Method for biological early warning detection of airborne particulate matter

Abstract

This invention discloses a method for early warning and detection of biological particulate matter in air, belonging to the field of environmental monitoring technology. It uses a high-speed, lensless diffraction imaging system to capture particulate matter diffraction ring images in real time, extracts gray-level co-occurrence matrix texture features and wavelet multi-scale features, fuses them to generate a real-time fusion vector, and constructs an index calculation model to calculate a biological index sequence. It identifies biological attributes and uses multi-class SVM for classification, statistically represents concentration by the number of events, and combines Poisson mutation detection to analyze the trend of change. When the concentration exceeds the standard or the trend is abnormal, it triggers a graded early warning. Through concentration adjustment feedback and multi-source data verification, it identifies the type of misjudgment and dynamically optimizes the equipment, model, and threshold. This method achieves high-precision identification, stable monitoring, and intelligent early warning of biological particulate matter in the air, improving detection sensitivity and robustness, and reducing the misjudgment rate.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to a method for early warning and detection of biological effects of airborne particulate matter. Background Technology

[0002] Particulate matter in the air is a significant component of environmental pollutants. This particulate matter includes not only inorganic substances but may also carry various microorganisms, such as bacteria, viruses, and fungal spores, forming airborne biological particles. These biological particles can lead to respiratory infections, allergic reactions, and other health problems.

[0003] A patent application with publication number CN119845904A discloses a method and system for locating and detecting viral infection based on diffraction optics. This method utilizes diffraction imaging principles to acquire diffraction fingerprints corresponding to all cells within an imaging area. The texture features of each diffraction fingerprint in the imaging area are extracted using the gray-level co-occurrence matrix method. The extracted texture features are compared with pre-stored standard texture features in a model relating cell morphological changes to diffraction fingerprints, thereby obtaining the "cell state" of each cell in the imaging area and locating and identifying the viral infection status of each cell. Applying this efficient optical diffraction spectroscopy system to viral infection detection satisfies the need for continuous, high-throughput assessment of cell status after viral infection.

[0004] Although existing virus infection detection technologies based on diffraction optics can achieve high-throughput identification at the cell level, they lack physical criteria for directly extracting biological structural features from diffraction rings. They cannot effectively determine the biological properties and classification of unknown particulate matter, nor can they sense the changing trends of biological particulate matter concentration in the air. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a biological early warning detection method for airborne particulate matter. This method utilizes lensless diffraction imaging technology to capture diffraction ring images in real time, employs gray-level co-occurrence matrix and wavelet transform to extract multi-scale features, combines logistic regression and SVM to achieve biological attribute identification and classification, and uses Poisson mutation detection and multi-level early warning response to concentration exceedances and abnormal trend changes. At the same time, a multi-source verification and feedback optimization mechanism is introduced to reduce the false judgment rate.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Methods for early warning and detection of biological effects of airborne particulate matter include:

[0008] A lensless diffraction imaging system was used to capture diffraction ring images of each particle in the area to be detected in real time.

[0009] Extract texture features from the gray-level co-occurrence matrix and wavelet multi-scale features, and fuse them to generate a real-time fusion vector; construct a historical fusion vector library based on historical data, and construct an index calculation model to calculate and generate a biological index sequence;

[0010] Biological index sequences are analyzed to identify the biological properties of particulate matter. Multi-class SVM is used for classification, and the number of events is counted to characterize the concentration. Poisson mutation analysis is combined with trend analysis. If the concentration exceeds the Poisson mutation detection threshold or the trend is abnormal, a graded warning is triggered; otherwise, monitoring continues.

[0011] By using a concentration adjustment feedback mechanism to screen for early warning errors, and combining multi-source data to verify and determine the type of misjudgment, the imaging equipment, exponential calculation model, and Poisson mutation detection threshold of the lensless diffraction imaging system are optimized in a categorized manner, and the optimization effect is verified regularly.

[0012] Specifically, the steps for generating the biological index sequence include:

[0013] Based on the real-time acquisition of diffraction ring images, features are extracted, gray-level co-occurrence matrices of each ring are constructed, texture feature vectors are generated, and wavelet transform is combined to construct multi-scale feature vectors and fuse them to generate real-time fused vectors.

[0014] Historical data is acquired, features are extracted and fused, and a historical fusion vector library is generated.

[0015] An index calculation model is constructed. The input layer receives real-time fused vectors, the weight calculation layer linearly combines vectors and weights, and the output layer outputs the biological index of the real-time fused vectors. The loss function is minimized through a historical fused vector library to optimize and obtain the optimal weight vector and bias term.

[0016] The real-time fusion vectors of each diffraction ring in the region to be detected are obtained and numbered to generate a vector sequence set of the region to be detected.

[0017] Using an index calculation model, the biological index of the real-time fusion vector is obtained, and a biological index sequence is generated.

[0018] Specifically, the steps for generating the real-time fusion vector include:

[0019] Preprocess the diffraction ring images of each particle, dynamically set the gray-level co-occurrence matrix parameters, and add radial and tangential directions in addition to the four angles of 0°, 45°, 90° and 180°.

[0020] Divide the annular region into equal parts Each ring band is used to extract contrast, energy, entropy, and correlation texture features;

[0021] The absolute difference in contrast between two adjacent rings in the same direction is calculated to obtain the ring texture gradient;

[0022] The correlation between adjacent rings is determined based on a preset gradient threshold, and the texture gradient of the correlated rings is associated with the current ring as an additional condition.

[0023] Obtain and calculate the absolute difference between the entropy values ​​in the radial and tangential directions to obtain the final directional entropy difference;

[0024] The internal structure of particulate matter is determined based on a preset entropy threshold, and particulate matter with disordered internal structure is eliminated.

[0025] The spacing within and between rings is set based on the average spacing of the diffraction rings;

[0026] Obtain the gray-level co-occurrence matrices of the intra-ring spacing and inter-ring spacing within the same ring zone, calculate the entropy of the two gray-level co-occurrence matrices and the entropy after averaging, and calculate the weighted average contribution value by combining the set entropy weighting coefficients.

[0027] Specifically, the steps for generating the real-time fusion vector further include:

[0028] The entropy and weighted average contribution values ​​of the two gray-level co-occurrence matrices are weighted and combined with square root operation to obtain the weighted Bach distance, which is used to measure the stability of the gray-level co-occurrence matrices.

[0029] If the stability is less than or equal to the preset stability threshold, the particle texture structure is determined to be highly similar and periodically stable at different spacings; otherwise, the particles are determined to be unstable and irregular in structure at different spacings and are discarded.

[0030] The Relief-F algorithm is used to preserve key features, which are then standardized and spliced ​​together to form the final texture feature vector.

