Method and system for predicting lifetime of a radioseparation resin based on machine learning

Abstract

The application provides a radioactive separation resin life prediction method and system based on machine learning, relates to the technical field of radioactive waste liquid treatment, and the method comprises the following steps: acquiring real-time online monitoring data of a radioactive separation resin in a running process, wherein the real-time online monitoring data comprises local adsorption capacity, local temperature and local radiation dose collected from three point positions, namely, upper, middle and lower positions, of the axial distribution of a resin column; based on the real-time online monitoring data, a microstructure deterioration proxy model is constructed, and a microstructure deterioration quantitative index of the resin under a current working condition is calculated; and the real-time online monitoring data and the microstructure deterioration quantitative index are subjected to feature fusion, and a fusion feature set is constructed. The application guarantees the continuous, stable and safe operation of a radioactive waste liquid treatment and nuclide separation system.

Description

Technical Field

[0001] This invention relates to the field of radioactive waste liquid treatment technology, and in particular to a method and system for predicting the lifetime of radioactive separation resins based on machine learning. Background Technology

[0002] In the field of spent fuel reprocessing and deep purification of radioactive waste liquid in the nuclear industry, radioactive separation resin is the core consumable for adsorbing and separating key radionuclides such as cesium and strontium. Its service life directly affects the system's operational safety and the effectiveness of nuclide separation.

[0003] Taking a radioactive waste liquid treatment system of a nuclear power plant as an example, this system uses styrene-based strong basic anion exchange resin for the adsorption and purification of radionuclides. Currently, the industry still mainly relies on periodic offline sampling to detect the overall exchange capacity of the resin and manual experience to determine the lifetime, obtaining average performance indicators of the resin. It is impossible to monitor the local adsorption capacity, temperature, and irradiation dose at the axial points of the upper, middle, and lower parts of the resin column in real time. There is also no quantitative characterization method for microstructure deterioration such as resin pore collapse and functional group shedding. Furthermore, there is no machine learning prediction model with generalization and enhancement capabilities for extreme conditions such as instantaneous strong irradiation, sudden changes in feed acidity, and sudden changes in competing ion concentration. It is also impossible to identify local deterioration inflection points and correct lifetime prediction results by combining axial state distribution curves. It has technical defects such as lack of online monitoring, inability to quantify microstructure deterioration, poor adaptability to extreme conditions, failure to consider the non-uniformity of axial spatial deterioration, and low lifetime prediction accuracy. It is very easy to misjudge the resin lifetime, resulting in a decrease in separation efficiency or even resin failure and leakage. Summary of the Invention

[0004] This invention provides a method and system for predicting the lifetime of radioactive separation resins based on machine learning, ensuring the continuous, stable, and safe operation of radioactive waste treatment and radionuclide separation systems.

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

[0006] Firstly, a machine learning-based method for predicting the lifetime of radioactive separation resins, the method comprising:

[0007] Real-time online monitoring data of the radioactive separation resin during operation is obtained. The real-time online monitoring data includes local adsorption capacity, local temperature and local irradiation dose collected from three points distributed axially along the resin column: the upper, middle and lower parts.

[0008] Based on real-time online monitoring data, the microstructure deterioration surrogate model was constructed to calculate the quantitative index of resin microstructure deterioration under the current working conditions.

[0009] Real-time online monitoring data is fused with quantitative indicators of microstructure deterioration to construct a fused feature set;

[0010] The fused feature set is input into a pre-trained extreme condition generalized enhanced machine learning model to obtain the baseline remaining life prediction value of the resin under the current working condition.

[0011] In real time, it is determined whether the current working condition belongs to an extreme working condition. Extreme working conditions include instantaneous strong radiation, sudden change in acidity of the feed solution, and sudden change in competitive ion concentration. If it belongs to an extreme working condition, a state distribution curve along the axial direction of the resin column is constructed based on the local adsorption capacity, local temperature, and local irradiation dose at three points: the upper, middle, and lower parts.

[0012] Curvature calculation and inflection point detection are performed on the state distribution curve to obtain the local degradation correction factor for each axial segment. The comprehensive spatial correction coefficient is generated by weighted fusion. The comprehensive spatial correction coefficient is used to correct the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value. If it does not belong to the range, the baseline remaining lifetime prediction value is directly used as the final remaining lifetime prediction value.

[0013] Furthermore, real-time online monitoring data of the radioactive separation resin during operation is obtained. This data includes local adsorption capacity, local temperature, and local irradiation dose collected from three points axially distributed along the resin column: the upper, middle, and lower sections.

[0014] Adsorption capacity sensor, temperature sensor and irradiation dose detector are respectively installed at the upper, middle and lower parts of the resin column to collect the original signals of local adsorption capacity, local temperature and local irradiation dose at each point in real time.

[0015] The acquired raw signals are converted from analog to digital to generate corresponding time series data. The time series data of local adsorption capacity, local temperature and local irradiation dose at the three points are synchronized and aligned with the timestamps to obtain synchronized data.

[0016] The synchronized data is processed using a sliding window method to detect and remove outliers, resulting in preprocessed real-time online monitoring data.

[0017] Furthermore, based on real-time online monitoring data, a microstructure deterioration proxy model was constructed to calculate quantitative indicators of resin microstructure deterioration under current operating conditions, including:

[0018] The pre-processed real-time online monitoring data is input into the pre-built microstructure deterioration proxy model. The microstructure deterioration proxy model is a regression model trained by a deep neural network based on the resin offline detection results under different working conditions in historical operating data.

[0019] The input data is subjected to multi-layer nonlinear mapping calculation by a microstructure degradation proxy model, and the output microstructure degradation quantification index includes the pore collapse coefficient, which characterizes the degree of pore collapse in the resin, and the functional group retention rate, which characterizes the degree of functional group detachment. The pore collapse coefficient is indirectly calculated by the decay rate of the resin specific surface area, and the functional group retention rate is indirectly calculated by the decay rate of the ion exchange capacity.

[0020] Furthermore, real-time online monitoring data is fused with microstructure degradation quantification indicators to construct a fused feature set, including:

[0021] The local adsorption capacity, local temperature, and local irradiation dose in the preprocessed real-time online monitoring data are spliced ​​with the pore collapse coefficient and functional group retention rate in a feature layer to generate an initial fusion feature vector; the initial fusion feature vector is then normalized to obtain a standardized fusion feature set.

[0022] Furthermore, the fused feature set is input into a pre-trained extreme condition generalization-enhanced machine learning model to obtain the baseline remaining life prediction of the resin under the current operating conditions, including:

[0023] The standardized fusion feature set is input into the extreme condition generalized enhanced machine learning model. The model is built on a deep neural network. It receives the fusion feature set through the input layer, performs nonlinear feature transformation and mapping through multiple hidden layers, and finally generates a one-dimensional value from the output layer as the baseline remaining life prediction value under the current operating condition. The extreme condition generalized enhanced machine learning model is optimized by introducing adversarial samples and physical consistency constraints during the training phase.

[0024] Furthermore, it determines in real time whether the current operating condition belongs to an extreme condition, which includes instantaneous strong radiation, sudden changes in feed solution acidity, and sudden changes in competing ion concentration. If it does, a state distribution curve along the resin column axis is constructed based on the local adsorption capacity, local temperature, and local irradiation dose at three points: the upper, middle, and lower sections.

