Server heat dissipation system based on single-phase immersed liquid cooling

Abstract

The invention relates to the technical field of server cooling, and discloses a server heat dissipation system based on single-phase immersed liquid cooling, which is characterized in that a liquid pool is arranged in a liquid cooling cabinet, a partition adjustable jet assembly is mounted at the bottom of the liquid cooling cabinet, and the partition adjustable jet assembly is composed of a jet orifice plate and a micro-valve array to form a plurality of independent jet areas; the cooling distribution unit comprises a heat exchanger, a circulating main pump, a high-temperature side coupling energy recovery unit and an oil product state monitoring unit; a liquid return port of the liquid pool is provided with a self-cleaning filtering unit; and the control unit operates a model prediction control algorithm based on the kinematic viscosity data, the water content data and the particle concentration data in combination with the temperature signal, the flow signal and the pressure difference signal, predicts oil product viscosity drift in combination with a space-time Transformer model, and outputs the opening degree of a micro valve in each injection area, the rotating speed of a bypass micro pump and the sequential control quantity of back flushing. According to the invention, the balance of the cooling liquid flow field, the real-time compensation of the physical property of the oil product and the closed-loop linkage of self-cleaning are realized, and the energy efficiency of the system is improved.

Description

Technical Field

[0001] This invention relates to the field of server cooling technology, and more specifically, to a server heat dissipation system based on single-phase immersion liquid cooling. Background Technology

[0002] With the increasing use of high-density computing servers in data centers, the heat load per server rack continues to rise. Traditional air-cooling methods, limited by low heat exchange efficiency, high noise, and low space utilization, are no longer sufficient to meet the thermal management requirements of high-power servers. In recent years, single-phase immersion liquid cooling technology has been widely studied and gradually applied to data center server cooling systems due to its advantages of high heat exchange efficiency, low noise, and ease of deployment. Its basic principle is to completely immerse the entire server or its main heat-generating components in insulating coolant, achieving efficient heat transfer through liquid convection. The heat is then recovered by the cooling distribution unit and exchanged with the secondary cooling system, thereby maintaining stable equipment operation.

[0003] However, existing patent CN118890861A – Single-phase immersion liquid cooling system for data centers and its adaptive control method – proposes a structure that sets up an internal liquid pool in the liquid cooling cabinet, achieves fluid circulation through injection pipes and filters, and uses reinforcement learning combined with regression algorithms to achieve temperature difference optimization control. Although this scheme improves the coolant distribution and temperature equalization problems to some extent, it still has the following shortcomings in practical applications: The injection components of CN118890861A have a fixed aperture structure, and the liquid distribution relies on a passive liquid equalization plate, which cannot adjust the injection flow rate in real time according to the local heat load, resulting in uneven temperature distribution and the risk of hot spots. Existing control strategies are mainly based on temperature difference and pump speed feedback, and do not consider in real time the impact of changes in coolant kinematic viscosity, water content, and particle concentration over time on flow resistance and heat transfer performance. A multi-stage filtration and backwashing structure is not formed, nor is a correlation judgment established between particle concentration and pressure difference changes, making it difficult to prevent nozzle clogging and flow deviation in a timely manner.

[0004] Therefore, it is necessary to design a server cooling system based on single-phase immersion liquid cooling to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a server heat dissipation system based on single-phase immersion liquid cooling, which aims to solve the problems of uneven flow field distribution and insufficient dynamic response, lack of control model for coolant property changes, and lack of self-cleaning and predictive maintenance mechanisms in the prior art.

[0006] This invention proposes a server heat dissipation system based on single-phase immersion liquid cooling, comprising: Liquid cooling cabinet, cooling distribution unit, self-cleaning filter unit and control unit, The liquid cooling cabinet is equipped with a liquid pool, and a zoned adjustable spray assembly is installed at the bottom of the liquid pool. The zoned adjustable spray assembly consists of a spray orifice plate and a micro valve array to form multiple independent spray zones. The equivalent diameter of the spray orifice in each spray zone is 2 to 10 mm and the opening is adjustable. Each spray zone is equipped with a bypass micro pump. The cooling distribution unit includes a heat exchanger, a circulating main pump, and an oil condition monitoring unit. The oil condition monitoring unit is used to detect the kinematic viscosity data, water content data, and particle concentration data of the coolant in real time. The return port of the liquid tank is equipped with the self-cleaning filter unit, which includes a two-stage filter element and a backwash solenoid valve for performing backwashing under control commands. The cooling distribution unit also includes a high-temperature side-coupled energy recovery unit, which consists of a thermoelectric power generation module and a DC-DC conversion module, and is used to convert the waste heat of the coolant into electrical energy. The control unit, based on the kinematic viscosity data, water content data, and particle concentration data output by the oil condition monitoring unit, and combined with temperature signals, flow signals, and differential pressure signals, runs a model predictive control algorithm. It uses a spatiotemporal Transformer model to predict oil viscosity drift and outputs the microvalve opening degree of each injection zone, the speed of the bypass micropump, and the timing control quantity of backflushing. When a flow deviation or particle abnormality is detected, it triggers self-cleaning and injection spectrum adjustment operations.

[0007] Furthermore, when the oil condition monitoring unit detects the kinematic viscosity, water content, and particle concentration of the coolant in real time, it includes: The oil condition monitoring unit includes a sampling branch connected in parallel with the return pipeline, a constant temperature measurement chamber, a micro-vibration beam viscosity sensor, a capacitive water content sensor, a laser scattering particle counter, a micropore debubbler, and a data fusion module. The sampling branch uses a micro-sampling pump to continuously sample at a flow rate of 0.1–1.0 L / min. The sample liquid enters the constant temperature measurement chamber after passing through the microporous deaerator and is measured at 35–45℃ with a temperature fluctuation ≤ ±0.5℃. The micro-vibrating beam viscosity sensor obtains kinematic viscosity data. The particle counter counts the 4–100 μm particle diameter channels and outputs particle concentration data. The front end of the laser scattering particle counter does not have a filter material with a pore size smaller than 4 μm. The capacitive water content sensor measures the change in dielectric constant of the coolant in the constant temperature measurement chamber and calculates the water content data based on the pre-stored calibration curve. The data fusion module outputs kinematic viscosity data in a 1–5 s cycle and outputs water content data and particle concentration data in a 10–30 s cycle.

[0008] Furthermore, before executing the model predictive control algorithm, the control unit also includes: Based on the kinematic viscosity data, particle concentration data, temperature signal, flow rate signal, and pressure difference signal, a state-space model including coolant flow resistance coefficient, zone pressure drop, and temperature gradient is constructed, and a state vector is calculated with a sampling period of 1 to 2 seconds. The state vector is used to characterize the coupling relationship between the zone flow field and the heat load in real time.

[0009] Furthermore, the model predictive control algorithm of the control unit takes the state vector as input and, within a prediction window of 30–120 s, uses chip junction temperature deviation, zone flow non-uniformity, and pump power weights as the objective function. The constraints include microvalve opening range of 0%–100%, bypass micropump speed change rate ≤20% / s, and pressure difference upper limit. The model predictive control algorithm incorporates a kinematic viscosity correction parameter, which is used to correct the influence of coolant viscosity changes on flow distribution in the prediction model.

[0010] Furthermore, the control unit combines the spatiotemporal Transformer model to perform multi-head self-attention calculation on the collected temperature signal, flow signal, pressure difference signal and kinematic viscosity data, extracts spatiotemporal related features to predict the changing trend of oil viscosity drift rate and zone drag coefficient in the next 30 to 120 seconds, and updates the kinematic viscosity correction parameters and flow resistance terms in the state space model according to the prediction results.

