Intelligent nuclear radiation dose real-time monitoring system based on deep learning

Abstract

The invention discloses an intelligent nuclear radiation dose real-time monitoring system based on deep learning, and the system comprises the steps: S1, a multi-mode sensing module which is used for outputting radiation intensity data with timestamps, environment image data and geographic space position information; s2, a preprocessing module used for outputting a structured feature sequence; s3, a fusion modeling module which is used for generating a unified fusion feature vector; s4, an end-edge reasoning module which is used for processing the fusion feature vector in stages and generating a nuclear radiation dose prediction result; s5, the task scheduling module is used for dynamically determining whether the reasoning task is executed on the end side or the edge node; s6, an anomaly detection module which is used for determining whether there is an abnormal change of the nuclear radiation dose; and S7, the system control module is used for coordinating the modules. The method has the advantages of being high in environmental adaptability, fast in reasoning response and high in anomaly recognition precision.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and particularly to an intelligent real-time nuclear radiation dose monitoring system based on deep learning. Background Art

[0002] With the continuous expansion of application scenarios such as nuclear energy technology, radiation medicine, nuclear facility decommissioning, and nuclear accident emergency response, higher real-time and precision requirements have been put forward for the dynamic perception and intelligent management of radiation dose in complex environments. In nuclear radiation scenarios, how to achieve rapid identification of sudden dose anomalies, comprehensive modeling of multi-modal environmental changes, and efficient scheduling of the reasoning process has become a key problem that urgently needs to be solved in radiation monitoring systems.

[0003] In the prior art, nuclear radiation dose monitoring mostly relies on data collection by a single sensor and an alarm strategy with a fixed threshold. Typical systems use Geiger counters, ionization chambers, or scintillators as radiation detectors, and cooperate with simple filtering and judgment logics for data processing, which is difficult to reflect the dose evolution process affected by multiple source factors in complex radiation environments. At the same time, traditional systems lack the ability to fuse and perceive environmental image and spatial position information, and cannot accurately identify the regional risk boundary in dynamic scenarios such as heterogeneous radiation distribution, mobile source interference, or sudden leakage. In addition, current methods generally adopt a centralized reasoning architecture, and the system response highly depends on network channels and backend processing capabilities, making it difficult to meet the edge-side collaboration requirements in scenarios with multiple devices, weak connections, or time-delay sensitivity. There are potential hazards such as response lag and control link failure in the reasoning results.

[0004] In terms of dose anomaly detection and early warning, mainstream methods still rely on static threshold judgment, lacking a deep learning-driven adaptive identification mechanism, which is difficult to capture potential pattern changes in dose data, and has insufficient recognition ability for non-linear growth, slow offset, or hidden abnormal events. As a result, the system often fails to issue an effective early warning at the initial stage of high-risk evolution, seriously affecting the nuclear safety prevention and control ability. In addition, some studies have tried to introduce neural networks for dose regression or radiation source identification, but generally stay in the model construction stage, lacking systematic integration with multi-modal input, reasoning path switching, and abnormal handling processes, and it is difficult to achieve a closed-loop intelligent response system from perception, modeling, reasoning to alarm.

[0005] In summary, there is an urgent need for an intelligent real-time nuclear radiation dose monitoring system that integrates multi-modal perception, edge-side collaborative reasoning, and a deep abnormal identification mechanism to solve the core problems such as low environmental perception accuracy, untimely reasoning response, and insufficient dose anomaly judgment ability existing in current nuclear radiation monitoring. Summary of the Invention

[0006] An object of the present invention is to propose an intelligent real-time monitoring system for nuclear radiation dose based on deep learning, a multi-modal perception modeling method that integrates radiation intensity, environmental images, and geographical location information, construct an inference architecture that coordinates a lightweight model on the edge side and a deep model on the edge, and introduce an anomaly detection mechanism based on reconstruction error to realize dose change trend recognition and multi-level early warning control, dynamically adjust the inference path through a task scheduling mechanism and combine the characteristics of time-series data to complete intelligent response, with the advantages of strong environmental adaptability, fast inference response, and high anomaly recognition accuracy.

[0007] An intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to an embodiment of the present invention, the system includes: S1. A multi-modal perception module, configured to collect nuclear radiation intensity data, environmental image data, and geographical space location information, and output radiation intensity data, environmental image data, and geographical space location information with timestamps; S2. A preprocessing module, configured to normalize the nuclear radiation intensity data, extract multi-scale features from the environmental image data, perform trajectory standardization processing on the geographical space location information, and output a structured feature sequence; S3. A fusion modeling module, configured to input the structured feature sequence into a deep fusion network based on an attention mechanism to generate a unified fusion feature vector; S4. An edge-side inference module, including a lightweight neural network model deployed on the edge side and a deep neural network model deployed on the edge node, configured to process the fusion feature vector in stages and generate a nuclear radiation dose prediction result; S5. A task scheduling module, configured to dynamically determine whether the inference task is executed on the lightweight neural network model on the edge side or the deep neural network model on the edge node based on the current network state and device load; S6. An anomaly detection module, configured to execute an anomaly recognition model based on reconstruction error on the nuclear radiation dose prediction result to determine whether there is an abnormal change in the nuclear radiation dose; S7. A system control module, configured to coordinate the communication, data transmission, and state scheduling among the various modules to achieve full-process control.

