Multi-index monitoring system for heavy metals in seawater direct current cooling and desulfurization discharge outlet

Abstract

The application discloses a seawater direct-flow cooling and desulfurization outlet heavy metal multi-index joint monitoring system, comprising: a sampling and pretreatment module; a mass spectrometry detection module; a time-frequency domain multi-index extraction module, which maps the original sequence data of each heavy metal element into a multi-dimensional index tensor containing time domain statistical characteristics and frequency domain energy distribution characteristics; a graph model index optimization module; a multi-index fusion module; an abnormal decision and early warning module. The seawater direct-flow cooling and desulfurization outlet heavy metal multi-index joint monitoring system provided by the application maps the heavy metal concentration time sequence into a time-frequency domain multi-dimensional index, adopts a graph model to perform topological structure characterization and dynamic median graph matching on the index sequence, obtains a similarity score based on an edge weight distance, and then adaptively and weightedly fuses into a single comprehensive monitoring index, finally outputs an early warning signal based on a statistical confidence interval and a continuous decision rule, and forms a complete processing link from original detection data to a final decision.

Description

Technical Field

[0001] This application belongs to the field of environmental monitoring technology, specifically involving a joint monitoring system for multiple indicators of heavy metals at seawater direct cooling and desulfurization discharge outlets. Background Technology

[0002] Seawater direct-flow cooling systems and desulfurization systems are important components of coastal thermal power plants, nuclear power plants, and large chemical enterprises. The warm wastewater discharged from seawater direct-flow cooling outlets may carry corrosion products (such as nickel, copper, and zinc) from heat exchange equipment, while the wastewater from desulfurization systems accumulates various heavy metals such as chromium, arsenic, cadmium, mercury, and lead due to the flue gas scrubbing process. These heavy metals, discharged into the nearshore environment with the wastewater, pose a potential risk to the marine ecosystem. Therefore, joint monitoring of heavy metals at both types of discharge outlets is of significant environmental importance.

[0003] Currently, existing monitoring technologies have the following main shortcomings:

[0004] First, the sampling and pretreatment processes failed to simultaneously address the dual characteristics of seawater (high salinity and high flow rate) and desulfurization wastewater (acidity and high suspended solids). Traditional methods often use hydrochloric acid or perchloric acid for digestion, resulting in extremely high Cl⁻ concentrations in the digestate, which can cause problems during subsequent inductively coupled plasma mass spectrometry (ICP-MS) analysis. 40 Ar 35 Cl + Severe polyatomic ion interference directly affects the accurate determination of elements such as arsenic, vanadium, and chromium; at the same time, volatile elements such as mercury and germanium are easily lost by forming chlorides in chlorine-containing systems, reducing the recovery rate of trace elements.

[0005] Second, most existing monitoring systems only output instantaneous concentration values ​​of each heavy metal, without in-depth analysis of the concentration time series. In reality, heavy metal emissions from discharge outlets often exhibit non-stationary, nonlinear, and multi-scale coupled characteristics, making it difficult to capture early degradation signals and intermittent abnormal events with simple concentration thresholds or trend analysis.

[0006] Third, there is a lack of effective information fusion mechanisms among multiple heavy metal elements and multiple statistical indicators. The concentration magnitudes of different elements vary greatly, and a single indicator or simple weighting can easily lead to the phenomenon of "large numbers eating small numbers". Moreover, fixed weights cannot adapt to changes in operating conditions, resulting in insufficient robustness of anomaly detection.

[0007] Fourth, anomaly decisions often rely on single-point threshold judgments, which can easily lead to frequent false alarms due to sensor noise or occasional fluctuations in water quality. At the same time, they lack dynamic adaptive statistical benchmarks based on historical normal conditions, making it difficult to meet the requirements of industrial sites for both low false alarm rates and high detection rates. Summary of the Invention

[0008] This application provides a joint monitoring system for multiple indicators of heavy metals at seawater DC cooling and desulfurization discharge outlets, aiming to solve the problems in the existing technology, such as trace heavy metal loss and mass spectrometry interference caused by chlorine-containing digestion systems, difficulty in capturing non-stable emission characteristics by single concentration thresholds, feature submersion caused by magnitude differences when multiple elements and multiple indicators are fused, and high false alarm rate caused by single-point judgment.

[0009] A combined monitoring system for multiple heavy metal indicators at seawater direct current cooling and desulfurization discharge outlets includes:

[0010] The sampling and pretreatment module collects water samples from the seawater DC cooling outlet and desulfurization outlet and performs digestion pretreatment to output a homogeneous, low-matrix-interference sample solution.

[0011] The mass spectrometry detection module, whose input is connected to the output of the sampling and preprocessing module, performs simultaneous quantitative detection of multiple target heavy metal elements in the sample solution under kinetic energy discrimination collision mode, and outputs the original sequence data of the concentration of each element changing over time.

[0012] The time-frequency domain multi-index extraction module has its input end connected to the output end of the mass spectrometry detection module. It maps the original sequence data of each heavy metal element into a multi-dimensional index tensor containing time-domain statistical features and frequency-domain energy distribution features. The multi-dimensional index tensor is organized according to heavy metal element, sliding window and index type.

[0013] The graph model index optimization module has its input end connected to the output end of the time-frequency domain multi-index extraction module. It independently constructs an undirected complete graph for each index sequence in the multi-dimensional index tensor, calculates the graph model sequence, and obtains the similarity score between each graph model and the historical normal graph model based on edge weight distance using dynamic median graph matching, and outputs the similarity score tensor.

[0014] The multi-indicator fusion module, whose input is connected to the output of the graph model index optimization module, normalizes the similarity score sequences of all heavy metal elements and all time-frequency domain indicators, calculates the point-by-point mean and point-by-point maximum, and uses the standard deviation method to adaptively determine the weight factor to fuse the point-by-point mean and point-by-point maximum, and outputs a single comprehensive monitoring index time series.

[0015] The abnormal decision-making and early warning module has its input end connected to the output end of the multi-indicator fusion module. It constructs a confidence interval based on the statistical parameters of the comprehensive monitoring indicator time series in the preset historical normal segment, performs hypothesis testing on the indicator value at the current moment, and outputs multi-level early warning signals after confirming the abnormality by continuously exceeding the counter.

[0016] Optionally, the sampling and preprocessing module includes:

[0017] The sampling unit is used to install automatic synchronous samplers triggered by the same timing controller at the outlet of the main drainage channel of the seawater direct cooling system and the wastewater clarification tank of the desulfurization system. The inner wall of the sampling pipeline is lined with polytetrafluoroethylene or perfluoroalkoxy resin.

[0018] An online filtration unit, whose input end is connected to the output end of the sampling unit, is used to remove suspended particulate matter with a particle size greater than 0.45μm and is equipped with an automatic backflushing cleaning function;

[0019] The microwave digestion unit has its input end connected to the output end of the online filtering unit, and uses... The ternary digestion system, without the addition of any chlorine-containing reagents, digests the filtered water sample through a stepped temperature increase program. After the digestion solution is brought to a final volume, the test sample solution is output.

[0020] Optionally, the mass spectrometry detection module is specifically used for:

[0021] High-purity helium was introduced into a hexapole collision reactor, and the reactor was operated in kinetic energy discrimination mode. The helium flow rate was preset to ensure... 57 The signal-to-back ratio of Fe reaches its peak flow rate value;

[0022] For each target heavy metal element, one or more low-interference isotopes are selected, and a high-resolution scanning method is used.

