Air quality monitoring system and method based on deep learning

Abstract

The invention discloses an air quality monitoring system and method based on deep learning, and relates to the technical field of air quality monitoring. According to the air quality monitoring system and method based on deep learning, a hardware-level encryption identifier and a federated learning architecture are adopted to ensure that sovereignty ownership is clear during data cross-border circulation, and ownership dispute caused by data format difference in a traditional scheme is avoided; the dual-channel analysis module fuses GAN spectral decomposition and morphological waveform detection, realizes precise separation of industrial / agricultural pollution sources in a single PM2.5 data stream, and reduces the misjudgment rate; the dynamic weight distributor is combined with terrain correction parameters, so that the monitoring network automatically optimizes space coverage in a complicated border landform, and the equipment blind area is reduced compared with a fixed station layout; the differential privacy and gradient desensitization technology constructs multiple protections, and the data leakage risk is effectively controlled on the premise of keeping the model precision.

Description

Technical Field

[0001] The present invention relates to the technical field of air quality monitoring, and particularly to an air quality monitoring system and method based on deep learning. Background Art

[0002] As the core carrier of cross-border industrial chain collaboration, a cross-border industrial park gathers diverse pollution sources such as coal-fired power plants, chemical plants, and cross-border logistics. Affected by terrain barriers, administrative jurisdiction division, monitoring standard differences, etc., in such areas;

[0003] In traditional detection schemes, most fixed stations sample once per hour, making it difficult to capture pollution pulse fluctuations caused by truck customs clearance, factory shift changes, etc. When industrial coal burning and agricultural burning occur simultaneously, chemical composition analysis methods may produce misjudgments due to cross-interference between sulfides and potassium ions. In addition, cross-border data sharing requires dedicated line transmission, and border areas often face network outages or protocol barriers, resulting in deviation in the traceability conclusions of the same pollution event by the monitoring systems of both parties due to inconsistent data formats. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] The present invention provides an air quality monitoring system and method based on deep learning to solve the problems that existing solutions rely on the deployment of high-density monitoring terminals, there are still data collection blind spots in infrastructure-weak areas such as border uninhabited areas, and complex algorithms have high computing power requirements for edge computing devices.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides an air quality monitoring system based on deep learning, which includes,

[0008] A pollution source separation module, including a fundamental frequency tracking algorithm processor and a pulse peak detection circuit, for receiving PM2.5 time-series data streams uploaded by monitoring terminals in at least two sovereign jurisdictions;

[0009] A dual-channel time-series analysis module, connected to the pollution source separation module, including an industrial feature extraction channel and an agricultural feature extraction channel. The industrial feature extraction channel uses a fundamental frequency tracking algorithm to identify 24-hour periodic fluctuation patterns, and the agricultural feature extraction channel is configured to detect pulsed peaks in the morning and evening periods;

[0010] A sovereign identifier output unit, which generates a pollution contribution report with sovereign jurisdiction codes according to the output results of the dual-channel time-series analysis module.

[0011] As a preferred embodiment of the air quality monitoring system based on deep learning according to the present invention, wherein: the industrial feature extraction channel is built-in with a spectrum decomposition generative adversarial network (GAN), and its generator decomposes the input PM2.5 data into industrial fundamental frequency components and noise components; the generator learns the time-frequency distribution characteristics of industrial pollution fundamental frequency components through adversarial training;

[0012] The agricultural feature extraction channel includes a morphological waveform detector configured to identify steep rising edge signals with a duration less than 2 hours.

[0013] As a preferred embodiment of the air quality monitoring system based on deep learning according to the present invention, wherein: in the morphological waveform detector of the agricultural feature extraction channel, the candidate signal determination method is:

[0014] P(t) = 1{s'(t)>a, d(t)<b, I m (t)>c},

[0015] wherein, P(t) represents the candidate signal detection result at time t, s'(t) represents the instantaneous slope at time t, a represents the slope threshold for determining a steep rising edge, d(t) represents the duration of the candidate pulse signal at time t, b represents the maximum allowable duration of the pulse signal in hours, I m (t) represents the morphological similarity index between the candidate signal and the ideal pulse waveform at time t, and c represents the determination threshold of the morphological similarity index;

[0016] Perform a similarity rating between the candidate signal and the predefined ideal pulse morphology, and the formula is:

[0017]

[0018] wherein, I m (t) represents the morphological similarity index of the candidate signal at time t, N represents the total number of sampling points, s(t i ) represents the measured signal value at the i-th sampling time, represents the mean value of the measured signals within the candidate signal window, s ideal (t i ) represents the ideal pulse waveform value at the i-th sampling time, represents the mean value of the ideal pulse waveform, and t i represents the i-th sampling time.

