Real-time high-precision radiation environment monitor based on autoregression encoder

Abstract

The invention discloses a real-time high-precision radiation environment monitor based on an autoregression encoder. The real-time high-precision radiation environment monitor comprises a data acquisition module and an upper computer, the data acquisition module comprises a gamma dose rate sensor, a temperature sensor and a humidity sensor; and the data acquisition module and the upper computer are connected in sequence. The upper computer comprises a data processing module, a real-time prediction module, a data storage unit and a result display module. The real-time high-precision radiation environment monitor based on the regression encoder has the advantages of being high in real-time performance and high in prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of environmental monitoring, and particularly relates to a real-time high-precision radiation environment monitor based on an auto-regressive encoder. Background Art

[0002] Radiation environment monitoring has important application values in the fields of nuclear energy management, environmental protection, public safety, etc. Accurately predicting the change of radiation signals can not only provide an important basis for environmental protection, but also be crucial for nuclear accident early warning and disaster emergency. However, the prediction of radiation signals faces numerous challenges, which mainly come from the following aspects:

[0003] Firstly, the sources of radiation signals are diverse, including natural radiation, industrial radiation, and human activities, etc. The intensity and distribution of these signals are affected by various factors such as meteorological conditions and geographical locations, so they have high complexity and uncertainty. Secondly, the existing radiation signal collection points are relatively sparse, resulting in a limited amount of available historical data. Traditional prediction methods are difficult to make full use of these limited data for efficient prediction. In addition, the radiation environment has highly dynamic characteristics, requiring the prediction system to not only respond in real time, but also effectively identify complex change trends and emergencies.

[0004] Currently, traditional radiation signal prediction methods mainly rely on statistical analysis and empirical models. These methods have poor effects in dealing with non-linear and high-dimensional data and are difficult to meet the requirements of real-time and high-precision. With the development of deep learning technology, more and more intelligent algorithms are applied to radiation signal prediction. However, most of the existing deep learning models are difficult to effectively capture the complex spatio-temporal characteristics in radiation signals under limited data conditions and cannot provide sufficiently accurate prediction results under the requirement of high real-time. Therefore, the core problem of radiation signal prediction lies in how to accurately extract effective features in radiation signals under limited data conditions and provide high-precision and real-time prediction results.

[0005] In order to overcome these technical bottlenecks, the present invention innovatively introduces a framework based on an auto-regressive encoder and combines a multi-scale feature extraction mechanism, which can not only maintain high prediction accuracy when dealing with complex environmental data, but also have strong real-time performance and robustness. Specifically, the technical innovation point of the present invention is to adopt an auto-regressive mechanism to dynamically optimize and adjust the prediction results at each step, further improving the prediction accuracy. In addition, through multi-scale processing technology, the present invention can extract the long-term trends and periodic characteristics of radiation signals at multiple levels, effectively suppressing short-term noise and interference, thereby improving the overall stability and accuracy of the model.

[0006] The implementation of the present invention can significantly improve the accuracy and timeliness of radiation signal prediction, providing a new technical path for radiation environment monitoring, and having important theoretical value and practical application prospects. Summary of the Invention

[0007] In order to overcome the problem of accurately extracting effective features in radiation signals under limited data conditions and providing high-precision and real-time prediction results, the object of the present invention is to provide a real-time high-precision radiation environment monitor based on an autoencoder for radiation environment monitoring, which can significantly improve the accuracy and timeliness of radiation signal prediction.