[0008] Optionally, the S3 specifically includes: S31. A modal feature alignment unit, configured to map the nuclear radiation intensity feature, environmental image feature, and geographical space location information feature in the structured feature sequence to a unified feature dimension respectively to generate an initial modal feature set , where represents the nuclear radiation intensity feature vector after unified encoding, represents the environmental image feature vector after being processed by a convolutional network, Represents the geospatial location information feature vector after spatial encoding; S32. An attention weighted encoding unit, which is used to introduce a cross-modal attention mechanism between the initial modal feature sets and calculate the attention distribution weights between each modality: ; Among them, Represents the modal feature The attention distribution weight for the modal feature , Is a correlation scoring function based on dot product, which is used to measure the feature dependence relationship between modalities; S33. A fusion representation generation unit, which is used to perform weighted combination on all modal features based on the attention distribution weights to generate a unified fusion feature vector , and its calculation method is as follows: ; S34. A regularization processing unit, which is used to perform layer normalization and non-linear activation function operations on the fusion feature vector .

[0009] Optionally, the S4 specifically includes: S41. An edge-side inference unit, which is used to deploy a lightweight neural network model , receive the fusion feature vector output by the fusion modeling module , and generate an intermediate feature representation , the Contains multiple layers of linear mapping and normalization modules, and the calculation relationship is: ; Among them, Is the weight matrix, Is the bias term, Is the non-linear activation function; S42. An intermediate feature buffer unit, which is used to store the intermediate feature representation In the local buffer area, and judge whether to upload according to the inference distribution instruction issued by the system control module; S43. An edge node inference unit, which is used to deploy a deep neural network model , and perform subsequent inference processes after receiving the intermediate feature representation To generate the final nuclear radiation dose prediction value : ; Among them, Contains multiple convolution, residual connection and attention mechanism modules, which are used to extract deep association features; S44. Inference stage switching control unit, configured to control the switching of the inference process between the lightweight neural network model and the deep neural network model according to the real-time allocation result returned by the system task scheduling module, and mark the current inference state; S45. Model parameter synchronization unit, configured to periodically check the model version number and parameter verification code of the edge side and the edge node, and complete the consistency update of the weight data through the system control module; S46. Abnormal interruption handling unit, configured to automatically trigger the edge node to take over the complete inference task process when the inference process or data transmission on the edge side is interrupted, and transmit the prediction status label back to the system control module after recovery.

[0010] Optionally, the S41 specifically includes: S411. Input feature receiving unit, configured to receive the fused feature vector output by the fusion modeling module , where represents the feature dimension, and represents the real number field; S412. Linear mapping processing unit, configured to input the fused feature vector into the lightweight neural network model . The lightweight neural network model is composed of multiple linear mapping layers. Among them, the structure of the th layer mapping is: Among them, represents the weight matrix of the th layer, represents the bias vector, is the output of the previous layer, and the initial input satisfies ; S413. Activation and normalization unit, configured to perform non-linear activation and normalization processing after some linear mapping layers; S414. Output generation unit, configured to output the intermediate feature representation at the end of the model, where represents the total number of layers of the neural network model, and is sent to the intermediate feature buffer unit.

[0011] Optionally, the S43 specifically includes: S431. Edge feature receiving unit, configured to receive the intermediate feature representation from the edge side, where represents the intermediate feature dimension, and send it into the deep neural network model of the edge node; S432. A convolutional feature extraction unit for performing initial spatial dimension expansion on the intermediate feature representation and extracting local context features through multiple two-dimensional convolutional layers. The convolutional calculation structure is as follows: ; where represents a standard convolution operation, is the output of the th convolutional layer; S433. A residual connection unit for introducing a residual skip connection structure between the two-dimensional convolutional layers to establish a direct mapping relationship between the input and output, satisfying , so as to enhance the cross-layer information retention ability, where is the output of the previous convolutional layer; S434. An attention enhancement unit for embedding a channel attention mechanism in several specified layers to perform adaptive weighted reconstruction on the channel dimension. The attention weights are generated by normalizing the input features after full connection mapping; S435. An output generation unit for generating the final nuclear radiation dose prediction value at the end of the deep neural network model and transmitting the nuclear radiation dose prediction value to the anomaly detection module and the system control module.

[0012] Optionally, the S44 specifically includes: S441. An inference state perception unit for receiving the real-time allocation result returned by the task scheduling module. The real-time allocation result is a scheduling status data packet, and the scheduling status data packet contains the network bandwidth , the end-side idle computing power and the task real-time requirement ; S442. An inference switching decision unit for constructing an inference selection function based on the scheduling status data, where this function satisfies: ; where is the pre-designed computing resource and bandwidth threshold, and outputs the model path to be executed currently; S443. An inference path execution unit for controlling the distribution of the fused feature vector to the corresponding lightweight neural network model and deep neural network model according to the result of the inference selection function and synchronously generating the path label of this inference round, where represents end-side inference, represents edge inference; S444, a status marking management unit, is configured to record the path label together with the timestamp, task number, and prediction result identifier to form a structured inference status log, and periodically synchronize it to the system control module.

[0013] Optionally, the S6 specifically includes: S61, a prediction result caching unit, is configured to receive and temporarily store the nuclear radiation dose prediction values generated by the edge node inference unit and store them in chronological order to form a sequence sample set , where is the sample time window length; S62, a reconstruction model inference unit, is configured to input the sequence sample set into a preset anomaly recognition model to generate a reconstructed sequence , where represents the reconstructed sequence result of the th prediction value; S63, an error calculation and determination unit, is configured to calculate a reconstruction error sequence based on the difference between the nuclear radiation dose prediction value and the reconstructed sequence at each moment , where the error value is defined as: ; and compare the error value with a set threshold . If there is , then mark this moment as an anomaly point; S64, an anomaly label generation unit, is configured to convert the moment numbers marked as anomalies into an anomaly event label set and send this set to the system control module.