[0011] Furthermore, the control unit uses the first control quantity of the model predictive control algorithm as the microvalve opening setting value and bypass micropump speed setting value for each injection zone, and applies acceleration / deceleration limits and a minimum holding time of 0.5 to 2 seconds to the setting values, and calculates the balancing bias based on the unevenness of the zoned flow to correct the microvalve opening.

[0012] Furthermore, the control unit generates the timing control quantity for backwashing based on the change rate of particle concentration data, the change rate of differential pressure signal, and the change trend of zone resistance coefficient. When the triggering condition is met, the micro-valve opening setting value of the corresponding spray area is lowered to no more than 10%, while the corresponding bypass micro-pump speed setting value is increased by 10% to 30%, and the backwashing solenoid valve is controlled to execute in a sequence of opening every 0.5 to 1.5 seconds and cycling 2 to 5 times at intervals of 30 to 120 seconds.

[0013] Furthermore, the control unit also generates a pre-adjustment control quantity based on the changing trends of the oil viscosity drift rate and the zone resistance coefficient. The pre-adjustment control quantity includes the baseline bias of the microvalve opening setting value for each injection zone and the baseline bias of the bypass micropump speed setting value, and coordinates with the timing control quantity of the backflushing in a priority manner; within a predetermined recovery time of 2 to 10 seconds, the setting value is linearly returned to the setting value calculated by the model predictive control algorithm.

[0014] Furthermore, when the control unit triggers the self-cleaning and jet spectrum adjustment operation upon detecting flow deviation or particle abnormality, it includes: When the control unit detects that the unevenness of the zone flow exceeds a set threshold or that either the rate of change of particle concentration data or the rate of change of differential pressure signal exceeds a set threshold, it generates a self-cleaning trigger command and a jet spectrum adjustment command. In the case of a slight abnormality, the jet spectrum adjustment operation is executed and maintained for 5 to 30 seconds. If the unevenness of the zone flow does not fall back to within the threshold, the timing control quantity of the backwash is additionally executed. In the case of a severe abnormality, the jet spectrum adjustment command and the timing control quantity of the backwash are issued simultaneously.

[0015] Furthermore, the injection spectrum adjustment operation is performed by the microvalve array to periodically open and close the microvalve opening setpoint of each injection area in a pulse manner. The oscillation opening and closing period is 0.5 to 3 seconds and the duty cycle is 20% to 60%. During the oscillation, the corresponding bypass micropump speed setpoint is increased by 5% to 20%. Within a recovery time of 2 to 10 seconds after the oscillation ends, the control unit linearly returns the microvalve opening setpoint and the bypass micropump speed setpoint to the setpoint calculated by the model predictive control algorithm. At the same time, the duty cycle is positively correlated with the oil viscosity drift rate to compensate for the increase in flow resistance caused by the increase in viscosity.

[0016] Compared with existing technologies, the advantages of this invention are as follows: Through the synergy of a zoned adjustable injection component and a bypass micro-pump, the coolant is dynamically distributed within the liquid pool of the liquid-cooled cabinet according to the local heat load of the server. The micro-valve array and the injection orifice plate work together to achieve real-time adjustment of the orifice equivalent diameter and opening, improving the uniformity of zoned flow and suppressing hot spots from the source. The oil condition monitoring unit within the cooling distribution unit continuously outputs kinematic viscosity data, water content data, and particle concentration data. The control unit fuses these data with temperature, flow, and differential pressure signals, and then, based on a model predictive control algorithm and combined with a spatiotemporal Transformer model, analyzes the oil viscosity drift and zoned resistance changes. Forward-looking prediction and online compensation enable adaptive optimization of flow distribution and heat exchange capacity according to oil properties and load fluctuations, thereby maintaining stable chip junction temperature, reducing pump power and overcooling redundancy during long-term operation; the self-cleaning filter unit adopts a two-stage filter element and a backwash solenoid valve, and uses the change rate of particle concentration data and differential pressure signal as the trigger criterion, combined with the injection spectrum adjustment operation to perform pulse oscillation flushing on the injection orifice plate, which can prevent scale and micro-clogging of the injection orifice, shorten maintenance downtime and extend component life; the high-temperature side-coupled energy recovery unit consists of a thermoelectric power generation module and a DC-DC conversion module, which converts the waste heat of the coolant into electrical energy to feed back into the sensing and control links, improving system energy efficiency and independence. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a structural block diagram of a server heat dissipation system based on single-phase immersion liquid cooling, provided for an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] In traditional single-phase immersion liquid-cooled server cooling systems, passive liquid distribution plates with fixed orifice diameters cannot dynamically adjust the jet flow rate according to local heat load, resulting in uneven coolant distribution in different areas of the liquid-cooled cabinet. Real-time changes in coolant kinematic viscosity, water content, and particle concentration are not incorporated into the control strategy, leading to dynamic deviations in flow resistance coefficient and heat transfer efficiency. The return pipeline lacks multi-stage filtration and backwashing mechanisms, causing particulate matter accumulation, nozzle blockage, and flow deviations, further exacerbating the unevenness of the temperature field distribution.

[0020] For example, in scenarios where high-density computing servers operate continuously at full load, the kinematic viscosity of the coolant increases due to long-term circulation, and micron-sized particles gradually deposit at the spray nozzles, creating localized flow resistance. The flow distribution at the bottom of the coolant pool becomes unbalanced, and the increased flow resistance in the spray areas corresponding to some high heat flux density chips leads to a decrease in coolant flow rate, reduced heat exchange efficiency, and localized temperature rise. The particle concentration in the return pipeline continues to increase without timely monitoring, filter blockage causes increased pressure differential, increased power consumption of the main circulation pump, and deterioration of system energy efficiency.

[0021] If the above issues are not addressed, the temperature gradient within the liquid-cooled cabinet will exceed the safe junction temperature threshold of the server chips, triggering thermal protection mechanisms to force frequency reduction or even causing hardware damage. The cumulative effects of coolant viscosity drift and particulate matter deposition will lead to a non-linear increase in flow resistance of the jet assembly, causing the main circulation pump to operate under overload conditions for extended periods, accelerating mechanical wear and system aging. The lack of self-cleaning functionality will force maintenance personnel to frequently shut down the system to replace filters, reducing data center operational continuity and increasing maintenance costs.

[0022] For this, please refer to Figure 1 As shown, this application proposes a server cooling system based on single-phase immersion liquid cooling, including: a liquid-cooled cabinet, a cooling distribution unit, a self-cleaning filtration unit, and a control unit. A liquid tank is installed inside the liquid-cooled cabinet, and a zoned adjustable spray assembly is installed at the bottom of the liquid tank. The zoned adjustable spray assembly consists of a spray orifice plate and a micro-valve array, forming multiple independent spray zones. The equivalent diameter of the spray orifices in each spray zone is 2–10 mm, and the opening is adjustable. A bypass micro-pump is provided for each spray zone. The cooling distribution unit includes a heat exchanger, a circulating main pump, and an oil condition monitoring unit. The oil condition monitoring unit is used to monitor the kinematic viscosity, water content, and particle concentration data of the coolant in real time. A self-cleaning filtration unit is installed at the return port of the liquid tank. The self-cleaning filtration unit includes a two-stage filter element and a backwash solenoid valve, used to perform backwashing under control commands. The cooling distribution unit also includes a high-temperature side-coupled energy recovery unit, which consists of a thermoelectric power generation module and a DC-DC converter module, used to convert waste heat from the coolant into electrical energy. Based on the kinematic viscosity data, water content data, and particle concentration data output by the oil condition monitoring unit, and combined with temperature signals, flow signals, and differential pressure signals, the control unit runs a model predictive control algorithm. It combines a spatiotemporal Transformer model to predict oil viscosity drift and outputs the microvalve opening degree of each injection zone, the speed of the bypass micropump, and the timing control quantity of backflushing. When a flow deviation or particle abnormality is detected, it triggers self-cleaning and injection spectrum adjustment operations.

