Rare earth ion chromatography online analysis and detection method and system

Abstract

The application relates to the technical field of analytical chemistry detection, and specifically discloses a rare earth ion chromatographic online analysis and detection method and system, which adopts a chromatography-mass spectrometry combined system, realizes multi-scale time-frequency decomposition of signals by constructing an adaptive wavelet base function library, accurately identifies and classifies interference types in combination with a convolutional neural network and vacuum degree coupling analysis; a U-Net generator with an attention mechanism and a multi-scale discriminator are designed for signal repair aiming at repairable interference; a state transition model containing a mass-to-charge ratio database is established, empirical mode decomposition and adaptive filtering technology are used to realize dynamic calibration of a mass axis; quantitative accuracy is ensured through double internal standard correction and a triple verification mechanism; when unrepairable interference is detected, a hierarchical self-checking program is started, and fault diagnosis is carried out in combination with spectrum fingerprint analysis.

Description

Technical Field

[0001] This invention relates to the field of analytical chemistry detection technology, specifically to an online analytical detection method and system for rare earth ion chromatography. Background Technology

[0002] Rare earth elements, as important strategic resources, are widely used in new energy, electronic information, and national defense industries. With the increasing demand for rare earth applications, the requirements for the accuracy and real-time performance of their component analysis are constantly rising. Chromatographic analysis technology, due to its excellent separation performance, has become one of the main methods for rare earth component detection. However, in actual industrial testing environments, factors such as instrument electromagnetic interference, plasma instability, and complex sample matrices seriously affect the reliability of analytical results. Especially in online continuous detection scenarios, traditional offline analysis methods struggle to meet the demands of real-time monitoring.

[0003] The existing technology has the following shortcomings:

[0004] Electromagnetic pulse interference (EMP) causes instability in the radio frequency voltage of the quadrupole mass analyzer, resulting in a mass axis shift and misidentification of the mass-to-charge ratio of the target rare earth ions. On the other hand, the quenching reaction between CO2 and plasma leads to a sharp drop in electron density and a significant decrease in ionization efficiency, creating periodic signal blind zones. These interferences prevent the accurate capture of the rare earth ion current signal after chromatographic separation, causing concentration misjudgments in online monitoring systems and severely impacting the reliability of detection in high-EMP industrial environments (such as rare earth metal arc melting and NdFeB sintering). Existing technologies can only partially suppress high-frequency noise interference and lack effective countermeasures against low-frequency EMP and chemical interference (COx). There is an urgent need to develop novel anti-interference detection methods and adaptive calibration techniques to solve this technical challenge. Summary of the Invention

[0005] The purpose of this invention is to provide an online analysis and detection method and system for rare earth ion chromatography to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A rare earth ion chromatography online analysis and detection method includes the following steps:

[0008] S1: Real-time acquisition of rare earth ion flow signals after rare earth ion chromatography separation;

[0009] S2: Perform multi-scale time-frequency decomposition on the acquired rare earth ion current signal, extract the interference feature interval in the rare earth ion current signal, and classify the interference type into repairable interference and unrepairable interference according to the interference features.

[0010] S3: For repairable interference, construct a signal repair model, process signal segments containing interference features through adversarial training, and output the repaired pure ion flow signal.

[0011] S4: Establish a state transition model based on historical signal quality offset data, and correct the quality axis deviation of the current detected signal in real time according to the type of interference;

[0012] The historical signal quality offset data includes: a pre-established rare earth ion mass-charge ratio database and real-time detected mass axis offset records;

[0013] S5: Match the repaired signal with the chromatographic retention time, calculate the concentration of each rare earth ion, and optimize the signal reconstruction parameters based on the repair effect.

[0014] S6: When the number of consecutive occurrences of identified unrepairable interference exceeds a preset threshold, the system self-check process is triggered and an abnormal warning is output.

[0015] As a further aspect of the present invention: the multi-scale time-frequency decomposition of the acquired rare-earth ion current signal specifically includes:

[0016] An adaptive wavelet basis function library is constructed, and the optimal wavelet basis is dynamically selected based on the characteristic absorption peak positions of rare earth elements. Symmetric compactly supported wavelets are used for cerium group elements, and asymmetric oscillatory wavelets are used for yttrium group elements.

[0017] The low-frequency channel extracts baseline drift features through 7-layer wavelet packet decomposition, while the high-frequency channel captures transient pulse interference through an improved S-transform. The time-frequency resolution is dynamically adjusted according to the signal-to-noise ratio.

[0018] Calculate the energy distribution dispersion of each sub-band at each scale, and mark an abnormal interval when the energy entropy of a specific frequency band exceeds three times the standard deviation of the historical mean.

[0019] Phase reconstruction and amplitude normalization are performed on the marked interval to separate the coupling components in the mixed interference.

[0020] As a further aspect of the present invention: the interference type differentiation specifically includes:

[0021] A dual-threshold judgment model is constructed: the first threshold is based on the ratio of the interference duration to the element characteristic period, and if the ratio is less than 0.1, it is judged as transient and repairable interference; the second threshold analyzes the morphological characteristics of the interference waveform through a convolutional neural network and outputs the probability of being unrepairable.

[0022] Perform vacuum environment coupling analysis and simultaneously monitor the vacuum degree change curve of the ion trap. When the correlation coefficient between the interference signal and the vacuum degree fluctuation exceeds 0.7, activate the irreparable flag bit.

[0023] Design a self-healing verification protocol, inject test pulse sequences into repairable interference, and if the recovery rate of the characteristic absorption peak exceeds 90%, it is confirmed as a repairable type;

[0024] Establish a residual energy spectrum library to store the spectral fingerprints of unrepairable interference. When a new interference matches the spectrum library with a degree of over 85%, it is directly classified as an unrepairable type.

[0025] As a further aspect of the present invention: the construction of the signal repair model specifically includes:

[0026] The generator employs a U-Net structure with an embedded attention gating mechanism. The U-Net structure contains 12 coding layers and 12 decoding layers, and retains the time-frequency features of the interference interval during the coding stage. The discriminator constructs a multi-scale convolutional pyramid to simultaneously evaluate the fidelity of the time-domain waveform and the integrity of the frequency-domain absorption peak.

[0027] The cerium group elements are light rare earth elements, and the yttrium group elements are heavy rare earth elements.

[0028] By introducing rare earth element characteristic absorption peak position locking into the generator loss function, the amplitude error of the signal in the preset peak position region after forced repair is less than 5%.

[0029] The generator weights are initialized using a pre-trained rare-earth energy spectrum encoder, and general repair knowledge is transferred to the current instrument environment.

[0030] The time-frequency mask matrix is ​​automatically generated based on the type of interference, guiding the generator to focus on repairing the interference area while keeping the clean area unchanged.

[0031] As a further aspect of the present invention: the specific process of the adversarial training includes:

[0032] First, basic repair capabilities are trained on a simulated dataset. Second, real instrument noise is injected for parameter fine-tuning. Finally, real-time repair performance is dynamically optimized through online reinforcement learning.

