Dust monitoring data processing and classification method based on multi-source sensor fusion

Abstract

The present application relates to a kind of dust monitoring data processing and classification method based on multi-source sensing fusion, specifically as follows: first, the data of target area deployment multi-source dust monitoring sensor node is collected, and is labeled, generates dataset;Subsequently, the data collected is carried out continuous wavelet transform, extracts wavelet energy spectrum and shannon entropy, and the enhanced feature tensor is spliced;Then through two-stage fusion and coding strategy, first wavelet packet decomposition is extracted frequency band energy feature, then the time sequence mode coding feature is formed by calculating multi-channel cross correlation, statistical moment, ratio feature, combined with local time sequence feature matrix realizes multi-source heterogeneous feature fusion;Subsequently, the deep learning model containing multi-scale time sequence feature extraction and dynamic fusion module is constructed, and the fusion feature matrix is inputted training;Finally, the preprocessed new monitoring data is inputted into the trained model, and the dust source and pollution grade classification result is outputted.The present application can effectively improve dust monitoring data classification accuracy and dust source identification precision.

Description

Technical Field

[0001] This invention relates to the field of dust monitoring data processing technology, and in particular to a method for processing and classifying dust monitoring data based on multi-source sensor fusion. Background Technology

[0002] With the accelerating pace of urbanization and industrialization, activities such as construction, road traffic, industrial production, and exposed ground surfaces are becoming increasingly frequent, making dust pollution a significant environmental problem affecting urban air quality and residents' health. Particulate matter contained in dust, especially PM2.5 and PM10, can remain suspended in the atmosphere for extended periods and disperse with air currents, easily triggering respiratory and cardiovascular diseases, while also adversely impacting the urban ecological environment and visibility. Therefore, how to conduct refined monitoring of dust pollution, accurately identify dust source types, and promptly assess the degree of pollution is a key technical issue in environmental supervision and pollution prevention.

[0003] Existing technologies objectively suffer from the following shortcomings: Firstly, existing dust monitoring methods largely rely on data from single or limited sensors, lacking a systematic fusion of multi-source sensor information and external environmental context. This makes it difficult to comprehensively reflect the true mechanisms of dust generation and diffusion, resulting in limited accuracy in classification and source tracing. Secondly, existing technologies generally employ raw time series or simple statistical features, lacking in-depth analysis of the multi-scale time-frequency characteristics of signals. This makes it difficult to simultaneously capture short-term emergencies and long-term trends, and insufficient in distinguishing complex dust sources and mixed pollution scenarios. Thirdly, in terms of feature fusion and dimensionality reduction, most methods only use simple splicing or linear dimensionality reduction techniques, failing to effectively model the intrinsic relationships between different sensor channels and different feature types, easily leading to information redundancy or loss of key discriminative information. Fourthly, existing classification models are mostly single-task structures, and their loss functions do not adequately consider class imbalance and temporal consistency, easily biasing towards dust source types with larger sample sizes and producing unstable results in time series prediction, affecting the reliability of practical applications.

[0004] Therefore, this invention proposes a method for processing and classifying dust monitoring data based on multi-source sensor fusion to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention develops a method for processing and classifying dust monitoring data based on multi-source sensor fusion. This invention effectively improves the accuracy of dust monitoring data classification and the precision of dust source identification.

[0006] The technical solution of this invention to solve the technical problem is a method for processing and classifying dust monitoring data based on multi-source sensor fusion, comprising the following steps:

[0007] S1. Deploy multi-source dust monitoring sensor nodes in the target monitoring area. Each node synchronously collects time-series data from multiple sensor channels. The data within a fixed monitoring time window is used as a sample. At the same time, external environmental context information is collected. The collected data is then labeled with pollution level and dust source type, and the data is integrated into a dataset.

[0008] S2. Perform continuous wavelet transform on the original time series data of each sensing channel to convert it to the time-frequency domain. Then extract the wavelet energy spectrum and Shannon entropy from the time-frequency representation and concatenate them with the original time series data to obtain the enhanced feature tensor.

[0009] S3. Perform multi-source heterogeneous feature fusion, specifically adopting a two-stage feature fusion and coding strategy. In the first stage, the original time-series data of the key channels are decomposed into wavelet packet to extract frequency band energy features, resulting in wavelet packet energy features. In the second stage, the cross-correlation, statistical moments, and ratio features between multiple channels are calculated, and the three types of features are concatenated into time-series pattern coding features. Then, a preliminary local transformation is performed on the enhanced feature tensor to obtain a local time-series feature matrix. Finally, the local time-series feature matrix, the time-series pattern coding features, and the wavelet packet energy features are fused to obtain a fused feature matrix.

[0010] S4. Construct and train a deep learning model for dust monitoring data classification. The model includes a multi-scale temporal feature extraction and dynamic fusion module, a dust monitoring data classification and loss function calculation module. Input the fused feature matrix into the model for classification and calculate the loss function. Then iteratively train and update the trainable parameters of the deep learning model to obtain the trained model.

[0011] S5. After preprocessing, the newly collected dust monitoring data is input into the trained model, which outputs the probability distribution of various dust sources and pollution levels. The category with the highest probability is taken as the final result.

[0012] S1 is as follows:

[0013] The monitoring areas include the areas surrounding construction sites, main urban roads, industrial parks, and open areas;

[0014] The data collected by the multi-source dust monitoring sensor nodes include PM2.5 concentration, PM10 concentration, temperature, and humidity. The sampling time interval is set, and the data collection covers different meteorological conditions, seasonal periods, and dust source activity periods. Samples are constructed according to a fixed monitoring time window.

[0015] Simultaneously, the external environmental context information corresponding to each sample is recorded, including the average wind speed, prevailing wind direction, time period, and whether it is a working day during the monitoring period.

[0016] Environmental experts manually labeled each collected sample by combining video recordings from monitoring points, on-site inspection logs, and multi-source data models. The dust source type labels included construction dust, road dust, industrial emissions, and soil wind erosion, and the pollution level labels included light, moderate, and heavy.

[0017] All samples and their corresponding four-channel raw time-series data, external environmental context information, dust source type labels, and pollution level labels are integrated to form a dataset.

[0018] S2 is as follows:

[0019] S2.1. Perform continuous wavelet transform independently on the time-series data of each sensor channel for each sample, mapping the one-dimensional time-series signal into a two-dimensional time-frequency coefficient matrix. The elements in the two-dimensional time-frequency coefficient matrix are represented as follows: , Indicates the first The first sample The element in the m-th row and n-th column of the continuous wavelet transform time-frequency coefficient matrix of each sensing channel. The scale index represents the wavelet transform. A time index representing a time series;

[0020] S2.2 For the time-frequency coefficient matrix of each channel, aggregate the energy at each scale along the time dimension to obtain the total energy intensity of that scale over the entire time range, and compress the dynamic range through logarithmic operation to obtain the wavelet energy spectrum of each sample and each sensing channel at each scale.

[0021] S2.3 Treat the normalized energy of the time-frequency coefficient matrix as a probability distribution, and measure the time-frequency distribution of the channel signal by calculating its Shannon entropy. High entropy value corresponds to noise interference or complex mixed dust source mode, while low entropy value corresponds to dust source event with single mode.

[0022] S2.4. The original time-series data of each sensing channel, the extracted multi-scale wavelet energy spectrum features, and the Shannon entropy features are spliced ​​together in the feature dimension to form an enhanced feature tensor that simultaneously contains time-domain, frequency-domain energy, and complexity information.

[0023] S3 is as follows:

[0024] S3.1, The raw time-series data of the two key particulate matter channels, PM2.5 concentration and PM10 concentration, are processed separately. Layered wavelet packet decomposition is used to obtain multi-layer node coefficients of a tree structure. The energy of each node coefficient is calculated, and the coefficient with the highest energy is selected. The energy of each node constitutes a feature vector;

[0025] Then, the feature vectors of the two key particulate matter channels, PM2.5 concentration and PM10 concentration, are concatenated to obtain the wavelet packet energy feature vector of each sample.

[0026] S3.2 Calculate the time-series pattern coding features by capturing the dynamic correlation and joint statistical characteristics between different sensing channels. The time-series pattern coding features include the maximum cross-correlation coefficient between channels and their lag time, the statistical moments of each channel and the statistics of the particulate matter concentration ratio sequence. The three types of features are spliced ​​together to form the time-series pattern coding feature vector.

