A cooling working substance online purification system for a two-phase immersion liquid cooling system

Abstract

The application discloses a cooling working medium online purification system for a two-phase immersion liquid cooling system, and relates to the technical field of heat dissipation management. The cooling working medium online purification system comprises a trend analysis module, a label identification module, a state evaluation module and a decision optimization module. The trend analysis module performs purity trend analysis on the cooling working medium based on start-up calibration time series data to generate a working medium purity trend graph. The label identification module analyzes the working medium purity trend graph to obtain filter differential pressure and working medium dew point, and determines a safety bypass state label by using an anomaly detection algorithm. The state evaluation module fuses the safety bypass state label and the working medium purity trend graph, and generates a health state evaluation report through multi-source information analysis. The decision optimization module analyzes the health state evaluation report to generate a system maintenance work order, and optimizes the anomaly detection algorithm to generate a maintenance decision report. The application converts the nonlinear coupling relationship of parameters such as pressure difference mutation and dew point drift into a visual spatiotemporal graph, and realizes visual monitoring of the purity attenuation process.

Description

Technical Field

[0001] This invention relates to the field of heat dissipation management technology, and in particular to an online purification system for the cooling medium in a two-phase immersion liquid cooling system. Background Technology

[0002] Two-phase immersion liquid cooling systems, as a core solution for heat dissipation in high-density computing equipment, have seen large-scale application in data centers and supercomputing centers in recent years. The principle of two-phase immersion liquid cooling systems relies on the phase change process of a low-boiling-point cooling medium to absorb heat, achieving efficient heat transfer. With increasing technological complexity, maintaining the purity of the cooling medium has become a key challenge. This is mainly achieved through periodic offline monitoring and maintenance, using gas chromatography to analyze the medium's composition or mechanical filters to adsorb impurities. In recent years, online monitoring technologies have gradually developed, with typical solutions including real-time impurity detection based on electrochemical sensors and differential pressure-driven bypass filtration technology. These solutions attempt to build dynamic maintenance mechanisms by continuously collecting medium state parameters.

[0003] However, the multi-dimensional nature of the evolution of working fluid purity requires more refined collaborative analysis capabilities, while existing methods still have two limitations in cross-modal data fusion and nonlinear dynamic modeling. Firstly, they lack the ability to quantify dynamic purity trends; traditional methods rely on discrete sampling points to assess the working fluid state, making it difficult to capture the spatiotemporal correlation of purity decay, resulting in insufficient early anomaly identification capabilities. Secondly, maintenance decisions are disconnected from system status; existing purification systems rely on fixed thresholds to trigger maintenance, ignoring the dynamic correlation between equipment health status and working fluid purity, leading to false alarm data failing to drive algorithm evolution and exacerbating resource waste. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online purification system for the cooling medium in a two-phase immersion liquid cooling system to solve the problem of dynamic purity multi-dimensional collaborative modeling and adaptive maintenance decision-making closed-loop optimization.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an online purification system for the cooling medium in a two-phase immersion liquid cooling system, comprising:

[0008] The preprocessing module collects and preprocesses equipment status signals to generate startup verification timing data. The trend analysis module, based on the startup verification timing data, performs purity trend analysis on the cooling medium to generate a medium purity trend graph. The tag identification module parses the medium purity trend graph to obtain filter differential pressure and medium dew point, and uses an anomaly detection algorithm to determine the safety bypass status tag. The status assessment module integrates the safety bypass status tag and the medium purity trend graph, and generates a health status assessment report through multi-source information analysis. The decision optimization module parses the health status assessment report to generate a system maintenance work order, and optimizes the anomaly detection algorithm to generate a maintenance decision report.

[0009] As a preferred embodiment of the online purification system for the cooling working fluid in a two-phase immersion liquid cooling system according to the present invention, the equipment status signals include physical operating parameter signals, working fluid purity monitoring signals, equipment health status signals, and consumable status signals.

[0010] The preprocessing includes filtering and denoising, normalization calibration, time alignment, and feature extraction.

[0011] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system according to the present invention, the startup verification timing data includes equipment operating parameters, working medium purity, equipment health status, and consumable status characteristic data.

[0012] The purity trend analysis of the cooling medium based on the startup verification timing data is performed in the following steps.

