Biological sample detection method and system based on AIOT

Abstract

The invention relates to the technical field of biological detection, and particularly discloses a biological sample detection method and system based on AIOT. According to the method, sample data is collected through a temperature and humidity sensor, a spectrum sensor, a geographic coordinate sensor and other multi-mode sensors, after preprocessing of sliding window filtering, z-score standardization and the like, a logic regression model is constructed, a mode gating mechanism is introduced to achieve feature level fusion, and a judgment threshold value is dynamically adjusted according to sample category distribution and business risks. The system adopts an edge computing and cloud analysis architecture, comprises a multi-modal acquisition module, a modal fusion engine and a distributed storage platform, and supports sensor expansion and cross-mechanism federated learning. The embodiment shows that the accuracy rate of the scheme reaches 92.5% and is improved by more than 10% compared with that of a single-mode model, the recall rate is improved by 40% in a sample imbalance scene, intelligence, automation and security of biological sample attribute judgment are achieved, and the method is suitable for scenes of wild animal monitoring, medical research and the like.

Description

Technical Field

[0001] The present invention relates to the field of biological detection technology, and in particular to a biological sample detection method and system based on AIOT. Background Art

[0002] Limitations of existing single-modal analysis in biological sample testing include: Relying solely on time-series data on temperature and humidity cannot distinguish differences in sample activity (e.g., samples collected at different altitudes are affected by air pressure); spectral analysis ignores differences in collection site (e.g., ear swabs and arm blood have different spectral signature distributions). Cross-modal correlations are missing: The potential impact of geographic coordinates (e.g., latitude) and climate type (e.g., monsoon regions) on sample pathogen carriage rates is not modeled; and the temporal correlation between collection timestamps (e.g., rainy season vs. dry season) and soil composition (e.g., heavy metal content) is not explored. Feature fusion is crude: Traditional methods directly splice multi-source data without considering data morphological differences between modalities (e.g., continuous-valued spectra vs. discrete labels), resulting in slow model convergence and weak generalization. Biological sample properties are influenced by multiple factors, but existing technologies lack differentiated modeling and deep fusion algorithms for time-series dynamic features, spectral material characteristics, and geographic environmental characteristics, making it difficult to capture nonlinear correlations between complex modalities.

[0003] Based on this, a biological sample detection method and system based on AIOT is now provided, which can eliminate the disadvantages of existing devices. Summary of the Invention

[0004] The purpose of the present invention is to provide a biological sample detection method and system based on AIOT, which solves the problem of inconvenience in use in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions: A biological sample detection method based on AIOT, comprising the following steps: Step 1: Multi-source data collection: Collect sensor data, equipment information, and geographic information of the sample, including: temperature time series obtained by temperature and humidity sensors and humidity time series ; pH sensor obtains the pH scalar value ; Spectral vector obtained by the spectral sensor ; Collection timestamp , Collection site classification label and collection method classification labels ; Geographic coordinates , altitude , climate type , soil composition vector and vegetation type labels ; Step 2: Data preprocessing: Clean, standardize and encode the data in step (1), including: using sliding window median filtering to denoise the temperature and humidity time series, replacing outliers exceeding ±3 times the standard deviation of the mean for spectral data; performing z-score standardization on numerical features , normalize the spectrum vector to Interval; One-hot encoding of classification labels; Step 3: Build a logistic regression model: based on the preprocessed feature vector , build a logistic regression model

[0006] Optimize parameters by maximum likelihood estimation , and solve it using gradient descent method or quasi-Newton method; Step 4: Sample attribute judgment: Based on the probability value output by the model , set the threshold ,when When , the sample is judged to belong to the target category, otherwise it belongs to other categories; Step 5: Accuracy feedback: Calculate the accuracy of the training set and test set to provide staff with a reference for model performance.

