Multi-source data fusion method and system in radionuclide separation process

Abstract

The application provides a multi-source data fusion method and system in a radionuclide separation process, and relates to the technical field of data processing.The method comprises the following steps: based on the statistical characteristics of the characteristic element points in the element interpretation block, and in combination with the relationship between the process condition parameter data and the waste liquid physical property parameter data corresponding to the element interpretation block, a dynamic nuclear interpretation calibration primitive is generated; the multi-source monitoring data is corrected by using the dynamic nuclear interpretation calibration primitive to obtain a fusion data base; the fusion data base is subjected to feature extraction to extract feature information related to the radionuclide type and concentration, and a multi-dimensional feature vector is generated; the multi-dimensional feature vector is input into a pre-trained pattern recognition model for classification and fusion recognition to obtain the type recognition result of the radionuclide and the concentration prediction value.The application improves the reliability of the data in the radionuclide separation process.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for multi-source data fusion during the separation of radionuclides. Background Technology

[0002] In the process of separating radionuclides, accurately identifying the types of nuclides and precisely detecting their concentrations are the core prerequisites for ensuring efficient process optimization, safe waste disposal, and environmental radiation monitoring. They are directly related to the stability and safety of the separation process. Currently, the mainstream methods for qualitative and quantitative detection of nuclides mainly rely on gamma-ray energy dispersive spectroscopy. This method, with its ability to identify the characteristic energy peaks of nuclides, can meet basic detection needs in conventional scenarios.

[0003] However, in actual radionuclide separation operations, the processing of gamma-ray energy spectrum data is easily affected by various complex factors, leading to insufficient data processing accuracy and consequently affecting the stability of nuclide identification and concentration detection. For example, spatiotemporal fluctuations in ambient background radiation can introduce a large amount of noise into the raw energy spectrum data, increasing the difficulty of energy spectrum signal denoising and characteristic peak identification, and easily causing the loss of characteristic information of low-concentration nuclides. The performance drift caused by long-term operation of the detector can lead to peak position shifts and peak shape distortions in the energy spectrum data, making it difficult to accurately carry out subsequent data correction, feature extraction, and other processing steps. At the same time, the dynamic changes in operating parameters such as temperature, pressure, and feed flow rate of the separation process, as well as physical properties such as acidity, alkalinity, and viscosity of the waste liquid, can indirectly affect the acquisition quality of energy spectrum data, further increasing the complexity of data preprocessing and potentially reducing the reliability of data processing results. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for multi-source data fusion in the process of radionuclide separation, thereby improving the reliability of data in the process of radionuclide separation.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for multi-source data fusion during the separation of radionuclides, the method comprising:

[0007] Step 1: Collect multi-source monitoring data generated during the separation of radionuclides. The multi-source monitoring data includes gamma-ray energy spectrum data, process operating condition parameter data, and waste liquid physical property parameter data.

[0008] Step 2: Convert multi-source monitoring data into feature element clusters, construct a configuration field based on the feature element clusters, and determine the core identification window in the configuration field; within the core identification window, perform extreme value search based on the geometric distribution boundary of the feature element clusters to determine the support boundary that can completely enclose the feature element clusters and has the minimum external geometric shape.

[0009] Step 3: Divide the core identification window into multiple element interpretation blocks using the support boundary as the segmentation benchmark, and project the feature element clusters into the corresponding element interpretation blocks; Based on the statistical characteristics of the feature element points within the element interpretation blocks, and combined with the relationship between the process condition parameter data and waste liquid physical property parameter data corresponding to the element interpretation blocks, generate dynamic kernel calibration units; Use dynamic kernel calibration units to correct the multi-source monitoring data to obtain the fused data base;

[0010] Step 4: Extract features from the fused data base to extract feature information related to the type and concentration of radionuclides and generate a multi-dimensional feature vector;

[0011] Step 5: Input the multidimensional feature vector into the pre-trained pattern recognition model for classification and fusion recognition to obtain the identification results of radionuclides and their concentration prediction values.

[0012] Secondly, a multi-source data fusion system for the separation of radionuclides includes:

[0013] The acquisition module is used to collect multi-source monitoring data generated during the separation of radionuclides. The multi-source monitoring data includes gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data.

[0014] The identification module is used to convert multi-source monitoring data into feature element clusters, construct a configuration field based on the feature element clusters, and determine the core identification window on the configuration field. Within the core identification window, extreme value search is performed based on the geometric distribution boundary of the feature element clusters to determine the support boundary that can completely enclose the feature element clusters and has the minimum external geometry.

[0015] The correction module is used to divide the core identification window into multiple feature interpretation blocks based on the support boundary, and project the feature feature clusters into the corresponding feature interpretation blocks; based on the statistical characteristics of the feature feature points inside the feature interpretation blocks, and combined with the relationship between the process condition parameter data and waste liquid physical property parameter data corresponding to the feature interpretation blocks, a dynamic kernel calibration unit cell is generated; the dynamic kernel calibration unit cell is used to correct the multi-source monitoring data to obtain the fused data base;

[0016] The extraction module is used to extract features from the fused data base, extract feature information related to the type and concentration of radionuclides, and generate multi-dimensional feature vectors.

[0017] The output module is used to input multidimensional feature vectors into a pre-trained pattern recognition model for classification and fusion recognition, so as to obtain the identification results of radionuclides and their concentration prediction values.

[0018] Thirdly, a computing device includes:

[0019] One or more processors;

[0020] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0021] The above-described solution of the present invention has at least the following beneficial effects:

[0022] This invention deeply correlates feature element clusters with process operating parameters and waste liquid physical property parameters by dividing the element interpretation blocks and generating dynamic nucleolysis calibration units. Utilizing the dynamic changes in these parameters, it specifically corrects multi-source monitoring data, forming a fused data base. This solves the problem of disconnected multi-source data and the inability to achieve synergistic effects through simple superposition. Furthermore, by extracting features from the fused data base and classifying and fusion-identifying the pattern recognition model, it fully explores the inherent correlations between various source data, improving the accuracy of nuclide identification and concentration prediction, effectively addressing the issues of unstable nuclide identification and large concentration detection deviations under complex operating conditions. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the multi-source data fusion method during the separation of radionuclides provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of a multi-source data fusion system during the separation of radionuclides provided in an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] like Figure 1 As shown, embodiments of the present invention propose a multi-source data fusion method during the separation of radionuclides, the method comprising the following steps:

[0027] Step 1: Collect multi-source monitoring data generated during the separation of radionuclides. The multi-source monitoring data includes gamma-ray energy spectrum data, process operating condition parameter data, and waste liquid physical property parameter data.

[0028] Step 2: Convert multi-source monitoring data into feature element clusters, construct a configuration field based on the feature element clusters, and determine the core identification window in the configuration field; within the core identification window, perform extreme value search based on the geometric distribution boundary of the feature element clusters to determine the support boundary that can completely enclose the feature element clusters and has the minimum external geometric shape.

[0029] Step 3: Divide the core identification window into multiple element interpretation blocks using the support boundary as the segmentation benchmark, and project the feature element clusters into the corresponding element interpretation blocks; Based on the statistical characteristics of the feature element points within the element interpretation blocks, and combined with the relationship between the process condition parameter data and waste liquid physical property parameter data corresponding to the element interpretation blocks, generate dynamic kernel calibration units; Use dynamic kernel calibration units to correct the multi-source monitoring data to obtain the fused data base;

[0030] Step 4: Extract features from the fused data base to extract feature information related to the type and concentration of radionuclides and generate a multi-dimensional feature vector;

[0031] Step 5: Input the multidimensional feature vector into the pre-trained pattern recognition model for classification and fusion recognition to obtain the identification results of radionuclides and their concentration prediction values.

[0032] In this embodiment of the invention, the invention deeply correlates feature element clusters with process condition parameters and waste liquid physical property parameters by dividing the element interpretation blocks and generating dynamic nucleolysis calibration units. By utilizing the dynamic change patterns of the condition parameters and physical property parameters, the invention makes targeted corrections to multi-source monitoring data to form a fused data base. This solves the problem of disconnected multi-source data and the inability to achieve synergistic effects through simple superposition. By extracting features from the fused data base and classifying and fusion-identifying the pattern recognition model, the invention can fully explore the inherent correlation between the data from each source, improve the accuracy of nuclide identification, and increase the accuracy of concentration prediction values. This effectively solves the problems of unstable nuclide identification and large concentration detection deviations under complex operating conditions.