[0008] The technical solution adopted by the present invention to solve its technical problems is: a real-time high-precision radiation environment monitor based on an autoencoder, including a data acquisition module and a host computer; the data acquisition module includes a gamma dose rate sensor, a temperature sensor and a humidity sensor, which are respectively used to collect radiation signal data, temperature data and humidity data in the environment in real time; the host computer includes a data processing module, a real-time prediction module and a data storage unit;

[0009] The data processing module is used to perform standardization processing on the radiation signal, temperature and humidity data in the data acquisition module;

[0010] The real-time prediction module realizes real-time prediction through the analysis of historical environment data and real-time environment data by the prediction model. The environment data includes radiation signal data, temperature data and humidity data; and a multi-step optimization and correction mechanism is introduced: the prediction model not only uses historical environment data for trend prediction, but also takes the prediction result of each step as input to correct and optimize the subsequent prediction results;

[0011] The data storage unit is used to store historical environment data, real-time environment data and prediction results generated by the real-time prediction module.

[0012] Furthermore, the data processing module is completed by the following process:

[0013] In the data processing module, the radiation signal, temperature and humidity data collected by the data acquisition module need to be standardized to eliminate the influence of different dimensions and magnitudes on the subsequent processing process. The formula for the standardization process is as follows:

[0014] For each dimension of data x i , that is, radiation signal, temperature data or humidity data, its standardization process is:

[0015]

[0016] where x t represents the data after standardization processing, xi Denote the \(i\)-th data point, and \(\mu\) is the mean of this data set, with the calculation formula as follows:

[0017]

[0018] \(\sigma\) is the standard deviation of this data set, with the calculation formula as follows:

[0019]

[0020] Furthermore, the real-time prediction module is completed by the following process:

[0021] The prediction model innovatively introduces the design of an autoregressive framework, uses historical environmental data for multi-step prediction, and gradually improves the accuracy and stability of the prediction by continuously feedbacking and optimizing the prediction results of each step. The prediction at each time step \(t\) of the prediction model is expressed as:

[0022]

[0023] where \(n\) represents the length of the historical environmental data, represents the prediction result at time step \(t\), \(x\) t-1 , \(x\) t-2 , …, \(x\) t-n is the historical environmental data sequence, and \(f(·)\) is the prediction function of the prediction model. By taking the result of each prediction as the new input, the system can gradually correct and optimize subsequent predictions:

[0024]

[0025] where \(k\) represents the length of future prediction.

[0026] Furthermore, the real-time prediction module also includes a seasonal decomposition module, which is used to extract long-term trends and periodic characteristics, can better handle complex and variable environmental data, improve the prediction accuracy and stability, and is especially suitable for real-time monitoring scenarios with strong seasonal fluctuations and noise interference such as radiation environment; it is completed by the following process:

[0027] Adopt a multi-scale decomposition method to decompose the data at multiple frequencies to adapt to different periodic characteristics. Specifically, use the Empirical Mode Decomposition (EMD) method to decompose the data layer by layer from low frequency (long-term trend) to high frequency (short-term fluctuation):

[0028]

[0029] where \(x\) t is the data after standardized processing; \(T\) tis the long-term trend component, representing the long-term change trend of the data; S t,i is the seasonal fluctuation component of the i-th layer, which can capture the periodic fluctuations on different time scales; R t is the residual component, representing the noise or short-term fluctuations that cannot be explained by the trend and seasonal fluctuations.

[0030] It also includes: an adaptive period detection mechanism is introduced; the specific period detection is realized through the following process:

[0031] Frequency Spectrum = FFT(T t )

[0032] Then, the seasonal period length p is dynamically determined by the frequency component with the maximum energy in the frequency spectrum:

[0033]

[0034] where P is the set of all possible period lengths, and FFT p is the amplitude corresponding to a certain period in the frequency spectrum. The periodic component S t,i is reconstructed from the detected period length P and the frequency spectrum analysis result of the original signal:

[0035]

[0036] where A and B are the amplitude coefficients obtained through frequency spectrum analysis.

[0037] It also includes: a non-linear noise suppression algorithm is introduced, which combines the adaptive filter and Kalman Filter technology to model and suppress the high-frequency noise components; specifically:

[0038] By introducing a filter in the high-frequency components, the residual R t is corrected and optimized:

[0039] R′ t = Kalman(R t , Q, R)

[0040] where Q and R represent the covariance matrices of the process noise and the observation noise respectively, and the Kalman filter can adaptively adjust the residual component according to the historical data and the noise characteristics.

