Adaptive control method for liquid cathode glow discharge optical emission spectroscopy

Abstract

This invention relates to the field of plasma spectral detection technology, specifically providing an adaptive control method for the spectral detection of liquid cathode glow discharge. The method includes: real-time acquisition of multimodal state data of the discharge process, inputting this data into a pre-trained stability evaluation model, and outputting a stability score that quantitatively characterizes the discharge state. The score is compared with preset first threshold, second threshold, and fault threshold: if the score is ≥ the second threshold, a stable state is determined and spectral acquisition and analysis are triggered; if the score is between the first and second thresholds, a critically stable state is determined and monitoring continues; if the score is between the fault threshold and the first threshold, an unstable state is determined and at least two discharge parameters are coordinated and adjusted until stability is restored; if the score is ≤ the fault threshold, a fault state is determined and a maintenance alarm is triggered. This invention significantly improves the detection stability and automation level under complex operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of plasma spectral detection technology, specifically relating to an adaptive control method for liquid cathode glow discharge spectral detection. Background Technology

[0002] Liquid cathode glow discharge spectroscopy (SCGD-OES) technology has broad application prospects in heavy metal detection and mineral analysis due to its advantages such as simple equipment and the ability to perform online analysis. However, this technology faces severe challenges in the detection of high-salinity complex matrix samples: high salt concentrations cause drastic changes in solution conductivity, salt crystals adhere to the electrode surface, random bubble entry, and electrode thermal drift, all of which can disrupt the dynamic equilibrium of the plasma, causing discharge flicker, contraction, or even extinction. Plasma instability directly leads to changes in excitation conditions, random fluctuations in characteristic spectral line intensity, and increased spectral background noise, severely reducing the repeatability and accuracy of detection results. This results in frequent system shutdowns or reliance on manual intervention, making it difficult to meet the needs of industrial online continuous detection.

[0003] To address this issue, existing technologies primarily enhance the intrinsic stability of plasma by optimizing the discharge cavity structure and employing pulsed power supplies, or by monitoring a single electrical signal and discarding abnormal data after detection. However, these solutions can only provide a reactive or rough assessment of instability, failing to achieve real-time, comprehensive, and quantitative sensing of plasma stability. Due to the lack of precise assessment capabilities for the discharge state, the system cannot provide timely warnings and interventions in the early stages of instability, and it is even more difficult to proactively adjust parameters based on assessment results to maintain stable operation. Summary of the Invention

[0004] To address the aforementioned shortcomings of the prior art, this invention provides an adaptive control method for the spectral detection of liquid cathode glow discharge, thereby solving the aforementioned technical problems.

[0005] In a first aspect, the present invention provides an adaptive control method for detecting glow discharge spectroscopy in a liquid cathode, comprising:

[0006] Real-time acquisition of multimodal state data during the glow discharge process of a liquid cathode, including discharge voltage signal, discharge current signal, plasma luminescence image, and continuous spectral background signal; The multimodal state data is input into a pre-trained stability assessment model, which outputs a stability score of the plasma for quantitatively assessing the current discharge state. The stability score is compared with at least one preset stability threshold to determine the current discharge state; Based on the determined discharge state, the corresponding closed-loop control operation is executed: if the discharge is determined to be in a stable state, spectral acquisition is triggered and the acquired spectrum is analyzed to obtain element detection results; if the discharge is determined to be in an unstable state, at least two discharge parameters are adjusted in a coordinated manner until the discharge returns to a stable state; if the discharge is determined to be in a fault state, a maintenance alarm is triggered.

[0007] In one optional implementation, multi-mode state data during the glow discharge process of the liquid cathode are acquired in real time, including: Real-time acquisition of voltage and current signals from the discharge circuit; Real-time image signals of the plasma luminescent area are acquired using an industrial camera; The continuous spectral background signal of plasma emission is acquired in real time using a spectrometer.

[0008] In an optional implementation, the stability score is generated by directly processing the multimodal state data through an end-to-end deep neural network model; the input layer of the deep neural network model receives a raw multimodal time-series data stream that is time-synchronized and aligned without manual feature extraction, including raw high-frequency waveforms of discharge voltage and discharge current, raw image sequence fragments of the plasma luminescent region, and raw intensity sequences of continuous spectral background in selected bands.

[0009] In an optional implementation, the deep neural network model includes a multimodal spatiotemporal attention encoder, the multimodal spatiotemporal attention encoder comprising: The branch feature extraction subnetwork is used to extract electrical fluctuation features from voltage waveforms and current waveforms, spatiotemporal evolution features from image sequences, and temporal evolution features from spectral background sequences, respectively. The cross-modal attention fusion module is used to perform a cross-attention mechanism between high-level features extracted from different branches to adaptively learn and quantify the coupling relationship and consistency between different modal signals, and use the fused features as a high-dimensional feature vector characterizing the discharge state. The training process of the deep neural network model introduces auxiliary constraint terms based on discharge physics knowledge as part of the loss function; the auxiliary constraint terms include at least: energy balance constraints for constraining the relationship between the internal characterization of the model and the instantaneous input power, and morphological-electrical correlation constraints for constraining the correlation between plasma morphological characteristics and discharge circuit impedance characteristics.

[0010] In an optional implementation, the stability score is generated synchronously through a multi-task diagnostic model; the multi-task diagnostic model adopts a network architecture combining a shared feature extraction layer and independent task heads; the shared feature extraction layer is used to extract temporal dynamic features from the current sampling window; the multi-task diagnostic model includes a first task head and a second task head; the first task head is a regression head, which uses a sigmoid activation function to output a refined stability score in the range of 0-1; the second task head is a classification head, which uses a softmax activation function to output the probability distribution of multiple preset root cause categories.

[0011] In an optional implementation, the temporal dynamic features include: Electrical dynamic characteristics include the instantaneous phase difference between the discharge voltage and current waveforms, the impedance change trend of the discharge circuit, and the main noise power of the high-frequency components of the current. The spatiotemporal features of the image include the jitter amplitude of the plasma bright spot centroid, the rate of change of the bright spot area, and the proportion of edge shadows used to indicate the crystal adhesion at the electrode tip; Spectral and auxiliary characteristics, including abrupt changes in continuous spectral background intensity, rate of change of electrode temperature, and setpoints for solution flow rate; Cross-modal mutual information features are used to quantify the degree of synchronization between high-frequency components of electrical signals and image texture entropy.