[0014] The beneficial effects of the present invention are: (1) By constructing a multi-modal perception mechanism that integrates radiation intensity, environmental images, and geospatial information, the present invention realizes the joint modeling of the dose distribution state in a complex radiation environment, breaks through the limitations of traditional reliance on single-sensor data, and improves the environmental perception stability and nuclear radiation dose measurement accuracy of the system under conditions of spatial heterogeneity, target occlusion, and sudden interference.

[0015] (2) The present invention designs an edge-cloud collaborative inference architecture, introduces a hierarchical inference process of a lightweight neural network model and a deep neural network model, and dynamically selects an inference path through a resource scheduling function, effectively reducing the system's dependence on the central node's computing and communication, adapting to the real-time response requirements under different network conditions, and enhancing the flexibility and processing efficiency in the edge deployment scenario.

[0016] (3)The present invention constructs a deep anomaly recognition mechanism based on reconstruction error, reconstructs and analyzes the difference of the nuclear radiation dose prediction result, and generates high-frequency dynamic anomaly labels in combination with a threshold determination strategy. Compared with the fixed threshold alarm mechanism, it can identify the hidden, persistent and slow-rising dose anomaly changes, and enhances the detection sensitivity and early warning forward ability of the system in the initial stage of nuclear accidents.

[0017] (4)The present invention proposes a linkage mechanism between inference state marking and path labels, establishes a unified inference state log management model, realizes the structured tracking management of the whole process from feature input, model call to prediction output, has high auditability and the ability to guarantee the closed-loop operation of the model, and provides basic data support for subsequent safety audit and operation and maintenance diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings: Figure 1 is the overall structural framework diagram of an intelligent nuclear radiation dose real-time monitoring system based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0020] Refer to Figure 1 , an intelligent nuclear radiation dose real-time monitoring system based on deep learning, the system includes: S1. A multi-modal perception module, which is used to collect nuclear radiation intensity data, environmental image data and geospatial location information, and output the radiation intensity data, environmental image data and geospatial location information with time stamps attached; In the S1 multi-modal perception module of the present invention, the collected nuclear radiation intensity data, environmental image data, and geospatial location information constitute the core input content of the system. Among them, the nuclear radiation intensity data is collected by radiation detectors, including dose rate values, pulse count values, sampling timestamps, and optional energy spectrum distribution information, which is used to reflect the radiation level and its change trend in the current area; the environmental image data is obtained through visible light cameras or infrared imaging devices, covering RGB image frames, continuous image sequences, image metadata (such as shooting time, perspective, etc.), and infrared images used in specific tasks, which is used to capture structural features, lighting conditions, and abnormal events in the monitoring environment; the geospatial location information is provided by the GNSS positioning module and / or IMU components, including three-dimensional coordinates (latitude, longitude, and altitude), continuous trajectory data, positioning accuracy indicators, and timestamps, etc., which is used to accurately calibrate the geographical location and movement path of the monitoring device. The above three types of data are all attached with a unified high-precision timestamp, and through a synchronization mechanism, a structured multi-source observation data stream is formed, which serves as the standard input for subsequent preprocessing and model modeling processes, providing a comprehensive and synchronized perception basis for the spatial modeling, dynamic perception, and intelligent early warning of nuclear radiation doses.

[0021] In this embodiment, the S1 multi-modal perception module includes a nuclear radiation intensity perception unit, an environmental image acquisition unit, and a geospatial positioning unit. The nuclear radiation intensity perception unit is used to collect nuclear radiation dose-related data in the monitoring area, and through continuous sampling of the surrounding radiation intensity, obtain time series information representing the change of radiation level. The environmental image acquisition unit is used to obtain image data in the current monitoring environment, and through the equipped imaging device, regularly photograph the monitoring area, which is used to characterize the structure, occlusion, lighting, or other potential influencing factors in the environment. The geospatial positioning unit is used to determine the spatial position where the perception device is located, and through the positioning component, obtain the geographical coordinate information of the device, realizing the one-to-one correspondence between the perception data and the spatial position. After the above various types of perception data are collected, they are subjected to time alignment processing by the synchronization control module, and all data are attached with a unified timestamp to form a structured multi-source fusion data set. This multi-modal perception module supports periodic collection and output, and as the input entrance of the entire system, it provides a synchronous and complete perception information basis for subsequent data preprocessing, feature extraction, and model inference.

[0022] S2. A preprocessing module, which is used to normalize the nuclear radiation intensity data, extract multi-scale features from the environmental image data, perform trajectory standardization processing on the geospatial location information, and output a structured feature sequence; In this embodiment, the S2 preprocessing module includes a radiation data normalization processing unit, an environmental image feature extraction unit, and a spatial trajectory standardization processing unit, which are used to perform format conversion and feature regularization on the data collected by the multi-modal perception module, and generate a structured multi-source input feature sequence. Among them, the radiation data normalization processing unit is used to perform scale conversion on the dose data collected by the nuclear radiation intensity perception unit, unify the data distributions output by different sensors, and reduce the input inconsistency caused by device drift, environmental interference, or data fluctuations. The environmental image feature extraction unit is used to extract multi-level visual structure information from the collected image frames, perform multi-scale processing on the image data through the feature encoding module, and convert the original image into an image feature vector representing texture, edges, and shapes, which is convenient for the subsequent model to discriminate the environmental state. The spatial trajectory standardization processing unit is used to perform path regularization processing on the position data collected by the geospatial positioning unit, and through methods such as trajectory interpolation, coordinate remapping, and smoothing filtering, unify the spatial resolution and sampling density, and ensure that the spatial information is consistent with other modal data in terms of time and sequence structure. After the processing is completed, various types of data are converted into structured feature items in a unified format and organized into a standardized feature sequence in chronological order, serving as the input data set for the fusion modeling module, providing a standardized and clean input feature basis for subsequent multi-modal joint modeling.