[0031] The db4 wavelet was used to perform discrete wavelet transform on the preprocessed diffraction ring image to generate a two-level wavelet decomposition coefficient sequence.

[0032] For the components of the two-layer wavelet decomposition coefficient sequence, a double loop is used to traverse the pixels in the diffraction ring image of each component, calculate the wavelet coefficient value of each pixel, and sum them to obtain the energy of the corresponding component.

[0033] Calculate the standard deviation and mean of the corresponding components, and combine the energy, standard deviation and mean of all components in order to construct a wavelet domain multi-scale feature vector;

[0034] The final texture feature vector and wavelet domain multi-scale feature vector are concatenated and Z-score normalized to generate a real-time fused vector.

[0035] Specifically, the steps for analyzing the biological index sequence, identifying the biological properties of particulate matter, and classifying it using a multi-class SVM include:

[0036] Set a judgment threshold and determine the particulate matter category as abiotic particulate matter, biotic particulate matter, or particulate matter to be reviewed based on the biological index. Particulate matter to be reviewed requires manual intervention.

[0037] Based on the judgment results, a set of abiotic particles and a set of biotic particles are generated;

[0038] For a set of biological particles, the real-time fusion vector corresponding to each particle in the set of biological particles is saved to a set of biological feature vectors;

[0039] Construct an SVM multi-classification model;

[0040] The SVM multi-classification model is trained by calling up a biological particle sample database and inputting the multi-dimensional feature vectors of the samples to learn the mapping relationship between feature vectors and categories.

[0041] Traverse the biological feature vector set, extract the feature vectors of biological particles in turn, input them into the trained SVM multi-classification model, output the category to which each biological particle belongs, and label the category of each biological particle in the biological particle set.

[0042] Specifically, the steps for combining Poisson mutation analysis to analyze the trend of change include:

[0043] Based on the biological particulate matter category labeling, the concentration values ​​of the same type and the number of events of each type of biological particulate matter are counted, and the concentration of classified biological aerosols is obtained by dividing by the volume of the area to be detected. The real-time concentration sequence is generated by sorting by the time dimension.

[0044] If the concentration of classified biological aerosols is greater than or equal to the preset concentration standard threshold, the current concentration is determined to exceed the standard, and a graded warning is triggered; otherwise, the current concentration is determined not to exceed the standard, and the concentration change trend is judged.

[0045] The diffraction ring images of the area to be detected are continuously acquired to obtain an image sequence, and the corresponding biological particle set is acquired simultaneously to obtain the concentration values ​​of various biological particles in each diffraction ring image;

[0046] Based on the concentration values ​​of various biological particles, calculate the concentration difference between the current time and the previous time, and divide it by the concentration value of the previous time to obtain the concentration change rate.

[0047] The number of events for various types of biological particulate matter follows a Poisson distribution. The number of events and the expected number of events corresponding to the Poisson distribution probability are determined, and the probability within the current time period is calculated by combining the results.

[0048] Specifically, the steps for combining Poisson mutation analysis to analyze the trend of change also include:

[0049] An adaptive sliding window strategy is used to select historical windows, and the baseline window length, minimum window size, and maximum window size are set.

[0050] Calculate the JS divergence between the current window and the history windows. If the JS divergence is greater than the preset divergence threshold, shorten the window to the smallest window; otherwise, expand the window to the largest window.

[0051] If the probability within the current time period is less than the preset significance threshold, it is marked as a suspected mutation event; otherwise, it is judged as normal fluctuation and the historical window continues to be updated.

[0052] Based on the concentration mutation judgment, if the absolute difference between the expected number of the current event and the expected number of the historical events is greater than the preset Poisson mutation detection threshold and the probability within the current time period is less than the preset significance threshold, it is judged as a concentration Poisson mutation event; otherwise, it is judged as normal fluctuation and the historical window continues to be updated.

[0053] If the concentration change rate is greater than the preset concentration change rate threshold and a concentration Poisson mutation event is detected, it is determined that the concentration is on an upward trend, triggering a graded warning; otherwise, it is determined that there is no upward trend, and concentration fluctuations continue to be monitored.

[0054] Specifically, the steps for triggering the tiered early warning include:

[0055] Receive concentration warning signal and trend anomaly warning signals ;

[0056] This indicates that a concentration warning has been triggered. This indicates that it has not been triggered.

[0057] Only when The difference between the actual concentration and the preset warning threshold is calculated and divided by the preset warning threshold to obtain the concentration warning quantification value.

[0058] This indicates that an abnormal trend warning has been triggered. This indicates that it has not been triggered.

[0059] Only when The difference between the actual trend change rate and the normal trend rate is calculated and divided by the normal trend rate to obtain the trend warning quantitative value.

[0060] Concentration warnings have higher priority than trend warnings, and the following rules should be followed when integrating warnings:

[0061] when When the trend warning is used as an auxiliary positive term, the final warning value is the product of the warning concentration quantification value and the trend warning quantification value and the weight.

[0062] when At that time, the final warning value is the trend warning quantification value;

[0063] Three threshold levels are set and combined with the final warning value to trigger tiered warnings, including Level 3 warning, Level 2 warning, Level 1 warning, and no warning.

[0064] Specifically, the steps for determining the type of misjudgment include:

[0065] Obtain concentration adjustment feedback data after the early warning, and determine the error type based on the concentration adjustment and concentration change: over-warning error, ineffective adjustment error, missed reporting error, and false alarm trend error.

[0066] The original diffraction ring image corresponding to the error time, the biological index recalculated based on the final texture feature vector and wavelet domain multi-scale feature vector, and the Poisson mutation detection statistic are extracted and used as multi-source data input to the verification decision tree to determine the misjudgment type according to priority.

[0067] For the original diffraction ring image, the Laplacian variance is used to determine whether it is blurred; if the Laplacian variance is less than the preset blur threshold, it is determined that the image is blurred and marked as a data quality misjudgment.

[0068] If the absolute difference between the recalculated biological index and the biological index is greater than the preset index error, it is marked as a feature fusion misjudgment.

[0069] Recalculate the Poisson mutation event detection statistic. If the absolute difference between the statistic and the expected number of events is less than or equal to the preset allowable error, it is marked as a mutation detection misjudgment.

[0070] Specifically, the steps for periodically verifying the optimization effect include:

[0071] When a data quality misjudgment is detected, the imaging device's self-cleaning function is triggered.

[0072] When a misclassification is identified as a feature fusion error, the final texture feature vector and wavelet domain multi-scale feature vector of the misclassified sample are extracted, the real-time fusion vector is updated, and the exponential calculation model is retrained.

[0073] When a mutation detection misjudgment is identified, the Poisson mutation detection threshold is dynamically adjusted.