[0027] S3.3 Input the enhanced feature tensor into a lightweight one-dimensional convolutional layer to perform preliminary local feature extraction and transformation, and obtain the local temporal feature matrix;

[0028] S3.4. Fuse the local temporal feature matrix with the wavelet packet energy feature vector and the temporal pattern coding feature vector;

[0029] First, the wavelet packet energy feature vector and temporal pattern encoding feature vector are projected onto a unified subspace using independent learnable affine transformations. Then, the projected global features are copied in the time dimension. Then, it is concatenated with the local temporal feature matrix in the feature dimension to obtain the fused feature matrix.

[0030] The specific process for calculating the feature vector of temporal pattern encoding is as follows:

[0031] (1) Calculate the cross-correlation sequence between each pair of sensing channels, take the maximum cross-correlation number and its corresponding leading or lagging relationship, and obtain the cross-correlation characteristics;

[0032] (2) Calculate the four statistical moments of the time series data for each channel, namely mean, variance, skewness and kurtosis, which constitute the statistical moment features and are used to describe the central tendency, dispersion, distribution asymmetry and peak degree of the data in each channel.

[0033] (3) Calculate the point-by-point ratio sequence of PM2.5 concentration to PM10 concentration, and then calculate the mean and variance of the ratio sequence to obtain the 2-dimensional ratio characteristics;

[0034] (4) The cross-correlation feature, statistical moment feature and ratio feature are concatenated to obtain the time-series pattern coding feature vector, which is used to comprehensively characterize the dynamic relationship and statistical characteristics between multiple channels.

[0035] S4 is as follows:

[0036] A deep learning model for dust monitoring data classification was constructed and trained. The fused feature matrix was input into the model, which first passed through a multi-scale temporal feature extraction and dynamic fusion module. This module first used parallel causal convolutional layers with different dilation rates to extract features from micro to macro scales, obtaining a mixed feature matrix. Then, a multi-scale attention mechanism was used to dynamically evaluate and fuse features from different scales, obtaining attention-adjusted scale-related features. Finally, a weighted feature concatenation was used to output the multi-scale fused feature matrix. The dust monitoring data classification and loss function calculation module adopted a classification module that integrates environmental context, performs dual-task prediction, and uses an adaptive multi-objective loss function. This module encodes external meteorological and temporal information, predicts dust source type and pollution level in parallel, and optimizes multi-objective loss. Then, the model was optimized, trained, and its parameters were updated.

[0037] The multi-scale temporal feature extraction and dynamic fusion module is as follows:

[0038] (1) Input the fusion feature matrix into multiple one-dimensional causal convolutional layers with different dilation rates. Each convolutional layer works independently to capture local patterns within the receptive field at a specific time. Then, the outputs of all convolutional layers are concatenated in the feature dimension to form a hybrid feature containing multi-scale information.

[0039] (2) The attention mechanism is used to dynamically fuse features at different scales. The attention mechanism first calculates the global context coefficient for each time step, and then calculates the importance weight of each scale feature at that time step based on the global context coefficient and the local features at each scale.

[0040] Specifically, global average pooling is first applied to the hybrid feature matrix along the feature dimension to obtain the global context coefficients at each time step; then, the hybrid feature matrix is ​​divided into multiple groups according to the inflation rate of the source, and for each time step and each inflation rate, a learnable attention network is used to calculate its attention weights.

[0041] (3) Use attention weights to weight the original multi-scale features step by step and scale by scale to obtain the scale-related features after attention adjustment;

[0042] (4) All the weighted features corresponding to the expansion rate are concatenated again in the feature dimension to obtain a weighted feature matrix. Then, nonlinear transformation and feature selection are performed through the gated activation unit to obtain the multi-scale fusion feature matrix output by multi-scale temporal feature extraction and dynamic fusion.

[0043] The dust monitoring data classification and loss function calculation module is as follows:

[0044] (1) A lightweight external feature encoder is used to encode wind speed, wind direction, time period and weekday information into feature vectors, which are then concatenated with the multi-scale fusion feature matrix in the feature dimension to form a classification feature matrix;

[0045] (2) A dual-task classification head is used for classification prediction. The main task predicts the dust source type and the auxiliary task predicts the pollution level. The two tasks share the main feature extraction backbone but have independent classifiers. More general and discriminative features are captured through joint learning.

[0046] Specifically, firstly, key time step information is aggregated through attention-weighted global pooling to obtain attention-weighted global feature vectors. Then, the global feature vectors are input into two independent classifiers to obtain dual-task prediction probability distributions, including dust source type prediction probability distributions and pollution level prediction probability distributions.

[0047] (3) A multi-objective adaptive loss function is adopted, including category adaptive weighted cross-entropy loss, standard cross-entropy loss and temporal smoothness consistency loss.

[0048] The training process of the model is as follows:

[0049] The deep learning model is iteratively trained and all trainable parameters are updated using data from the dataset. In each training iteration, a batch of sample data is randomly selected from the dataset and forward propagation is performed sequentially. After preprocessing in steps S2 and S3, the data is input into the model to obtain the predicted probability distribution of dust source type and pollution level of the current batch of samples. Then, the total loss function is calculated.

[0050] Then, backpropagation is performed to calculate the gradient of the total loss with respect to all trainable parameters in the model. An adaptive moment estimator optimizer is used to update all these trainable parameters based on the calculated gradients, minimizing the total loss function.

[0051] Set an iteration stopping condition, and when the condition is met, end the training to obtain the trained model.

[0052] S5 is detailed below:

[0053] For new monitoring data, after preprocessing, it is input into the trained model, which outputs the probability distribution of various dust sources and the probability distribution of each pollution level. The category with the highest probability is taken as the final dust source type classification result and pollution level determination result. Based on the dust source type classification result, the dust source types in different regions are quickly identified and traced for analysis, and graded early warning and targeted control are carried out according to the pollution level.

[0054] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0055] This invention discloses a method for processing and classifying dust monitoring data based on multi-source sensor fusion. It employs a multi-source sensor fusion modeling approach for dust monitoring, systematically fusing multi-channel time-series data of PM2.5, PM10, temperature, and humidity with environmental context information such as wind speed, wind direction, time period, and weekday. This achieves a comprehensive characterization of dust formation and diffusion processes, significantly improving the ability to depict complex dust source scenarios. Furthermore, it utilizes a multi-scale time-frequency feature enhancement mechanism combining continuous wavelet transform and wavelet packet decomposition. While preserving the original time-series information, it further extracts multi-scale energy spectra, Shannon entropy, and fine-band energy features, enabling the model to simultaneously perceive instantaneous disturbances and periodicity. The study of changes and long-term trends improves the separability between different dust sources. An adaptive fusion and dimensionality reduction coding strategy using multi-source heterogeneous features is employed. Through joint modeling of local temporal convolution, global statistics, and frequency band features, combined with a learnable weighted fusion mechanism, the feature dimensions are effectively compressed while maintaining physical meaning, avoiding information redundancy and enhancing the expression of key features. A deep classification model integrating multi-scale causal convolution, dynamic attention mechanism, and dual-task learning is used, combined with a multi-objective loss function that incorporates adaptive weighting of categories and temporal smoothness consistency constraints. This achieves joint prediction of dust source type and pollution level, maintaining high stability and robustness even under conditions of category imbalance and temporal fluctuations. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0057] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0058] Figure 2 The image shows the energy distribution heatmaps corresponding to different dust sources for the multi-scale time-frequency feature extraction method of the present invention.

[0059] Figure 3 The time-frequency Shannon entropy corresponds to the four dust source signals.

[0060] Figure 4 This is a scatter plot showing the distribution of different dust source samples in the feature space after feature dimensionality reduction. Detailed Implementation

[0061] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0062] Example 1

[0063] like Figure 1As shown, a method for processing and classifying dust monitoring data based on multi-source sensor fusion includes the following steps:

[0064] S1. Deploy multi-source dust monitoring sensor nodes in the target monitoring area. Each node synchronously collects time-series data from multiple sensor channels. The data within a fixed monitoring time window is used as a sample. At the same time, external environmental context information is collected. The collected data is then labeled with pollution level and dust source type, and the data is integrated into a dataset.

[0065] S2. Perform continuous wavelet transform on the original time series data of each sensing channel to convert it to the time-frequency domain. Then extract the wavelet energy spectrum and Shannon entropy from the time-frequency representation and concatenate them with the original time series data to obtain the enhanced feature tensor.