[0041] Recombine each component to obtain the reconstructed x t ′ :

[0042]

[0043] Furthermore, it also includes: introducing the GQA mechanism, i.e., the grouped query attention mechanism, which can perform fine-grained attention allocation between different features and time steps when processing complex time-series data, improve the prediction accuracy, and enhance the adaptability and robustness of the model; the GQA mechanism groups the queries, and each group represents the features on a certain data type or time scale. The reconstructed x t ′ is input into the linear transformation layer to map it to a high-dimensional feature space; the specific formula is as follows:

[0044] Q t,g = W Q ·x′ t + b Q

[0045] K t,g = W K ·x′ t + b K

[0046] where W Q and W K are the query weight matrix and the key weight matrix respectively; b Q and b K are the bias vector of the query and the bias vector of the key respectively; Q t,g is the query vector within the g-th group at the t-th time step; K t,g is the key vector within the g-th group at the t-th time step; Q i,g is the query vector within the g-th group at the i-th time step; K i,g is the key vector within the g-th group at the i-th time step; T represents the total number of time steps; the attention calculation of GQA is expressed as:

[0047]

[0048] α t,g is the weight of the query Q t,g in the g-th group; score(·) is the similarity between the query vector and the key vector, calculated by the dot product method.

[0049] It also includes introducing an adaptive adjustment mechanism; the dynamic adjustment is carried out through the following formula:

[0050] Q′ t,g = Linear(x t,g + λ·H t,g )

[0051] where x t,g is the input feature of the current time step y for the g-th group; H t,g is the historical information x t ′, representing the understanding degree of the prediction model for this set of features; λ is a dynamic weight adjustment factor, which is automatically adjusted according to the influence of historical environmental data; Q t ′ ,g is the updated query vector. According to Q t ′ ,g , recalculate the attention to obtain α t ′ ,g . According to the attention weight α t ′ ,g , perform weighted combination on the reconstructed time series x t ′ to generate the prediction result Its formula is:

[0052]

[0053] Furthermore, the data storage unit is used to calculate the prediction error by storing the historical environmental data and the prediction results generated by the real-time prediction module, and is used to dynamically optimize the prediction model parameters; specifically:

[0054] 4.1 Historical environmental data storage

[0055] The data storage unit first stores the historical environmental data and the corresponding real-time environmental data, as well as the prediction results at each time step; these data are used to provide the training and prediction input of the prediction model. For each data point, the system will record its original value and the corresponding predicted value.

[0056] 4.2 MSE calculation and storage

[0057] Calculate the error MSE between the true value and the predicted value of the historical environmental data:

[0058]

[0059] where n is the number of stored data points; x t is the historical environmental data, that is, the data after normalization processing; is the historical prediction result. The calculated MSE will be used as a feedback signal to the prediction model for parameter update and optimization. Furthermore, the host computer also includes a result display module, which is used to provide various forms of real-time display interfaces, including graphical trend analysis, abnormal alarm prompt, and visual prediction data report, to facilitate users to view and manage the monitoring data.

[0060] The result display module presents real-time environmental data and prediction results in a graphical manner, providing time series charts including multi-dimensional data, where the multi-dimensional data are radiation signals, temperature data, and humidity data; the predicted values and actual values of each data type will be identified with different colors or graphics to help users intuitively identify trend changes and error fluctuations. On each monitoring interface, users can see the change trends of environmental data in real time and the comparison with the prediction results; the data types are radiation signals, temperature data, and humidity data.

[0061] The beneficial effects of the present invention are as follows: 1) The innovative introduction of the design of the autoregressive framework enables the use of historical environmental data for multi-step prediction, and by continuously feedbacking and optimizing the prediction results of each step, the prediction accuracy and stability can be gradually improved.