[0033] In a preferred embodiment of the present invention, step 1 involves collecting multi-source monitoring data generated during the separation of radionuclides. The multi-source monitoring data includes gamma-ray energy spectrum data, process operating parameter data, and waste liquid physical property parameter data, and may include:

[0034] Step 101: Collect radiation signals from radionuclides in the separation process flow channel. After analyzing and processing the radionuclides' radiation signals, generate gamma-ray energy spectrum data. Specifically, this includes: deploying a radiation signal acquisition sensor on the inner wall of the flow channel in the radionuclide separation process; the sensor continuously collects radiation signals released by the radionuclides within the flow channel; maintaining a stable fit between the sensor and the inner wall of the flow channel during the acquisition process, with the fit gap controlled between 0.1 and 0.3 mm; the collected radiation signals contain ray count data in different energy ranges, ranging from 0.1 to 10 MeV; and then analyzing and processing the radiation signals. Specifically, Fourier transform is used to process the acquired raw radiation signal. First, the radiation signal in the time domain is converted into a frequency domain signal to clarify the distribution of different frequency components in the signal. The frequency range corresponding to high-frequency noise is screened out. The high-frequency noise frequency is set to 500 to 1000 Hz. Then, the signal components in this high-frequency range are removed by filtering, and the effective signals related to the radiation of radionuclides are retained. After noise removal, the denoised signals are classified and screened according to the known characteristic radiation energy range of radionuclides. Only the signal data that conforms to the radiation energy characteristics of the nuclide are retained, and finally, gamma-ray energy spectrum data is generated.

[0035] Step 102: Based on the gamma-ray energy dispersive spectroscopy (EDS) data, synchronously acquire the original process parameters during the separation process. Perform feature-weighted filtering on the original process parameters according to the GAS data to generate process condition parameter data. This process condition parameter data includes temperature data, pressure data, flow rate data, and liquid level data. Specifically, this involves deploying multiple sets of original process parameter acquisition devices at key nodes in the separation process flow channel. These key nodes include the inlet, middle, and outlet ends of the flow channel. Temperature data is acquired using a PT100 platinum resistance temperature sensor, with a range of 0 to 100℃ and an accuracy of ±0.1℃. Pressure data is acquired using a diffused silicon pressure transmitter, with a range of 0 to 1.6 MPa and an accuracy of ±0.5%FS. Flow rate data is acquired using an electromagnetic flow meter, with a range of 0.1 to 10 m³ / h and an accuracy of ±0.2%FS. Liquid level data is acquired using an immersion-type... The liquid level transmitter collects data with a range of 0 to 5 m and an accuracy of ±0.1%FS. All acquisition devices synchronously and in real time acquire the original process parameters during the separation process. After acquisition, the original process parameters are processed by feature weighting and screening in conjunction with gamma-ray energy spectrum data. Specifically, the count fluctuation amplitude of each energy range in the gamma-ray energy spectrum data is analyzed first, and energy ranges with significant count fluctuations and related to the nuclide separation reaction are screened out. Correspondingly, process parameter ranges with synchronous fluctuations and high correlation with the energy range are determined. Then, weight coefficients are assigned to the original process parameters in the corresponding ranges. The weight coefficients range from 0.3 to 0.9. The allocation of weight coefficients is determined based on the ratio of parameter fluctuation amplitude to energy spectrum fluctuation amplitude. The larger the ratio, the higher the correlation, and the larger the weight coefficient. Finally, redundant parameter data with smooth fluctuations and low correlation with the nuclide separation state are screened out to generate process condition parameter data.

[0036] Step 103: Based on the process operating condition parameter data, synchronously collect the original detection values ​​of waste liquid samples during the separation process, and calibrate the original detection values ​​of waste liquid samples according to the process operating condition parameter data to generate waste liquid physical property parameter data. The waste liquid physical property parameter data includes pH data, conductivity data, and suspended solids concentration data. Specifically, this includes: deploying a waste liquid sample detection device at the waste liquid collection end of the separation process, wherein pH data is collected using a glass electrode pH meter, with a collection range of 0 to 14 pH and an accuracy of ±0.01 pH; conductivity data is collected using a conductivity electrode sensor, with a collection range of 0 to 10000 μS / cm and an accuracy of ±1% FS; suspended solids concentration data is collected using a laser turbidimeter, with a collection range of 0 to 1000 NTU and an accuracy of ±2% FS. The device synchronously collects the original detection values ​​of waste liquid samples during the separation process; strict control measures are implemented during the collection process. The collection time of the waste liquid samples is synchronized with the collection time of the process condition parameter data, with a time deviation of no more than 100 milliseconds. After the collection is completed, the original detection values ​​of the waste liquid samples are calibrated according to the process condition parameter data. Specifically, standard values ​​of acidity, alkalinity, conductivity, and suspended solids concentration of the waste liquid are statistically analyzed under different temperature and pressure conditions. The temperature statistical range is 0 to 100℃, and the pressure statistical range is 0 to 1.6MPa. The influence of temperature and pressure changes on the physical properties of the three types of waste liquids is clarified. Then, combined with the real-time temperature and pressure data collected in step 102, the original detection values ​​of the waste liquid samples are corrected according to the influence law. For example, for every 10℃ increase in temperature, the baseline for measuring the conductivity of the waste liquid decreases by 3% to 5%. The baseline for the original detection value of conductivity needs to be adjusted according to the corresponding influence law based on the collected real-time temperature data to complete the calibration and finally generate the physical property parameter data of the waste liquid.

[0037] Step 104 involves performing spatiotemporal registration processing on the gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data according to a unified time benchmark to generate multi-source monitoring data. Specifically, this includes: acquiring the gamma-ray energy spectrum data generated in Step 101, the process condition parameter data generated in Step 102, and the waste liquid physical property parameter data generated in Step 103; uniformly setting millisecond-level timestamps for the three types of data with a timestamp accuracy of 1 millisecond, matching the acquisition frequency of various acquisition devices. The acquisition frequency of devices such as radiation signal acquisition sensors and PT100 platinum resistance temperature sensors is set to once per second, and the timestamps synchronously correspond to the acquisition time with a time deviation of no more than 100 milliseconds; performing spatiotemporal registration processing according to a unified time benchmark, specifically by comparing the timestamps of various types of data one by one, and correlating the gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data at the same millisecond time node to form multi-source monitoring data with temporal continuity.

[0038] This embodiment constructs a complete multi-source monitoring data system by synchronously acquiring and spatiotemporally registering gamma-ray energy spectrum data, process operating condition data, and waste liquid physical property data, thus solving the problem that single energy spectrum data acquisition is easily affected by operating conditions and physical property parameters.

[0039] In a preferred embodiment of the present invention, step 2 involves converting multi-source monitoring data into feature element clusters, constructing a configuration field based on the feature element clusters, and determining a core identification window within the configuration field. Within the core identification window, an extreme value search is performed based on the geometric distribution boundary of the feature element clusters to determine a support boundary that can completely enclose the feature element clusters and has the minimum bounding geometry. This step may include:

[0040] Step 201: Aggregate and classify the data points in the multi-source monitoring data based on the feature similarity to generate a feature element cluster; extract the geometric distribution boundary based on the coordinate distribution of each data point in the feature element cluster, construct a closed spatial morphology based on the geometric distribution boundary, and generate a field domain for the closed spatial morphology to obtain the configuration field. Specifically, this includes: acquiring multi-source monitoring data (acquiring the gamma-ray energy spectrum data generated in step 101, the process condition parameter data generated in step 102, and the waste liquid physical property parameter data generated in step 103, setting millisecond-level timestamps for the three types of data with a timestamp accuracy of 1 millisecond, matching the acquisition frequency of various acquisition devices, including radiation signal acquisition sensors, PT100 platinum resistance temperature sensors, etc.). The sampling frequency of sensors and other equipment is set to once per second, with timestamps synchronously corresponding to the sampling time, and a time deviation not exceeding 100 milliseconds. Spatiotemporal registration is performed according to a unified time benchmark. Specifically, using timestamps as the core, the timestamps of various data types are compared one by one, and gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data at the same millisecond time node are correlated one-to-one, integrating them into multi-source monitoring data with temporal continuity. Each data point in the multi-source monitoring data is extracted, and each data point contains attribute information in three dimensions: gamma-ray energy spectrum characteristic value, process condition parameter value, and waste liquid physical property parameter value. The feature similarity between each data point in the multi-source monitoring data is calculated using the following formula: (in Let be the similarity between data point i and data point j. Let i be the attribute value of data point i in the k-th dimension. (where the attribute value of data point j in the k-th dimension is a factor), the smaller the numerical difference, the better. The closer the similarity is to 1, the higher the similarity; set a similarity threshold. =0.75, Data points are grouped into the same category, forming multiple feature element clusters. After aggregation and classification, the coordinate distribution of each data point in each feature element cluster is extracted. The coordinate distribution is constructed in three dimensions using the values ​​of the three dimensions as coordinate axes. Then, the geometric distribution boundary is extracted based on the coordinate distribution. The geometric distribution boundary is the closed contour formed by connecting the outermost data points of the feature element cluster in three-dimensional space. Subsequently, a closed spatial form is constructed based on the geometric distribution boundary. The closed spatial form is enclosed by the geometric distribution boundary. Finally, the closed spatial form is used to generate a field, resulting in a configuration field.