[0074] Calculate the false positive rate. If the false positive rate is greater than the preset false positive rate threshold, increase the Poisson mutation detection threshold; otherwise, decrease the Poisson mutation detection threshold.

[0075] Each processing The system detects and evaluates the false alarm rate and early warning response time in response to early warning events.

[0076] If the false positive rate is less than the preset false positive rate threshold and the warning response time is less than the preset response time threshold, then the optimization effect of the imaging equipment, exponential calculation model and Poisson mutation detection threshold of the current lensless diffraction imaging system is deemed to meet the standard; otherwise, optimization continues.

[0077] The beneficial effects of this invention are:

[0078] 1. Real-time particulate matter images are acquired using a lensless diffraction imaging system. Gray-level co-occurrence matrix texture features and wavelet multi-scale features are extracted and fused to generate a biological index sequence using an exponential calculation model. This accurately distinguishes biological attributes, and multi-category SVM is used to further subdivide biological particulate matter. Poisson mutation analysis is combined with concentration trend analysis, and an adaptive sliding window and JS divergence optimization window are used to double-verify abnormal trends. Graded warning values ​​are generated according to the logic of concentration warning priority and trend warning auxiliary, reducing false alarms and missed alarms.

[0079] 2. A closed-loop optimization system is constructed through a concentration adjustment feedback mechanism: Early warning errors are screened based on whether concentration adjustments are performed, and error types are further subdivided; multi-source data is then extracted, and a validation decision tree is used to accurately determine the three root causes of misjudgments: data quality, feature fusion, and mutation detection; finally, targeted optimization and regular verification of the misjudgment rate and early warning response time are implemented. This closed loop of error screening, root cause localization, classification optimization, and effect verification continuously adapts to changes in scenarios and improves stability and reliability. Attached Figure Description

[0080] Figure 1 This is a flowchart illustrating the biological early warning detection method for airborne particulate matter disclosed in this application;

[0081] Figure 2 This is a flowchart illustrating the identification of the biological properties of particulate matter in this application;

[0082] Figure 3 This is a flowchart illustrating the triggering of tiered early warnings in this application;

[0083] Figure 4 This is a flowchart for determining the type of misjudgment in this application. Detailed Implementation

[0084] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0085] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a method for early warning detection of biological effects of airborne particulate matter, including the following steps:

[0086] Step S1: Using a high-speed triggered lensless diffraction imaging system, capture diffraction ring images of each particle in the area to be detected in real time;

[0087] Step S2: Spatial texture features are extracted using gray-level co-occurrence matrix, multi-scale detail features are obtained by combining wavelet transform, real-time fusion vectors are generated by fusion, and historical data is obtained to build a historical fusion vector library. An exponential calculation model is constructed through an exponential calculation model, and a biological exponential sequence is generated through the exponential calculation model.

[0088] Step S3: Identify the biological attributes of particulate matter by analyzing the biological index sequence, and classify biological particulate matter using an SVM multi-classification model. Count the number of events to characterize the concentration, and analyze the trend of change by combining Poisson mutation detection. When the concentration exceeds the Poisson mutation detection level or the trend of change is abnormal, a graded early warning is triggered; otherwise, monitoring continues.

[0089] Step S4: Screen early warning errors through concentration adjustment feedback mechanism, combine multi-source data verification and determine the type of misjudgment, classify and optimize the imaging equipment, exponential calculation model and Poisson mutation detection threshold of lensless diffraction imaging system, and verify the optimization effect regularly.

[0090] Specifically, in step S2, the specific steps for constructing the biological index sequence include:

[0091] Based on real-time acquired diffraction ring images, feature extraction is performed to construct gray-level co-occurrence matrices for each ring, thereby generating texture feature vectors. Simultaneously, wavelet transform is combined to construct multi-scale feature vectors, and feature fusion is performed to generate real-time fused vectors.

[0092] Historical data, including known and unknown biological particulate matter, is acquired and its features are extracted and fused to generate a historical fusion vector library.

[0093] An index calculation model is constructed using logistic regression, including an input layer, a weight calculation layer, and an output layer.

[0094] The input layer is used to receive real-time fused vectors as input to the exponential calculation model;

[0095] The weight calculation layer is responsible for weighted summation of the input real-time fusion vector and adding a bias term to achieve a linear combination of the input vector and the weight vector. The specific process includes: first, normalizing and calibrating each feature component of the real-time fusion vector to ensure that each feature component is on the same order of magnitude and to avoid feature weight shift; second, initializing the weight coefficients of each feature component based on the correlation between each feature and the bioactivity index in the historical fusion vector library, with higher correlations resulting in larger weight coefficients; next, multiplying each feature component of the normalized and calibrated real-time fusion vector by its corresponding weight coefficient and summing the results to obtain the linear combination value; finally, adding a bias term, with the initial value of the bias term set to 0.1 based on industry expert experience to correct linear fitting bias and complete the linear transformation of the input vector.

[0096] The output layer outputs the biometric index of the real-time fused vector and optimizes the weight vector by minimizing the loss function and using a historical fused vector library. and bias terms ;

[0097] The training set consisted of 80% of the samples in the historical fusion vector database, and the validation set consisted of 20% of the samples. In the 80% of samples, the known biological particulate fusion vectors were labeled with corresponding biological indices, while the unknown biological particulate fusion vectors were labeled with 0. The cross-entropy loss function was used as the loss function, and the index calculation model was trained using an Adam optimizer with an initial learning rate of 0.001, a decay coefficient of 0.95, and a batch size of 32. During the training process, the accuracy of the index calculation model was verified every 100 iterations until the accuracy of the validation set stabilized above 95%. After training, the real-time fusion vector to be detected was input into the index calculation model. The model received the vector through the input layer, transformed it through a linear combination of weights in the weight calculation layer, and mapped it through the output layer. The output was the biological index of the corresponding real-time fusion vector, with a value ranging from 0 to 1. The closer the biological index was to 1, the stronger the biologicality of the particulate matter.

[0098] Since there are multiple diffraction ring images in the region to be detected, the real-time fusion vectors of each diffraction ring in the region to be detected are obtained and numbered to generate a vector sequence set of the region to be detected. ,in, For the first Real-time fusion vector of diffraction rings, ,and The total number of diffraction rings is given; the real-time fusion vectors are output using an exponential calculation model. Biological index And generate sequences of biological indices. .