[0025] The system determines in real time whether the current operating condition is an extreme condition, which includes instantaneous strong radiation, sudden changes in feed solution acidity, and sudden changes in competing ion concentration. If it is, the local adsorption capacity, local temperature, and local radiation dose at the three points are normalized to obtain normalized values ​​for adsorption capacity, temperature, and radiation dose. The normalized values ​​for adsorption capacity, temperature, and radiation dose are then weighted and summed according to preset weights to calculate the comprehensive state index value for each point.

[0026] Based on the comprehensive state index value, the axial coordinates corresponding to the upper, middle and lower parts of the resin column are used as interpolation nodes, and the comprehensive state index value of each node is used as the node function value. A continuous state distribution curve along the axial direction of the resin column is constructed using the cubic spline interpolation method. Specifically, a cubic polynomial is constructed between every two adjacent nodes, and all cubic polynomials have the same function value, first derivative value and second derivative value at the nodes. By solving the three moment equations established by the continuity condition of the second derivative at the nodes, the coefficients of each piecewise cubic polynomial are determined, thereby obtaining the state distribution curve composed of piecewise cubic polynomials.

[0027] Furthermore, curvature calculation and inflection point detection are performed on the state distribution curve to obtain the local degradation correction factor for each axial segment. A comprehensive spatial correction coefficient is generated through weighted fusion. This comprehensive spatial correction coefficient is used to correct the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value. If it does not belong to the range specified in the original text, the baseline remaining lifetime prediction value is directly used as the final remaining lifetime prediction value, including:

[0028] The curvature of the state distribution curve is calculated to obtain the curvature value corresponding to each point on the curve; extreme points of curvature are detected, and the axial positions corresponding to the local maxima of curvature are marked as potential deterioration mutation points, and the potential deterioration mutation points are used as inflection points of the state distribution curve.

[0029] The resin column is divided into several continuous deterioration sections along the axial direction based on the inflection point. The curvature value is integrated within each deterioration section to obtain the cumulative deterioration intensity of the section. The ratio of the cumulative deterioration intensity to the section length is used as the local deterioration correction factor of the section. The local deterioration correction factors of each section are weighted and summed according to the proportion of each deterioration section to the total height of the resin column to obtain the comprehensive spatial correction coefficient.

[0030] Multiply the comprehensive spatial correction coefficient by the baseline remaining life prediction value to obtain the final remaining life prediction value after spatial distribution characteristic correction; if it is determined in real time that the current operating condition is not an extreme operating condition, the baseline remaining life prediction value is directly output as the final remaining life prediction value.

[0031] Secondly, a machine learning-based system for predicting the lifetime of radioactive separation resins includes:

[0032] The acquisition module is used to acquire real-time online monitoring data of the radioactive separation resin during operation. The real-time online monitoring data includes local adsorption capacity, local temperature and local irradiation dose collected from three points distributed axially along the resin column: the upper, middle and lower parts.

[0033] The calculation module is used to calculate the quantitative index of microstructure deterioration of resin under the current working conditions based on real-time online monitoring data and through the constructed microstructure deterioration proxy model.

[0034] The fusion module is used to fuse real-time online monitoring data with microstructure deterioration quantification indicators to construct a fusion feature set.

[0035] The prediction module is used to input the fused feature set into a pre-trained extreme condition generalized enhanced machine learning model to obtain the baseline remaining life prediction value of the resin under the current working conditions.

[0036] The judgment module is used to determine in real time whether the current working condition belongs to an extreme working condition. Extreme working conditions include instantaneous strong radiation, sudden change in acidity of the feed solution, and sudden change in competitive ion concentration. If it belongs to an extreme working condition, the state distribution curve along the axial direction of the resin column is constructed based on the local adsorption capacity, local temperature, and local radiation dose at the upper, middle, and lower points.

[0037] The correction module is used to calculate the curvature and detect inflection points of the state distribution curve, obtain the local degradation correction factor of each axial segment, generate a comprehensive spatial correction coefficient through weighted fusion, and use the comprehensive spatial correction coefficient to correct the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value; if it does not belong to the category, the baseline remaining lifetime prediction value is directly used as the final remaining lifetime prediction value.

[0038] Thirdly, a computing device includes:

[0039] One or more processors;

[0040] 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.

[0041] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0042] The above-described solution of the present invention has at least the following beneficial effects:

[0043] By using online monitoring to collect local adsorption capacity, local temperature, and local irradiation dose at three points along the axial direction of the resin column (upper, middle, and lower), and combining this with a microstructure degradation proxy model to quantify the resin's microscopic degradation indicators, the monitoring data and degradation indicators are fused together and input into an optimized extreme-condition generalized enhanced machine learning model to obtain the baseline remaining lifetime. Furthermore, by constructing an axial state distribution curve under extreme conditions, calculating curvature, and detecting inflection points to generate a comprehensive spatial correction coefficient to correct the lifetime prediction value, this technique overcomes the technical problems of existing technologies in predicting the lifetime of radioactive separation resins, such as the lack of online monitoring, inability to quantify microstructure degradation, poor adaptability to extreme conditions, lack of consideration for axial spatial degradation heterogeneity, low lifetime prediction accuracy, and susceptibility to lifetime misjudgment. This achieves the technical effect of accurately predicting the remaining lifetime of the resin, avoiding resource waste caused by premature resin replacement or the risk of radionuclide leakage due to prolonged service life, and ensuring the continuous, stable, and safe operation of radioactive waste treatment and radionuclide separation systems. Attached Figure Description

[0044] Figure 1 This is a schematic flowchart of a machine learning-based method for predicting the lifetime of radioactive separation resins, provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of a machine learning-based system for predicting the lifetime of radioactive separation resins, provided in an embodiment of the present invention. Detailed Implementation

[0046] 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, these embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art.

[0047] like Figure 1 As shown, embodiments of the present invention propose a machine learning-based method for predicting the lifetime of radioactive separation resins, the method comprising the following steps:

[0048] Step 1: Obtain real-time online monitoring data of the radioactive separation resin during operation. The real-time online monitoring data includes local adsorption capacity, local temperature, and local irradiation dose collected from three points axially distributed in the upper, middle, and lower parts of the resin column.

[0049] Step 2: Based on real-time online monitoring data, the microstructure deterioration quantification index of the resin under the current working conditions is calculated through the constructed microstructure deterioration proxy model.

[0050] Step 3: Perform feature fusion between real-time online monitoring data and microstructure deterioration quantification indicators to construct a fused feature set;

[0051] Step 4: Input the fused feature set into the pre-trained extreme working condition generalization enhanced machine learning model to obtain the baseline remaining life prediction value of the resin under the current working condition.

[0052] Step 5: Determine in real time whether the current working condition belongs to an extreme working condition. Extreme working conditions include instantaneous strong radiation, sudden change in acidity of the feed solution, and sudden change in competitive ion concentration. If it belongs to an extreme working condition, construct a state distribution curve along the axial direction of the resin column based on the local adsorption capacity, local temperature, and local irradiation dose at the three points of the upper, middle, and lower parts.

[0053] Step 6: Calculate the curvature and detect the inflection point of the state distribution curve to obtain the local degradation correction factor for each axial segment. Generate a comprehensive spatial correction coefficient through weighted fusion. Use the comprehensive spatial correction coefficient to correct the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value. If it does not belong to the range, the baseline remaining lifetime prediction value is directly used as the final remaining lifetime prediction value.