[0033] Three types of instrument-specific perturbations are randomly added to the discriminator input: baseline drift with limited slope, high-frequency glitches with a pulse width of less than ten microseconds, and harmonic distortion with a distortion rate of no more than 8%.

[0034] We utilize convolutional autoencoders to extract the core features of failed repair samples, construct a difficult sample library, and dynamically expand the training data.

[0035] A lightweight verification network is run concurrently during the signal repair process. If the signal-to-noise ratio improvement of the repaired signal does not reach 15 dB, the network automatically reverts to the previous stable model version and outputs the repaired pure ion flow signal.

[0036] As a further aspect of the present invention: the establishment of the state transition model specifically includes:

[0037] Extract the mass deviation vectors under different interference types from historical signals, store the offset patterns in partitions according to ion charge-mass ratio, and establish an offset-time decay curve for each partition.

[0038] The row dimension represents the interference type index, the column dimension represents the device runtime, and the matrix elements contain the mean quality offset and confidence interval. The matrix parameters are dynamically updated every eight hours.

[0039] By monitoring the rate of change of vacuum level in the ion trap, the calculation weights of the state transition probability are corrected. For every order of magnitude decrease in vacuum level, the confidence interval of the offset expands by 40%.

[0040] A standard ion beam of known mass is injected, and transfer matrix reconstruction is triggered when the measured offset deviates from the predicted value by more than five parts per million.

[0041] As a further aspect of the present invention: the real-time correction of the mass axis deviation of the current detection signal specifically includes:

[0042] The current signal quality deviation is decomposed into a reference offset and a high-frequency fluctuation component, and piecewise linear fitting and empirical mode processing are used respectively; the reference offset ranges from 0.01 to 0.1 amu, and the high-frequency fluctuation ranges from greater than 0.001 amu.

[0043] For repairable interference, a finite impulse response filter is used to correct the reference offset, while for unrepairable interference, a Kalman predictor is switched to compensate for high-frequency fluctuations.

[0044] When the correction amplitude exceeds twice the historical extreme value for three consecutive sampling periods, the system automatically switches to safe mode and locks the quality axis.

[0045] Record the parameter change path for each correction and generate a heatmap of quality stability evolution for system diagnosis.

[0046] As a further aspect of the present invention, the specific steps of S5 include:

[0047] An adaptive time window is generated based on the elution characteristics of rare earth elements. A forward-shifting contraction window is used for lanthanide light rare earth elements, with a window width of ±0.3 min. A backward-extending expansion window is used for heavy rare earth elements, with a window width of ±0.7 min.

[0048] An equivalent amount of internal standard signal is introduced as a benchmark factor, and a weighted calculation formula is constructed by combining the repair signal integrity coefficient. When the signal integrity is lower than the threshold, the weight of the internal standard is automatically increased to 80%.

[0049] The signal-to-noise ratio improvement and feature peak symmetry of the repaired signal are used as optimization indicators to establish a parameter adjustment rule base. If the peak symmetry deteriorates by more than 5%, the generator learning rate is reduced. If the signal-to-noise ratio improvement is less than 15 dB, the adversarial training rounds are increased.

[0050] It stores the time-frequency characteristic values, chromatographic matching deviations, and concentration errors before and after the repair. When a new interference mode is detected, it automatically generates a parameter optimization scheme and outputs the final ion concentration data.

[0051] As a further aspect of the present invention: the triggering system self-test process specifically includes:

[0052] When an unrepairable interference occurs three times consecutively, a primary self-test is initiated to scan for voltage fluctuations in the ion source lens; when it occurs five times consecutively, an intermediate self-test is activated to verify the electromagnetic field parameters of the mass analyzer; and when it occurs eight times consecutively, an advanced self-test is performed to comprehensively diagnose the vacuum system and detector gain.

[0053] Based on the automatic matching detection of the spectrum fingerprint of unrepairable interference, if the interference energy is concentrated in the high frequency band (greater than 10kHz), the focus is on detecting the electron multiplier; if it is concentrated in the low frequency band (less than 10kHz), the focus is on the stability of the radio frequency power supply.

[0054] The two-photon excitation detection channel and the conventional channel are operated synchronously. When the interference correlation between the two channels exceeds 90%, it is determined to be a hardware failure; when it is less than 40%, it is determined to be sample contamination.

[0055] A fault mode map is constructed based on a historical fault database. The self-inspection data features are matched in real time, and a fault location report with confidence level is output and an abnormal early warning signal is generated.

[0056] A rare earth ion chromatography online analysis and detection system includes:

[0057] A signal acquisition module is used to acquire rare earth ion flow signals after rare earth ion chromatography separation in real time.

[0058] The interference identification module performs multi-scale time-frequency decomposition on the acquired rare earth ion flow signal, extracts the interference feature interval in the rare earth ion flow signal, and classifies the interference type into repairable interference and unrepairable interference according to the interference features.

[0059] The signal repair module constructs a signal repair model for repairable interference, processes signal segments containing interference features through adversarial training, and outputs a repaired pure ion flow signal.

[0060] The dynamic calibration module establishes a state transition model based on historical signal quality offset data and corrects the quality axis deviation of the current detection signal in real time according to the type of interference.

[0061] The concentration calculation and feedback module matches the repaired signal with the chromatographic retention time, calculates the concentration of each rare earth ion, and optimizes the signal reconstruction parameters based on the repair effect.

[0062] The self-test and early warning module triggers the system self-test process and outputs an abnormal warning when the number of consecutive occurrences of the identified unrepairable interference exceeds a preset threshold.

[0063] The beneficial effects of this invention are:

[0064] (1) This invention achieves multi-scale time-frequency decomposition by constructing an adaptive wavelet basis function library, and accurately identifies and classifies repairable and unrepairable interferences in chromatographic signals by combining convolutional neural networks and vacuum degree coupling analysis. For repairable interferences, an adversarial repair model composed of a U-Net generator with embedded attention gating mechanism and a multi-scale discriminator is adopted. The high-fidelity repair of the signal is achieved through a three-stage training strategy (simulated data pre-training, real data fine-tuning and online reinforcement learning). At the same time, the integrity of key spectral regions is ensured by a rare earth element characteristic absorption peak position locking module. By establishing a state transition model containing a mass-charge ratio database and real-time offset records, and using empirical mode decomposition to separate mass deviation into a reference offset and high-frequency fluctuation components, dynamic compensation is performed by an adaptive FIR filter and an improved Kalman predictor, respectively. With the help of a dual internal standard correction system and a triple verification mechanism (peak area ratio, characteristic peak position offset and separation degree verification), the accuracy and reliability of rare earth element detection under complex working conditions are improved.

