollow convolutional neural network extraction multiscale time sequence feature establishes dust raising branch, by Transformer encoder extraction long-range spectral dependence establishes noise branch, by attention fusion mechanism carries out cross-modal feature interaction, obtains double-flow deep neural network model;The noise reduction monitoring data is input into model and is identified, and event classification probability and decision factor are output;Intelligent early warning is carried out according to evaluation result.Effectively improve complex environmental event identification accuracy, and the identification accuracy of construction dust raising event is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental monitoring, and in particular to a dust and noise monitoring data intelligent analysis method based on deep learning. BACKGROUND

[0002] Traditional dust and noise monitoring technology mainly relies on single sensors and simple threshold judgment, which has significant limitations. Single sensors can only obtain single type data, and cannot comprehensively consider the influence of meteorological conditions, environmental interference and other factors on the monitoring results, resulting in insufficient accuracy and comprehensiveness of the data. Traditional dust monitoring equipment is prone to misjudgment of non-pollution source particles raised by wind force as dust pollution, causing distortion of monitoring data. Simple threshold judgment lacks deep analysis ability for complex environmental characteristics, and the analysis accuracy is not high. SUMMARY

[0003] The present application is designed to solve the above problems, and provides a dust and noise monitoring data intelligent analysis method based on deep learning.

[0004] To achieve the above purpose, the technical scheme of the present application is as follows: further, in the dust and noise monitoring data intelligent analysis method based on deep learning, the dust and noise monitoring data intelligent analysis method comprises the following steps:

[0005] Obtain initial multi-source monitoring data, use wavelet threshold to denoise the dust data in the data, use adaptive frequency band filtering to denoise the non-environmental interference in the data, and obtain denoised monitoring data;

[0006] Use 1D-CNN hollow convolutional neural network to extract multi-scale time sequence features to establish a dust branch, use Transformer encoder to extract long-range spectral dependence to establish a noise branch, and use attention fusion mechanism for cross-modal feature interaction to obtain a CNN-Transformer dual-flow deep neural network model;

[0007] Input the denoised monitoring data into the CNN-Transformer dual-flow deep neural network model for recognition, output event classification probability and decision factor, evaluate the environmental state according to the event classification probability and decision factor, and perform intelligent early warning according to the evaluation result.

[0008] Further, in the dust and noise monitoring data intelligent analysis method based on deep learning, the obtaining of the initial multi-source monitoring data, the use of wavelet threshold to denoise the dust data in the data, and the use of adaptive frequency band filtering to denoise the non-environmental interference in the data to obtain denoised monitoring data comprises:

[0009] The multi-source monitoring data is obtained by synchronously collecting dust concentration data, noise spectrum data and meteorological parameters through multi-source sensors,

[0010] The original data of each sensor is subjected to outlier detection, and the time point data of a single sensor is judged. If the value exceeds 3 times the standard deviation of the historical data mean of the sensor, it is judged as an outlier, and the outlier is marked.

[0011] For the marked outliers, interpolation method is used to repair the adjacent time point data to obtain the processed multi-source monitoring data.

[0012] Further, in the above-mentioned dust and noise monitoring data intelligent analysis method based on deep learning, the initial multi-source monitoring data is obtained, the dust data in the data is denoised by wavelet threshold, and the non-environmental interference in the data is denoised by adaptive frequency band filtering to obtain denoised monitoring data, comprising:

[0013] The processed multi-source monitoring data is obtained, and the preprocessed dust concentration data is subjected to 3-layer wavelet decomposition by using db4 wavelet base function to obtain detail coefficients and approximation coefficients of different frequency levels.

[0014] The threshold value of each layer of detail coefficients is calculated by using the threshold value, the detail coefficients are processed by using the soft threshold function, and the denoised dust concentration data is obtained by wavelet reconstruction of the processed coefficients.

[0015] Further, in the above-mentioned dust and noise monitoring data intelligent analysis method based on deep learning, the initial multi-source monitoring data is obtained, the dust data in the data is denoised by wavelet threshold, and the non-environmental interference in the data is denoised by adaptive frequency band filtering to obtain denoised monitoring data, comprising:

[0016] The processed multi-source monitoring data is obtained, and the noise spectrum data in the initial multi-source monitoring data and the time domain signal collected by the noise sensor are converted into frequency domain signals by using FFT fast Fourier transform, and real-time spectrum analysis is performed to obtain the frequency components of the noise to obtain the spectrum analysis result.

[0017] The frequency band range of the environmental interference is determined according to the spectrum analysis result. If the periodic noise generated by the industrial equipment is determined to be in the frequency band of 50-200Hz, the frequency band is set as the filter-out frequency band.

[0018] The IIR infinite impulse response filter is used for adaptive frequency band filtering of the filter-out frequency band, and the band-stop filter is used to set the passband and stopband parameters of the filter according to the determined frequency band range, and the non-environmental interference is denoised to obtain the denoised monitoring data.