[0066] S3. Perform multi-source heterogeneous feature fusion, specifically adopting a two-stage feature fusion and coding strategy. In the first stage, the original time-series data of the key channels are decomposed into wavelet packet to extract frequency band energy features, resulting in wavelet packet energy features. In the second stage, the cross-correlation, statistical moments, and ratio features between multiple channels are calculated, and the three types of features are concatenated into time-series pattern coding features. Then, a preliminary local transformation is performed on the enhanced feature tensor to obtain a local time-series feature matrix. Finally, the local time-series feature matrix, the time-series pattern coding features, and the wavelet packet energy features are fused to obtain a fused feature matrix.

[0067] S4. Construct and train a deep learning model for dust monitoring data classification. The model includes a multi-scale temporal feature extraction and dynamic fusion module, a dust monitoring data classification and loss function calculation module. Input the fused feature matrix into the model for classification and calculate the loss function. Then iteratively train and update the trainable parameters of the deep learning model to obtain the trained model.

[0068] S5. After preprocessing, the newly collected dust monitoring data is input into the trained model, which outputs the probability distribution of various dust sources and pollution levels. The category with the highest probability is taken as the final result.

[0069] In a specific implementation, S1 is as follows:

[0070] Multi-source dust monitoring sensor nodes are deployed in the target monitoring area (such as around construction sites, main urban roads, industrial parks, and open areas). Each node simultaneously collects time-series data from four sensor channels: PM2.5 concentration, PM10 concentration, temperature, and humidity. The sampling time interval is as follows: The default setting is 1 minute to ensure that minute-level concentration change details are captured. Data collection covers different meteorological conditions, seasonal periods, and dust source activity periods to fully reflect the diversity and representativeness of the data.

[0071] To construct the dataset for model training and validation, each sample is constructed based on a fixed monitoring time window, such as a continuous 24-hour period (i.e., the time series length of a single sample). The synchronous four-channel data (at each time point) is used as a sample instance.

[0072] While collecting raw time-series data, it is necessary to simultaneously record the external environmental context information corresponding to each sample, including the average wind speed, prevailing wind direction, time period (such as early morning, morning, afternoon, or night), and whether it is a working day.

[0073] For each collected sample, environmental experts need to manually label it based on video recordings from monitoring points, on-site inspection logs, and multi-source data patterns. In one embodiment, the labeling includes two types of tags: one is the dust source type tag, which mainly includes construction, road dust, industrial emissions, and soil wind erosion; the other is the pollution level tag, which is classified into light, moderate, and heavy levels according to the average concentration level of PM2.5 or PM10 within the monitoring window, with reference to the national ambient air quality standards.

[0074] All samples and their corresponding four-channel raw time-series data, external environmental context information, dust source type labels, and pollution level labels are integrated and divided into training set, validation set, and test set, thereby completing the construction of the dataset and providing a foundation for subsequent model training and evaluation.

[0075] In a specific implementation, S2 is as follows:

[0076] Dust monitoring data based on multi-source sensors exhibits significant temporal correlation, local fluctuations, and multi-scale characteristics. Different dust sources show varying patterns at different time scales. Conventional techniques typically use the original sequence directly or perform simple moving averages, making it difficult to simultaneously capture instantaneous high-frequency events and long-term trends. Furthermore, these techniques are sensitive to sensor noise, making it difficult for classification models to learn discriminative multi-scale features. This invention performs continuous wavelet transform on the raw time-series data of each sensor channel, converting it to the time-frequency domain to simultaneously reveal the frequency domain characteristics of the signal and its changes over time. Then, wavelet energy spectrum and Shannon entropy are extracted from the time-frequency representation as enhancing features. This integrates multi-scale frequency domain energy distribution and signal complexity information while preserving the original time-series information. The specific steps are as follows:

[0077] 1) Generation of time-frequency matrix using multi-channel continuous wavelet transform

[0078] Continuous wavelet transform is performed independently on the time-series data of the four sensor channels for each sample, mapping the one-dimensional time-series signal into a two-dimensional time-frequency coefficient matrix to capture the joint distribution of the signal's energy evolution over time at multiple scales. The continuous wavelet transform representation for the sensor reading of a single sensor channel is as follows:

[0079]

[0080] In the formula, Indicates the first The first sample The element in the m-th row and n-th column of the continuous wavelet transform time-frequency coefficient matrix of each sensing channel; this element value represents the original signal at the scale index. Corresponding frequency band and time index The time-frequency joint energy distribution intensity at the corresponding moment; For sample index For sensor channel indexing, These correspond to four sensing channels: PM2.5, PM10, temperature, and humidity. For scale indexing, ; The total number of wavelet transform scales corresponds to different resolution levels for frequency analysis of the signal, and is preferably set to 64. For time indexing, ; To distinguish from Time index, ; The time series length of a single sample, i.e., the total number of time points; For the first The scale parameters corresponding to each scale index; For the first The first sample Each sensor channel at a specific time point The sensor readings come from the first The first sample Timing data of each sensor channel ,Right now, ; wavelet mother function Considering the complex conjugate of the Morlet wavelet, which has good energy concentration in both the time and frequency domains, the Morlet wavelet is used as the wavelet mother function. This represents the sampling time interval.

[0081] In practice, the specific value of the time series length T of a single sample depends on the monitoring time window. For example, if the sampling time interval... Minutes, and each sample represents 24 hours of data, then Typically, T can be set to a fixed value, such as 60 (1 hour) or 1440 (24 hours), depending on the application requirements to ensure time sequence consistency; since dust concentration usually changes slowly, the sampling time interval... The default setting is 1 minute. Minute-level sampling balances timing details and data volume, while being compatible with environmental monitoring standards and facilitating multi-source sensor synchronization; The scale parameters corresponding to each scale index Obtained through logarithmically spaced sampling to better match the broadband characteristics of dust signals, the calculation method is expressed as follows: ,in, This represents the smallest scale parameter, corresponding to the highest frequency in the analysis. This represents the maximum scale parameter, corresponding to the lowest frequency in the analysis. For example, let... , , ,but Uniformly distributed on a logarithmic scale, covering multiple resolution levels from 1 to 128.

[0082] It should be noted that, The term serves as an energy normalization factor to ensure the comparability of wavelet transform energy at different scales and avoid amplitude deviations introduced by scale variations. The preferred setting of the total number of wavelet transform scales is 64, which is based on the frequency band characteristics of typical dust monitoring data. By using logarithmic scale sampling to cover a wide frequency range from high-frequency noise to low-frequency trends, the completeness of multi-scale feature extraction is ensured, while avoiding excessive computational complexity. In practical applications, the total number of wavelet transform scales can be adjusted according to the sampling frequency. For example, when the sampling frequency is high, the total number of wavelet transform scales can be appropriately increased.

[0083] 2) Calculate the multi-scale wavelet energy spectrum

[0084] For the time-frequency coefficient matrix of each channel, the energy at each scale is aggregated along the time dimension to obtain the total energy intensity of that scale over the entire time range. Then, the dynamic range is compressed through logarithmic operations to enhance numerical stability, expressed as:

[0085]

[0086] In the formula, To indicate the first The first sample The first sensing channel in the first The wavelet energy spectrum at each scale is a scalar, characterizing the signal of that channel at different scales. The total energy intensity within the corresponding frequency band reflects the significance of the signal components in that frequency band; It is a logarithmic function, with the default base being the natural constant.

[0087] It should be noted that the numerical energy spectrum values ​​are adopted. The original energy value is compressed in the form of [formula] to reduce the impact of extreme values ​​and ensure numerical stability, while avoiding taking the logarithm of input values ​​that may be zero.

[0088] 3) Calculate the Shannon entropy of the time-frequency matrix.

[0089] Treating the normalized energy of the time-frequency coefficient matrix as a probability distribution, its Shannon entropy is calculated to measure the complexity and uncertainty of the time-frequency distribution of the channel signal. High entropy values ​​may correspond to noise interference or complex mixed dust source patterns, while low entropy values ​​correspond to dust source events with a single pattern, expressed as:

[0090]

[0091] In the formula, Indicates the first The Shannon entropy of the time-frequency matrix of a sensing channel is a scalar used to quantify the degree of disorder or information content of the time-frequency energy distribution of the signal in that channel. The higher the entropy value, the more dispersed and uncertain the time-frequency energy distribution is. The normalized time-frequency energy probability is represented by , where represents the probability of the i-th sample and channel. In the entire time-frequency matrix, located at scale and time The proportion of energy at a given point to the total time-frequency energy is calculated as follows: ; Represents the logarithmic function with base 2; Indicates difference from Scale index, ; Indicates difference from and Time index, ; Indicates the first The first sample The first wavelet transform time-frequency coefficient matrix of the nth sensing channel Line number Column elements.