[0019] As a preferred embodiment of the air quality monitoring system based on deep learning according to the present invention, it further includes:

[0020] A cross-border data cleaning interface configured to convert data uploaded by monitoring terminals in different sovereign jurisdictions into a standardized tensor with a unified timestamp, which retains the administrative division code of the original data collection location;

[0021] The standardized tensor refers to a three-dimensional data matrix that has been subjected to minimum-maximum normalization and dimension alignment;

[0022] A dynamic weight allocator automatically adjusts the weight coefficients of monitoring terminals in pollution contribution calculations based on the geographical density of monitoring terminals in each sovereign jurisdiction.

[0023] As a preferred solution of the deep learning-based air quality monitoring system described in the present invention, in the dynamic weight allocator, the average spacing is used to determine the geographical density of monitoring terminals in each sovereign jurisdiction, and the weight coefficient calculation formula is:

[0024]

[0025] Among them, w represents the weight coefficient of the monitoring terminals in each sovereign jurisdiction, α represents the proportional scaling constant in the inverse proportional relationship, D represents the average spacing between monitoring terminals in the jurisdiction, β represents the terrain undulation correction parameter, and δ represents a dimensionless indicator reflecting the degree of terrain undulation. The larger the value, the more drastic the terrain change.

[0026] In a second aspect, the present invention provides an air quality monitoring method based on deep learning, comprising:

[0027] Step S1: Continuously collect PM2.5 concentration time series data through monitoring terminals deployed in each sovereign jurisdiction. The sovereign encryption chip built into the terminal adds an irreversible jurisdiction identifier to the original data;

[0028] The jurisdiction identification generated by the sovereign encryption chip includes dual verification information of the geographic location hash value and the time zone offset;

[0029] Step S2: Using a transfer learning framework to train a regional pollution feature extraction model, a teacher model is used to learn the chemical characteristics of pollution sources in multiple countries based on mass spectrometer data, and a student model is used to perform knowledge distillation using PM2.5 concentration time series data.

[0030] During the knowledge distillation process, the output layer of the teacher model is replaced with a probability distribution vector containing two types of pollution sources: industrial emissions and agricultural burning;

[0031] In step S3, a federated learning strategy is used to update the feature extraction model. The monitoring terminals in each sovereign jurisdiction only upload the model gradient difference, and the gradient data is desensitized using a differential privacy algorithm.

[0032] As a preferred solution of the air quality monitoring method based on deep learning according to the present invention, wherein: in step S1, the PM2.5 raw data X collected by each monitoring terminal is processed by an embedded sovereign encryption chip, and an irreversible jurisdiction identifier is added. During the process, a national cryptographic algorithm is used to generate a hardware fingerprint, and data binding is achieved by combining the digital certificate and GPS coordinates of the device. The steps include:

[0033] Connect the unique hardware fingerprint, digital certificate and GPS coordinates of the device, and use the SM3 hash function in the national cryptographic algorithm to generate an irreversible identifier. The process is expressed as:

[0034] L = SM3(p1||p2||p3),

[0035] wherein, p1 represents the device hardware fingerprint, which is the digest generated by the unique physical identifier collected by the embedded chip, p2 represents the device digital certificate, which is the authentication data generated at the time of device factory and bound to the device information, p3 represents the device GPS coordinate data, which is obtained in real time by the built-in positioning module of the terminal. SM3 is the hash function for implementing message digest generation in the national cryptographic algorithm, ‖ represents the data connection operation, and L represents the generated irreversible jurisdiction identifier for binding the raw data;

[0036] Combine the irreversible identifier L with the raw data X to form the data X' after identifier processing:

[0037] X' = {X, L};

[0038] Verify the device location using the real-time GPS data p'3. The verification formula is:

[0039] V = 1{|p3 - p′3| < ∈},

[0040] wherein, p'3 represents the detected real-time GPS coordinates, ∈ is the deviation threshold allowed for GPS data, and V is the verification result. If the condition is met, output 1 to indicate that the verification is passed.

