A nuclear radiation edge real-time anomaly detection and early warning method based on multi-source data fusion

Abstract

The application discloses a kind of nuclear radiation edge real-time exception detection and early warning method based on multi-source data fusion.Deploy STM32 master control and multiple types of sensors in edge node, collect radiation value, temperature and humidity, geographic position data;Radiation value is preprocessed by Kalman filtering, and multi-dimensional feature vector is constructed;Improved isolation forest algorithm is used for preliminary detection of radiation anomaly, sliding window sampling, feature weighting and adaptive threshold mechanism are introduced;The first evidence body is constructed with the exception score, and the second evidence body is constructed with the environmental deviation degree, and the decision is fused by weighted D-S evidence theory;Real anomaly triggers local sound and light alarm and uploads preferentially through NB-IoT, and cloud platform implements three-level grading early warning push according to the degree of anomaly.The application effectively reduces false alarm rate by more than 80% through multi-source data fusion and edge intelligent calculation, and the detection accuracy reaches 97.8%.

Description

Technical Field

[0001] This invention belongs to the field of nuclear radiation monitoring and Internet of Things technology, specifically involving a method for real-time detection and early warning of nuclear radiation edge anomalies based on multi-source data fusion, which is particularly suitable for real-time radiation safety monitoring in scenarios such as the surrounding environment of nuclear facilities, radioactive material transportation, and nuclear medicine sites. Background Technology

[0002] With the widespread application of nuclear technology in industry, medicine, energy, and other fields, the risk of nuclear radiation leakage has increasingly become a major threat to public safety. From Chernobyl to Fukushima, the tragic lessons of nuclear accidents underscore the necessity of establishing a 24 / 7, high-precision radiation monitoring network. However, traditional nuclear radiation detection equipment primarily relies on single GM counters or scintillator detectors, which can only provide radiation dose rate data and suffer from the following technical limitations:

[0003] 1. Single data source: Only radiation values ​​are collected, without considering the impact of environmental factors such as temperature, humidity, and geographical location on radiation detection. The counting efficiency of GM counter tubes is significantly affected by temperature and humidity. Single radiation data is prone to measurement deviation and cannot determine whether anomalies are caused by environmental factors, resulting in a high false alarm rate.

[0004] 2. Insufficient processing capacity: Existing devices mostly adopt a centralized processing architecture of "collection-upload-analysis," where all data needs to be uploaded to the cloud for analysis. This architecture suffers from response latency (typically tens of seconds to minutes), cannot meet the needs of real-time early warning, and its function fails when the network is unstable.

[0005] 3. Outdated anomaly detection algorithms: Most methods use a fixed threshold comparison method, where an alarm is triggered when the radiation value exceeds a preset threshold. This method cannot adapt to slow fluctuations in background radiation values ​​(such as weather changes or cosmic ray fluctuations) and sudden environmental changes (such as drastic changes in temperature and humidity), leading to frequent missed or false alarms.

[0006] 4. Lack of multi-source fusion decision-making mechanism: Existing systems do not effectively integrate multi-dimensional information such as radiation values, temperature and humidity, and geographical location, making it impossible to comprehensively determine the authenticity of anomalies. For example, an increase in radiation values ​​may originate from a radioactive source leak, or it may simply be due to changes in the performance of the counting tube caused by increased humidity; a single threshold method is insufficient to distinguish between these factors.

[0007] 5. Inadequate early warning mechanism: Most devices only have local audible and visual alarms and lack the ability to provide tiered early warnings and remote push notifications, which cannot meet the differentiated response needs under different levels of urgency.