[0065] (2) This invention establishes a state transition model based on Markov chains, analyzes historical mass shift data and current instrument status in real time, and dynamically adjusts signal processing strategies: automatically matches the optimal repair intensity for different types of interference; continuously optimizes key parameters through online reinforcement learning mechanism, enabling the system to intelligently adapt to complex and variable sample matrices and instrument status fluctuations; combined with a dual-channel verification system and a fault mode map library, it realizes self-verification and self-optimization of processing strategies, reduces the need for frequent manual parameter adjustment in traditional methods, improves the efficiency of routine sample analysis, and ensures the robustness of rare earth ion chromatography online analysis and detection methods under different working conditions. Attached Figure Description

[0066] The invention will now be further described with reference to the accompanying drawings.

[0067] Figure 1 This is a flowchart of an online rare earth ion chromatography analysis and detection method according to the present invention;

[0068] Figure 2 This is a flowchart of an online rare earth ion chromatography analysis and detection system according to the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1, please refer to Figure 1 As shown, this invention provides an online rare earth ion chromatography analysis and detection method, comprising the following steps:

[0071] S1: Real-time acquisition of rare earth ion flow signals after rare earth ion chromatography separation;

[0072] S2: Perform multi-scale time-frequency decomposition on the acquired rare earth ion current signal, extract the interference feature interval in the rare earth ion current signal, and classify the interference type into repairable interference and unrepairable interference according to the interference features.

[0073] S3: For repairable interference, construct a signal repair model, process signal segments containing interference features through adversarial training, and output the repaired pure ion flow signal.

[0074] S4: Establish a state transition model based on historical signal quality offset data, and correct the quality axis deviation of the current detected signal in real time according to the type of interference;

[0075] The historical signal quality offset data includes: a pre-established rare earth ion mass-charge ratio database and real-time detected mass axis offset records;

[0076] S5: Match the repaired signal with the chromatographic retention time, calculate the concentration of each rare earth ion, and optimize the signal reconstruction parameters based on the repair effect.

[0077] S6: When the number of consecutive occurrences of identified unrepairable interference exceeds a preset threshold, the system self-check process is triggered and an abnormal warning is output.

[0078] In S1, the rare earth ion current signal after rare earth ion chromatography separation is acquired in real time, specifically including:

[0079] A high-performance liquid chromatography-inductively coupled plasma mass spectrometry (HPLC-ICP-MS) system was used as the basic detection platform. A sulfonic acid-based cation exchange column (such as a Dionex IonPac CS5A) was selected, with the column temperature controlled at 30±2℃. The mobile phase was an aqueous solution containing 5-20 mM α-hydroxyisobutyric acid (HIBA), and the pH was adjusted to 3.8-4.2 using ammonia. A six-way valve autosampler was used, with an injection volume of 20-100 μL and a stable flow rate of 1.0 mL / min.

[0080] During mass spectrometry detection, the plasma power was maintained at 1350-1550 W, the sampling depth at 8-10 mm, and the nebulizer gas flow rate at 0.8-1.2 L / min. The rare-earth ion current signal after chromatographic separation was detected by a quadrupole mass analyzer with a signal acquisition frequency of no less than 10 Hz. The analog signal was converted to a digital signal using a 16-bit high-precision analog-to-digital converter (ADC) to ensure a dynamic range covering 10 Hz. 4 Scale. All collected data are accompanied by timestamps accurate to the millisecond level, providing a time series reference for subsequent time-frequency analysis.

[0081] To ensure signal integrity, the system employs a full PEEK flow path from the column outlet to the detector inlet, with an inner diameter of 0.15-0.25 mm and a total length not exceeding 30 cm, minimizing dead volume. Simultaneously, a pulse damper (such as a 1.5 μL spring-loaded damper) is installed before the plasma torch to effectively suppress baseline fluctuations caused by infusion pump pulsation. During acquisition, the system back pressure (range 3-15 MPa) is monitored in real time, and a flow calibration program is automatically triggered when the pressure deviates from the set value by ±10%.

[0082] In S2, multi-scale time-frequency decomposition is performed on the acquired rare-earth ion current signal to extract interference feature regions from the rare-earth ion current signal. Based on the interference characteristics, the interference type is classified into repairable interference and unrepairable interference, specifically including:

[0083] During the signal preprocessing stage, the system constructs a wavelet basis function library. For the characteristic absorption peak positions of different rare earth element groups, the system automatically selects the optimal wavelet basis: for cerium group elements (La-Eu), the sym5 wavelet basis function is selected; for yttrium group elements (Gd-Lu), the bior3.5 biorthogonal wavelet is selected.

[0084] The signal decomposition adopts a seven-layer wavelet packet decomposition architecture, with the low-frequency channel (0-1.56Hz) used to extract baseline drift features. The high-frequency component (12.5~50Hz) processing adopts an improved S-transform algorithm, and the time-frequency resolution is dynamically adjusted according to the real-time signal-to-noise ratio: when the signal-to-noise ratio is higher than 30dB, a wide window (256 points) is used to improve the frequency resolution; when the signal-to-noise ratio is lower than 20dB, a narrow window (64 points) is switched.

[0085] During the interference feature extraction stage, the system calculates the energy distribution dispersion index of each sub-band at each scale. This index is defined as the standard deviation of the ratio of sub-band energy to global energy. When the dispersion of a specific frequency band (usually cD4, 3.125~6.25Hz) exceeds a preset threshold (empirical value is 0.35), that time period is marked as an abnormal interval.

[0086] Interference type differentiation employs a dual-threshold determination model. The first threshold is based on the absolute value of the interference duration: when the interference duration is less than 1 ms, it is determined to be transient and repairable interference. The second threshold is achieved by monitoring changes in the ion trap vacuum level: the system calculates the Pearson correlation coefficient between the interference signal and the vacuum level curve in real time; when the correlation coefficient exceeds 0.7 and lasts for more than 50 ms, an unrepairable flag is activated.

[0087] Self-healing verification protocol: For interference regions marked as repairable, inject a standard test pulse sequence (amplitude 30% of full scale, pulse width 1ms). Confirm the repair effect by comparing the characteristic absorption peak area ratio before and after repair. When the Nd³⁺ 575nm peak recovery rate is less than 90%, this type of interference is reclassified as unrepairable.

[0088] The residual energy spectrum library stores the spectral characteristics of unrepairable interference, organized using a binary tree structure. Each spectrum node contains characteristic parameters such as: fundamental frequency component amplitude (0-50Hz range, 0.5Hz resolution), harmonic distortion coefficients (2nd to 5th harmonics), and transient response attenuation constant. The matching algorithm employs an improved Dynamic Time Warping (DTW) method; when a new interference signal matches the records in the spectrum library by more than 85%, it is directly classified as unrepairable. The spectrum library is automatically updated weekly, and newly added interference patterns are entered into the system after confirmation by engineers.

[0089] The system maintenance interface provides interference analysis visualization functions, including: time-frequency distribution heatmaps (displaying areas of concentrated interference energy), phase trajectory diagrams (showing instantaneous phase changes of the signal), and vacuum coupling time series comparison diagrams. All diagnostic data is stored in HDF5 format, containing complete timestamps, device status parameters, and operation logs, supporting offline in-depth analysis. The configuration parameters of the interference identification module are managed through encrypted configuration files, including key parameters such as wavelet decomposition layer number, entropy threshold, and CNN model path, allowing authorized engineers to perform on-site optimization.