[0099] Specifically, the steps for generating the real-time fusion vector include:

[0100] The acquired particle-by-particle diffraction ring images are preprocessed, including adaptive grayscale compression and ring mask extraction, to remove background interference;

[0101] The gray-level co-occurrence matrix parameters are dynamically set, and radial and tangential directions are added in addition to the traditional angles to enhance the sensitivity to the orientation of biological particle structure. The traditional angles are fixed directions for extracting texture features from the gray-level co-occurrence matrix, including 0°, 45°, 90°, and 180°, which correspond to the pixel arrangement directions of horizontal, upper right to lower left, vertical, and upper left to lower right, respectively.

[0102] Divide the annular region into equal parts Each annular band is used to extract contrast from the gray-level co-occurrence matrix of the corresponding region. ,energy ,entropy and correlation Four basic texture features;

[0103] Obtain two adjacent rings in the same direction respectively Contrast on and The absolute difference between the two contrast values ​​is calculated to obtain the final annular texture gradient. To quantify the trend of texture change from the center to the edge, the expression is as follows:

[0104] ;

[0105] in, For a ring belt in direction The contrast is below. Adjacent rings are in the same direction The contrast ratio is lower;

[0106] The value reflects the magnitude of texture variation. Based on the statistical distribution of historical biological particle diffraction ring texture gradients, the 90th percentile is selected as the benchmark to set the gradient threshold. , The value range is 0.15~0.25. In this embodiment, due to the gradual change in the diffraction ring texture of biological particles and the statistical result of the 90th percentile value being 0.2, A value of 0.2 is used to determine whether adjacent circumferential zones are correlated; if If the adjacent rings are not correlated, then they are determined to be correlated; otherwise, they are determined to be correlated. This is associated with the current annulus as an additional condition;

[0107] Obtain the entropy values ​​in the radial and tangential directions respectively. and Calculate the absolute difference between the two entropy values ​​to obtain the final directional entropy difference. The expression is as follows:

[0108] ;

[0109] An entropy threshold is set based on the statistical characteristics of directional entropy difference corresponding to the ordered internal structure of biological particles. , The value range is 0.10~0.20. In this embodiment, since the absolute difference between the radial and tangential entropy values ​​of the diffraction rings of biological particles is concentrated around 0.20, and the entropy threshold can effectively distinguish between highly ordered biological particles and disordered impurities, then... A value of 0.20 is used to determine the orderliness of the internal structure of particulate matter; if If the internal structure of the particulate matter is determined to be highly ordered, then the particulate matter is determined to be disordered and the particulate matter is discarded.

[0110] Based on the average spacing of the diffraction rings, the intra-ring spacing and inter-ring spacing are set, where the intra-ring spacing is half of the average diffraction ring spacing, and the inter-ring spacing is equal to the average diffraction ring spacing. Within the same ring zone, the gray-level co-occurrence matrices of the intra-ring spacing and inter-ring spacing are obtained, and labeled as follows: and ;calculate and entropy and synchronous calculation and Entropy of the average gray-level co-occurrence matrix Set the entropy weighting coefficient, combined with The weighted average contribution value was calculated by analyzing the two entropies. The weighted average contribution value is then weighted and calculated, and the result is obtained through a square root operation. and Weighted Bach distance between To measure the stability of the co-occurrence matrix The expression is as follows:

[0111] ;

[0112] in and They are respectively with and The relevant weighting coefficients are set based on the differential analysis of the contribution of intra-ring spacing and inter-ring spacing to texture features. Based on the actual characteristic that the inter-ring spacing is more sensitive to changes in diffraction ring texture than the intra-ring spacing, the following settings are made: ; These are joint weighting coefficients used for correction. and and The system deviation between them, the impact of quantified texture consistency on overall stability, and the setting based on the physical properties of the high correlation between the texture within and between biological particle diffraction rings. ;

[0113] Based on the stability of historical biological particle diffraction rings The statistical distribution was used as a benchmark to set a stability threshold. , The value range is 0.10~0.20. In this embodiment, due to the high consistency of the texture within and between the biological particle diffraction rings, the stability is... If the concentration is around 0.15, then... The value is 0.15, used to distinguish between high and low stability; if If the particle texture structure is highly similar and periodically stable under both intra-ring spacing and inter-ring spacing, then it is determined that the particle texture structure is highly similar and periodically stable. If the particle texture is unstable and irregular at different spacings, the particles will be removed.

[0114] The Relief-F algorithm is used to remove redundant features and retain key features, including , ,energy Correlation as well as After standardization, the data are spliced ​​and merged to form the final texture feature vector. ; Comprehensive characterization of the internal structural properties of biological particles; among which, Include Values ​​and additional conditions;

[0115] Wavelet transform is used to decompose the detail information of diffraction ring images at different scales to capture the subtle differences of biological particles in the frequency domain. The preprocessed diffraction ring images are subjected to two-level two-dimensional discrete wavelet transform using db4 wavelet. Each level is decomposed into a low-frequency approximation component LL and three high-frequency detail components, including horizontal LH, vertical HL, and diagonal HH, thereby generating a two-level wavelet decomposition coefficient sequence.

[0116] For each component in the two-level wavelet decomposition coefficient sequence, a double loop is used to traverse every pixel in the image of each component, calculating the wavelet coefficient value of each pixel. The energy of the corresponding component is obtained based on the sum of the wavelet coefficient values ​​of all pixels. Simultaneously, for the wavelet coefficient values ​​of all pixels within each component, the arithmetic mean of the wavelet coefficient values ​​is calculated as the mean, and the square root of the average of the sum of squared deviations of each wavelet coefficient value from the mean is calculated as the standard deviation. The statistical features of all components are combined sequentially to construct a multi-scale feature vector in the wavelet domain. Among them, statistical characteristics include the energy, mean, and standard deviation of the corresponding components;

[0117] For the final texture feature vector With wavelet domain multi-scale eigenvectors Feature concatenation is performed, and the concatenated feature vectors are standardized using Z-score to eliminate dimensional differences and generate real-time fused vectors.

[0118] Specifically, in step S3, the specific steps of analyzing the biological index sequence, identifying the biological attributes of particulate matter, and classifying it using an SVM multi-classification model include:

[0119] Based on the statistical distribution of bioactivity indices of historical biological and non-biological particulate matter samples, a judgment threshold is set using the boundary of the 95% confidence interval for both types of samples as a benchmark. , The value range is 0.15~0.25. In this embodiment, due to the difference in the distribution of the bioactivity index between historical biological particulate matter samples and historical non-biological particulate matter samples, the boundary value of the 95% confidence interval corresponds to 0.2. The value is 0.2, combined with the biological index. To determine the particle type of each diffraction ring, including abiotic and biotic particles; if If the biological characteristics do not meet the judgment criteria, it is judged as non-biological particulate matter; if If the biological characteristics meet the criteria, it is classified as biological particulate matter; if If the biometrics are low, it indicates that the biological characteristics are difficult to distinguish clearly, making it difficult to determine whether the organism is biological or non-biological, and the confidence level of the biometric index is low. Therefore, the particulate matter is marked as needing to be reviewed and artificial intervention is required.