[0062] 2) The innovative introduction of the seasonal decomposition function to extract long-term trends and periodic characteristics can better handle complex and variable environmental data, improve prediction accuracy and stability, and is especially suitable for real-time monitoring scenarios such as radiation environments with strong seasonal fluctuations and noise interference.

[0063] 3) The innovative introduction of the grouped query attention mechanism can perform fine-grained attention allocation between different features and time steps when processing complex time series data, improve prediction accuracy, and enhance the adaptability and robustness of the model.

[0064] 4) This predictor can significantly improve the accuracy and timeliness of radiation signal prediction, providing a new technical path for radiation environment monitoring, and has important theoretical value and practical application prospects. Description of the Drawings

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0066] Figure 1 It is the functional structure diagram of a real-time high-precision radiation environment monitor based on an autoregressive encoder proposed by the present invention. Detailed Embodiments

[0067] The following will describe the present invention in detail with reference to the drawings. Without conflict, the features in the following embodiments and implementation manners can be combined with each other.

[0068] Embodiment 1

[0069] Refer to Figure 1, A real-time high-precision radiation environment monitor based on an auto-regressive encoder, comprising a data acquisition module 4 and a host computer; the data acquisition module 4 includes a gamma dose rate sensor 1, a temperature sensor 2, and a humidity sensor 3, which are respectively used to collect radiation signal data, temperature data, and humidity data in the environment, and these data are the basis for subsequent processing and environmental detection. The data acquisition module 4 and the host computer are connected in sequence. The host computer includes a data processing module 5, a real-time prediction module 6, a data storage unit 7, and a result display module 8.

[0070] The data processing module 5 is responsible for preprocessing the radiation signal, temperature, and humidity data in the data acquisition module to ensure the stability and reliability of data acquisition.

[0071] In the data processing module 5, the radiation signal, temperature, and humidity data collected by the data acquisition module 4 need to be standardized to eliminate the influence of different dimensions and magnitudes on the subsequent processing process. The formula for the standardization process is as follows:

[0072] For each dimension of data x i (such as radiation signal, temperature, or humidity), its standardization process is:

[0073]

[0074] where x represents the data after standardization, x i represents the i-th data point, μ is the mean of this data set, and the calculation formula is:

[0075]

[0076] σ is the standard deviation of this data set, and the calculation formula is:

[0077]

[0078] Through the above standardization process, the data will be converted into a standard normal distribution with a mean of 0 and a standard deviation of 1. This processing step can effectively avoid the adverse effects of different feature scale differences on the model training process, ensure equal weights of various data in the model, and thus improve the prediction accuracy and the training efficiency of the model.

[0079] 2. The real-time prediction module 6 is completed by the following process:

[0080] 2.1 Core auto-regressive framework and multi-step optimization mechanism:

[0081] The core innovation of the real-time prediction module 6 is the design of an autoregressive framework, namely the prediction model. The prediction model uses historical environmental data and real-time environmental data for multi-step prediction, and gradually improves the prediction accuracy and stability by continuously feedbacking and optimizing the prediction results of each step. The prediction at each time step t of the prediction model can be expressed as:

[0082]

[0083] where n represents the length of historical environmental data, represents the prediction result at time step t, x t-1 , x t-2 , …, x t-n is the historical environmental data sequence, and f(·) is the prediction function of the prediction model. By taking the result of each prediction as the new input, the subsequent predictions are gradually corrected and optimized:

[0084]

[0085] where k represents the length of future prediction. This optimization mechanism based on the autoregressive framework ensures that each prediction result can be used to adjust future predictions, reduce the accumulation of errors, and improve the accuracy.

[0086] 2.2 Seasonal decomposition function: Extract long-term trends and periodic characteristics

[0087] The seasonal decomposition module is a key innovation in the real-time prediction module 6. Its core goal is to extract and separate the complex time series components in the radiation signal, temperature, and humidity data through multi-level decomposition, so as to more accurately capture the long-term trends and periodic changes, and at the same time effectively remove noise and short-term fluctuations, improving the prediction accuracy and stability. Traditional seasonal decomposition methods generally split the data into three parts: long-term trend, seasonal fluctuation, and residual, while this design adopts a complex multi-scale decomposition method and an adaptive prediction model to cope with the changing characteristics of environmental data.