[0090] In S3, for repairable interference, a signal repair model is constructed. This model processes signal segments containing interference features through adversarial training, outputting a repaired, clean ion flow signal. Specifically, this includes:

[0091] The generator network employs an improved U-Net architecture, comprising 12 encoding layers and 12 decoding layers, each embedding an attention gating mechanism. The encoder uses dilated convolutions to progressively expand the receptive field, with feature-preserving nodes in the fourth and eighth layers specifically storing time-frequency features within interference regions. The decoder fuses low-level detail features through skip connections, and adds rare-earth element absorption peak locking in the final layer, containing 15 parallel one-dimensional convolutional kernels (corresponding to the absorption peaks of 15 rare-earth ions). After forced restoration, the amplitude error of the signal in the preset peak region is less than 5%. The network initialization uses pre-trained rare-earth energy spectrum encoder weights, trained on a dataset containing 500,000 standard rare-earth spectra.

[0092] The discriminator is designed as a multi-scale convolutional pyramid structure, comprising three parallel branches: the time-domain branch uses a 5-layer one-dimensional convolutional network (kernel lengths of 128, 64, 32, 16, and 8 points) to evaluate waveform fidelity; the frequency-domain branch is transformed by Fast Fourier Transform and then connected to a 4-layer two-dimensional convolutional network (kernel size 3×3) to verify the integrity of absorption peaks; the time-frequency branch uses continuous wavelet transform and then connects to a 3-layer three-dimensional convolutional network (kernel size 3×3×3) to analyze the rationality of the time-frequency distribution. The outputs of the three branches are fused by a fully connected layer to output the probability of signal authenticity (range 0-1). During discriminator training, three types of instrument-specific perturbations are injected: baseline drift with a slope not exceeding 0.5mV / s, high-frequency glitches with a pulse width of 2-10μs, and waveform distortion with a total harmonic distortion rate of 5%–8%, to enhance model robustness.

[0093] The adversarial training process was implemented in three phases: The first phase used a simulated dataset (containing 100,000 artificially injected perturbation samples) to train the base model with a learning rate of 0.0002, a batch size of 64, and the Adam optimizer. The second phase introduced noisy data (5000 samples) collected by real instruments for fine-tuning, reducing the learning rate to 0.00005 and focusing on optimizing the generator's attention gating parameters. The third phase deployed an online reinforcement learning mechanism to evaluate the repair effect in real time and adjust network parameters accordingly, updating the model weights every 100 actual samples processed. The training loss function used a four-element combination: adversarial loss (40% weight), feature matching loss (30%), peak locking loss (20%), and perceptual loss (10%).

[0094] The time-frequency mask matrix generation algorithm dynamically adjusts according to the interference type: for electromagnetic pulse interference, a rectangular time-domain mask is generated (pulse width extended by 20%); for plasma scintillation interference, a Gaussian frequency-domain mask is constructed (center frequency ±5Hz range). The mask matrix is ​​embedded into the decoder's path through a 1×1 convolutional layer, guiding the network to focus on repairing the interference region. A mask supervision node is set in the fourth layer of the encoder to ensure that the feature transfer loss in the clean signal region is less than 0.01.

[0095] The repair quality verification system comprises a lightweight CNN network (with only 3 convolutional layers) and traditional signal processing units. The CNN network calculates the structural similarity (SSIM) metric of the signal before and after repair in real time, triggering an alarm when it falls below 0.85. The traditional units detect the signal-to-noise ratio improvement value, automatically reverting to the previous stable model version if it does not reach the 15dB threshold. The system maintenance terminal displays visual information about the repair process, including: original / repaired signal comparison graphs, time-frequency distribution difference heatmaps, attention weight distribution graphs, etc., assisting engineers in quality assessment.

[0096] The difficult example management system continuously collects failed repair cases (SSIM less than 0.7 or SNR improvement less than 10dB), extracts core features through a convolutional autoencoder (compressed to a 128-dimensional vector), and builds a feature codebook using K-means clustering (k=50). It automatically generates 2000 augmentation samples weekly (through feature recombination and noise injection) to dynamically expand the training dataset. The system retains the 10 most recent model versions and automatically selects the optimal version for deployment based on validation set performance (average SSIM and peak position error).

[0097] The signal output interface undergoes triple verification: first, high-frequency artifacts are eliminated using a digital filter (cutoff frequency 0.8 times the Nyquist frequency); second, characteristic peak position shift is verified using standard samples (allowing ±0.1nm deviation); and finally, the peak area integral repeatability of the entire signal is calculated (RSD less than 1% is acceptable). All repair records are stored in a blockchain database, containing information such as the original signal, repair parameters, and operators, ensuring process traceability.

[0098] In S4, a state transition model is established based on historical signal quality offset data. According to the type of interference, the quality axis deviation of the currently detected signal is corrected in real time, specifically including:

[0099] During the database construction phase, the system pre-stores precise mass-to-charge ratio data for 15 rare earth ions, sourced from the NIST standard reference database. The real-time migration recording system uses a circular buffer to store mass migration data from the past 30 days, recording a complete mass scan every 5 minutes (covering a range of 100-1000 amu, with a resolution of 0.01 amu). Data storage employs a hierarchical structure: the raw data layer stores unprocessed instrument readings, the feature extraction layer stores migrations after wavelet noise reduction, and the analysis layer contains statistical features categorized by interference type.

[0100] The state transition model is constructed as follows: The system first extracts quality deviation vectors under different disturbance types from historical data, dividing the quality intervals into 10 AMU intervals. An independent offset pattern database is established for each interval, containing the mean offset, standard deviation, and time decay coefficient. The time decay curve is fitted using a three-parameter exponential model, with the fitting parameters automatically updated every 8 hours. The row dimension of the transition probability matrix encodes 12 preset disturbance types (such as electromagnetic pulse, plasma scintillation, etc.), while the column dimension is divided into 24 time periods (one hour per period) based on the equipment's operating time. Matrix elements include the 95% confidence interval of the offset and the probability of transition Markov chain states. A Bayesian smoothing algorithm is used to handle data mutations during each update.

[0101] The vacuum compensation subsystem monitors changes in the ion trap vacuum level in real time, with a sampling frequency of 10Hz. When the vacuum level decreases from the reference value (5×10⁻ ... 5 When the vacuum level (Pa) decreases by an order of magnitude, the system automatically expands the offset confidence interval by 40%. The compensation algorithm employs an adaptive weighting strategy: when the vacuum level is within the normal range, the confidence interval weight is set to 1.0; for each order of magnitude decrease, the weight decreases by 0.15. Cross-correlation analysis between vacuum level data and mass offset is performed every 15 minutes, and compensation parameter optimization is triggered when the correlation coefficient exceeds 0.6.