[0023] A dual internal standard online calibration strategy was adopted, with the internal standard solution containing 5 μg / L of [unclear - likely a specific ingredient or solution]. 103 Rh and 5 μg / L 193 Ir is mixed online with the sample solution at a 1:1 volume ratio through a three-way valve and then enters the nebulizer;

[0024] Output the time-element concentration matrix to the time-frequency domain multi-index extraction module. The time-element concentration matrix includes the sampling timestamp, sample identification code, and concentration values ​​of each heavy metal element.

[0025] Optionally, the time-frequency domain multi-index extraction module is specifically used for:

[0026] First-order difference detrending preprocessing was performed on the original concentration time series of each heavy metal element;

[0027] A fixed-length sliding window is used to frame the detrended sequence. The window length L = 5000 data points and the window overlap length R = L / 2.

[0028] For each concentration subsequence within a window, time-domain statistical indicators are calculated, including mean, variance, root mean square, skewness, kurtosis, and maximum absolute value.

[0029] The power spectral density is obtained by performing a discrete Fourier transform on the concentration subsequence within each window, and the root mean square frequency is extracted as a frequency domain index.

[0030] The time-domain statistical indicators and frequency-domain indicators of each window are combined into a feature vector, and the multi-dimensional indicator tensors are output to the graph model indicator optimization module in the order of heavy metal elements and windows.

[0031] Optionally, the graph model index optimization module is specifically used for:

[0032] For each input single index sequence, a fixed-length sliding window is used to divide the data into frames. Each window contains j=10 consecutive data points, and the window sliding step size d=1, resulting in K-j+1 data segments.

[0033] For each data segment, construct an undirected complete graph, where nodes correspond to the values ​​of data points, and the weight of an edge is defined as the absolute difference between the corresponding data values ​​of the two connected nodes.

[0034] The most recent graph model before using the current graph model A median graph is calculated for each graph model, wherein the median graph is the graph model that minimizes the sum of edge-weighted distances to all other graph models;

[0035] The edge-weighted distance between the current graph model and the median graph is calculated as a similarity score, whereby the edge-weighted distance is defined as the sum of the absolute values ​​of the differences in the weights of corresponding edges between the two graph models.

[0036] The similarity scores of all heavy metal elements and all time-frequency domain indicators are organized into similarity score tensors and output to the multi-indicator fusion module.

[0037] Optionally, the multi-indicator fusion module is specifically used for:

[0038] For each similarity score sequence, perform min-max normalization independently to map its value range to the [0,1] interval;

[0039] At the same time, calculate the pointwise mean and pointwise maximum of all normalized similarity scores;

[0040] Calculate the standard deviation of a point-by-point mean series and the standard deviation of the pointwise maximum sequence and through Define adaptive weighting factors ;

[0041] The point-by-point mean and the point-by-point maximum value are... By integrating the data, a comprehensive monitoring index can be obtained. The time series of comprehensive monitoring indicators is output to the anomaly decision-making and early warning module.

[0042] Optionally, the anomaly decision-making and early warning module is specifically used for:

[0043] Take the first part of the comprehensive monitoring indicator sequence Using a point or a manually calibrated time period as the initial historical normal segment, calculate its sample mean μ and sample standard deviation σ;

[0044] Based on the Gaussian distribution assumption, the confidence interval for the normal state is defined as follows: , where k is the confidence factor, and its value ranges from 3 to 6;

[0045] Hypothesis testing is performed on the comprehensive monitoring indicators at the current moment: if the indicator value falls outside the confidence interval and exceeds it consecutively a certain number of times... If this happens, it is confirmed as an anomaly;

[0046] The warning levels are determined based on the degree to which the value exceeds the confidence interval: Level 1 warning is issued when the excess amount Δ > 0 and the number of consecutive excesses is 1 or 2; Level 2 warning is issued when the number of consecutive excesses is ≥ 3; and Level 3 warning is issued when the excess amount Δ > 3σ or the index value > 0.95.

[0047] The warning signal is sent to the outside world through digital output, human-machine interface display and event logging.

[0048] Optionally, the anomaly decision-making and early warning module is also configured with a sliding baseline window strategy: during continuous system operation, only the most recent baseline window is used. The sample mean μ and sample standard deviation σ are dynamically updated using a comprehensive monitoring indicator value that is judged to be normal, and abnormal points are not included in the update of the historical normal segment during the abnormal warning period.

[0049] Compared with the prior art, this application has at least the following beneficial effects:

[0050] The heavy metal multi-index joint monitoring system for seawater DC cooling and desulfurization discharge outlets provided in this application maps the time series of heavy metal concentrations into multi-dimensional time-frequency domain indicators, and uses a graph model to perform topological structure representation and dynamic median graph matching on the indicator series to obtain a similarity score based on edge weight distance. Then, it is adaptively weighted and fused into a single comprehensive monitoring indicator. Finally, it outputs an early warning signal based on statistical confidence intervals and continuous judgment rules, forming a complete processing link from raw detection data to final decision.

[0051] By utilizing the cascaded processing of time-frequency domain feature extraction and graphical model optimization, high-frequency random noise and baseline drift in the original concentration sequence are effectively suppressed, while the detectability of early degradation features and intermittent anomalies is enhanced, solving the problem of insufficient identification capability of traditional thresholding methods for non-stationary and nonlinear emission signals.

[0052] By employing maximum-minimum normalization and standard deviation method to adaptively weight multiple indicators, the problem of feature submersion caused by differences in magnitude among different heavy metal elements and statistical indicators is balanced, while also taking into account overall drift sensitivity and response speed to sudden anomalies.

[0053] The dynamic confidence interval based on the Gaussian distribution assumption, combined with the judgment rule of continuous exceedance counter, significantly reduces false triggers caused by sensor noise or occasional fluctuations in water quality at the preset false alarm rate. At the same time, it supports adaptive updates of the sliding reference window, which can adapt to baseline drift caused by equipment aging and seasonal changes. Attached Figure Description

[0054] Figure 1 A schematic diagram of the module connection of a combined monitoring system for multiple heavy metal indicators at a seawater DC cooling and desulfurization discharge outlet provided in one embodiment of this application;

[0055] Figure 2 To analyze isotopes and their main sources of interference and elimination methods. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0057] The seawater direct cooling and desulfurization discharge outlet heavy metal multi-index joint monitoring system provided in this application, such as Figure 1 As shown, it includes: a sampling and preprocessing module, a mass spectrometry detection module, a time-frequency domain multi-index extraction module, a graphical model index optimization module, a multi-index fusion module, and an anomaly decision-making and early warning module;

[0058] The sampling and pretreatment module is located at the upstream input of the entire monitoring system. Its output is a homogeneous, low-matrix-interference sample solution that can be directly imported into the mass spectrometry detection module. This module must simultaneously meet the sampling representativeness requirements of high-salt, high-flow-rate water bodies at seawater direct-cooling outlets and acidic, high-suspended-solids water bodies at desulfurization outlets, as well as the cleanliness requirements for subsequent KEDICP-MS determination of trace heavy metals (especially chromium, arsenic, cadmium, mercury, lead, cobalt, nickel, copper, zinc, gallium, and germanium).