[0023] S3. The fusion modeling module is used to input the structured feature sequence into a deep fusion network based on the attention mechanism to generate a unified fusion feature vector; In this embodiment, S3 specifically includes: S31. The modal feature alignment unit is used to map the nuclear radiation intensity feature, environmental image feature, and geospatial position information feature in the structured feature sequence to a unified feature dimension respectively, and generate an initial modal feature set , where represents the nuclear radiation intensity feature vector after unified encoding, represents the environmental image feature vector after being processed by the convolutional network, represents the geospatial position information feature vector after spatial encoding; S32. The attention weighted encoding unit is used to introduce a cross-modal attention mechanism among the initial modal feature sets, and calculate the attention distribution weights among each modality: ; where represents the modal feature 's attention distribution weight to the modal feature , , is a correlation scoring function based on dot product, which is used to measure the feature dependence relationship between modalities; Specifically, for a given modal feature and , the system calculates the feature matching degree between the two through a correlation evaluation function , and obtains its attention allocation weight through Softmax normalization operation , indicating that in the current fusion process, modality pays attention to modality . The calculation method of this weight is as shown in the formula, ensuring that the sum of all allocated weights in the same modality is 1, meeting the requirements of probability normalization.

[0024] S33. Fusion representation generation unit, which is used to perform weighted combination on all modal features based on the attention allocation weight to generate a unified fusion feature vector , and its calculation method is as follows: ; This fusion process introduces an attention weighting mechanism in the modal dimension, and assigns weights determined by the relationship between each modal feature to other modalities . In specific implementation, the system linearly weights the other modal features that each modality pays attention to, and performs summation integration on all modal dimensions to obtain the final fusion feature representation. The above process is as shown in the formula, adopting a double summation structure to ensure that all interaction relationships between modalities are included in the fusion calculation.

[0025] S34. Regularization processing unit, which is used to apply layer normalization and non-linear activation function operations to the fusion feature vector .

[0026] In this embodiment, by constructing a modal feature alignment unit, an attention weighted coding unit, a fusion representation generation unit and a regularization processing unit, the unified feature coding and dynamic weighted fusion of nuclear radiation intensity, environmental image and geospatial information are realized. It can adaptively adjust the information interaction weight according to the correlation between modalities, thereby effectively integrating the key features in multi-source heterogeneous data, improving the representation ability and discriminability of the fusion representation, providing a more robust and environment-adaptive input basis for the subsequent inference module, and enhancing the intelligent recognition ability and processing accuracy of the system in complex nuclear radiation scenarios.

[0027] S4. Edge-side inference module, including a lightweight neural network model deployed on the edge side and a deep neural network model deployed on the edge node, which is used to process the fusion feature vector in stages and generate a nuclear radiation dose prediction result; In this embodiment, S4 specifically includes: S41. Edge-side inference unit, which is used to deploy a lightweight neural network model , receive the fused feature vector output by the fusion modeling module , and generate an intermediate feature representation based on the fused feature vector , the includes a multi-layer linear mapping and normalization module, and the calculation relationship is: ; Among them, is the weight matrix, is the bias term, is the non-linear activation function; This formula performs feature mapping processing on the input vector through a combination of linear transformation and non-linear activation function. Among them, the fused feature vector first undergoes linear mapping of the weight matrix and translation adjustment of the bias term to form a new feature space representation; subsequently, the activation function is introduced to perform non-linear processing on the mapping result to enhance the model's expression ability for complex patterns. The finally obtained represents the output result of the edge-side inference module, which will be passed as input to the edge node inference module or the cache module.

[0028] S42. Intermediate feature buffer unit, used to store the intermediate feature representation in the local cache area, and determine whether to transmit it upward according to the inference distribution instruction issued by the system control module; S43. Edge node inference unit, used to deploy a deep neural network model , and perform subsequent inference processes after receiving the intermediate feature representation to generate the final nuclear radiation dose prediction value : ; Among them, includes multiple convolution, residual connection and attention mechanism modules, used to extract deep correlation features; The formula represents the forward calculation process of the deep model , that is, taking as input, passing through a multi-layer neural structure composed of convolution, residual connection and attention mechanism modules, extracting information with more semantic depth and local context correlation, and finally outputting a dose estimation result in scalar or vector form . This structure can effectively fuse the capabilities of local feature enhancement and global dependence modeling, and improve the discrimination accuracy of the system for complex radiation patterns.

[0029] In specific use, the model It supports continuous operation on edge nodes with computing resources. When the system task scheduling module decides to transfer the inference task to edge processing, the deep prediction process of nuclear radiation dose is completed in the manner defined by this formula. The generated will be synchronously transmitted to the anomaly detection module and the system control module for subsequent risk identification and response decision-making.

[0030] S44. Inference phase switching control unit, used to control the switching of the inference process between the lightweight neural network model and the deep neural network model according to the real-time allocation result returned by the system task scheduling module, and mark the current inference state; S45. Model parameter synchronization unit, used to periodically check the model version number and parameter verification code of the edge node and the end side, and complete the consistency update of the weight data through the system control module; S46. Abnormal interruption handling unit, used to automatically trigger the edge node to take over the complete inference task process when the inference process or data transmission at the end side is interrupted, and transmit the prediction status label back to the system control module after recovery.