[0041] As a preferred solution of the air quality monitoring method based on deep learning according to the present invention, wherein: in step S3, the steps of updating the feature extraction model using the federated learning strategy include:

[0042] In a single jurisdiction, let the model gradient difference uploaded by the i-th terminal be Δθ i , then the gradient aggregation within the jurisdiction adopts average calculation:

[0043] wherein, Δθ i represents the gradient difference of the i-th terminal, N r represents the number of terminals participating in the federated learning within the jurisdiction r, and ΔΘ r represents the aggregated gradient difference within the jurisdiction r;

[0044] Cross-jurisdictional global gradient update is performed through a hierarchical topology structure, and the update formula is:

[0045] where R represents the total number of sovereign jurisdictions, and w r represents the weight coefficient corresponding to jurisdiction r, reflecting the data and equipment characteristics within the jurisdiction, represents the aggregated gradient of the jurisdiction after differential privacy protection;

[0046] In the differential privacy mechanism, each jurisdiction adaptively adjusts the noise according to the data volume. Let the amplitude of the Gaussian noise added by jurisdiction r be σ r = γ / N r , where γ represents the noise scaling constant, and σ r represents the standard deviation of the noise of jurisdiction r, which is inversely proportional to the data volume N r ;

[0047] The gradient after differential privacy processing is:

[0048]

[0049] where represents Gaussian distribution noise with a mean of 0 and a variance of ;

[0050] As a preferred solution of the air quality monitoring method based on deep learning described in the present invention, it further includes:

[0051] Construct a sovereign credibility evaluation module, whose input is the historical data consistency index and device calibration record of each jurisdiction's monitoring terminal, and the output is a dynamic credibility weight; the dynamic credibility weight is used to reduce the pollution attribution deviation caused by abnormal device calibration;

[0052] When detecting conflicts in the monitoring data of adjacent jurisdictions, use the data of the jurisdiction with a higher credibility weight as the pollution tracing benchmark;

[0053] A sovereign policy loading unit, configured to dynamically enable or disable specific data processing processes according to the legal requirements of the device deployment location; when the instruction is executed, a sovereign compliance log containing multiple language versions is automatically generated, and the data operation records of different jurisdictions in the log are stored in independent encrypted partitions.

[0054] As a preferred solution of the air quality monitoring method based on deep learning described in the present invention, wherein: in the sovereign policy loading unit, the log partition encryption adopts an independent encryption protocol authenticated by each sovereign jurisdiction; the data operation record contains a reference index of territorial legal provisions.

[0055] The beneficial effects of the present invention are as follows: The present invention uses hardware-level encryption identification and a federated learning architecture to ensure clear sovereignty attribution when data is circulated across borders, avoiding ownership disputes caused by differences in data formats in traditional solutions; the dual-channel analysis module combines GAN spectrum decomposition and morphological waveform detection to accurately separate industrial / agricultural pollution sources in a single PM2.5 data stream, reducing the misjudgment rate; the dynamic weight allocator combines terrain correction parameters to automatically optimize the spatial coverage of the monitoring network in complex border terrains, reducing the blind area of equipment compared to fixed-site layouts; differential privacy and gradient desensitization technologies build multiple protections to effectively control the risk of data leakage while maintaining the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 It is a schematic framework diagram of an air quality monitoring system based on deep learning in Embodiment 1.

[0058] Figure 2 It is a schematic flowchart of an air quality monitoring method based on deep learning in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the drawings of the specification.

[0060] Many specific details are set forth in the following description in order to fully understand the present invention, but the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0061] Secondly, the so-called "one embodiment" or "embodiment" refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that mutually excludes other embodiments.

[0062] Embodiment 1, referring to Figure 1 and Figure 2 This embodiment provides an air quality monitoring system based on deep learning, including:

[0063] The pollution source separation module, which includes a fundamental frequency tracking algorithm processor and a pulse peak detection circuit, is used to receive the PM2.5 time series data stream uploaded by the monitoring terminals in at least two sovereign jurisdictions.

[0064] The dual-channel time series analysis module is connected to the pollution source separation module and includes an industrial feature extraction channel and an agricultural feature extraction channel. Among them, the industrial feature extraction channel uses the fundamental frequency tracking algorithm to identify the 24-hour periodic fluctuation pattern, and the agricultural feature extraction channel is configured to detect the pulsed peaks in the morning and evening periods.