[0059] The sampling and preprocessing module specifically includes the following units:

[0060] The sampling unit is equipped with automatic synchronous samplers at the main drainage channel of the seawater direct-flow cooling system and the outlet of the wastewater clarification tank of the desulfurization system. Both samplers are triggered by the same timing controller to ensure that water samples taken at the same time point reflect the combined discharge status of the two outlets. The inner wall of the sampling pipeline is lined with polytetrafluoroethylene (PTFE) or perfluoroalkoxy resin (PFA) to prevent metal adsorption and precipitation. At least 200 mL of water sample is collected per sampling cycle (e.g., every 1 to 4 hours), of which 100 mL is used for analysis within this system, and the remainder is retained for future reference.

[0061] After sampling, the water sample is filtered in an online filtration unit to remove suspended particulate matter larger than 0.45 μm. The filtration unit uses a disposable polyethersulfone (PES) membrane and is equipped with an automatic backflushing cleaning function to prevent cross-contamination caused by membrane clogging. The filtered water sample is then placed in a low-temperature storage tank (4℃±2℃) and sent to a microwave digestion unit within 30 minutes to maintain the stability of heavy metal speciation.

[0062] The pretreatment unit (microwave digestion) takes into account the seawater matrix. The content is extremely high (approximately 19 g / L). If hydrochloric acid (HCl) is introduced during digestion, or if a high concentration is present in the sample system, this could lead to serious contamination. This will cause serious problems during subsequent ICP-MS analysis. 40 Ar 35 Cl + interference 75 As + ,as well as 35 Cl 16 O + Interference 51 V + , 52 Cr + Meanwhile, to avoid the loss of volatile elements such as germanium (Ge) and cadmium (Cd) due to the formation of chlorides at high temperatures, this module does not use hydrochloric acid or perchloric acid, nor does it add any chlorine-containing reagents.

[0063] The digestion system adopts Ternary system. Specifically, accurately transfer 50.0 mL of filtered water sample into a PTFE digestion vessel, place it in a fume hood, and then add 5 mL of concentrated distillate that has undergone sub-boiling distillation. (ρ=1.42g / mL), 2mL high-purity (ρ=1.15g / mL) and 1mL of analytical grade (30%). The addition of [the substance] can enhance the oxidation capacity of the system, accelerate the decomposition of organic matter and some sulfides, and reduce the matrix effect of residual carbon in the digestion solution on subsequent mass spectrometry.

[0064] After sealing the digestion vessel with a lid, it was placed into the microwave digester. The digestion program employed stepped temperature control to avoid a rapid pressure rise due to a violent reaction. The set temperature-time program is as follows:

[0065] Step 1: Increase the room temperature to 120°C at a rate of 5°C / min and hold for 5 minutes;

[0066] Step 2: Increase the temperature to 180°C at a rate of 4°C / min and hold for 15 minutes;

[0067] Step 3: Increase the temperature to 200°C at a rate of 3°C / min and hold for 10 minutes.

[0068] The heating rate corresponding to the above procedure can be described by the following formula:

[0069]

[0070] In the formula, r is the heating rate (°C / min). For the target temperature, This is the starting temperature for the current step. Set the time for the heating stage. This stepped heating can effectively prevent the loss of low-boiling-point elements (such as Hg and Ge) due to instantaneous high temperature volatilization.

[0071] After digestion, the system automatically cools to below 60°C. Open the digestion vessel and transfer the digest to a 100 mL PFA volumetric flask. Use 2%... Rinse the inside of the digestion vessel three times, combine the washings in a volumetric flask, and finally dilute with 2%... Make up to the mark. If trace amounts of precipitate appear in the digestion solution (mainly undissolved silicates or fluorides), filter it with a 0.2 μm PTFE needle filter before making up to the mark.

[0072] Furthermore, to compensate for matrix effects and instrument drift in subsequent ICP-MS measurements, an internal standard mixture was introduced online into the final volume of the sample solution. The internal standard solution contained 10 μg / L of [a specific ingredient / component]. 103 Rh and 193 Ir, medium is 2% The internal standard solution and the sample solution are mixed online at a volume ratio of 1:1 through a three-way valve and then directly enter the nebulizer of the mass spectrometry detection module.

[0073] The final output sample solution has the following characteristics: nitric acid concentration of approximately 2% to 4%, and total dissolved solids (TDS) of less than 0.2% (mass fraction). The residual concentration was reduced to a negligible level (<50 mg / L), thus providing a low-interference, highly stable injection medium for KEDICP-MS. This solution also ensures that elements such as Cr, As, Cd, Hg, Pb, Co, Ni, Cu, Zn, Ga, and Ge do not precipitate or change their valence state at room temperature for at least 72 hours.

[0074] Furthermore, in monitoring scenarios requiring higher sensitivity, a pre-concentration module can be added after the microwave digestion step: using a solid-phase extraction (SPE) column (such as Chelex-100 or iminodiacetic acid type resin) to selectively enrich trace heavy metals in seawater, and then using 1 mol / L... After elution, bring the volume to a suitable level. In this mode, the method detection limit can be reduced by 1 to 2 orders of magnitude, but an additional eluent blank correction procedure is required.

[0075] For quality control, a spiked recovery sample (5 μg / L) was inserted every 10 samples, and a standard reference material (such as nearshore sediment or seawater standard GBW(E)080040) was analyzed simultaneously to monitor digestion recovery and matrix interference. If the recovery rate of any element exceeds the range of 85%–115%, or the internal standard... 103 If the Rh signal strength drops by more than 30% from the initial calibration value, the system determines that the batch of data is invalid and automatically triggers the re-decryption procedure.

[0076] The mass spectrometry detection module receives homogeneous sample solutions from the sampling and preprocessing module and operates in kinetic energy discrimination (KED) collision mode to simultaneously and quantitatively detect 11 target metallic elements, including chromium, arsenic, cadmium, mercury, lead, cobalt, nickel, copper, zinc, gallium, and germanium. It outputs raw sequence data of each element's concentration over time for use by the subsequent time-frequency domain multi-index extraction module, as detailed below:

[0077] A quadrupole inductively coupled plasma mass spectrometer was selected, equipped with ultrapure argon as the plasma and collision gas. The mass spectrometer was installed in a temperature and humidity controlled laboratory (temperature 20℃±2℃, relative humidity 40%~60%). The main operating parameters were set as follows:

[0078] To eliminate interference from the high-salinity seawater matrix and polyatomic ions introduced by the digestion acid (mainly for) 40 Ar 16 O + right 56 Fe + , 40 Ar 35 Cl + right 75 As + , 35 Cl16 16O + right 51 V + or 52 Cr + The mass spectrometry detection module introduces high-purity helium into the hexapole collision cell and employs a kinetic energy discrimination mode. This mode utilizes the property that the collision cross-section of polyatomic ions is larger than that of monoatomic ions: when helium collides with the ion beam, polyatomic ions lose more kinetic energy and cannot overcome the energy barrier of the subsequent quadrupole, thus being filtered out.

[0079] The helium flow rate was determined through optimization experiments. A 10 μg / L iron standard solution was used as the test object, and measurements were taken at different helium flow rates. 54 Fe、 56 Fe、 57 The signal-to-background ratio (SNR) of Fe (the ratio of signal strength to background count). Experiments show that when the helium flow rate is 4.7 mL / min, 57 The signal-to-background ratio of Fe reaches its peak. At this point, the Co / ClO ratio built into the instrument is used as the tuning indicator parameter, with a value of 10.6 ± 0.2. The definition of Co / ClO is:

[0080]

[0081] The higher this ratio, the stronger the ability of the collision cell to suppress oxide interference. In subsequent daily analyses, the system should be automatically tuned to Co / ClO ≥ 10.5 after powering on each day before sample testing can begin.