[0008] In recent years, the development of edge computing and multi-source data fusion technologies has provided new ideas for intelligent nuclear radiation monitoring. For example, published patent CN202510110244.3 proposes a nuclear radiation environment monitoring system based on unmanned aerial vehicles (UAVs), which divides the area to be monitored into multiple sub-regions and sets up multiple radiation dose monitoring points within each sub-region. However, existing technologies have yet to present a complete, edge-oriented method for real-time anomaly detection and graded early warning of nuclear radiation that integrates improved isolated forest and evidence theory. Summary of the Invention

[0009] The purpose of this invention is to overcome the above-mentioned defects in the existing technology and provide a method for real-time detection and early warning of nuclear radiation edge anomalies based on multi-source data fusion. By integrating multi-source sensors at the edge and combining the improved isolated forest algorithm and weighted evidence theory, real-time and accurate detection and graded early warning of radiation anomalies can be achieved.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A method for real-time detection and early warning of nuclear radiation edge anomalies based on multi-source data fusion, characterized by the following steps:

[0012] Step 1: Multi-source data acquisition and preprocessing

[0013] An STM32F103C8T6 main control chip is deployed at the edge detection node, and a GM counter tube, a DHT11 temperature and humidity sensor, an ATGM336H Beidou positioning module, and an M5310-ANB-IoT communication module are integrated. Radiation pulse counts, temperature and humidity values, and positioning data in NMEA-0183 format are collected. The radiation pulse counts are smoothed by Kalman filtering, the temperature and humidity are filtered by moving average, and the positioning data is validated and parsed to generate a multi-dimensional feature vector.

[0014] Step 2: Preliminary detection of radiation anomalies based on improved isolated forests

[0015] An unsupervised anomaly detection model based on isolated forests is constructed, and a sliding window sampling strategy is used to extract historical data to train the model. A feature weighting mechanism is introduced during feature splitting to increase the weight of radiation value and its rate of change. The current feature vector is input into the model to calculate the anomaly score and compare it with a dynamic adaptive threshold. If the score exceeds the threshold, it is marked as a suspected anomaly and proceeds to the next step of fusion decision.

[0016] Step 3: Multi-source fusion decision based on weighted evidence theory

[0017] Building a recognition framework ,in Indicates an anomaly. The condition is considered normal. The basic probability allocation of the first evidence body is constructed based on the abnormality score in step 2. The basic probability allocation of the second evidence body is constructed based on the deviation of the current temperature and humidity from the historical average. The environmental dynamic weighting factor is introduced to weight and correct the two evidence bodies. The basic probability allocation after fusion is calculated using the DS evidence theory fusion rule. If the abnormal probability after fusion is greater than the preset warning threshold, it is determined to be a real abnormality.

[0018] Step 4: Edge-Cloud Collaborative Tiered Early Warning

[0019] Once a genuine anomaly is identified, the edge node immediately triggers a local audible and visual alarm and prioritizes uploading the abnormal data packet to the OneNet cloud platform via the NB-IoT module. The cloud platform sets three warning levels based on the degree of radiation anomaly and pushes them to the monitoring terminal in a tiered manner via SMS, WeChat, telephone, etc. At the same time, historical data is stored in the cloud, supporting trend analysis and visualization.

[0020] Furthermore, the state equation and observation equation for the Kalman filter in step 1 are as follows:

[0021] (1)

[0022] (2)

[0023] in for The true value of radiation at any given moment. for The measured value at time, For process noise, obey distributed, To measure noise, obey Distribution. The recursive formula for Kalman filtering is:

[0024] Prediction phase:

[0025] (3)

[0026] (4)

[0027] Update phase:

[0028] (5)

[0029] (6)

[0030] (7)

[0031] in For Kalman gain, Let be the error covariance.

[0032] Furthermore, in step 1, the temperature and humidity are averaged using a sliding window with a length of 5, and the calculation formula is as follows:

[0033] (8)

[0034] (9)

[0035] Furthermore, in step 1, the multidimensional feature vector ,in This is the filtered radiation value. For temperature, For humidity, Longitude Latitude The rate of change of radiation value within the time window. For the rate of temperature change, This represents the rate of change in humidity. The formula for calculating the rate of change is:

[0036] (10)

[0037] in This represents the sampling time interval.

[0038] Furthermore, in step 2, the feature weight allocation for the improved isolated forest is: radiance value Weight rate of change of radiation Weight ,temperature Weight ,humidity Weight Temperature change rate Weight Humidity change rate Weight The probability that a feature is selected as a splitting feature is:

[0039] (11)

[0040] in For feature dimensions.