[0023] Specifically, a liquid-cooled cabinet refers to a closed cabinet that houses servers and stores coolant. It can be implemented using a composite structure of a metal frame and thermal insulation materials. Inside the cabinet, a liquid pool is installed to immerse the servers and store the coolant. A zoned adjustable spray assembly is installed at the bottom of the pool to provide dynamic flow regulation. The zoned adjustable spray assembly is an adjustable spray device composed of a spray orifice plate and a micro-valve array. Specifically, it uses a combination of micro-valves with independently controllable opening in each zone and a multi-aperture spray orifice plate to form multiple independent spray zones. By adjusting the opening of the micro-valve in each zone and the speed of the bypass micro-pump, differentiated coolant supply is achieved for different heat load zones. The cooling distribution unit is a system module for coolant circulation and heat exchange. It is implemented using a combination of a heat exchanger, a main circulation pump, and an oil condition monitoring unit. The heat exchanger exchanges heat between the coolant and the secondary cooling system, the main circulation pump drives coolant circulation, and the oil condition monitoring unit monitors the kinematic viscosity, water content, and particle concentration of the coolant in real time to assess changes in fluid performance. The self-cleaning filtration unit refers to the filtration device installed at the return port of the coolant pool. It employs a combination of a dual-stage filter element and a backwashing solenoid valve. The dual-stage filter element intercepts particulate contaminants of different sizes, while the backwashing solenoid valve removes filter blockages through reverse flushing to maintain filtration efficiency. The high-temperature side-coupled energy recovery unit converts waste heat from the coolant into electrical energy. It utilizes a combination of a thermoelectric generator module and a DC-DC converter module. The thermoelectric generator produces electricity based on the temperature difference effect, while the DC-DC converter converts the electricity into a stable DC output to power internal system equipment. The control unit is a computational module that achieves intelligent regulation based on multi-source signals. It employs a computational architecture that fuses model predictive control algorithms with a spatiotemporal Transformer model. By real-time acquisition of temperature, flow rate, differential pressure, and oil condition data, it predicts coolant viscosity drift and flow resistance trends, dynamically adjusting the micro-valve opening, bypass micro-pump speed, and backwashing sequence to optimize system performance.

[0024] This application achieves synergistic optimization of dynamic coolant flow distribution and self-cleaning function by combining a zoned adjustable injection component with a multi-parameter fusion control strategy. By real-time monitoring of coolant kinematic viscosity, water content, and particle concentration, and combining model predictive control algorithms with a spatiotemporal Transformer model, injection parameters and backwashing timing are dynamically corrected. This effectively solves problems such as flow resistance changes, uneven heat distribution, and nozzle clogging caused by coolant performance degradation, thereby improving system heat dissipation efficiency and reliability.

[0025] The working process and principle of this application are as follows: a liquid tank is set up inside the liquid-cooled cabinet, and a zoned adjustable spray assembly is installed at the bottom of the liquid tank. The zoned adjustable spray assembly consists of a spray orifice plate and a micro-valve array, forming multiple independent spray zones. The equivalent diameter of the spray orifice in each spray zone is 2-10mm and the opening is adjustable, and a bypass micro-pump is provided accordingly. This design allows for precise adjustment of the coolant spray volume according to the heat dissipation requirements of different zones.

[0026] The cooling distribution unit includes a heat exchanger, a main circulating pump, and an oil condition monitoring unit. The oil condition monitoring unit monitors the kinematic viscosity, water content, and particle concentration of the coolant in real time. These data reflect changes in the physical properties of the coolant and have a significant impact on system performance.

[0027] The liquid return port of the liquid tank is equipped with a self-cleaning filter unit, including a two-stage filter element and a backwash solenoid valve. Under control command, backwashing can be performed to effectively remove accumulated impurities.

[0028] The cooling distribution unit also includes a high-temperature side-coupled energy recovery unit, consisting of a thermoelectric power generation module and a DC-DC converter module. This design can convert waste heat from the coolant into electrical energy, improving system energy efficiency.

[0029] The control unit is the core of the system. Based on data output from the oil condition monitoring unit, combined with temperature, flow rate, and differential pressure signals, it runs a model predictive control algorithm. Simultaneously, it uses a spatiotemporal Transformer model to predict oil viscosity drift. Based on these inputs and predictions, the control unit outputs the microvalve opening degree for each injection zone, the bypass micropump speed, and the timing control parameters for backflushing. When flow deviations or particle abnormalities are detected, it also triggers self-cleaning and injection spectrum adjustment operations.

[0030] It achieves precise control of coolant distribution, real-time monitoring of oil status, system self-cleaning function, and energy recovery and utilization. By comprehensively considering various factors and adopting advanced control algorithms, it can adapt to different operating conditions and maintain efficient and stable operation.

[0031] As a preferred embodiment, the solution of this application is specifically implemented as follows: The liquid-cooled cabinet houses a liquid tank. At the bottom of the tank is a zoned adjustable spray assembly, consisting of a stainless steel spray orifice plate and a micro-valve array of a microelectromechanical system (MEMS). The spray assembly is divided into eight independent spray zones, each containing 100 nozzles with an equivalent nozzle diameter of 5 mm. Each spray zone is equipped with a corresponding magnetically driven bypass micro-pump.

[0032] The heat exchanger in the cooling distribution unit adopts a plate heat exchanger structure, and the main circulating pump is a variable frequency centrifugal pump. The oil condition monitoring unit includes a micro-vibration beam viscosity sensor, a capacitive water content sensor, and a laser scattering particle counter, which are used to measure kinematic viscosity, water content, and particle concentration, respectively.

[0033] The self-cleaning filter unit uses a two-stage filter element made of stainless steel mesh with pore sizes of 100μm and 20μm respectively. The backwash solenoid valve is normally closed and has a rated power of 10W.

[0034] The high-temperature side-coupled energy recovery unit uses a semiconductor thermoelectric power generation module with a thermoelectric conversion efficiency of 5%. The DC-DC converter module boosts the low-voltage DC output from the thermoelectric module to 48V.

[0035] The control unit uses an industrial-grade embedded computer, running a Python-based model predictive control algorithm and a spatiotemporal Transformer model. The control cycle is 1 second, and the prediction window is 60 seconds.

[0036] During system operation, the control unit dynamically adjusts the opening of the micro-valve in the spray zone and the speed of the bypass micro-pump based on data from various sensors to achieve precise flow distribution. When an abnormally high particle concentration is detected, a backwashing procedure is triggered to clean the filter element. Simultaneously, the spray spectrum is adjusted by rapidly opening and closing the micro-valve to prevent nozzle clogging.