[0102] The standard ion beam validation system automatically runs a calibration procedure weekly: seven mass calibration solutions (containing Li, Na, K, Rb, Cs, Tl, and Bi elements) are injected sequentially, covering a scan range of 50-500 amu. When the deviation between the measured mass offset and the model prediction exceeds 5 ppm, the system initiates a matrix reconstruction process: first, the real-time calibration service is frozen; then, the transfer model is retrained based on data from the most recent 72 hours; and finally, it is re-uploaded after validation set testing (requiring a mean absolute error of less than 0.002 amu). The reconstruction process is fully automated and takes no more than 15 minutes.

[0103] The real-time correction system operates as follows: The quality deviation of the current signal is first separated into a baseline offset (range 0.01-0.1 amu) and high-frequency fluctuations (greater than 0.001 amu) using Empirical Mode Decomposition (EMD). The baseline offset correction employs an FIR filter, with coefficients dynamically adjusted according to the interference type: a 51st-order Hamming window design with a cutoff frequency of 0.1 Hz is used for repairable interference; an unrepairable interference is switched to a 127th-order Blackman window with a cutoff frequency of 0.05 Hz. High-frequency fluctuation compensation uses an improved Kalman predictor, incorporating the vacuum degree change rate as a control variable in the state equation. The measurement noise covariance matrix is ​​updated every 30 seconds.

[0104] The safety protection mechanism includes a three-level response: when the correction magnitude exceeds the historical extreme value (twice the maximum value in the past 24 hours) for three consecutive sampling cycles (1 second interval), the system first switches to safety mode (locking the mass axis to the most recent stable value); secondly, it initiates an automatic diagnostic program to check critical components such as the ion source and analyzer; finally, it generates an emergency maintenance report and notifies the engineer. All operations are recorded in the mass stability log, including fields such as timestamp, offset, correction parameters, and operation type.

[0105] The visualization diagnostic system generates a real-time heatmap of quality stability: the horizontal axis represents the mass number (100-1000 AMU), the vertical axis represents time (the last 24 hours), and the color depth represents the magnitude of the offset (blue indicates negative offset, red indicates positive offset). The heatmap refreshes every 5 minutes and supports zooming and region selection to view detailed data. The system also provides trend analysis curves to show the offset changes of specific mass numbers over different time periods, assisting engineers in assessing the system's status.

[0106] The performance monitoring interface displays key operating indicators: mass axis stability index (average offset over the past hour), correction success rate (effective percentage of the last 100 corrections), vacuum compensation strength, etc. All data is uploaded to the central monitoring system via the OPC-UA interface, supporting remote diagnostics and maintenance.

[0107] In S5, the repaired signal is matched with the chromatographic retention time, the concentration of each rare earth ion is calculated, and the signal reconstruction parameters are optimized based on the repair effect. Specifically, this includes:

[0108] A smart time window allocation system was established based on the elution characteristics of rare earth elements. For lanthanide light rare earth elements (La-Eu), the system adopts a forward-shifting and contracting time window strategy, with the window center point set 0.1 minutes before the standard retention time and the window width fixed at ±0.3 minutes. For heavy rare earth elements (Gd-Lu), a backward-shifting and expanding time window is used, with the window center point delayed by 0.2 minutes and the window width expanded to ±0.7 minutes. The time window boundaries are Gaussian smoothed (σ=0.05 minutes) to avoid truncation effects. The window parameters are dynamically adjusted based on the average retention time of the most recent 100 analyses.

[0109] The internal standard calibration system uses Y (89Y) and In (115In) as dual internal standard elements, which are added during the sample pretreatment stage. The internal standard signal integrity assessment includes three indicators: peak height recovery rate (target ≥ 95%), peak area consistency (RSD ≤ 3%), and retention time offset (≤ 0.05 minutes). When the repair signal integrity coefficient is lower than the preset threshold (0.7), the system automatically adjusts the weighted calculation formula, increasing the internal standard weight from the baseline value of 50% to 80%. The weighted calculation uses piecewise linear interpolation to smoothly transition the weight ratio within the integrity coefficient range of 0.7-0.9.

[0110] The parameter optimization engine monitors two key performance indicators in real time: signal-to-noise ratio (SNR) improvement (target ≥ 15dB) and feature peak symmetry (rate of change ≤ 5%). The system's built-in rule base contains 32 parameter adjustment rules. For example, when peak symmetry deteriorates by more than 5%, the generator's learning rate decreases by 0.0001 steps; when the SNR improvement is less than 15dB, the adversarial training rounds are increased from the standard 100 to 150. Once a rule is triggered, the system automatically generates a parameter adjustment scheme.

[0111] The data tracking system meticulously records repair data across four dimensions: time-frequency feature values ​​(including wavelet energy distribution and instantaneous frequency trajectory), chromatographic matching deviations (retention time difference and peak width ratio), concentration errors (compared to standard methods), and reconstruction parameters (learning rate, training epochs, etc.). This data is stored in a relational database, supporting SQL queries and trend analysis. When a new interference pattern is detected (with less than 70% similarity to historical records), the parameter optimization process is automatically initiated: first, the interference feature fingerprint is extracted; then, different parameter combinations are tested in a simulation environment; finally, the optimal solution is recommended, and a technical report is generated.

[0112] Before outputting concentration calculation results, triple verification is required: First, check the ratio of the peak area integral of each rare earth element to the internal standard (allowable fluctuation range ±5%); second, verify the characteristic peak position shift (less than or equal to 0.2 nm); finally, confirm the separation degree between adjacent elements (R greater than or equal to 1.5). All concentration data are accompanied by a quality assessment label (AE level), where A-level data must simultaneously meet the following requirements: signal-to-noise ratio greater than or equal to 25 dB and symmetry greater than or equal to 0.9. The system automatically generates a complete report containing the original data, repair process, calculation method, and quality rating, supporting export in PDF and CSV formats.

[0113] The feedback optimization system employs a closed-loop control architecture: after each analysis, the system compares the repair effect with the expected target and calculates the parameter adjustment amount using a fuzzy logic algorithm. The adjustment signal is transmitted to the signal repair module through a secure channel, using an incremental update method (each adjustment increment less than or equal to 10%) to ensure system stability. For parameter combinations that consistently perform poorly (failing to meet the target for three consecutive times), the system automatically triggers an expert diagnostic mode, generating a detailed report containing a problem description, root cause analysis, and solution suggestions for engineers' reference and decision-making.

[0114] In S6, when the number of consecutive occurrences of identified unrepairable interference exceeds a preset threshold, the system self-check process is triggered and an anomaly warning is output, specifically including:

[0115] When the system detects three consecutive instances of unrepairable interference, it initiates a primary self-test procedure. This procedure first scans the voltage fluctuation of the ion source lens, with a sampling frequency set to 10kHz and a continuous monitoring time of 30 seconds, recording the voltage fluctuation range, frequency characteristics, and transient pulse events. The detection parameters include the DC high voltage (range 0-5000V) and radio frequency superposition component (frequency 1-100kHz) of the extraction lens, focusing lens, and deflection lens, with an allowable fluctuation threshold set at ±0.5% of the nominal value. Simultaneously, it checks the ripple coefficient of the lens power supply (required to be less than or equal to 0.1%) and temperature stability (within ±1℃).