[0013] The startup verification timing data includes equipment operating parameters, working fluid purity, equipment health status, and consumable status characteristic data;

[0014] Based on the startup verification time series data, the multi-scale entropy algorithm is applied to calculate the working fluid purity entropy value and generate a multi-scale entropy value sequence.

[0015] The multi-scale entropy sequence is converted into a recursive graph, and a recursive quantized feature vector is formed by quantifying the stability and anomaly degree of the cooling fluid purity.

[0016] As a preferred embodiment of the online purification system for the cooling working fluid in a two-phase immersion liquid cooling system described in this invention, the generation of the working fluid purity trend map refers to constructing the correlation between the recursive quantification feature vector and the consumable state feature data, and performing spatiotemporal alignment to generate the working fluid purity trend map.

[0017] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system described in this invention, the step of obtaining the filter pressure difference and working medium dew point by analyzing the working medium purity trend graph refers to generating the filter pressure difference and working medium dew point by analyzing the equipment operating parameters through a depth-first traversal algorithm based on the topological structure of the working medium purity trend graph.

[0018] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system described in this invention, the specific steps for determining the safety bypass status tag using an anomaly detection algorithm are as follows:

[0019] The dynamic fluctuation characteristics of the filter pressure difference and the periodic characteristics of the working fluid dew point are extracted, and a fusion matrix is ​​generated through spatiotemporal alignment.

[0020] The fusion matrix is ​​transformed into a recursive graph, and the nonlinear coupling relationship of differential pressure and dew point is obtained by quantizing the recursive features.

[0021] The local density deviation is calculated based on the nonlinear coupling relationship of differential pressure and dew point to generate a density deviation anomaly score. The state level is divided by manifold space clustering method to generate a safety bypass state label.

[0022] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system described in this invention, the specific steps for integrating the safety bypass status tag and the working fluid purity trend graph are as follows:

[0023] The timestamps of the safety bypass status labels are dynamically aligned with the time axis of the working fluid purity trend graph to construct a causal reasoning graph.

[0024] Multimodal feature fusion of causal reasoning graph and working fluid purity trend graph is performed to generate time-series topological causal data.

[0025] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system described in this invention, the step of generating a health status assessment report through multi-source information analysis refers to extracting health features from time-series topological causal data, calculating a dynamic health index, and generating a health status assessment report in conjunction with a causal inference graph.

[0026] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system described in this invention, wherein: the step of generating a system maintenance work order by parsing the health status assessment report refers to extracting abnormal equipment events from the health assessment report, constructing a maintenance decision network, calculating the event correlation strength, and obtaining the system maintenance work order.

[0027] As a preferred embodiment of the online purification system for the cooling medium in a two-phase immersion liquid cooling system described in this invention, the specific steps for optimizing the anomaly detection algorithm to generate a maintenance decision report are as follows:

[0028] Execute system maintenance work orders and generate parameter correction vectors by comparing the real-time output of the anomaly detection algorithm;

[0029] By fusing parameter correction vectors and equipment anomaly events, maintenance decision reports are generated through feature state reconstruction.

[0030] The beneficial effects of this invention are as follows: by dynamically capturing the evolution law of working fluid purity, the nonlinear coupling relationship of parameters such as pressure difference change and dew point drift is transformed into a visual spatiotemporal map, realizing the visual monitoring of the purity decay process; at the same time, the algorithm parameters are calibrated in real time through work order execution feedback, and the fixed rules are replaced by dynamic thresholds. Combined with the feature reconstruction of historical false alarm data, the algorithm is driven to self-evolve throughout its life, eliminating the over-maintenance problem caused by static strategies. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of an online purification system for the cooling medium in a two-phase immersion liquid cooling system.

[0033] Figure 2 A flowchart for decision optimization.

[0034] Figure 3 This is a flowchart for trend analysis.

[0035] Figure 4 A flowchart for label recognition. Detailed Implementation

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0038] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0039] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an online purification system for the cooling working fluid in a two-phase immersion liquid cooling system, comprising the following steps:

[0040] The preprocessing module collects and preprocesses device status signals to generate startup verification timing data.

[0041] Equipment status signals include physical operating parameter signals, working fluid purity monitoring signals, equipment health status signals, and consumable status signals;

[0042] It should be noted that the physical operating parameter signals are acquired in real time by various sensors installed on the liquid cooling system. These include temperature sensors monitoring the contact temperature between the working fluid and the equipment, pressure sensors monitoring the pressure in the pipelines and chambers, flow meters monitoring the circulation rate of the working fluid, and pump speed and power sensors. These physical quantities are converted into continuous electrical signal outputs, forming a basic operating status data stream.