[0019] Further, in the above-mentioned deep learning-based intelligent analysis method for dust and noise monitoring data, the 1D-CNN hollow convolutional neural network is used to extract multi-scale time sequence features to establish a dust branch, a Transformer encoder is used to extract long-range spectral dependence to establish a noise branch, and an attention fusion mechanism is used for cross-modal feature interaction to obtain a CNN-Transformer double-flow deep neural network model, which includes:

[0020] The 1D-CNN hollow convolutional neural network at least includes an input layer, a hollow convolutional layer, a batch normalization layer, and a pooling layer.

[0021] The input layer is used to receive dust data and meteorological parameters in the noise reduction monitoring data, and the dust data and the meteorological parameters are spliced into one-dimensional time sequence data.

[0022] Three hollow convolutional layers are set, the first layer has an expansion rate of 1, a convolution kernel size of 3, and an output channel number of 64; the second layer has an expansion rate of 2, a convolution kernel size of 3, and an output channel number of 128; and the third layer has an expansion rate of 4, a convolution kernel size of 3, and an output channel number of 256, so as to obtain multi-scale time sequence features through different expansion rates.

[0023] A batch normalization layer is added after each hollow convolutional layer, and an average pooling layer is added after the last hollow convolutional layer, with a pooling window size of 2.

[0024] Further, in the above-mentioned deep learning-based intelligent analysis method for dust and noise monitoring data, the 1D-CNN hollow convolutional neural network is used to extract multi-scale time sequence features to establish a dust branch, a Transformer encoder is used to extract long-range spectral dependence to establish a noise branch, and an attention fusion mechanism is used for cross-modal feature interaction to obtain a CNN-Transformer double-flow deep neural network model, which includes:

[0025] The Transformer encoder at least includes an input layer, a position encoding layer, a Transformer encoder layer, and a normalization layer.

[0026] The input layer is used to input the noise spectrum data after noise reduction as input, and a sine-cosine position encoding is used to add position information to the noise spectrum data.

[0027] Six Transformer encoders are set, each of which contains a multi-head self-attention mechanism and a feedforward neural network, the multi-head self-attention mechanism allows the model to focus on different parts in different representation subspaces to capture long-range spectral dependence, and a layer normalization layer is added before and after the multi-head self-attention mechanism and the feedforward neural network.

[0028] Further, in the above-mentioned deep learning-based dust noise monitoring data intelligent analysis method, the denoised monitoring data is input into the CNN-Transformer dual-flow deep neural network model for recognition, and event classification probability and decision factor are output; environmental state assessment is performed according to the event classification probability and decision factor, and intelligent early warning is performed according to the assessment result, including:

[0029] The event classification probability at least includes construction dust, traffic noise, industrial noise and normal environment, and the decision factor is used to measure the credibility of the model prediction;

[0030] The classification probability threshold of each category event is set to 0.7, and when the classification probability of the category event is greater than or equal to 0.7, it is considered that the possibility of the event occurring is high;

[0031] The decision factor ranges from 0 to 1, and the decision factor threshold is set to 0.6, and when the decision factor is greater than or equal to 0.6, it is considered that the model prediction result is reliable.

[0032] Further, in the system for implementing a deep learning-based dust noise monitoring data intelligent analysis method, the system includes the following sub-modules:

[0033] A monitoring data acquisition module is configured to acquire initial multi-source monitoring data, denoise dust data in the data using a wavelet threshold, and denoise non-environmental interference in the data through adaptive frequency band filtering to obtain denoised monitoring data;

[0034] A network model establishment module is configured to extract multi-scale time sequence features using a 1D-CNN hollow convolutional neural network to establish a dust branch, extract long-range spectral dependence through a Transformer encoder to establish a noise branch, and perform cross-modal feature interaction through an attention fusion mechanism to obtain a CNN-Transformer dual-flow deep neural network model;

[0035] A monitoring intelligent analysis module is configured to input the denoised monitoring data into the CNN-Transformer dual-flow deep neural network model for recognition, output event classification probability and decision factor, perform environmental state assessment according to the event classification probability and decision factor, and perform intelligent early warning according to the assessment result.

[0036] Further, in the system for implementing a deep learning-based dust noise monitoring data intelligent analysis method, the system includes the following sub-modules:

[0037] A collection sub-module is configured to synchronously collect dust concentration data, noise spectrum data and meteorological parameters through multi-source sensors to obtain multi-source monitoring data,

[0038] A judgment sub-module is configured to perform outlier detection on the raw data of each sensor, and to judge the data of a single sensor at a time point as an outlier if the value of the data is more than 3 times the standard deviation of the mean value of historical data of the sensor, and to mark the outlier;

[0039] A repair sub-module is configured to repair the marked outliers by using an adjacent time point data interpolation method to obtain processed multi-source monitoring data.