[0037] Through the above-described scheme, this application achieves precise control and adaptive adjustment of the liquid cooling system. The zoned adjustable injection assembly can dynamically adjust the coolant distribution according to the heat dissipation needs of different areas, effectively suppressing hotspot formation. The combination of real-time oil condition monitoring and model prediction algorithms enables the system to respond promptly to changes in the physical properties of the coolant, maintaining optimal cooling performance. The self-cleaning filter unit and injection spectrum adjustment function effectively prevent system clogging and extend maintenance cycles. The high-temperature side energy recovery unit improves the system's energy utilization efficiency. The synergistic effect of these technical features enhances the performance and reliability of the single-phase immersion liquid-cooled server heat dissipation system, providing strong support for the stable operation of high-density computing servers.

[0038] In some of the solutions described above in this application, the oil condition monitoring unit needs to detect the kinematic viscosity, water content and particle concentration of the coolant in real time. However, in actual operation, the coolant sampling process may cause measurement errors due to temperature fluctuations or bubble interference. The difference in detection cycle of different parameters may affect the response speed of the control system. The particle detection accuracy is limited by the pre-filter material and cannot accurately reflect the true concentration.

[0039] This application further proposes an oil condition monitoring unit for real-time detection of the kinematic viscosity, water content, and particle concentration of coolant. The oil condition monitoring unit includes a sampling branch connected in parallel with the return pipeline, a constant-temperature measurement chamber, a micro-vibration beam viscosity sensor, a capacitive water content sensor, a laser scattering particle counter, a microporous deaerator, and a data fusion module. The sampling branch uses a micro-sampling pump to continuously sample at a flow rate of 0.1–1.0 L / min. The sample liquid passes through the microporous deaerator and enters the constant-temperature measurement chamber, where measurements are taken at 35–45℃ with temperature fluctuations ≤ ±0.5℃. The micro-vibration beam viscosity sensor obtains the kinematic viscosity data. The particle counter counts particles in channels with diameters of 4–100 μm and outputs particle concentration data. The laser scattering particle counter does not have a filter material with a pore size smaller than 4 μm at its front end. The capacitive water content sensor measures the change in dielectric constant of the coolant in the constant temperature measurement chamber and calculates the water content data based on the pre-stored calibration curve. The data fusion module outputs kinematic viscosity data in a 1-5s cycle and water content data and particle concentration data in a 10-30s cycle.

[0040] The parallel connection of the sampling branch and return pipeline avoids the impact of main circulation flow fluctuations on sampling stability, and the flow range of the micro-sampling pump matches the detection requirements under different operating conditions. The microporous deaerator eliminates the interference of air bubbles on viscosity and dielectric constant measurements through physical interception. The constant temperature measurement chamber uses a precision temperature control device to maintain a constant temperature environment, eliminating the impact of temperature drift on viscosity and moisture content detection. The laser scattering particle counter is directly exposed to the sample liquid flow path, avoiding the retention of small particles by the pre-filter material and ensuring the integrity of particle concentration detection above 4μm. The data fusion module performs time synchronization and error compensation on the raw data from different sensors. Kinematic viscosity data is output at a high frequency to support real-time control, while moisture content and particle concentration data are output at a low frequency to balance the system load.

[0041] Specifically, a miniature sampling pump introduces coolant from the return line into the sampling branch at a constant flow rate. As the sample passes through a microporous deaerator, bubbles larger than 50 μm in diameter are intercepted and expelled. The deaerated sample flows into a constant-temperature measurement chamber, where a temperature control module maintains a constant temperature, eliminating interference from temperature changes on the viscosity and moisture content sensors. The micro-vibrating beam viscosity sensor calculates kinematic viscosity by measuring changes in vibration frequency, updating its results every 1–5 seconds. The capacitive moisture content sensor detects changes in the dielectric constant between two plates and, combined with pre-stored dielectric constant curves corresponding to different moisture contents, calculates the current moisture content, updating its results every 10–30 seconds. A laser scattering particle counter directly analyzes the scattered light intensity of particles in the sample without a pre-filter, classifying and statistically analyzing particle concentrations within the 4–100 μm particle size range, with data updates synchronized with moisture content detection. The data fusion module performs timestamp alignment and outlier removal on viscosity, moisture content, and particle concentration data. Kinematic viscosity data is preferentially transmitted to the control unit for real-time flow regulation, while moisture content and particle concentration data are smoothed for long-term trend analysis. This structural design ensures the accuracy and timeliness of key parameter detection, providing reliable input for subsequent model predictive control.

[0042] As a preferred embodiment, the solution of this application is specifically implemented as follows: The oil condition monitoring unit includes a sampling branch connected in parallel with the return pipeline, a constant-temperature measurement chamber, a micro-vibration beam viscosity sensor, a capacitive water content sensor, a laser scattering particle counter, a microporous deaerator, and a data fusion module. The sampling branch uses a micro-sampling pump to continuously sample at a flow rate of 0.5 L / min. The sample liquid passes through the microporous deaerator and enters the constant-temperature measurement chamber, where measurements are taken at 40℃ with a temperature fluctuation of ±0.3℃. The micro-vibration beam viscosity sensor obtains kinematic viscosity data. The particle counter counts particles in channels with diameters of 4μm, 6μm, 14μm, 21μm, 38μm, and 70μm and outputs particle concentration data. The laser scattering particle counter does not have a filter material with a pore size smaller than 4μm at its front end. The capacitive water content sensor measures the change in dielectric constant of the coolant within the constant-temperature measurement chamber and calculates the water content data based on a pre-stored calibration curve. The data fusion module outputs kinematic viscosity data at 3-second intervals and water content and particle concentration data at 20-second intervals.

[0043] Through the above technical solutions, this application achieves real-time and accurate monitoring of coolant kinematic viscosity, water content, and particle concentration. The constant-temperature measurement chamber ensures the stability of the measurement environment, while the micro-vibration beam viscosity sensor provides high-precision kinematic viscosity data. The laser scattering particle counter counts multiple particle size channels, making particle concentration analysis more comprehensive. The capacitive water content sensor, combined with a calibration curve, enables accurate determination of water content. The data fusion module outputs various parameters at different cycles, balancing the data update frequency with system load. This monitoring data provides reliable input for subsequent control strategies, helping to promptly detect changes in coolant performance and prevent potential reductions in heat dissipation efficiency and equipment failure risks.

[0044] In some of the solutions described above in this application, when constructing a state-space model that includes coolant flow resistance coefficient, zone pressure drop, and temperature gradient, the data acquisition cycle and the real-time nature of the state vector are not clearly defined, which makes the model unable to accurately represent the coupling relationship between the dynamic flow field and the heat load, thereby affecting the prediction accuracy of the control algorithm.

[0045] This application further proposes to construct a state-space model based on kinematic viscosity data, particle concentration data, temperature signal, flow signal and pressure difference signal, which includes coolant flow resistance coefficient, zone pressure drop and temperature gradient, and calculates the state vector with a sampling period of 1 to 2 seconds. The state vector is used to characterize the coupling relationship between the zone flow field and the heat load in real time.

[0046] The state-space model's input parameters include kinematic viscosity data, particle concentration data, temperature signal, flow rate signal, and pressure difference signal. The output parameters are the coolant flow resistance coefficient, zone pressure drop, and temperature gradient. The data acquisition period is set to 1–2 seconds, and the input parameters are averaged over a fixed time window. The state vector, composed of the flow resistance coefficient, zone pressure drop, and temperature gradient, is updated in real-time through matrix operations.