[0054] In this embodiment of the invention, the following technical means are employed: real-time online acquisition of local adsorption capacity, local temperature, and local irradiation dose at three points along the upper, middle, and lower parts of the resin column; calculation of resin microstructure degradation quantification index using a microstructure degradation proxy model; feature fusion of monitoring data and degradation quantification index to construct a fused feature set; input of the fused feature set into a pre-trained extreme condition generalized enhanced machine learning model to obtain a baseline remaining lifetime prediction value; construction of a state distribution curve along the resin column axis under extreme conditions; obtaining a local degradation correction factor through curvature calculation and inflection point detection; generating a comprehensive spatial correction coefficient to correct the baseline prediction value; and direct output of the baseline prediction value under non-extreme conditions. Therefore, this approach overcomes the technical problems of incomplete online monitoring, inability to quantify microstructure degradation, poor adaptability to extreme conditions, failure to consider axial spatial degradation heterogeneity, low prediction accuracy, and susceptibility to lifetime misjudgment in existing radioactive separation resin lifetime prediction methods. This achieves the technical effect of accurately obtaining the remaining resin lifetime, avoiding resource waste caused by premature resin replacement or the risk of radionuclide leakage due to overdue service, and ensuring the continuous, stable, and safe operation of radioactive waste treatment and radionuclide separation systems.

[0055] In a preferred embodiment of the present invention, step 1 above may include:

[0056] Step 1.1 involves installing an adsorption capacity sensor, a temperature sensor, and an irradiation dose detector at the upper, middle, and lower parts of the resin column, respectively, to collect the raw signals of local adsorption capacity, local temperature, and local irradiation dose at each point in real time. Specifically, this includes real-time online monitoring of the local adsorption capacity, temperature, and irradiation dose at axial points in the upper, middle, and lower parts of the resin column. Adsorption capacity sensors, temperature sensors, and irradiation dose detectors are installed at the upper, middle, and lower axial parts of the radioactive separation resin column, respectively. The adsorption capacity sensor collects the raw signal of the local adsorption capacity at the corresponding point in real time, the temperature sensor collects the raw signal of the local temperature at the corresponding point in real time, and the irradiation dose detector collects the raw signal of the local irradiation dose at the corresponding point in real time, thus completing the synchronous acquisition of the raw signals of local adsorption capacity, local temperature, and local irradiation dose at the three axial points of the resin column.

[0057] Step 1.2 involves performing analog-to-digital conversion on the acquired raw signals to generate corresponding time-series data. The local adsorption capacity, local temperature, and local irradiation dose time-series data at the three locations are then time-stamped and aligned to obtain synchronized data. Specifically, this includes transmitting all acquired analog raw signals to the data processing unit, performing analog-to-digital conversion, converting the analog signals into digital signals, and generating local adsorption capacity time-series data, local temperature time-series data, and local irradiation dose time-series data for the upper, middle, and lower locations, respectively.

[0058] Using a unified time reference as the standard, the above nine sets of time series data are synchronized and aligned with timestamps. This ensures that the local adsorption capacity data, local temperature data, and local irradiation dose data at the same time point correspond one-to-one, eliminating time deviations between different points and different parameter data, and finally obtaining synchronized data with complete time dimension matching.

[0059] Step 1.3 involves using a sliding window method to detect and remove outliers from the synchronized data, resulting in preprocessed real-time online monitoring data. Specifically, this includes selecting a fixed-length sliding window to iterate through the synchronized data segment by segment, first calculating the average of all data within a single sliding window, then calculating the standard deviation of the data within that window relative to the average, identifying data values ​​greater than the sum of the average and three times the standard deviation, and data values ​​less than the difference between the average and three times the standard deviation as outliers, directly removing all identified outliers, retaining valid and stable data, and finally obtaining the preprocessed real-time online monitoring data.

[0060] In this embodiment of the invention, an adsorption capacity sensor, a temperature sensor, and an irradiation dose detector are respectively installed at the upper, middle, and lower parts of the resin column to collect raw signals at each point in real time. The raw signals are converted from analog to digital to generate time series data, and the timestamps of the three points are synchronized and aligned. Then, a sliding window method is used to detect and remove outliers from the synchronized data. Therefore, this invention overcomes the technical problems of existing technologies, such as the single data collection point of radioactive separation resin monitoring, the inability of raw signals to be directly used for subsequent calculations, the asynchrony of data at different points, and the presence of outliers in the data, which lead to inaccurate and unreliable monitoring data and the inability to provide effective basic data for resin lifetime prediction. This invention achieves the acquisition of accurate, synchronized, and reliable pre-processed real-time online monitoring data.

[0061] In a preferred embodiment of the present invention, step 2 above may include:

[0062] Step 2.1: Input the preprocessed real-time online monitoring data into the pre-built microstructure deterioration proxy model. The microstructure deterioration proxy model is a regression model trained using a deep neural network, based on the offline resin detection results under different operating conditions in historical operating data. Specifically, it includes: inputting the preprocessed real-time online monitoring data after acquisition, conversion, synchronization, and anomaly removal into the pre-built microstructure deterioration proxy model. This microstructure deterioration proxy model is a regression model specifically built to characterize the microscopic deterioration state of radioactive separation resin. Its construction and training process specifically involves: collecting historical operating data of the radioactive separation resin under various operating conditions such as historical normal operation, instantaneous strong irradiation, fluctuations in feed acidity, and changes in competing ion concentration. Simultaneously, offline detection results of resin sampling under various operating conditions are obtained. The offline detection results include the initial specific surface area, current specific surface area, initial ion exchange capacity, current ion exchange capacity, pore structure parameters, and functional group content parameters of the resin. Historical operating data is used as the model input sample, and the corresponding offline detection degradation parameters are used as the model output label. A deep neural network is used for training. The deep neural network receives input samples through the input layer, completes feature extraction and nonlinear fitting through multiple hidden layers, and continuously iterates and optimizes the internal weights and biases of the network to gradually reduce and converge the error between the model output value and the actual offline detection value. Finally, a microstructure degradation proxy model that can directly map the resin microstructure degradation state through online monitoring data is obtained.

[0063] Step 2.2: The input data is processed using a microstructure degradation proxy model through multi-layer nonlinear mapping calculations. The microstructure degradation quantification index is output, including a pore collapse coefficient (characterizing the degree of pore collapse within the resin) and a functional group retention rate (characterizing the degree of functional group detachment). The pore collapse coefficient is indirectly calculated using the decay rate of the resin's specific surface area, and the functional group retention rate is indirectly calculated using the decay rate of the ion exchange capacity. Specifically, the microstructure degradation proxy model performs multi-layer nonlinear mapping calculations on the received pre-processed real-time online monitoring data. Through a multi-layer network structure, the intrinsic correlation between local adsorption capacity, local temperature, local irradiation dose, and resin microstructure degradation in the monitoring data is mined and fitted. Finally, the microstructure degradation quantification index of the resin under the current operating conditions is output. The microstructure degradation quantification index includes the pore collapse coefficient and the functional group retention rate. The pore collapse coefficient, used to characterize the degree of pore collapse within the resin, is indirectly calculated using the decay rate of the resin's specific surface area. The specific calculation formula is: K c =1-S÷S0 where K c S is the pore collapse coefficient, S is the current specific surface area of ​​the resin, and S0 is the initial specific surface area of ​​the resin. The larger the value of the pore collapse coefficient, the more severe the pore collapse inside the resin. The functional group retention rate is used to characterize the degree of functional group detachment and is indirectly calculated by the decay rate of ion exchange capacity. The specific calculation formula is: R g =Q÷Q0 where R g Q represents the retention rate of functional groups, Q represents the current ion exchange capacity of the resin, and Q0 represents the initial ion exchange capacity of the resin. The smaller the value of the retention rate of functional groups, the more severe the loss of functional groups from the resin. Through the above calculation process, the deterioration state of the microstructure of the radioactive separation resin can be accurately quantified.