[0120] Based on the judgment results, sets of abiotic particles and biotic particles are generated. Since abiotic particles do not possess dynamic characteristics, the abiotic particle set is not processed. For the biotic particle set… Collect biological particles The real-time fusion vectors corresponding to each particulate matter are saved to a set of biological feature vectors. In China; among them For the first Individual biological particles, For the first The feature vector corresponding to each biological particulate matter. , This represents the total number of biological particulate matter.

[0121] Using a one-to-many approach as a multi-class classification strategy, an SVM multi-class classification model with a multinomial kernel function is employed to classify the multidimensional feature vectors of biological particles. Mapping to a high-dimensional feature space completes the feature space mapping;

[0122] Call the database of biological particle samples with known category labels, input the multidimensional feature vectors of the samples to train the SVM multi-classification model, and learn the mapping relationship between feature vectors and categories;

[0123] Traversing the set of biological feature vectors The feature vector corresponding to each biological particulate matter is extracted sequentially and input into a trained SVM multi-classification model to output the specific category to which each biological particulate matter belongs, and to classify the biological particulate matter set. Each biological particle in the set is categorized and labeled to determine the category of all biological particles in the set.

[0124] Specifically, in step S3, the specific steps for combining Poisson mutation analysis to determine the trend include:

[0125] Based on the labels of each biological particle in the biological particulate matter ensemble, the concentration values ​​of biological particles of the same type are counted, and the event counts of each type of biological particulate matter are combined. The volume of the area to be detected is obtained through equipment calibration. Calculate the concentration of classified biological aerosols. :

[0126] ;

[0127] in The time dimension is used to distinguish statistical results from different time windows; different time periods are used to differentiate statistical results from different time windows. Corresponding classification of biological aerosol concentration Arranged in chronological order, generating real-time concentration sequences. ;

[0128] Based on the statistical distribution of historical concentration data, concentration standard thresholds are set according to industry standards and the statistical quantiles of historical exceedance events. , The range of values ​​is the national standard limit for the corresponding target biological particulate matter. The range, such as that of common mold-based bioaerosols, is 50~150. In this embodiment, since the area to be tested is an indoor public place, the corresponding national standard limit is 100. Furthermore, if the 90th percentile of historical exceedance events corresponds to the national standard limit, then... Value 100 To determine whether the concentration exceeds the standard; if Then determine the current time. The concentration exceeded the standard, triggering a tiered warning; if If the concentration does not exceed the standard, then the concentration change trend is judged.

[0129] In order to identify the changes in various biological particles, all diffraction ring images of the area to be detected are continuously acquired to obtain a diffraction ring image sequence of the area to be detected, and the corresponding biological particle set is obtained for each diffraction ring image, thereby obtaining the concentration values ​​of various biological particles in the corresponding diffraction ring image.

[0130] Calculate the rate of change of concentration based on the concentration values ​​of various biological particles; determine the current time. Compared to the previous moment The concentration difference between the two values ​​is divided by the concentration value at the previous time. The concentration change rate was obtained. ;

[0131] ;

[0132] Number of events involving various types of biological particulate matter Constructing the probability mass function, which follows a Poisson distribution:

[0133] ;

[0134] in, The expected number of events within the current time period; determine the number of events corresponding to the calculated probability. Calculate the exponential part separately and power part The product of the two parts divided by The factorial gives the probability within the current time period. Used to quantify the number of actual events within the current time period. The degree of abnormality;

[0135] An adaptive sliding window strategy is used to select historical windows, and the baseline window length is set based on prior knowledge. Minimum window and the largest window The JS divergence is obtained by calculating the difference in probability distributions between the current window and historical windows using the formula for calculating JS divergence. The divergence threshold was determined using Monte Carlo simulation. , The value range is 0.10~0.20. In this embodiment, since the critical divergence value at 95% confidence level in Monte Carlo simulation is 0.15, then... The value is 0.15; if Then shorten the window to the smallest window. Otherwise, expand the window to the maximum window size. The Monte Carlo simulation process includes: generating a large number of randomly sampled window data samples based on historical window data; calculating the JS divergence statistics between each window data sample; extracting the divergence critical value with a 95% confidence level; and determining the divergence critical value as the divergence threshold. ;

[0136] Based on the probability distribution of historical windows, a significance threshold is set according to the significance level requirements of statistical hypothesis testing. , The value range is 0.01~0.05. In this embodiment, since the detection of sudden changes in bioaerosol concentration needs to consider both the false negative rate and the false positive rate, the industry-standard significance level is 0.05. The value is 0.05; if This indicates the number of events within the current time period. If the event deviates significantly from the historical expectation, it is marked as a suspected mutation event; otherwise, it is judged as a normal fluctuation event, and the historical window is updated continuously.

[0137] Combined with the formula for judging concentration mutation Further confirmation is needed to determine whether this is a concentration Poisson mutation event; among which, For the expected number of historical events, The absolute difference of the expected number. The Poisson mutation detection threshold is set based on the statistical standard of the number of events in historical periods without mutations. The value range is 2~5. This embodiment requires high detection sensitivity and requires minimal historical data fluctuation. The value is 4;

[0138] If both conditions are met Furthermore, the absolute difference in the expected number is greater than the Poisson mutation detection threshold. If the result is positive, it is determined to be a concentration Poisson mutation event; otherwise, it is determined to be a normal fluctuation event, and the historical window is updated continuously.

[0139] A concentration change rate threshold is set according to industry control requirements for bioaerosol concentration early warning. , The value range is 10% to 30%. In this embodiment, since the area to be detected is an indoor public place, the industry requirement is that an increase in concentration of 20% triggers an alarm. Take the value 20%; if If a Poisson mutation event is detected, it is determined that the concentration is on an upward trend, and a graded warning is triggered; otherwise, it is determined that there is no upward trend, and the fluctuation of biological particulate matter concentration continues to be monitored.