[0043] The purity monitoring signal of the working fluid is acquired by an online component analyzer directly immersed in the cooled working fluid. A laser spectroscopy probe is used to detect the absorption spectrum of impurity gases in the working fluid in real time, or an electrochemical sensor is used to monitor changes in dielectric constant and ion concentration. The chemical properties of the working fluid components are converted into quantifiable electrical signals through optical or electrochemical principles, enabling continuous tracking of purity indicators.

[0044] Equipment health status signals originate from diagnostic monitoring of the operating characteristics of core components. Vibration sensors collect the mechanical vibration spectrum of pumps and compressors, acoustic sensors capture abnormal noise characteristics in two-phase flow, or high-frequency current clamps monitor the harmonic distortion rate of drive motors. After spectrum analysis and feature extraction, the mechanical wear and electrical aging status of the equipment can be indirectly reflected.

[0045] Consumable status signals primarily assess the performance of filter and adsorbent components. Differential pressure sensors monitor pressure drop changes at the filter inlet and outlet to determine the degree of clogging. Capacitive level sensors track the expansion and contraction of the adsorbent material, or RFID tags integrated into the filter cartridge read cumulative operating time and temperature exposure history. These collectively constitute the dynamic assessment basis for consumable performance degradation.

[0046] Preprocessing includes filtering and denoising, normalization calibration, time alignment, and feature extraction.

[0047] It should be noted that the physical operating parameter signals (temperature, pressure, and flow rate, etc.) are processed by wavelet denoising to eliminate high-frequency interference signals and retain the true fluctuation characteristics of equipment operation; standardization is performed to unify the dimensions and compensate for the inherent errors of the sensors; time alignment is then used to achieve time synchronization of multi-source signals to ensure the continuity of data on a unified time axis; statistical features (such as mean and variance) and dynamic features (such as pressure instantaneous change rate markers) are extracted to form the equipment operating parameters.

[0048] The working fluid purity monitoring signal is filtered to smooth out noise interference and retain chemical characteristic peaks. The working fluid purity monitoring signal is then conditioned, amplified, and converted to generate the physical quantity of working fluid concentration. After being aligned with the timestamp of the equipment operating parameters, chemical characteristic indicators (such as the intensity of specific gas absorption peaks and the dew point drift velocity) are extracted to generate the working fluid purity.

[0049] The equipment health status signal is filtered by bandpass to remove mechanical resonance interference and focus on the effective diagnostic frequency band; the sound pressure level and current harmonic distortion rate are normalized to a uniform dimension range; the high-frequency signal is downsampled and synchronized with the liquid cooling system time axis; frequency domain features (such as the frequency energy distribution of rotating component fault characteristics) are extracted to form the equipment health status.

[0050] The consumable status signal is filtered by moving average to eliminate instantaneous fluctuation interference; the physical expansion height of the adsorbent is mapped to the saturation status parameter; the cumulative running time recorded by RFID is fused with real-time monitoring data; dynamic performance degradation indicators (such as pressure difference growth trend and adsorbent saturation rate) are extracted to quantify the consumable life status and generate consumable status characteristic data.

[0051] The equipment operating parameters, working fluid purity, equipment health status, and consumable status characteristic data are synchronized and aligned in time to generate startup verification timing data.

[0052] The trend analysis module, based on the startup verification timing data, performs purity trend analysis on the cooling working fluid and generates a working fluid purity trend chart.

[0053] The startup verification timing data includes equipment operating parameters, working fluid purity, equipment health status, and consumable status characteristic data;

[0054] Specifically, monitoring equipment operating parameters (such as filter differential pressure and working fluid dew point) to assess the physical conditions of the liquid cooling system provides a fundamental guarantee for the stability of the liquid cooling system by capturing transient events such as sudden changes in differential pressure. For example, abnormal pressure slope can provide early warning of pipeline blockage risk, and fluctuations in average temperature reflect changes in heat dissipation efficiency.