[0007] On the basis of the above technical solutions, the present invention also provides the following optional technical solutions: In an optional solution: in step 1: the acquisition frequency of the temperature time series and the humidity time series is once per minute; the spectral vector Dimension Corresponding wavelength range spectral intensity; The collection site classification label Includes at least one of "arms", "legs" and "ears".

[0008] In an optional solution: In step 2: the window size of the sliding window median filter is 5 to 10 data points; the spectral data normalization formula is: .

[0009] In one optional solution: In step 3: the parameters The optimization process includes: constructing the log-likelihood function ,in ; Iteratively update parameters by dynamically adjusting the learning rate gradient descent method , until the likelihood function converges.

[0010] In an optional solution: in step (5): the accuracy rate is calculated as follows:

[0011] in is the total number of samples, is the predicted category, For the true category; when the accuracy of the test set is lower than the preset threshold (such as 80%), the model parameter re-optimization process is triggered.

[0012] An AIOT-based biological sample detection system is characterized by including the following modules: a multi-source data acquisition module: including a temperature and humidity sensor, a pH sensor, a spectrum sensor, a GPS positioning unit, and an acquisition information input interface, for synchronously acquiring sensor data, device parameters, and geographic information of the sample; Data preprocessing module: including time series filter, outlier replacement unit and normalization processor, used for denoising, encoding and normalizing the raw data; Logistic regression model building module: equipped with a parameter optimization processor and a model storage unit, used to train and store the coefficient matrix of the logistic regression model; Attribute judgment module: contains probability calculation unit and threshold comparator, which is used to determine the sample category according to the model output probability; Feedback module: includes an accuracy calculator and an alarm trigger unit, which are used to output model performance indicators and trigger the retraining process.

[0013] In an optional solution: in the data preprocessing module: The time series filter is configured as a sliding window median filter with dynamically adjustable window size (5-10 points); The outlier replacement unit has a built-in standard deviation calculator, and the judgment conditions are ; The normalization processor supports both z-score normalization and Min-Max normalization modes.

[0014] In an optional solution: in the attribute judgment module: The threshold comparator supports dynamic threshold setting. The threshold adjustment rules are as follows: in and are the number of positive and negative samples respectively.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The solution of the present invention integrates multi-source data such as temperature, humidity, spectrum, and geographic coordinates to construct a hierarchical feature engineering and dynamic threshold adjustment mechanism, significantly improving the accuracy and adaptability of monkey biological sample attribute judgment: Multimodal fusion breakthrough: Integrating 10+ dimensional features such as temporal dynamics, spectral materials, and geographical environment to address the one-sidedness of traditional single-modal analysis, the measured accuracy rate has been increased to over 92%, an increase of over 10% compared to a single spectral model.

[0016] Enhanced adaptive capabilities: Through the modal gating mechanism, weights are automatically assigned (for example, the weight of the temporal modal in high-altitude scenarios is increased to 50%), and thresholds are dynamically adjusted based on sample category distribution and business risk. This improves the recall rate by 40% in scenarios with sample imbalance (the target category accounts for 15%), and the specificity for high-risk scenarios reaches 99%.

[0017] Full-process efficiency optimization: Automation is achieved from data collection to model inference, with single-sample processing at the edge taking less than 10ms. Cloud-based batch training efficiency is increased by 6.8 times, which is more than 20 times more efficient than traditional manual analysis. At the same time, blockchain evidence storage and federated learning are used to ensure data security and compliance, meeting the needs of cross-institutional collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a logic block diagram of the detection method of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0020] Scenario 1: Real-time monitoring of wildlife habitat samples Deployment environment: A national nature reserve (26°15′N, 118°30′E, 300-1200m above sea level, subtropical monsoon climate). The goal was to monitor the health of 500 macaques, focusing on the early identification of samples carrying mosquito-borne infectious diseases (such as monkeypox virus). Front-end acquisition system construction Intelligent sampling device: Main structure: Stainless steel waterproof shell (IP67 protection level), built-in: Timing sensor group: SHT45 temperature and humidity sensor (measurement range -40°C to 125°C, accuracy ±0.1°C / ±1.5% RH), collects temperature and humidity data at a frequency of 1 time per second and is equipped with a 32KB ring buffer to store the last 30 seconds of data; MS5611 air pressure sensor (resolution 0.01mbar), calculates altitude using the International Standard Atmosphere Model (formula .