[0092] 4) Construct an enhanced feature tensor for multi-feature fusion

[0093] The raw time-series data from the four sensor channels, the extracted multi-scale wavelet energy spectrum features, and the Shannon entropy features are concatenated along the feature dimension to form an enhanced feature tensor that simultaneously contains time-domain, frequency-domain energy, and complexity information. Specifically, the raw data matrix of the i-th sample is defined as follows: , dimension Each column corresponds to the timing data of the sensing channel;

[0094] According to the The first sample The first sensing channel in the first Wavelet energy spectrum at various scales The calculation method involves obtaining the wavelet energy spectrum matrix of the four channels, expanding the four channels into row vectors in channel order to obtain the expanded wavelet energy spectrum eigenvectors; combining the Shannon entropy of the time-frequency matrices of the four sensing channels to obtain the Shannon entropy eigenvectors; and then... The original data matrix, length The expanded wavelet energy spectrum eigenvectors are replicated in the time dimension. The matrix formed in this step, and the Shannon entropy eigenvector of length 4, are replicated in the time dimension. The matrices formed in this way are connected along the feature dimensions to form the enhanced feature tensor of the i-th sample. , dimension The enhancement feature tensor is constructed by sequentially concatenating the original temporal features, multi-scale wavelet energy spectrum features, and time-frequency matrix Shannon entropy features along the feature dimension. This enhances the discriminative power of the features and helps the model learn multi-scale dust source patterns.

[0095] In a specific implementation, S3 is as follows:

[0096] The enhanced feature tensor has high dimensionality and contains heterogeneous features. However, this feature set does not consider the physical coupling relationships between different sensing channels and more refined frequency band information. Conventional feature fusion methods, such as simple concatenation or principal component analysis, fail to achieve effective information interaction and compression; the former may destroy the physical meaning of the features and lose crucial classification details. This invention proposes a two-stage feature fusion and encoding strategy: In the first stage, wavelet packet decomposition is performed on key channels to extract more refined frequency band energy features. In the second stage, the temporal cross-correlation and statistical moments between multiple channels are calculated to encode their joint behavior patterns. Finally, through a learnable adaptive weighting layer, the original enhanced features, refined frequency band features, and temporal pattern features are projected and fused to generate a low-dimensional fused feature matrix that combines information density and physical meaning. The specific steps are as follows:

[0097] 1) Fine-grained frequency band energy feature extraction

[0098] The raw time-series data of the two key particulate matter channels, PM2.5 and PM10, were respectively analyzed. Layered wavelet packet decomposition is used to obtain multi-layer node coefficients of a tree structure. The energy of each node coefficient is calculated, and the coefficient with the highest energy is selected. The energy of each node forms a feature vector to capture the fine-band signal intensity associated with specific dust source activity. Specifically, for the PM2.5 or PM10 channel data of the i-th sample, its wavelet packet decomposition... Layer The energy calculation method for each node is expressed as follows:

[0099]

[0100] In the formula, Indicates the first For each sample in the specified particle channel, wavelet packet decomposition is performed on the _th _ ... Layer The average energy corresponding to each node is a scalar, representing the energy intensity of the particulate matter concentration signal within the specific fine frequency band represented by that node. This represents the layer index of the wavelet packet decomposition. ; This represents the total number of decomposition levels; an example value is 5. Indicates the first Node index in the layer, ; Indicates the first For each sample in the specified particle channel, wavelet packet decomposition is performed on the _th _ ... Layer The coefficient sequence of each node; Indicates the first For each sample in the specified particle channel, wavelet packet decomposition is performed on the _th _ ... Layer The coefficient sequence of the nth node One element; Indicates the index of an element within the coefficient sequence. ; Represents the coefficient sequence The length.

[0101] In the specific implementation process, after performing the above calculations for PM2.5 and PM10 channels respectively, the channel with the highest energy value is selected from the average energy set of all nodes in each channel. These constitute the feature vector of the channel.

[0102] It should be noted that specifying a particulate matter channel means processing the raw time-series data of the two key particulate matter concentration sensing channels, PM2.5 and PM10, independently. That is, when calculating the wavelet packet decomposition energy features, wavelet packet decomposition needs to be performed separately on the sensor data of the PM2.5 channel and the PM10 channel to extract the fine frequency band energy features of each channel.

[0103] By concatenating the feature vectors of the two key particulate matter channels, PM2.5 and PM10, we obtain the first... Wavelet packet energy feature vector of each sample , dimension ,in, Indicates the number of nodes selected. It should be less than the total number of nodes generated by wavelet packet decomposition; an example value is 10.

[0104] 2) Multi-channel timing pattern coding

[0105] To capture the dynamic correlation and joint statistical characteristics between different sensing channels, a set of time-series pattern coding features is calculated. These features include the maximum cross-correlation coefficient and its lag time between channels, the statistical moments of each channel, and the statistics of the particulate matter concentration ratio sequence. Specifically, for the i-th sample, the following three types of features are calculated and concatenated to form a time-series pattern coding feature vector, as follows:

[0106] a) Cross-correlation characteristics: Calculate the cross-correlation coefficient sequence between each pair of the four sensing channels, and take the maximum cross-correlation coefficient. and their corresponding leading or lagging relationships This reflects the similarity and phase difference of signals between channels;

[0107] Based on this, the four sensing channels are paired up into six pairs, each pair including the maximum cross-correlation number and its corresponding leading or lagging relationship, resulting in 12-dimensional cross-correlation features.

[0108] in, Indicates channel and channels The maximum cross-correlation coefficient between time series data, taking values ​​in the range of Between these values, a larger absolute value indicates a stronger linear correlation, specifically calculated through the channel. and Obtaining the cross-correlation coefficient sequence of time series data This indicates the index of the first sensing channel, distinct from 'c'. These correspond to PM2.5, PM10, temperature, and humidity, respectively. This indicates the index of the second sensing channel, distinct from 'c'. These correspond to PM2.5, PM10, temperature, and humidity, respectively. This indicates that the maximum number of cross-correlation numbers has been reached. At that time, the channel Relative to channel The lag time reflects the leading or lagging relationship between two signals in time.

[0109] It should be noted that the time series data in the cross-correlation feature calculation refers to the original time series data of each sample, that is, the original time series of each sensing channel.

[0110] In the specific implementation process, the leading or lagging relationship Specifically, this is obtained by calculating the cross-correlation function sequence of the time-series data from two channels. First, the cross-correlation coefficient is calculated over a series of lag times. Then, the time at which the absolute value of the cross-correlation coefficient reaches its maximum value is identified. ,like , indicating channel Changes lag behind channel ,like This indicates a channel. Leading the way .

[0111] b) Statistical Moments Characteristics: Calculate four statistical moments for the time series data of each channel, namely: mean... ,variance skewness kurtosis It is used to describe the central tendency, dispersion, asymmetry of distribution, and peak degree of data in each channel;

[0112] The mean, variance, skewness, and kurtosis of each channel yield a statistical moment feature of 4 channels × 4 moments = 16 dimensions;

[0113] in, Indicates the first The mean of time series data for each channel; Indicates the first The variance of time-series data for each channel; Indicates the first Skewness of time series data for each channel; Indicates the first Kurtosis of time series data for each channel.

[0114] c) Ratio characteristics: Calculate the point-by-point ratio series of PM2.5 to PM10, and then calculate the mean of this ratio series. and variance This reflects the ratio of fine particulate matter to coarse particulate matter and its fluctuations.

[0115] mean and variance Composed of 2D ratio characteristics;

[0116] It represents the mean of the point-by-point ratio sequence of PM2.5 to PM10, reflecting the average ratio of fine particulate matter to coarse particulate matter, which helps to identify the type of dust source; These represent the variances of the point-by-point ratio series of PM2.5 and PM10, respectively, reflecting the degree of fluctuation in the ratios. Large fluctuations may indicate mixed dust sources or rapidly changing events.

[0117] By concatenating the cross-correlation feature, statistical moment feature, and ratio feature, we obtain the first... Temporal pattern encoded feature vector of each sample With a dimension of 30, it is used to comprehensively characterize the dynamic relationships and statistical properties between multiple channels.