[0031] The specific content of S41 includes: S411. Input feature receiving unit, used to receive the fused feature vector output by the fusion modeling module , where represents the feature dimension, represents the real number field; S412. Linear mapping processing unit, used to input the fused feature vector into the lightweight neural network model . The lightweight neural network model consists of multiple linear mapping layers. Among them, the structure of the th layer mapping is: ; Among them, represents the weight matrix of the th layer, represents the bias vector, is the output of the previous layer, and the initial input satisfies ; This formula represents the linear mapping structure of the th layer in the lightweight model. Specifically, the output vector of the previous layer is multiplied by the current layer weight matrix , and the bias term is added to form the output of the current layer. This process is repeated in each layer of the model to extract the linear change pattern and dimension transformation law in the input features.

[0032] In the model initialization stage, the input vector satisfies , that is, directly receive the fused feature representation generated by the fusion modeling module, and then each layer of linear mapping unit performs feature mapping based on this input in sequence.

[0033] S413. An activation and normalization unit, used to perform non-linear activation and normalization processing after some linear mapping layers; S414. An output generation unit, used to output intermediate feature representations at the end of the model , where represents the total number of layers of the neural network model, and send to the intermediate feature buffer unit.

[0034] The S43 specifically includes: S431. An edge feature receiving unit, used to receive intermediate feature representations from the edge side , where represents the intermediate feature dimension, and send it into the deep neural network model of the edge node ; S432. A convolutional feature extraction unit, used to perform initial spatial dimension expansion on the intermediate feature representation , and extract local context features through multiple two-dimensional convolutional layers. The convolutional calculation structure is: ; ; where represents the standard convolution operation, is the output of the th layer of convolution; The formula defines the standard calculation structure of the th layer in the convolutional network, that is, take the convolutional output of the previous layer as the input of the current convolutional layer, perform feature extraction operations on it through the convolutional kernel, and output the feature representation of the current layer. The whole process starts from the initial input and unfolds layer by layer iteratively.

[0035] In the formula, represents the standard two-dimensional convolution operation, whose function is to scan the input feature map in the form of a sliding window and extract regional pattern information. This operation can set different convolutional kernel sizes, strides, and numbers of channels in each layer to adapt to the extraction requirements of multi-scale features.

[0036] S433. A residual connection unit, used to introduce a residual skip connection structure between the two-dimensional convolutional layers, establish a direct mapping relationship between the input and output, and satisfy , to enhance the cross-layer information retention ability, where is the convolutional output of the previous layer; S434, the attention enhancement unit, is used to embed the channel attention mechanism in several specified layers, and adaptively weight and reconstruct the channel dimension. The attention weight is generated by normalizing the input features through a fully connected mapping; S435, the output generation unit, is used to generate the final nuclear radiation dose prediction value at the end of the deep neural network model , and transmit the nuclear radiation dose prediction value to the anomaly detection module and the system control module.

[0037] The S44 specifically includes: S441, the inference state perception unit, is used to receive the real-time allocation result returned by the task scheduling module. The real-time allocation result is a scheduling status data packet, and the scheduling status data packet contains the network bandwidth , the end-side idle computing power and the real-time requirement of the task ; S442, the inference switching decision unit, is used to construct an inference selection function based on the scheduling status data , where the function satisfies: ; Among them, is the pre-designed computing resource and bandwidth threshold, output the model path to be executed currently; The function defined in the formula represents a conditional control strategy. When the remaining computing resources at the end side are greater than or equal to the preset threshold , and the current network bandwidth is lower than the limit threshold , preferentially select the lightweight neural network model to execute the inference process; otherwise, the system automatically switches to the deep neural network model of the edge node to complete the remaining computing tasks.

[0038] S443, the inference path execution unit, is used to control the fusion feature vector to be distributed to the corresponding lightweight neural network model and deep neural network model according to the result of the inference selection function , and synchronously generate the path label of this inference round , where represents end-side inference, represents edge inference; S444, the status label management unit, is used to use the path label Recorded together with the timestamp, task number, and prediction result identifier to form a structured inference status log, and periodically synchronized to the system control module.

[0039] In this embodiment, by constructing a collaborative inference architecture of a lightweight neural network model on the edge side and a deep neural network model on the edge, a phased processing mechanism for fused feature vectors is realized. It can not only quickly generate intermediate feature representations on the edge side to meet the initial response requirements in latency-sensitive scenarios, but also further complete in-depth feature mining and dose prediction at the edge node to support the inference ability of complex models. At the same time, the system ensures the inference consistency and stability between different computing levels through an inference path switching control and model parameter synchronization mechanism, effectively balancing computational efficiency and prediction accuracy, and enhancing the adaptive ability and continuous operation ability of the nuclear radiation monitoring system under multi-scenario deployments.

[0040] S5. A task scheduling module, configured to dynamically determine whether to execute the inference task on the lightweight neural network model on the edge side or the deep neural network model on the edge node based on the current network state and device load; In this embodiment, the S5 task scheduling module includes a status information acquisition unit, a scheduling decision generation unit, and a scheduling instruction broadcasting unit, which are used to dynamically regulate the execution location of the inference task and realize the intelligent allocation of the model path.

[0041] The status information acquisition unit is configured to continuously receive the operation status data from the edge side and the edge node, and the collected content includes but is not limited to scheduling-related parameters such as the current available computing power, memory load, system idle rate, and current network channel bandwidth of the edge device.