[0088] Traditional seasonal decomposition methods usually analyze based on a fixed time scale, but in the actual environment, data such as radiation signals, temperature, and humidity may exhibit seasonal variations of multiple different frequencies. For example, the radiation signal may show annual periodic fluctuations, while temperature and humidity may have strong seasonal fluctuations in a shorter cycle (such as day and night changes). Therefore, this module adopts a multi-scale decomposition method to decompose the data at multiple frequencies to adapt to different periodic characteristics. Specifically, the empirical mode decomposition (EMD) method is used to decompose the data layer by layer from low frequency (long-term trend) to high frequency (short-term fluctuation):

[0089]

[0090] Among them, x t is the data after standardization; T t is the long-term trend component, representing the long-term change trend of the data; S t,i is the seasonal fluctuation component of the i-th layer, which can capture the periodic fluctuations on different time scales; R t is the residual component, representing the noise or short-term fluctuations that cannot be explained by the trend and seasonal fluctuations. This multi-scale decomposition can not only accurately extract the periodic fluctuations of different frequencies, but also finely model the trend, thus helping the prediction model to better understand the structure of the data.

[0091] To further improve the adaptability to different data cycles, this module introduces an adaptive cycle detection mechanism. In traditional seasonal decomposition, the cycle length is usually preset and cannot flexibly cope with the dynamic changes of the cycles in the data. In this module, however, the cycle detection is dynamically adaptive, and the optimal cycle length is automatically determined according to the characteristics of the data. This mechanism relies on the spectral analysis of the data (e.g., through the fast Fourier transform, FFT), and determines the cycle by detecting the significant frequency components in the data. The specific cycle detection can be achieved through the following process:

[0092] Frequency Spectrum = FFT(T t )

[0093] Then, the seasonal cycle length p is dynamically determined by the frequency component with the maximum energy in the spectrum:

[0094]

[0095] where P is the set of all possible cycle lengths, and FFT p is the amplitude corresponding to a certain cycle in the spectrum. The periodic component S t,i can be reconstructed from the detected cycle length P and the spectral analysis result of the original signal:

[0096]

[0097] where A and B are the amplitude coefficients obtained through spectral analysis. Through adaptive cycle detection, the system can automatically adjust the cycle of seasonal decomposition according to the changes in environmental data, enabling the prediction model to accurately capture the potential seasonal patterns in the data.

[0098] During the seasonal decomposition process, there are often some inevitable noises or short-term fluctuations. Traditional methods usually classify these fluctuations as "residuals". However, in some cases, there may be useful information in these noises, especially for non-linear and sudden environmental changes. Therefore, this module introduces a non-linear noise suppression algorithm, which combines adaptive filter and Kalman Filter technology to accurately model and suppress high-frequency noise components.

[0099] This method introduces a filter in the high-frequency components to correct and optimize the residual R t to avoid the influence of noise interference on the prediction results:

[0100] R' t = Kalman(R t , Q, R)

[0101] where Q and R represent the covariance matrices of process noise and observation noise respectively. The Kalman filter can adaptively adjust the residual components according to historical environmental data and noise characteristics.

[0102] Recombine each component to obtain the reconstructed x': t :

[0103]

[0104] Through the above innovative method, the seasonal decomposition function can better handle complex and variable environmental data, improve the prediction accuracy and stability, and is especially suitable for real-time monitoring scenarios such as radiation environment with strong seasonal fluctuations and noise interference.