[0116] When the number of unrepairable interferences accumulates to 5, the system automatically activates the intermediate self-test procedure. This procedure focuses on verifying the electromagnetic field parameters of the mass analyzer, including the matching accuracy of the quadrupole's RF voltage (0-3000V) and DC voltage (0-1000V) (requiring a ratio error of less than or equal to 0.01%), and the stability of the magnetic field strength (for magnetic mass spectrometry) (fluctuation less than or equal to 5ppm). During the verification process, the system injects a standard calibration gas (containing NaCl and CsI cluster ions), and the electromagnetic field performance is evaluated by monitoring the broadening of the characteristic mass peaks (requiring an FWHM of less than or equal to 0.8amu) and their positional offset (less than or equal to 0.05amu). All parameter checks are performed in real time via the Modbus protocol, interacting with the power controller and reading the status register every 100ms.

[0117] When unrepairable interference occurs 8 times, an advanced self-test mode is triggered, and the system performs a comprehensive diagnostic of the vacuum system and detectors. Vacuum detection includes three levels of monitoring: backing pump pressure (0-10 Torr range, accuracy 0.1%), high vacuum gauge reading (1×10⁻⁻⁶), and high vacuum gauge reading (1×10⁻⁶). 8 -1×10⁻ 4Pa (sampling interval 1 second) and residual gas analysis (RGA scan 1-200 amu). Electron multiplier gain (stepped up from 500V, 50V increments), dark current (required to be less than or equal to 10 cps), and pulse height distribution (PHD, to verify single-electron response peak) are tested sequentially. The entire process takes approximately 15 minutes, during which the system automatically pauses sample analysis and enters diagnostic mode.

[0118] The spectral fingerprint analysis subsystem monitors the frequency domain characteristics of unrepairable interference in real time, employing a 1024-point FFT transform (Hanning window, frequency resolution 50Hz). When the interference energy is mainly distributed in the high-frequency band above 10kHz (energy proportion greater than or equal to 70%), the self-test program focuses on checking the electron multiplier components, including checking the high-voltage connector impedance (required to be greater than or equal to 100MΩ), testing the impulse response rise time (target less than or equal to 5ns), and verifying the gain linearity (within 10³-10). 7 The deviation within the range is less than or equal to 5%. For low-frequency dominant interference less than 10kHz (energy percentage greater than or equal to 60%), focus on checking the RF power supply system and measure carrier frequency stability (within ±10Hz), harmonic distortion (THD less than or equal to 1%), and load matching (VSWR less than or equal to 1.5).

[0119] The dual-channel validation system operates synchronously with a conventional detection channel (electron multiplier output) and a two-photon excitation detection channel (photodiode receiver). The signals from both channels are time-aligned (accuracy ±100ns) and then cross-correlation coefficients are calculated. The sampling window width is set to 1 second, and the sliding step size is 100ms. If the correlation coefficient exceeds 0.9 for 3 consecutive seconds, it is considered a hardware fault (such as detector aging or amplifier malfunction); if the correlation coefficient is below 0.4 for more than 5 seconds, it is considered sample matrix contamination. During validation, the system automatically records the baseline noise (required to be less than or equal to 1mV), pulse count rate difference (allowable ±5%), and waveform distortion (cross-validation error less than or equal to 3%) for both channels.

[0120] The intelligent diagnostic engine constructs a fault pattern map based on a database containing 5,000 historical fault cases. Each fault record includes: a description of the fault phenomenon (text + code), equipment status parameters (50-dimensional feature vector), maintenance measures, and effectiveness evaluation. Real-time self-test data is matched with the map database after feature extraction, and a modified nearest neighbor algorithm (k=5, distance weighted by feature importance) is used to calculate similarity, outputting a fault location report. The report includes a predicted faulty component (e.g., ion source lens, RF generator), a confidence score (0%–100%), and suggested measures (cleaning, replacement, or calibration). Conclusions with a confidence score greater than or equal to 80% directly trigger an automatic maintenance work order; conclusions with a confidence score between 50% and 80% require engineer confirmation; and conclusions with a confidence score less than 50% initiate an expert consultation mode.

[0121] The early warning signal generation system implements a three-level response based on the severity of the fault: Level 1 (yellow) warnings are for temporary faults that can be automatically recovered (such as brief vacuum fluctuations), which are displayed through the system interface and logged; Level 2 (orange) warnings are for potential faults requiring manual intervention (such as lens voltage drift), which include an SMS notification to the responsible engineer; and Level 3 (red) warnings are for serious faults that immediately endanger equipment safety (such as vacuum system leaks), which automatically cut off the high-voltage power supply and trigger audible and visual alarms. All warning information is accompanied by a complete diagnostic data package, including equipment status records for 5 minutes before and after the fault, self-test results, and preliminary analysis conclusions, which can be accessed remotely via a secure link. The system maintenance interface provides warning record query, statistical analysis, and response tracking functions to ensure closed-loop management.

[0122] Please see Figure 2 As shown, a rare earth ion chromatography online analysis and detection system includes:

[0123] A signal acquisition module is used to acquire rare earth ion flow signals after rare earth ion chromatography separation in real time.

[0124] The interference identification module performs multi-scale time-frequency decomposition on the acquired rare earth ion flow signal, extracts the interference feature interval in the rare earth ion flow signal, and classifies the interference type into repairable interference and unrepairable interference according to the interference features.

[0125] The signal repair module constructs a signal repair model for repairable interference, processes signal segments containing interference features through adversarial training, and outputs a repaired pure ion flow signal.

[0126] The dynamic calibration module establishes a state transition model based on historical signal quality offset data and corrects the quality axis deviation of the current detection signal in real time according to the type of interference.

[0127] The concentration calculation and feedback module matches the repaired signal with the chromatographic retention time, calculates the concentration of each rare earth ion, and optimizes the signal reconstruction parameters based on the repair effect.

[0128] The self-test and early warning module triggers the system self-test process and outputs an abnormal warning when the number of consecutive occurrences of the identified unrepairable interference exceeds a preset threshold.

[0129] Example 2: The present invention also includes specific experimental data:

[0130] 1. Signal Acquisition (corresponding to S1)

[0131] A high-performance liquid chromatography (HPLC) system equipped with inductively coupled plasma mass spectrometry (ICP-MS) was used. The chromatographic column was a Dionex IonPac CS5A cation exchange column, with a column temperature of 30 °C. The mobile phase was an aqueous solution containing 10 mM α-hydroxyisobutyric acid (HIBA) (pH=4.0), with a flow rate of 1.0 mL / min and an injection volume of 50 μL. The ICP-MS plasma power was set to 1450 W, the sampling depth to 9 mm, and the nebulizer gas flow rate to 1.0 L / min. Under these conditions, the retention times of Nd³⁺, Sm³⁺, and Gd³⁺ were approximately 5.2 min, 7.8 min, and 10.5 min, respectively. The signal acquisition frequency was 20 Hz.