[0047] Specifically, during model construction, kinematic viscosity and particle concentration data are used to calculate the dynamic changes in the coolant drag coefficient. Temperature and flow signals are combined with differential pressure signals to generate pressure drop and temperature gradients for each zone. The data acquisition cycle is set to 1–2 seconds to ensure the real-time performance of the state vector while avoiding excessive computational load caused by high-frequency sampling. The state vector describes the coupling relationship between the flow field and heat load through a system of linear equations. The drag coefficient is related to coolant viscosity and particle concentration, the zone-level pressure drop reflects the differences in flow distribution in each injection zone, and the temperature gradient characterizes the heat load distribution. The input parameters are mean-filtered every 1–2 seconds to eliminate instantaneous noise interference. The filtered data is then substituted into the state-space equations to solve for the current state vector. This state vector serves as the input to the model predictive control algorithm, providing dynamic flow field parameters for subsequent micro-valve opening and pump speed adjustments.

[0048] As a preferred embodiment, the solution of this application is implemented as follows: Before executing the model predictive control algorithm, the control unit constructs a state-space model containing the coolant drag coefficient, zone pressure drop, and temperature gradient based on kinematic viscosity data, particle concentration data, temperature signal, flow rate signal, and differential pressure signal. This state-space model uses discrete-time representation, and the state variables include the flow rate, pressure drop, and temperature of each injection zone. The system inputs are the microvalve opening and the bypass micropump speed, and the system outputs are the measured temperature, flow rate, and differential pressure signals. The drag coefficient in the model is updated based on real-time kinematic viscosity data, and particle concentration data is used to correct the pressure drop calculation. The state equation and output equation are obtained through a system identification method. The control unit calculates the state vector with a sampling period of 1.5 s, and the state vector is used to characterize the coupling relationship between the zoned flow field and the thermal load in real time. The extended Kalman filter algorithm is used for state estimation to improve the estimation accuracy of the state vector.

[0049] Through the above technical solution, this application achieves accurate modeling of the dynamic characteristics of the liquid cooling system. By constructing a state-space model that includes flow resistance, pressure drop, and temperature gradient, and integrating data from multiple sensors, the control unit can monitor the system's operating status in real time. The use of a short 1.5s sampling period to calculate the state vector ensures timely capture of rapid system changes. This provides a reliable system model and state estimate for subsequent model predictive control, helping to improve control accuracy and system response speed. Simultaneously, by incorporating coolant property changes into the model, the adaptability of the control strategy to changes in oil condition is enhanced.

[0050] In some of the solutions described above in this application, the model predictive control algorithm makes predictions based on the state-space model. However, the change in coolant viscosity over time can lead to deviations in flow distribution predictions, affecting the accuracy of chip junction temperature control and energy efficiency.

[0051] This application further proposes a model predictive control algorithm that uses a state vector as input and, within a prediction window of 30 to 120 seconds, uses chip junction temperature deviation, zone flow non-uniformity, and pump power weights as the objective function. Constraints include a microvalve opening range of 0 to 100%, a bypass micropump speed change rate not exceeding 20% ​​per second, and a pressure differential limit. The model predictive control algorithm incorporates a kinematic viscosity correction parameter, which is used to correct the impact of coolant viscosity changes on flow distribution in the prediction model.

[0052] The objective function balances temperature uniformity, flow distribution, and energy consumption through multi-objective optimization. The chip junction temperature deviation term uses the root mean square error, the zone flow non-uniformity term uses the standard deviation, and the pump power weight term is dynamically adjusted based on the current system energy consumption mode. Among the constraints, the microvalve opening range corresponds to the physical actuator limit, the limitation on the bypass micropump speed change rate prevents hydraulic shock, and the differential pressure upper limit protects pipeline safety. The kinematic viscosity correction parameter is updated through an online identification algorithm, and its value forms a closed-loop feedback with the kinematic viscosity data output by the oil condition monitoring unit.

[0053] Specifically, the model predictive control algorithm predicts the system dynamics over the next 30 to 120 seconds based on the state vector within each control cycle, and calculates the optimal setpoints for the microvalve opening and bypass micropump speed through rolling optimization. Kinematic viscosity correction parameters are embedded in the flow resistance calculation module of the predictive model to compensate for flow distribution errors caused by changes in coolant viscosity in real time. For example, when an increase in kinematic viscosity leads to an increase in flow resistance, the correction parameter adjusts the viscosity term in the predictive model, causing the optimization algorithm to increase the bypass micropump speed or increase the microvalve opening in the corresponding region in advance. Constraints are forcibly satisfied through a quadratic programming solver to ensure that the rate of change and absolute value of the control variables are within a safe range. The weighting coefficient of the chip junction temperature deviation term in the objective function is adaptively adjusted according to the heat load intensity, prioritizing the reduction of temperature fluctuations under high-temperature conditions and focusing on reducing pump power consumption under low-temperature conditions.

[0054] As a preferred embodiment, the solution of this application is implemented as follows: The model predictive control algorithm of the control unit takes the state vector as input and, within a 60-second prediction window, uses the chip junction temperature deviation, zone flow non-uniformity, and pump power weight term as the objective function. Constraints include a micro-valve opening range of 0%–100%, a bypass micro-pump speed change rate not exceeding 15% / s, and a pressure difference upper limit. A kinematic viscosity correction parameter is added to the model predictive control algorithm to correct the impact of coolant viscosity changes on flow distribution in the prediction model. Specifically, the control unit first obtains the state vector, which includes information such as flow rate, pressure, and temperature for each zone. Then, within a 60-second prediction window, the optimal control strategy is determined by solving an optimization problem. The objective function consists of three parts: a chip junction temperature deviation term to ensure the chip temperature is within a safe range, a zone flow non-uniformity term to balance the cooling effect of each zone, and a pump power weight term to reduce system energy consumption. Constraints limit the range of micro-valve opening and bypass micro-pump speed changes to ensure system stability. Furthermore, a kinematic viscosity correction parameter is incorporated into the prediction model to compensate for flow distribution deviations caused by changes in coolant viscosity. As a result, the control unit can predict future trends based on the system's real-time status and calculate the optimal control parameters, achieving precise adjustment of the cooling system.

[0055] Through the above technical solution, this application enables precise predictive control of the server cooling system. The model predictive control algorithm considers several key factors, including chip temperature, flow uniformity, and energy consumption, ensuring both cooling performance and system efficiency. The constraint settings ensure smooth changes in the control quantity, avoiding system oscillations. The introduction of a kinematic viscosity correction parameter effectively compensates for the impact of coolant property changes on the system, improving control accuracy. Compared to traditional feedback control, this model-based predictive control method exhibits better dynamic performance and robustness, adapting to server load changes and environmental disturbances, ensuring the stable and efficient operation of the cooling system.

[0056] In some of the above-mentioned schemes in this application, the model predictive control algorithm relies on the state-space model to predict and generate control quantities. However, the coolant viscosity drift and the change of the zone resistance coefficient have spatiotemporal dynamic characteristics. Traditional prediction models are difficult to capture the nonlinear coupling relationship between multiple variables, resulting in the lag in updating the viscosity correction parameters and flow resistance terms, which affects the accuracy of flow distribution and the thermal equilibrium effect.

[0057] This application further proposes that the control unit, combined with the spatiotemporal Transformer model, performs multi-head self-attention calculation on the collected temperature signal, flow signal, pressure difference signal and kinematic viscosity data, extracts spatiotemporal correlation features to predict the changing trend of oil viscosity drift rate and zone resistance coefficient in the next 30 to 120 seconds, and updates the kinematic viscosity correction parameters and the flow resistance term in the state space model based on the prediction results.