[0105] 2.3 Grouped Query Attention Mechanism

[0106] To further improve the adaptability of the real-time prediction module 6 to complex environmental data, this module introduces the Grouped Query Attention (GQA) mechanism. As an innovative attention model, when dealing with complex time-series data, it can perform fine-grained attention allocation between different features and time steps, improve the prediction accuracy, and enhance the adaptability and robustness of the model. In this design, the GQA mechanism groups the queries, and each group represents the features on a certain data type or time scale. The query vectors within each group are specifically used to process the data related to this feature type, so that the model can adaptively adjust its attention allocation in different data dimensions. The reconstructed x' is input into the linear transformation layer to map it to a high-dimensional feature space. The specific formula is as follows:

[0107] Q t,g = W Q ·x' t + bQ

[0108] K t,g = W K ·x′ t + b K

[0109] Wherein, W Q and W K are the query weight matrix and the key weight matrix respectively; b Q and b K are the bias vector of the query and the bias vector of the key respectively; Q t,g is the query vector in the g-th group at the t-th time step; K t,g is the key vector in the g-th group at the t-th time step; Specifically, the attention calculation of Grouped Query Attention can be expressed as:

[0110]

[0111] α t,g is the weight of the query Q t,g in the g-th group; score(·) is the similarity between the query vector and the key vector, calculated by the dot product method.

[0112] In addition to grouping the queries, the innovation of this module also lies in its adaptive adjustment mechanism. According to the changes in real-time environmental data, the prediction model can automatically determine which features need more attention. Through adaptive learning, the system can adjust the weights of each group at different time steps, so as to flexibly respond to different environmental changes. For example, when a certain dimension (such as temperature) becomes more influential on the radiation signal, the system will increase the query weight of this dimension; while at other times, the system will automatically reduce its attention to this dimension.

[0113] This adaptive mechanism is dynamically adjusted by the following formula:

[0114] Q′ t,g = Linear(x t,g + λ·H t,g )

[0115] Wherein, x t,h is the input feature (such as temperature, humidity, etc.) of the g-th group at the current time step t; H t,g is the historical information predicted before, indicating the understanding degree of the prediction model for the features of this group; λ is the dynamic weight adjustment factor, automatically adjusted according to the influence of historical environmental data; Q′ t,g is the updated query vector. According to Q′ t,g , recalculate the attention to obtain α′ t,g . According to the attention weight α′t,g , perform weighted combination on the reconstructed time series x' t to generate the prediction result The formula is as follows:

[0116]

[0117] Through this mechanism, the prediction model can automatically adjust the query vectors of each group according to the change of feature importance to ensure accurate capture of key features. By introducing Grouped Query Attention, this module can perform fine-grained attention allocation between different features and time steps when processing complex time series data, improve the prediction accuracy, and enhance the adaptability and robustness of the prediction model.

[0118] 3. Data storage unit 7 is completed through the following process:

[0119] Data storage unit 7 plays a crucial role in the entire system, responsible for storing historical environmental data, real-time environmental data, and real-time prediction results, and providing data support for the autoregressive framework. This unit is not only used for storing and maintaining data, but also provides the basis for the optimization process of autoregressive prediction. By storing historical environmental data and prediction results, it calculates and feedbacks the mean square error (MSE) to achieve the optimization and accuracy improvement of the prediction model. Data storage unit 7 can ensure the stability and efficiency of the prediction model and is a key component of the autoregressive framework and prediction optimization.

[0120] 3.1 Storage of historical environmental data

[0121] Data storage unit 7 first stores all historical environmental data and corresponding real-time environmental data (such as radiation signals, temperature, humidity, etc.), as well as prediction results at each time step. These data are used to provide training and prediction input for the prediction model. For each data point, the system records its original value and the corresponding predicted value.

[0122] 3.2 Calculation and storage of MSE

[0123] In the autoregressive framework, the error between the true value and the predicted value of historical environmental data is crucial for the optimization process. Data storage unit 7 is responsible for calculating and storing the MSE of historical time steps. The error calculation formula is as follows:

[0124]

[0125] where n is the number of stored data points; x t is the historical environmental data; These are historical prediction results. The calculated MSE will be passed as a feedback signal to the prediction model for parameter update and optimization in the autoregressive framework. Specifically, the storage unit will store the MSE value corresponding to each time step t for subsequent evaluation and adjustment of the prediction model performance.