[0042] The scheduling decision generation unit is configured to construct a task scheduling logic function according to the currently collected status data, and determine whether it has the ability to complete the inference task on the edge side by comparing with a preset threshold. When the edge resources meet the lightweight inference conditions and the network state is not suitable for data uplink, the system preferentially completes the inference process locally; otherwise, the system will instruct the inference task to be completed by the edge node.

[0043] The scheduling instruction broadcasting unit is configured to send the decision result to the inference path execution module in the form of a control instruction, and update the task path record table of the system control module to ensure that the subsequent modules receive and process data on the correct path.

[0044] In this embodiment, by introducing a task scheduling module based on resource awareness and bandwidth judgment, a flexible switching control mechanism for inference tasks between the edge side and the edge node is realized, which can adjust the model execution location in real time according to the system operation state, balance inference efficiency and resource utilization rate, avoid performance degradation problems caused by single-node computing bottlenecks or network fluctuations, and enhance the scalability, stability, and response speed of the system in various scenarios.

[0045] S6. Anomaly detection module, which is used to execute an anomaly recognition model based on reconstruction error on the nuclear radiation dose prediction result to determine whether there is an abnormal change in the nuclear radiation dose; S61. Prediction result caching unit, which is used to receive and temporarily store the nuclear radiation dose prediction values generated by the edge node inference unit and store them in chronological order to form a sequence sample set , where is the sample time window length; S62. Reconstruction model inference unit, which is used to input the sequence sample set into a preset anomaly recognition model to generate a reconstructed sequence , where represents the reconstructed sequence result of the th predicted value; S63. Error calculation and determination unit, which is used to calculate a reconstruction error sequence based on the difference between the nuclear radiation dose prediction value and the reconstructed sequence at each moment , where the error value is defined as: ; And compare the error value with a set threshold . If there is , then mark this moment as an abnormal point; Specifically, the error value calculation formula measures the difference through the Euclidean distance (L2 norm) between the predicted value and the reconstructed value. This error can reflect whether the prediction result at the current moment deviates from the normal mode. When the error value exceeds the set threshold , the system marks this moment as an abnormal point.

[0046] S64. Abnormal label generation unit, which is used to convert the moment numbers marked as abnormal into an abnormal event label set and send this set to the system control module.

[0047] In this embodiment, by constructing an inference path selection function based on the real-time scheduling state, the comprehensive judgment of the current network bandwidth, edge-side computing power, and task real-time requirements is realized. The edge-side lightweight model or the edge deep model can be dynamically selected for inference task allocation, which not only improves the resource adaptation ability of the model in multiple scenarios, but also enhances the stability and response flexibility of the system under network fluctuations or device load changes, effectively avoiding the problems of latency anomalies or resource waste caused by fixed inference paths, and ensuring the high availability and processing efficiency of the nuclear radiation monitoring system in the edge-side collaboration scenario.

[0048] S7, the system control module, is used to coordinate the communication, data transmission and status scheduling among the various modules, and realize the full-process control.

[0049] In this embodiment, the S7 system control module includes a communication coordination unit, a data synchronization management unit and a status scheduling control unit, which are used to uniformly manage the interaction process among the internal functional modules of the system, and ensure the stability and orderliness of the multi-stage processing process.

[0050] The communication coordination unit is used to establish a data communication channel among the modules, and ensure that the information transmission among the end-side inference module, the edge inference module, the task scheduling module and the anomaly detection module can be carried out with low latency and low packet loss. This unit supports the message queue mechanism and transmission status monitoring, and can automatically adjust the packet size and sending frequency according to the network conditions.

[0051] The data synchronization management unit is used to organize and cache key data streams such as multi-modal perception data, model output results and prediction labels, and perform timing marking on the processing tasks in different time periods, so as to ensure the consistency and traceability of data in the processes of inference, anomaly recognition and early warning output.

[0052] The status scheduling control unit is used to coordinate the model switching, module scheduling and task instruction broadcasting functions. According to the information feedback from the task scheduling module and the system status log, it dynamically adjusts the module execution priority and running status, and periodically transmits the overall system running status back to the upper-layer platform interface.

[0053] Through this system control module, all sub-modules can complete the closed-loop operation process of "acquisition - modeling - inference - recognition - early warning" under the unified scheduling system, and realize the whole-process coordinated control of task scheduling, data flow and status maintenance.

[0054] In this embodiment, by introducing the system control module to centrally manage the communication, data and scheduling tasks, the process docking and control status consistency among the functional units are realized, and the overall operation stability and module collaboration ability of the system are improved; at the same time, based on the module status monitoring and log feedback mechanism, the whole process of the inference path, model output and abnormal events can be recorded and scheduled for tracking, which enhances the manageability, maintainability of the system and the adaptability to complex working conditions, and provides a strong system-level guarantee for the nuclear radiation monitoring task.

[0055] Example: To verify the feasibility of the present invention, the intelligent real-time nuclear radiation dose monitoring system based on deep learning proposed by the present invention was applied to a mobile radiation monitoring drill task deployed by a national nuclear emergency response technical support center in the coastal area of Shandong. This task simulated a scenario of sudden radiation leakage from a nuclear facility, requiring the monitoring system to perform high-frequency, low-latency, and intelligent perception and analysis of nuclear radiation dose in a dynamic mobile environment, and automatically issue warning signals in case of sudden abnormalities to achieve on-site decision support and remote collaborative linkage.

[0056] The drill scenario was set as follows: A nuclear facility malfunctioned and there was a suspected risk of radioactive material diffusion. The emergency response unit needed to quickly deploy an unmanned monitoring vehicle fleet within a 5-kilometer radius of the incident site for data collection and real-time analysis. Due to factors such as dense buildings, complex traffic, and uneven signal coverage in the area, the traditional static point monitoring system was difficult to meet the requirements of rapid response and high-precision discrimination, and the existing equipment could not effectively handle the real-time fusion modeling problems of multi-modal data such as images, position information, and nuclear radiation intensity, seriously restricting the on-site emergency monitoring efficiency and judgment accuracy.