[0041] Furthermore, the formula for calculating the anomaly score in step 2 is as follows:

[0042] (12)

[0043] in For the sample The average path length of all trees in an isolated forest. For a given subsample size The path length normalization factor is calculated using the following formula:

[0044] (13)

[0045] in For harmonic numbers, , The Euler constant is approximately 0.5772156649.

[0046] Furthermore, the method for determining the dynamic adaptive threshold in step 2 is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The 95th percentile of abnormal scores in a normal sample is used as the current threshold, and this threshold is set every preset period. Recalculation and Update:

[0047] (14)

[0048] Furthermore, the basic probability allocation of the first piece of evidence in step 3 is as follows:

[0049] (15)

[0050] in These are the normalized outlier scores. The normalization method is as follows:

[0051] (16)

[0052] Furthermore, the environmental deviation in step 3 The calculation formula is:

[0053] (17)

[0054] in , The historical mean and standard deviation of temperature. , This represents the historical mean and standard deviation of humidity. Historical statistics are updated using an exponentially weighted moving average.

[0055] (18)

[0056] (19)

[0057] in The forgetting factor is set to 0.95.

[0058] Furthermore, the basic probability allocation of the second piece of evidence in step 3 is as follows:

[0059] (20)

[0060] Furthermore, in step 3, the environmental dynamic weighting factor is determined based on the rate of change of temperature and humidity. An environmental severity index is defined as follows:

[0061] (twenty one)

[0062] in , The preset threshold is used. Therefore, the weighting factor is:

[0063] (twenty two)

[0064] The weighted basic probability allocation is as follows:

[0065] (twenty three)

[0066] (twenty four)

[0067] Furthermore, the DS evidence theory fusion rule in step 3 is as follows:

[0068] (25)

[0069] (26)

[0070] (27)

[0071] in Conflict factors:

[0072] (28)

[0073] Furthermore, the determination rule in step 3 is: if and If it is, then it is determined to be a true anomaly, where The warning threshold is set to 0.6.

[0074] Furthermore, the criteria for classifying the three warning levels in step 4 are as follows:

[0075] Level 3 Warning (Attention Level): Fusion Anomaly Probability and The radiation level did not exceed twice the safety limit, and was only recorded by the platform and notified via SMS.

[0076] Level 2 Warning (Alert Level): Fusion Anomaly Probability and If the radiation level exceeds the safety limit by 2 times but not by 5 times, a notification will be sent to the on-duty personnel via WeChat.

[0077] Level 1 Warning (Danger Level): Fusion Anomaly Probability If the radiation level exceeds the safety limit by more than 10 times, the system will automatically make a phone call and activate the emergency plan.

[0078] Furthermore, in step 4, the edge node executes a local alarm, including driving a buzzer and an LED indicator. The buzzer control logic is as follows:

[0079] (29)

[0080] Furthermore, the improved sliding window sampling strategy for isolated forests in step 2 is as follows: For each training iteration, the nearest... One sample, This enables online model updates. The model update cycle is... Hour.

[0081] Beneficial effects

[0082] Compared with the prior art, the present invention has the following beneficial effects:

[0083] 1. Improved Detection Accuracy: By fusing multi-source data and incorporating environmental information such as temperature, humidity, and geographical location, and combining this with DS evidence theory for comprehensive decision-making, the false alarm rate caused by environmental interference is significantly reduced. Experiments show that in scenarios with drastic temperature and humidity changes, the traditional threshold method results in 5 false alarms per 24 hours, while this method reduces the false alarm rate to 1 per 24 hours, a reduction of over 80%.

[0084] 2. Real-time Response: Data preprocessing and anomaly detection are completed at the edge, with an average time of less than 200ms from data acquisition to fusion decision-making, meeting real-time early warning requirements. The NB-IoT module adopts PSM low-power mode, with a current of only 5μA during non-transmission periods, an average power consumption of less than 100mW, and a 10000mAh battery that can operate continuously for more than 15 days.

[0085] 3. Strong algorithm adaptability: The improved isolated forest algorithm introduces dynamic sampling and feature weighting mechanisms, which can adapt to the slow drift of the radiation background value. The dynamic threshold update strategy of Equation (14) is adopted to ensure the robustness of the model under different seasons and weather conditions. Experimental data show that the accuracy of the model remains above 96% in all four seasons.