[0058] The spatiotemporal Transformer model processes time-series data of temperature, flow rate, pressure difference, and kinematic viscosity in parallel using a multi-head self-attention mechanism. It calculates the correlation weights of different sensor signals in the temporal and spatial dimensions, generating a hidden layer vector containing spatiotemporal characteristics. The prediction module outputs the changing trends of oil viscosity drift rate and zone drag coefficient based on the hidden layer vector, with the prediction window set to 30–120 s to adapt to the dynamic response speed under different operating conditions. The kinematic viscosity correction parameter is dynamically adjusted at a rate of 0.1%–0.5% / s based on the viscosity drift rate. The flow resistance term adopts an incremental update strategy in the state-space model, with the update amplitude limited to ±5% in each iteration to avoid oscillations.

[0059] Specifically, temperature, flow, and differential pressure signals are input into the spatiotemporal Transformer model with a sampling period of 1–2 seconds, while kinematic viscosity data is input synchronously with a period of 1–5 seconds. A multi-head self-attention layer divides the time series of each signal into 8–16 attention heads, calculates the correlation matrix for different time steps and partition locations, and generates a spatiotemporal feature vector through weighted fusion. The prediction module uses a fully connected network to map the feature vector to viscosity drift rate and drag coefficient change rate, and outputs the prediction results to the model predictive control algorithm with a period of 10–30 seconds. Viscosity correction parameters are applied to the kinematic viscosity baseline value with a linear compensation term based on the drift rate. The flow resistance term is replaced in the state-space equation with a weighted average of the predicted drag coefficient and the real-time differential pressure signal, with weighting coefficients set to 0.6–0.8 to balance the reliability of predicted and measured data. This process enables the model predictive control algorithm to compensate for flow distribution deviations caused by viscosity changes in advance and adaptively adjust the flow resistance parameters to match the actual partition heat load distribution, thereby improving the dynamic response accuracy and stability of the injection zone flow control.

[0060] As a preferred embodiment, the solution of this application is specifically implemented as follows: The control unit uses a spatiotemporal Transformer model to perform multi-head self-attention calculations on the collected temperature, flow, pressure differential, and kinematic viscosity data. First, these signals and data are arranged in a time series to form an input matrix. Then, through the multi-head self-attention mechanism, weights are calculated for different time points and data types to extract temporal and spatial correlation features.

[0061] Specifically, the multi-head self-attention mechanism comprises multiple parallel attention heads, each independently calculating attention weights. Each attention head transforms the input into three matrices—query, key, and value—through a linear transformation, then calculates the dot product between the query and key matrices and performs softmax normalization to obtain the attention weights. Finally, the attention weights are multiplied by the value matrix to obtain the head's output. The outputs of multiple heads are then concatenated and linearly transformed to form the final self-attention output.

[0062] Furthermore, the spatiotemporal Transformer model also includes feedforward neural network layers and layer normalization for further feature extraction. The model extracts higher-level spatiotemporal feature representations layer by layer by stacking multiple Transformer encoder layers.

[0063] Therefore, the model can capture the complex spatiotemporal relationships between temperature, flow rate, pressure difference, and viscosity. For example, an increase in temperature in a certain region may lead to a decrease in local viscosity, which in turn affects the flow rate distribution. Or, a change in pressure difference at a certain moment may predict future viscosity trends.

[0064] Based on the extracted spatiotemporal features, the model predicts the changing trends of oil viscosity drift rate and zone drag coefficient over the next 30–120 seconds. The prediction uses a regression head to map the extracted features to specific predicted values. The prediction results are used to update the kinematic viscosity correction parameters and the flow resistance term in the state-space model.

[0065] Specifically, the kinematic viscosity correction parameter is a dynamically adjusted coefficient used to correct for the impact of coolant viscosity changes on flow distribution in the prediction model. When a future increase in viscosity is predicted, the correction parameter increases accordingly to compensate for the increase in flow resistance. Conversely, it decreases. The flow resistance term in the state-space model is also updated based on the predicted trend of the zone drag coefficient changes to ensure that the model accurately reflects the actual system state.

[0066] Through the above technical solution, this application can achieve accurate prediction and dynamic correction of oil viscosity and zoned resistance. By employing a spatiotemporal Transformer model, the system can capture complex spatiotemporal correlations, improving prediction accuracy. Timely updates to the prediction results allow the control strategy to anticipate impending changes, thereby achieving more proactive and precise flow control. This not only improves the system's response speed but also enhances the stability of heat dissipation, effectively preventing the formation of localized hot spots. Simultaneously, by dynamically adjusting the kinematic viscosity correction parameters and flow resistance terms, the system can adapt to long-term changes in oil properties, maintaining long-term operational reliability and efficiency.

[0067] In some of the solutions described above in this application, the first control quantity output by the model predictive control algorithm is directly used as the setpoint for the microvalve opening in each injection zone and the setpoint for the bypass micropump speed. However, sudden changes in the setpoints may cause frequent actuator movements, leading to flow fluctuations and pressure pulsations. Furthermore, relying solely on the model predictive control algorithm may not be able to promptly correct flow distribution deviations caused by differences in flow resistance in the injection zones.

[0068] This application further proposes to use the first control quantity of the model predictive control algorithm as the microvalve opening setpoint and bypass micropump speed setpoint for each injection zone, and to apply acceleration / deceleration limits and a minimum holding time of 0.5 to 2 seconds to the setpoints, and to calculate the balancing bias based on the non-uniformity of the zoned flow to correct the microvalve opening.

[0069] Acceleration and deceleration limits are achieved by setting an upper limit for the microvalve opening change rate to prevent sudden changes in opening commands exceeding 20% / second within adjacent control cycles. The minimum hold time requirement is set to remain unchanged for 0.5 to 2 seconds after adjustment to avoid frequent start-stop of the actuator. The balancing offset is generated based on the real-time calculated zonal flow unevenness. When the unevenness exceeds 5%, the microvalve opening is increased by 1% to 3% in the injection area with low flow, while the opening is reduced by the same amount in the area with high flow.

[0070] Specifically, after the model predictive control algorithm generates the microvalve opening setpoint and the bypass micropump speed setpoint, the control unit first applies a rate limiting process to the setpoints to ensure that the microvalve opening change rate does not exceed 20% per second. Subsequently, the setpoints enter a holding phase, remaining constant for 0.5–2 seconds to stabilize the flow field. Simultaneously, the control unit calculates the deviation between the flow rate and the average value in each injection zone in real time. When the flow non-uniformity in a zone exceeds a set threshold, a proportional-integral algorithm is used to calculate the balancing offset. This offset is added to the original setpoint, causing the microvalve opening in high-flow-rate areas to decrease by 1%–3%, and the opening in low-flow-rate areas to increase by the same amount. After the holding phase, the new setpoints are output after another rate limiting, forming a closed-loop regulation. This process effectively suppresses flow oscillations and improves the flow balance between zones by slowing down the rate of change of the setpoints and introducing dynamic offset compensation.

[0071] As a preferred embodiment, the solution of this application is specifically implemented as follows: The control unit uses the first control variable from the model predictive control algorithm as the microvalve opening setpoint and bypass micropump speed setpoint for each injection zone. Acceleration / deceleration limits are applied to these setpoints, with a minimum hold time of 1 second. A trim bias is calculated based on the zone flow non-uniformity to correct the microvalve opening.