[0012] In one optional implementation, at least two discharge parameters are coordinated and adjusted, including: Extract the maximum probability value and its corresponding root cause category from the probability distribution of root cause categories, and compare the maximum probability value with a preset confidence threshold: If the maximum probability value is higher than the confidence threshold, the current primary root cause is determined to be the root cause category corresponding to the maximum probability, and an adaptive control strategy for the root cause category is triggered. If the maximum probability value is not higher than the confidence threshold, the current disturbance is determined to be a composite disturbance or an unknown disturbance, and a preset general control strategy is triggered.

[0013] In one optional implementation, the general control strategy includes: A preset control algorithm is used to calculate the adjustment amount of at least two parameters among discharge voltage, solution flow rate and electrode spacing; the preset control algorithm is at least one of incremental proportional-integral-derivative control algorithm, Takagi-Sugeno fuzzy control algorithm or model predictive control algorithm based on quadratic programming solution; According to the adjustment amount, the at least two parameters are coordinated to suppress plasma flickering, contraction or extinction, so that the plasma can be restored to a stable discharge state.

[0014] In an optional implementation, the stability score is compared with at least one preset stability threshold to determine the current discharge state, including: The stability score is compared with a preset first threshold, a second threshold, and a fault threshold, and the current discharge state is determined based on the comparison results. If the stability score is greater than or equal to the second threshold, the discharge is determined to be in a stable state. If the first threshold is less than or equal to the stability score and less than the second threshold, the discharge is determined to be in a critical stable state. The current parameters are maintained and monitoring continues. If the fault threshold is less than the stability score and less than the first threshold, the discharge is determined to be in an unstable state. If the stability score is less than or equal to the fault threshold, the discharge is determined to be in a fault state, triggering a maintenance alarm.

[0015] In one optional implementation, the acquired spectrum is analyzed to obtain elemental detection results, including: Preprocessing operations on spectral data; the preprocessing operations include at least one of the following: background subtraction based on polynomial least squares fitting; wavelength correction based on multi-point linear regression of internal standard spectral lines; smoothing based on Savitzky-Golay filtering; threshold denoising based on Daubechies wavelet; multi-frame spectral stacking and averaging; intensity normalization and segment standardization; The preprocessed spectral data is input into a pre-trained quantitative analysis model; the quantitative analysis model is a one-dimensional convolutional neural network regression model, used to realize the nonlinear mapping from spectral intensity sequence to target element concentration; the model outputs the concentration detection result of at least one target element and its corresponding prediction uncertainty.

[0016] The beneficial effects of this invention lie in the adaptive control method for spectral detection of liquid cathode glow discharge provided by this invention. This invention acquires multimodal state data such as discharge voltage, discharge current, plasma emission images, and spectral background signals in real time, and directly generates a stability score based on this raw data, achieving comprehensive quantitative perception of the plasma state. The principle is that discharge instability often manifests as coordinated anomalies in multiple dimensions—high-frequency jitter in the voltage waveform, texture changes in the plasma image, and noise enhancement in the spectral background. These phenomena are intrinsically correlated in different modes. This invention processes multimodal raw data simultaneously through an end-to-end deep neural network, enabling it to capture early signs of instability that cannot be identified by a single human eye or single-parameter monitoring, such as the spatiotemporal coupling characteristics of microsecond-level electrical fluctuations and millisecond-level image bright spot contraction. Compared to traditional methods based on a single signal or post-event judgment, this invention can provide continuous quantitative scores at the initial stage of instability, providing accurate basis for subsequent control decisions. It fundamentally solves the technical problem of the inability to quantitatively assess the plasma state in real time, significantly improving the stability and detection reliability of discharge operation under complex conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0019] Figure 2 This is a diagram of the glow discharge structure in a liquid cathode glow discharge spectroscopy detection system.

[0020] Figure 3 This is a structural diagram of the waste discharge base in a liquid cathode glow discharge spectroscopy detection system.

[0021] Figure 4 This is a schematic diagram of the structure and DCS communication relationship of the liquid cathode glow discharge spectroscopy detection system.

[0022] Figure 5 This is another illustrative flowchart of a method according to an embodiment of the present invention.

[0023] Figure 6 This is a flowchart of spectral data preprocessing and signal-to-noise ratio enhancement according to an embodiment of the present invention.

[0024] Figure 7 This is a schematic diagram of a quantitative spectral analysis according to an embodiment of the present invention.

[0025] Among them, 1 is a tungsten rod; 2 is a carbon rod; 3 is a ceramic tube; 4 is a waste discharge base; and 41 is a waste discharge port. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0028] The method provided in this embodiment of the invention is executed by a computer device, and correspondingly, the adaptive control system for liquid cathode glow discharge spectral detection runs in the computer device.

[0029] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The executing entity can be an adaptive control system for detecting glow discharge spectroscopy in a liquid cathode. Depending on the specific requirements, the order of steps in this flowchart can be changed, and some steps can be omitted.

[0030] like Figure 1 As shown, the method includes: S1. Real-time acquisition of multi-modal state data during the glow discharge process of the liquid cathode, wherein the multi-modal state data includes discharge voltage signal, discharge current signal, plasma luminescence image and continuous spectrum background signal; S2. Input the multimodal state data into a pre-trained stability evaluation model, and have the stability evaluation model output a stability score for quantitatively evaluating the plasma in the current discharge state; S3. Compare the stability score with at least one preset stability threshold to determine the current discharge state; S4. Based on the determined discharge state, execute the corresponding closed-loop control operation: if the discharge is determined to be in a stable state, trigger spectral acquisition and analyze the acquired spectrum to obtain element detection results; if the discharge is determined to be in an unstable state, coordinately adjust at least two discharge parameters until the discharge returns to a stable state; if the discharge is determined to be in a fault state, trigger a maintenance alarm.

[0031] Figure 2 and Figure 3 Together they demonstrated and explained the structure of the core hardware component of the liquid cathode glow discharge spectroscopy (SCGD-OES) detection system in this invention—the discharge unit.

[0032] Figure 2 The discharge structure shown includes: Tungsten rod 1: As a high-voltage anode, it is usually connected to the positive terminal of a DC high-voltage power supply and is located at the top of the device.

[0033] Carbon rod 2: Serves as a liquid cathode and is located at the bottom of the device. The solution to be tested flows out through the interior or surface of the carbon rod, forming a thin liquid film at the end of the carbon rod.