[0086] 4. Fusion Decision Science: The weighted evidence theory fusion mechanism using equations (17)-(28) is adopted, which considers both the statistical characteristics of radiation data and the correlation of environmental factors. When the environment changes drastically, the weight of environmental evidence is dynamically reduced through equation (22) to avoid false alarms caused by sudden environmental changes. Theoretical derivation proves that this fusion mechanism satisfies the commutative and associative laws and has a good mathematical foundation.

[0087] 5. Improved Early Warning Mechanism: A three-tiered early warning system has been established, achieving local-cloud collaboration. The average response time for Level 1 early warning is <3 seconds, Level 2 <5 seconds, and Level 3 <10 seconds. The cloud platform supports multi-node access and big data analysis, facilitating the construction of a regional nuclear radiation monitoring network.

[0088] 6. The system has good scalability: the edge nodes support remote parameter configuration and algorithm model updates, and the feature weights of equation (11) and the threshold parameters of equation (22) can be dynamically adjusted through the NB-IoT downlink channel. The cloud adopts the OneNet platform, which supports multiple protocols such as MQTT and LwM2M, making it easy to integrate with other IoT systems. Attached Figure Description

[0089] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0090] Figure 2 This is a block diagram of the edge node hardware structure of the present invention.

[0091] Figure 3 This is a flowchart of the multi-source data acquisition and preprocessing process of the present invention.

[0092] Figure 4 The flowchart of the improved isolated forest algorithm of this invention is shown.

[0093] Figure 5 This is a schematic diagram of the isolated tree construction of the present invention.

[0094] Figure 6 This is a flowchart of the weighted evidence theory fusion decision-making process for this invention.

[0095] Figure 7 This is a schematic diagram of the graded early warning mechanism of the present invention.

[0096] Figure 8 This is a physical structural diagram of the edge node of the present invention.

[0097] Figure 9 This is a comparison chart of radiation values ​​before and after Kalman filtering in this invention.

[0098] Figure 10 This is a comparison chart of ROC curves for anomaly detection using different algorithms of this invention.

[0099] Figure 11 This is a bar chart comparing the false alarm rates under temperature and humidity interference according to the present invention. Detailed Implementation

[0100] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. This embodiment provides detailed implementation methods and specific operation processes based on the technical solution of the present invention, but the scope of protection of the present invention is not limited to the following embodiment.

[0101] Example 1: Edge Node Hardware Deployment

[0102] like Figure 2 As shown, the edge detection node uses the STM32F103C8T6 as the main control chip. This chip is based on the Cortex-M3 core, has a main frequency of 72MHz, and has 37 GPIOs, 3 serial ports, and multiple timers to meet the requirements of multiple sensor access.

[0103] The GM counter uses a J305β / γ Geiger-Miller tube, operating at 400V with an output pulse width of approximately 50μs. The counter's output signal is amplified and shaped by a transistor before being input to the main controller's TIM3_CH1 pin (PA6) for pulse counting via input capture mode. The boost circuit uses an MC34063 chip to boost the 3.7V lithium battery voltage to 400V.

[0104] The DHT11 temperature and humidity sensor's data pin is connected to PA9. It uses a single-bus protocol to read temperature and humidity data. The measurement range is: temperature 0-50℃±2℃, humidity 20-90%RH±5%RH. A 4.7kΩ pull-up resistor ensures that the bus remains high when idle.

[0105] The ATGM336H-5N Beidou positioning module communicates with the main controller via serial port 2 (PA2-TX, PA3-RX) at a baud rate of 9600bps, outputting data in NMEA-0183 format. The module features built-in antenna detection and short-circuit protection, and supports GPS / Beidou dual-mode positioning.

[0106] The M5310-ANB-IoT module communicates with the main controller via serial port 1 (PA9-TX, PA10-RX), supports Band 3 / Band 5 / Band 8 frequency bands, and uses an LCC package. The module is controlled via AT command set, supports multiple protocols such as LwM2M and MQTT, and can be directly connected to the OneNet cloud platform.