[0072] Specifically, the control unit first receives a first control quantity output by the model predictive control algorithm. This control quantity includes the target opening value of the microvalve in each injection zone and the target rotational speed value of the bypass micropump. Further, the control unit imposes acceleration and deceleration limits on these target values. For example, the rate of change of the microvalve opening is limited to within 20% per second, and the rate of change of the bypass micropump rotational speed is limited to within 15% per second.

[0073] Therefore, the control unit ensures smooth operation of each actuator, avoiding drastic fluctuations. Simultaneously, the control unit sets a minimum hold time of 1 second. This means that even if a new target value is reached, the actuator will maintain its current state for at least 1 second before making adjustments. As a preferred implementation, this minimum hold time can be fine-tuned according to the system response characteristics, ranging from 0.5 to 2 seconds.

[0074] Furthermore, the control unit calculates the balancing bias based on the unevenness of the flow rate in each zone. Specifically, it first calculates the deviation between the actual flow rate and the average flow rate in each injection zone. Then, based on the magnitude and direction of the deviation, it corrects the micro-valve opening setpoint. For example, in areas with lower flow rates, the micro-valve opening will be appropriately increased, while in areas with higher flow rates, the micro-valve opening will be appropriately decreased. Through this dynamic balancing mechanism, the system can better balance the cooling effect in each zone.

[0075] Through the above technical solutions, this application achieves precise control over the opening degree of the micro-valve in the injection zone and the rotational speed of the bypass micro-pump. The setting of acceleration / deceleration limits and minimum hold time effectively suppresses drastic fluctuations in the control parameters, improving system stability. The real-time correction mechanism for zoned flow non-uniformity further optimizes coolant distribution, resulting in a more uniform temperature distribution across different zones. This multi-level control strategy enhances the adjustment accuracy and response speed of the liquid cooling system, providing a reliable guarantee for the stable operation of the server.

[0076] In some of the solutions described above in this application, the control unit generates the microvalve opening and bypass micropump speed setpoints through model predictive control algorithms. However, in actual operation, the accumulation of particulate matter in the coolant may cause blockage of the injection orifice plate, resulting in abnormal flow deviation and pressure difference. Existing solutions cannot promptly determine the degree of blockage and perform targeted flushing operations, which poses a risk of local overheating.

[0077] This application further proposes a control unit that generates timing control quantities for backwashing based on the rate of change of particle concentration data, the rate of change of differential pressure signal, and the trend of change of zone resistance coefficient. When the triggering condition is met, the micro-valve opening setting value of the corresponding spray area is lowered to no more than 10%, while the corresponding bypass micro-pump speed setting value is increased by 10% to 30%, and the backwashing solenoid valve is controlled to execute in a sequence of opening every 0.5 to 1.5 seconds and cycling 2 to 5 times at intervals of 30 to 120 seconds.

[0078] The particle concentration change rate was calculated using particle concentration data output by a laser scattering particle counter at 10–30 s intervals. The differential pressure signal change rate was calculated using differential operations based on real-time monitoring values ​​of zoned pressure drop in the state-space model. The trend of zoned resistance coefficient change was derived from the dynamic update results of the flow resistance term predicted by the spatiotemporal Transformer model. When the microvalve opening setting was lowered to below 10%, the flow rate in the corresponding injection zone was suppressed, and the bypass micropump speed was increased by 10%–30% to compensate for flow loss and avoid fluctuations in the total system flow. The backwash solenoid valve opened for 0.5–1.5 s and then closed, repeating this process 2–5 times at 30–120 s intervals, flushing away particle deposits on the filter element surface with short-duration high-pressure reverse liquid flow.

[0079] Specifically, when the rate of change in particle concentration exceeds a threshold or the trend of the differential pressure signal change and resistance coefficient indicates an abnormal increase in flow resistance, the control unit determines to trigger backflushing. At this time, the opening of the micro-valve in the corresponding spray area is limited to below 10%, reducing the flow load in the clogged area. Simultaneously, the bypass micro-pump speed is increased by 10%–30% to maintain overall system flow balance. The backflushing solenoid valve opens with short pulses of 0.5–1.5 seconds, forming periodic high-pressure backflushing at intervals of 30–120 seconds. 2–5 cycles can remove particulate matter from the filter element surface. For example, when the rate of change in particle concentration exceeds 5% / min and the rate of increase in differential pressure exceeds 200 Pa / s, the solenoid valve opens for 1 second, cycles 3 times at 60-second intervals, the micro-valve opening decreases to 8%, and the bypass micro-pump speed is increased by 20%. This timing control effectively removes blockages and maintains system stability through the synergistic effect of short-duration high-frequency backflushing and flow compensation.

[0080] As a preferred embodiment, the solution of this application is specifically implemented as follows: The control unit generates the timing control quantity for backwashing based on the rate of change of particle concentration data, the rate of change of differential pressure signal, and the trend of change of zone resistance coefficient. When the triggering condition is met, the micro-valve opening setting value of the corresponding spray zone is reduced to 10%, while the corresponding bypass micro-pump speed setting value is increased by 20%, and the backwashing solenoid valve is controlled to execute in a sequence of opening 1 second and cycling 3 times at 60-second intervals.

[0081] Specifically, the control unit first acquires the particle concentration data output by the oil condition monitoring unit and calculates its rate of change. Simultaneously, it acquires the differential pressure signal from the differential pressure sensor and calculates the rate of change of differential pressure. Furthermore, based on the previously established state-space model, it calculates the variation trend of the zone drag coefficient for each injection zone.

[0082] Therefore, the control unit comprehensively analyzes the above three parameters, and triggers a reverse flushing operation when any parameter exceeds a preset threshold. For example, the triggering condition is determined to be met when the particle concentration change rate exceeds 5% / min, the pressure difference change rate exceeds 2% / min, or the zone resistance coefficient shows a continuous upward trend.

[0083] Upon triggering backflushing, the control unit first lowers the microvalve opening setting for the corresponding spray area to 10%. This operation aims to reduce the forward flow, creating conditions for backflushing. Simultaneously, the rotational speed setting of the corresponding bypass micropump is increased by 20% to increase the backflushing flow rate.

[0084] Next, the control unit sends a control command to the backflushing solenoid valve, causing it to open and close according to a preset sequence. In this embodiment, the solenoid valve cycles three times, opening every 1 second and then at 60-second intervals. This intermittent flushing effectively removes blockages while avoiding interference with the system caused by prolonged backflushing.

[0085] After the backwash is completed, the control unit gradually restores the micro-valve opening and bypass micro-pump speed to normal operating conditions, completing one self-cleaning cycle.

[0086] Through the above technical solution, this application achieves automated cleaning and maintenance of the cooling system. By monitoring changes in particle concentration, pressure differential, and resistance in real time, potential blockage risks can be detected promptly. The backflushing method effectively removes deposits from the injection holes and pipes, preventing long-term blockages that could lead to a decrease in cooling efficiency. The intermittent flushing strategy ensures cleaning effectiveness while minimizing disruption to normal operation. Simultaneously, by adjusting the micro-valve opening and bypass micro-pump speed, the backflushing flow rate is optimized, improving cleaning efficiency. This self-cleaning mechanism extends the system's maintenance cycle, improves operational reliability, and reduces manual maintenance costs.

[0087] In some of the solutions described above in this application, when the control unit performs backwashing and jet spectrum adjustment operations, the lack of a coordination mechanism between the pre-adjustment control quantity and the real-time control command may cause a conflict between the set values ​​of the micro-valve opening and the bypass micro-pump speed, resulting in flow fluctuations or response lag.