[0034] Ceramic tube 3: It is fitted over the tungsten rod and serves as insulation and positioning to ensure that no breakdown discharge occurs between the high-voltage anode and surrounding components.

[0035] Waste discharge base 4: Located at the bottom of the device, used to collect waste liquid after discharge.

[0036] like Figure 3 As shown, the waste outlet 41 is connected to the waste outlet base 4 to discharge waste liquid from the system.

[0037] Figure 4 This document illustrates the physical architecture, hardware module composition, and communication relationship between the entire intelligent detection system described in this invention and its data interaction with the upper-level industrial control system. The entire system consists of a field layer (SCGD-OES detection unit), a control layer (control server), and a monitoring and management layer (host computer / DCS system). The SCGD-OES detection unit, located in the field layer, is the physical hardware that directly contacts the sample. Its core components include a discharge chamber that generates plasma, a high-voltage power supply that provides the excitation voltage, a solution delivery device that controls the flow rate, a spectrometer for acquiring spectra, an industrial camera for capturing plasma images, and various sensors. Its function is to achieve the multimodal data acquisition described in the document. The control layer, located at the core of the system, is the control server. Its hardware core is the "MCU main control unit + FPGA co-processing unit" collaborative architecture emphasized in the document. It runs all core algorithms, including state assessment, adaptive control (such as PID and fuzzy control algorithms), and spectral analysis (including preprocessing and 1D-CNN-based quantitative analysis). The MCU and FPGA communicate internally at high speed via SPI or EtherCAT bus. The top-level monitoring and management layer is the host computer or distributed control system (DCS) interface module. It acts as a bridge between the system and the outside world, connecting with remote PCs or factory DCSs through industrial protocols such as Modbus-TCP and OPC-UA, uploading detection results and status, and receiving process instructions.

[0038] Please refer to Figure 5 The following provides a detailed description of each step in the adaptive control method for detecting liquid cathode glow discharge spectra provided by the present invention.

[0039] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0040] S101. The voltage and current signals of the discharge circuit are synchronously acquired through a high-frequency sampling module. The high-frequency sampling module includes a voltage divider connected to both ends of the discharge electrodes and a current transformer connected in series with the discharge circuit. The sampling frequency is set to no less than 1 kHz to ensure that high-frequency dynamic fluctuations during the discharge process can be captured. The acquired analog signals are converted from analog to digital and then transmitted to a data buffer in digital form for subsequent stability assessment.

[0041] S102. Image signals of the plasma luminescence region are acquired in real time using an industrial camera. The industrial camera is equipped with a suitable lens, facing the plasma discharge region, and continuously acquires a sequence of grayscale images at a frame rate of no less than 100 fps. To reduce data volume and focus on key areas, the camera can be configured with a region of interest, acquiring only local images containing the plasma bright spot and its surrounding electrodes. Each frame is accurately timestamped for alignment with other modal data.

[0042] S103. The continuous spectral background signal emitted by the plasma is acquired by a spectrometer. The plasma emission is collected by an optical fiber probe and then introduced into the spectrometer, which has a wavelength coverage of 200–800 nm and an optical resolution of no more than 0.3 nm, capable of resolving fine structures in complex matrix backgrounds. The spectrometer operates in external trigger mode, synchronized with image acquisition and electrical sampling, and outputs a raw sequence of continuous spectral background intensity in selected bands (e.g., 225–235 nm and 400–410 nm) at fixed intervals.

[0043] To ensure the temporal consistency of multimodal data, all the aforementioned acquisition units are connected to the same clock synchronization system, and each data frame or sampling point is assigned a globally unique timestamp. The acquired raw voltage waveforms, current waveforms, image sequences, and spectral background sequences are packaged into data windows of fixed duration (e.g., 2 seconds) in chronological order and directly used as input to the subsequent stability evaluation model without any manual feature extraction. This raw signal-level synchronous acquisition method preserves the instantaneous correlation and high-frequency details across modes during the discharge process.

[0044] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0045] Adopt a stability evaluation model based on Gradient Boosting Decision Tree (GBDT). The input features of this model include: (1) Discharge signal fluctuation characteristics: the standard deviation σV and coefficient of variation CVV of the discharge voltage, the fluctuation amplitude ΔI and peak-to-peak value IPP of the discharge current; (2) Image texture entropy value: the gray texture entropy Htexture of the plasma emission image, calculated through the Gray Level Co-occurrence Matrix (GLCM); (3) Spectral background variance: the variance σbg and peak noise ratio SNRbg of the continuous spectral background signal in the 200 - 250 nm band.

[0046] The model is trained with historical data to map the above 8-dimensional feature vector into a stability score S between 0 and 1. The higher the score, the more stable the plasma. Specifically, S > 0.75 represents a high-stable state, 0.65 < S ≤ 0.75 represents a critical stable state, and S ≤ 0.65 represents an unstable state that requires triggering control adjustment.

[0047] In this embodiment, the first threshold S1 is set to 0.65, the second threshold S2 is set to 0.75, and the fault threshold Sf is set to 0.50. When S < S1, the MCU main control unit immediately starts adaptive control; when S > S2 and lasts for more than 5 seconds, the system determines that it has recovered stability and triggers spectral acquisition; when S < Sf and lasts for more than 30 seconds, the system generates a maintenance alarm.

[0048] In practical applications, a Support Vector Machine (SVM) model can also be used to replace the GBDT model. The features are mapped to a high-dimensional space through a Radial Basis Function (RBF) kernel function for classification or regression, and the calculation of the stability score can also be achieved.

[0049] To further improve the accuracy and robustness of the stability score, in another preferred embodiment of the present invention, the stability score is not obtained by manually extracting features and then inputting them into a traditional machine learning model, but is directly generated by an end-to-end deep neural network model for processing the original multi-modal data. The core of this embodiment is: enabling the model to autonomously learn and extract high-dimensional features related to stability from the original signal, avoiding possible information loss or bias in the manual feature design process.

[0050] The input layer of the deep neural network model receives the original multi-modal time-series data stream that is not subjected to any manual feature extraction and is strictly time-synchronized and aligned. For a sampling window with a length of seconds (for example seconds), the input data includes: Electrical signal: Discharge voltage and discharge current The original high-frequency waveform, sampling rate Therefore, each window contains Each sampling point constitutes two one-dimensional temporal tensors. , .

[0051] Image sequence: Grayscale image sequence of the plasma luminescent region, frame rate Each window contains Frame images. Each frame image is scaled or cropped to a fixed size. This constitutes a four-dimensional tensor. (Time, Height, Width, Channel).