[0140] Specifically, in step S3, the specific steps for triggering the graded early warning include:

[0141] Receive warning category signals, including concentration warnings and abnormal trend warnings; set concentration warning signals. Used to indicate concentration warning status, among which This indicates that a concentration warning has been triggered. This indicates that it has not been triggered;

[0142] Simultaneously set a concentration warning quantification value. This indicates that the concentration exceeds the warning threshold. The degree, and only if Calculations are performed in real time to determine the concentration warning quantification value. The expression is as follows:

[0143] ;

[0144] in, This represents the actual concentration. This is the difference between the actual concentration and the warning threshold; the warning threshold is set according to industry control standards for particulate matter and environmental requirements of the area to be monitored. ; The range of values ​​is In this embodiment, since the area to be detected is an indoor public environment, and considering the requirements for biological particulate matter control, then... Value ;

[0145] Set up trend anomaly warning signal ,in This indicates that an abnormal trend warning has been triggered. The time indicates that it has not been triggered; at the same time, a trend warning quantification value is set. This reflects the degree of abnormality in the rate of change of the trend, and only when... Calculations are performed in real time to determine the quantitative value of trend warning. The expression is as follows:

[0146] ;

[0147] in, This represents the actual rate of change of the trend. This is the normal trend rate;

[0148] Since concentration warnings directly reflect the current state, they have higher priority than trend warnings, which reflect the pattern of change. The following rules are followed when integrating warnings: When a concentration warning signal... At that time, trend warning is used as an auxiliary positive factor, combined with concentration warning, through weighting. The impact of dynamic adjustment trend warning on the overall warning level; when the concentration warning signal At that time, the warning strength is determined solely by the trend warning value; the expression is as follows:

[0149] ;

[0150] in, This is the final warning value. The weight of trend warnings is determined based on the priority difference between concentration warnings and trend warnings, and the warning response requirements of biological particulate matter monitoring scenarios. ; The value range is 0.2~0.4. In this embodiment, because it is necessary to strengthen the leading role of concentration early warning and weaken the auxiliary influence of trend early warning, The value is 0.3;

[0151] Three threshold levels were set based on expert experience. , and ,in and Based on the three-level threshold Mapping of early warning levels; when When it is determined to be a Level III warning, an immediate response and emergency measures are required; when When it is determined to be a Level II warning, close monitoring is required; when When it is determined to be a Level 1 warning, there is a potential risk and continuous monitoring is required; when At that time, it was determined that there was no warning and the current status was normal.

[0152] Specifically, step S4 includes the following steps:

[0153] Obtain concentration adjustment feedback data after the warning, and filter out the warning error cases based on whether there is a concentration adjustment step;

[0154] When adjusting the concentration, if the adjusted concentration is lower than the warning threshold... If the concentration does not change after adjustment, and adjustment continues, it is judged as an excessive warning error; if the concentration does not change after adjustment, it is judged as an invalid adjustment error.

[0155] If a concentration increase triggers an alert without concentration adjustment, it is considered a missed error; if a trend alert misjudges a normal fluctuation, it is considered a false trend error.

[0156] Extracting multi-source data includes: retrieving the original diffraction ring image corresponding to step S1 at the error time and checking the image quality; extracting the final texture feature vector and wavelet domain multi-scale feature vector of particles corresponding to step S2, and recalculating the bioactivity index. ; and obtain the statistics for Poisson mutation detection, thereby filtering out misjudged data in the early warning;

[0157] Multi-source data is input into the verification decision tree, and the misjudgment type of the misjudged data is specifically determined according to priority, namely, data quality misjudgment, feature fusion misjudgment, and mutation detection misjudgment.

[0158] For the original diffraction ring image, the blur threshold is obtained through device calibration. And using Laplace variance Determine if it is ambiguous; if If the image is blurred, the feature extraction error is determined to be due to the image blur and marked as a data quality misjudgment; otherwise, it is determined that there is no data quality misjudgment.

[0159] Recalculate the biological index in step S2 Based on the historical statistical fluctuation range of biological indices and the required detection accuracy, the index error was set to [value missing]. , The value range is 3% to 8% of the historical average of the biological index. In this embodiment, since the detection accuracy requirement is 5%, then... The value is taken as 5% of the historical average of the biological index; if the biological index is recalculated... With biological index The absolute difference between them exceeds ,Right now If the condition is met, it is marked as a feature fusion misjudgment; otherwise, it is determined that there is no feature fusion misjudgment.

[0160] Recalculate the statistics for Poisson mutation event detection Based on the statistical characteristics of the Poisson distribution and the confidence requirements for mutation detection, the allowable error is set to [value missing]. , The value range is 1.5 to 2.5 times the historical standard deviation of the Poisson statistic. In this embodiment, since the allowable error for the statistics at a 95% confidence level is 2 times the historical standard deviation of the Poisson statistic, then... The value is twice the historical standard deviation of the Poisson statistic; if If the result is positive, it is determined that the normal fluctuation was mistakenly identified as a mutation in the original detection and marked as a mutation detection misjudgment; otherwise, it is determined that there is no mutation detection misjudgment.

[0161] The verification results are categorized and labeled according to the root cause of the misjudgment, and corresponding measures are taken: when it is labeled as a data quality misjudgment, the imaging equipment self-cleaning is triggered;

[0162] When a sample is marked as a misclassified feature fusion error, the final texture feature vector of the misclassified sample is extracted. With wavelet domain multi-scale eigenvectors Update the real-time fusion vector in step S2 and retrain the exponential calculation model, so that... ;

[0163] When a mutation detection misclassification is detected, the Poisson mutation detection threshold is dynamically adjusted. : Set the current Poisson mutation detection threshold Multiplying by the adjustment factor yields the updated Poisson mutation detection threshold. ;

[0164] ;

[0165] in, For the false positive rate, As an adjustment factor, this embodiment sets the adjustment factor based on the statistical distribution characteristics of the misjudgment rate and the step size control strategy of dynamic adjustment of the threshold.

[0166] The false positive rate threshold is set based on the industry false positive rate control standard for Poisson mutation detection and the statistical average of historical early warning events. , The value range is 5% to 15%. In this embodiment, the optimization goal of the detection system is to reduce the false positive rate and improve the recognition accuracy. Take the value 10%; if Then expand Otherwise shrink ;

[0167] Each processing Warning events were detected, and the false positive rate was verified. Regarding the early warning response time, a response time threshold is set based on the industry acceptance standards for lensless diffraction imaging systems and the latency requirements for real-time detection. , The range of values ​​is In this embodiment, due to the need to meet the response requirements of high-precision real-time early warning, Value ;like Less than and the early warning response time is less than If the optimization of the imaging equipment, exponential calculation model, and Poisson mutation detection threshold of the current lensless diffraction imaging system is satisfactory, then the optimization will continue.