[0041] Step 202: Perform density distribution scanning on the configuration field to obtain density distribution data of the feature element clusters. Perform extreme point detection and gradient analysis on the density distribution data to generate density distribution extreme points and gradient change characteristics. Delineate the core analysis region based on the density distribution extreme points and gradient change characteristics, and define the core analysis region using a window to generate a core identification window. Specifically, this includes dividing the configuration field into several uniformly sized three-dimensional sub-regions, with sub-region side lengths... Given ∈[0.5,2] data feature units, count the number of feature points n contained in each sub-region, and calculate the feature point density of each sub-region. (in (For the sub-region volume); then construct multiple nested polygons. The first layer of polygons covers the sub-region with the highest ρ value, the second layer covers the sub-region with the highest ρ value, and the third layer covers the sub-region with the highest ρ value. The vertices of each layer of polygons are derived from the geometric center of the corresponding density interval sub-region. The geometric center is calculated by connecting the parts. , , ( (Given the coordinates of the m-th feature point within the sub-region), calculate the area-to-perimeter ratio of each layer of polygons. (S is the area of ​​the polygon, L is the perimeter), number of vertices By analyzing feature parameters, the density distribution data of the feature element clusters and the polygon nesting hierarchy features are finally obtained. Extreme point detection and gradient analysis are then performed on the density distribution data, with the extreme points being... For the corresponding sub-regions, gradient analysis is used to calculate the density change rate between adjacent sub-regions. , The distance between the centers of subregions i and j is . ,in Let x be the x-coordinate of the geometric center of the i-th sub-region. Let x be the x-coordinate of the geometric center of the j-th sub-region. Let y be the y-coordinate of the geometric center of the i-th sub-region. Let y be the geometric center of the j-th sub-region. Let z be the z-coordinate of the geometric center of the i-th sub-region. Let z be the geometric center of the j-th sub-region. Determine the density gradient direction based on G, and then filter based on the polygon nesting hierarchy features. The nested layers generate density distribution extreme points and gradient change characteristics; based on the density distribution extreme points and gradient change characteristics, the core analysis region is delineated. The core analysis region is the spatial range centered on the density extreme points and covering the nested layers with significant gradient changes. Then, the core analysis region is windowed to generate the core identification window.

[0042] Step 203: Taking the outermost distribution point of the feature element cluster within the core identification window as the search starting point, perform successive approximation searches along each dimension to obtain the boundary point set. Specifically, this includes: determining the outermost distribution point of the feature element cluster within the core identification window, where the outermost distribution point is the point corresponding to the coordinate extreme value in each dimension. ,in The maximum coordinate value of the feature cluster in the x-axis. The minimum coordinate value in the x-dimension. , The maximum / minimum coordinate values ​​in the y-dimension. , (These are the maximum / minimum coordinates in the z-axis), and these points are selected as the starting point for the search to determine the search step size. (where α is the step size coefficient, with a value range of [0.01, 0.05];) The core identification window corresponds to a certain dimension length (i.e., the side length of the core identification window in the x, y, and z dimensions). The search proceeds sequentially along each dimension of the core identification window: starting from the starting point, each advancement... Record the data point position (x, y, z), and simultaneously verify whether the position is within a valid nested layer (i.e., the sub-region where the position is located). ,in The feature point density of this sub-region. (The maximum density value in all sub-regions) until the boundary data points of the feature feature cluster are found, and all boundary points are aggregated to form a boundary point set.

[0043] Step 204a involves calculating the connection order between adjacent boundary points based on their spatial coordinates, and then connecting the boundary points sequentially to generate an initial envelope contour. This process includes: acquiring a set of boundary points and extracting the specific coordinates of each boundary point in the three dimensions of x, y, and z in 3D space; calculating the connection order of these boundary points by first calculating the Euclidean distance from each boundary point to the density extremum point and sorting all boundary points from near to far according to this distance; then, using the density extremum point as the center, calculating the polar angle (ranging from 0 to 360 degrees) corresponding to each boundary point, and rearranging adjacent boundary points in a clockwise order according to the polar angle to ensure that a continuous contour trajectory is formed after the arrangement, and that the contour conforms to the shape of the outer boundary of the nested polygon; finally, connecting all boundary points in this order to generate the initial envelope contour.

[0044] Step 204b involves detecting the initial envelope contour line and connecting the endpoints based on the detected opening positions of the contour line. The resulting closed contour line serves as the boundary of the envelope to be filled. Specifically, this includes: checking whether the initial envelope contour line has an opening: calculating the sum of the absolute values ​​of the coordinate differences between the first and last endpoints of the contour line in the x, y, and z dimensions. If this value is greater than 0.1 times the search step size, it is determined that an opening exists. If an opening is detected, the two endpoints of the opening are identified, and the straight-line distance between these two endpoints is calculated (this distance is required to be no more than 0.1 times the length of the corresponding dimension of the core recognition window). The two endpoints are connected along this shortest straight-line path to form a closed contour line, which is the boundary of the envelope to be filled.

[0045] Step 204c: The internal region enclosed by the boundary to be filled is subdivided to obtain grid cells; the grid cells are filled to generate the spatial boundary to be confirmed, specifically including: setting three levels of grid cells based on multi-layer nested polygons, the grid cells in the first layer of nested polygons have a side length of 0.1 to 0.3 data feature units (data feature units are the basic units that characterize the dimensions of multi-source monitoring data), the second layer has a side length of 0.3 to 0.5 data feature units, and the third layer has a side length of 0.5 to 0.8 data feature units; each grid cell is checked one by one to determine whether it contains feature point (i.e., the coordinates of feature points fall within the coordinate range of the grid cell), the grid cells containing feature points are marked as valid grid cells, and the grid cells not containing feature points are marked as invalid grid cells, and finally the spatial boundary to be confirmed is generated.

[0046] Step 204d involves extracting the circumscribed morphological parameters of each boundary dimension of the spatial boundary to be confirmed. An initial shrinkage step size is set based on these parameters, and the boundary dimensions are shrunk according to this step size to generate a preliminary shrunk boundary. Specifically, this includes extracting the length, width, and height of the spatial boundary in the x-axis, y-axis, and z-axis, and setting the initial shrinkage step size: the initial shrinkage step size is the boundary's corresponding dimension parameter (length / width / height) multiplied by a shrinkage coefficient. The shrinkage coefficient ranges from 0.005 to 0.02, with a smaller coefficient for higher nesting levels (higher density). Specifically, the coefficient for the first-level nested polygon is 0.02, for the second level it is 0.01, and for the third level it is 0.005. The boundary is then shrunk inwards along the x, y, and z dimensions by the initial shrinkage step size. Simultaneously, the shrinkage direction is adjusted by referencing the contours of the inner nested polygons to ensure that the distance between the shrunk boundary and the vertices of the inner nested polygons does not exceed 0.1 times the initial shrinkage step size. After shrinkage, a preliminary shrunk boundary is generated.

[0047] Step 204e: Perform integrity verification on the initial contraction boundary to obtain the verification result; generate a contraction step size correction amount based on the verification result; adjust the initial contraction step size based on the correction amount to generate a corrected contraction step size; continue dimensional contraction on the initial contraction boundary based on the corrected contraction step size to generate a secondary contraction boundary. Specifically, this includes: checking the coordinates of each feature point one by one to determine whether they fall within the coordinate range of the contracted boundary (i.e., the coordinate values ​​of the feature points in the x, y, and z dimensions all satisfy the condition that the minimum value of the contraction boundary in that dimension is ≤ the feature point coordinate value is ≤ the maximum value of the contraction boundary in that dimension). If all feature points satisfy this condition, the verification passes and the boundary exists. If any feature point does not meet the requirements, the verification fails. Based on the verification results, the shrinkage step size is adjusted, and the shrinkage coefficient correction is calculated. The shrinkage coefficient correction is equal to the initial shrinkage coefficient multiplied by the correction coefficient, where the correction coefficient ranges from 0.1 to 0.3. When the verification fails, the correction coefficient is 0.3 (positive value), and when the verification passes, the correction coefficient is -0.1 (negative value). The adjusted shrinkage step size is then calculated: the adjusted shrinkage step size is equal to the initial shrinkage step size plus (the shrinkage coefficient correction multiplied by the boundary dimension parameter). The initial shrinkage boundary is then shrunk again according to the adjusted shrinkage step size, conforming to the contour shape of the inner nested polygons during the shrinkage process, ultimately generating a secondary shrinkage boundary.

[0048] Step 204f involves repeatedly performing integrity verification and boundary dimension shrinkage on the secondary shrinking boundary until the shrunken boundary exactly encloses all feature element clusters and further shrinkage would not be able to completely enclose them. The iteration stops when this happens, and the supporting boundary is obtained. Specifically, this includes repeating steps 204d to 204e on the secondary shrinking boundary. After each shrinkage, the fit between the shrinking boundary and the nested layer is calculated (fit is the number of vertices that overlap between the shrinking boundary and the nested layer divided by the total number of vertices in the nested layer, and the fit is required to be no less than 90%). If the fit meets the standard, the shrinkage continues; otherwise, it stops. This process is repeated until the boundary can just completely enclose all feature element points and further shrinkage would not be able to completely enclose them. The final boundary obtained is the supporting boundary.