[0052] Spectral background: A continuous sequence of raw intensity values ​​of the spectral background in selected bands (e.g., 225-235 nm and 400-410 nm). Each window contains... Each spectral sampling point constitutes a one-dimensional time series tensor. ,in The number of wavelength channels for the selected band.

[0053] At the core of the model is a multimodal spatiotemporal attention encoder, the structure of which is as follows: Branch feature extraction subnetwork: Electrical Engineering Branch: Analysis of the spliced ​​voltage and current waveforms A one-dimensional convolutional neural network (1D-CNN) is applied. Through multiple layers of convolution and pooling, local fluctuation patterns are extracted, and an electrical feature vector is output. .

[0054] Image branch: on the image sequence tensor This method applies a lightweight 3D convolutional neural network (3D-CNN) to simultaneously capture the spatial texture features of a single frame and the spatiotemporal evolution features between frames, outputting an image feature vector. .

[0055] Spectral branch: for spectral background sequences By applying recurrent neural networks (RNNs) or their variants (such as LSTMs), the temporal evolution patterns of background noise are captured, and spectral feature vectors are output. .

[0056] Cross-modal attention fusion module: Extracts high-level features from each branch. Subsequently, a cross-attention mechanism is introduced to deeply fuse multimodal information. Taking image features as a query to retrieve electrical features as an example, the calculation process is as follows:

[0057] in This is a learnable weight matrix. In this way, the model can adaptively learn and quantify the coupling relationships between signals of different modalities. For example, the model can learn whether "a sharp current spike corresponds precisely in time to a brightness change in a specific region of an image," and this degree of cross-modal "consistency" or "mismatch" is itself a strong indicator of stability. Finally, through multiple cross-attention operations, the features of all branches are fused into a high-dimensional feature vector. This serves as a comprehensive representation of the current discharge state.

[0058] To guide the model to learn physical representations and avoid generating physically impossible stability scores, an auxiliary constraint term based on discharge physics is introduced during the model training process. The overall loss function... Defined as:

[0059] It is the main loss function (e.g., the mean squared error used for rating regression). It is a weighting coefficient that balances the various losses.

[0060] Energy balance constraints This constraint requires that the model learn some kind of representation internally (e.g., from fused features). The scalar obtained by projection through an auxiliary fully connected layer. The estimated instantaneous input power should be consistent with the actual input power. There is a smooth correlation. The loss can be defined as the negative value of the Pearson correlation coefficient between the two within the time window, which encourages the model to learn a representation consistent with the trend of input power change.

[0061]

[0062] Morphological-electrical correlation constraints This constraint utilizes a known physical prior: the impedance of the discharge circuit typically increases as plasma contracts. It defines the feature component representing the plasma "area" in the image features. and the characteristic component representing "impedance" in electrical characteristics. The constraint states that the two are negatively correlated. The loss can be defined as... That is, if the two are actually positively correlated, then a penalty is imposed.

[0063]

[0064] Through the aforementioned end-to-end model with physical constraints, the system can autonomously learn more fundamental and robust stability representations from massive amounts of unlabeled data, and can provide more reasonable stability scores even when faced with new perturbations that have not appeared in the training set.

[0065] To accurately pinpoint the root causes of instability while assessing discharge stability, thereby improving the efficiency and targeting of subsequent parameter coordination adjustments, a multi-task diagnostic model is introduced in another preferred embodiment of the invention to perform simultaneous stability scoring and root cause analysis. The core of this embodiment lies in learning the common instability characteristics of different root causes through a shared feature extraction layer, and then outputting refined scores and root cause probability distributions through independent task heads.

[0066] During the offline training phase, perform the following steps: 1. Data preparation and time-series dynamic feature extraction This implementation first constructs a dataset for model training. For each sampling window of 2-5 seconds, a set of high-level temporal dynamic features is extracted from the synchronously acquired multimodal raw data. These features are designed to characterize subtle changes in the discharge state from different dimensions, specifically including the following four categories: Electrical dynamic characteristics: Instantaneous phase difference The discharge voltage was analyzed using the Hilbert transform. and discharge current Perform analytical signal transformation, extract the instantaneous phase, calculate the phase difference sequence, and then take the mean or standard deviation within the window to reflect the change in impedance angle.

[0067]

[0068] Impedance variation trend Calculate the discharge circuit impedance within the window. The linear regression slope captures the slow drift of impedance.

[0069] Current noise power For the detrended current signal Perform a Fast Fourier Transform to extract the average power of the main noise peaks in the 50Hz–1kHz frequency band.

[0070] Spatiotemporal characteristics of images: Center of gravity jitter amplitude Threshold segmentation is performed on each frame of the plasma bright spot image, and the centroid coordinates of the bright spot are calculated. Then calculate the standard deviation of the centroid's trajectory in the X and Y directions within the window.

[0071] Area change rate Calculate the area of ​​the bright spot The rate of change over time within the window is calculated as the slope of the linear regression divided by the average area.

[0072] Edge shadow ratio Canny edge detection is performed on the image to count the proportion of newly appearing non-plasma edge pixels in the fixed neighborhood of the electrode tip, which is used to quantitatively indicate the degree of crystal adhesion on the electrode surface.

[0073] Spectral and auxiliary characteristics: Background intensity mutation index Calculate the difference in spectral background intensity of the selected band (e.g., 225–235 nm) at the beginning and end of the window, and then normalize it.

[0074] Electrode temperature change rate Electrode temperature signal Linear fitting is performed within the window to obtain the rate of temperature change.

[0075] Flow rate setpoint The current solution flow rate setting value in the window is used as the input for operating condition context information.

[0076] Cross-modal mutual information features: Electro-optical Interconnection Calculate the high-frequency components of the current signal. Image frame-by-frame texture entropy Mutual information quantifies the degree of synchronicity between electrical and optical perturbations. The formula for calculating mutual information is:

[0077] in and Represent and Discretized values.

[0078] 2. Root Cause Category and Stability Score Labeling Label each sampling window with two types of labels: Refined stability rating labels This is derived by inversely calculating the signal-to-noise ratio (SNR) or relative standard deviation (RSD) of characteristic peak intensities from the spectral data subsequently acquired within this window. For example, the mapping rule can be set as follows: RSD < 2% corresponds to a score of 0.9, RSD > 10% corresponds to a score of 0.4, and linear interpolation is used in the intermediate interval.