[0055] The purity of the working fluid (such as peak impurity concentration and dew point drift rate) quantifies the evolution of the chemical properties of the cooling medium, directly characterizing the degree of contaminant accumulation and the decline trend of phase change efficiency. Real-time monitoring of peak impurity concentration and dew point drift rate indicates that the difference between the peak impurity concentration readings is continuously increasing, which indicates that the working fluid contamination is aggravated. If the dew point drift rate reading is continuously decreasing, it exposes the risk of temperature control failure. This provides input for the generation of purity trend charts and drives the anomaly detection algorithm to accurately locate the degradation of chemical performance.

[0056] Equipment health status (such as vibration frequency energy and harmonic distortion rate) diagnoses mechanical wear and electrical aging of core components such as pumps. An abnormal increase in vibration frequency energy indicates bearing wear, and a surge in harmonic distortion rate indicates motor winding insulation failure. This forms the core basis for reliability assessment of liquid cooling systems, supporting fault root cause tracing and health index calculation.

[0057] The status of consumables (such as filter clogging rate and adsorbent saturation) is dynamically tracked to track the performance degradation trajectory of the filter components. Filter clogging rate quantifies the loss of filter cartridge efficiency, and adsorbent saturation reflects the material's adsorption capacity, accurately determining the timing of consumable replacement and avoiding over-maintenance or delayed failure. This provides a direct basis for the decision optimization module to generate maintenance work orders.

[0058] Based on the startup verification time series data, the multi-scale entropy algorithm is applied to calculate the working fluid purity entropy value and generate a multi-scale entropy value sequence.

[0059] It should be noted that, based on the start-up verification time series data, the working fluid purity feature data is extracted, the working fluid purity feature data is divided into multiple scales, and the total number of working fluid purity feature data in a single scale is counted. The time series of working fluid purity feature data in a single scale is divided into non-overlapping windows and the total number of non-overlapping windows is counted. The ratio of the total number of working fluid purity feature data to the total number of non-overlapping windows is used as the average value, and a coarse-grained sequence is generated.

[0060] Extract the sample entropy and working fluid purity entropy values ​​from the coarse-grained sequence, traverse all scales, output the total sample entropy values, and arrange them in scale order to form a multi-scale entropy value sequence.

[0061] The multi-scale entropy sequence is converted into a recursive graph, and a recursive quantized feature vector is formed by quantifying the stability and anomaly degree of the purity of the cooling working fluid.

[0062] It should be noted that the sample entropy working fluid purity entropy value is extracted from the multi-scale entropy value sequence; the sample entropy working fluid purity entropy value is used as the cooling working fluid purity; continuous data points with similar fluctuation characteristics in the cooling working fluid purity in the multi-scale entropy value sequence are identified, and these continuous data points with similar fluctuation characteristics are marked as recursive points. A binary recursion matrix is ​​constructed based on the distribution of recursive points; for example, a matrix element value of 1 marks a recursive point, and 0 marks a non-recursive point, generating a recursion graph.

[0063] Based on the recursion graph, the proportion of recursive points to the multi-scale entropy sequence is statistically analyzed to generate a recursion rate value representing the overall stability of the sequence. Diagonal structures (straight lines formed by consecutive recursive points) are identified in the recursion graph, and the proportion of diagonal lines to the total number of recursive points is statistically analyzed to generate a deterministic value representing the strength of periodicity. Vertical line segments (consecutive recursive points in the vertical direction) are detected in the recursion graph, and the proportion of vertical line segments to the total number of recursive points is statistically analyzed to generate a laminar flow value indicating abnormal state stagnation. The lengths of all diagonals in the recursion graph are collected, and the probability distribution of these lengths is statistically analyzed to assess the complexity of the liquid cooling system.

[0064] The four quantitative indicators—recursion rate, determinism, laminar flow, and length probability—are combined in a fixed order to form a feature vector. Normalization is then performed to unify the dimensional range, resulting in a recursive quantitative feature vector.

[0065] The correlation between recursive quantization feature vectors and consumable status feature data is constructed and spatiotemporally aligned to generate a working fluid purity trend map.

[0066] It should be noted that, based on the recursive quantization feature vector and the consumable status feature data, dynamic time axis calibration is performed to match the timestamp sequence of the recursive quantization feature vector with the timestamp sequence of the consumable status feature data. Dynamic time warping ensures that the two types of feature data are synchronously obtained with aligned timestamp sequences under a unified time reference.