[0040] Further, in the system for implementing the dust noise monitoring data intelligent analysis method based on deep learning, the system comprises the following sub-modules:

[0041] A decomposition sub-module is configured to obtain the processed multi-source monitoring data, perform 3-layer wavelet decomposition on the preprocessed dust concentration data by using a db4 wavelet base function to obtain detail coefficients and approximation coefficients of different frequency levels;

[0042] A reconstruction sub-module is configured to calculate a threshold value of each layer of detail coefficients by using a threshold value, process the detail coefficients by using a soft threshold function, and perform wavelet reconstruction on the processed coefficients to obtain the noise-reduced dust concentration data.

[0043] The beneficial effects are as follows: the initial multi-source monitoring data is obtained, the dust data in the data is denoised by using a wavelet threshold value, the non-environmental interference in the data is denoised by using an adaptive frequency band filter to obtain denoised monitoring data; a 1D-CNN hollow convolutional neural network is used to extract multi-scale time sequence features to establish a dust branch, a Transformer encoder is used to extract long-range spectral dependence to establish a noise branch, and an attention fusion mechanism is used to perform cross-modal feature interaction to obtain a CNN-Transformer dual-flow deep neural network model; the denoised monitoring data is input into the CNN-Transformer dual-flow deep neural network model for identification, and event classification probability and decision factors are output; the environmental state is evaluated according to the event classification probability and the decision factors, and intelligent early warning is performed according to the evaluation result. 1. Compared with a traditional single sensor scheme, the data dimension is improved, the pollution source characteristics and environmental influence factors are completely covered, and full-dimensional data support is provided for subsequent analysis. 2. The complex environmental event recognition accuracy is effectively improved, the construction dust event recognition accuracy is improved, and the comprehensive performance is improved compared with the traditional scheme in complex scenes such as construction sites and industrial parks. BRIEF DESCRIPTION OF DRAWINGS

[0044] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not considered a limitation of the present application.

[0045] Figure 1A first embodiment schematic view of a dust-raising noise monitoring data intelligent analysis method based on deep learning in an embodiment of the present application is shown in the figure.

[0046] Figure 2 A second embodiment schematic view of a dust-raising noise monitoring data intelligent analysis method based on deep learning in an embodiment of the present application is shown in the figure.

[0047] Figure 3 A third embodiment schematic view of a dust-raising noise monitoring data intelligent analysis method based on deep learning in an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0049] Those skilled in the art can understand that, unless specifically stated, the singular forms "a", "an" and "the" used herein can also include the plural forms. It should be further understood that the use of the word "comprise" in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but do not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0050] The present application will be described in detail below in combination with the drawings, such as Figure 1 A dust-raising noise monitoring data intelligent analysis method based on deep learning, which comprises the following steps:

[0051] Step 101, obtaining initial multi-source monitoring data, using wavelet threshold to denoise the dust data in the data, and using adaptive frequency band filtering to denoise the non-environmental interference in the data to obtain denoised monitoring data;

[0052] Specifically, in the present embodiment, multi-source sensor is used to synchronously collect dust concentration data, noise spectrum data and meteorological parameters to obtain multi-source monitoring data,

[0053] The original data of each sensor is subjected to outlier detection, and the time data of a single sensor is judged. If the value exceeds 3 times the standard deviation of the historical data mean value of the sensor, it is judged as an outlier, and the outlier is marked.

[0054] For the marked outliers, interpolation method is used to repair the adjacent time point data to obtain the processed multi-source monitoring data.

[0055] The processed multi-source monitoring data is acquired, the pre-processed dust concentration data is decomposed by 3 layers of wavelet based on db4 wavelet base function, and detail coefficients and approximation coefficients of different frequency levels are obtained;

[0056] Threshold values of the detail coefficients of each layer are calculated by using a threshold value, the detail coefficients are processed by using a soft threshold function, the processed coefficients are reconstructed by wavelet, and the dust concentration data after noise reduction is obtained.

[0057] The processed multi-source monitoring data is acquired, the noise spectrum data in the initial multi-source monitoring data and the time domain signal collected by the noise sensor are converted into frequency domain signals by using FFT fast Fourier transform, real-time spectrum analysis is performed to obtain the frequency components of the noise, and the spectrum analysis result is obtained;

[0058] The frequency band range of the environmental interference is determined according to the spectrum analysis result, if the periodic noise generated by the industrial equipment is determined as the 50-200Hz frequency band, the frequency band is set as the filtering frequency band;

[0059] The IIR infinite impulse response filter is used for adaptive frequency band filtering of the filtering frequency band, the passband and stopband parameters of the filter are set according to the determined frequency band range by using the band-stop filter, the non-environmental interference is processed by noise reduction, and the noise reduction monitoring data is obtained.