[0064] In this embodiment of the invention, preprocessed real-time online monitoring data is input into a microstructure deterioration proxy model trained by a deep neural network based on historical operating data and offline resin detection results. This model is used to perform multi-layer nonlinear mapping calculations and outputs the pore collapse coefficient and functional group retention rate, which can respectively characterize the degree of resin pore collapse and the degree of functional group detachment. This overcomes the technical problems of existing technologies that cannot quantitatively characterize the microstructure deterioration of radioactive separation resins and that the state of pore collapse and functional group detachment within the resin is difficult to reflect accurately in real time. Thus, online quantification and accurate characterization of the resin microstructure deterioration state are achieved.

[0065] In a preferred embodiment of the present invention, step 3 above may include:

[0066] Step 3.1 involves splicing the local adsorption capacity, local temperature, and local irradiation dose from the preprocessed real-time online monitoring data with the pore collapse coefficient and functional group retention rate to generate an initial fusion feature vector. The initial fusion feature vector is then normalized to obtain a standardized fusion feature set. Specifically, this involves splicing the preprocessed real-time online monitoring data with the calculated microstructure degradation quantification index. The spliced ​​features include the local adsorption capacity, local temperature, and local irradiation dose corresponding to the upper, middle, and lower parts of the resin column axis, as well as the pore collapse coefficient and functional group retention rate characterizing the resin's microstructure degradation state. All these features are then sequentially combined in a fixed order, integrating the dispersed independent features into a continuous one-dimensional feature sequence to generate the initial fusion feature vector, thus achieving a comprehensive fusion of the resin's macroscopic operating state characteristics and microstructure degradation characteristics.

[0067] The generated initial fused feature vector is normalized to eliminate computational interference caused by excessive differences in dimensions and numerical ranges between different features. The min-max normalization method is used to standardize each feature component in the initial fused feature vector. The specific calculation formula is as follows: In the formula, These are the normalized standard eigenvalues. This represents the original value of a feature component in the initial fused feature vector. This is the minimum value of the feature component across all historical samples. This represents the maximum value of the feature component across all historical samples. The above formula is used to normalize each feature component within the initial fused feature vector, mapping all feature values ​​to a uniform range of 0 to 1, ultimately yielding the standardized fused feature set.

[0068] In this embodiment of the invention, a technique is employed to splice preprocessed real-time online monitoring data, including local adsorption capacity, local temperature, local irradiation dose, and microstructure degradation quantification indicators, including pore collapse coefficient and functional group retention rate, into a feature layer to generate an initial fusion feature vector. This initial fusion feature vector is then normalized to obtain a standardized fusion feature set. This overcomes the technical problems in existing technologies, such as the failure to combine macroscopic resin operation monitoring data with microscopic degradation indicators, the single feature dimension, and inconsistent feature dimensions leading to non-standardized subsequent model inputs, low feature utilization, and impact on lifetime prediction accuracy. Ultimately, this achieves the effect of integrating the macroscopic operating state and microscopic degradation characteristics of the resin, eliminating differences in feature dimensions, and obtaining a standardized and comprehensive fusion feature set.

[0069] In a preferred embodiment of the present invention, step 4 above may include:

[0070] Step 4.1: Input the standardized fused feature set into the extreme condition generalization enhanced machine learning model. The model is built on a deep neural network. It receives the fused feature set through the input layer, performs nonlinear feature transformation and mapping through multiple hidden layers, and finally generates a one-dimensional value from the output layer as the baseline remaining lifetime prediction value under the current operating conditions. The extreme condition generalization enhanced machine learning model is optimized during the training phase by introducing adversarial samples and physical consistency constraints. Specifically, the extreme condition generalization enhanced machine learning model is a regression prediction model built on a deep neural network, specifically designed to address the problems of insufficient lifetime prediction generalization ability and large prediction deviations of radioactive separation resins under extreme conditions such as instantaneous strong radiation, sudden changes in feed acidity, and sudden changes in competing ion concentration. This model adopts a multi-layer deep neural network structure, which from top to bottom includes an input layer, multiple hidden layers, and an output layer. The functions and structures of each layer are as follows:

[0071] Input layer: Used to receive the standardized fusion feature set. The number of neurons in the input layer is completely consistent with the feature dimensions of the fusion feature set, ensuring complete reception of all fusion features, including the local adsorption capacity, local temperature, local irradiation dose of the upper, middle and lower parts of the resin column, as well as the pore collapse coefficient and functional group retention rate, and converting the standardized feature data into an input signal that can be recognized by the neural network.

[0072] Hidden Layers: Multiple consecutive hidden layers are set, each containing several neurons. These layers are connected via fully connected layers. The core function is to perform nonlinear feature transformation and deep feature mining on the fused input features, capturing the intrinsic correlation between different features and the remaining lifespan of the resin. Features include macroscopic operational features and microscopic degradation features. The output calculation process for each hidden layer is as follows: Let the input of a certain hidden layer be... The weight matrix of the neurons in this layer is: The bias vector is The ReLU activation function is used. After performing a nonlinear transformation, the output H of this hidden layer is calculated using the following formula: In the formula, This represents a matrix multiplication operation between the weight matrix and the input vector. Specifically, it involves multiplying each feature component of the input vector by its corresponding weight, summing all the products, and then multiplying the sum by the bias vector. Perform addition to obtain a linear output result; The ReLU activation function is used to introduce nonlinear features and eliminate the limitations of linear transformation. Its calculation rule is that when the input value is greater than 0, the output value is equal to the input value; when the input value is less than or equal to 0, the output value is 0. Multiple hidden layers perform the above operation in sequence to gradually complete the deep extraction and nonlinear mapping of features, and transform the original fused features into high-order features that directly reflect the remaining life of the resin.

[0073] Output layer: Employs a single neuron structure to output the baseline remaining lifetime prediction. The calculation process of the output layer is as follows, assuming the output of the last hidden layer is... The weights of the output layer are , bias is The output layer does not use an activation function (to maintain linear output), and the baseline remaining lifetime prediction value is... The calculation formula is: In the formula, This represents the product of the output layer weights and the output of the last hidden layer, then multiplied by the bias. Adding them together yields a one-dimensional value, which is the baseline remaining life prediction of the resin under the current operating conditions.

[0074] The core of model training is to iteratively optimize the weights and biases of each layer through a large number of samples, while introducing adversarial samples and physical consistency constraints to improve the model's generalization ability and prediction reliability under extreme conditions. The specific process is as follows:

[0075] Training sample preparation: Historical data of radioactive separation resin under various operating conditions, including normal operating conditions, instantaneous strong irradiation operating conditions, sudden change in feed acidity operating conditions, and sudden change in competing ion concentration operating conditions, were collected. The historical data includes real-time online monitoring data after pretreatment, quantitative indicators of microstructure deterioration, and the corresponding actual remaining life of the resin, which were determined through offline detection and actual service records. The historical data were divided into training set, validation set, and test set in a ratio of 7:2:1 for model training, parameter optimization, and performance verification.