[0082] For matrix elements that may coexist in seawater and desulfurization wastewater, one or more low-interference isotopes were selected for each target element, prioritizing mass-to-charge ratios with higher abundance and less significant influence from isotopes. The final selected analytical isotopes and their main sources of interference and elimination methods are detailed below. Figure 2 , Figure 2 All isotopes in the KED mode were scanned using a high-resolution (0.5 amu) method to further filter out the tailing interference of adjacent mass peaks;

[0083] To counteract matrix effects (especially incompletely digested residual organic matter and salts) and sensitivity drift caused by prolonged instrument operation, a dual internal standard online calibration strategy was employed. The internal standard solution contained 5 μg / L... 103 Rh and 5 μg / L 193 2% of Ir The medium is mixed online with the sample solution at a 1:1 volume ratio through a three-way valve and then enters the nebulizer.

[0084] Criteria for selecting internal standard elements: 103 Rh and 93Ir content in natural samples is extremely low, its mass number is close to that of the target element, and there is no spectral overlap, indicating similar ionization behavior. The correction principle is based on the positive correlation between the internal standard signal intensity and the suppression / enhancement effect caused by the sample matrix. For the i-th sample, the corrected concentration of the target element x is... Calculated by the following formula:

[0085]

[0086] In the formula:

[0087] The original concentration is calculated directly by the instrument based on the external standard curve;

[0088] The average signal intensity of the internal standard (or Rh or Ir) in the standard solution sequence;

[0089] The signal intensity of the corresponding internal standard in the i-th sample;

[0090] and These are the dilution factor correction terms for standard solutions and samples, respectively (usually set to 1, corrected when the dilution ratios are different).

[0091] In actual operation, if the internal standard signal intensity of any sample is lower than ±30% of the average value of the standard sequence, the system will automatically determine that the sample matrix is ​​overloaded and trigger the "retest after dilution" command.

[0092] Calibration curves were prepared at eight concentration points using a serially diluted mixed standard stock solution, covering a range of 0.1 μg / L to 1000 μg / L (with the highest Hg concentration at 2.0 μg / L). Each concentration point was measured three times, and the average signal value was used. The standard curves were fitted using linear regression, and the equation is as follows:

[0093]

[0094] In the formula The net signal strength (cps) after subtracting the blank. Here, 'a' represents concentration, 'a' represents sensitivity coefficient, and 'b' represents intercept. The linear correlation coefficient must be ≥0.999.

[0095] The limit of detection (LOD) is defined as: the detection limit of a sample blank (2%). ) Perform 11 consecutive measurements and calculate the standard deviation of the signal. The detection limit is:

[0096]

[0097] The typical detection limits for each element in KED mode are as follows: Cr 0.02 μg / L, Co 0.005 μg / L, Ni 0.03 μg / L, Cu 0.01 μg / L, Zn 0.08 μg / L, Ga 0.01 μg / L, Ge 0.02 μg / L, As 0.04 μg / L, Cd 0.005 μg / L, Hg 0.002 μg / L, Pb 0.01 μg / L (all concentrations are based on the original water sample).

[0098] The mass spectrometry detection module operates in fully quantitative scanning mode, with each sample analysis taking 90 seconds (including 60 seconds of stable injection and 30 seconds of data acquisition). Integration time: 0.1 seconds / mass point for elements with a concentration >1 μg / L, and 0.5 seconds / mass point for trace elements (such as Cd and Hg) to reduce counting and statistical errors;

[0099] After each sample measurement, use 2% sequentially. Ultrapure water and 5% Clean the injection system for 60 seconds each time to reduce memory effect (especially for Hg and As). After every 10 samples, insert a standard solution (intermediate concentration point) for retesting. If the relative error exceeds ±10%, recalibrate.

[0100] Finally, the mass spectrometry detection module outputs a time-element concentration matrix, which is transmitted to the time-frequency domain multi-index extraction module in CSV format. This matrix contains the sampling timestamp, sample identification code, and concentration values ​​(unit: μg / L) of 11 heavy metals. Simultaneously, the internal standard signal intensity, tuning report, and calibration curve parameters for each measurement are encapsulated as metadata for subsequent data quality traceability.

[0101] The time-frequency domain multi-index extraction module is located after the mass spectrometry detection subsystem. Its input is the time-elemental concentration matrix output by this subsystem (i.e., the original sequence of heavy metal concentration changes over time), and its output is a multi-dimensional index matrix that can be processed by the graph model index optimization module. The core task of this module is to map the one-dimensional concentration time series into a high-dimensional feature space containing both time-domain statistical features and frequency-domain energy distribution features, thereby providing a rich information foundation for subsequent graph modeling and anomaly detection. The workflow of the time-frequency domain multi-index extraction module is as follows:

[0102] 1) Input data preprocessing: Before sliding window processing, necessary preprocessing is performed on the original concentration sequence. Let the original concentration time series of the m-th heavy metal element (m=1,2,…,M, where M=11, corresponding to Cr, Co, Ni, Cu, Zn, Ga, Ge, As, Cd, Hg, Pb) be:

[0103]

[0104] Where N is the total number of sampling points (i.e., the number of concentration data points output by the mass spectrometry detection subsystem; each sampling point corresponds to one water sample analysis result, and the sampling interval is typically 1–4 hours). To eliminate the influence of potential large-scale trend terms (such as equipment aging or seasonal emission fluctuations) on subsequent time-frequency domain indicators, the sequence is subjected to first-order difference detrending processing:

[0105]

[0106] If the test shows no obvious trend in the sequence, then the original sequence is used directly. The detrended sequence is denoted as... (The superscript m is omitted for brevity);

[0107] 2) Sliding Window Segmentation and Parameter Settings: To approximate the non-stationary concentration change sequence as a segmented stationary signal, a fixed-length sliding window is used to segment x into frames. Let the window length (i.e., the number of consecutive sampling points in each segment) be L, and the window overlap length be R. Then, the sample index range covered by the k-th window is:

[0108]

[0109] Wherein, the window starting index is The ending index is The choice of window length L needs to balance temporal resolution and statistical robustness: if L is too small, the indicators are easily affected by short-term noise; if L is too large, early degradation information will be smoothed out. In this system, based on the sampling interval and the typical degradation cycle in industrial settings, L = 5000 data points are set (corresponding to approximately 5000 water sample analyses; if the sampling interval is 2 hours, it covers approximately 416 days). The overlap length is set to R = L / 2 = 2500 to enhance feature continuity. For different application scenarios, L can also be configured by the operator between 2000 and 10000, and adjusted in conjunction with the number of nodes j in the graph model optimization module.

[0110] 3) Time-domain index calculation, for the concentration subsequence within each window. (Length L), calculate the following 6 time-domain statistical indicators;

[0111] Mean:

[0112]

[0113] The mean reflects the average concentration level within the window and is sensitive to slow drift;

[0114] variance:

[0115]

[0116] Variance (sample variance) characterizes the amplitude of concentration fluctuations. An increase in variance often indicates unstable emissions or equipment malfunctions.

[0117] Root mean square (RMS):

[0118]

[0119] The root mean square (RMS) is sensitive to changes in amplitude and is a commonly used indicator for measuring energy levels. In mechanical condition monitoring, it has been proven to have excellent early warning capabilities for slow degradation.