[0067] Based on the time-aligned recursive quantization feature vector and consumable state feature data, the co-evolution of the four-dimensional indicators (recursion rate, determinism, laminar flow rate, and length probability) in the recursive quantization feature vector with the filter clogging rate and adsorbent saturation in the consumable state feature data is analyzed. For example, an abnormal increase in the laminar flow rate and the accelerated deterioration of the filter clogging rate exhibit a synchronous and predictable positive co-evolution pattern, allowing for direct early warning of accelerated filter clogging rate deterioration through an abnormal increase in the laminar flow rate. Strongly correlated feature pairs between the recursive quantization feature vector and the consumable state feature data are obtained.

[0068] By integrating the aligned timestamp sequence with strongly correlated feature pairs, the time dimension is the aligned timestamp sequence, and the feature dimension is the four-dimensional index of recursion rate, determinism, laminarity, and length probability, along with the fused data of filter clogging rate and adsorbent saturation, a trend map of working fluid purity is constructed.

[0069] The label recognition module analyzes the working fluid purity trend graph to obtain the filter pressure difference and working fluid dew point, and uses an anomaly detection algorithm to determine the safety bypass status label;

[0070] Based on the topology of the working fluid purity trend graph, the equipment operating parameters are analyzed by a depth-first traversal algorithm to generate filter pressure difference and working fluid dew point;

[0071] It should be noted that, based on the working fluid purity trend graph, the sub-nodes are visited sequentially along the branch path (condenser → pump → filter → evaporator). When visiting each node, the embedded equipment operating parameters are read.

[0072] Scan the equipment's operating parameters, extract labels with the field names "filter differential pressure" and "working fluid dew point", and extract the corresponding timestamp sequences for each. Match the timestamp sequence of the filter differential pressure to the timestamp sequence of the working fluid dew point, and finally output the filter differential pressure and working fluid dew point corresponding to the timestamp.

[0073] The dynamic fluctuation characteristics of the filter pressure difference and the periodic characteristics of the working fluid dew point are extracted, and a fusion matrix is ​​generated through spatiotemporal alignment.

[0074] It should be noted that the instantaneous fluctuation intensity of the filter pressure difference is captured to generate a pressure difference change rate sequence. The extreme points of the filter pressure difference in the pressure difference change rate sequence are identified as dynamic fluctuation characteristics of the filter pressure difference and marked as pressure difference mutation events. Simultaneously, the spectral characteristics of the working fluid dew point are decomposed to obtain the dominant frequency component and amplitude value. The product of the dominant frequency component and the amplitude value is used as the periodicity characteristic of the working fluid dew point to obtain the working fluid dew point periodic parameter sequence.

[0075] The timestamps of the differential pressure change rate sequence, differential pressure abrupt events, and working fluid dew point periodic parameter sequence are aligned to a unified time grid using linear interpolation. Under the unified time grid, these are integrated into four-dimensional features: differential pressure change rate value, differential pressure abrupt events, working fluid dew point principal periodic parameter, and amplitude value. A two-dimensional fusion matrix is ​​constructed with the time point sequence as the row dimension and the four types of feature values ​​as the column dimension, comprehensively representing the co-evolutionary relationship between differential pressure dynamic fluctuations and dew point periodicity.

[0076] The fusion matrix is ​​transformed into a recursive graph, and the nonlinear coupling relationship of differential pressure and dew point is obtained by quantizing the recursive features.

[0077] It should be noted that the fusion matrix is ​​scanned row by row, and the numerical distribution of adjacent row vectors in the fusion matrix is ​​compared. If the numerical distributions of adjacent row vectors in the fusion matrix highly overlap, they are marked as similar states and recursion points. The recursion points marked as similar states in the fusion matrix are counted, and each recursion point is filled into a two-dimensional matrix according to its timestamp to generate a recursion graph.

[0078] Based on the recursive graph, a recursive feature quantification operation is performed. The proportion of diagonal structures (straight lines formed by consecutive recursive points) in the recursive graph is statistically analyzed to generate deterministic values ​​characterizing the periodic coupling strength of pressure difference and dew point. The proportion of vertical line segments (consecutive recursive points in the vertical direction) in the recursive graph is statistically analyzed to generate laminar flow values ​​quantifying the state of stagnant coupling phenomena. The distribution of diagonal lengths in the recursive graph is statistically analyzed. If the diagonal length distribution is more dispersed, it indicates higher system complexity and lower predictability. If the diagonal length distribution is more concentrated, it indicates lower system complexity and higher predictability.