[0021] Spectral detection module: AS7341 spectral sensor (11 channels, wavelength range 365-1000nm, integration time settable to 100ms), paired with a diffuse reflectance sampling probe (5cm from the sample), and a built-in 450nm blue light source to stimulate bioluminescence; Each sampling is repeated 5 times and the average value is taken to reduce noise.

[0022] Positioning and marking unit: Beidou III module (BDS-3, single-point positioning accuracy 2.5m), outputs latitude and longitude, altitude, and UTC time (NMEA-0183 protocol); UHF RFID reader / writer (860-960MHz) to read animal collar tags (which store individual number, age, etc.).

[0023] Power supply: Solar panel (5W) + lithium battery (12V / 10Ah), supports 7 consecutive days of operation in rainy weather; Mobile sampling terminal: Inspection personnel are equipped with handheld devices (based on iPad Pro modification) with: 7-inch capacitive screen (supports glove operation) displays temperature and humidity curves and spectral intensity waveforms in real time; Bluetooth adapter (connected to sampling device, transmission distance 100m); Built-in thermal printer to print sample labels containing QR codes and RFID chips (coding format: protected area code + year + serial number, such as "NH20250001") Edge computing node configuration Fixed deployment node: The protected area management station sets up an edge computing server (NVIDIA Jetson AGX Orin) with the following configuration: 64-core ARM CPU, 200TOPS GPU, 16GB memory, 256GB storage; Implementation functions: The temperature and humidity series were subjected to sliding median filtering (window size = 10) to eliminate abnormal fluctuations; Spectral data successive projection algorithm (SPA) feature selection, screening 5 key wavelengths from 11 channels; The geographic coordinates are Fourier encoded to generate a 64-dimensional spatial semantic vector.

[0024] Mobile computing unit: The handheld device is equipped with a lightweight modal fusion engine that implements modal gating and logistic regression inference based on TensorFlow Lite, with single-sample processing time less than 10ms.

[0025] Scenario 2: Batch analysis of laboratory samples Deployment environment: A biomedical laboratory processes an average of 200 monkey tissue samples per day and requires automated determination of sample activity (the target category is "normal activity," accounting for approximately 30%).

[0026] 1. High-throughput acquisition system Sample delivery module: The conveyor belt speed is 50mm / s, and the photoelectric sensor detects the sample in place; The robotic arm (accuracy ±0.1mm) automatically grabs the sample tube and places it in the spectrum detection station (with integrating sphere accessory to ensure spectral repeatability <1%).

[0027] Multimodal sensor arrays: Temperature and humidity sensor: integrated into the -80°C sample refrigerator, collecting environmental data every 10 minutes; AVASpec-NIR256-2.5 near-infrared spectrometer (900-1700 nm, 2.5 nm resolution) was used to analyze the water content of the samples (positively correlated with activity).

[0028] 2. Cloud-based analytics platform Server cluster: 8 AWS c5.4xlarge instances (16-core CPU, 64GB memory), equipped with the PyTorch distributed training framework; Storage architecture: Time series data: The InfluxDB cluster stores historical temperature and humidity data (retained for 3 years); Sample archive: PostgreSQL database (including PostGIS extension), supporting "collection location + climate type" combined query; Graph analysis: Neo4j builds a "sample-environment-device" association network and supports Cypher statements to query dependencies between modalities.

[0029] Implementation of the entire data processing process 1. Multimodal Data Acquisition Protocol Time series data: Generates a continuous sequence of temperature and humidity with a 30-second window, and simultaneously records air pressure values ​​for altitude calculation and environmental characteristic modeling.