[0107] The 0.96-inch OLED screen is connected via SPI interfaces (PB3-SCLK, PB4-MISO, PB5-MOSI, PB6-DC, PB7-CS), with a resolution of 128×64, and is used for local data display.

[0108] The system is powered by a 3.7V 18650 lithium battery, which is stepped down to 3.3V via an RT9193-3.3V LDO to power the main controller and other modules. The power management circuit includes overcharge and over-discharge protection and power detection functions, and monitors the battery voltage through the PA1 ADC pin.

[0109] The edge node circuitry employs a double-layer board design, measuring 80mm × 50mm. Component layout follows a "power-digital-analog" partitioning principle, with the GM counter's high-voltage circuitry isolated by separate copper plating to reduce interference. The physical structure is shown below. Figure 8 As shown, it uses an ABS plastic shell, with a window on the front revealing the OLED screen and indicator lights, and antennas and sensor interfaces on the side.

[0110] Example 2: Multi-source data acquisition and preprocessing

[0111] After the system is powered on, the main controller initializes all peripherals. Timer 1 is configured for a 1ms interrupt for system timing and task scheduling; Timer 3 is configured for input capture mode, triggered on the falling edge, to count pulses on the GM counter. The total number of pulses is calculated every minute and converted to the radiation dose rate according to equation (30). ):

[0112] (30)

[0113] in For 1 minute pulse count, The conversion factor is used for the J305 counting tube. (correspond unit).

[0114] Apply Kalman filtering to the radiation value sequence. Assume the initial estimate is... Initial error covariance Process noise covariance Measure noise covariance Calculate recursively according to formulas (1)-(7). Figure 9 The comparison of radiation values ​​before and after filtering is shown. The original data showed obvious statistical fluctuations, while the filtered curve was smooth and the true trend of change was preserved.

[0115] Temperature and humidity are averaged using a sliding window of length 5, calculated according to equations (8) and (9). When there is one outlier in five consecutive samples, the moving average can effectively suppress impulse interference.

[0116] Location data parsing process as follows Figure 3 As shown. The serial port interrupts to receive data. When the character '$' is detected, reception begins. After recognizing the "GNRMC" frame, it is saved to the buffer. The UTC time, latitude, longitude, and positioning status fields are parsed out. The latitude and longitude are only used for subsequent processing when the positioning status is 'A' (valid positioning). The latitude format is ddmm.mmmm, and the longitude format is dddmm.mmmm. They are converted to decimal format according to formulas (31) and (32):

[0117] (31)

[0118] (32)

[0119] Feature vector construction: Generate a set of feature vectors every 5 seconds. The rate of change is calculated according to formula (10). Seconds. The feature vector is stored in a circular buffer for use by the anomaly detection module.

[0120] Example 3: Improved Anomaly Detection in Isolated Forests

[0121] In the initial stage of system deployment, 24 hours of normal operating condition data were continuously collected to construct the initial training set D0, containing 17280 samples (sampled every 5 seconds). Feature Dimensions (Latitude and longitude coordinates will not be used for anomaly detection at this time).

[0122] Parameters for building an isolated forest: number of trees Subsampling size The maximum depth of each tree Set the feature weights according to formula (11), where each feature weight is... .

[0123] The process of constructing an isolated tree is as follows: Figure 5 As shown: Randomly sampled from the training set From a sample, randomly select a feature. (by probability) (Selection), randomly choose a split value between the minimum and maximum values ​​of this feature. The samples are divided into a left subtree (feature value). ) and right subtree (eigenvalues) Recursively partition until all samples are indivisible or the maximum depth is reached.

[0124] During online detection, feature vectors are extracted for each new set of data collected. Traverse all isolated trees and calculate the average path length. The path length of each tree is calculated as follows: starting from the root node, traverse downwards to the leaf node based on the feature value and the split point, and the number of edges traversed is the path length. Anomaly scores are calculated using equations (12) and (13). .

[0125] Dynamic threshold The initial value is set to 0.6. Every hour, the 95th percentile is recalculated using the anomaly scores of the most recent 1000 normal samples according to equation (14), and updated. .like If it is marked as a suspected anomaly, it will be transferred to the fusion decision process.