[0118] 3) Preliminary local transformation of enhanced features

[0119] The enhanced feature tensor is input into a lightweight one-dimensional convolutional layer for preliminary local feature extraction and transformation to extract basic local temporal patterns, represented as:

[0120]

[0121] In the formula, Indicates the first The local temporal feature matrix of the sample is the _th ... The output matrix of the enhanced features of each sample after transformation by a one-dimensional convolutional layer has a dimension of . It retains the original number of time steps. However, the high-dimensional enhanced features are derived from... Dimensional compression or transformation to This allows for preliminary dimensionality reduction and local feature abstraction. This represents a one-dimensional convolution operation that extracts local temporal patterns by performing sliding window computation in the time dimension. The weight parameter matrix of a one-dimensional convolutional layer is a trainable parameter. This represents the learnable bias parameter vector of a one-dimensional convolutional layer; these are trainable parameters. This represents the feature dimension of the output of the one-dimensional convolutional layer. It is a preset hyperparameter that determines the dimension of the features after the initial transformation. An example value is 64.

[0122] 4) Adaptive weighted feature fusion

[0123] The local temporal feature matrix is ​​fused with the wavelet packet energy feature vector and the temporal pattern coding feature vector. Specifically, the wavelet packet energy feature vector and the temporal pattern coding feature vector are first projected onto a unified subspace through independent learnable affine transformations. Then, the projected global features are copied in the time dimension. This is then combined with the local temporal feature matrix along the feature dimension to achieve time-step fusion of global features and local temporal feature matrices, represented as:

[0124]

[0125] In the formula, Indicates the first The fusion feature matrix of each sample has a dimension of . It integrates local time series, fine frequency band and global statistical pattern information; Indicates a splicing operation; The weight matrix used for the affine transformation of wavelet packet features is a trainable parameter. The bias vector used for the affine transformation of wavelet packet features is a trainable parameter. This represents the weight matrix used for the affine transformation of temporal pattern coding features, and is a trainable parameter; The bias vector represents the affine transformation used for temporal pattern coding features and is a trainable parameter; This means copying the vector in the time dimension. Second-rate; This represents the dimension after the wavelet packet feature is projected; it is a hyperparameter with an example value of 16. This represents the dimension of the time-series pattern encoded features after projection; it is a hyperparameter with an example value of 16.

[0126] S4 is detailed below, including model construction and training:

[0127] In a specific implementation, the multi-scale temporal feature extraction and dynamic fusion module is as follows:

[0128] The fused feature matrix contains rich information from the time domain, frequency domain, and statistical patterns. However, its temporal dependencies are complex and multi-scale. Conventional convolutional neural networks are limited by fixed-size convolutional kernels, making it difficult to effectively capture long-range dependencies. Recurrent neural networks, while capable of processing sequences, suffer from low computational efficiency and gradient vanishing problems, and cannot dynamically assess the importance of features at different time scales to the current classification task. This invention employs a core feature extraction module combining hybrid dilated causal convolution and multi-scale attention. This module first uses a set of parallel causal convolutional layers with different dilation rates to efficiently extract multi-scale temporal patterns from micro to macro levels while ensuring temporal causality. Then, a multi-scale attention mechanism is used to dynamically evaluate and fuse features from different scales, enabling the model to adaptively focus on the scale information most relevant to the current context. The specific steps are as follows:

[0129] 1) Parallel multi-scale causal convolution feature extraction

[0130] The fused feature matrix is ​​simultaneously input into multiple one-dimensional causal convolutional layers with different dilation rates. Each convolutional layer operates independently, responsible for capturing local patterns within its receptive field at a specific time. The outputs of all convolutional layers are concatenated along the feature dimension to form a hybrid feature containing multi-scale information, represented as:

[0131]

[0132] In the formula, The mixed feature matrix represents the mixed feature matrix extracted by parallel multi-scale convolution, with a dimension of [dimension value missing] for a single sample. It includes various time-series patterns ranging from fine-grained to coarse-grained; Indicates the expansion rate The output feature matrix of the causal convolutional layer has a dimension of [dimensionality missing] for a single sample. This characterizes the local temporal pattern of the original signal at a specific time scale, and the calculation method for a single sample is expressed as follows: ; The expansion rate is a predefined set. The elements in the kernel determine the spacing between elements when processing a sequence. For standard convolution, It can exponentially expand the receptive field without increasing parameters or the number of layers; Represents a predefined set of expansion rates, for example The set contains different inflation rates, enabling the model to capture patterns across different time spans in parallel. Represents the set of expansion rates The size of the number of parallel convolution branches; Indicates the corresponding expansion rate The weight parameters of a one-dimensional causal convolutional layer are trainable parameters; Indicates the expansion rate for One-dimensional causal convolution operation, causality ensures time step The output depends only on Inputting information in real time and prior to prevent future information leaks; This is the number of channels output by each convolutional layer; an example value is 32. This indicates a concatenation operation along the feature dimension.

[0133] 2) Multi-scale attention weight calculation

[0134] An attention mechanism is used to dynamically fuse features at different scales. The attention mechanism first calculates the global context coefficients at each time step, and then calculates the importance weights of each scale feature at that time step based on the global context coefficients and the local features at each scale.

[0135] Specifically, firstly, global average pooling is applied to the hybrid feature matrix along its feature dimensions to obtain the global context coefficients for each time step. These coefficients represent the aggregated information of all scale features at that moment, expressed as:

[0136]

[0137] In the formula, Indicates time step The global context coefficient is a scalar that integrates the average information of all multi-scale features at that moment and serves as a reference benchmark for calculating attention weights. Indicates the feature dimension index. .

[0138] The hybrid feature matrix is ​​divided according to the inflation rate of its source. For each group, and for each time step and each inflation rate, its attention weights are computed using a learnable attention network, denoted as:

[0139]

[0140] In the formula, Indicates at time step For expansion rate The attention weights assigned to the scale features, with values ​​between 0 and 1, represent the scale features at time step [number of steps]. ,scale The importance of a feature to the current context is considered; a higher weight indicates that the feature at that scale is more critical. This represents the attention scoring function, which calculates the relevance score between the current-scale feature and the global context. A higher score indicates that the feature at that scale is more important at the current time step. The calculation method is expressed as follows: ; Represents the hyperbolic tangent activation function; Represents the natural exponential function; Indicates at time step , from the inflation rate The feature vector of the convolution branch has a dimension of [dimensionality] for a single sample. This represents local information at a specific scale at that moment, and is the expansion rate. The output feature matrix of the causal convolutional layer The Line, that is, from the inflation rate Time steps extracted from convolution branches eigenvectors; The weight matrix, representing the weights used to transform local feature vectors in the attention network, is a trainable parameter. The weight vector used to transform the global context scalar in the attention network is a trainable parameter. The bias vector representing the attention network is a trainable parameter; The weight vector in the attention network, used to map the transformed joint representation to scalar scores, represents trainable parameters. express The transpose of .

[0141] 3) Adaptive weighting and fusion of scale features

[0142] Attention weights are used to weight the original multi-scale features step-by-step and scale-by-scale. The weighted features emphasize important scale information and suppress unimportant scale information, as shown below:

[0143]

[0144] In the formula, Indicates at time step After attention weighting, the value derived from the inflation rate The feature vectors of the branches represent the scale-related features after attention adjustment.

[0145] 4) Weighted feature splicing and gated activation output

[0146] The weighted features corresponding to all expansion rates are concatenated along the feature dimension to obtain a weighted feature matrix. This matrix is ​​then subjected to nonlinear transformation and feature selection via a gated activation unit to obtain the feature matrix for multi-scale temporal feature extraction and dynamic fusion, as shown below:

[0147]

[0148] In the formula, This represents the multi-scale fusion feature matrix, which is the output feature matrix of the multi-scale temporal feature extraction and dynamic fusion module. The dimension for a single sample is... It contains deep temporal features refined through multi-scale extraction, dynamic attention weighting, and nonlinear gating; This represents the output feature dimension of the multi-scale fused feature matrix, which is set to [value] by default. ; The weighted feature matrix is ​​a concatenated matrix obtained by weighting all scaled features, with a dimension of [dimension value missing] for a single sample. The multi-scale temporal information after dynamic fusion is represented by the following calculation method: ; Indicates the expansion rate The weighted feature matrix of all time steps, i.e. , dimension ; Indicates at time step After attention weighting, the value derived from the inflation rate eigenvectors of the branch transpose; This represents the Sigmoid activation function, and its output value is in Between them, gating signals are used to control the flow of information; The weight matrix representing the gated activation unit used to generate the gated signal is a trainable parameter. This represents the bias vector used by the gated activation unit to generate the gated signal, and it is a trainable parameter; The weight matrix represents the gated activation unit used to perform nonlinear transformations on the features, and is a trainable parameter; The bias vector representing the gated activation unit used to perform nonlinear transformations on the features is a trainable parameter. This indicates element-wise multiplication.