[0079] Root cause category label Combining operation logs, temporal characteristic trends, and original signal recordings, domain experts comprehensively determine the dominant root cause within the window. Preset root cause categories may include: C1 - Electrode surface crystal adhesion, C2 - Solution flow rate fluctuations, C3 - Bubble interference, C4 - Thermal drift of electrode spacing, and C5 - Sudden changes in sample matrix salinity.

[0080] 3. Construction of a multi-task diagnostic model The model adopts a multi-task learning architecture with a shared underlying layer and independent task heads: Input layer: Receives the concatenated data. 3D time-series dynamic feature vector .

[0081] Shared feature extraction layer: Consists of 2-3 fully connected network layers, each followed by a ReLU activation function and Dropout regularization, used to learn unstable representations of common root causes. .

[0082] Task Head 1 – Stability Scoring Regression Head: A fully connected layer is added after the shared layer, with one output node and a sigmoid activation function. The output is mapped to the range of 0-1 to obtain a refined stability score. .

[0083] Task Header 2 – Root Cause Classification Header: A fully connected layer follows the shared layer, and the number of output nodes is the preset number of root cause categories. (In this embodiment) The activation function is Softmax, and the output is the root cause probability distribution. ,in .

[0084] 4. Loss Function and Model Training The model's total loss function is a weighted sum of the regression loss and the classification loss:

[0085] in and To balance the weighting coefficients of the importance of the two tasks (e.g.) (emphasizing the accuracy of the scoring).

[0086] Regression loss The system employs a smooth L1 loss (Huber Loss) to penalize the rating prediction error, balancing the stability of the mean squared error with the robustness of the absolute error to outliers.

[0087]

[0088] Classification loss We employ labeled smooth cross-entropy loss to improve the model's generalization ability and prevent overfitting.

[0089]

[0090] in , This is the label smoothing coefficient (usually taken as 0.1).

[0091] The model is trained using the AdamW optimizer, and the overall loss is monitored on the validation set. An early stopping strategy is implemented to prevent overfitting. After training, the model has the ability to simultaneously output a refined stability score and root cause probability distribution.

[0092] During the online real-time inference phase, the following steps are performed: In actual system operation, the online inference process is as follows: 1. Real-time data buffering and feature computation The system maintains a first-in, first-out (FIFO) data buffer, continuously loading the latest multimodal raw data. Each time a new batch of data is received (e.g., every 0.1 seconds), feature calculation is triggered. Based on the raw data within the current sliding window (2-5 seconds in length), all the aforementioned time-series dynamic features are updated to obtain a real-time feature vector. .

[0093] 2. Model Inference Real-time feature vectors Input a pre-loaded multi-task diagnostic model; after forward propagation, output a refined stability score. Root cause probability distribution vector

[0094] 3. Decision-making logic The control unit makes the following decisions based on the model output: Stability assessment: If For a sustained period (e.g., more than 0.5 seconds) below a preset action threshold If the value is 0.65, the system is determined to have entered an unstable state, triggering the subsequent adaptive control process.

[0095] Root cause diagnosis: from probability distribution Find the highest probability and their corresponding root cause categories .Will Compared with the preset confidence threshold (For example Compare: like Then it is determined that the current principal root is because It has a high confidence level. The system categorizes root causes. Send to the control strategy generation module, which is used to trigger targeted adjustment strategies (for example, if electrode crystallization is diagnosed, the flow rate is preferentially increased).

[0096] If , it is determined that the current is a composite disturbance or an unknown disturbance, and the confidence level is low. At this time, the system can return to the preset general PID adjustment strategy or adopt a conservative adjustment strategy with probability weighting combination to avoid incorrect adjustment caused by misdiagnosis.

[0097] In an embodiment of the present invention, based on step S3, a possible embodiment will be given below to non-restrictively elaborate on its specific implementation scheme.

[0098] First, the system pre-sets three stability thresholds, namely the second threshold S2, the first threshold S1, and the fault threshold S f , and the three satisfy S2 > S1 > S f . According to the actual application scenario, in this embodiment, S2 = 0.75, S1 = 0.65, S f = 0.50. These thresholds can be adaptively adjusted according to the sample matrix characteristics or operating conditions.

[0099] The system obtains the current stability score S in each control cycle (for example, every 0.1 second) and compares it with the above three thresholds in sequence, and determines the discharge state according to the following rules: S301. Stable state determination If the stability score S ≥ S2, it is determined that the discharge is in a stable state. At this time, the plasma fluctuation is slight, the excitation condition is stable, and it is suitable for spectral quantitative analysis. The system enters the steady-state acquisition mode, triggers the spectrometer to collect emission spectral data, and performs subsequent preprocessing and concentration analysis. At the same time, the system maintains the current discharge parameters unchanged and only performs routine monitoring.

[0100] S302. Critical stable state determination If S1 ≤ S < S2, it is determined that the discharge is in a critical stable state. At this time, there is a slight fluctuation in the plasma, but it has not reached the level that requires active intervention. The system does not trigger parameter adjustment, nor does it perform spectral acquisition, but maintains the current parameters and continues to monitor. If the score rises above S2 in subsequent cycles, it transfers to the stable state; if it continues to drop below S1, it transfers to the unstable state. The setting of this critical region avoids frequent adjustment caused by instantaneous fluctuations and improves the stability of the system.

[0101] S303. Unstable state determination If S fIf S < S1, it is determined that the discharge is in an unstable state. At this time, the plasma has shown obvious scintillation, contraction or jitter. If the spectrum is collected in this state, it will lead to unreliable quantitative analysis results. The system immediately triggers an adaptive control process to coordinately adjust at least two parameters among the discharge voltage, solution flow rate and electrode spacing to suppress the unstable phenomenon. During the adjustment process, the system continuously updates the stability score and monitors its change trend in real time.

[0102] S304. Fault state determination If S ≤ S f it is determined that the discharge is in a fault state. At this time, the plasma has been seriously unstable and may even go out. Conventional parameter adjustment is difficult to restore. The system immediately triggers a maintenance alarm, sends an alarm message to the host computer or DCS system through the communication interface, and at the same time generates an operation log containing the current operating parameters, historical score curve and diagnostic information, prompting the operator to perform electrode cleaning, solution replacement or other maintenance operations.

[0103] To prevent misjudgment caused by instantaneous noise interference, duration verification can be introduced into all the above state determinations. For example, only when the stability score meets the conditions of a certain interval for 3 consecutive control cycles, the corresponding discharge state is formally determined. This mechanism effectively improves the reliability and anti-interference ability of state determination.