[0168] In summary, this invention combines lensless diffraction imaging with intelligent analysis to achieve real-time identification, concentration monitoring, and adaptive control of biological particles in the air. First, in the data acquisition phase, a high-speed triggered lensless imaging system is used to capture real-time images of diffraction rings within the detection area, one particle at a time. Based on image features, a biological discrimination model integrating spatial texture and frequency domain details is constructed. In the feature extraction phase, a gray-level co-occurrence matrix is ​​used to extract the spatial texture features of the diffraction rings, introducing radial and tangential direction parameters to enhance sensitivity to the structural directionality of biological particles. The annular region is divided into multiple rings, and the contrast, energy, entropy, and correlation of each ring are calculated. A ring texture gradient is also designed. directional entropy difference and co-occurrence matrix stability Features such as [list of features]. Simultaneously, a db4 wavelet is used for two-level decomposition to extract the energy, mean, and standard deviation of each component, constructing multi-scale frequency domain features; redundant features are removed using the Relief-F algorithm, ultimately generating a standardized real-time fusion vector; simultaneously, a historical fusion vector library is constructed to train the index calculation model, outputting a biological index sequence to achieve quantitative assessment of the biological attributes of each particulate matter. Based on the biological index sequence, biological attribute identification is performed, setting a dual-threshold mechanism to classify particulate matter into three categories: biological, non-biological, or pending verification. For the biological particulate matter set, a one-to-many strategy is further used to construct an SVM multi-classification model, combined with a multinomial kernel function to achieve high-dimensional space classification, completing the determination of the specific category of biological particulate matter. To achieve dynamic concentration monitoring, the number of events for each type of biological particle within a unit time window is counted, and the aerosol concentration of the classification is calculated based on the detection area volume, constructing a real-time concentration sequence. Furthermore, a Poisson mutation detection mechanism is introduced, selecting historical data based on an adaptive sliding window, calculating the expected number of events and assessing the anomaly probability of the current event count, and dynamically adjusting the window length using JS divergence to improve detection sensitivity. The early warning mechanism is triggered when the concentration exceeds a preset threshold or when a significant Poisson mutation is detected with an upward trend. The mechanism employs a tiered response strategy, combining the degree of concentration exceeding the standard with the quantitative value of abnormal trends. A final early warning value is generated through weighted fusion and mapped to Level 1, Level 2, and Level 3 warning levels, with corresponding measures implemented to achieve refined risk management. Finally, a multi-source verification mechanism is introduced to retrospectively analyze early warning events, identifying error types such as image blurring, feature fusion deviation, or misjudgment of mutations. Based on this, it triggers equipment self-cleaning, model retraining, or threshold adaptive optimization, forming a closed-loop feedback loop to continuously improve accuracy and stability.

[0169] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for early warning and detection of biological effects of airborne particulate matter, characterized in that, include: A lensless diffraction imaging system was used to capture diffraction ring images of each particle in the area to be detected in real time. Extract texture features from the gray-level co-occurrence matrix and wavelet multi-scale features, and fuse them to generate a real-time fusion vector; construct a historical fusion vector library based on historical data, and construct an index calculation model to calculate and generate a biological index sequence; Biological index sequences are analyzed to identify the biological properties of particulate matter. Multi-class SVM is used for classification, and the number of events is counted to characterize the concentration. Poisson mutation analysis is combined with trend analysis. If the concentration exceeds the Poisson mutation detection threshold or the trend is abnormal, a graded warning is triggered. Otherwise, continue monitoring; By using a concentration adjustment feedback mechanism to screen for early warning errors, and combining multi-source data to verify and determine the type of misjudgment, the imaging equipment, exponential calculation model, and Poisson mutation detection threshold of the lensless diffraction imaging system are optimized in a categorized manner, and the optimization effect is verified regularly.

2. The method for early warning and detection of biological airborne particulate matter according to claim 1, characterized in that, The specific steps for generating the biological index sequence include: Based on the real-time acquisition of diffraction ring images, features are extracted, gray-level co-occurrence matrices of each ring are constructed, texture feature vectors are generated, and wavelet transform is combined to construct multi-scale feature vectors and fuse them to generate real-time fused vectors. Historical data is acquired, features are extracted and fused, and a historical fusion vector library is generated. An index calculation model is constructed. The input layer receives real-time fused vectors, the weight calculation layer linearly combines vectors and weights, and the output layer outputs the biological index of the real-time fused vectors. The loss function is minimized through a historical fused vector library to optimize and obtain the optimal weight vector and bias term. The real-time fusion vectors of each diffraction ring in the region to be detected are obtained and numbered to generate a vector sequence set of the region to be detected. Using an index calculation model, the biological index of the real-time fusion vector is obtained, and a biological index sequence is generated.

3. The method for early warning and detection of biological effects of airborne particulate matter according to claim 2, characterized in that, The specific steps for generating the real-time fusion vector include: Preprocess the diffraction ring images of each particle, dynamically set the gray-level co-occurrence matrix parameters, and add radial and tangential directions in addition to the four angles of 0°, 45°, 90° and 180°. Divide the annular region into equal parts Each ring band is used to extract contrast, energy, entropy, and correlation texture features; The absolute difference in contrast between two adjacent rings in the same direction is calculated to obtain the ring texture gradient; The correlation between adjacent rings is determined based on a preset gradient threshold, and the texture gradient of the correlated rings is associated with the current ring as an additional condition. Obtain and calculate the absolute difference between the entropy values ​​in the radial and tangential directions to obtain the final directional entropy difference; The internal structure of particulate matter is determined based on a preset entropy threshold, and particulate matter with disordered internal structure is eliminated. The spacing within and between rings is set based on the average spacing of the diffraction rings; Obtain the gray-level co-occurrence matrices of the intra-ring spacing and inter-ring spacing within the same ring zone, calculate the entropy of the two gray-level co-occurrence matrices and the entropy after averaging, and calculate the weighted average contribution value by combining the set entropy weighting coefficients.

4. The method for early warning and detection of biological effects of airborne particulate matter according to claim 3, characterized in that, The specific steps for generating the real-time fusion vector also include: The entropy and weighted average contribution values ​​of the two gray-level co-occurrence matrices are weighted and combined with square root operation to obtain the weighted Bach distance, which is used to measure the stability of the gray-level co-occurrence matrices. If the stability is less than or equal to the preset stability threshold, the particle texture structure is determined to be highly similar and periodically stable at different spacings; otherwise, the particles are determined to be unstable and irregular in structure at different spacings and are discarded. The Relief-F algorithm is used to preserve key features, which are then standardized and spliced ​​together to form the final texture feature vector. The db4 wavelet was used to perform discrete wavelet transform on the preprocessed diffraction ring image to generate a two-level wavelet decomposition coefficient sequence. For the components of the two-layer wavelet decomposition coefficient sequence, a double loop is used to traverse the pixels in the diffraction ring image of each component, calculate the wavelet coefficient value of each pixel, and sum them to obtain the energy of the corresponding component. Calculate the standard deviation and mean of the corresponding components, and combine the energy, standard deviation and mean of all components in order to construct a wavelet domain multi-scale feature vector; The final texture feature vector and wavelet domain multi-scale feature vector are concatenated and Z-score normalized to generate a real-time fused vector.