[0149] It should be noted that in the process of calculating the multi-scale fusion feature matrix, the output of the Sigmoid gate signal determines the proportion of features after hyperbolic tangent function transformation that are allowed to pass through, thereby achieving adaptive information filtering.

[0150] In a specific implementation, the dust monitoring data classification and loss function calculation module is as follows:

[0151] The output feature matrix of the multi-scale temporal feature extraction and dynamic fusion module contains deep temporal dependencies and pattern information, but lacks environmental context and pollution severity information crucial for dust classification. Furthermore, conventional classification methods employ a single global pooling and fully connected layer, making it difficult to simultaneously model the overall sequence pattern and key local events. Additionally, dust data suffers from class imbalance, with significant differences in the number of samples from different dust sources. Conventional cross-entropy loss biases towards the majority class, affecting the model's ability to identify a few dust sources, such as specific industrial emissions. This invention employs a classification module that integrates environmental context, performs dual-task prediction, and uses an adaptive multi-objective loss function. By encoding external meteorological and temporal information, predicting dust source types and pollution levels in parallel, and optimizing multi-objective loss, it improves classification performance and robustness. The specific steps are as follows:

[0152] 1) Contextual Feature Concatenation

[0153] Considering that dust generation and dispersion heavily depend on external meteorological conditions and time factors, a lightweight external feature encoder is adopted to encode wind speed, wind direction, time period, and weekday information into feature vectors. These vectors are then concatenated with a multi-scale fusion feature matrix along the feature dimension to form a classification feature matrix. The calculation method for the environmental context feature vector is as follows:

[0154]

[0155] In the formula, Indicates the first The environmental context feature vector of each sample has a dimension of . It is a low-dimensional representation that integrates multiple external factors, used to provide environmental background for time-series features and enhance the model's understanding of environmental factors related to dust source formation and diffusion; Indicates the first The average wind speed of each sample, after standardization, is a scalar quantity used to eliminate the influence of dimensions. Indicates the first The encoding vector of the dominant wind direction for each sample. If the wind direction is regarded as a continuous angle, the encoding is a two-dimensional vector composed of its sine and cosine values. If it is regarded as a categorical variable, one-hot encoding is used. Indicates the first The one-hot encoded vector of each sample for each time period, with a dimension equal to the total number of time period categories; Indicates the first The one-hot encoded vector of whether a sample is a weekday or not, with a dimension of 2; The weight matrix representing the external feature encoder is a trainable parameter. The bias vector representing the external feature encoder is a trainable parameter; This represents the modified linear unit activation function, used to introduce nonlinear transformation capabilities.

[0156] In practice, the lightweight external feature encoder is structured as a simple fully connected neural network layer with trainable weight parameters of [value missing]. The trainable bias parameters are It employs a modified linear unit activation function, whose input is the concatenated external feature vector, and whose output is a low-dimensional environmental context feature vector. The lightweight external feature encoder has a lightweight structure and can perform effective nonlinear feature transformations; if the first The average wind speed reading of the sample over 24 hours is 2.5 m / s, while the average wind speed of the entire training dataset is 2.0 m / s with a standard deviation of 1.0 m / s. Therefore, the... The standardized value of the average wind speed of each sample Furthermore, if wind direction is divided into 8 directions, the first... The prevailing wind direction for the sample is "east," corresponding to category 3. Therefore, the... Encoding vector of the prevailing wind direction for each sample It is an 8-dimensional vector, with the third element being 1 and the rest being 0; and the time period is determined according to the start time of the sample. For example, a day is divided into 4 time periods, including early morning (0-6 am), morning (6-12 am), afternoon (12-18 pm), and night (18-24 pm). If the time period of the sample is "morning", then the first sample... One-hot encoded vector of each sample for a given time period It is a 4-dimensional vector, with the second element being 1; and, it determines whether it is a working day based on the sample collection date, for example, if the first element is 1... If the first sample is collected on a weekday, then the second sample... One-hot encoded vector of whether a sample is a weekday or not If it is a non-working day, then the first... One-hot encoded vector of whether a sample is a weekday or not .

[0157] Copy the environmental context feature vector in the time dimension After that, it is concatenated with the multi-scale fused feature matrix to obtain the classification feature matrix, and the th feature matrix is ​​defined. Classification feature matrix of each sample , dimension It also contains deep multi-scale temporal patterns and environmental context information, providing a comprehensive basis for classification decisions.

[0158] 2) Dual-task classification head prediction

[0159] To fully utilize the rich features learned and improve the model's discriminative power, a dual-task classification head is used for classification prediction. The main task predicts the dust source type, and the auxiliary task predicts the pollution level. The two tasks share the main feature extraction backbone but have independent classifiers. Through joint learning, more general and discriminative features are captured.

[0160] Specifically, key time step information is first aggregated through attention-weighted global pooling, and then input into the classifier to obtain the predicted probabilities, as follows:

[0161]

[0162] In the formula, Indicates the first The attention-weighted global feature vector of each sample, with dimension [ ]. It dynamically aggregates information from all time steps and pays more attention to local events that are important for classification. Indicates the first The attention weights at each time step are scalars with values ​​between 0 and 1, representing the importance of the features at that time step for the final classification. They are calculated as follows: ; Indicates the first Classification feature matrix of each sample In the Feature vectors at each time step; The weight matrix representing the attention mechanism consists of trainable parameters; The bias vector representing the attention mechanism is a trainable parameter; The query vector representing the attention mechanism consists of trainable parameters. express The transpose of .

[0163] The global feature vector is input into two independent classifiers to obtain the dual-task prediction probability distributions, which are expressed as follows:

[0164]

[0165]

[0166] In the formula, Indicates the first The probability distribution of dust source type prediction for each sample is A dimensional vector, whose dimensional vector is the first dimensional vector. The element indicates that the sample belongs to the first element. The probability of a dust-like source; Indicates the number of dust source types; The category index represents the type of dust source. ; Indicates the first The probability distribution for predicting the pollution level of a sample is: A dimensional vector, whose dimensional vector is the first dimensional vector. The element indicates that the sample belongs to the first element. The probability of Level 1 pollution; Indicates the number of pollution level categories; A level index representing the degree of pollution. ; The weight matrix representing the dust source classification head is a trainable parameter; The bias vector representing the dust source classification head is a trainable parameter; The weight matrix representing the pollution level classification head is a trainable parameter; The bias vector representing the pollution level classification head is a trainable parameter; This represents the Softmax activation function, which transforms the input vector into a probability distribution.

[0167] In practical implementation, dust sources include: construction dust, road dust, industrial emissions, and soil erosion. Therefore, [the relevant regulations / measures] are set up... Furthermore, in this embodiment, the pollution level category is based on PM2.5 or PM10 concentration, specifically divided into "light, moderate, and heavy" pollution. Therefore, the following settings are provided: .

[0168] 3) Calculation of multi-objective adaptive loss function

[0169] To effectively train the dual-task model and address the issues of class imbalance and prediction temporal consistency, a multi-objective adaptive loss function is employed. This function includes class-adaptive weighted cross-entropy loss, standard cross-entropy loss, and temporal smoothness consistency loss, which are adaptively balanced using learnable weights. The total loss function is expressed as follows:

[0170]

[0171] In the formula, This represents the total loss function, which is the total loss used for backpropagation to optimize the model; The category-adaptive weighted cross-entropy loss represents the dust source classification and is used to alleviate the problem of imbalance in dust source type data. The standard cross-entropy loss, representing the pollution level classification, is used to supervise the model's learning of pollution level predictions, prompting the model to learn more general and discriminative feature representations. This represents the temporal smoothness consistency loss, encouraging the model to make consistent dust source type predictions for adjacent time steps with similar features; This represents the loss weight parameter used to balance the loss of dust source classification; it is a trainable parameter. This represents the learnable loss weight parameters used to balance the pollution level classification loss; these are trainable parameters. This represents the learnable loss weight parameters used to balance the temporal smoothness loss; these are trainable parameters.

[0172] It should be noted that, The term is used as a regularization term to prevent the weight parameters from becoming infinitely large, thereby encouraging the model to automatically learn a balanced combination of loss weights.