[0120] Skewness:

[0121]

[0122] Skewness measures the asymmetry of the distribution. A positive value indicates a heavier right tail (high concentration anomalies), while a negative value indicates a heavier left tail. In desulfurization discharge outlets, abrupt changes in skewness often correspond to intermittent high-concentration emission events.

[0123] kurtosis:

[0124]

[0125] Here, excess kurtosis is used to make the kurtosis of the normal distribution zero. Kurtosis is extremely sensitive to impulsive pulse anomalies and can capture sudden concentration spikes caused by pipe erosion or the shedding of corrosion debris;

[0126] Maximum absolute value:

[0127]

[0128] The maximum absolute value directly reflects the extreme situation of the concentration within the window, which is of intuitive significance for detecting excessive emissions.

[0129] The above six time-domain indicators are calculated for the k-th window respectively, resulting in the time-domain indicator vector for that window. ;

[0130] 4) Frequency domain index calculation: To extract periodic fluctuation components from the concentration sequence (such as frequency disturbances from the desulfurizer dosing pump or the effect of seawater tides on the dilution of the cooling outlet), a Discrete Fourier Transform (DFT) is performed on the subsequence within each window. Let the signal length within the window be L, then its DFT is defined as:

[0131]

[0132] Where q is the frequency index, corresponding to the physical frequency. Frequency resolution , This is the equivalent sampling rate (i.e., the sampling frequency of concentration data, unit: 1 / hour). Only the positive frequency component is considered. ), calculate the power spectral density:

[0133]

[0134] (Factor 2 is used to compensate for the energy of a single-sided spectrum).

[0135] The root mean square frequency (RMSF) is extracted from the power spectrum as the sole frequency domain index, and its definition is as follows:

[0136]

[0137] The root mean square frequency (RMSF) characterizes the second moment of the power spectrum on the frequency axis, reflecting the dominant frequency band of the signal. When periodic fluctuations occur in the concentration sequence (such as regular backwashing or dosing), the RMSF will converge to a narrow band; when the sequence tends towards white noise or aperiodic perturbations, the RMSF broadens. Compared to simple peak frequencies, the RMSF is more robust to noise and spectral leakage.

[0138] For each window k, a scalar frequency domain index is calculated. After adding it to the time-domain index vector, the complete feature vector of the window is obtained. The dimension is 7;

[0139] Finally, a multidimensional index matrix is ​​output. For a specific heavy metal element m, all K sliding windows are processed sequentially to obtain the feature matrix of that element.

[0140]

[0141] The feature matrices of all M heavy metal elements are stacked along the channel dimension to form the output of the time-frequency domain multi-index extraction module—the multidimensional index tensor:

[0142]

[0143] This tensor has dimensions M×K×7. In practical system implementations, it is typically stored element-wise as several two-dimensional matrices and passed to the graph model index optimization module. The p-th column (p=1,2,…,7) in the table represents the time series of the p-th indicator as a function of the window index k, denoted as... These sequences are the input signals for the graphical model optimization module;

[0144] Furthermore, near the beginning and end of the sequence, the sliding window may not be able to completely cover L points. For the initial segment (k=1), if but Then, the sequence is symmetrically extended to point L using reflection padding; for the last segment, if the number of remaining points is less than L, the incomplete window is discarded (or a smaller window is used, but the number of nodes j in the graphical model needs to be adjusted to maintain consistency). Furthermore, before calculating the index, the original concentration data within each window is processed. Outlier replacement: If a certain point Then replace its value with (i.e., the mean within the window) to eliminate the excessive influence of occasional spikes in the sensor on indicators such as kurtosis;

[0145] The time-frequency domain multi-index extraction module acquires concentration matrices in batches from the mass spectrometry detection subsystem via the data bus. After completing the full segment calculation for each heavy metal element, it immediately... The data is sent to the graph model index optimization module. Simultaneously, this module records the following metadata into the system database: window length L, overlap length R, detrending method, and start and end timestamps for each window. This metadata is used for subsequent traceability analysis of the results. The output multidimensional index tensor H retains the dynamic characteristics of the original concentration data while significantly reducing the data dimensionality (compressed from N sampling points to M×K×7 features), laying the foundation for efficient computation of the graph model.

[0146] The graphical model index optimization module takes as input the time-varying sequence of individual indices for each heavy metal element (e.g., the RMS value of Cr alone over time) and outputs a similarity score sequence optimized by the graphical model. It uses the graphical model to characterize the topological structure of non-stationary time series and effectively suppresses noise disturbances and short-term random fluctuations in the original time-frequency domain indices through dynamic median plot matching, thereby enhancing the detectability of early degradation features. The specific process is as follows:

[0147] Let the time-frequency domain index (p=1,2,…,M, M=11) of the m-th heavy metal element (m=1,2,…,7, corresponding to mean, variance, RMS, skewness, kurtosis, maximum absolute value, and root mean square frequency) have a time series length of K (i.e., the number of windows output by the time-frequency domain extraction module). This index sequence is denoted as:

[0148]

[0149] The graph model index optimization module processes each (m, p) group independently, generating a corresponding similarity score sequence. (The specific length depends on the sliding window segmentation method). For brevity, the superscript (m, p) will be omitted below, and a single index sequence will be denoted as... ;

[0150] To extract the local topology of the index sequence, a fixed-length sliding window is used to divide h into frames. Each window contains j consecutive data points (i.e., the number of nodes in the graph model), and the window step size (the interval between the starting points of two adjacent windows) is d. In this system, j=10 and d=1 (i.e., sliding one data point at a time, with overlapping window heights), thus obtaining K-j+1 data segments. The t-th segment is:

[0151]

[0152] The selection of the number of nodes j=10 is based on the following considerations: (1) Too small a j (such as 3~5) will make the graph model too sparse and unable to reflect the long-range correlation of the sequence; (2) Too large a j (such as more than 20) will significantly increase the amount of computation, and the time span covered by each segment will be too long, which may smooth out local abnormal changes. Through experiments, j=10 achieves the best balance between maintaining the richness of graph structure information and computational efficiency;

[0153] For each data segment Construct an undirected complete graph ,in:

[0154] Node set ,node Corresponding to The i-th data point ;

[0155] Edge set An edge is an edge that connects any two distinct nodes.

[0156] The weight of an edge is defined as the Euclidean distance (or absolute difference in the one-dimensional case) between the corresponding data values ​​of the two connected nodes:

[0157]

[0158] The weight matrix described above is a j×j symmetric matrix with diagonal elements of 0. (Graphical model) Completely determined by its weight matrix (adjacency matrix) express:

[0159]

[0160] Through the above construction, the original one-dimensional index sequence h is transformed into a graphical model sequence. ;

[0161] To quantify the extent to which the current graph model deviates from the "normal" state, a reference graph model representing historical normal behavior needs to be defined. This module uses a median graph as a dynamic reference, defined as: among a given set of graph models, the graph model that minimizes the sum of its distances to all other graph models.

[0162] For the t-th graph model, we only use its nearest predecessor. The median map is calculated for each graph model (excluding the current graph). In this system, the median map is taken as follows: Even using To calculate the median plot. For In the initial stage, the median map is computed using all existing graphical models. The definition of is:

[0163]

[0164] in The distance metric between the two graph models is determined using edge weight-based distance (Distance of Edge Weight Value, DEWV), which will be detailed below. Since the median graph must be a graph in the input graph set (not an arbitrary one), the above minimization can be achieved through exhaustive comparison. The solution involves multiple candidate graphs with extremely low computational cost.