[0079] Peak deterministic values ​​indicate a strong periodic correlation between pressure differential abrupt changes and dew point drift; peak laminar flow values ​​expose abnormal working fluid retention caused by filter blockage; the lowest liquid cooling system complexity reflects a highly predictable coupling relationship; a nonlinear coupling relationship between pressure differential and dew point is constructed using three sets of quantitative indicators: deterministic values, laminar flow values, and liquid cooling system complexity.

[0080] The local density deviation is calculated based on the nonlinear coupling relationship of differential pressure and dew point to generate a density deviation anomaly score. The state level is divided by manifold space clustering method to generate a safety bypass state label.

[0081] It should be noted that the density deviation anomaly score is calculated by locating the total number of nearest neighbors of each data point in the characteristic space of the nonlinear coupling relationship between pressure difference and dew point. The expression is as follows:

[0082] ;

[0083] in, This represents the density deviation anomaly score. For the number of nearest neighbors, To represent data points The The feature vectors of the nearest neighbors, Representing data points The The feature vectors of the nearest neighbors, The index value is the number of nearest neighbors. This represents the minimum average nearest neighbor distance for all data points. This represents the maximum value of the average nearest neighbor distance for all data points.

[0084] Cluster analysis is performed using density deviation anomaly scores as input parameters, and clusters are created based on density deviation anomaly scores. For example, (0 ≤ density deviation anomaly score < 0.3): marked as "safe" state; density deviation anomaly score cluster (0.3 ≤ density deviation anomaly score < 0.7): marked as "warning" state; density deviation anomaly score cluster (0.7 ≤ density deviation anomaly score ≤ 1): marked as "dangerous" state. The clustering results are mapped to discrete state labels, and a safe bypass state label sequence is output according to the aligned timestamp sequence.

[0085] The status assessment module integrates safety bypass status tags and working fluid purity trend graphs, and generates a health status assessment report through multi-source information analysis.

[0086] The timestamps of the safety bypass status labels are dynamically aligned with the time axis of the working fluid purity trend graph to construct a causal reasoning graph.

[0087] It should be noted that the timestamp sequence of the safety bypass status label and the time axis sequence of the working fluid purity trend graph are extracted (as shown by the monitoring time points marked in the graph). The timestamps of the timestamp sequence of the safety bypass status label and the time axis sequence of the working fluid purity trend graph are normalized and calibrated, and the time-aligned safety bypass status label sequence and working fluid purity trend graph are output.

[0088] State transition events (such as "safe → dangerous" transition) are defined based on the aligned safety bypass state labels; purity anomaly events (such as a sudden drop in recursion rate exceeding 20%) are defined based on the topology of the working fluid purity trend graph. If a purity anomaly event appears in a state transition event, it is marked as a purity event; state transition events and purity anomaly events are used as nodes, and all purity events are used as causal edges to construct a causal reasoning graph.

[0089] Multimodal feature fusion of causal reasoning graph and working fluid purity trend graph is performed to generate time-series topological causal data;

[0090] It should be noted that event type markers (such as state transition types "safe → dangerous") are extracted from the causal reasoning graph, and recursive quantization feature vectors are extracted from the working fluid purity trend graph. The timestamp sequence of the causal reasoning graph and the time axis of the working fluid purity trend graph are aligned to a unified time grid. Under the unified time grid, the recursive quantization features and event type markers are integrated according to time points to generate a fused feature vector.

[0091] When the feature vectors are arranged according to the time series, a two-dimensional matrix is ​​constructed with the row dimension being the aligned time point sequence and the column dimension being the fused feature vectors, generating time-series topological causal data.

[0092] Health features are extracted from time-series topological causal data to calculate a dynamic health index, and a health status assessment report is generated by combining the causal inference graph.

[0093] It should be noted that, based on time-series topological causal data, the moving step size of the statistical recursion rate value generates purity stability, the mean of the statistical deterministic value generates periodic regularity, and the peak value of the laminar flow property generates state retention risk. A weighted summation of purity stability, periodic regularity, and state retention risk is performed to calculate the dynamic health index, expressed as:

[0094] ;

[0095] in, For dynamic health index, As a purity and stability weight, It is the moving average of the recurrence rate values ​​within the time window. For periodic regularity weights, The arithmetic mean of the deterministic values, Assigning a risk weight to the state of being stuck. This represents the global maximum value of the laminar flow property. This represents the peak value of the laminar flow property within the time window.