[0173] In the specific implementation, each sub-loss is defined as follows:

[0174] a) The category-adaptive weighted cross-entropy loss for dust source classification is calculated as follows:

[0175]

[0176] In the formula, This represents the total number of samples in the training batch; Indicates the first The true dust source type label for each sample is an integer index; This indicates an indicator function; its value is 1 when the condition within the parentheses is true, and 0 otherwise. Indicates the first The loss weight for dust-like sources is used to give more attention to low-frequency categories, and its calculation method is expressed as follows: ; Indicates the first The frequency of dust-type sources in the entire training dataset, i.e., the number of samples of that class divided by the total number of samples; It is the median frequency of all categories, a scalar constant, obtained from statistics on the training set before training.

[0177] b) The standard cross-entropy loss for pollution level classification is calculated as follows:

[0178]

[0179] In the formula, Indicates the first The true contamination level label for each sample is an integer index; Indicates the first Probability distribution of pollution level prediction for each sample The Each element.

[0180] c) The calculation method for the temporal smoothness consistency loss is expressed as follows:

[0181]

[0182] In the formula, Indicates the first The predicted logistic value matrix of dust source types for all time steps of each sample, with dimensions of [missing information]. ,in ; Indicates the first Each sample at time step The dust source type prediction logistic value vector, i.e., the original output without Softmax activation, has a dimension of The calculation method is expressed as ; Indicates the first Each sample at time step Dust source type prediction logical value vector; Indicates the first Classification feature matrix of each sample In the Feature vectors at each time step; Indicates the first Classification feature matrix of each sample In the Feature vectors at each time step; Represents the L2 norm; It is a temperature hyperparameter, greater than 0, used to control the influence of feature differences on the smoothness constraint strength, with an example value of 1.0.

[0183] It should be noted that the trainable parameters used in calculating the predicted logical value vector of dust source type are... and The trainable parameters used in the dust source classification head prediction are the same set of parameters. Using the same classifier weights and biases, the logical values ​​corresponding to the features at each time step are calculated to construct constraints for the consistency of predictions between adjacent time steps.

[0184] In a specific implementation, the iterative training of the deep learning model and the update of trainable parameters are as follows:

[0185] Using the training set, the trained deep learning model is iteratively trained to update all trainable parameters. In each training iteration, a batch of sample data is randomly selected from the training set, and the forward propagation process is executed sequentially.

[0186] The original time-series data of the samples and the external environmental context information are input into the model. After passing through the multi-scale time-frequency feature extraction and enhancement, multi-source heterogeneous feature fusion and dimensionality reduction encoding, and multi-scale time-series feature extraction and dynamic fusion modules, the dust source type and pollution level prediction probability distribution of the current batch of samples are obtained through the dust monitoring data classification and loss function calculation module. Then, based on the prediction probability and the true label of the samples, the multi-objective adaptive total loss function is calculated.

[0187] Using the backpropagation algorithm, the gradient of the total loss with respect to all trainable parameters in the model is calculated. An adaptive moment estimator optimizer is then used to update all these trainable parameters based on the calculated gradient in order to minimize the total loss function.

[0188] The training process is repeated cyclically. Each training cycle consists of completing all samples in the training set. During training, the model performance is evaluated periodically on an independent validation set, and the changes in loss value and classification accuracy are monitored. When the performance on the validation set no longer improves within several consecutive training cycles, or when the preset maximum number of training cycles is reached, the iterative training process is stopped, and the model parameters obtained at this time are saved as the final trained model.

[0189] In a specific implementation, S5 is as follows:

[0190] Once the model is trained, it can be used to automatically classify newly collected dust monitoring data.

[0191] Specifically, for new monitoring data, the synchronous raw time-series data from PM2.5, PM10, temperature, and humidity sensors are first organized in the same way as S1, with a fixed time window (e.g., 24 hours), and the wind speed, wind direction, time period, and weekday information for the corresponding time period are collected.

[0192] Then, this new sample is input into the pre-trained complete model processing flow: its original time-series data undergoes multi-scale time-frequency feature extraction and enhancement in S2, converting it into an enhanced feature tensor; this tensor is then fused and dimensionality-reduced in the multi-source heterogeneous feature fusion and dimensionality reduction coding module in S3, along with the fine-band energy features and time-series pattern coding features calculated from the new sample data; the generated fused feature matrix and the encoded environmental context information are then sequentially processed through the multi-scale time-series feature extraction and dynamic fusion module and the classification module. Finally, the model outputs the probability distribution of the new sample belonging to various dust sources (such as construction, road dust, etc.) and the probability distribution of its belonging to various pollution levels, taking the category with the highest probability as the final dust source type classification result and pollution level determination result. Based on the classification results, regulatory authorities can achieve rapid identification and source tracing analysis of dust source types in different areas, and conduct graded early warning and targeted control according to pollution levels, thereby improving the accuracy and efficiency of dust pollution control.

[0193] Example 2

[0194] like Figures 2 to 4 As shown, the specific performance of multi-scale time-frequency feature extraction on different types of dust signals is analyzed.

[0195] Figure 2 The heatmap shown illustrates the energy distribution characteristics of four dust sources across 64 wavelet scales. The horizontal axis represents the wavelet scale index, ranging from 1 to 64, corresponding to the analysis range from high to low frequencies. The vertical axis represents the four dust source types, with color intensity indicating normalized energy intensity. It can be observed that construction dust sources exhibit a concentrated energy distribution in the mid-frequency region, consistent with the intermittent operation of construction machinery. Road dust displays a mixed pattern across multiple frequency bands. Industrial emissions show sustained high energy levels even in the low-frequency range, reflecting continuous emissions. Soil erosion exhibits a high energy distribution in the mid-to-high frequency bands.

[0196] like Figure 3 The bar chart shown illustrates the time-frequency Shannon entropy of the four dust source signals. The vertical axis is in bits and is used to quantify the complexity of the time-frequency distribution of the signals. The results show that industrial emissions have the highest Shannon entropy, indicating that their signal pattern is the most complex.

[0197] like Figure 4 The scatter plot shown illustrates the distribution of different dust source samples in the feature space after feature dimensionality reduction. The horizontal and vertical axes represent the two main feature dimensions, and the points of different colors represent different dust source categories. It can be seen that the various types of samples form relatively clear clustering regions, and there is some overlap at the boundaries, reflecting the complexity of the actual dust source signals.

[0198] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A multi-source sensing fusion-based dust monitoring data processing and classification method, characterized in that, Includes the following steps: S1. Deploy multi-source dust monitoring sensor nodes in the target monitoring area. Each node synchronously collects time-series data from four sensor channels: PM2.5 concentration, PM10 concentration, temperature, and humidity, as well as external environmental context information. Using data within a fixed monitoring time window as samples, the collected data is labeled with pollution level and dust source type, and integrated into a dataset. S2. Perform continuous wavelet transform on the original time series data of each sensor channel to convert it to the time-frequency domain, extract the wavelet energy spectrum and Shannon entropy, and concatenate them with the original time series data to obtain the enhanced feature tensor. S3. A two-stage feature fusion and coding strategy is adopted. First, wavelet packet decomposition is performed on the original time series data of the key channels to extract frequency band energy features. Then, the cross-correlation, statistical moments, and ratio features between multiple channels are calculated and spliced ​​into time series pattern coding features. Then, the enhanced feature tensor is subjected to preliminary local transformation to obtain the local time series feature matrix. The three are fused to obtain the fused feature matrix. S3 is as follows: S3.1, The raw time-series data of the two key particulate matter channels, PM2.5 concentration and PM10 concentration, are processed separately. Layered wavelet packet decomposition is used to obtain multi-layer node coefficients of a tree structure. The energy of each node coefficient is calculated, and the coefficient with the highest energy is selected. The energy of each node constitutes a feature vector; Then, the feature vectors of the two key particulate matter channels, PM2.5 concentration and PM10 concentration, are concatenated to obtain the wavelet packet energy feature vector of each sample. S3.2 Calculate the time-series pattern coding features by capturing the dynamic correlation and joint statistical characteristics between different sensing channels. The time-series pattern coding features include the maximum cross-correlation coefficient between channels and their lag time, the statistical moments of each channel and the statistics of the particulate matter concentration ratio sequence. The three types of features are spliced ​​together to form the time-series pattern coding feature vector. S3.3 Input the enhanced feature tensor into a lightweight one-dimensional convolutional layer to perform preliminary local feature extraction and transformation, and obtain the local temporal feature matrix; S3.