[0165] Graph model similarity calculation (DEWV distance), the DEWV distance is defined as the sum of the absolute values ​​of the differences in the weights of corresponding edges between two graph models. Let the adjacency matrices of the two graph models be respectively... and ,but:

[0166]

[0167] in,

[0168]

[0169] Since the graph is undirected and its diagonal is zero, the above formula actually calculates... The sum of the absolute values ​​of the weight differences of the edges. This distance metric satisfies nonnegativity, symmetry, and the triangle inequality, and can be used as an effective similarity index.

[0170] Current graph model With median plot The DEWV distance between them is the similarity score at that moment:

[0171]

[0172] The larger the score, the more serious the deviation of the current graph model's structure from the historical normal pattern, that is, the significant change in the local topology of the indicator sequence, corresponding to possible abnormal or degenerate events;

[0173] For each index sequence By performing the above steps, a similarity score sequence is obtained:

[0174]

[0175] Organize the similarity scores of all M elements and P=7 indicators according to the original dimensions to form the optimized similarity tensor:

[0176]

[0177] The tensor has dimensions M×P×(K-j+1). In actual system implementation, it is usually stored as multiple one-dimensional sequences in element-index pairs and then passed to the multi-index fusion module.

[0178] Compared with the original time-frequency domain index sequence, the optimized similarity score sequence has the following advantages: (1) The graph modeling process has low-pass filtering characteristics through global comparison of distances between nodes, which can suppress high-frequency random noise; (2) The DEWV distance is not sensitive to changes in the local gain of the signal (because the edge weight reflects the relative difference rather than the absolute amplitude), thus resisting baseline drift; (3) The dynamic median map update mechanism enables the reference model to adapt to slow environmental changes (such as seasonal water temperature fluctuations), avoiding misjudging normal drift as abnormal.

[0179] In the sequence initiation stage ( The candidate set of the median plot is insufficient. At this point, all existing graph models are used to calculate the median graph, and the similarity scores of these initial stages are labeled as "warm-up data," which are not used in subsequent anomaly decisions. Although a window sliding step size of d=1 generates a large number of overlapping graph models, the adjacency matrix of each graph model is only j×j=100, and the DEWV distance calculation involves... Subtraction and absolute value operations can be performed in real time by modern embedded processors or industrial control computers. To further reduce computational load, the step size d can be set to 2 or 3, but this will correspondingly reduce the time resolution.

[0180] The multi-indicator fusion module takes as input a set of similarity score sequences corresponding to all heavy metal elements and their time-frequency domain indicators, and outputs a single comprehensive monitoring indicator time series. It effectively aggregates optimized information from different elements and feature dimensions to form a dimensionless scalar sequence that comprehensively reflects the combined state of the seawater direct cooling system and desulfurization outlet, thereby reducing the complexity of downstream anomaly decision-making and improving detection robustness. Specifically:

[0181] Suppose the graph model index optimization module processes M heavy metal elements (M=11 in this system, including Cr, Co, Ni, Cu, Zn, Ga, Ge, As, Cd, Hg, Pb) and P time-frequency domain indices (P=7, i.e., mean, variance, root mean square, skewness, kurtosis, maximum absolute value, and root mean square frequency), then there are a total of M×P independent similarity score sequences. The length of each sequence is T=K-j+1, where K is the number of windows output by the time-frequency domain extraction module, and j is the number of graph model nodes (j=10). All these sequences are aligned in time (the same sampling time corresponds to the same window index t) to form the following multidimensional array:

[0182] For each time t (t=1,2,…,T), there exists an M×P matrix. Its elements are:

[0183]

[0184] in Let be the similarity score of the m-th element and the p-th index at time t;

[0185] Because the concentration levels of different heavy metals and the numerical ranges of different indicators (such as kurtosis and root mean square frequency) vary significantly, directly performing an arithmetic average would lead to a "large number eats up small number" phenomenon, meaning that indicators with larger magnitudes dominate the fusion result, while indicators with smaller magnitudes but potentially equally important values ​​(such as the similarity scores of trace elements like Ge and Ga) are submerged. Therefore, we first perform max-min normalization independently on each similarity score sequence, mapping its value range to the [0,1] interval;

[0186] For the m-th element and the p-th index, its normalized value The calculation is as follows:

[0187]

[0188] In the formula, the denominator is the range of the index series over the entire time axis. If the denominator is zero (i.e., the index series is always constant), then it is defined as follows: After normalization, all indicator sequences have the same value range, eliminating the influence of dimensions;

[0189] It is worth noting that maximum-minimum normalization depends on the global maximum and minimum values ​​of the entire time series. In online monitoring systems, extreme values ​​can be calculated based on historical normal data in the initial stage and dynamically updated as the running time increases (e.g., recalculated every week of new data), or an incremental extreme value update strategy can be adopted.

[0190] After normalization, at the same time t, all M×P normalized similarity scores are aggregated. The point-by-point mean at this time is defined as:

[0191]

[0192] The point-by-point mean reflects the overall average level of all elements and all indicators, and is more sensitive to uniform and slow overall changes;

[0193] The pointwise maximum value at this moment is defined as:

[0194]

[0195] The maximum value at each point can capture sudden and significant anomalies in any element or indicator. Even if other indicators are within the normal range, as long as one channel deviates strongly, the maximum value will rise, thus ensuring a rapid response to local faults.

[0196] When merging the point-by-point mean and point-by-point maximum into a single comprehensive index, it is necessary to determine the weighting coefficient between the two. Fixed weights cannot adapt to changes in the relative importance of the mean and maximum under different operating stages and conditions. Therefore, this module uses the standard deviation method to adaptively calculate the weighting factor α.

[0197] The basic idea of ​​the standard deviation method is: if the degree of variation of a variable (measured by standard deviation) is large, it means that the variable has a strong ability to distinguish different times and should be given a higher weight; conversely, if the variable is almost unchanged, its contribution to anomaly detection is small and should be given a lower weight.

[0198] First, calculate the point-by-point mean sequence. Standard deviation and pointwise maximum sequence Standard deviation :

[0199]

[0200]

[0201] The weighting factor α is defined as follows:

[0202]

[0203] Clearly, 0 ≤ α ≤ 1. When When α is relatively large, it approaches 1, and the comprehensive index focuses more on the point-by-point mean; conversely, when α is relatively small... When the value is large, α approaches 0, and the comprehensive index focuses more on the maximum value point by point;

[0204] Ultimately, the comprehensive monitoring indicators at each time point t Calculate using the following formula:

[0205]

[0206] Here, α is the weighting factor. It is important to emphasize that α is a global value calculated once based on the entire historical data sequence (or all currently available data), and it remains constant throughout the monitoring process.

[0207] If the system has been running for a sufficiently long time, the addition of new data will cause... or Significant changes (relative changes exceeding 20%) can trigger a recalculation of α and an update of the composite index series. However, it is generally recommended to keep α constant throughout a continuous monitoring task to ensure consistency of the indicators.