[0030] Spectral data: Obtain light intensity values ​​in the 365-1000nm band. The near-infrared band is used for moisture analysis, and the ultraviolet-visible light band is used for biomolecule feature recognition.

[0031] Discrete data: Labels such as collection location and climate type use predefined enumeration values ​​(such as "ear" and "subtropical_monsoon"); The BeiDou module outputs latitude and longitude (decimal), altitude (meters) and UTC timestamp.

[0032] 2. Implementation of feature-level fusion algorithm (1) Time series modal processing Noise filtering: A sliding median filter with a window size of 10 is applied to the temperature and humidity series to remove outliers caused by instantaneous interference.

[0033] Statistical feature extraction: Calculate the mean, standard deviation, and range to reflect the data fluctuation range; Linear regression is used to fit the pressure change trend and the slope is extracted as the dynamic feature; Divide the sub-window to calculate the kurtosis and skewness to capture the data distribution shape.

[0034] (2) Spectral modal processing Outlier replacement: Spectral intensity values ​​exceeding ±3 standard deviations of the mean are replaced with the channel mean.

[0035] Characteristic wavelength selection: The five characteristic wavelengths (such as 700nm and 850nm) that are most relevant to the sample properties are screened through the Successive Projection Algorithm (SPA).

[0036] Part-related weighting: Based on the collection site (such as "ear"), the characteristic wavelengths in the near-infrared band are given a higher weight (increased by 20%) to enhance the part-specific characteristics.

[0037] (3) Discrete modal processing Geographic coordinate encoding: The longitude and latitude are normalized to the interval [0, 1]; A two-dimensional Fourier transform is used to generate a 64-dimensional vector to capture the periodicity of spatial positions (such as the correlation between latitude and temperature).

[0038] Label semantic enhancement: Labels such as climate type and vegetation type are pre-defined with weights based on domain knowledge (e.g., “tropical rainforest” has a weight of 0.8), and 16-dimensional vectors are generated through weighted one-hot encoding.

[0039] 3. Modal Fusion and Model Input Feature splicing: Splice in the order of "time series statistical features (36 dimensions) → spectral screening features (5 dimensions) → geocoding vectors (64 dimensions) → label weighted vectors (16 dimensions)" to form a 121-dimensional fusion feature.

[0040] Modal gating mechanism: Dynamically adjusts the weight of each modality through training. For example, in high-altitude scenarios, the weight of the temporal modality (temperature, humidity, and air pressure) is increased to 50%, the spectral modality is reduced to 30%, and the geographic modality remains at 20%.

[0041] 4. Logistic regression model training Parameter optimization: Maximize the log-likelihood function and use a dynamic learning rate gradient descent (initial 0.01, decaying by 0.95 every 50 rounds); L2 regularization (C=0.5) suppresses overfitting.

[0042] Threshold Adjustment: When the sample is unbalanced (the target class is less than 20%), the threshold is lowered to 0.4-0.5; The threshold of 0.9 is manually set for high-risk scenarios to ensure high specificity Experimental verification and performance data 1. Wildlife monitoring scenario Dataset: 1,200 samples were collected from January to March 2025, including 180 target categories (accounting for 15%) and 121 fusion features.

[0043] Comparison results

[0044] Practical application: 3 monkeypox virus carriers were identified 3 days in advance. The edge model is automatically updated daily, and the accuracy fluctuation during the rainy season is less than 2%. Laboratory testing scenario Optimization points: The weight of the spectral moisture feature is increased to 0.5, and the lagged correlation coefficient between the acquisition time and the refrigerator temperature and humidity is introduced.

[0045] Performance indicators: An average of 220 cases are processed daily, with a misdiagnosis rate of <5%; Training 200 epochs on 8 GPUs takes 45 minutes, which is 6.8 times faster than training on a single GPU.