[0065] The industrial feature extraction channel is built-in with a spectral decomposition generative adversarial network (GAN). Its generator decomposes the input PM2.5 data into industrial fundamental frequency components and noise components; the generator learns the time-frequency distribution characteristics of the industrial pollution fundamental frequency components through adversarial training.

[0066] The agricultural feature extraction channel includes a morphological waveform detector, which is configured to identify the steep rising edge signals with a duration less than 2 hours.

[0067] In the morphological waveform detector of the agricultural feature extraction channel, the candidate signal determination method is:

[0068] P(t) = 1{s'(t)>a, d(t)<b, I m (t)>c},

[0069] where P(t) represents the detection result of the candidate signal at time t, s'(t) represents the instantaneous slope at time t, a represents the slope threshold for determining the steep rising edge, d(t) represents the duration of the candidate pulse signal at time t, b represents the maximum allowable duration of the pulse signal in hours, I m (t) represents the morphological similarity index between the candidate signal and the ideal pulse waveform at time t, and c represents the determination threshold of the morphological similarity index.

[0070] Perform the similarity rating between the candidate signal and the predefined ideal pulse morphology. The formula is:

[0071]

[0072] where I m (t) represents the morphological similarity index of the candidate signal at time t, N represents the total number of sampling points, s(t i ) represents the measured signal value at the i-th sampling time, represents the mean value of the measured signals within the candidate signal window, s ideal (t i ) represents the ideal pulse waveform value at the i-th sampling time, represents the mean value of the ideal pulse waveform, t i represents the i-th sampling time;

[0073] The sovereignty identification output unit generates a pollution contribution report with sovereignty jurisdiction coding based on the output results of the dual-channel timing analysis module;

[0074] The air quality monitoring system also includes:

[0075] A cross-border data cleaning interface is configured to convert data uploaded by monitoring terminals in different sovereign jurisdictions into a standardized tensor with a unified timestamp. The standardized tensor retains the administrative division code of the original data collection location;

[0076] A normalized tensor is a three-dimensional data matrix that has been min-max normalized and dimensionally aligned.

[0077] A dynamic weight allocator that automatically adjusts the weight coefficients of monitoring terminals in pollution contribution calculations based on the geographical density of monitoring terminals in each sovereign jurisdiction;

[0078] In the dynamic weight allocator, the average spacing is used to determine the geographical density of monitoring terminals in each sovereign jurisdiction. The weight coefficient calculation formula is:

[0079]

[0080] Where w represents the weight coefficient of the monitoring terminals in each sovereign jurisdiction, α represents the scaling constant in the inverse proportional relationship, D represents the average spacing between monitoring terminals in the jurisdiction, β represents the terrain undulation correction parameter, and δ represents a dimensionless index reflecting the degree of terrain undulation. A larger value indicates a more drastic terrain change.

[0081] Specifically, the instantaneous slope, signal duration and morphological similarity index are used to identify the real pulse signal and equipment fault noise, and the signal is distinguished through threshold control; the statistical correlation calculation between the sampling data and the ideal template is used to improve the accuracy of signal matching; and based on the monitoring terminal spacing and the dimensionless index of terrain undulation, an inverse proportional relationship between the weight and the spatial density is established, supplemented by a correction coefficient, so that the weight distribution under complex terrain conditions is more reasonable.

[0082] This embodiment also provides a monitoring method for the above-mentioned deep learning-based air quality monitoring system, including:

[0083] Step S1: PM2.5 concentration time series data is continuously collected through monitoring terminals deployed in each sovereign jurisdiction. The sovereign encryption chip built into the terminal adds an irreversible jurisdiction identifier to the original data.

[0084] The jurisdiction identification generated by the sovereign encryption chip contains dual verification information of the geographic location hash value and time zone offset;

[0085] In step S1, the original PM2.5 data X collected by each monitoring terminal is processed by the built-in sovereign encryption chip, and an irreversible jurisdiction identifier is added. During the process, a national cryptographic algorithm is used to generate a hardware fingerprint, and data binding is achieved by combining the device's digital certificate and GPS coordinates. The steps include:

[0086] Connect the unique hardware fingerprint, digital certificate, and GPS coordinates of the device, and use the SM3 hash function in the national cryptographic algorithm to generate an irreversible identifier. The process is expressed as:

[0087] L = SM3(p1||p2||p3),

[0088] where p1 represents the device hardware fingerprint, which is the digest generated from the unique physical identifier collected by the built-in chip; p2 represents the device digital certificate, which is the authentication data generated at the time of device factory and bound to the device information; p3 represents the device GPS coordinate data, which is obtained in real time by the built-in positioning module of the terminal. SM3 is the hash function in the national cryptographic algorithm to implement message digest generation, ‖ represents the data connection operation, and L represents the generated irreversible jurisdiction identifier for binding the original data;

[0089] Combine the irreversible identifier L with the original data X to form the data X' after identifier processing:

[0090] X' = {X, L};

[0091] Verify the device location using the real-time GPS data p'3. The verification formula is:

[0092] V = 1{|p3 - p′3| < ∈},

[0093] where p'3 represents the detected real-time GPS coordinates, ∈ is the deviation threshold allowed for GPS data, and V is the verification result. If the condition is met, output 1 to indicate that the verification is passed;

[0094] Specifically, a national cryptographic algorithm is used to generate a hardware fingerprint and a device digital certificate, and combined with the GPS coordinates to form an irreversible jurisdiction identifier. The technical path determines the device source in the data collection link. The irreversible hash function ensures that even if the identifier is made public, the original information cannot be restored externally, protecting data privacy. The real-time GPS coordinate verification provides a location accuracy check for device data, preventing forgery or false reporting;

[0095] In step S2, use the transfer learning framework to train the regional pollution feature extraction model. The teacher model learns the chemical characteristics of pollution sources in multiple countries based on mass spectrometer data, and the student model uses the PM2.5 concentration time series data for knowledge distillation;

[0096] During the knowledge distillation process, the output layer of the teacher model is replaced with a probability distribution vector containing two types of pollution sources: industrial emissions and agricultural combustion;

[0097] Step S3: Update the feature extraction model using the federated learning strategy. The monitoring terminals in each sovereign jurisdiction only upload the model gradient differences, and desensitize the gradient data through the differential privacy algorithm;

[0098] In step S3, the steps of updating the feature extraction model using the federated learning strategy include:

[0099] In a single jurisdiction, let the model gradient difference uploaded by the i-th terminal be Δθ i , then the gradient aggregation within the jurisdiction uses average calculation:

[0100] where, Δθ i represents the gradient difference of the i-th terminal, N r represents the number of terminals participating in federated learning within jurisdiction r, and ΔΘ r represents the aggregated gradient difference within jurisdiction r;

[0101] Perform global gradient update across jurisdictions through a hierarchical topology structure. The update formula is:

[0102] where, R represents the total number of sovereign jurisdictions, w r represents the weight coefficient corresponding to jurisdiction r, reflecting the data and device characteristics within the jurisdiction, represents the aggregated gradient of the jurisdiction after differential privacy protection;

[0103] In the differential privacy mechanism, each jurisdiction adaptively adjusts the noise according to the data volume. Let the amplitude of the Gaussian noise added by jurisdiction r be σ r =γ / N r , where, γ represents the noise scaling constant, and σ r represents the standard deviation of the noise in jurisdiction r, which is inversely proportional to the data volume N r ;

[0104] The gradient after differential privacy processing is:

[0105]

[0106] where, represents Gaussian distributed noise with a mean of 0 and a variance of ,

[0107] Specifically, each jurisdiction reduces the direct exposure risk of individual terminal data by averaging and aggregating local gradient differences. The hierarchical topology structure can utilize jurisdiction-level aggregation to achieve efficient model updates while fully dispersing cross-jurisdictional data. The differential privacy mechanism dynamically adjusts the noise addition amplitude based on the data volume within each jurisdiction. The noise is lower when the data volume is large and higher when the data volume is small, thus balancing the relationship between privacy protection and model accuracy.

[0108] The monitoring method also includes building a sovereign credibility assessment module, whose input is the historical data consistency index and equipment calibration records of monitoring terminals in each jurisdiction, and the output is a dynamic credibility weight. The dynamic credibility weight is used to reduce the pollution attribution bias caused by equipment calibration anomalies.

[0109] When conflicting monitoring data from adjacent jurisdictions is detected, the data from the jurisdiction with a higher credibility weight is used as the benchmark for pollution source tracing;

[0110] A sovereignty policy loading unit is configured to dynamically enable or disable specific data processing processes based on the legal requirements of the device's deployment location. A sovereignty compliance log containing multiple language versions is automatically generated when instructions are executed, and data operation records for different jurisdictions in the log are stored in independent encrypted partitions.