5. The method for early warning and detection of biological effects of airborne particulate matter according to claim 4, characterized in that, The specific steps for analyzing the biological index sequence, identifying the biological attributes of particulate matter, and classifying it using a multi-class SVM include: Set a judgment threshold and determine the particulate matter category as abiotic particulate matter, bioparticulate matter, or particulate matter pending review based on the bioactivity index; Based on the judgment results, a set of abiotic particles and a set of biotic particles are generated; For a set of biological particles, the real-time fusion vector corresponding to each particle in the set of biological particles is saved to a set of biological feature vectors; Construct an SVM multi-classification model; The SVM multi-classification model is trained by calling up a biological particle sample database and inputting the multi-dimensional feature vectors of the samples to learn the mapping relationship between feature vectors and categories. Traverse the biological feature vector set, extract the feature vectors of biological particles in turn, input them into the trained SVM multi-classification model, output the category to which each biological particle belongs, and label the category of each biological particle in the biological particle set.

6. The method for early warning and detection of biological effects of airborne particulate matter according to claim 5, characterized in that, The specific steps for combining Poisson mutation analysis to analyze the trend of change include: Based on the biological particulate matter category labeling, the concentration values ​​of the same type and the number of events of each type of biological particulate matter are counted, and the concentration of classified biological aerosols is obtained by dividing by the volume of the area to be detected. The real-time concentration sequence is generated by sorting by the time dimension. If the concentration of classified biological aerosols is greater than or equal to the preset concentration standard threshold, the current concentration is determined to exceed the standard, and a graded warning is triggered; otherwise, the current concentration is determined not to exceed the standard, and the concentration change trend is judged. The diffraction ring images of the area to be detected are continuously acquired to obtain an image sequence, and the corresponding biological particle set is acquired simultaneously to obtain the concentration values ​​of various biological particles in each diffraction ring image; Based on the concentration values ​​of various biological particles, calculate the concentration difference between the current time and the previous time, and divide it by the concentration value of the previous time to obtain the concentration change rate. The number of events for various types of biological particulate matter follows a Poisson distribution. The number of events and the expected number of events corresponding to the Poisson distribution probability are determined, and the probability within the current time period is calculated by combining the results.

7. The method for early warning and detection of biological effects of airborne particulate matter according to claim 6, characterized in that, The specific steps for combining Poisson mutation analysis to analyze trends also include: An adaptive sliding window strategy is used to select historical windows, and the baseline window length, minimum window size, and maximum window size are set. Calculate the JS divergence between the current window and the history windows. If the JS divergence is greater than the preset divergence threshold, shorten the window to the smallest window; otherwise, expand the window to the largest window. If the probability within the current time period is less than the preset significance threshold, it is marked as a suspected mutation event; otherwise, it is judged as normal fluctuation and the historical window continues to be updated. Based on the concentration mutation judgment, if the absolute difference between the expected number of the current event and the expected number of the historical events is greater than the preset Poisson mutation detection threshold and the probability within the current time period is less than the preset significance threshold, it is judged as a concentration Poisson mutation event; otherwise, it is judged as normal fluctuation and the historical window continues to be updated. If the concentration change rate is greater than the preset concentration change rate threshold and a concentration Poisson mutation event is detected, it is determined that the concentration is on an upward trend, triggering a graded warning; otherwise, it is determined that there is no upward trend, and concentration fluctuations continue to be monitored.

8. The method for early warning and detection of biological effects of airborne particulate matter according to claim 7, characterized in that, The specific steps for triggering the graded early warning include: Receive concentration warning signal and trend anomaly warning signals ; This indicates that a concentration warning has been triggered. This indicates that it has not been triggered. Only when The difference between the actual concentration and the preset warning threshold is calculated and divided by the preset warning threshold to obtain the concentration warning quantification value. This indicates that an abnormal trend warning has been triggered. This indicates that it has not been triggered. Only when The difference between the actual trend change rate and the normal trend rate is calculated and divided by the normal trend rate to obtain the trend warning quantitative value. Concentration warnings have higher priority than trend warnings, and the following rules should be followed when integrating warnings: when When the trend warning is used as an auxiliary positive term, the final warning value is the concentration warning quantification value plus the product of the trend warning quantification value and the weight. when At that time, the final warning value is the trend warning quantification value; Three threshold levels are set and combined with the final warning value to trigger tiered warnings, including Level 3 warning, Level 2 warning, Level 1 warning, and no warning.

9. The method for early warning and detection of biological effects of airborne particulate matter according to claim 8, characterized in that, The specific steps for determining the type of misjudgment include: Obtain concentration adjustment feedback data after the early warning, and determine the error type based on the concentration adjustment and concentration change: over-warning error, ineffective adjustment error, missed reporting error, and false alarm trend error. The original diffraction ring image corresponding to the error time, the biological index recalculated based on the final texture feature vector and wavelet domain multi-scale feature vector, and the Poisson mutation detection statistic are extracted and used as multi-source data input to the verification decision tree to determine the misjudgment type according to priority. For the original diffraction ring image, the Laplacian variance is used to determine whether it is blurred; if the Laplacian variance is less than the preset blur threshold, it is determined that the image is blurred and marked as a data quality misjudgment. If the absolute difference between the recalculated biological index and the biological index is greater than the preset index error, it is marked as a feature fusion misjudgment. Recalculate the Poisson mutation event detection statistic. If the absolute difference between the statistic and the expected number of events is less than or equal to the preset allowable error, it is marked as a mutation detection misjudgment.

10. The method for early warning and detection of biological effects of airborne particulate matter according to claim 9, characterized in that, The specific steps for periodically verifying the optimization effect include: When a data quality misjudgment is detected, the imaging device's self-cleaning function is triggered. When a misclassification is identified as a feature fusion error, the final texture feature vector and wavelet domain multi-scale feature vector of the misclassified sample are extracted, the real-time fusion vector is updated, and the exponential calculation model is retrained. When a mutation detection misjudgment is identified, the Poisson mutation detection threshold is dynamically adjusted. Calculate the false positive rate. If the false positive rate is greater than the preset false positive rate threshold, increase the Poisson mutation detection threshold; otherwise, decrease the Poisson mutation detection threshold. Each processing The system detects and evaluates the false alarm rate and early warning response time in response to early warning events. If the false positive rate is less than the preset false positive rate threshold and the warning response time is less than the preset response time threshold, then the optimization effect of the imaging equipment, exponential calculation model and Poisson mutation detection threshold of the current lensless diffraction imaging system is deemed to meet the standard; otherwise, optimization continues.

Images