[0126] Figure 4The flowchart of the improved Isolation Forest algorithm is shown, including three parallel tasks: model initialization, online detection, and model update.

[0127] Example 4: Weighted Evidence Theory Fusion Decision Making

[0128] Building a recognition framework For suspected anomalous samples, the anomalous scores are normalized according to equation (16). The allocation of the first body of evidence, BPA, is calculated according to formula (15).

[0129] Calculate historical temperature and humidity statistics. Take the average temperature over the past 24 hours. Standard deviation Average humidity Standard deviation Calculate the environmental deviation using formula (17). For example, the current temperature ,humidity ,calculate:

[0130]

[0131] Calculate the BPA allocation for the second body of evidence according to formula (20):

[0132]

[0133] Calculate the severity of the environmental conditions. Let... , Current rate of temperature change Humidity change rate Calculate according to formula (21):

[0134]

[0135] Calculate the weighting factor according to formula (22):

[0136]

[0137] The weighted BPA is calculated according to formulas (23) and (24):

[0138]

[0139]

[0140] Calculate the conflict factor according to formula (28):

[0141]

[0142] Calculate the fused BPA according to formulas (25)-(27):

[0143]

[0144] After fusion, the probability of anomaly is 0.643, the probability of normality is 0.092, and the probability of uncertainty is 0.265. A warning threshold is then set. ,because and It was determined to be a genuine anomaly.

[0145] Figure 6 It demonstrates a flowchart of the weighted evidence theory fusion decision-making process, clearly presenting the complete process from evidence construction to fusion judgment.

[0146] Example 5: Implementation of Tiered Early Warning

[0147] After determining that it is a genuine anomaly, the edge node performs the following operations:

[0148] 1. Local Records: Write the abnormal moment data (timestamp, radiation value, temperature and humidity, latitude and longitude, abnormal score, fusion probability) into the Flash memory, starting from address 0x08040000. Each record is 32 bytes, and the latest 1000 records are stored in a loop.

[0149] 2. Audible and visual alarm: Drive the PA7 pin to output a high level according to formula (29) to trigger the active buzzer to sound (frequency 2kHz, sounding for 1 second / 1 second interval); drive the PB5 pin to output a low level to light up the red LED indicator (flashing frequency 1Hz).

[0150] 3. Data reporting: Send AT commands to the M5310-A ​​module via serial port 1, set QoS=1, and ensure successful reception by the platform.

[0151] After receiving abnormal data, the OneNet cloud platform processes it in a tiered manner according to preset rules:

[0152] Example of a Level 3 warning: Radiation value (Safety Limits) The fusion probability is 0.65, triggering a Level 3 warning. The platform sent an SMS to the monitoring mobile phone: "[Nuclear Radiation Warning] Node ID: 001, Time: 2026-02-14 10:23:45, Radiation Value..." The fusion probability is 0.65; please pay attention.

[0153] Example of a Level II warning: Radiation value The fusion probability is 0.85, triggering a level-two warning. The platform pushes a notification via WeChat: "[Nuclear Radiation Warning] Node ID: 001, Location: 29.63°N 106.50°E, Radiation Value..." (Exceeded limit by 2.4 times), fusion probability 0.85, please check immediately.

[0154] Example of Level 1 Warning: Radiation Value (Exceeding the limit by 17 times), fusion probability 0.98, triggering a Level 1 warning. The platform automatically dialed a preset number via Alibaba Cloud's communication API and played a voice message: "Emergency nuclear radiation alarm! Radiation levels at node 001 are severely exceeded. Please activate the emergency plan immediately!"

[0155] Figure 7 The diagram illustrates the tiered early warning mechanism, including edge processing flow and cloud-based tiered push logic.

[0156] Example 6: Online Model Update

[0157] Every hour, the system executes a model update procedure:

[0158] 1. Read the most recent 1000 normal data entries (marked as normal samples) from Flash, extract feature vectors to form a new training set. .

[0159] 2. Update the isolated forest using incremental learning: retain 80% (80 trees) of the original 100 trees, generate 20 new trees with new data, and replace the 20 worst-performing old trees. The tree performance evaluation metric is the average path length variance.