[0076] Adversarial sample generation: To improve the model's generalization ability under extreme conditions, adversarial samples are introduced during the training phase. The specific generation process is as follows: The original fused feature set in the training set... Add a small perturbation ΔX to generate adversarial examples The calculation formula is: In the formula, This is the perturbation coefficient, ranging from 0.01 to 0.05. It is used to control the magnitude of the perturbation and prevent excessive perturbation from distorting the samples. For loss function For input features gradient, This is a sign function; it outputs 1 when the gradient is positive, -1 when the gradient is negative, and 0 when the gradient is zero. These are the model's predicted values. To determine the actual remaining lifespan of the resin, adversarial samples are generated to allow the model to learn abnormal features under extreme conditions during training, thereby improving the model's adaptability to extreme conditions.

[0077] Physical consistency constraints are introduced: To ensure that the baseline remaining lifetime predictions output by the model conform to physical laws, such as the remaining lifetime of resin not being negative and the remaining lifetime decreasing as degradation progresses, a physical consistency constraint term is introduced during the training phase and added to the loss function to construct a new loss function. The calculation formula is: In the formula, The mean squared error loss function measures the deviation between the model's predicted values ​​and the actual values. Its calculation formula is as follows: , The number of training samples. Indicates all Sum the squared deviations of each sample. The constraint coefficients range from 0.1 to 0.3 and are used to adjust the weights of the constraint terms. For physical consistency constraints, the calculation rule is: when When <0, =∣ |, penalizing negative lifetime predictions; when > hour, = - The penalty exceeded the predicted value of the resin's initial life. For the initial design life of the resin; when As the degree of microstructure deterioration increases, = - The predicted value that is opposite to the trend of punishment and deterioration. The predicted value is from the previous time step; otherwise... =0.

[0078] Model iterative training: Input the training set samples, including the original samples and adversarial samples, into the model, and calculate the loss function using the backpropagation algorithm. The weights of each hidden layer and output layer are continuously updated iteratively. Bias until the loss function Once the model converges to a preset threshold, such as 0.001, training is complete. Then, the model's hyperparameters, such as the number of hidden layers, the number of neurons, and the perturbation coefficients, are adjusted using a validation set. Constraint coefficient This ensures the model's generalization ability; finally, the model's performance is verified through a test set. When the prediction error is less than a preset value, such as 5%, the model is officially put into use, which is the pre-trained extreme condition generalization enhancement machine learning model.

[0079] The standardized fused feature set is then input into a pre-trained extreme condition generalization enhancement machine learning model through the model input layer. The input layer converts the standardized feature data into electrical signals recognizable by the neural network and transmits them to the first hidden layer. The first hidden layer then... The calculation formula performs linear operations and nonlinear activation transformations on the input features to obtain the output features of the first hidden layer. This output feature is then passed to the next hidden layer, and so on, sequentially completing nonlinear feature transformations and deep feature mining for all hidden layers. Redundant features are gradually removed, and core high-order features directly related to the remaining lifespan of the resin are extracted. Finally, the output of the last hidden layer is... Transmitted to the output layer; the output layer receives the output of the last hidden layer. ,according to The calculation formula is used to perform linear operations to generate a one-dimensional value, which is the baseline remaining lifetime prediction value of the radioactive separation resin under the current operating conditions. This baseline remaining lifetime prediction value comprehensively considers the macroscopic operating state of the resin, including local adsorption capacity, local temperature, local irradiation dose, and microscopic deterioration state, including pore collapse coefficient and functional group retention rate. It has been optimized by adversarial sample and physical consistency constraints to achieve accurate prediction under normal operating conditions.

[0080] In this embodiment of the invention, the standardized fusion feature set is input into an extreme condition generalization enhancement machine learning model constructed based on a deep neural network and optimized by introducing adversarial samples and physical consistency constraints during the training phase. Through multi-layer nonlinear feature transformation and mapping, the model outputs a baseline remaining lifetime prediction value. This overcomes the technical problems of traditional models, such as poor generalization ability under extreme conditions, lack of physical rationality constraints, and insufficient prediction accuracy and reliability. As a result, the model's adaptability and prediction accuracy under complex conditions are improved, and a stable, reliable, and physically consistent baseline remaining lifetime prediction value is output.

[0081] In a preferred embodiment of the present invention, step 5 above may include:

[0082] Step 5.1: Real-time determination of whether the current operating condition belongs to extreme conditions, including instantaneous strong radiation, sudden changes in feed acidity, and sudden changes in competing ion concentration. If so, normalize the local adsorption capacity, local temperature, and local irradiation dose at the three points to obtain normalized values ​​for adsorption capacity, temperature, and irradiation dose. Weight the normalized values ​​for adsorption capacity, temperature, and irradiation dose according to preset weights to calculate the comprehensive state index value for each point. Specifically, after real-time determination that the current operating condition belongs to extreme conditions of instantaneous strong radiation, sudden changes in feed acidity, and sudden changes in competing ion concentration, first, the upper part of the resin column, The local adsorption capacity, local temperature, and local irradiation dose at the three locations in the middle and lower parts were subjected to minimum-maximum normalization to eliminate calculation interference caused by different dimensions and large differences in numerical ranges among different monitoring parameters. The normalization calculation process for each parameter is to obtain the normalized value of a certain monitoring parameter, which is equal to the real-time acquired value of the parameter minus the minimum value of the parameter in the historical operation process, and then divided by the maximum value of the parameter in the historical operation process minus the minimum value of the parameter in the historical operation process. Through the above calculation rules, the normalized values ​​of adsorption capacity, temperature, and irradiation dose for the three locations are obtained respectively.

[0083] After obtaining the normalized values ​​of each parameter, the three normalized values ​​for the same location are weighted and summed according to preset weighting coefficients to calculate the comprehensive state index value of that location. The weighted summation formula is as follows: In the formula, This represents the comprehensive status index value for the current location. The preset weighting coefficients are the values ​​corresponding to the local adsorption capacity. This is the normalized value of the adsorption capacity at the current location. These are the preset weighting coefficients corresponding to local temperatures. This is the normalized temperature value at the current location. The preset weighting coefficients corresponding to the local irradiation dose. The normalized value of the irradiation dose at the current point is given, and the sum of all weight coefficients is 1. Following the above calculation process, the comprehensive state index values ​​of the three points at the top, middle and bottom of the resin column are calculated respectively, and the comprehensive state index values ​​corresponding to the three points are obtained.

[0084] Step 5.2: Based on the comprehensive state index value, using the axial coordinates corresponding to the upper, middle, and lower parts of the resin column as interpolation nodes, and the comprehensive state index value of each node as the node function value, a continuous state distribution curve along the axial direction of the resin column is constructed using the cubic spline interpolation method. Specifically, a cubic polynomial is constructed between every two adjacent nodes, ensuring that all cubic polynomials have the same function value, first derivative value, and second derivative value at the nodes. By solving the three-moment equation system established by the continuity condition of the second derivative at the nodes, the coefficients of each piecewise cubic polynomial are determined, thus obtaining the state distribution curve composed of piecewise cubic polynomials. Specifically, using the three uniformly distributed monitoring points at the upper, middle, and lower parts of the resin column as interpolation nodes, an independent cubic polynomial is constructed between every two adjacent nodes. Since there are three interpolation nodes, a total of two cubic polynomials are constructed. The first cubic polynomial corresponds to the axial interval from the upper node to the middle node, and the second cubic polynomial corresponds to the axial interval from the middle node to the lower node. The two cubic polynomials are combined to form a complete axial state distribution curve. During the construction of the polynomial, it is necessary to ensure that the two cubic polynomials meet the constraint requirement of continuous smoothness at the middle node. The relevant calculation and constraint rules are as follows: First, the state values ​​calculated by the two polynomials at the middle node are completely equal, and the values ​​are consistent with the comprehensive state index value at the middle point. Second, the slope values ​​of the curves obtained by performing first-order derivative calculations on the two polynomials are completely equal at the middle node, ensuring that there are no abrupt changes in the curve transition. Third, the curvature change rate values ​​of the curves obtained by performing second-order derivative calculations on the two polynomials are completely equal at the middle node, improving the overall smoothness of the curve. At the same time, in order to ensure the stability of the two ends of the curve, the second-order derivative results at the upper and lower nodes are set to be zero, that is, the curvature change rate is zero, which is used as the boundary constraint condition.