[0049] This embodiment constructs a configuration field and determines the minimum circumscribed support boundary, thereby achieving precise definition of the effective range of the feature element cluster. Through density scanning, extreme value search, and iterative shrinkage optimization, it ensures that the support boundary can completely enclose the feature element cluster and eliminate redundant and invalid data, thus improving the accuracy and rationality of the feature region definition.

[0050] In a preferred embodiment of the present invention, step 3 involves dividing the core identification window into multiple element interpretation blocks using the support boundary as the segmentation criterion, and projecting the feature element clusters into the corresponding element interpretation blocks; based on the statistical characteristics of the feature element points within the element interpretation blocks, and combined with the relationship between the process condition parameter data and waste liquid physical property parameter data corresponding to the element interpretation blocks, generating dynamic kernel calibration units; and using the dynamic kernel calibration units to correct the multi-source monitoring data to obtain the fused data base, which may include:

[0051] Step 301: Using the support boundary as the segmentation benchmark, the core identification window is spatially divided. Based on the geometric orientation of the support boundary, the core identification window is divided into multiple initial sub-regions to obtain candidate interpretation blocks. Specifically, this includes: acquiring all geometric feature information of the support boundary, including the extension direction, inflection point coordinates, segment length, and curvature of each line segment; analyzing the spatial distribution complexity of the feature element clusters, specifically by statistically analyzing the distribution range of feature element points in the three-dimensional space of the core identification window and the average distance between points. If the distribution range of feature element points covers more than 80% of the core identification window and the average distance is less than 0.3 data feature units, it is considered high complexity; if the distribution range covers 40% to... If 80% of the area is covered and the average spacing is between 0.3 and 0.6 data feature units, it is classified as medium complexity. If the distribution range covers less than 40% of the core identification window and the average spacing is greater than 0.6 data feature units, it is classified as low complexity. Based on the geometric orientation of the support boundary and the complexity of the feature element cluster, the number of initial segmented sub-regions is determined. High complexity corresponds to no less than 8 initial segmented sub-regions, medium complexity corresponds to 4 to 7, and low complexity corresponds to 2 to 3. Finally, according to the determined number of sub-regions, the core identification window is divided into multiple continuous and non-overlapping initial segmented sub-regions along the geometric orientation of the support boundary. Each initial segmented sub-region is enclosed by continuous line segments of the support boundary to form a closed spatial region, thereby obtaining candidate interpretation blocks.

[0052] Step 302 involves detecting overlapping regions between adjacent candidate decoded blocks and adjusting their boundaries to obtain preliminary decoded blocks. Specifically, this includes: extracting the boundary coordinates of all adjacent candidate decoded blocks; comparing the boundary segments of adjacent blocks segment by segment to detect overlapping regions; and calculating the overlap length and intersection area of ​​the boundary segments of adjacent blocks. If the overlap length of the boundary segments is greater than 0.5 data feature units, or the area of ​​the intersection region is greater than the square of 0.2 data feature units, then an overlapping region is determined to exist. For the detected overlapping regions... The adjustment direction is determined to be perpendicular to the overlapping boundary line segment. The adjustment distance is determined according to the degree of overlap. For every 0.1 data feature units exceeding the overlap length, the adjustment distance increases by 0.1 data feature units, but the maximum adjustment distance does not exceed 1 data feature unit. According to the determined adjustment direction and adjustment distance, the boundaries of adjacent candidate interpretation blocks are fine-tuned segment by segment. During the fine-tuning process, the spacing between adjacent block boundaries is checked in real time to ensure that the boundaries of adjacent blocks do not overlap after adjustment and the spacing does not exceed 0.1 data feature units. Finally, after the adjustment of all overlapping areas is completed, a preliminary interpretation block with clear boundaries, no overlap and tight connection is obtained.

[0053] Step 303: The shape of the preliminary interpretation block is normalized, and the boundary of the preliminary interpretation block is fine-tuned to obtain the feature interpretation block. Spatial coordinates are extracted for each feature point in the feature feature cluster, and the spatial positional relationship between the feature point and each feature interpretation block is calculated. Based on the spatial positional relationship, the relationship data between the feature point and the feature interpretation block is determined. Specifically, this includes: normalizing the shape of the preliminary interpretation block; based on the three-dimensional coordinates of all feature points within the preliminary interpretation block, extracting the maximum and minimum coordinate values ​​for each dimension; and using these maximum and minimum coordinate values ​​as a reference, extending outwards... The boundary range of the normalized block is extended by 0.2 data feature units, and the initial interpretation block is adjusted into a regular geometric shape of cube or cuboid. Then, the boundary of the normalized block is fine-tuned, with the adjustment range controlled within 0.1 data feature units. During the fine-tuning process, it is verified one by one whether all feature element points in the block are enclosed, to ensure that the adjusted block boundary can completely cover the corresponding feature element points, and finally obtain the feature interpretation block with regular shape and clear boundary coordinates. Then, the absolute coordinate value of each feature element point in the feature element cluster in the three-dimensional coordinate system of the core identification window is extracted, and the specific values ​​of each feature element point in the x, y, and z dimensions are recorded.

[0054] Each feature point's 3D coordinates are compared with the boundary coordinate range of each feature interpretation block to determine if the feature point's coordinates are within the feature interpretation block's coordinate range. If the feature point's x, y, and z coordinates are all between the minimum and maximum boundary coordinates of that dimension of the corresponding feature interpretation block, then the feature point is determined to be inside the feature interpretation block. If the feature point's coordinates in one dimension are equal to the boundary coordinates of that dimension of the feature interpretation block, then the feature point is determined to be on the boundary of the feature interpretation block. If the feature point's coordinates in any dimension exceed the boundary coordinate range of that dimension of the feature interpretation block, then the feature point is determined to be outside the feature interpretation block. This process obtains the spatial relationship between the feature points and each feature interpretation block. Finally, based on the spatial relationship, the unique feature interpretation block corresponding to each feature point is determined, forming the relationship data between feature points and feature interpretation blocks.

[0055] Step 304: Based on the relationship data between feature points and feature interpretation blocks, project the feature points to the corresponding feature interpretation blocks. Specifically, this includes: retrieving the relationship data between feature points and feature interpretation blocks determined in step 303, and matching feature points with their corresponding feature interpretation blocks one by one according to this data; then, for feature points determined to be located inside or on the boundary of a feature interpretation block, directly assign them to the corresponding feature interpretation block; for feature points determined to be located outside a feature interpretation block, calculate the Euclidean distance from the feature point to the boundary of each feature interpretation block, that is, calculate the three-dimensional straight-line distance between the feature point and all points on the boundary of each feature interpretation block, and take the minimum distance value as the distance from the feature point to the corresponding feature interpretation block, and assign the feature point to the nearest feature interpretation block; finally, after completing the assignment matching of all feature points, associate and store the coordinate information of each feature point with the corresponding feature interpretation block to realize the projection of feature points to corresponding feature interpretation blocks.

[0056] Step 305: Statistically analyze the spatial distribution density of feature points within each feature interpretation block to generate density distribution data. Based on the density distribution data, calculate the geometric mean position of all feature points and determine the geometric mean position as the initial calibration base point for the feature interpretation block. Specifically, this includes: counting the number of projected feature points within each feature interpretation block and recording the total number of feature points in each block; then calculating the volume of each feature interpretation block. If the feature interpretation block is a cube, multiply the block's side length by the side length and then by the side length again to obtain the volume; if it is a cuboid, multiply the block's length by the width and then by the height to obtain the volume; and finally, using the feature points within each feature interpretation block... Dividing the number of feature points by the volume of the block yields the spatial distribution density of feature points within each block. The spatial distribution density data of all blocks are then aggregated to generate density distribution data. Based on this density distribution data, the geometric mean position of all feature points within each feature interpretation block is calculated. Specifically, the coordinate values ​​of all feature points in the x-axis are summed, and then divided by the total number of feature points to obtain the average coordinate value in the x-axis. Similarly, the average coordinate values ​​in the y-axis and z-axis are calculated. The average coordinate values ​​in these three dimensions together constitute the geometric mean position of the feature points within the block. Finally, the calculated geometric mean position is determined as the initial calibration base point for the feature interpretation block.

[0057] Step 306: Using the initial calibration base point as the center, analyze the spatial dispersion of all feature points relative to the initial calibration base point, and calculate the dispersion feature value of the feature interpretation block; determine the dynamic influence radius based on the dispersion feature value, and construct the initial kernel calibration space with the initial calibration base point as the core and the dynamic influence radius. Specifically, this includes: using the initial calibration base point determined in step 305 as the center, calculating the three-dimensional spatial straight-line distance between each feature point and the initial calibration base point, that is, calculating the difference between the coordinate values ​​of each feature point in the x, y, and z dimensions and the corresponding coordinate values ​​of the initial calibration base point, squaring the difference in each dimension, adding them together, and then taking the square root of the sum to obtain the distance between the feature point and the initial calibration base point; summing all the distance values ​​between the feature points and the initial calibration base point, and calculating the square root of these distance values. The mean is calculated by summing all distance values ​​and dividing by the total number of feature points. The difference between each distance value and this mean is calculated, the squares of each difference are summed, and the sum is divided by the total number of feature points. The square root of the sum is then taken to obtain the dispersion characteristic value of the feature block. The larger the dispersion characteristic value, the more dispersed the spatial distribution of feature points within the block. The dynamic influence radius is then determined based on the dispersion characteristic value. The scaling factor is set to a range of 1.2 to 1.5. If the dispersion characteristic value is greater than 0.5, the scaling factor is 1.5; if the dispersion characteristic value is not greater than 0.5, the scaling factor is 1.2. The dynamic influence radius is obtained by multiplying the dispersion characteristic value by the selected scaling factor. Finally, a spherical initial kernel calibration space is constructed in three-dimensional space, with the initial calibration base point as the center and the calculated dynamic influence radius as the radius.