4. Fuse the local temporal feature matrix with the wavelet packet energy feature vector and the temporal pattern coding feature vector; First, the wavelet packet energy feature vector and temporal pattern encoding feature vector are projected onto a unified subspace using independent learnable affine transformations. Then, the projected global features are copied in the time dimension. Then, it is concatenated with the local temporal feature matrix in the feature dimension to obtain the fused feature matrix; S4. Construct and train a deep learning model for dust monitoring data classification. The model includes a multi-scale temporal feature extraction and dynamic fusion module, a dust monitoring data classification and loss function calculation module. Input the fused feature matrix into the model for classification and calculate the loss function. Then iteratively train and update the trainable parameters of the deep learning model to obtain the trained model. S5. After preprocessing, the newly collected dust monitoring data is input into the trained model, which outputs the probability distribution of various dust sources and pollution levels. The category with the highest probability is taken as the final result.

2. The method for processing and classifying dust monitoring data based on multi-source sensor fusion according to claim 1, characterized in that, S1 is as follows: The monitoring areas include the areas surrounding construction sites, main urban roads, industrial parks, and open areas; The data collected by the multi-source dust monitoring sensor nodes include PM2.5 concentration, PM10 concentration, temperature, and humidity. The sampling time interval is set, and the data collection covers different meteorological conditions, seasonal periods, and dust source activity periods. Samples are constructed according to a fixed monitoring time window. Simultaneously, the external environmental context information corresponding to each sample is recorded, including the average wind speed, prevailing wind direction, time period, and whether it is a working day during the monitoring period. Environmental experts manually labeled each collected sample by combining video recordings from monitoring points, on-site inspection logs, and multi-source data models. The dust source type labels included construction dust, road dust, industrial emissions, and soil wind erosion, and the pollution level labels included light, moderate, and heavy. All samples and their corresponding four-channel raw time-series data, external environmental context information, dust source type labels, and pollution level labels are integrated to form a dataset.

3. The method for processing and classifying dust monitoring data based on multi-source sensor fusion according to claim 1, characterized in that, S2 is as follows: S2.

1. Perform continuous wavelet transform independently on the time-series data of each sensor channel for each sample, mapping the one-dimensional time-series signal into a two-dimensional time-frequency coefficient matrix. The element in the m-th row and n-th column of the two-dimensional time-frequency coefficient matrix is... ,Right now, Indicates the first The first sample The element in the m-th row and n-th column of the continuous wavelet transform time-frequency coefficient matrix of each sensing channel. The scale index represents the wavelet transform. A time index representing a time series; S2.2 For the time-frequency coefficient matrix of each channel, aggregate the energy at each scale along the time dimension to obtain the total energy intensity of that scale over the entire time range, and compress the dynamic range through logarithmic operation to obtain the wavelet energy spectrum of each sample and each sensing channel at each scale. S2.3 Treat the normalized energy of the time-frequency coefficient matrix as a probability distribution, and measure the time-frequency distribution of the channel signal by calculating its Shannon entropy. High entropy value corresponds to noise interference or complex mixed dust source mode, while low entropy value corresponds to dust source event with single mode. S2.

4. The original time-series data of each sensing channel, the extracted multi-scale wavelet energy spectrum features, and the Shannon entropy features are spliced ​​together in the feature dimension to form an enhanced feature tensor that simultaneously contains time-domain, frequency-domain energy, and complexity information.

4. The dust monitoring data processing and classification method based on multi-source sensor fusion according to claim 1, characterized in that, time series... The specific process for calculating the pattern encoding feature vector is as follows: (1) Calculate the cross-correlation sequence between each pair of sensing channels, take the maximum cross-correlation number and its corresponding leading or lagging relationship, and obtain the cross-correlation characteristics; (2) Calculate the four statistical moments of the time series data for each channel, namely mean, variance, skewness and kurtosis, which constitute the statistical moment features and are used to describe the central tendency, dispersion, distribution asymmetry and peak degree of the data in each channel. (3) Calculate the point-by-point ratio sequence of PM2.5 concentration to PM10 concentration, and then calculate the mean and variance of the ratio sequence to obtain the 2-dimensional ratio characteristics; (4) The cross-correlation feature, statistical moment feature and ratio feature are concatenated to obtain the time-series pattern coding feature vector, which is used to comprehensively characterize the dynamic relationship and statistical characteristics between multiple channels.

5. The method for processing and classifying dust monitoring data based on multi-source sensor fusion according to claim 1, characterized in that, S4 is as follows: A deep learning model for dust monitoring data classification was constructed and trained. The fused feature matrix was input into the model, which first passed through a multi-scale temporal feature extraction and dynamic fusion module. This module first used parallel causal convolutional layers with different dilation rates to extract features from micro to macro scales, obtaining a mixed feature matrix. Then, a multi-scale attention mechanism was used to dynamically evaluate and fuse features from different scales, obtaining attention-adjusted scale-related features. Finally, a weighted feature concatenation was used to output the multi-scale fused feature matrix. The dust monitoring data classification and loss function calculation module adopted a classification module that integrates environmental context, performs dual-task prediction, and uses an adaptive multi-objective loss function. This module encodes external meteorological and temporal information, predicts dust source type and pollution level in parallel, and optimizes multi-objective loss. Then, the model was optimized, trained, and its parameters were updated.

6. The dust monitoring data processing and classification method based on multi-source sensor fusion according to claim 1, characterized in that, The multi-scale temporal feature extraction and dynamic fusion module is as follows: (1) Input the fusion feature matrix into multiple one-dimensional causal convolutional layers with different dilation rates. Each convolutional layer works independently to capture local patterns within the receptive field at a specific time. Then, the outputs of all convolutional layers are concatenated in the feature dimension to form a hybrid feature containing multi-scale information. (2) The attention mechanism is used to dynamically fuse features at different scales. The attention mechanism first calculates the global context coefficient for each time step, and then calculates the importance weight of each scale feature at that time step based on the global context coefficient and the local features at each scale. First, global average pooling is applied to the hybrid feature matrix along the feature dimension to obtain the global context coefficients at each time step. Then, the hybrid feature matrix is ​​divided into multiple groups according to the inflation rate of the source. For each time step and each inflation rate, a learnable attention network is used to calculate its attention weights. (3) Use attention weights to weight the original multi-scale features step by step and scale by scale to obtain the scale-related features after attention adjustment; (4) All the weighted features corresponding to the expansion rate are concatenated again in the feature dimension to obtain a weighted feature matrix. Then, nonlinear transformation and feature selection are performed through the gated activation unit to obtain the multi-scale fusion feature matrix output by multi-scale temporal feature extraction and dynamic fusion.

7. The method for processing and classifying dust monitoring data based on multi-source sensor fusion according to claim 1, characterized in that, The dust monitoring data classification and loss function calculation module is as follows: (1) A lightweight external feature encoder is used to encode wind speed, wind direction, time period and weekday information into feature vectors, which are then concatenated with the multi-scale fusion feature matrix in the feature dimension to form a classification feature matrix; (2) A dual-task classification head is used for classification prediction. The main task predicts the dust source type and the auxiliary task predicts the pollution level. The two tasks share the main feature extraction backbone but have independent classifiers. More general and discriminative features are captured through joint learning. First, key time step information is aggregated through attention-weighted global pooling to obtain attention-weighted global feature vectors. Then, the global feature vectors are input into two independent classifiers to obtain dual-task prediction probability distributions, including dust source type prediction probability distribution and pollution level prediction probability distribution. (3) A multi-objective adaptive loss function is adopted, including category adaptive weighted cross-entropy loss, standard cross-entropy loss and temporal smoothness consistency loss.

8. The method for processing and classifying dust monitoring data based on multi-source sensor fusion according to claim 1, characterized in that, The training process of the model is as follows: The deep learning model is iteratively trained and all trainable parameters are updated using data from the dataset. In each training iteration, a batch of sample data is randomly selected from the dataset and forward propagation is performed sequentially. After preprocessing in steps S2 and S3, the data is input into the model to obtain the predicted probability distribution of dust source type and pollution level of the current batch of samples. Then, the total loss function is calculated. Then, backpropagation is performed to calculate the gradient of the total loss with respect to all trainable parameters in the model. An adaptive moment estimator optimizer is used to update all these trainable parameters based on the calculated gradients, minimizing the total loss function. Set an iteration stopping condition, and when the condition is met, end the training to obtain the trained model.

9. The method for processing and classifying dust monitoring data based on multi-source sensor fusion according to claim 1, characterized in that, S5 is detailed below: For new monitoring data, after preprocessing, it is input into the trained model, which outputs the probability distribution of various dust sources and the probability distribution of each pollution level. The category with the highest probability is taken as the final dust source type classification result and pollution level determination result. Based on the dust source type classification result, the dust source types in different regions are quickly identified and traced for analysis, and graded early warning and targeted control are carried out according to the pollution level.

Images