[0085] To determine all the unknown coefficients of the two cubic polynomials, the second derivative of each polynomial is first calculated. The rate of change of curvature at each node is obtained through the second derivative and used as the bending moment parameter. According to the boundary constraints, the bending moment parameters corresponding to the upper and lower nodes are both zero; only the bending moment parameter corresponding to the middle node needs to be calculated. Based on the constraint of continuous rate of change of curvature at each node and the requirement of matching node state values, a three-moment calculation relationship is established. Since only three interpolation nodes are used, this calculation relationship can be simplified to a single linear calculation rule, where adjacent nodes... Substituting the condition of equal axial spacing between nodes into the linear calculation rule, a simplified calculation is performed: First, the right side of the linear calculation rule is calculated by subtracting the comprehensive state index of the middle node from the comprehensive state index of the lower node, and then subtracting the comprehensive state index of the upper node from the comprehensive state index of the middle node, and multiplying by a fixed coefficient to obtain the calculated value on the right side; the left side of the linear calculation rule is the axial spacing between adjacent nodes multiplied by a fixed coefficient, and then multiplied by the bending moment parameter of the middle node; finally, the calculated value on the right side is divided by the fixed product on the left side to obtain the bending moment parameter value of the middle node. At this point, the bending moment parameters of all three interpolation nodes are determined.

[0086] After obtaining all bending moment parameter values, the bending moment parameters, axial spacing between adjacent nodes, axial coordinates of each node, and comprehensive state index values ​​of each node are substituted into the fixed coefficient calculation rules of the piecewise cubic polynomial to complete the coefficient calculation item by item: First, for the first segment of the cubic polynomial from the upper to the middle section, the coefficient of each term is calculated sequentially, and the corresponding parameters are substituted into the rules to determine all unknown parameters through addition, subtraction, multiplication, and division operations; then, for the second segment of the cubic polynomial from the middle to the lower section, the same calculation method is used, substituting the corresponding parameters to complete the calculation of all coefficients, and finally determining all coefficients of the two segments of the cubic polynomial; in the two segments of the cubic polynomial After all coefficients are calculated and determined, the specific expression of the two polynomials is completely determined. Combining the two polynomials yields a state distribution curve that is continuously distributed along the resin column axis. This curve covers the entire axial range from the top to the bottom of the resin column. The state values ​​are calculated using the first cubic polynomial from the top to the middle section, and the state values ​​are calculated using the second cubic polynomial from the middle to the bottom section. The two curves satisfy the following conditions at the middle node: continuous state values, continuous slope values, and continuous rate of change of curvature values. Through all the above calculations and constraints, a smooth, continuous state distribution curve that can truly reflect the axial deterioration distribution of the resin is finally obtained.

[0087] In this embodiment of the invention, the local adsorption capacity, local temperature, and local irradiation dose at three points are normalized and weighted to obtain a comprehensive state index value. The axial continuous state distribution curve is constructed by using the axial coordinate of the resin column as a node and cubic spline interpolation. Therefore, the technical problems of discrete monitoring points being unable to reflect the continuous axial degradation distribution of the resin column and the spatial state being incomplete and unsmooth are overcome. Thus, the technical effect of continuously, accurately, and smoothly characterizing the axial service state of the resin column is achieved.

[0088] In a preferred embodiment of the present invention, step 6 above may include:

[0089] Step 6.1: Calculate the curvature of the state distribution curve to obtain the curvature value corresponding to each point on the curve; perform extreme point detection on the curvature values, mark the axial position corresponding to the local maximum curvature point as the potential deterioration mutation point, and use the potential deterioration mutation point as the inflection point of the state distribution curve. Specifically, this includes: based on the constructed state distribution curve that is continuous along the axial direction of the resin column, which is composed of piecewise cubic polynomials, calculate the curvature value of each point on the curve. The curvature characterizes the degree of change in the axial state of the resin. The larger the curvature, the more significant the abrupt change in the resin deterioration state at that location. Then, locate the deterioration mutation point, i.e., the inflection point, through extreme point detection. The state distribution curve consists of two segments of cubic polynomials, from the upper part to the middle of the resin column: From middle to lower part: ,in, The coefficient of the cubic term determines the weight of the cubic term in the curve, directly affecting the curvature and overall trend of the curve. The coefficients of the quadratic term represent the parabolic curvature of the dominant curve, reflecting the second-order trend of state change. The coefficient of the first term determines the linear slope of the curve and reflects the first-order trend of state change; The constant term represents the initial value of the resin deterioration state at the starting position of this curve segment. Taking the first and second derivatives of the two polynomials respectively reflects the rate of change of the curve's slope:

[0090] The first polynomial, , This represents the upper axial coordinate. For the central axial coordinates: First derivative slope: Rate of change of the slope of the second derivative: .

[0091] Second polynomial , (Lower axial coordinates): First derivative (slope): Second derivative (rate of change of slope): .

[0092] Based on the first and second derivatives, and using the formula for calculating the curvature of a plane curve, the curvature of any point on the curve is calculated. curvature value The formula is: In the formula, The absolute value of the second derivative. It is the square of the first derivative. Taking the cube root of the result of adding the square of the first derivative to 1 (i.e., the result raised to the power of 1.5), and then calculating the state distribution curve point by point along the entire axial range of the resin column using the above formula,... The curvature values ​​at all points within the curve are used to obtain the curvature curve distributed along the axial direction. The larger the curvature value, the more significant the change in the service condition of the resin at that axial position from normal to deterioration.

[0093] For curvature curve Local maxima detection is performed to identify key locations of abrupt changes in resin axial degradation. The specific process is as follows: A sliding window method is used to traverse the curvature curve, with the window length set to 5% of the total axial length of the resin column to ensure detection accuracy and stability. The maximum curvature value and corresponding axial position within each window are calculated. The detected local maxima are then filtered. If a local maxima is greater than twice the average curvature value within its two adjacent windows, it is determined to be a local curvature maximum, and its corresponding axial position is marked as a potential degradation abrupt change point. All potential degradation abrupt change points are identified as inflection points on the state distribution curve. Inflection points are the core locations where abrupt changes occur in the axial degradation state of the resin.