[0058] Step 307: Retrieve pre-stored process condition parameter data and waste liquid physical property parameter data that have a spatial mapping relationship with the element interpretation block; analyze both to calculate the coupling coefficient between the condition and the property; adjust the boundary morphology of the initial kernel calibration space using the coupling coefficient to generate a dynamic kernel calibration unit cell; match and correct each data point in the multi-source monitoring data with the dynamic kernel calibration unit cell to generate a fused data base. Specifically, this includes: retrieving pre-stored process condition parameter data and waste liquid physical property parameter data, where the spatial mapping relationship between these data and the element interpretation block is that the spatial location of the element interpretation block corresponds to the installation location of the process condition parameter and waste liquid physical property parameter acquisition sensors; the element interpretation block's... The time dimension corresponds to the acquisition time of the parameters, and the time deviation does not exceed 100 milliseconds. Next, the retrieved process condition parameter data and waste liquid physical property parameter data are analyzed to calculate the coupling coefficient between the condition and the property. Specifically, the fluctuation range of the process condition parameter data and the fluctuation range of the waste liquid physical property parameter data are calculated separately. The fluctuation range of the process condition parameter is the maximum value minus the minimum value of the parameter in the corresponding time period, and the fluctuation range of the waste liquid physical property parameter is calculated in the same way. The correlation between the fluctuation ranges of the two is analyzed. By calculating the covariance of the fluctuation ranges of the two ÷ the standard deviation of the fluctuation range of the process condition parameter × the standard deviation of the fluctuation range of the waste liquid physical property parameter, the Pearson correlation coefficient is obtained. This coefficient is the coupling coefficient between the condition and the property, and its value ranges from -1 to 1.

[0059] The boundary morphology of the initial kernel calibration space is adjusted using the coupling coefficient between operating conditions and physical properties. Specifically, the dynamic influence radius of the initial kernel calibration space is multiplied by (1 + the absolute value of the coupling coefficient between operating conditions and physical properties) to obtain the adjusted dynamic influence radius. Using the initial calibration base point as the center and the adjusted dynamic influence radius as the radius, the spatial boundary is redefined, generating a dynamic kernel calibration unit cell. The morphology of this unit cell dynamically changes with the coupling relationship between process conditions and waste liquid properties. Multi-source monitoring data, after spatiotemporal registration processing, is acquired. The acquisition and registration process involves acquiring the gamma-ray energy spectrum data generated in step 101, the process condition parameter data generated in step 102, and the waste liquid property parameter data generated in step 103. Millisecond-level timestamps are uniformly set for these three types of data, with a timestamp accuracy of 1 millisecond, matching the acquisition frequency of various acquisition devices. Specifically, the acquisition frequency of devices such as the radiation signal acquisition sensor and the PT100 platinum resistance temperature sensor is set to once per second, with the timestamp synchronously corresponding to the acquisition time, and the time deviation not exceeding 100 milliseconds. Subsequently, data is collected according to a unified time... The benchmark undergoes spatiotemporal registration processing. Specifically, it uses timestamps as the core, comparing the timestamps of various data types one by one. Gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data at the same millisecond time node are correlated one-to-one, integrating them into multi-source monitoring data with temporal continuity. Next, each data point in the multi-source monitoring data is matched with a dynamic nucleus calibration cell to determine whether the three-dimensional coordinates of each data point are within the spatial range of the dynamic nucleus calibration cell, i.e., whether the three-dimensional straight-line distance from the data point to the initial calibration base point does not exceed the adjusted dynamic influence radius. For data points within the dynamic nucleus calibration cell range, numerical correction is performed, specifically by multiplying the original data point value by [1 - (distance from the data point to the initial calibration base point ÷ adjusted dynamic influence radius) × 0.1] to obtain the corrected value. For data points not within the dynamic nucleus calibration cell range, their original values ​​remain unchanged. Finally, all corrected and uncorrected data points are integrated and arranged in timestamp order to generate a fused data base.

[0060] In this embodiment, by rationally dividing the element interpretation blocks and accurately projecting the feature element points, the initial calibration base point and dynamic influence radius are generated based on the statistical characteristics of the feature points within the blocks. Combined with the coupling coefficient of operating conditions and physical properties, the morphology of the calibration unit cell is dynamically adjusted, thereby realizing adaptive correction of multi-source monitoring data and effectively suppressing the interference caused by operating condition fluctuations, physical property changes and detector drift in the separation process.

[0061] In a preferred embodiment of the present invention, step 4, which involves feature extraction from the fused data base to extract feature information related to the type and concentration of radionuclides and generate a multidimensional feature vector, may include:

[0062] Step 401 involves separating the data sources from the fused data base, extracting the gamma-ray energy spectrum data sequence, process condition parameter time series, and waste liquid physical property parameter dataset. Specifically, this includes acquiring the fused data base, which is a multi-source monitoring data with temporal continuity formed after dynamic kernel calibration and unit cell correction. It includes the gamma-ray energy spectrum data generated in step 101, the process condition parameter data generated in step 102, and the waste liquid physical property parameter data generated in step 103. All three types of data have been uniformly timestamped at the millisecond level with a timestamp accuracy of 1 millisecond, consistent with the acquisition frequency of once per second by devices such as radiation signal acquisition sensors and PT100 platinum resistance temperature sensors. The data is matched in terms of rate, with the deviation between the timestamp and the acquisition time not exceeding 100 milliseconds. The three types of data at the same millisecond time node are correlated one by one. The data source of the fused data base is separated. According to the data type and the object of representation, all data belonging to gamma-ray energy spectrum in the fused data base are selected and arranged in order of timestamp to form a gamma-ray energy spectrum data sequence. All data belonging to process condition parameters are selected and arranged in order of timestamp to form a process condition parameter time series, which includes four types of parameter data: temperature, pressure, flow rate, and liquid level. All data belonging to waste liquid physical property parameters are selected and classified according to data type to form a waste liquid physical property parameter dataset.

[0063] Step 402: Smooth the gamma-ray energy spectrum data sequence to obtain a smoothed energy spectrum curve; perform background subtraction on the smoothed energy spectrum curve to obtain net count energy spectrum data; identify characteristic peaks in the net count energy spectrum data to obtain a set of characteristic peak positions; perform Gaussian fitting on each characteristic peak in the set of characteristic peak positions to extract peak energy, peak area, net count, and half-width at half-maximum (WHM) parameters, and classify them to obtain a set of nuclide characteristic parameters; divide the time series of process operating parameters into equal-duration windows to obtain an operating condition analysis window sequence; analyze the temperature and pressure within each operating condition analysis window. The mean, variance, and trend slope of flow rate and liquid level data were calculated separately and combined to obtain the dynamic feature set of the operating conditions. Statistical analysis was performed on the pH, conductivity, and suspended solids concentration data in the waste liquid physical property parameter dataset, calculating the mean, standard deviation, and correlation coefficients among the three. This data were then integrated to obtain the physical property statistical feature vector. Specifically, this involved processing the gamma-ray energy spectrum data sequence, using a moving average method to smooth and filter the gamma-ray energy spectrum data sequence, selecting a fixed-length sliding window, summing all energy spectrum data within the window, and then dividing by the number of data points within the window. Smoothed data at the center of the window is obtained. The sliding window is moved sequentially to traverse the entire gamma-ray energy spectrum data sequence, eliminating random noise in the energy spectrum data and obtaining a smooth energy spectrum curve. Background subtraction is performed on the smooth energy spectrum curve. The smooth energy spectrum curve's flat regions between characteristic peaks are analyzed, and multiple sampling points representing the background are selected. A continuous curve reflecting the background trend of the energy spectrum is constructed using these sampling points. The corresponding background curve data point is subtracted from each data point in the smooth energy spectrum curve to obtain the net count energy spectrum data, eliminating the influence of background interference on characteristic peak identification. Characteristic peak identification is performed on the net count energy spectrum data. A reasonable peak identification threshold is set, and the net count energy spectrum data is traversed. When the value of a data point exceeds the set threshold, and the values ​​of the adjacent data points before and after it are all less than the value of the data point, the data point is determined to be the peak point of the characteristic peak. The energy position corresponding to the peak point is recorded. The energy positions corresponding to all identified peak points are summarized to obtain a characteristic peak position set. Gaussian fitting is performed on each characteristic peak in the characteristic peak position set to construct a Gaussian fitting function. The expression of the Gaussian fitting function is: Where y represents the count value of the net count energy spectrum at a certain energy position, and A represents the peak height parameter of the characteristic peak. This indicates the energy position to be fitted. The peak energy represents the characteristic peak, F represents half the full width at half maximum (FWHM) parameter of the characteristic peak, and C represents the background correction parameter. The energy spectrum data corresponding to each characteristic peak is substituted into the Gaussian fitting function. By iteratively calculating and adjusting the parameters in the function, the constructed Gaussian fitting function is made to fit the energy spectrum data corresponding to the characteristic peak as closely as possible. Then, the three core parameters of each characteristic peak are extracted: peak energy, net peak area count, and FWHM. These three parameters of all characteristic peaks are sorted and classified according to peak energy to form a set of nuclide characteristic parameters.