[0126] 4. The host computer further includes: a result display module 8.

[0127] The result display module 8 displays the real-time environmental data and prediction results in a graphical manner, providing time series charts of multi-dimensional data including radiation signals, temperature, humidity, etc. The predicted values and actual values of each data type (such as radiation signals, temperature, humidity) will be identified with different colors or graphics to help users visually identify trend changes and error fluctuations. On each monitoring interface, users can see the change trends of environmental data in real time and the comparison with the prediction results. This real-time display function can help users promptly grasp the dynamic changes of the radiation environment, especially providing very valuable information during environmental monitoring, early warning, and response.

[0128] The above embodiments are only used to illustrate the design concept and characteristics of the present invention, and the purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made according to the principles and design concepts disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A real-time high-precision radiation environment monitor based on an autoregressive encoder, characterized in that: It includes a data acquisition module and a host computer; the data acquisition module includes a gamma dose rate sensor, a temperature sensor and a humidity sensor, which are respectively used to collect radiation signal data, temperature data and humidity data in the environment in real time; the host computer includes a data processing module, a real-time prediction module and a data storage unit; The data processing module is used to perform standardization processing on the radiation signal, temperature and humidity data in the data acquisition module; The real-time prediction module realizes real-time prediction by analyzing historical environmental data and real-time environmental data through the prediction model, and the environmental data includes radiation signal data, temperature data and humidity data; and introduces a multi-step optimization and correction mechanism: the prediction model not only uses historical environmental data for trend prediction, but also takes the prediction results of each step as input to correct and optimize the subsequent prediction results; The data storage unit is used to store historical environmental data, real-time environmental data and prediction results generated by the real-time prediction module.

2. A real-time high-precision radiation environment monitor based on an autoregressive encoder according to claim 1, characterized in that: The data processing module is completed using the following process: In the data processing module, the radiation signal, temperature and humidity data collected in the data acquisition module need to be standardized; the formula of the standardization process is as follows: For each dimension of data x i , that is, radiation signal, temperature data or humidity data, the standardization process is: where x t represents the data after normalization, x i represents the i-th data point, μ is the mean of the data set, and the calculation formula is: σ is the standard deviation of the data set, and the calculation formula is:

3. A real-time high-precision radiation environment monitor based on an autoregressive encoder according to claim 1, characterized in that: The real-time prediction module is completed using the following process: The prediction model uses historical environmental data to make multi-step predictions, and gradually improves the accuracy and stability of the prediction by continuously feeding back and optimizing the prediction results of each step; the prediction of the prediction model at each time step t is expressed as: Where n represents the length of historical environmental data, represents the prediction result at time step t, x t-1 ,x t-2 ,…,x t-n is the historical environmental data sequence, f(·) is the prediction function of the prediction model; As new inputs are given, the system is able to progressively correct and optimize subsequent predictions: Here, k represents the length of future predictions.

4. A real-time high-precision radiation environment monitor based on an autoregressive encoder according to claim 3, characterized in that: The real-time prediction module also includes a seasonal decomposition module, which is completed using the following process: Using the Empirical Mode Decomposition (EMD) method, the data is decomposed layer by layer from low frequency (long-term trend) to high frequency (short-term fluctuation): Among them, x t is the data after standardization; T t S is the long-term trend component, representing the long-term trend of data; t,i is the seasonal fluctuation component of the i-th layer, which can capture the periodic fluctuations on different time scales; R t is the residual component, representing noise or short-term fluctuations that cannot be explained by trend and seasonal fluctuations; It also includes: introducing an adaptive period detection mechanism; the specific period detection is implemented through the following process: Frequency Spectrum=FFT(T t ) Then, the seasonal cycle length p is dynamically determined by the frequency component with the maximum energy in the spectrum: Where P is the set of all possible cycle lengths, FFT p is the amplitude corresponding to a certain period in the spectrum; the periodic component S t,i Reconstructed from the detected period length P and the spectrum analysis results of the original signal: Among them, A and B are the amplitude coefficients obtained by spectrum analysis; It also includes: introducing a nonlinear noise suppression algorithm, combining adaptive filters and Kalman filter technology to model and suppress high-frequency noise components; specifically: By introducing a filter in the high frequency component, the residual R t Corrections and optimizations: R′ t =Kalman(R t ,Q,R) Among them, Q and R represent the covariance matrices of process noise and observation noise respectively. The Kalman filter can adaptively adjust the residual component according to historical data and noise characteristics; Recombining the components, we get the reconstructed x′ t :