[0104] In an embodiment of the present invention, based on step S4, a possible embodiment will be given below to non-restrictively elaborate on its specific implementation scheme.

[0105] According to the determined discharge state, corresponding closed-loop control operations are executed: if it is determined that the discharge is in a stable state, trigger spectrum collection and analyze the collected spectrum to obtain element detection results; if it is determined that the discharge is in an unstable state, coordinately adjust at least two discharge parameters until the discharge returns to a stable state; if it is determined that the discharge is in a fault state, trigger a maintenance alarm.

[0106] 1. Root cause confidence judgment and strategy selection When the system triggers adaptive control, it first obtains the root cause probability distribution vector output by the multi-task diagnostic model at the current moment , which respectively correspond to five preset root cause categories: crystal attachment on the electrode surface, solution flow rate fluctuation, bubble interference, thermal drift of electrode spacing, and sudden change in sample matrix salinity. The system extracts the maximum probability value and its corresponding root cause category .

[0107] Compare with the preset confidence threshold (in this embodiment, take ): High-confidence diagnosis: If Then it is determined that the current principal root is because The diagnostic results are highly reliable. The system is based on root cause category. Trigger targeted control strategies and call the corresponding adjustment schemes from the preset root cause-control mapping library.

[0108] Low confidence diagnosis: If If the current disturbance is determined to be a complex disturbance or an unknown disturbance, and a single root cause cannot fully explain the current unstable state, the system will revert to the preset general control strategy and use a classic control algorithm for multi-parameter coordinated adjustment to avoid misdiagnosis and incorrect adjustment.

[0109] 2. Root cause-driven targeted control strategies The system pre-sets differentiated control strategies for different root cause categories, as illustrated below: If the diagnosis is crystal adhesion on the electrode surface (C1): the system prioritizes increasing the solution flow rate to inhibit further crystal growth by increasing the liquid scouring force, while simultaneously fine-tuning the voltage to maintain the plasma excitation intensity. Flow rate adjustment amount It is positively correlated with the severity of crystallization; the more severe the crystallization, the greater the increase in flow rate.

[0110] If the diagnosis is solution flow rate fluctuation (C2): the system prioritizes adjusting the drive parameters of the micro-injection pump, such as the PID coefficient or the stepper motor pulse frequency, to stabilize the flow rate feedback. Simultaneously, the electrode spacing is appropriately increased to broaden the plasma stability range and enhance the tolerance to flow rate fluctuations.

[0111] If the diagnosis is bubble interference (C3): the system will first briefly increase the discharge voltage to enhance the plasma energy to "break down" the instantaneous circuit caused by the bubble, while simultaneously fine-tuning the electrode spacing. Once the score recovers, the original parameters will be restored.

[0112] If the diagnosis is electrode spacing thermal drift (C4): The system dynamically adjusts the electrode spacing based on the temperature change rate fed back by the electrode temperature monitoring unit to compensate for the spacing change caused by thermal expansion. Compensation amount With temperature change A linear relationship exists: ,in The thermal compensation coefficient is determined experimentally.

[0113] If the diagnosis is a sudden change in sample matrix salinity (C5): the system adopts a combined voltage and flow rate regulation strategy, appropriately increasing the voltage and simultaneously decreasing the flow rate to expand the plasma stability range and enhance adaptability to high-salt matrices.

[0114] 3. General Control Strategy When insufficient confidence triggers the general control strategy, the system employs a classical control algorithm to coordinately adjust at least two parameters among the discharge voltage, solution flow rate, and electrode spacing. This implementation supports multiple control algorithms, which can be selected or switched according to actual needs: (1) Incremental PID control algorithm The system uses stability scoring bias As input, where Score the stability of the target (e.g., take the second threshold). The incremental adjustment increment of each parameter output by the incremental PID controller:

[0115] in The PID coefficients for each parameter were obtained through experimental tuning. The advantage of incremental PID is that the output is only an increment, so even in the event of a fault, the parameters will not change drastically, resulting in higher safety.

[0116] (2) Takagi-Sugeno fuzzy control algorithm The system will determine the stability score deviation. and its rate of change As fuzzy input variables, the adjustment amounts of each parameter are obtained through fuzzification, fuzzy inference, and defuzzification. The fuzzy rule base is constructed based on expert experience; for example, if... For negative large and If the negative value is large, then significantly increase the voltage and slightly increase the spacing; if negative and If it is zero, then the voltage and flow rate are increased by a moderate amount; if For negative small and If the value is small, then the voltage is fine-tuned while other parameters are maintained.

[0117] The advantage of fuzzy control algorithms lies in their ability to handle the nonlinearity and time-varying nature of the discharge process well without relying on precise mathematical models.

[0118] (3) Model predictive control algorithm based on quadratic programming The system establishes a simplified prediction model for the discharge process, using the current moment as the starting point to predict the future. The stability score of each step changes. Under the premise of satisfying physical constraints (e.g., voltage range 200-1100V, flow rate range 0.1-2.0 mL / min, spacing range 0.6-1.2 mm), a quadratic programming problem is solved to optimize the future... The algorithm employs a step-by-step control sequence to make the predicted score as close as possible to the target value, while penalizing drastic changes in the control input. It performs rolling optimization within each control cycle, executing only the first step of the control input, demonstrating foresight and optimality.

[0119] 4. Physical constraints and execution of coordinated regulation Regardless of the control algorithm used, the system adheres to the following physical constraints when outputting the adjustment amount: The adjustment range of each parameter is strictly limited to the safe range allowed by the hardware; The change rate of each parameter per unit time is limited to the preset maximum change rate to prevent overload or sudden discharge of the actuator. During the adjustment process, the stability score is continuously monitored. If the score rises back to the stable range, the adjustment is terminated. If there is no improvement after multiple adjustments and the score continues to drop below the fault threshold, a maintenance alarm is triggered.

[0120] After the adjustment amount is calculated, the MCU main control unit sends it to the FPGA coprocessing unit via the SPI bus. The FPGA converts the digital control amount into a PWM signal or a DAC analog signal, which drives the high-voltage power supply module, the micro-injection pump and the stepper motor actuator respectively to complete the coordinated adjustment of the discharge voltage, solution flow rate and electrode spacing.