[0064] Simultaneously, the time series of process operating parameters is processed. Based on the actual process operation, fixed equal-length windows are set. The window duration is determined in conjunction with the acquisition frequency to ensure that each window contains a sufficient number of process operating parameter data. Windows are sequentially divided according to timestamp order, dividing the entire process operating parameter time series into multiple continuous and non-overlapping operating condition analysis windows, forming an operating condition analysis window sequence. Statistical calculations are performed on the temperature, pressure, flow, and liquid level data within each operating condition analysis window. When calculating the mean of the temperature data, all temperature data within the window are summed and then divided by the number of temperature data points within the window. When calculating the variance of the temperature data, the mean of all temperature data within the window is calculated first, then each temperature data point is subtracted from the mean, and the squares of each difference are summed and then divided by the number of temperature data points within the window. When calculating the trend slope of the temperature data, a linear regression equation is constructed with the timestamp as the horizontal axis and the temperature data as the vertical axis. The expression of the linear regression equation is: Y = K × T + B, where Y represents the value of the process operating parameter at the corresponding timestamp, and T represents the time interval. In this context, K represents the slope of the linear regression equation, i.e., the trend slope of the process operating parameters, and B represents the intercept of the linear regression equation. The trend slope of the temperature data is obtained by calculating the slope of the linear regression equation. Using the same calculation method, the mean, variance, and trend slope of the pressure, flow rate, and liquid level data within each operating condition analysis window are calculated separately. The mean, variance, and trend slope of the temperature, mean, variance, and trend slope of the pressure, mean, variance, and trend slope of the flow rate, mean, variance, and trend slope of the flow rate, and mean, variance, and trend slope of the liquid level are all combined to form the operating condition feature corresponding to that window. After summarizing the operating condition features of all windows, the dynamic feature set of the operating conditions is obtained, reflecting the dynamic change law of the process operating conditions.

[0065] Step 403 involves normalizing the set of nuclide feature parameters, the dynamic feature set of operating conditions, and the statistical feature vector of physical properties to obtain an initial fusion feature vector. Specifically, this includes: acquiring the set of nuclide feature parameters, the dynamic feature set of operating conditions, and the statistical feature vector of physical properties; clarifying the value range and data magnitude of the three types of features; and, due to the different representation objects and units of measurement of the three types of features, the data magnitudes differ significantly, requiring normalization to eliminate the impact of these magnitude differences on subsequent feature fusion and dimensionality reduction. A linear normalization method is then used to normalize each parameter in the set of nuclide feature parameters. The process begins by identifying the maximum and minimum values ​​of the parameter within the entire set. The minimum value of each parameter is then subtracted from its value, and the result is divided by the difference between the maximum and minimum values ​​to obtain the normalized value. All nuclide feature parameters are processed in this way to obtain a normalized set of nuclide feature parameters. Using the same linear normalization method, each working condition feature in the dynamic feature set and each physical property feature in the physical property statistical feature vector are normalized to obtain a normalized set of dynamic feature parameters and a normalized set of physical property statistical feature vectors. These are then concatenated sequentially, arranging all feature parameters in order to form a vector containing all normalized features; this vector is the initial fused feature vector.

[0066] Step 404: Dimensionality reduction is performed on the initial fused feature vector. The covariance matrix is ​​calculated to obtain eigenvalues, which are then sorted by size. Based on a preset threshold, eigenvectors corresponding to the eigenvalues ​​are selected to form a transformation matrix. The initial fused feature vector is linearly transformed to a lower-dimensional space to obtain a multi-dimensional feature vector. Specifically, this includes: obtaining the initial fused feature vector; collecting the initial fused feature vectors of multiple samples; constructing a feature matrix, where each row corresponds to the initial fused feature vector of a sample, and each column corresponds to a feature parameter; calculating the covariance matrix of the feature matrix. When calculating the covariance matrix, the mean of each feature parameter is first calculated. The mean of the feature parameter is then subtracted from the value of that feature parameter for each sample to obtain the deviation value. The deviation values ​​of all samples are used to construct a deviation matrix. The transpose of the deviation matrix is ​​multiplied by the deviation matrix, and then divided by the difference between the number of samples and one to obtain the covariance matrix. The covariance matrix reflects the correlation between the various feature parameters. The covariance... The matrix is ​​decomposed into eigenvalues ​​to find all eigenvalues ​​and corresponding eigenvectors of the covariance matrix. Each eigenvalue corresponds to an eigenvector, and the magnitude of the eigenvalue reflects the amount of feature information carried by the corresponding eigenvector. All eigenvalues ​​are sorted in descending order; the larger the eigenvalue, the more information the corresponding eigenvector carries. A preset threshold is set, which is determined according to the actual feature extraction requirements, to select eigenvectors with sufficient information. The sorted eigenvalues ​​are traversed, and all eigenvalues ​​greater than or equal to the preset threshold are selected. The corresponding eigenvectors are extracted and arranged in descending order to form a transformation matrix. The initial fused eigenvectors are linearly transformed with the transformation matrix by multiplying the initial fused eigenvectors by the transformation matrix, mapping the initial fused eigenvectors to a low-dimensional space composed of the selected eigenvectors, resulting in a dimensionality-reduced eigenvector, which is the multidimensional eigenvector.

[0067] This embodiment achieves accurate extraction of features related to the type and concentration of radionuclides by separating the data sources of the fused data base, extracting features from each data source, normalizing features, and reducing dimensions. It effectively removes redundant features, retains core feature information, and ensures that the extracted multidimensional feature vectors can accurately reflect the characteristics of radionuclides.

[0068] In a preferred embodiment of the present invention, step 5, which involves inputting the multidimensional feature vector into a pre-trained pattern recognition model for classification and fusion recognition to obtain the identification results of radionuclides and their concentration prediction values, may include:

[0069] Step 501: Normalize the multidimensional feature vector to generate a normalized feature vector; input the normalized feature vector into a pre-trained pattern recognition model. The pattern recognition model includes an input layer, multiple hidden layers, a classification output layer, and a regression output layer connected in sequence. Specifically, this includes: obtaining the multidimensional feature vector generated in step 404. Since the multidimensional feature vector is obtained after dimensionality reduction, the value range of each feature parameter may still differ. To further optimize the model input and ensure the accuracy of model recognition, the multidimensional feature vector is normalized. The same linear normalization method as in step 403 is used. First, the maximum and minimum values ​​of each feature parameter in the multidimensional feature vector are found. The minimum value of each feature parameter is subtracted from its value, and then divided by the difference between the maximum and minimum values ​​of the parameter to obtain the normalized value of each feature parameter. After all feature parameters are normalized, a normalized feature vector is generated.

[0070] A pattern recognition model is constructed, comprising an input layer, multiple hidden layers, a classification output layer, and a regression output layer connected sequentially. Specifically, the number of neurons in the input layer matches the dimension of the normalized feature vector, ensuring the input layer can fully receive all feature information from the normalized feature vector. Multiple hidden layers are set, the specific number determined by the number of samples and feature complexity. The number of neurons in each hidden layer gradually decreases, with the first hidden layer having more neurons than the input layer, and subsequent hidden layers having fewer neurons than the previous one. Each hidden layer uses the ReLU activation function to perform a non-linear transformation of the normalized feature vector, extracting deep features. The number of neurons in the classification output layer matches the preset number of radionuclide types, with each neuron corresponding to one preset radionuclide. The Softmax activation function is used to output the probability of the normalized feature vector corresponding to each radionuclide. The regression output layer has one neuron, using a linear activation function, to output the predicted concentration value of the radionuclide. All layers are fully connected, meaning each neuron in the previous layer is connected to every neuron in the next layer, ensuring sufficient transmission of feature information.