[0057] In this scenario, the system described in the present invention was jointly operated through the end-side perception and reasoning modules deployed on 6 monitoring vehicles and the edge computing nodes set up along the road. The monitoring vehicles were equipped with multi-modal perception units, which periodically collected data including nuclear radiation dose, surrounding environment images, and GNSS positioning information, and judged whether the data was processed by the end-side model or uploaded to the edge node for reasoning through the task scheduling module. After receiving different modal features, the built-in deep fusion module in the system used the attention mechanism for weighted modeling, extracted stable feature vectors and input them into the neural network model. For the possible abnormal dose evolution trend, the system used the reconstruction error strategy for intelligent identification, automatically marked the abnormal period and issued multi-level warning information, which was synchronously transmitted to the mobile command platform in real time.

[0058] During the actual test period, the drill task lasted for 4 hours in total, covering a monitoring radius of about 8.7 kilometers. The system collected a total of more than 32,000 groups of data time series. The system accurately detected three simulated radiation anomaly sources and completed the warning response within no more than 12 seconds after the start of each significant radiation level fluctuation. To clearly show the actual comparison effect before and after the deployment, the following specific data tables are given.

[0059] Table 1: Statistical results of nuclear radiation anomaly detection drill based on the system of the present invention

[0060] It can be seen from Table 1 that in the actual exercise, the system deployed by the present invention showed significant advantages in the whole process of data collection, reasoning judgment and response feedback. Especially in terms of the response speed of nuclear radiation anomaly detection, the average delay dropped from 96 seconds of the traditional system to 11.7 seconds, which greatly shortened the judgment and response time window in the early stage of the accident and improved the efficiency of emergency response. At the same time, under complex terrain and unstable network connection conditions, the system can dynamically adjust the reasoning execution path through the task scheduling module, reasonably allocate end-side reasoning resources, and significantly reduce the impact of end-side resource bottlenecks on judgment speed and accuracy. The overall fusion recognition accuracy of the system reached 97.6%, which is more than 15% higher than the traditional method.

[0061] In a specific scenario, the system successfully identified a simulated abnormal source hidden in a low-lying corner that was difficult to capture by conventional means. The error between its predicted location and the actual source point was only 7.5 meters, which is far lower than the warning error range of traditional equipment. At the same time, during the entire data collection process, the system automatically marked more than 90% of the abnormal event time periods, and cooperated with the multi-channel warning system to send event information to the remote control platform as soon as possible, ensuring the real-time and full coverage of the exercise mission.

[0062] According to feedback from technicians and test reports, after deploying the system of the present invention, the entire nuclear radiation monitoring task has a higher level of automation, recognition accuracy and system stability, which is especially suitable for scenarios with sudden environmental changes, unstable signal conditions or multi-terminal heterogeneous perception. The system not only supports flexible deployment and scalable maintenance, but also has complete closed-loop information processing and intelligent reasoning capabilities, which can significantly improve the recognition efficiency, judgment accuracy and response decision timeliness of abnormal nuclear radiation events in nuclear emergency response tasks.

[0063] In summary, the intelligent nuclear radiation dose real-time monitoring system based on deep learning proposed in the present invention has excellent performance in real environments, especially in data fusion, adaptive reasoning and abnormal intelligent detection, compared with traditional systems, it has significant advantages, which fully proves the practicality, advancement and scalability of this technical solution in complex radiation monitoring tasks.

[0064] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent real-time monitoring system for nuclear radiation dose based on deep learning, characterized in that: The system comprises: S1, a multimodal sensing module, used to collect nuclear radiation intensity data, environmental image data and geographic space location information, and output the radiation intensity data, environmental image data and geographic space location information with timestamps; S2, preprocessing module, used to normalize the nuclear radiation intensity data, perform multi-scale feature extraction on the environmental image data, perform trajectory standardization on the geographic space location information, and output a structured feature sequence; S3, fusion modeling module, used to input the structured feature sequence into the deep fusion network based on the attention mechanism to generate a unified fusion feature vector; S4, the edge inference module, including a lightweight neural network model deployed on the edge and a deep neural network model deployed on the edge node, which is used to process the fused feature vectors in stages and generate nuclear radiation dose prediction results; S5, task scheduling module, used to dynamically determine whether the inference task is executed on the lightweight neural network model on the end side or the deep neural network model on the edge node based on the current network status and device load; S6, anomaly detection module, used to execute an anomaly recognition model based on reconstruction error on the nuclear radiation dose prediction result to determine whether there is an abnormal change in nuclear radiation dose; S7, system control module, is used to coordinate the communication, data transmission and status scheduling among the modules to achieve full process control.

2. According to claim 1, the intelligent real-time monitoring system for nuclear radiation dose based on deep learning is characterized in that: The S3 specifically includes: S31, a modal feature alignment unit, used to map the nuclear radiation intensity feature, the environmental image feature and the geographic space location information feature in the structured feature sequence to a unified feature dimension, and generate an initial modal feature set ,in, represents the uniformly coded nuclear radiation intensity feature vector, represents the feature vector of the environment image after being processed by the convolutional network, Represents the feature vector of the spatially encoded geographic location information; S32, an attention weighted encoding unit, used to introduce a cross-modal attention mechanism between the initial modal feature sets and calculate the attention allocation weights between the modalities: ； in, Representing modal characteristics Modal characteristics The attention allocation weight is , It is a dot product-based correlation scoring function used to measure the feature dependency between modalities; S33, a fusion representation generating unit, used to perform weighted combination of all modal features based on the attention allocation weights to generate a unified fusion feature vector , which is calculated as follows: ； S34, a regularization processing unit, used to process the fused feature vector Apply layer normalization and non-linear activation function operations.