[0121] In another preferred embodiment of the present invention, after the system determines that the discharge is in a stable state and triggers spectral acquisition, a series of preprocessing operations are performed on the acquired emission spectral data, and then the data is input into a deep learning model for quantitative analysis to obtain the concentration detection results of the target element.

[0122] 1. Spectral data preprocessing Please refer to Figure 6 The raw spectral data acquired by the spectrometer contains various interfering components, including continuous background drift, wavelength shift, random noise, and short-term fluctuations. To improve the accuracy and stability of subsequent quantitative analysis, this implementation method designs a multi-stage preprocessing workflow, where each step can be adaptively combined based on the actual signal-to-noise ratio and stability score.

[0123] A polynomial least squares fitting method was used to subtract continuous background and dark current. For each spectral window, a background region without characteristic spectral lines (such as the 200-220 nm band) was selected as the fitting baseline, and a second- or third-order polynomial was used to fit the background to obtain the background curve. Then from the original spectral intensity After subtracting the background, the background-corrected spectrum is obtained. This method can effectively eliminate slow baseline rise caused by factors such as salt spray and specular scattering.

[0124] Due to changes in ambient temperature or mechanical drift of the spectrometer, characteristic spectral lines may experience slight wavelength shifts, affecting the accuracy of quantitative analysis. This implementation method employs a multi-point linear regression method based on internal standard spectral lines for wavelength correction. Specifically, 2-3 known and stable internal standard spectral lines (such as Ar spectral lines or matrix element spectral lines of known concentration in the sample) are selected, their actual pixel positions are measured, and a linear regression model is established with the standard wavelength positions. Then, the wavelength axis of the entire spectrum is corrected to ensure that key characteristic lines are located in the correct pixel range.

[0125] The Savitzky-Golay convolutional filtering method is used to smooth the spectrum, suppressing high-frequency random noise while preserving the peak shape and position. The filter parameters (window width and polynomial order) can be adaptively adjusted according to the spectral resolution and noise level. In this embodiment, the typical settings are a window width of 11 points and a polynomial order of 3.

[0126] To further improve the signal-to-noise ratio, especially enhancing the recognizability of weak spectral lines, a Daubechies wavelet basis is used for multi-scale decomposition of the spectrum. Soft or hard thresholding is applied to the high-frequency detail coefficients in the wavelet coefficients, setting coefficients below the threshold to zero before wavelet reconstruction. This method effectively separates signal from noise, and is particularly suitable for extracting weak signals under strong background noise.

[0127] Based on the current stability score and signal-to-noise ratio threshold, multi-frame spectra are adaptively superimposed. When the stability score is high (e.g. ), with fewer frames superimposed (e.g., 3-5 frames) to improve analysis speed; when the stability score is within the critical range (e.g. To ensure a good signal-to-noise ratio, the number of superimposed frames is increased (e.g., 8-10 frames). The superimposed spectral intensity is the arithmetic mean of the wavelengths corresponding to the spectra of each frame, effectively reducing the impact of short-term random fluctuations.

[0128] To eliminate absolute intensity variations caused by fluctuations in light source intensity and differences in sample volume, the spectrum is normalized. Normalization can be based on the total ion current intensity or proportionally based on the intensity of a stable internal standard line. Subsequently, segmental standardization is performed on specific analytical bands to ensure that the intensity mean of each wavelength point is 0 and the standard deviation is 1, facilitating the training and inference of subsequent deep learning models.

[0129] After the above preprocessing steps are executed sequentially, the signal-to-noise ratio of typical characteristic lines can be increased to more than twice the original value, and weak spectral lines under high salt and high background interference can still maintain distinguishable peak shape and peak area.

[0130] 2. Quantitative Analysis Model Please refer to Figure 7The preprocessed spectral data is input into a one-dimensional convolutional neural network regression model to achieve an end-to-end mapping from spectral intensity sequence to target element concentration.

[0131] The model's input consists of two parts: Spectral intensity sequence: The spectral intensity values ​​of the preprocessed full spectrum or selected characteristic bands form a one-dimensional vector. ,in This represents the number of wavelength points.

[0132] Matrix eigenvectors: To compensate for spectral interference caused by different sample matrices, auxiliary features reflecting sample characteristics are introduced, including total dissolved solids concentration, conductivity, average background wavelength, magnesium-lithium ratio, etc., forming an auxiliary vector. .

[0133] The input layer concatenates the spectral sequence and matrix features into a complete feature vector. .

[0134] Network structure: A one-dimensional convolutional neural network consists of the following layers: Input layer: Receives the concatenated feature vectors.

[0135] Multiple convolutional-pooling modules: Each module contains a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, and a max-pooling layer. The convolutional layers extract local spectral features, such as peak shape, half-width at half-maximum (WHM), and slope, using multiple convolutional kernels. The pooling layers reduce the feature dimensionality and introduce translation invariance. In this embodiment, three convolutional-pooling modules are used, with the number of convolutional kernels being 32, 64, and 128 respectively, and the kernel size being 5-7 points.

[0136] Global average pooling layer: The feature map output by the last convolutional layer is globally average pooled along the wavelength dimension, which condenses and fuses key information from different bands, while reducing the parameters of the fully connected layer to prevent overfitting.

[0137] Fully connected hidden layer: contains 128 neurons, uses the ReLU activation function, and adds Dropout (dropout rate 0.3) to further improve generalization ability.

[0138] Output layer: Contains two branches: one is the target element concentration output, with the number of neurons equal to the number of elements to be tested (e.g., Li, Mg, K, Na, Ca, a total of 5), and the output concentration estimate is obtained using a linear activation function. Secondly, predict the uncertainty output by using the Softplus activation function to output a non-negative uncertainty estimate. , which represents the confidence level of the predicted concentration of each element.

[0139] The model is trained using a large amount of sample spectral data covering different concentration ranges and matrix compositions. The loss function is the joint loss of concentration prediction error and uncertainty estimation.

[0140] This loss function encourages the model to provide a reasonable uncertainty estimate for the prediction results while ensuring accuracy: the uncertainty increases when the prediction error is large, and decreases when the prediction error is small. The model is trained using the Adam optimizer, with an initial learning rate of 0.001 that gradually decreases with each training epoch.

[0141] In the actual detection process, the system inputs the preprocessed spectral data into the trained model in real time, and the model outputs the data synchronously after forward propagation. Predicted concentration values ​​of each target element

[0142] Prediction uncertainty of each target element

[0143] Concentration results are sent to a host computer or DCS system via industrial communication protocols (such as Modbus-TCP, OPC-UA) for process monitoring and linkage control; uncertainty information is used to assess the reliability of the test results. If the uncertainty is too high, the system can automatically trigger a retest or prompt manual verification.