[0071] The pattern recognition model is trained as follows: A large amount of sample data is collected, each containing a multi-dimensional feature vector, a corresponding radionuclide label, and an actual concentration value. The sample data is divided into training and testing sets proportionally. The training set is used for model training, and the testing set is used for model performance validation. Training parameters are set, including the learning rate, number of training iterations, batch size, and loss function. The learning rate is set to a suitable fixed value to ensure stable convergence of the model training. The number of training iterations is set to the number of iterations required for the model's loss function to converge. The batch size is determined based on hardware performance and the number of samples. The loss function is a weighted sum of classification and regression loss functions. The classification loss function measures the error in the model's species recognition, and the regression loss function measures the error in the model's concentration prediction. After normalizing the multi-dimensional feature vectors from the training set, they are input into the constructed pattern recognition model. The model receives the normalized feature vectors through the input layer, performs nonlinear transformations layer by layer through multiple hidden layers to extract deep fusion features, and then inputs them to the classification output layer and the regression output layer, respectively. The model is divided into two layers: a classification output layer that outputs the probability of each radionuclide, and a regression output layer that outputs the concentration prediction value. The model's output probability of species and concentration prediction value are compared with the actual species labels and actual concentration values ​​corresponding to the training set samples. The loss function value is calculated, and the connection weights and biases of each layer of the model are adjusted according to the loss function value through backpropagation algorithm to gradually reduce the loss function value. The above training process is repeated until the number of training iterations reaches a preset value or the loss function value converges to a preset threshold, at which point training stops. The trained model is then validated using a test set. The multidimensional feature vector of the test set is input into the model to obtain the model's recognition result and prediction value. This is compared with the actual value of the test set samples to calculate the recognition accuracy and concentration prediction error. If both the accuracy and error meet the preset requirements, the model training is complete, and a pre-trained pattern recognition model is obtained. If the requirements are not met, the model structure or training parameters are adjusted, and retraining is performed until the model performance meets the standards. The generated normalized feature vector is input into the pre-trained pattern recognition model, and the model is started to recognize and predict.

[0072] Step 502: The input layer receives the normalized feature vector and passes it to the first hidden layer; multiple hidden layers perform layer-by-layer nonlinear transformations on the normalized feature vector to extract deep fusion features, and pass the deep fusion features to the classification output layer and the regression output layer respectively. Specifically, after the pattern recognition model is started, the input layer receives the normalized feature vector generated in step 501 and passes each feature value of the normalized feature vector to the corresponding neuron of the first hidden layer. Each neuron in the first hidden layer receives feature values ​​from the input layer. All received feature values ​​are multiplied by the corresponding connection weights of that neuron, and the neuron's bias is added to obtain its input value. This input value is then substituted into the ReLU activation function to calculate the neuron's output value. All neurons in the first hidden layer are calculated in this way, resulting in the output feature vector of the first hidden layer. This achieves the first nonlinear transformation and feature extraction of the normalized feature vector. The output feature vector of the first hidden layer is then passed to the second hidden layer. The second hidden layer uses the same calculation method as the first, performing a nonlinear transformation on the input feature vector to extract deeper features, resulting in the output feature vector of the second hidden layer. Following this process, the output feature vector of the previous hidden layer is passed to the next hidden layer in turn. Multiple hidden layers perform nonlinear transformations and feature extraction on the feature vectors layer by layer, gradually removing redundant features and retaining the core features related to the type and concentration of radionuclides, ultimately obtaining the deep fusion feature. After the deep fusion feature is generated, the model passes it to the classification output layer and the regression output layer, respectively.

[0073] Step 503: The classification output layer calculates the deep fusion features to obtain the probability distribution of the deep fusion features corresponding to each preset radionuclide type. The type with the highest probability value is determined as the radionuclide type identification result. Specifically, the classification output layer receives the deep fusion features transmitted from multiple hidden layers and processes the deep fusion features. Each neuron in the classification output layer corresponds to a preset radionuclide. Each neuron receives all feature values ​​of the deep fusion features, multiplies each feature value by the connection weight corresponding to the neuron, and adds the bias of the neuron to obtain the input value of the neuron. Substitutes the input values ​​of all neurons into the Softmax activation function, and converts the input value of each neuron into a probability value between 0 and 1 through calculation. The sum of the probability values ​​of all neurons is 1, forming the probability distribution of the deep fusion features corresponding to each preset radionuclide type. Traversing all probability values ​​in the probability distribution, the highest probability value is found. The preset radionuclide type corresponding to this probability value is the radionuclide type identification result.

[0074] Step 504: The regression output layer performs fully connected computation on the deep fusion features and obtains the concentration value corresponding to the deep fusion features through a linear activation function. The concentration value is used as the concentration prediction value of the radionuclide. The type identification result is associated with the concentration prediction value to generate the final radionuclide type identification result and its concentration prediction value. Specifically, the regression output layer receives the deep fusion features transmitted from multiple hidden layers and performs fully connected computation on the deep fusion features. The neurons of the regression output layer receive all feature values ​​of the deep fusion features, multiply each feature value by the connection weight corresponding to the neuron, and add the bias of the neuron to obtain the input value of the neuron. The input value is substituted into the linear activation function, and the linear activation function directly outputs the input value as the concentration value. This concentration value is the concentration prediction value of the radionuclide, which can reflect the approximate concentration level of the radionuclide. The radionuclide type identification result obtained in step 503 is associated with the concentration prediction value obtained in this step. The type identification result and the concentration prediction value are stored accordingly to clarify the concentration prediction value corresponding to each identified radionuclide and generate the final radionuclide type identification result and its concentration prediction value.

[0075] In this embodiment, a pattern recognition model is constructed, which includes an input layer, multiple hidden layers, a classification output layer, and a regression output layer. After sufficient training, the model can accurately identify and predict multi-dimensional feature vectors, and at the same time complete the identification of radionuclides and concentration prediction, thereby improving the accuracy and efficiency of identification and prediction.

[0076] like Figure 2 As shown, embodiments of the present invention also provide a multi-source data fusion system for the separation of radionuclides, comprising:

[0077] The acquisition module is used to collect multi-source monitoring data generated during the separation of radionuclides. The multi-source monitoring data includes gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data.

[0078] The identification module is used to convert multi-source monitoring data into feature element clusters, construct a configuration field based on the feature element clusters, and determine the core identification window on the configuration field. Within the core identification window, extreme value search is performed based on the geometric distribution boundary of the feature element clusters to determine the support boundary that can completely enclose the feature element clusters and has the minimum external geometry.

[0079] The correction module is used to divide the core identification window into multiple feature interpretation blocks based on the support boundary, and project the feature feature clusters into the corresponding feature interpretation blocks; based on the statistical characteristics of the feature feature points inside the feature interpretation blocks, and combined with the relationship between the process condition parameter data and waste liquid physical property parameter data corresponding to the feature interpretation blocks, a dynamic kernel calibration unit cell is generated; the dynamic kernel calibration unit cell is used to correct the multi-source monitoring data to obtain the fused data base;

[0080] The extraction module is used to extract features from the fused data base, extract feature information related to the type and concentration of radionuclides, and generate multi-dimensional feature vectors.

[0081] The output module is used to input multidimensional feature vectors into a pre-trained pattern recognition model for classification and fusion recognition, so as to obtain the identification results of radionuclides and their concentration prediction values.

[0082] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0083] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0084] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for multi-source data fusion during the separation of radionuclides, characterized in that, The method includes: Step 1: Collect radiation signals from radioactive nuclides in the separation process flow channel. Analyze and process these signals to generate gamma-ray energy spectrum data. Based on the gamma-ray energy spectrum data, simultaneously collect the original process parameters during the separation process. Perform feature-weighted filtering on these parameters using the gamma-ray energy spectrum data to generate process condition parameter data, including temperature, pressure, flow rate, and liquid level data. Based on the process condition parameter data, simultaneously collect the original detection values ​​of waste liquid samples during the separation process. Calibrate these values ​​using the process condition parameter data to generate waste liquid physical property parameter data, including pH, conductivity, and suspended solids concentration data. Perform spatiotemporal registration of the gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data according to a unified time reference to generate multi-source monitoring data. Step 2: Aggregate and classify the data points in the multi-source monitoring data based on the feature similarity to generate feature element clusters; extract the geometric distribution boundary based on the coordinate distribution of each data point in the feature element clusters; construct a closed spatial morphology based on the geometric distribution boundary; generate a field domain for the closed spatial morphology to obtain a configuration field; scan the density distribution of the configuration field to obtain the density distribution data of the feature element clusters; perform extreme point detection and gradient analysis on the density distribution data to generate density distribution extreme points and gradient change features; delineate the core analysis area based on the density distribution extreme points and gradient change features; define the core analysis area by windowing to generate a core identification window; take the outermost distribution point of the feature element cluster within the core identification window as the search starting point, and perform successive approximation searches along various dimensions to obtain a boundary point set; construct the envelope morphology of the boundary point set to generate the spatial boundary to be confirmed; optimize the minimum bounding morphology of the spatial boundary to be confirmed to generate a supporting boundary. Step 3: Divide the core identification window into multiple element interpretation blocks using the support boundary as the segmentation benchmark, and project the feature element clusters into the corresponding element interpretation blocks; Based on the statistical characteristics of the feature element points within the element interpretation blocks, and combined with the relationship between the process condition parameter data and waste liquid physical property parameter data corresponding to the element interpretation blocks, generate dynamic kernel calibration units; Use dynamic kernel calibration units to correct the multi-source monitoring data to obtain the fused data base; Step 4: Extract features from the fused data base to extract feature information related to the type and concentration of radionuclides and generate a multi-dimensional feature vector; Step 5: Input the multidimensional feature vector into the pre-trained pattern recognition model for classification and fusion recognition to obtain the identification results of radionuclides and their concentration prediction values.