[0132] 2. Interference identification and classification (corresponding to S2)

[0133] During the peak period of Nd³⁺ (~5.2 min), the system acquired a transient signal dip lasting 0.6 ms, and simultaneously detected synchronous fluctuations in the vacuum monitoring channel.

[0134] The system first uses the sym5 wavelet basis to perform 7-level wavelet packet decomposition on the Nd³⁺ ion current signal. In the cD4 subband (3.125-6.25 Hz), the energy dispersion was found to reach 0.41, exceeding the threshold of 0.35, so this period was marked as an abnormal interval.

[0135] Classification determination:

[0136] First threshold: The duration of the interference is 0.6 ms < 1 ms, which meets the condition of "transient repairable interference".

[0137] Second threshold: The Pearson correlation coefficient between the interference signal and the vacuum curve was calculated to be 0.75 (>0.7), but the duration did not reach 50 ms, so the unrepairable flag was not activated.

[0138] Based on comprehensive assessment, the system classifies this interference as "repairable interference".

[0139] 3. Signal Repair (corresponding to S3)

[0140] The system invokes a pre-trained adversarial repair model to handle the abnormal region. The generator is based on a U-Net architecture, and its loss function pre-sets the feature absorption peak position locking of Nd³⁺.

[0141] Before the repair, the signal-to-noise ratio (SNR) of the signal in this region was 18 dB, and the amplitude of the characteristic peak (e.g., m / z 146 for NdO⁺) was lost by approximately 35%. After model repair, the SNR of the output signal improved to 33 dB, the characteristic peak amplitude recovered to 98% of its original level, and the amplitude error was 2% (<5% of the preset requirement). The lightweight validation network confirmed that the SNR improvement (15 dB) met the threshold, so the repair result was adopted.

[0142] 4. Dynamic calibration of the mass axis (corresponding to S4)

[0143] During the same period of the disturbance, the state transition model predicts a reference offset of approximately +0.03 amu for the current mass axis based on the type of disturbance, "electromagnetic pulse".

[0144] The system uses a 51st-order FIR filter to correct the reference offset. After correction, verified by standard samples, the deviation between the measured and theoretical mass-to-charge ratio of the Nd³⁺ main isotope (m / z 146) decreased from 0.028 amu to 0.003 amu.

[0145] 5. Concentration Calculation and Feedback (corresponding to S5)

[0146] The repaired and calibrated Nd³⁺ ion current signal was integrated within its adaptive time window (5.2 ± 0.3 min), combined with the internal standard (¹¹). 5 Quantitative calculations are performed using In.

[0147] Ultimately, the error in the Nd³⁺ concentration calculation compared to the standard method (offline ICP-MS) decreased from -12.5% ​​before restoration to -1.8%. The system recorded a peak symmetry change rate of +1% and a signal-to-noise ratio improvement of 15 dB during this restoration, confirming that the current parameters were optimal and no feedback optimization was required.

[0148] The working principle of this invention is as follows: This invention employs a high-performance liquid chromatography-mass spectrometry (HPLC-MS) system to acquire rare earth ion current signals in real time, achieving efficient separation of rare earth elements through optimized chromatographic conditions (sulfonic acid-based cation exchange column, HIBA mobile phase). To address interference in the acquired signals, the system constructs an adaptive wavelet basis function library for multi-scale time-frequency decomposition, combining convolutional neural networks and vacuum coupling analysis to accurately distinguish between repairable and unrepairable interference types. For repairable interference, an adversarial repair model is designed, consisting of a U-Net generator with attention gating and a multi-scale discriminator, achieving high-fidelity signal repair through a three-stage training strategy. The system establishes a state transition model including a mass-charge ratio database and real-time offset records, dynamically correcting mass axis deviation using empirical mode decomposition and adaptive filtering techniques. During concentration calculation, an intelligent time window is designed based on the elution characteristics of rare earth elements, introducing dual internal standard correction and a parameter optimization engine to ensure quantitative accuracy. When unrepairable interference is detected, the system initiates a graded self-check procedure based on the cumulative number of occurrences, achieving precise fault location through spectral fingerprint analysis and dual-channel verification, ultimately ensuring equipment safety through a three-level early warning mechanism. This innovative method combines deep learning signal processing with traditional chromatographic analysis, improving the stability and reliability of rare earth element detection under complex operating conditions.

[0149] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A rare earth ion chromatography online analysis and detection method, characterized in that, Includes the following steps: S1: Real-time acquisition of ICP-MS ion current signals after rare earth ion chromatography separation; S2: Perform multi-scale time-frequency decomposition on the acquired rare earth ion current signal, extract the interference feature interval in the rare earth ion current signal, and classify the interference type into repairable interference and unrepairable interference according to the interference features. Interference type differentiation includes: a first threshold based on the ratio of interference duration to element characteristic period; a ratio less than 0.1 indicates transient, repairable interference; a second threshold analyzes the morphological characteristics of the interference waveform using a convolutional neural network to output an unrepairable probability; vacuum environment coupling analysis is implemented, synchronously monitoring the ion trap vacuum degree change curve; when the correlation coefficient between the interference signal and vacuum degree fluctuation exceeds 0.7, an unrepairable flag is activated; a self-healing verification protocol is designed, injecting test pulse sequences into repairable interference; if the characteristic absorption peak recovery rate exceeds 90%, it is confirmed as a repairable type; a residual energy spectrum library is established to store the spectral fingerprints of unrepairable interference; when a new interference matches the spectrum library with a matching degree exceeding 85%, it is directly classified as an unrepairable type. S3: For repairable interference, construct a signal repair model, process signal segments containing interference features through adversarial training, and output the repaired pure ion flow signal. The signal restoration model is constructed as follows: the generator retains the time-frequency characteristics of the interference region during the encoding stage; rare earth element characteristic absorption peak position locking is introduced into the generator loss function to force the amplitude error of the signal in the preset peak position region to be less than 5% after restoration; the generator weights are initialized using a pre-trained rare earth energy spectrum encoder to transfer general restoration knowledge to the current instrument environment; and a time-frequency mask matrix is ​​automatically generated according to the interference type to guide the generator to focus on restoring the interference region while keeping the clean region unchanged. S4: Establish a state transition model based on historical signal quality offset data, and correct the quality axis deviation of the current detected signal in real time according to the type of interference; The state transition model is established by: extracting mass deviation vectors under different interference types from historical signals, storing offset patterns in partitions according to ion charge-to-mass ratio, and establishing an offset-time decay curve for each partition; the row dimension represents the interference type index, the column dimension represents the device operating time, and the matrix elements contain the transition Markov chain state probabilities and confidence intervals; by monitoring the rate of change of the ion trap vacuum level, the calculation weights of the state transition probabilities are corrected, and the offset confidence interval is expanded by 40% for every order of magnitude decrease in vacuum level; Real-time correction includes: decomposing the current signal quality deviation into a reference offset and a high-frequency fluctuation component; the reference offset correction uses an FIR filter with coefficients dynamically adjusted according to the type of interference; and the high-frequency fluctuation compensation uses an improved Kalman predictor, with the vacuum degree change rate introduced as a control variable in the state equation. S5: Match the repaired signal with the chromatographic retention time, calculate the concentration of each rare earth ion, and optimize the signal reconstruction parameters based on the repair effect. S6: When the number of consecutive occurrences of identified unrepairable interference exceeds a preset threshold, the system self-check process is triggered and an abnormal warning is output.

2. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The multi-scale time-frequency decomposition of the acquired rare-earth ion current signal includes: The optimal wavelet basis is dynamically selected based on the characteristic absorption peak position of rare earth elements. Symmetric compactly supported wavelets are used for cerium group elements, while asymmetric oscillatory wavelets are used for yttrium group elements. The low-frequency channel extracts baseline drift features through 7-layer wavelet packet decomposition, while the high-frequency channel captures transient pulse interference through an improved S-transform. The time-frequency resolution is dynamically adjusted according to the signal-to-noise ratio. Calculate the energy distribution dispersion of each sub-band at each scale, and mark an abnormal interval when the energy entropy of a specific frequency band exceeds three times the standard deviation of the historical mean. Phase reconstruction and amplitude normalization are performed on the marked interval to separate the coupling components in the mixed interference.

3. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The construction of the signal repair model includes: The generator employs a U-Net structure with an embedded attention gating mechanism, which includes 12 coding layers and 12 decoding layers. The discriminator constructs a multi-scale convolutional pyramid to simultaneously evaluate the fidelity of the time-domain waveform and the integrity of the frequency-domain absorption peak.

4. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The specific process of the adversarial training includes: First, basic repair capabilities are trained on a simulated dataset. Second, real instrument noise is injected for parameter fine-tuning. Finally, real-time repair performance is dynamically optimized through online reinforcement learning. Three types of instrument-specific perturbations are randomly added to the discriminator input: baseline drift with limited slope, high-frequency glitches with a pulse width of less than ten microseconds, and harmonic distortion with a distortion rate of no more than 8%. We utilize convolutional autoencoders to extract the core features of failed repair samples, construct a difficult sample library, and dynamically expand the training data. A lightweight verification network is run concurrently during the signal repair process. If the signal-to-noise ratio improvement of the repaired signal does not reach 15 dB, the network automatically reverts to the previous stable model version and outputs the repaired pure ion flow signal.

5. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The establishment of the state transition model includes: A standard ion beam of known mass is injected, and transfer matrix reconstruction is triggered when the measured offset deviates from the predicted value by more than five parts per million.

6. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The real-time correction of the quality axis deviation of the current detection signal specifically includes: The current signal quality deviation is decomposed into a reference offset and a high-frequency fluctuation component, and piecewise linear fitting and empirical mode processing are used respectively; the reference offset ranges from 0.01 to 0.1 amu, and the high-frequency fluctuation ranges from greater than 0.001 amu. Repairable interference uses a 51st-order Hamming window design with a cutoff frequency of 0.1Hz; unrepairable interference switches to a 127th-order Blackman window with a cutoff frequency of 0.05Hz. When the correction amplitude exceeds twice the historical extreme value for three consecutive sampling periods, the system automatically switches to safe mode and locks the quality axis. Record the parameter change path for each correction and generate a heatmap of quality stability evolution for system diagnosis.

7. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The specific steps of S5 include: An adaptive time window is generated based on the elution characteristics of rare earth elements. A forward-shifting contraction window is used for lanthanide light rare earth elements, with a window width of ±0.3 min. A backward-extending expansion window is used for heavy rare earth elements, with a window width of ±0.7 min. An equivalent amount of internal standard signal is introduced as a benchmark factor, and a weighted calculation formula is constructed by combining the repair signal integrity coefficient. When the signal integrity is lower than the threshold, the weight of the internal standard is automatically increased to 80%. The signal-to-noise ratio improvement and feature peak symmetry of the repaired signal are used as optimization indicators to establish a parameter adjustment rule base. If the peak symmetry deteriorates by more than 5%, the generator learning rate is reduced. If the signal-to-noise ratio improvement is less than 15 dB, the adversarial training rounds are increased. It stores the time-frequency characteristic values, chromatographic matching deviations, and concentration errors before and after the repair. When a new interference mode is detected, it automatically generates a parameter optimization scheme and outputs the final ion concentration data.

8. The rare earth ion chromatography online analysis and detection method according to claim 1, characterized in that, The trigger system self-test process specifically includes: When an unrepairable interference occurs three times consecutively, a primary self-test is initiated to scan for voltage fluctuations in the ion source lens; when it occurs five times consecutively, an intermediate self-test is activated to verify the electromagnetic field parameters of the mass analyzer; and when it occurs eight times consecutively, an advanced self-test is performed to comprehensively diagnose the vacuum system and detector gain. Based on the automatic matching detection of the spectrum fingerprint of unrepairable interference, if the interference energy is concentrated in the high frequency band (greater than 10kHz), the focus is on detecting the electron multiplier; if it is concentrated in the low frequency band (less than 10kHz), the focus is on the stability of the radio frequency power supply. The two-photon excitation detection channel and the conventional channel are operated synchronously. When the interference correlation between the two channels exceeds 90%, it is determined to be a hardware failure; when it is less than 40%, it is determined to be sample contamination. A fault mode map is constructed based on a historical fault database. The self-inspection data features are matched in real time, and a fault location report with confidence level is output and an abnormal early warning signal is generated.

9. A rare earth ion chromatography online analysis and detection system, characterized in that, A rare earth ion chromatography online analysis and detection method as described in any one of claims 1-8, comprising: A signal acquisition module is used to acquire rare earth ion flow signals after rare earth ion chromatography separation in real time. The interference identification module performs multi-scale time-frequency decomposition on the acquired rare earth ion flow signal, extracts the interference feature interval in the rare earth ion flow signal, and classifies the interference type into repairable interference and unrepairable interference according to the interference features. The signal repair module constructs a signal repair model for repairable interference, processes signal segments containing interference features through adversarial training, and outputs a repaired pure ion flow signal. The dynamic calibration module establishes a state transition model based on historical signal quality offset data and corrects the quality axis deviation of the current detection signal in real time according to the type of interference. The concentration calculation and feedback module matches the repaired signal with the chromatographic retention time, calculates the concentration of each rare earth ion, and optimizes the signal reconstruction parameters based on the repair effect. The self-test and early warning module triggers the system self-test process and outputs an abnormal warning when the number of consecutive occurrences of the identified unrepairable interference exceeds a preset threshold.

Images