[0096] The value range of the purity stability weight is: Dynamically allocated based on equipment type (e.g., server racks require higher purity and stability); the range of values ​​for periodic regularity weights is... The risk weight of state dwell time is determined by inverting historical failure data of rotating equipment such as pumps / compressors; the value range of the state dwell time risk weight is... Based on the critical risk probability of working fluid retention caused by filter clogging.

[0097] Extract time-series topological causal data from the causal inference graph to generate event chains (such as the "filter blockage → pressure drop change → dangerous state" event chain), and annotate the dynamic health index during the event chain occurrence period. Align the dynamic health index during the event chain occurrence period with the timestamp of the time-series topological causal data, and combine the dynamic health index to generate a structured health status assessment report.

[0098] The decision optimization module parses the health status assessment report to generate system maintenance work orders and optimizes the anomaly detection algorithm to generate maintenance decision reports.

[0099] Extract equipment anomalies from health assessment reports to construct a maintenance decision network, calculate event correlation strength, and obtain system maintenance work orders;

[0100] It should be noted that the abnormal causal event chain is extracted from the health status assessment report, and the event type and occurrence timestamp are recorded (example: event chain = "filter blockage → pressure differential change", timestamp = "2023-05-10 14:30:00"). The event codes in the health status assessment report are parsed (example: "check filter blockage" → event code F101, "detect adsorbent saturation" → event code A203); a maintenance decision network is constructed based on the extracted abnormal causal event chain and event codes.

[0101] The system maintenance work orders are generated by prioritizing them according to the strength of the associated edges, based on the strength of the associated edges. (For example, a strength ≥ 0.8 generates a 24-hour emergency work order, 0.6 ≤ strength < 0.8 generates a 72-hour high-priority work order, and strength < 0.6 generates a 7-day regular work order.) The expression for calculating the strength of the associated edges is:

[0102] ;

[0103] in, For the strength of the associated edge, To indicate in the event Events under the condition of occurrence The conditional probability of occurrence To indicate an event and events The number of times they occur simultaneously To indicate an event The total number of times it occurred in historical data.

[0104] Execute system maintenance work orders and generate parameter correction vectors by comparing the real-time output of the anomaly detection algorithm;

[0105] It should be noted that the maintenance action field (such as "clean filter") in the system maintenance work order is read, and the corresponding maintenance operation is started (example: work order ID=MT20230510_F101, maintenance action=clean filter). During the execution of the maintenance action, the real-time output results of the anomaly detection algorithm for the same event are collected synchronously (example: timestamp when the maintenance action starts: "2023-05-10 15:00:00", anomaly detection algorithm output: "dangerous" status label).

[0106] Record equipment status changes after maintenance actions are performed (e.g., differential pressure drops from 0.5MPa to 0.1MPa after filter cleaning → maintenance successful). Compare and classify the real-time output of the anomaly detection algorithm with the maintenance results. Anomaly detection algorithms outputting "danger" and maintenance confirms the presence of an anomaly are marked as true positives; anomaly detection algorithms outputting "danger" but maintenance confirms the equipment is normal are marked as false positives; anomaly detection algorithms outputting "safe" but maintenance confirms the presence of an anomaly are marked as false negatives (e.g., anomaly detection algorithm outputs "danger" and maintenance confirms filter blockage → true positive). Generate parameter correction vectors by associating true positive, false positive, and false negative events with event codes.

[0107] By fusing parameter correction vectors and equipment anomaly events, maintenance decision reports are generated through feature state reconstruction.

[0108] It should be noted that the event code field in the parameter correction vector is matched with the maintenance action field in the maintenance decision network (example: event code F101 matches the maintenance action node "clean filter") to obtain the root cause event. The adjustment parameters in the parameter correction vector are attached to the node attribute set of the root cause event (example: add attributes to the root cause event node "filter blockage").

[0109] Extract event features (such as "filter clogging"), parameter features, and timeliness features (such as "24-hour emergency work order") from root cause events, and combine them to generate a maintenance decision report.