[0144] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A method for adaptive control of liquid cathode glow discharge optical emission spectroscopy, characterized in that, include: Real-time acquisition of multimodal state data during the glow discharge process of a liquid cathode, including discharge voltage signal, discharge current signal, plasma luminescence image, and continuous spectral background signal; The multimodal state data is input into a pre-trained stability assessment model, which outputs a stability score of the plasma for quantitatively assessing the current discharge state. The stability score is compared with at least one preset stability threshold to determine the current discharge state; Based on the determined discharge state, the corresponding closed-loop control operation is executed: if the discharge is determined to be in a stable state, spectral acquisition is triggered and the acquired spectrum is analyzed to obtain element detection results; if the discharge is determined to be in an unstable state, at least two discharge parameters are adjusted in a coordinated manner until the discharge returns to a stable state; if the discharge is determined to be in a fault state, a maintenance alarm is triggered.

2. The method of claim 1, wherein, Real-time acquisition of multi-mode state data during the glow discharge process of a liquid cathode, including: Real-time acquisition of voltage and current signals from the discharge circuit; Real-time image signals of the plasma luminescent area are acquired using an industrial camera; The continuous spectral background signal of plasma emission is acquired in real time using a spectrometer.

3. The method of claim 1, wherein, The stability score is generated by directly processing the multimodal state data through an end-to-end deep neural network model. The input layer of the deep neural network model receives a raw multimodal time-series data stream that is not manually extracted and is time-synchronized and aligned, including the raw high-frequency waveforms of discharge voltage and discharge current, raw image sequence fragments of the plasma luminescence region, and raw intensity sequences of the continuous spectral background of the selected band.

4. The method of claim 3, wherein, The deep neural network model includes a multimodal spatiotemporal attention encoder, which comprises: The branch feature extraction subnetwork is used to extract electrical fluctuation features from voltage waveforms and current waveforms, spatiotemporal evolution features from image sequences, and temporal evolution features from spectral background sequences, respectively. The cross-modal attention fusion module is used to perform a cross-attention mechanism between high-level features extracted from different branches to adaptively learn and quantify the coupling relationship and consistency between different modal signals, and use the fused features as a high-dimensional feature vector characterizing the discharge state. The training process of the deep neural network model introduces auxiliary constraint terms based on discharge physics knowledge as part of the loss function; the auxiliary constraint terms include at least: energy balance constraints for constraining the relationship between the internal characterization of the model and the instantaneous input power, and morphological-electrical correlation constraints for constraining the correlation between plasma morphological characteristics and discharge circuit impedance characteristics.

5. The method of claim 4, wherein, The stability score is generated synchronously through a multi-task diagnostic model; the multi-task diagnostic model adopts a network architecture that combines a shared feature extraction layer with independent task heads; the shared feature extraction layer is used to extract temporal dynamic features from the current sampling window. The multi-task diagnostic model includes a first task head and a second task head; the first task head is a regression head, which uses the Sigmoid activation function to output a refined stability score in the range of 0-1; the second task head is a classification head, which uses the Softmax activation function to output the probability distribution of multiple preset root cause categories.

6. The method of claim 5, wherein, The time-series dynamic features include: Electrical dynamic characteristics include the instantaneous phase difference between the discharge voltage and current waveforms, the impedance change trend of the discharge circuit, and the main noise power of the high-frequency components of the current. The spatiotemporal features of the image include the jitter amplitude of the plasma bright spot centroid, the rate of change of the bright spot area, and the proportion of edge shadows used to indicate the crystal adhesion at the electrode tip; Spectral and auxiliary characteristics, including abrupt changes in continuous spectral background intensity, rate of change of electrode temperature, and setpoints for solution flow rate; Cross-modal mutual information features are used to quantify the degree of synchronization between high-frequency components of electrical signals and image texture entropy.

7. The method of claim 5, wherein, Coordinated adjustment of at least two discharge parameters, including: Extract the maximum probability value and its corresponding root cause category from the probability distribution of root cause categories, and compare the maximum probability value with a preset confidence threshold: If the maximum probability value is higher than the confidence threshold, the current primary root cause is determined to be the root cause category corresponding to the maximum probability, and an adaptive control strategy for the root cause category is triggered. If the maximum probability value is not higher than the confidence threshold, the current disturbance is determined to be a composite disturbance or an unknown disturbance, and a preset general control strategy is triggered.

8. The method of claim 7, wherein, The general control strategy includes: A preset control algorithm is used to calculate the adjustment amount of at least two parameters among discharge voltage, solution flow rate and electrode spacing; the preset control algorithm is at least one of incremental proportional-integral-derivative control algorithm, Takagi-Sugeno fuzzy control algorithm or model predictive control algorithm based on quadratic programming solution; According to the adjustment amount, the at least two parameters are coordinated to suppress plasma flickering, contraction or extinction, so that the plasma can be restored to a stable discharge state.

9. The method of claim 1, wherein, The stability score is compared with at least one preset stability threshold to determine the current discharge state, including: The stability score is compared with a preset first threshold, a second threshold, and a fault threshold, and the current discharge state is determined based on the comparison results. If the stability score is greater than or equal to the second threshold, the discharge is determined to be in a stable state. If the first threshold is less than or equal to the stability score and less than the second threshold, the discharge is determined to be in a critical stable state. The current parameters are maintained and monitoring continues. If the fault threshold is less than the stability score and less than the first threshold, the discharge is determined to be in an unstable state. If the stability score is less than or equal to the fault threshold, the discharge is determined to be in a fault state, triggering a maintenance alarm.

10. The method of claim 1, wherein, The collected spectra are analyzed to obtain elemental detection results, including: Preprocessing operations on spectral data; the preprocessing operations include at least one of the following: background subtraction based on polynomial least squares fitting; wavelength correction based on multi-point linear regression of internal standard spectral lines; smoothing based on Savitzky-Golay filtering; threshold denoising based on Daubechies wavelet; multi-frame spectral stacking and averaging; intensity normalization and segment standardization; The preprocessed spectral data is input into a pre-trained quantitative analysis model; the quantitative analysis model is a one-dimensional convolutional neural network regression model, used to realize the nonlinear mapping from spectral intensity sequence to target element concentration; the model outputs the concentration detection result of at least one target element and its corresponding prediction uncertainty.

Images