[0111] In the sovereign policy loading unit, log partition encryption uses an independent encryption protocol certified by each sovereign jurisdiction; data operation records contain a reference index to local legal terms.

[0112] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An air quality monitoring system based on deep learning, characterized in that: include, A pollution source separation module, including a fundamental frequency tracking algorithm processor and a pulse peak detection circuit, is used to receive PM2.5 time series data streams uploaded by monitoring terminals in at least two sovereign jurisdictions; A dual-channel time series analysis module, connected to the pollution source separation module, includes an industrial feature extraction channel and an agricultural feature extraction channel. The industrial feature extraction channel uses a fundamental frequency tracking algorithm to identify 24-hour periodic fluctuation patterns, and the agricultural feature extraction channel is configured to detect pulse peaks in the morning and evening hours. The sovereignty identification output unit generates a pollution contribution report with a sovereignty jurisdiction code based on the output results of the dual-channel timing analysis module.

2. The air quality monitoring system based on deep learning according to claim 1, characterized in that: The industrial feature extraction channel has a built-in spectrum decomposition generative adversarial network (GAN), whose generator decomposes the input PM2.5 data into industrial fundamental frequency components and noise components. The generator learns the time-frequency distribution characteristics of the industrial pollution fundamental frequency components through adversarial training. The agricultural feature extraction channel includes a morphological waveform detector configured to identify steep rising edge signals with a duration of less than 2 hours.

3. The air quality monitoring system based on deep learning according to claim 2, characterized in that: In the morphological waveform detector of the agricultural feature extraction channel, the candidate signal determination method is: P(t) = 1{s'(t) > a, d(t) < b, I m (t) > c}, Among them, P(t) represents the candidate signal detection result at time t, s'(t) represents the instantaneous slope at time t, a represents the slope threshold for determining a steep rising edge, d(t) represents the duration of the candidate pulse signal at time t, b represents the maximum allowable duration of the pulse signal in hours, and I m (t) represents the morphological similarity index between the candidate signal and the ideal pulse waveform at time t, and c represents the determination threshold of the morphological similarity index; The similarity rating between the candidate signal and the predefined ideal pulse shape is performed using the formula: Among them, I m (t) represents the morphological similarity index of the candidate signal at time t, N represents the total number of sampling points, s(t i ) represents the measured signal value at the i-th sampling time, represents the mean value of the measured signal within the candidate signal window, s ideal (t i ) represents the ideal pulse waveform value at the i-th sampling time, represents the mean value of the ideal pulse waveform, t i represents the i-th sampling time.

4. The air quality monitoring system based on deep learning according to claim 1, characterized in that Also includes: A cross-border data cleaning interface configured to convert data uploaded by monitoring terminals in different sovereign jurisdictions into a standardized tensor with a unified timestamp, wherein the standardized tensor retains the administrative division code of the original data collection location; The standardized tensor refers to a three-dimensional data matrix that has been subjected to minimum-maximum normalization and dimension alignment; A dynamic weight allocator automatically adjusts the weight coefficients of monitoring terminals in pollution contribution calculations based on the geographical density of monitoring terminals in each sovereign jurisdiction.

5. The air quality monitoring system based on deep learning according to claim 4, characterized in that: In the dynamic weight allocator, the average spacing is used to determine the geographical density of monitoring terminals in each sovereign jurisdiction, and the weight coefficient calculation formula is: Among them, w represents the weight coefficient of the monitoring terminals in each sovereign jurisdiction, α represents the proportional scaling constant in the inverse proportional relationship, D represents the average spacing between monitoring terminals in the jurisdiction, β represents the terrain undulation correction parameter, and δ represents a dimensionless indicator reflecting the degree of terrain undulation. The larger the value, the more drastic the terrain change.