[0060] Specifically,

[0061] I. Multi-source sensor deployment

[0062] Sensor type: high-precision dust sensor (laser scattering dust sensor, measurement range 0-10mg / m³, accuracy ±10%), noise sensor (capacitive microphone noise sensor, measurement range 30-130dB, accuracy ±1.5dB), meteorological sensor (including temperature and humidity sensor, wind speed and direction sensor, barometric pressure sensor). The temperature and humidity sensor is selected as SHT30 type, the measurement temperature range is -40-125℃, the accuracy is ±0.3℃, the humidity range is 0-100%RH, and the accuracy is ±2%RH; the wind speed and direction sensor is ultrasonic, the wind speed measurement range is 0-60m / s, the accuracy is ±0.3m / s, the wind direction measurement range is 0-360°, and the accuracy is ±3°; the barometric pressure sensor is selected as MS5611 type, the measurement range is 10-1200hPa, and the accuracy is ±0.12hPa.

[0063] Sensor layout: uniformly deploy sensor nodes in the monitoring area, each node contains various sensors described above, and the interval is set to 50-100 meters according to the monitoring accuracy requirement. For the key monitoring area (construction site, traffic trunk road), the sensor deployment is appropriately densified.

[0064] II. Synchronization acquisition

[0065] A high-precision GPS timing module (u-blox NEO-M8N) is used to provide a unified time reference for each sensor, with a time synchronization accuracy of nanoseconds. Each sensor collects data at a sampling period of 1 second, and the collection time is triggered synchronously by the GPS timing module to ensure that the multi-source data is strictly aligned on the time axis.

[0066] III. Preprocessing process

[0067] Data format conversion:

[0068] Analog signal processing: The analog signals output by the dust sensor and some meteorological sensors are converted to digital signals by a 24-bit analog-to-digital converter (ADC), with conversion accuracy meeting the resolution requirements of monitoring data.

[0069] Unified data format: The digital signals output by all sensors are converted to the system-specified JSON format, with data structures including timestamp, sensor type, monitoring value, unit, and other fields. For example:

[0070] Outlier detection and processing:

[0071] Z-score method is used to detect outliers. For each sensor's monitoring data sequence, the mean and standard deviation are calculated.

[0072] Outlier processing: If the outlier is an isolated point, the average value of the adjacent two normal data points is used for interpolation processing; if multiple consecutive data points are outliers, they are marked as data missing, and the sliding window mean is used for filling in the subsequent model input.

[0073] Noise reduction processing

[0074] I. Wavelet threshold denoising of dust data

[0075] Wavelet basis selection and decomposition level: db4 wavelet basis is selected, which has good regularity and vanishing moments, suitable for dust data denoising. According to the sampling frequency (1 Hz) and noise characteristics of dust data, the decomposition level is determined to be 4 layers.

[0076] Wavelet decomposition and threshold processing:

[0077] The preprocessed dust data is decomposed into 4 layers of wavelet, obtaining the low-frequency coefficients and high-frequency coefficients of each layer.

[0078] The soft threshold function is used to process the high-frequency coefficients, and the threshold calculation uses VisuShrink wavelet reconstruction: the processed coefficients of each layer are reconstructed by wavelet to obtain the denoised dust data.

[0079] II. Non-environmental interference adaptive frequency band filter denoising

[0080] Noise spectrum analysis: Real-time spectrum analysis is performed on the noise spectrum data in the initial multi-source monitoring data (time domain signals collected by noise sensors are converted into frequency domain signals using Fast Fourier Transform (FFT), with a transform point number of 1024 and a frequency resolution of 0.9766 Hz) to obtain the main frequency components of the noise.

[0081] Adaptive frequency band determination: Based on the spectrum analysis results, the main frequency band range of environmental interference is dynamically determined. For example, if periodic noise generated by industrial equipment is mainly concentrated in the 50-200Hz frequency band, this frequency band is set as the frequency band that needs to be filtered out.

[0082] Filtering: An infinite impulse response (IIR) filter is used for adaptive frequency band filtering. The filter type is selected as a band-stop filter, and the passband and stopband parameters of the filter are set according to the determined frequency band range to achieve noise reduction processing of non-environmental interference.

[0083] Step 102, using 1D-CNN hollow convolutional neural network to extract multi-scale time sequence features to establish dust branch, extracting long-range spectral dependence through Transformer encoder to establish noise branch, and performing cross-modal feature interaction through attention fusion mechanism to obtain CNN-Transformer dual-flow deep neural network model;

[0084] Specifically, the 1D-CNN hollow convolutional neural network in the embodiment at least includes an input layer, a hollow convolutional layer, a batch normalization layer and a pooling layer;

[0085] The input layer is used to receive dust data and meteorological parameters in the noise reduction monitoring data, and the dust data and meteorological parameters are spliced into one-dimensional time series data.