[0160] 3. The parameters for generating the new tree are the same as those for the initial training: subsampling size. The feature weights are configured according to equation (11).

[0161] 4. After the update, the model is stored in external Flash memory, starting at address 0x08080000, occupying approximately 50KB of space.

[0162] 5. Dynamic threshold synchronization update: (The last part is incomplete and likely refers to a different function or feature.) The 95th percentile of the abnormal scores of all samples was used as the new threshold. Write to the Flash parameter area.

[0163] The model update process does not affect online detection and adopts a double buffer mechanism: the current model is used for detection, a new model is generated for updating, and the pointer is switched atomically after the update is completed.

[0164] Example 7: Performance Testing and Result Analysis

[0165] System performance testing was conducted in a laboratory environment.

[0166] Test 1: Accuracy Test

[0167] The test set contains 1000 normal data sets and 50 abnormal data sets (using a Cs-137 radioactive source, activity). Comparing the three methods:

[0168] Method A: Fixed threshold method (threshold) )

[0169] Method B: Standard Isolation Forest (Unweighted, Fixed Threshold)

[0170] Method C: The method of this invention (improved isolated forest + fusion decision)

[0171] The results are shown in Table 1:

[0172] Table 1. Comparison of the three methods

[0173] method accuracy Recall rate F1 score AUC value Fixed threshold method 92.1% 84.0% 0.878 0.921 Standard isolated forest 95.3% 92.0% 0.936 0.973 Method of the present invention 97.8% 96.0% 0.969 0.991

[0174] Figure 10 The ROC curves of three methods are shown, and the method of this invention is also presented. This method is significantly superior to the comparison method.

[0175] Test 2: Anti-interference capability test

[0176] Using a heater to simulate temperature disturbances ( (Heating up in 10 minutes), humidifier simulates humidity interference ( Number of false alarms within 24 hours:

[0177] In an interference-free environment: Method A generated 2 false alarms, Method B generated 1 false alarm, and the method of this invention generated 0 false alarms.

[0178] Temperature interference environment: Method A generated 8 false alarms, Method B generated 5 false alarms, and the method of this invention generated 1 false alarm.

[0179] Humidity interference environment: Method A had 6 false alarms, Method B had 4 false alarms, and the method of this invention had 0 false alarms.

[0180] Simultaneous interference from temperature and humidity: Method A generated 15 false alarms, Method B generated 9 false alarms, and the method of this invention generated 2 false alarms.

[0181] Figure 11 The bar chart comparing false alarm rates visually demonstrates the superiority of the method of this invention under environmental interference.

[0182] Test 3: Real-time performance test

[0183] From radiation pulse input to completion of fusion decision output, the time consumed at each stage was measured using a logic analyzer:

[0184] Data acquisition and preprocessing: 35ms (including Kalman filtering)

[0185] Anomaly detection in isolated forests: 128ms (traversing 100 trees)

[0186] Evidence theory fusion decision: 0.5ms

[0187] Total time: 163.5ms

[0188] It meets the requirements for real-time detection (<200ms). The reporting time of NB-IoT depends on the network conditions, averaging 3-5 seconds.

[0189] Test 4: Power Consumption Test

[0190] Measure power consumption in each operating mode using a power analyzer:

[0191] Working mode (data acquisition + detection):

[0192] NB-IoT transmission mode: 220mA (peak, lasting 2 seconds)

[0193] PSM Low Power Mode: 5μA

[0194] Based on a data transmission rate of once per minute, the average power consumption is approximately 15mW, and the 10000mAh battery can operate continuously for about 30 days.

[0195] Example 8: Application Scenario

[0196] Scenario 1: Environmental monitoring around nuclear facilities

[0197] Fifty edge nodes were deployed around a nuclear power plant to form a monitoring network. The nodes were spaced 500 meters apart, covering an area with a radius of 5 kilometers. On a certain day in December 2025, node 23 detected radiation levels ranging from [previous value]. Rise to At the same time, the temperature rose by 3°C and the humidity increased. The traditional threshold method did not trigger an alarm (the threshold was not exceeded). However, the method of this invention detected abnormal scores. Integrated decision-making The incident triggered a Level 3 alert. An investigation revealed that the dust was caused by nearby construction work and did not result in any actual leakage, but the alert highlighted a potential risk.