6. A method for monitoring air quality based on deep learning, based on the air quality monitoring system based on deep learning according to any one of claims 1 to 5, characterized in that, include: Step S1: Continuously collect PM2.5 concentration time series data through monitoring terminals deployed in each sovereign jurisdiction. The sovereign encryption chip built into the terminal adds an irreversible jurisdiction identifier to the original data; The jurisdiction identification generated by the sovereign encryption chip includes dual verification information of the geographic location hash value and the time zone offset; Step S2: Using a transfer learning framework to train a regional pollution feature extraction model, a teacher model is used to learn the chemical characteristics of pollution sources in multiple countries based on mass spectrometer data, and a student model is used to perform knowledge distillation using PM2.5 concentration time series data. During the knowledge distillation process, the output layer of the teacher model is replaced with a probability distribution vector containing two types of pollution sources: industrial emissions and agricultural burning; In step S3, a federated learning strategy is used to update the feature extraction model. The monitoring terminals in each sovereign jurisdiction only upload the model gradient difference, and the gradient data is desensitized using a differential privacy algorithm.

7. The air quality monitoring method based on deep learning according to claim 6, characterized in that: In step S1, the original PM2.5 data X collected by each monitoring terminal is processed by an embedded sovereign encryption chip, and an irreversible jurisdiction identifier is added. During the process, a national cryptography algorithm is used to generate a hardware fingerprint, and data binding is achieved by combining the device's digital certificate and GPS coordinates. The steps include: Connect the unique hardware fingerprint, digital certificate, and GPS coordinates of the device, and use the SM3 hash function in the national cryptography algorithm to generate an irreversible identifier. The process is expressed as: L = SM3(p1||p2||p3), where p1 represents the device hardware fingerprint, which is the digest generated from the unique physical identifier collected by the embedded chip; p2 represents the device digital certificate, which is the authentication data generated at the time of device factory and bound to the device information; p3 represents the device GPS coordinate data, which is obtained in real time by the terminal's built-in positioning module. SM3 is the hash function in the national cryptography algorithm for generating message digests, ‖ represents the data connection operation, and L represents the generated irreversible jurisdiction identifier for binding the original data; Combine the irreversible identifier L with the original data X to form the data X' after identifier processing: X' = {X, L}; Use the real-time GPS data p'3 to verify the device location. The verification formula is: V = 1{|p3 - p ′3 |<∈}, where p'3 represents the detected real-time GPS coordinates, ∈ is the deviation threshold allowed for GPS data, and V is the verification result. If the condition is met, output 1 to indicate that the verification is passed.

8. The air quality monitoring method based on deep learning according to claim 6, characterized in that: In step S3, the steps for updating the feature extraction model using the federated learning strategy include: In a single jurisdiction, let the model gradient difference uploaded by the $i$-th terminal be $\Delta\theta$ i , then the gradient aggregation within the jurisdiction adopts average calculation: Among them, Δθ i represents the gradient difference of the i-th terminal, and N r represents the number of terminals participating in federated learning within the jurisdiction r, and ΔΘ r represents the aggregated gradient difference of the jurisdiction r; Perform global gradient update across jurisdictions through a hierarchical topology structure. The update formula is: where R represents the total number of sovereign jurisdictions, and w r represents the weight coefficient corresponding to jurisdiction r, reflecting the data and device characteristics within the jurisdiction, represents the aggregated gradient of the jurisdiction after differential privacy protection; In the differential privacy mechanism, each jurisdiction adaptively adjusts the noise according to the data volume. Let the amplitude of the Gaussian noise added by jurisdiction r be σ r = γ / N r , where γ represents the noise scaling constant, and σ r represents the standard deviation of the noise in jurisdiction r, which is inversely proportional to the data volume N r ; The gradient after differential privacy processing is: Among them, represents Gaussian distributed noise with a mean of 0 and a variance of .

9. The air quality monitoring method based on deep learning according to claim 6, characterized in that, It also includes: Construct a sovereign credibility assessment module, whose input is the historical data consistency index and device calibration record of each jurisdiction's monitoring terminal, and the output is the dynamic credibility weight; The dynamic credibility weight is used to reduce the pollution attribution bias caused by abnormal device calibration; When detecting conflicts in monitoring data between adjacent jurisdictions, use the data of the jurisdiction with a higher credibility weight as the pollution source tracing benchmark; A sovereign policy loading unit, configured to dynamically enable or disable specific data processing processes according to the legal requirements of the device deployment location; automatically generate a sovereign compliance log containing multiple language versions when the instruction is executed, and the data operation records in different jurisdictions in the log are stored in independent encrypted partitions.

10. The air quality monitoring method based on deep learning according to claim 9, characterized in that: In the sovereign policy loading unit, the log partition encryption uses an independent encryption protocol authenticated by each sovereign jurisdiction; the data operation records contain the reference index of the territorial legal provisions.

Images