[0094] Step 6.2: Divide the resin column axially into several continuous deterioration segments based on the inflection points. Integrate the curvature value within each deterioration segment to obtain the cumulative deterioration intensity of the segment. Use the ratio of the cumulative deterioration intensity to the segment length as the local deterioration correction factor for the segment. Weight the local deterioration correction factors of each segment according to the proportion of each deterioration segment to the total height of the resin column to obtain the comprehensive spatial correction coefficient. Specifically, based on all the obtained inflection points, divide the resin column axially into continuous segments according to the axial coordinate position of the resin column corresponding to the inflection points. If multiple inflection points are identified, divide the resin column axially into multiple continuous deterioration segments sequentially from top to bottom. The first segment is the area between the upper part of the resin column and the first inflection point; the second segment is the area between the first and second inflection points; and so on for subsequent segments, with the last segment being the area between the last inflection point and the lower part of the resin column. If no inflection point is detected, the entire axial region of the resin column from top to bottom is directly considered as a complete deterioration segment. For each deterioration segment, a definite integral is performed on the curvature values ​​corresponding to all axial positions within the segment. The result of the integral is the cumulative deterioration intensity of the deterioration segment. The cumulative deterioration intensity is used to characterize the total degree of resin deterioration abrupt change within the corresponding segment. The specific integration process is as follows: using the starting axial coordinate of the deterioration segment as the lower limit of integration and the ending axial coordinate of the deterioration segment as the upper limit of integration, the curvature values ​​at any axial position within the segment are continuously integrated. If the direct analytical solution of the integral result is too complex, the trapezoidal numerical integration method can be used. The entire deterioration segment is divided into multiple small axial steps. The product of the curvature value within each small step and the corresponding step length is calculated segment by segment. Then, all the product results are added together in sequence. The approximate sum is the cumulative deterioration intensity of the deterioration segment.

[0095] The cumulative degradation intensity of each degradation segment is divided by the axial length of the corresponding degradation segment. The resulting ratio is the local degradation correction factor for that degradation segment. The local degradation correction factor characterizes the average degradation abrupt change intensity per unit axial length within the corresponding segment. The axial length of the degradation segment is the difference between the ending axial coordinate and the starting axial coordinate of the segment. A larger local degradation correction factor indicates a more severe degradation abrupt change per unit length of resin within the degradation segment, and a greater impact on the overall remaining life of the resin. First, the proportion of the axial length of each degradation segment to the total axial height of the resin column is calculated. The total axial height of the resin column is the difference between the lower and upper axial coordinates of the resin column. Then, the length proportion corresponding to each degradation segment is multiplied by the local degradation correction factor for that segment to obtain the weighted result for the corresponding segment. Finally, the weighted results of all deteriorated sections are summed sequentially, and the summation value is the comprehensive spatial correction coefficient. The comprehensive spatial correction coefficient is used to characterize the comprehensive impact of the axial deterioration mutation of the entire resin column. The value range of the comprehensive spatial correction coefficient is limited to between zero and one. If the calculated value is greater than one, the comprehensive spatial correction coefficient is directly set to one to avoid over-correction of the remaining life prediction value. The larger the value of the comprehensive spatial correction coefficient, the more significant the axial deterioration mutation phenomenon of the resin column is, and the greater the correction of the baseline remaining life prediction value is required.

[0096] Step 6.3: Multiply the comprehensive spatial correction coefficient by the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value after spatial distribution characteristic correction. If it is determined in real time that the current operating condition does not belong to the extreme operating condition, the baseline remaining lifetime prediction value is directly output as the final remaining lifetime prediction value. Specifically, this includes: collecting the current operating parameters of the radioactive separation resin in real time, mainly including irradiation dose, feed acidity, and competing ion concentration. The collected real-time operating parameters are compared with the pre-set extreme operating condition judgment thresholds. The extreme operating condition judgment thresholds are: instantaneous irradiation dose exceeding twice the normal operating value, feed acidity change exceeding ±0.5 pH units, and competing ion concentration change exceeding 30% of the normal operating value. Based on the comparison results, it is determined whether the current operating condition belongs to the extreme operating condition of instantaneous strong irradiation, feed acidity change, and competing ion concentration change.

[0097] If the current operating condition is determined to be an extreme condition, the calculated baseline remaining lifetime prediction is multiplied by the comprehensive spatial correction coefficient. The result is the final remaining lifetime prediction after correction for the resin's axial spatial distribution characteristics. This correction process incorporates the impact of uneven axial spatial degradation distribution of the resin column into the lifetime prediction system, solving the problem of low lifetime prediction accuracy in existing technologies due to the lack of consideration for uneven axial degradation. If the current operating condition is determined not to be an extreme condition, it indicates that the axial degradation distribution of the resin column is relatively uniform, and no spatial correction is required. The obtained baseline remaining lifetime prediction is directly output as the final remaining lifetime prediction. The final output remaining lifetime prediction comprehensively considers the macroscopic operating state of the resin, the degree of microstructural degradation, and the characteristics of uneven axial spatial degradation distribution under extreme conditions, improving the accuracy of remaining lifetime prediction for radioactive separation resins.

[0098] In this embodiment of the invention, the curvature of the state distribution curve is calculated and inflection point detection is performed to mark the deterioration mutation point. Based on the inflection point, the deterioration section is divided and the cumulative deterioration intensity and local deterioration correction factor are calculated. The comprehensive spatial correction coefficient is obtained by weighting and the baseline remaining life prediction value is corrected. The technical means of directly outputting the baseline prediction value under non-extreme working conditions overcomes the technical problems of being unable to identify the axial deterioration mutation area of ​​the resin column, difficulty in quantifying the local spatial deterioration difference, inability to combine spatial distribution for correction under extreme working conditions, and large deviation between the prediction result and the actual deterioration state. Thus, it achieves the technical effects of accurately locating the axial deterioration mutation position of the resin, quantifying the deterioration intensity of each section, realizing accurate spatial correction of the remaining life under extreme working conditions, and improving the accuracy and reliability of resin life prediction.

[0099] like Figure 2 As shown, embodiments of the present invention also provide a machine learning-based system for predicting the lifetime of radioactive separation resins, comprising:

[0100] The acquisition module is used to acquire real-time online monitoring data of the radioactive separation resin during operation. The real-time online monitoring data includes local adsorption capacity, local temperature and local irradiation dose collected from three points distributed axially along the resin column: the upper, middle and lower parts.

[0101] The calculation module is used to calculate the quantitative index of microstructure deterioration of resin under the current working conditions based on real-time online monitoring data and through the constructed microstructure deterioration proxy model.

[0102] The fusion module is used to fuse real-time online monitoring data with microstructure deterioration quantification indicators to construct a fusion feature set.

[0103] The prediction module is used to input the fused feature set into a pre-trained extreme condition generalized enhanced machine learning model to obtain the baseline remaining life prediction value of the resin under the current working conditions.

[0104] The judgment module is used to determine in real time whether the current working condition belongs to an extreme working condition. Extreme working conditions include instantaneous strong radiation, sudden change in acidity of the feed solution, and sudden change in competitive ion concentration. If it belongs to an extreme working condition, the state distribution curve along the axial direction of the resin column is constructed based on the local adsorption capacity, local temperature, and local radiation dose at the upper, middle, and lower points.

[0105] The correction module is used to calculate the curvature and detect inflection points of the state distribution curve, obtain the local degradation correction factor of each axial segment, generate a comprehensive spatial correction coefficient through weighted fusion, and use the comprehensive spatial correction coefficient to correct the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value; if it does not belong to the category, the baseline remaining lifetime prediction value is directly used as the final remaining lifetime prediction value.

[0106] 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.

[0107] 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.