[0086] Three hollow convolutional layers are set, the first layer has an expansion rate of 1, a convolution kernel size of 3 and an output channel number of 64; the second layer has an expansion rate of 2, a convolution kernel size of 3 and an output channel number of 128; and the third layer has an expansion rate of 4, a convolution kernel size of 3 and an output channel number of 256, and multi-scale time sequence features are obtained through different expansion rates.

[0087] A batch normalization layer is added after each hollow convolutional layer, and an average pooling layer is added after the last hollow convolutional layer, with a pooling window size of 2.

[0088] The Transformer encoder at least includes an input layer, a position encoding layer, a Transformer encoder layer and a normalization layer;

[0089] The input layer is used to input the noise spectrum data after noise reduction as input, and position information is added to the noise spectrum data based on sine-cosine position encoding.

[0090] Six Transformer encoders are set, each of which contains a multi-head self-attention mechanism and a feed-forward neural network. The multi-head self-attention mechanism allows the model to focus on different parts in different representation subspaces, capturing long-range spectral dependencies. Layer normalization layers are added before and after the multi-head self-attention mechanism and the feed-forward neural network.

[0091] Specifically,

[0092] Dust branch: 1D-CNN hollow convolutional neural network

[0093] Network structure: input layer: receives the denoised dust data and meteorological parameters (temperature, humidity, wind speed, wind direction, air pressure), and splices them into one-dimensional time series data.

[0094] Dilated convolutional layer: three dilated convolutional layers are set. The first layer has an expansion rate of 1, a convolution kernel size of 3, and an output channel number of 64; the second layer has an expansion rate of 2, a convolution kernel size of 3, and an output channel number of 128; the third layer has an expansion rate of 4, a convolution kernel size of 3, and an output channel number of 256. Different expansion rates are used to obtain multi-scale time series features.

[0095] Batch normalization layer (BatchNormalization): a batch normalization layer is added after each dilated convolutional layer to speed up network convergence and reduce overfitting.

[0096] Activation function: ReLU activation function is used to increase the non-linear representation ability of the network.

[0097] Pooling layer: an average pooling layer is added after the last dilated convolutional layer, with a pooling window size of 2, to reduce the feature dimension.

[0098] Noise branch: Transformer encoder

[0099] Input layer: the denoised noise spectrum data (frequency domain signal after FFT conversion, with 512 frequency points) is taken as input, with an input dimension of \(T\times512\).

[0100] Position encoding: since the Transformer model does not have time sequence order information, sine-cosine position encoding is used to add position information to the noise spectrum data. The position encoding formula is:

[0101] Transformer encoder layer: six Transformer encoders are set, each of which contains a multi-head self-attention mechanism (with 8 heads) and a feed-forward neural network. The multi-head self-attention mechanism allows the model to focus on different parts in different representation subspaces, capturing long-range spectral dependencies.

[0102] Layer Normalization: Layer normalization layers are added before and after the multi-head self-attention mechanism and the feedforward neural network to stabilize network training.

[0103] Attention Fusion Mechanism

[0104] Feature Extraction: The feature vector output by the dust branch is, with a dimension of ; the feature vector output by the noise branch is, with a dimension of. The dimensions of the two are unified to (set to 1024) through a fully connected layer.

[0105] Cross-modal Attention Calculation: A bidirectional attention mechanism is used to calculate the attention weight of dust features on noise features and the attention weight of noise features on dust features

[0106] Feature Fusion: The features of the two branches are weighted and fused through the attention weight to obtain the fused features.

[0107] Model Identification and Output

[0108] I. Data Input Processing

[0109] The denoised monitoring data (including dust data, noise spectrum data, and meteorological parameters) are normalized by the minimum-maximum normalization method, scaling the data to the [0, 1] interval. The data is divided into input samples according to the time window (length 100, sliding step 50).

[0110] II. Model Identification Process

[0111] The normalized input samples are input into the CNN-Transformer dual-flow deep neural network model. The dust branch extracts multi-scale time sequence features through the dilated convolutional neural network, and the noise branch extracts long-range spectral dependence features through the Transformer encoder. The two branches interact through the attention fusion mechanism. After processing by the fully connected layer and the softmax function, the event classification probability is output, with 4 categories including construction dust, traffic noise, industrial noise, and normal environment. At the same time, the decision factor is output through the decision factor calculation module (implemented by a fully connected layer), which is used to measure the confidence of the model prediction.

[0112] Step 103, input the denoised monitoring data into the CNN-Transformer dual-flow deep neural network model for identification, output the event classification probability and the decision factor; evaluate the environmental state according to the event classification probability and the decision factor, and perform intelligent early warning according to the evaluation result.