3. According to claim 2, the intelligent real-time monitoring system for nuclear radiation dose based on deep learning is characterized in that: The S4 specifically includes: S41, end-side inference unit, used to deploy lightweight neural network models , receiving the fusion feature vector output by the fusion modeling module , and generate intermediate feature representation based on the fused feature vector , It includes multi-layer linear mapping and normalization modules, and the calculation relationship is: ； in, is the weight matrix, is the bias term, is a nonlinear activation function; S42, an intermediate feature buffer unit, used to represent the intermediate feature Stored in the local cache area, and determine whether to transmit uplink according to the inference allocation instruction issued by the system control module; S43, edge node inference unit, used to deploy deep neural network models , after receiving the intermediate feature representation Then execute the subsequent reasoning process to generate the final nuclear radiation dose prediction value : ； in, Contains multiple convolution, residual connection and attention mechanism modules to extract deep correlation features; S44, the inference phase switching control unit is used to control the inference process in the lightweight neural network model according to the real-time allocation result returned by the system task scheduling module and deep neural network models Switch between and mark the current reasoning state; S45, a model parameter synchronization unit, used to periodically check the model version number and parameter verification code of the end side and the edge node, and complete the consistency update of the weight data through the system control module; S46, abnormal interrupt processing unit, is used to automatically trigger the edge node to take over the complete reasoning task process when the end-side reasoning process or data transmission is interrupted, and return the prediction status label to the system control module after recovery.

4. According to claim 3, the intelligent real-time monitoring system for nuclear radiation dose based on deep learning is characterized in that: The S41 specifically includes: S411, input feature receiving unit, used to receive the fusion feature vector output by the fusion modeling module ,in, represents the feature dimension, represents the field of real numbers; S412, a linear mapping processing unit, used to input the fused feature vector into a lightweight neural network model The lightweight neural network model consists of multiple linear mapping layers, where The structure of the layer map is: ； in, Indicates The weight matrix of the layer, represents the bias vector, is the output of the previous layer, and the initial input satisfies ; S413, an activation and normalization unit, used to perform nonlinear activation and normalization processing after the partial linear mapping layer; S414, output generation unit, used to output intermediate feature representation at the end of the model ,in, Represents the total number of layers of the neural network model, and Send to the intermediate feature buffer unit.

5. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 3 is characterized in that: The S43 specifically includes: S431, edge feature receiving unit, used to receive the intermediate feature representation from the end side ,in, Represents the intermediate feature dimension and feeds it into the deep neural network model of the edge node middle; S432, convolution feature extraction unit, used to represent the intermediate features The initial spatial dimension is expanded, and local context features are extracted through multiple two-dimensional convolutional layers. The convolution calculation structure is: ； in, represents the standard convolution operation, For the Layer convolution output; S433, a residual connection unit, used to introduce a residual skip connection structure between the two-dimensional convolutional layers to establish a direct mapping relationship between input and output, satisfying , to enhance the cross-layer information retention capability, where It is the convolution output of the previous layer; S434, an attention enhancement unit, used to embed a channel attention mechanism in a number of designated layers, and perform adaptive weighted reconstruction on the channel dimension, wherein the attention weight is generated by normalizing the input features through full-connection mapping; S435, output generation unit, used to generate the final nuclear radiation dose prediction value at the end of the deep neural network model And transmit the predicted nuclear radiation dose value to the anomaly detection module and the system control module.

6. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 3 is characterized in that: The S44 specifically includes: S441, the reasoning state perception unit is used to receive the real-time allocation result returned by the task scheduling module, the real-time allocation result is a scheduling state data packet, and the scheduling state data packet contains the network bandwidth , Idle computing power on the client side Real-time requirements of tasks ; S442, inference switching decision unit, used to build an inference selection function based on the scheduling state data , where the function satisfies: ； in, To preset computing resources and bandwidth thresholds, Output the model path that should be executed currently; S443, a reasoning path execution unit, configured to select a function according to the reasoning The result controls the fusion feature vector Distribute to the corresponding lightweight neural network model and deep neural network model, and simultaneously generate the path label for this round of reasoning ,in, Indicates client-side reasoning. represents marginal reasoning; S444, a status tag management unit is used to set the path tag It is recorded together with the timestamp, task number and prediction result identifier to form a structured reasoning status log, and is periodically synchronized to the system control module.

7. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 5 is characterized in that: The S6 specifically includes: S61, prediction result cache unit, used to receive and temporarily store the nuclear radiation dose prediction value generated by the edge node reasoning unit , and store them in time order to form a sequence sample set ,in, is the length of the sample time window; S62, reconstructing a model inference unit, for inputting the sequence sample set into a preset anomaly recognition model , generating the reconstruction sequence ,in, Indicates The reconstructed sequence results of predicted values; S63, error calculation and determination unit, used to calculate the reconstruction error sequence based on the difference between the nuclear radiation dose prediction value at each moment and the reconstruction sequence , where the error value is defined as: ； The error value is compared with the set threshold Compare, if there is , then mark this moment as an abnormal point; S64, an abnormal label generation unit, used to number the moments marked as abnormal Convert to abnormal event label set , and sends the collection to the system control module.

Images