[0108] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0109] 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 predicting the lifetime of radioactive separation resins based on machine learning, characterized in that, The method includes: Real-time online monitoring data of the radioactive separation resin during operation is obtained. The real-time online monitoring data includes local adsorption capacity, local temperature and local irradiation dose collected from three points distributed axially along the resin column: the upper, middle and lower parts. Based on real-time online monitoring data, the microstructure deterioration surrogate model was constructed to calculate the quantitative index of resin microstructure deterioration under the current working conditions. Real-time online monitoring data is fused with quantitative indicators of microstructure deterioration to construct a fused feature set; The fused feature set is input into a pre-trained extreme condition generalized enhanced machine learning model to obtain the baseline remaining life prediction value of the resin under the current working condition. In real time, it is determined whether the current working condition belongs to an extreme working condition. Extreme working conditions include instantaneous strong radiation, sudden change in acidity of the feed solution, and sudden change in competitive ion concentration. If it belongs to an extreme working condition, a state distribution curve along the axial direction of the resin column is constructed based on the local adsorption capacity, local temperature, and local irradiation dose at three points: the upper, middle, and lower parts. The curvature of the state distribution curve is calculated to obtain the curvature value corresponding to each point on the curve; extreme points of curvature are detected, and the axial positions corresponding to the local maxima of curvature are marked as potential deterioration mutation points, and the potential deterioration mutation points are used as inflection points of the state distribution curve. The resin column is divided into several continuous deterioration sections along the axial direction based on the inflection point. The curvature value is integrated within each deterioration section to obtain the cumulative deterioration intensity of the section. The ratio of the cumulative deterioration intensity to the section length is used as the local deterioration correction factor of the section. The local deterioration correction factors of each section are weighted and summed according to the proportion of each deterioration section to the total height of the resin column to obtain the comprehensive spatial correction coefficient. Multiply the comprehensive spatial correction coefficient by the baseline remaining life prediction value to obtain the final remaining life prediction value after spatial distribution characteristic correction; if it is determined in real time that the current operating condition is not an extreme operating condition, the baseline remaining life prediction value is directly output as the final remaining life prediction value.

2. The method for predicting the lifetime of radioactive separation resin based on machine learning according to claim 1, characterized in that, Real-time online monitoring data of the radioactive separation resin during operation is obtained. This data includes local adsorption capacity, local temperature, and local irradiation dose collected from three points axially distributed along the resin column: the upper, middle, and lower sections. Adsorption capacity sensor, temperature sensor and irradiation dose detector are respectively installed at the upper, middle and lower parts of the resin column to collect the original signals of local adsorption capacity, local temperature and local irradiation dose at each point in real time. The acquired raw signals are converted from analog to digital to generate corresponding time series data. The time series data of local adsorption capacity, local temperature and local irradiation dose at the three points are synchronized and aligned with the timestamps to obtain synchronized data. The synchronized data is processed using a sliding window method to detect and remove outliers, resulting in preprocessed real-time online monitoring data.

3. The method for predicting the lifetime of radioactive separation resin based on machine learning according to claim 2, characterized in that, Based on real-time online monitoring data, a microstructure deterioration proxy model was constructed to calculate quantitative indicators of resin microstructure deterioration under current operating conditions, including: The pre-processed real-time online monitoring data is input into the pre-built microstructure deterioration proxy model. The microstructure deterioration proxy model is a regression model trained by a deep neural network based on the resin offline detection results under different working conditions in historical operating data. The input data is subjected to multi-layer nonlinear mapping calculation by a microstructure degradation proxy model, and the output microstructure degradation quantification index includes the pore collapse coefficient, which characterizes the degree of pore collapse in the resin, and the functional group retention rate, which characterizes the degree of functional group detachment. The pore collapse coefficient is indirectly calculated by the decay rate of the resin specific surface area, and the functional group retention rate is indirectly calculated by the decay rate of the ion exchange capacity.

4. The method for predicting the lifetime of radioactive separation resin based on machine learning according to claim 3, characterized in that, Real-time online monitoring data is fused with quantitative indicators of microstructural degradation to construct a fused feature set, including: The local adsorption capacity, local temperature, and local irradiation dose in the preprocessed real-time online monitoring data are spliced ​​with the pore collapse coefficient and functional group retention rate in a feature layer to generate an initial fusion feature vector; the initial fusion feature vector is then normalized to obtain a standardized fusion feature set.

5. The method for predicting the lifetime of radioactive separation resin based on machine learning according to claim 4, characterized in that, The fused feature set is input into a pre-trained extreme condition generalization-enhanced machine learning model to obtain the baseline remaining life prediction of the resin under the current operating conditions, including: The standardized fusion feature set is input into the extreme condition generalized enhanced machine learning model. The model is built on a deep neural network. It receives the fusion feature set through the input layer, performs nonlinear feature transformation and mapping through multiple hidden layers, and finally generates a one-dimensional value from the output layer as the baseline remaining life prediction value under the current operating condition. The extreme condition generalized enhanced machine learning model is optimized by introducing adversarial samples and physical consistency constraints during the training phase.

6. The method for predicting the lifetime of radioactive separation resin based on machine learning according to claim 5, characterized in that, Real-time determination of whether the current operating condition belongs to extreme conditions, including instantaneous strong radiation, sudden changes in feed solution acidity, and sudden changes in competing ion concentration; if so, a state distribution curve along the resin column axis is constructed based on the local adsorption capacity, local temperature, and local irradiation dose at three points: the upper, middle, and lower sections. The system determines in real time whether the current operating condition is an extreme condition, which includes instantaneous strong radiation, sudden changes in feed solution acidity, and sudden changes in competing ion concentration. If it is, the local adsorption capacity, local temperature, and local radiation dose at the three points are normalized to obtain normalized values ​​for adsorption capacity, temperature, and radiation dose. The normalized values ​​for adsorption capacity, temperature, and radiation dose are then weighted and summed according to preset weights to calculate the comprehensive state index value for each point. Based on the comprehensive state index value, the axial coordinates corresponding to the upper, middle and lower parts of the resin column are used as interpolation nodes, and the comprehensive state index value of each node is used as the node function value. A continuous state distribution curve along the axial direction of the resin column is constructed using the cubic spline interpolation method. Specifically, a cubic polynomial is constructed between every two adjacent nodes, and all cubic polynomials have the same function value, first derivative value and second derivative value at the nodes. By solving the three moment equations established by the continuity condition of the second derivative at the nodes, the coefficients of each piecewise cubic polynomial are determined, thereby obtaining the state distribution curve composed of piecewise cubic polynomials.

7. A machine learning-based system for predicting the lifetime of radioactive separation resins, 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 acquire real-time online monitoring data of the radioactive separation resin during operation. The real-time online monitoring data includes local adsorption capacity, local temperature and local irradiation dose collected from three points distributed axially along the resin column: the upper, middle and lower parts. The calculation module is used to calculate the quantitative index of microstructure deterioration of resin under the current working conditions based on real-time online monitoring data and through the constructed microstructure deterioration proxy model. The fusion module is used to fuse real-time online monitoring data with microstructure deterioration quantification indicators to construct a fusion feature set. The prediction module is used to input the fused feature set into a pre-trained extreme condition generalized enhanced machine learning model to obtain the baseline remaining life prediction value of the resin under the current working conditions. The judgment module is used to determine in real time whether the current working condition belongs to an extreme working condition. Extreme working conditions include instantaneous strong radiation, sudden change in acidity of the feed solution, and sudden change in competitive ion concentration. If it belongs to an extreme working condition, the state distribution curve along the axial direction of the resin column is constructed based on the local adsorption capacity, local temperature, and local radiation dose at the upper, middle, and lower points. The correction module is used to calculate the curvature and detect inflection points of the state distribution curve, obtain the local degradation correction factor of each axial segment, generate a comprehensive spatial correction coefficient through weighted fusion, and use the comprehensive spatial correction coefficient to correct the baseline remaining lifetime prediction value to obtain the final remaining lifetime prediction value; if it does not belong to the category, the baseline remaining lifetime prediction value is directly used as the final remaining lifetime prediction value.

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.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images