[0113] Specifically, the event classification probability in the embodiment at least includes construction dust, traffic noise, industrial noise and normal environment, and the decision factor is used to measure the credibility of the model prediction;

[0114] The classification probability threshold of each category event is set to 0.7, and when the classification probability of the category event is greater than or equal to 0.7, it is considered that the event is more likely to occur;

[0115] The decision factor ranges from 0 to 1, and the decision factor threshold is set to 0.6. When the decision factor is greater than or equal to 0.6, it is considered that the model prediction result is reliable.

[0116] Specifically,

[0117] Dust branch processing: The dust concentration sequence is spliced with 5 meteorological parameters such as temperature and humidity, wind speed and direction, and air pressure into a 100x6 two-dimensional time sequence matrix, and features are extracted layer by layer through 3 layers of hollow convolution layers. The first layer of convolution captures short-term fluctuations of 1-3 seconds (instantaneous wind speed changes affect dust), the second layer expands the perception range to 5-7 seconds through a convolution kernel with an expansion rate of 2, and the third layer captures medium and long-term trends of 30-40 seconds (the cumulative effect of humidity changes on particulate matter deposition) through an expansion rate of 4. Finally, a time sequence feature vector containing multi-scale environmental correlations is output.

[0118] Noise branch processing: Convert the 1-second noise time domain signal into 512-point frequency domain features to form a 100x512 frequency spectrum matrix. After injecting time sequence information through position coding, global dependence modeling is performed through 6 layers of Transformer encoder. Each layer of encoder calculates the self-attention of 8 different subspaces at the same time, capturing the energy distribution correlation of high-frequency noise (mechanical vibration) and low-frequency noise (traffic roar) in long time, and outputting a frequency spectrum feature vector containing cross-band and cross-time point dependence relationship. The feature vectors output by the two branches are fused through a bidirectional attention mechanism: first calculate the attention of the dust features to the noise features (whether high-concentration dust period is accompanied by specific frequency band noise enhancement), then calculate the contribution of the noise features to the dust features in the reverse direction, dynamically adjust the weights of the two, and then splice them into a comprehensive feature vector. Through a fully connected layer, it is mapped to a 4-class space, and the Softmax function is used to output the occurrence probability of 4 types of events (construction dust, traffic noise, industrial noise, and normal environment) (the probability of construction dust is 0.82, indicating that the current scene is 82% likely to be a construction dust event). At the same time, a fully connected layer is separately set to evaluate the credibility of the comprehensive features, outputting a decision factor in the range of 0-1 (the higher the value, the more certain the model prediction), which is used to filter ambiguous judgments with low confidence.

[0119] I. Environmental state evaluation standard

[0120] Dual criteria are used to ensure evaluation reliability:

[0121] Event classification probability: Set the trigger threshold to 0.7 for each of the 4 types of events. When the probability of a certain type of event is ≥0.7, it indicates that the model based on historical data matching, the possibility of the occurrence of the event is significantly higher than other categories (traffic noise probability 0.75, means that the current noise characteristics and historical traffic noise sample matching degree more than 75%).

[0122] Decision factor: Value range 0-1, obtained by optimizing the prediction entropy value during model training. When the decision factor ≥0.6, it means that the model has a clear feature recognition for the current input sample, and the prediction result is reliable (decision factor 0.8 means that the model has low uncertainty for this judgment). Only events that meet both the probability threshold and the decision factor threshold will be included in the final evaluation results, to avoid false positives caused by data anomalies or ambiguous model judgments.

[0123] II. Intelligent early warning mechanism

[0124] Single event immediate warning: When the output result of a certain time window meets the double threshold conditions, the system immediately triggers the corresponding event warning. For example, if the construction dust probability of a window is 0.78 and the decision factor is 0.65, and the monitoring location is near the construction site, it is determined as a construction dust event, and an immediate warning message is generated. Warning information includes event type ("industrial noise anomaly"), occurrence time (accurate to seconds), monitoring node location (through sensor built-in GPS or deployment coordinate positioning), and preliminary severity (based on probability value, divided into "mild" and "moderate", such as probability 0.7-0.8 for mild, 0.8-1 for moderate).

[0125] Multi-window trend analysis and warning escalation: Establish a dynamic verification mechanism for 3 consecutive time windows. If the same monitoring node triggers the same type of event warning (traffic noise) within 150 seconds (3 50-second sliding windows) and the decision factor of each window is ≥0.6, it is determined as a persistent pollution event, and the warning level is upgraded to "severe". The first window triggers a mild traffic noise warning, and the subsequent two windows continue to trigger, the system automatically upgrades the warning to severe, and marks that the event has lasted for 5 minutes, indicating that there may be long-term pollution sources such as traffic congestion or equipment abnormalities.

[0126] Diversified warning methods: According to the warning level, adopt a hierarchical notification strategy:

[0127] Mild warning: Pop-up reminder through the monitoring platform, and send a short message to the on-site manager (including event type, time, location), to facilitate timely troubleshooting.