[0208] Comprehensive monitoring indicators It is a dimensionless positive real number with a theoretical value range of [0,1]. This index integrates graphical model optimization information of all heavy metal elements and all time-frequency domain features, which not only retains the sensitivity of the mean to the overall state drift, but also takes into account the timely response of the maximum value to sudden anomalies, and achieves the optimal trade-off between the two through adaptive weights;

[0209] The multi-indicator fusion module outputs a one-dimensional time series:

[0210]

[0211] Each sampling time t corresponds to a comprehensive index value. This sequence is sent to the anomaly decision-making and early warning module in the form of structured data (such as CSV or time series database). At the same time, this module synchronously records the following auxiliary information: the extreme values ​​of each index used for normalization, the calculated α value, and the historical statistics (mean, standard deviation) of the pointwise mean and pointwise maximum value sequence, in order to facilitate subsequent model calibration and result reproduction;

[0212] The anomaly decision-making and early warning module takes as input the time series of comprehensive monitoring indicators output by the module. The output is a warning signal reflecting the combined status of the seawater direct cooling system and the desulfurization outlet (including multiple levels of indication such as normal, attention, alarm, and emergency). Statistical hypothesis testing is performed on the comprehensive monitoring indicators. Under a preset false alarm rate, abnormal points exceeding the normal fluctuation range are automatically identified. False triggers caused by random noise are filtered out through continuous judgment rules. Finally, reliable alarm information is sent to operators or the upper-level control system, as detailed below:

[0213] In the initial stage of online monitoring, the system lacks prior knowledge of the "normal state." Therefore, a period of historical data is needed as a reference for establishing a statistical baseline. There are two methods for selecting the historical normal period:

[0214] Fixed initial segment method: Take the first segment of the comprehensive monitoring index sequence. Points (taken in this system) This period (corresponding to approximately 100 sampling times) serves as the initial normal segment. During this segment, the system is assumed to be in a healthy operating state, with all indicators within the baseline range.

[0215] Manual calibration method: If it is known in advance that the system will operate normally for a certain period of time (for example, a week after the equipment has just been repaired or cleaned), the operator can specify the start and end times on the system interface and calibrate the data within that period as the historical normal period;

[0216] During continuous system operation, historical normal periods can be dynamically updated using a sliding reference window strategy: only the most recent data is used. Statistical parameters are estimated using a subset of points deemed normal (rather than all historical data) to prevent slow drift from causing the baseline to gradually deviate from the true normal range. The sliding baseline window is updated as follows: whenever 10 consecutive points do not trigger an anomaly warning, these 10 points are included in the baseline window, and the first 10 points to enter the window are removed, maintaining the same number of normal points, while preserving the window length. constant;

[0217] Let the set of comprehensive monitoring index values ​​within the historical normal period (fixed initial period or sliding reference window) be denoted as . Calculate its sample mean μ and sample standard deviation σ:

[0218]

[0219]

[0220] It should be noted that the comprehensive monitoring indicators After normalization and fusion, its theoretical distribution can be approximated as a Gaussian distribution (the central limit theorem guarantees that the fused values ​​of multiple indicators are asymptotically normal). In actual verification, normality tests (such as the Shapiro-Wilk test) are performed on historical normal data. If there is a significant deviation from normality, the median and interquartile range (IQR) can be used instead of the mean and standard deviation. However, this system defaults to processing the data using a Gaussian model.

[0221] Based on the Gaussian distribution assumption, the confidence interval (or control limit) for the normal state is defined as follows:

[0222]

[0223] Where k is the confidence factor. According to the properties of the Gaussian distribution, when k=3, the probability of a single point falling outside the interval is approximately 0.27%; when k=6, this probability drops to approximately... This can greatly reduce false alarms caused by random fluctuations. Considering both the tolerance for false alarms and the stringent requirements for missed alarms in industrial settings, this system uses k=6. In practical applications, k can also be configured by the operator within the range of 3 to 6, depending on the required monitoring sensitivity.

[0224] For the current moment Comprehensive monitoring indicators Construct the following hypothesis test:

[0225]

[0226] like If true, the system determines that the current state is normal; if... Points exceeding the confidence interval (i.e., greater than μ+kσ or less than μ-kσ) are initially identified as candidate points for anomalies. It should be noted that, due to the comprehensive monitoring indicators... After normalization, with 0 as the lower bound and 1 as the upper bound, cases below μ-kσ are extremely rare (unless k is very large and μ is close to 0). Therefore, we usually only need to focus on the upper bound (i.e., ...). This indicates an increased deviation of the system's state from normal;

[0227] Single-point exceedances of the confidence interval may be caused by transient sensor interference, occasional fluctuations in water quality, or purely random noise. Directly triggering an alarm would result in an unacceptable false alarm rate. Therefore, this module introduces a continuous exceedance counter. The principle of "continuous exceedances are required to confirm an anomaly" is adopted.

[0228] The specific rules are as follows:

[0229] initialization ;

[0230] At each new moment ,like ,but ;otherwise ;

[0231] when Time (This plan takes) This is identified as an anomaly and an alert is triggered.

[0232] Once an anomaly is confirmed, the counter retains its current value and will not trigger new alarms again during the warning period, until the system state recovers (i.e., continuous...). (All points fall within the confidence interval)

[0233] consecutive times The selection is based on the following analysis: With k=6, the probability of a single false alarm is extremely low, and theoretically, the probability of three consecutive false alarms is negligible; however, if k is changed to 3, the probability of three consecutive false alarms is approximately... This is still within an acceptable range. This parameter is also configurable.

[0234] To provide more refined decision support, this module, after confirming anomalies, further classifies warning levels based on the degree to which the anomaly exceeds the confidence interval. Let the excess amount be:

[0235]

[0236] Define three levels of early warning:

[0237] Level 1 Warning (Attention): (The continuous confirmation threshold has not yet been reached; operators are only being prompted to pay attention.)

[0238] Level 2 Warning (Alarm): (An anomaly has been confirmed; manual review or increased monitoring frequency is recommended.)

[0239] Level 3 Warning (Emergency): (i.e., the magnitude exceeding the upper bound of the confidence interval is more than one standard deviation), or (Approaching the normalization upper limit), an emergency shutdown suggestion is triggered and automatic sampling and evidence retention are activated;

[0240] Warning signals are output in the following ways:

[0241] Digital output: Sends alarm codes (0=normal, 1=caution, 2=alarm, 3=emergency) to the Distributed Control System (DCS) or Programmable Logic Controller (PLC) via RS485, Modbus TCP, or OPCUA protocol.

[0242] Human-machine interface display: The comprehensive monitoring index curve is displayed on the monitoring screen in different colors (green, yellow, orange, red), and the warning level is automatically marked at abnormal points;

[0243] Event Log: The start and end times, peak values, duration of exceeding limits, and heavy metal elements involved in each warning are stored in the historical database for accident recall and analysis.

[0244] Furthermore, after the system issues a confirmed anomaly warning, it is necessary to wait for the comprehensive monitoring indicators to fall back to the normal confidence range and remain there continuously. Three points (this system uses three points) are required to determine "abnormal recovery". After recovery, the counter... The alarm signal is removed once the counter is reset to zero. If intermittent exceedances occur during the recovery process, the recovery counter will also be reset accordingly, ensuring that the alarm is only lifted when the recovery is continuous and stable.

[0245] For the sliding baseline window strategy, during anomaly alert periods, outliers are not included in the updates of historical normal periods to avoid polluting the baseline statistics. Sliding updates can only resume after the system returns to normal.

[0246] Furthermore, considering the different tolerances for false alarms in different monitoring scenarios (such as nearshore emission compliance areas vs. areas surrounding nature reserves), the aforementioned key parameters ( These parameters should be designed as remotely configurable variables. Operators can modify these parameters via host computer software without restarting the system. Furthermore, after the system has been running for a certain period (e.g., 3 months), accumulated historical normal data can be used to re-estimate μ and σ and update the confidence intervals to eliminate baseline drift caused by equipment aging or seasonal changes, achieving long-term adaptability.