[0046] Association discovery: Through the graph database, it was found that samples with "altitude > 800m and soil potassium content > 150ppm" had a 42% higher probability of reduced mitochondrial activity.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A biological sample detection method based on AIOT, characterized in that: The following steps are involved: Step 1: Multi-source data collection: Collect sensor data, equipment information, and geographic information of the sample, including: Temperature time series obtained by temperature and humidity sensors and humidity time series ; pH sensor obtains the pH scalar value ; Spectral vector obtained by the spectral sensor ; Collection timestamp , Collection site classification label and collection method classification labels ; Geographic coordinates , altitude , climate type , soil composition vector and vegetation type labels ; Step 2: Data preprocessing: Clean, standardize and encode the data in step (1), including: The temperature and humidity time series were denoised using a sliding window median filter, and the spectral data were replaced with outliers exceeding the mean ± 3 times the standard deviation; Perform z-score normalization on numerical features , the spectrum vector is normalized to interval; One-hot encode the classification labels; Step 3: Build a logistic regression model: based on the preprocessed feature vector , build a logistic regression model Optimize parameters by maximum likelihood estimation , and solve it using gradient descent method or quasi-Newton method; Step 4: Sample attribute judgment: Based on the probability value output by the model , set the threshold ,when When , the sample is judged to belong to the target category, otherwise it belongs to other categories; Step 5: Accuracy feedback: Calculate the accuracy of the training set and test set to provide staff with a reference for model performance.

2. The AIOT-based biological sample detection method and system according to claim 1, characterized in that: In step 1: the acquisition frequency of the temperature time series and humidity time series is once per minute; the spectral vector Dimension Corresponding wavelength range spectral intensity; The collection site classification label Includes at least one of "arms", "legs", and "ears".

3. The biological sample detection method based on AIOT according to claim 1, characterized in that: In step 2: the window size of the sliding window median filter 5~10 data points; The spectral data normalization formula is: .

4. The biological sample detection method based on AIOT according to claim 1, characterized in that: In step 3: the parameters The optimization process includes: constructing the log-likelihood function ,in ;- Iteratively update parameters by gradient descent with dynamic adjustment of learning rate , until the likelihood function converges.

5. The biological sample detection method and system based on AIOT according to claim 1, characterized in that: In step 5: the accuracy rate is calculated as follows: in is the total number of samples, is the predicted category, For the true category; when the accuracy of the test set is lower than the preset threshold (such as 80%), the model parameter re-optimization process is triggered.

6. A biological sample detection system based on AIOT, characterized in that: Includes the following modules: Multi-source data acquisition module: includes temperature and humidity sensors, pH sensors, spectrum sensors, GPS positioning units and acquisition information input interfaces, used to synchronously acquire sensor data, equipment parameters and geographic information of samples; Data preprocessing module: including time series filter, outlier replacement unit and normalization processor, used for denoising, encoding and normalizing the raw data; Logistic regression model building module: equipped with a parameter optimization processor and a model storage unit, used to train and store the coefficient matrix of the logistic regression model; Attribute judgment module: contains probability calculation unit and threshold comparator, which is used to determine the sample category according to the model output probability; Feedback module: includes an accuracy calculator and an alarm trigger unit, which are used to output model performance indicators and trigger the retraining process.

7. The AIOT-based biological sample detection system according to claim 6, characterized in that: In the data preprocessing module: The time series filter is configured as a sliding window median filter with dynamically adjustable window size (5-10 points); The outlier replacement unit has a built-in standard deviation calculator, and the judgment conditions are ; The normalization processor supports both z-score normalization and Min-Max normalization modes.

8. The AIOT-based biological sample detection system according to claim 6, characterized in that: In the attribute judgment module: The threshold comparator supports dynamic threshold setting. The threshold adjustment rules are as follows: in and are the number of positive and negative samples respectively.

Images