[0198] Scenario 2: Monitoring of Radioactive Material Transportation

[0199] Edge nodes were installed on vehicles transporting radioactive materials. During transport, the vehicles passed through high-altitude areas (3000m), where enhanced cosmic rays caused radiation levels to rise. (Plains area) The method of this invention learns from historical data and dynamically adjusts the threshold accordingly, thus avoiding false alarms. Simultaneously, the BeiDou positioning module uploads the trajectory in real time, allowing for viewing of the entire transportation route's radiation data in the cloud.

[0200] Scenario 3: Real-time early warning for nuclear medicine sites

[0201] The system of this invention has been deployed in the nuclear medicine department of a hospital. After a patient receives an injection of radiopharmaceutical, the radiation level in the waiting area briefly rises to [a certain value]. The method of this invention detected abnormal scores. Fusion probability This triggers a Level 2 alert, notifying staff to guide patients to a dedicated waiting area to prevent other people from being exposed for extended periods.

[0202] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time detection and early warning of nuclear radiation edge anomalies based on multi-source data fusion, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition and preprocessing. Deploy the main control chip and multiple types of sensors at the edge detection node to collect radiation values, temperature and humidity, and geographic location data. Perform Kalman filtering smoothing on the radiation pulse counts, moving average filtering on the temperature and humidity, and validity verification and parsing on the positioning data to generate multi-dimensional feature vectors. ,in This is the filtered radiation value. For temperature, For humidity, Longitude Latitude The rate of change of radiation value For the rate of temperature change, The rate of change of humidity; Step 2: Based on the improved isolated forest method for preliminary detection of radiation anomalies, an unsupervised anomaly detection model for isolated forests is constructed. A sliding window sampling strategy is used to extract historical data to train the model. A feature weighting mechanism is introduced during feature splitting, and the current feature vector is input into the model to calculate the anomaly score. With dynamic adaptive threshold In comparison, if It is then marked as a suspected anomaly; Step 3: Constructing an identification framework based on multi-source fusion decision-making using weighted evidence theory. abnormal scores Constructing the first body of evidence Environmental deviation Constructing a second body of evidence A dynamic weighting factor is introduced for weighted correction, and the fusion probability is calculated using the DS evidence theory fusion rule. ,like and Then it is determined to be a genuine anomaly; Step 4: Edge-cloud collaborative hierarchical early warning. Edge nodes execute local audible and visual alarms and upload them via NB-IoT with priority. The cloud platform implements three-level hierarchical early warning push based on the degree of anomaly.

2. The method according to claim 1, characterized in that, The recursive formula for the Kalman filter in step 1 is as follows:

3. The method according to claim 1, characterized in that, The probability that the feature is selected as a splitting feature in step 2 is: The feature weights are assigned as follows: radiation value Weight rate of change of radiation Weight The total weight of the remaining features is 0.

3.

4. The method according to claim 1, characterized in that, The dynamic adaptive threshold in step 2 Take history The 95th percentile of the abnormal scores of normal samples is recalculated and updated every preset period.

5. The method according to claim 1, characterized in that, Environmental deviation in step 3 The calculation formula is: in , The historical mean and standard deviation of temperature. , The values ​​represent the historical mean and standard deviation of humidity.

6. The method according to claim 1, characterized in that, In step 3, the dynamic weighting factor is based on the environmental severity index. Confirmed, weighting factor is , .

7. The method according to claim 1, characterized in that, The DS evidence theory fusion rule in step 3 is as follows: in .

8. The method according to claim 1, characterized in that, The criteria for classifying the three levels of early warning in step 4 are as follows: Level 3 Warning: And if the radiation level does not exceed twice the safety limit, a text message notification will be sent. Level 2 warning: The radiation level exceeded the safety limit by 2 times but not by 5 times, and this was pushed via WeChat. Level 1 Warning: If the radiation level exceeds the safety limit by more than 10 times, a phone call will be made automatically.

Images