[0128] Severe warning: In addition to the above methods, start the on-site sound and light alarm device (red light flashing on the construction site + bee sound prompt), and send an email warning to the environmental management department (with event duration, characteristic trend chart), triggering remote supervision intervention. All warning information is stored in the cloud database, supporting historical event query and spatio-temporal distribution analysis (viewing the concentration of construction dust warning in a certain area within a week through GIS map), providing data support for environmental management decisions.

[0129] The beneficial effects are that 1. Compared with the traditional single sensor scheme, the data dimension is improved, and the pollution source characteristics and environmental impact factors are completely covered, providing full-dimensional data support for subsequent analysis, 2. The complex environmental event recognition accuracy is effectively improved, and the construction dust event recognition accuracy is improved. In complex scenes such as construction sites and industrial parks, the comprehensive performance is improved compared with the traditional scheme.

[0130] See Figure 2 In a deep learning-based dust and noise monitoring data intelligent analysis method, initial multi-source monitoring data is obtained, wavelet threshold is used to denoise the dust data in the data, and adaptive frequency band filtering is used to denoise the non-environmental interference in the data, to obtain denoising monitoring data, including the following steps:

[0131] Step 201, synchronously collecting dust concentration data, noise spectrum data and meteorological parameters by multi-source sensors to obtain multi-source monitoring data,

[0132] Step 202, performing outlier detection on the original data of each sensor, judging the time point data of a single sensor, if the value exceeds 3 times the standard deviation of the historical data mean value of the sensor, it is judged as an outlier, and the outlier is marked;

[0133] Step 203, for the marked outliers, using the adjacent time point data interpolation method to repair, to obtain the processed multi-source monitoring data.

[0134] See Figure 3 In a deep learning-based dust and noise monitoring data intelligent analysis method, initial multi-source monitoring data is obtained, wavelet threshold is used to denoise the dust data in the data, and adaptive frequency band filtering is used to denoise the non-environmental interference in the data, to obtain denoising monitoring data, including the following steps:

[0135] Step 301, obtaining the processed multi-source monitoring data, using FFT fast Fourier transform to convert the noise spectrum data in the initial multi-source monitoring data and the time domain signal collected by the noise sensor into frequency domain signal, and performing real-time spectrum analysis to obtain the frequency component of the noise, to obtain the spectrum analysis result;

[0136] Step 302, judging the frequency band range of the environmental interference according to the spectrum analysis result, if the periodic noise generated by the industrial equipment is judged as the frequency band of 50-200Hz, the frequency band is set as the filtering frequency band;

[0137] Step 303, performing adaptive frequency band filtering on the filtering frequency band by using an IIR infinite impulse response filter, setting the passband and stopband parameters of the filter according to the determined frequency band range by using a band-stop filter, performing noise reduction processing on the non-environmental interference, and obtaining the noise reduction monitoring data.

[0138] The basic principle, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent analysis method for dust noise monitoring data, characterized in that, The dust-raising noise monitoring data intelligent analysis method comprises the following steps: Synchronous acquisition of dust-raising concentration data, noise spectrum data and meteorological parameters by multi-source sensors to obtain multi-source monitoring data; Denoising of the dust data in the data by wavelet threshold, and denoising of the non-environmental interference in the data by adaptive frequency band filtering to obtain denoised monitoring data; A dust branch is established by using a 1D-CNN hollow convolutional neural network to extract multi-scale time sequence features, the 1D-CNN hollow convolutional neural network at least comprising an input layer, a hollow convolutional layer, a batch normalization layer and a pooling layer; the input layer is used to receive the dust data and meteorological parameters in the denoised monitoring data, and the dust data and meteorological parameters are spliced into one-dimensional time sequence data; a noise branch is established by using a Transformer encoder to extract long-range spectral dependence, the Transformer encoder at least comprising an input layer, a position encoding layer, a Transformer encoder layer and a normalization layer; the input layer is used to input the denoised noise spectrum data as input, and position information is added to the noise spectrum data based on sine-cosine position encoding; the attention weights of the dust features to the noise features and the attention weights of the noise features to the dust features are calculated by using a bidirectional attention mechanism to obtain a CNN-Transformer double-flow deep neural network model; The denoised monitoring data is input into the CNN-Transformer double-flow deep neural network model for identification, and event classification probability and decision factors are output; the event classification probability at least comprises construction dust, traffic noise, industrial noise and normal environment, the environmental state is evaluated according to the event classification probability and the decision factors, and intelligent early warning is performed according to the evaluation result. 2.The deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, wherein, Further comprising: Abnormal value detection is performed on the original data of each sensor, the time point data of a single sensor is judged, if the value exceeds 3 times the standard deviation of the historical data mean value of the sensor, the value is judged as an abnormal value, and the abnormal value is marked; For the marked abnormal value, interpolation method of adjacent time point data is used for repair to obtain processed multi-source monitoring data. 3.The deep learning-based intelligent analysis method for dust noise monitoring data according to claim 2, wherein, Further comprising: The processed multi-source monitoring data is obtained, the preprocessed dust concentration data is subjected to 3-layer wavelet decomposition by using a db4 wavelet base function to obtain detail coefficients and approximation coefficients of different frequency levels; Threshold values of the detail coefficients of each layer are calculated by using a threshold value, the detail coefficients are processed by using a soft threshold function, and the processed coefficients are subjected to wavelet reconstruction to obtain denoised dust concentration data. 4.The deep learning-based intelligent analysis method for dust noise monitoring data according to claim 2, wherein, Further comprising: The processed multi-source monitoring data is obtained, and the noise spectrum data in the initial multi-source monitoring data and the time domain signal collected by the noise sensor are converted into frequency domain signals by using FFT fast Fourier transform, and real-time spectrum analysis is performed to obtain the frequency components of the noise to obtain a spectrum analysis result; The frequency band range of the environmental interference is judged according to the spectrum analysis result, if the periodic noise generated by the industrial equipment is judged as a 50-200Hz frequency band, the frequency band is set as a filtering frequency band; Adaptively filter the filtered frequency band by using an IIR infinite impulse response filter, and set the passband and stopband parameters of the filter according to the determined frequency band range by using a band-stop filter, to perform noise reduction processing on non-environmental interference, and obtain noise reduction monitoring data.