5. A real-time high-precision radiation environment monitor based on an autoregressive encoder according to claim 4, characterized in that: Also includes: Introduce the GQA mechanism; the GQA mechanism groups queries, each group represents a data type or feature on a time scale; the reconstructed x′ t It is input into the linear transformation layer to map it into a high-dimensional feature space; The specific formula is as follows: Q t,g =W Q ·x′ t +b Q K t,g =W K ·x′ t +b K Among them, W Q and W K are the query weight matrix and the key weight matrix respectively; b Q and b K are the query bias vector and the key bias vector respectively; Q t,g is the query vector in the gth group at the tth time step; K t,g is the key vector in the gth group at the tth time step; Q i,g is the query vector in the gth group at the i-th time step; K i,g is the key vector of the i-th time step in the g-th group; T represents the total time step; the attention calculation of GQA is expressed as: α t,g It is the query Q t,g The weight in the gth group; score(·) is the similarity between the query vector and the key vector, calculated using the dot product method; It also includes the introduction of an adaptive adjustment mechanism; dynamic adjustment is performed through the following formula: Q′ t,g =Linear(x t,g +λ·H t,g ) Among them, x t,g is the input feature of the gth group at the current time step t; H t,g is the historical information x′ predicted before t , which indicates the degree of understanding of the prediction model for this group of features; λ is a dynamic weight adjustment factor, which is automatically adjusted according to the influence of historical environmental data; Q′ t,g is the updated query vector; according to Q′ t,g , recalculate the attention and get α′ t,g ; According to the attention weight α′ t,g , for the reconstructed time series x′ t Perform weighted combination to generate prediction results The formula is:

6. A real-time high-precision radiation environment monitor based on an autoregressive encoder according to claim 1, characterized in that: The data storage unit is used to store historical environmental data and prediction results generated by the real-time prediction module, calculate prediction errors, and dynamically optimize prediction model parameters; specifically: 4.1 Historical environmental data storage The data storage unit first stores the historical environmental data and the corresponding real-time environmental data, as well as the prediction results of each time step; these data are used to provide training and prediction input for the prediction model; for each data point, the system records its original value and the corresponding predicted value; 4.2 MSE Calculation and Storage Calculate the error MSE between the actual value and the predicted value of historical environmental data: Where n is the number of stored data points; x t It is historical environmental data, that is, data that has been standardized; is the historical prediction result; the calculated MSE will be passed to the prediction model as a feedback signal for parameter updating and optimization.

7. A real-time high-precision radiation environment monitor based on an autoregressive encoder according to claim 1, characterized in that: The host computer also includes a result display module, which is used to provide real-time display interfaces in various forms, including graphical trend analysis, abnormal alarm prompts, and visual forecast data reports, so that users can view and manage monitoring data; The result display module displays environmental data and prediction results in a graphical way, providing a time series chart including multi-dimensional data, such as radiation signal, temperature data, and humidity data; the predicted value and actual value of each data type will be marked with different colors or graphics to help users intuitively identify trend changes and error fluctuations; on each monitoring interface, users can see the changing trend of environmental data in real time, as well as the comparison with the prediction results; Data types include radiation signals, temperature data, and humidity data.

Images