2. The multi-source data fusion method in the process of radionuclide separation according to claim 1, characterized in that, Envelope morphology construction is performed on the boundary point set to generate the spatial boundary to be confirmed. Minimum bounding morphology optimization is then performed on the spatial boundary to generate the supporting boundary, including: Calculate the connection order between adjacent boundary points based on their spatial coordinates, and connect the boundary points sequentially according to the connection order to generate the initial envelope contour line; The initial envelope contour is detected, and the endpoints are connected according to the opening position of the detected contour. The resulting closed contour is used as the boundary of the envelope to be filled. The internal region enclosed by the boundary to be filled is subdivided to obtain mesh cells; the mesh cells are then filled to generate the spatial boundary to be confirmed. The circumscribed morphological parameters of each boundary dimension of the spatial boundary to be confirmed are extracted. An initial shrinkage step size is set according to the circumscribed morphological parameters, and each boundary dimension of the spatial boundary to be confirmed is shrunk according to the initial shrinkage step size to generate a preliminary shrunk boundary. The integrity of the initial contraction boundary is verified to obtain the verification result; the contraction step size correction is generated based on the verification result; the initial contraction step size is adjusted based on the contraction step size correction to generate the corrected contraction step size; the initial contraction boundary is further dimensionally contracted based on the corrected contraction step size to generate the secondary contraction boundary. The integrity verification and boundary dimension shrinkage are repeated on the secondary shrinkage boundary until the shrunken boundary exactly encloses the entire feature element cluster and further shrinkage will not be able to completely enclose it. The iteration stops when this is done, and the supporting boundary is obtained.

3. The multi-source data fusion method in the process of radionuclide separation according to claim 2, characterized in that, Using the supporting boundary as the segmentation criterion, the core identification window is divided into multiple feature interpretation blocks, and feature feature clusters are projected into the corresponding feature interpretation blocks, including: Using the support boundary as the segmentation benchmark, the core identification window is divided into spatial regions. Based on the geometric orientation of the support boundary, the core identification window is divided into multiple initial segmentation sub-regions to obtain candidate interpretation blocks. The overlapping areas between adjacent candidate interpretation blocks are detected and processed, and the boundaries of adjacent candidate interpretation blocks are adjusted to obtain preliminary interpretation blocks; The shape of the preliminary interpretation block is normalized, and the boundary of the preliminary interpretation block is fine-tuned to obtain the feature interpretation block; the spatial coordinates of each feature point in the feature feature cluster are extracted, and the spatial positional relationship between the feature point and each feature interpretation block is calculated; the relationship data between the feature point and the feature interpretation block is determined based on the spatial positional relationship. Based on the relationship data between feature points and feature interpretation blocks, feature points are projected into the corresponding feature interpretation blocks.

4. The multi-source data fusion method in the process of radionuclide separation according to claim 3, characterized in that, Based on the statistical characteristics of feature points within the feature interpretation block, and combined with the relationship between process condition parameter data and waste liquid physical property parameter data corresponding to the feature interpretation block, a dynamic kernel calibration unit cell is generated. The multi-source monitoring data were corrected by using dynamic kernel decomposition calibration units to obtain a fused data base, including: The spatial distribution density of feature points within each feature interpretation block is statistically analyzed to generate density distribution data. Based on the density distribution data, the geometric mean position of all feature points is calculated, and the geometric mean position is determined as the initial calibration base point of the feature interpretation block. Centered on the initial calibration base point, the spatial dispersion of all feature points relative to the initial calibration base point is analyzed, and the dispersion feature value of the feature interpretation block is calculated. Based on the dispersion feature value, the dynamic influence radius is determined, and the initial kernel calibration space is constructed with the initial calibration base point as the core and the dynamic influence radius as the basis. Retrieve pre-stored process condition parameter data and waste liquid physical property parameter data that have spatial mapping relationships with the element interpretation blocks, analyze the two, and calculate the coupling coefficient between the condition and the physical property; use the coupling coefficient between the condition and the physical property to adjust the boundary morphology of the initial kernel calibration space to generate a dynamic kernel calibration unit cell; match and correct each data point in the multi-source monitoring data with the dynamic kernel calibration unit cell to generate a fused data base.

5. The multi-source data fusion method in the process of radionuclide separation according to claim 4, characterized in that, Feature extraction is performed on the fused data base to extract feature information related to the type and concentration of radionuclides, generating a multidimensional feature vector, including: The data sources of the fused data base were separated, and the gamma-ray energy spectrum data sequence, the time series of process operating parameters, and the dataset of waste liquid physical property parameters were extracted respectively. The gamma-ray energy spectrum data sequence was smoothed and filtered to obtain a smoothed energy spectrum curve. Background subtraction was performed on the smoothed energy spectrum curve to obtain net count energy spectrum data. Characteristic peaks were identified in the net count energy spectrum data to obtain a set of characteristic peak positions. Gaussian fitting was performed on each characteristic peak in the characteristic peak position set to extract peak energy, peak area, net count, and half-width at half-maximum (WHM) parameters. These parameters were then categorized to obtain a set of nuclide characteristic parameters. The time series of process operating parameters was divided into equal-length windows to obtain an operating condition analysis window sequence. The mean, variance, and trend slope of the temperature, pressure, flow rate, and liquid level data within each operating condition analysis window were calculated and combined to obtain a dynamic feature set of operating conditions. Statistical analysis was performed on the acidity, alkalinity, conductivity, and suspended solids concentration data in the waste liquid physical property parameter dataset. The mean, standard deviation, and correlation coefficient among the three parameters were calculated and integrated to obtain a statistical feature vector of physical properties. The set of nuclide characteristic parameters, the set of dynamic characteristics under operating conditions, and the statistical characteristic vector of physical properties are normalized to obtain the initial fused characteristic vector; The initial fused feature vector is dimensionality reduced by calculating the covariance matrix to obtain eigenvalues, which are then sorted by size. Based on a preset threshold, the eigenvectors corresponding to the eigenvalues ​​are selected to form a transformation matrix, which linearly transforms the initial fused feature vector to a low-dimensional space to obtain a multi-dimensional feature vector.

6. The multi-source data fusion method in the process of radionuclide separation according to claim 5, characterized in that, Multidimensional feature vectors are input into a pre-trained pattern recognition model for classification and fusion recognition, yielding the identification results of radionuclides and their concentration predictions, including: The multidimensional feature vectors are normalized to generate normalized feature vectors; the normalized feature vectors are then input into a pre-trained pattern recognition model, which includes an input layer, multiple hidden layers, a classification output layer, and a regression output layer connected in sequence. The input layer receives the normalized feature vector and passes it to the first hidden layer; Multiple hidden layers perform layer-by-layer nonlinear transformations on the normalized feature vectors to extract deep fusion features, which are then passed to the classification output layer and the regression output layer, respectively. The classification output layer calculates the deep fusion features to obtain the probability distribution of the deep fusion features corresponding to each preset radionuclide species, and determines the species with the highest probability value as the radionuclide species identification result. The regression output layer performs fully connected computation on the deep fusion features and obtains the concentration values ​​corresponding to the deep fusion features through a linear activation function. The concentration values ​​are used as the concentration prediction values ​​of radionuclides. The species identification results are correlated with the concentration prediction values ​​to generate the final radionuclide species identification results and their concentration prediction values.

7. A multi-source data fusion system for the separation of radionuclides, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to collect multi-source monitoring data generated during the separation of radionuclides. The multi-source monitoring data includes gamma-ray energy spectrum data, process condition parameter data, and waste liquid physical property parameter data. The identification module is used to convert multi-source monitoring data into feature element clusters, construct configuration fields based on feature element clusters, and determine the core identification window on the configuration field. Within the core identification window, extreme value search is performed based on the geometric distribution boundary of the feature element cluster to determine the support boundary that can completely enclose the feature element cluster and has the smallest bounded geometry. The correction module is used to divide the core identification window into multiple feature interpretation blocks based on the support boundary, and to project the feature feature clusters into the corresponding feature interpretation blocks. Based on the statistical characteristics of feature points within the feature interpretation block, and combined with the relationship between process condition parameter data and waste liquid physical property parameter data corresponding to the feature interpretation block, a dynamic kernel calibration unit cell is generated; the dynamic kernel calibration unit cell is used to correct the multi-source monitoring data to obtain the fused data base. The extraction module is used to extract features from the fused data base, extract feature information related to the type and concentration of radionuclides, and generate multi-dimensional feature vectors. The output module is used to input multidimensional feature vectors into a pre-trained pattern recognition model for classification and fusion recognition, so as to obtain the identification results of radionuclides and their concentration prediction values.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

Images