5. The deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, characterized in that, Further comprising: Three hollow convolutional layers are set, the first layer has an expansion rate of 1, a convolution kernel size of 3, and an output channel number of 64; the second layer has an expansion rate of 2, a convolution kernel size of 3, and an output channel number of 128; and the third layer has an expansion rate of 4, a convolution kernel size of 3, and an output channel number of 256, so as to obtain multi-scale time sequence features through different expansion rates; A batch normalization layer is added after each hollow convolutional layer, and an average pooling layer is added after the last hollow convolutional layer, with a pooling window size of 2.

6. The deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, wherein, Further comprising: Six Transformer encoders are set, each of which includes a multi-head self-attention mechanism and a feedforward neural network, the multi-head self-attention mechanism allows the model to focus on different parts in different representation subspaces and capture long-range spectral dependencies, and a layer normalization layer is added before and after the multi-head self-attention mechanism and the feedforward neural network.

7. The deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, characterized in that, Further comprising: The classification probability threshold of each type of event is set to 0.7, and when the classification probability of a type of event is greater than or equal to 0.7, it is considered that the event is more likely to occur; the decision factor range is [0, 1], and the decision factor threshold is set to 0.6, and when the decision factor is greater than or equal to 0.6, it is considered that the model prediction result is reliable.

8. The system for implementing the deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, characterized in that, The system comprises the following modules: A monitoring data acquisition module is configured to acquire initial multi-source monitoring data, denoise dust data in the data by using a wavelet threshold, and denoise non-environmental interference in the data by using adaptive frequency band filtering, to obtain noise reduction monitoring data; A network model establishment module is configured to extract multi-scale time sequence features by using a 1D-CNN hollow convolutional neural network to establish a dust branch, extract long-range spectral dependencies by using a Transformer encoder to establish a noise branch, and perform cross-modal feature interaction by using an attention fusion mechanism, to obtain a CNN-Transformer dual-flow deep neural network model; A monitoring intelligent analysis module is configured to input the noise reduction monitoring data into the CNN-Transformer dual-flow deep neural network model for identification, output event classification probability and decision factor, and perform environmental state evaluation according to the event classification probability and the decision factor, and perform intelligent early warning according to the evaluation result.

9. The system for implementing the deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, characterized in that, The system comprises the following sub-modules: An acquisition sub-module is configured to synchronously acquire dust concentration data, noise spectrum data, and meteorological parameters by using multi-source sensors, to obtain multi-source monitoring data, A judgment sub-module is configured to perform outlier detection on the original data of each sensor, and perform judgment on the moment data of a single sensor, if the value of the moment data exceeds 3 times the standard deviation of the historical data mean value of the sensor, the value is judged as an outlier, and the outlier is marked; A repair sub-module is configured to repair the marked outliers by using an adjacent time point data interpolation method, to obtain processed multi-source monitoring data.

10. The system for implementing the deep learning-based intelligent analysis method for dust noise monitoring data according to claim 1, wherein, The system comprises the following sub-modules: The decomposition submodule is configured to obtain the processed multi-source monitoring data, perform three-layer wavelet decomposition on the preprocessed dust concentration data by using a db4 wavelet base function, and obtain detail coefficients and approximation coefficients of different frequency levels; The reconstruction submodule is configured to calculate a threshold value of each layer of the detail coefficients by using a threshold value, process the detail coefficients by using a soft threshold function, perform wavelet reconstruction on the processed coefficients, and obtain the de-noised dust concentration data.

Images