[0247] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A combined monitoring system for multiple heavy metal indicators at seawater direct current cooling and desulfurization discharge outlets, characterized in that, include: The sampling and pretreatment module collects water samples from the seawater DC cooling outlet and desulfurization outlet and performs digestion pretreatment to output a homogeneous, low-matrix-interference sample solution. The mass spectrometry detection module, whose input is connected to the output of the sampling and preprocessing module, performs simultaneous quantitative detection of multiple target heavy metal elements in the sample solution under kinetic energy discrimination collision mode, and outputs the original sequence data of the concentration of each element changing over time. The time-frequency domain multi-index extraction module has its input end connected to the output end of the mass spectrometry detection module. It maps the original sequence data of each heavy metal element into a multi-dimensional index tensor containing time-domain statistical features and frequency-domain energy distribution features. The multi-dimensional index tensor is organized according to heavy metal element, sliding window and index type. The graph model index optimization module has its input end connected to the output end of the time-frequency domain multi-index extraction module. It independently constructs an undirected complete graph for each index sequence in the multi-dimensional index tensor, calculates the graph model sequence, and obtains the similarity score between each graph model and the historical normal graph model based on edge weight distance using dynamic median graph matching, and outputs the similarity score tensor. The multi-indicator fusion module, whose input is connected to the output of the graph model index optimization module, normalizes the similarity score sequences of all heavy metal elements and all time-frequency domain indicators, calculates the point-by-point mean and point-by-point maximum, and uses the standard deviation method to adaptively determine the weight factor to fuse the point-by-point mean and point-by-point maximum, and outputs a single comprehensive monitoring index time series. The abnormal decision-making and early warning module has its input end connected to the output end of the multi-indicator fusion module. It constructs a confidence interval based on the statistical parameters of the comprehensive monitoring indicator time series in the preset historical normal segment, performs hypothesis testing on the indicator value at the current moment, and outputs multi-level early warning signals after confirming the abnormality by continuously exceeding the counter.

2. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 1, characterized in that, The sampling and preprocessing module includes: The sampling unit is used to install automatic synchronous samplers triggered by the same timing controller at the outlet of the main drainage channel of the seawater direct cooling system and the wastewater clarification tank of the desulfurization system. The inner wall of the sampling pipeline is lined with polytetrafluoroethylene or perfluoroalkoxy resin. An online filtration unit, whose input end is connected to the output end of the sampling unit, is used to remove suspended particulate matter with a particle size greater than 0.45μm and is equipped with an automatic backflushing cleaning function; The microwave digestion unit has its input end connected to the output end of the online filtering unit, and uses... The ternary digestion system, without the addition of any chlorine-containing reagents, digests the filtered water sample through a stepped temperature increase program. After the digestion solution is brought to a final volume, the test sample solution is output.

3. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 1, characterized in that, The mass spectrometry detection module is specifically used for: High-purity helium was introduced into a hexapole collision reactor, and the reactor was operated in kinetic energy discrimination mode. The helium flow rate was preset to ensure... 57 The signal-to-back ratio of Fe reaches its peak flow rate value; For each target heavy metal element, one or more low-interference isotopes are selected and high-resolution scanning is used. A dual internal standard online calibration strategy was adopted, with the internal standard solution containing 5 μg / L of [unclear - likely a specific ingredient or solution]. 103 Rh and 5 μg / L 193 Ir is mixed online with the sample solution at a 1:1 volume ratio through a three-way valve and then enters the nebulizer; Output the time-element concentration matrix to the time-frequency domain multi-index extraction module. The time-element concentration matrix includes the sampling timestamp, sample identification code, and concentration values ​​of each heavy metal element.

4. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 1, characterized in that, The time-frequency domain multi-index extraction module is specifically used for: First-order difference detrending preprocessing was performed on the original concentration time series of each heavy metal element; A fixed-length sliding window is used to frame the detrended sequence. The window length L = 5000 data points and the window overlap length R = L / 2. For each concentration subsequence within a window, time-domain statistical indicators are calculated, including mean, variance, root mean square, skewness, kurtosis, and maximum absolute value. The power spectral density is obtained by performing a discrete Fourier transform on the concentration subsequence within each window, and the root mean square frequency is extracted as a frequency domain index. The time-domain statistical indicators and frequency-domain indicators of each window are combined into a feature vector, and the multi-dimensional indicator tensors are output to the graph model indicator optimization module in the order of heavy metal elements and windows.

5. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 1, characterized in that, The graph model index optimization module is specifically used for: For each input single index sequence, a fixed-length sliding window is used to divide the data into frames. Each window contains j=10 consecutive data points, and the window sliding step size d=1, resulting in K-j+1 data segments. For each data segment, construct an undirected complete graph, where nodes correspond to the values ​​of data points, and the weight of an edge is defined as the absolute difference between the corresponding data values ​​of the two connected nodes. The most recent graph model before using the current graph model A median graph is calculated for each graph model, wherein the median graph is the graph model that minimizes the sum of edge-weighted distances to all other graph models; The edge-weighted distance between the current graph model and the median graph is calculated as a similarity score, whereby the edge-weighted distance is defined as the sum of the absolute values ​​of the differences in the weights of corresponding edges between the two graph models. The similarity scores of all heavy metal elements and all time-frequency domain indicators are organized into similarity score tensors and output to the multi-indicator fusion module.

6. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 1, characterized in that, The multi-indicator fusion module is specifically used for: For each similarity score sequence, perform min-max normalization independently to map its value range to the [0,1] interval; At the same time, calculate the pointwise mean and pointwise maximum of all normalized similarity scores; Calculate the standard deviation of a point-by-point mean series and the standard deviation of the pointwise maximum sequence and through Define adaptive weighting factors ; The point-by-point mean and the point-by-point maximum value are... By integrating the data, a comprehensive monitoring index can be obtained. The time series of comprehensive monitoring indicators is output to the anomaly decision-making and early warning module.

7. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 1, characterized in that, The anomaly decision-making and early warning module is specifically used for: Take the first part of the comprehensive monitoring indicator sequence Using a point or a manually calibrated time period as the initial historical normal segment, calculate its sample mean μ and sample standard deviation σ; Based on the Gaussian distribution assumption, the confidence interval for the normal state is defined as follows: , where k is the confidence factor, and its value ranges from 3 to 6; Hypothesis testing is performed on the comprehensive monitoring indicators at the current moment: if the indicator value falls outside the confidence interval and exceeds it consecutively a certain number of times... If this happens, it is confirmed as an anomaly; The warning level is divided according to the degree of exceeding the confidence interval: when the excess amount Δ>0 and the number of consecutive excesses is 1 or 2, a level 1 warning is output; A level 2 warning will be issued when the number of consecutive exceedances is ≥3. A level 3 warning is issued when the excess Δ > 3σ or the index value > 0.95; The warning signal is sent to the outside world through digital output, human-machine interface display and event logging.

8. The combined monitoring system for multiple heavy metal indicators at seawater direct-flow cooling and desulfurization discharge outlets according to claim 7, characterized in that, The anomaly decision-making and early warning module is also configured with a sliding baseline window strategy: during continuous system operation, only the most recent baseline window is used. The sample mean μ and sample standard deviation σ are dynamically updated using a comprehensive monitoring indicator value that is judged to be normal, and abnormal points are not included in the update of the historical normal segment during the abnormal warning period.

Images