[0110] In summary, this invention achieves visualized monitoring of the purity decay process by dynamically capturing the evolution law of the working fluid purity and transforming the nonlinear coupling relationship of parameters such as pressure difference mutation and dew point drift into a visual spatiotemporal spectrum. At the same time, it uses work order execution feedback to calibrate algorithm parameters in real time, replaces fixed rules with dynamic thresholds, and drives the algorithm's lifelong self-evolution by combining feature reconstruction of historical false alarm data, thus eliminating the over-maintenance problem caused by static strategies.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An online purification system for the cooling working fluid in a two-phase immersion liquid cooling system, characterized in that: include, The preprocessing module collects and preprocesses device status signals to generate startup verification timing data. The trend analysis module, based on the startup verification timing data, performs purity trend analysis on the cooling working fluid and generates a working fluid purity trend chart. The label recognition module analyzes the working fluid purity trend graph to obtain the filter pressure difference and working fluid dew point, and uses an anomaly detection algorithm to determine the safety bypass status label; The status assessment module integrates safety bypass status tags and working fluid purity trend graphs, and generates a health status assessment report through multi-source information analysis. The decision optimization module parses the health status assessment report to generate system maintenance work orders and optimizes the anomaly detection algorithm to generate maintenance decision reports.

2. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 1, characterized in that: The equipment status signals include physical operating parameter signals, working fluid purity monitoring signals, equipment health status signals, and consumable status signals; The preprocessing includes filtering and denoising, normalization calibration, time alignment, and feature extraction.

3. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 2, characterized in that: The startup verification timing data includes equipment operating parameters, working fluid purity, equipment health status, and consumable status characteristic data; The purity trend analysis of the cooling medium based on the startup verification timing data is performed in the following steps. The startup verification timing data includes equipment operating parameters, working fluid purity, equipment health status, and consumable status characteristic data; Based on the startup verification time series data, the multi-scale entropy algorithm is applied to calculate the working fluid purity entropy value and generate a multi-scale entropy value sequence. The multi-scale entropy sequence is converted into a recursive graph, and a recursive quantized feature vector is formed by quantifying the stability and anomaly degree of the cooling fluid purity.

4. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 3, characterized in that: The generation of the working fluid purity trend map refers to constructing the correlation between the recursive quantification feature vector and the consumable state feature data, and performing spatiotemporal alignment to generate the working fluid purity trend map.

5. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 4, characterized in that: The process of obtaining filter pressure difference and working fluid dew point by analyzing the working fluid purity trend map refers to generating filter pressure difference and working fluid dew point by analyzing equipment operating parameters through a depth-first traversal algorithm based on the topological structure of the working fluid purity trend map.

6. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 5, characterized in that: The specific steps for determining the safety bypass status label using an anomaly detection algorithm are as follows. The dynamic fluctuation characteristics of the filter pressure difference and the periodic characteristics of the working fluid dew point are extracted, and a fusion matrix is ​​generated through spatiotemporal alignment. The fusion matrix is ​​transformed into a recursive graph, and the nonlinear coupling relationship of differential pressure and dew point is obtained by quantizing the recursive features. The local density deviation is calculated based on the nonlinear coupling relationship of differential pressure and dew point to generate a density deviation anomaly score. The state level is divided by manifold space clustering method to generate a safety bypass state label.

7. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 6, characterized in that: The specific steps for integrating the safety bypass status label and the working fluid purity trend graph are as follows. The timestamps of the safety bypass status labels are dynamically aligned with the time axis of the working fluid purity trend graph to construct a causal reasoning graph. Multimodal feature fusion of causal reasoning graph and working fluid purity trend graph is performed to generate time-series topological causal data.

8. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 7, characterized in that: The process of generating a health status assessment report through multi-source information analysis refers to extracting health features from time-series topological causal data, calculating a dynamic health index, and combining this with a causal inference graph to generate a health status assessment report.

9. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 8, characterized in that: The process of generating system maintenance work orders by parsing health status assessment reports refers to extracting abnormal equipment events from health assessment reports, constructing a maintenance decision network, calculating the correlation strength of events, and obtaining system maintenance work orders.

10. The online purification system for the cooling working fluid in a two-phase immersion liquid cooling system as described in claim 9, characterized in that: The specific steps for optimizing the anomaly detection algorithm to generate a maintenance decision report are as follows. Execute system maintenance work orders and generate parameter correction vectors by comparing the real-time output of the anomaly detection algorithm; By fusing parameter correction vectors and equipment anomaly events, maintenance decision reports are generated through feature state reconstruction.

Images