[0088] This application further proposes a control unit that generates pre-adjusted control quantities based on the changing trends of oil viscosity drift rate and zone resistance coefficient. These pre-adjusted control quantities include baseline biases for the microvalve opening setpoints of each injection zone and for the bypass micropump speed setpoints, and are prioritized and coordinated with the timing control quantities for backflushing. Within a predetermined recovery time of 2–10 seconds, the setpoints are linearly restored to the setpoints calculated by the model predictive control algorithm.

[0089] The pre-adjusted control quantity is generated using a predictive model of the changing trends of oil viscosity drift rate and zone resistance coefficient. The baseline bias is calculated as a linear function of the viscosity change rate and the resistance coefficient change rate. The priority coordination mechanism sets the backflushing timing control quantity as a high-priority command, freezing the superposition effect of the pre-adjusted control quantity during backflushing triggering and restoring it after the backflushing sequence is completed. The linear regression process uses a time window piecewise interpolation algorithm to gradually transition the setpoints of the microvalve opening and bypass micropump speed to the original setpoints of the model predictive control algorithm within 2–10 seconds with a fixed slope.

[0090] Specifically, when the oil viscosity drift rate exceeds a threshold, the control unit applies a baseline bias of -5% to +3% to the microvalve opening setpoint based on the viscosity change gradient, and simultaneously applies a baseline bias of +5% to +15% to the bypass micropump speed setpoint. During the backflushing trigger phase, the control unit marks the backflushing command as a high-priority task, suspends the superposition calculation of pre-adjusted control quantities, and prioritizes the operations of reducing the microvalve opening to no more than 10% and increasing the bypass micropump speed by 10% to 30%. After the backflushing solenoid valve completes flushing according to the preset pulse sequence, the control unit linearly increases the microvalve opening from 10% to the current setpoint output by the model predictive control algorithm at a rate of 2% per second within a 5-second recovery time, while simultaneously reducing the bypass micropump speed to the original setpoint at a rate of 4% per second. During this process, the baseline bias of the pre-adjusted control quantities re-activates after the recovery time, avoiding actuator oscillation caused by the superposition of multiple control quantities.

[0091] As a preferred embodiment, the solution of this application is implemented as follows: The control unit generates a pre-adjustment control quantity based on the changing trends of oil viscosity drift rate and zone resistance coefficient. This pre-adjustment control quantity includes a baseline bias for the microvalve opening setpoint of each injection zone and a baseline bias for the bypass micropump speed setpoint. The baseline bias is generated by inputting the changing trends of oil viscosity drift rate and resistance coefficient into a linear regression model to calculate the compensation amount for the microvalve opening setpoint and the increment of the bypass micropump speed. Further, the pre-adjustment control quantity and the timing control quantity for backflushing are prioritized and coordinated. When backflushing is triggered, the execution of the pre-adjustment control quantity is paused and resumed after the backflushing sequence is completed. Within a predetermined recovery time, the control unit linearly reverts the microvalve opening setpoint and the bypass micropump speed setpoint to the setpoint calculated by the model predictive control algorithm. For example, when the recovery time is set to 5 seconds, the control unit adjusts the microvalve opening from 20% after baseline offset to 30% of the model prediction output at a rate of 4% per second, while reducing the bypass micropump speed from 2500 r / min to 2000 r / min at a rate of 100 r / min per second.

[0092] Through the above technical solution, this application can dynamically compensate for the impact of coolant viscosity changes and flow channel resistance fluctuations on flow distribution, while coordinating the conflict between backflushing operations and conventional control, avoiding flow distribution instability caused by sudden particle blockage or viscosity mutations. Therefore, while maintaining chip junction temperature uniformity, it reduces mechanical wear caused by frequent microvalve adjustments and improves the system's robustness against coolant degradation and particulate contamination.

[0093] In some of the solutions described above in this application, when the non-uniformity of the zoned flow exceeds the set threshold or the rate of change of particle concentration data and the rate of change of differential pressure signal are abnormal, the existing control strategy relies on a single adjustment method and cannot dynamically coordinate the execution sequence and parameter matching of the jet flow adjustment and backwashing operation according to the level of abnormality, which may lead to sudden changes in local flow resistance or excessively long recovery time.

[0094] This application further proposes that when the detected zonal flow non-uniformity exceeds a set threshold, or when either the particle concentration data change rate or the differential pressure signal change rate exceeds a set threshold, a self-cleaning trigger command and a jet spectrum adjustment command are generated. Specifically, in the case of a minor anomaly, a jet spectrum adjustment operation is performed and maintained for 5–30 seconds. If the zonal flow non-uniformity does not return to within the threshold, a timing control quantity for backwashing is added. In the case of a severe anomaly, both the jet spectrum adjustment command and the backwashing timing control quantity are issued simultaneously.

[0095] The injection spectrum adjustment operation involves a microvalve array periodically opening and closing the microvalve opening setpoints for each injection zone in a pulse manner. The oscillation period is 0.5–3 seconds, and the duty cycle is 20%–60%. During the oscillation, the corresponding bypass micropump speed setpoint is increased by 5%–20%. In the recovery phase, a linear method is used to restore the microvalve opening setpoints and bypass micropump speed setpoints to the setpoints calculated by the model predictive control algorithm within 2–10 seconds. At the same time, the duty cycle is finely adjusted based on the oil viscosity drift rate.

[0096] Specifically, when flow deviation or particle abnormalities trigger self-cleaning, the abnormality level is first determined. For minor abnormalities, only injection spectrum adjustment is initiated. This involves altering the local flow field through periodic oscillations of the micro-valve opening, utilizing the pulse impact effect to peel away particulate deposits adhering to the nozzle inner wall. Maintaining the duty cycle within the 20%–60% range balances flushing intensity and flow stability, while increasing the bypass micro-pump speed by 5%–20% compensates for the instantaneous flow drop caused by oscillations. If the flow non-uniformity does not recover after adjustment, a reverse flushing operation is superimposed, using high-pressure backflushing to further clear blockages. For severe abnormalities, both operations are performed simultaneously to shorten recovery time. During the recovery phase, a linear regression approach is used to avoid abrupt changes in the setpoint. The duty cycle is dynamically adjusted based on the oil viscosity drift rate; for example, increasing the duty cycle when viscosity increases enhances flushing force and compensates for the increased flow resistance caused by increased viscosity. This graded response mechanism balances abnormality handling efficiency with system stability.

[0097] As a preferred embodiment, the solution of this application is implemented as follows: When the non-uniformity of the zone flow is detected to exceed a set threshold, or when either the rate of change of particle concentration data or the rate of change of differential pressure signal exceeds the set threshold, the control unit generates a self-cleaning trigger command and a jet spectrum adjustment command. In the case of mild abnormality, the microvalve array periodically opens and closes the set value of the microvalve opening in each jet area in a pulse manner. The oscillation opening and closing period is set to 0.5 seconds to 3 seconds, and the duty cycle is controlled in the range of 20% to 60%. At the same time, during the oscillation, the set value of the corresponding bypass micropump speed is increased by 5% to 20%. If the non-uniformity of the zone flow does not fall back to within the threshold after being maintained for 5 to 30 seconds, a backwash operation is additionally executed. In the case of severe abnormality, the jet spectrum adjustment command and the timing control quantity of backwash are issued synchronously, and the backwash solenoid valve is executed in a sequence of opening every 0.5 seconds and cycling 5 times at 30-second intervals.

[0098] Through the above technical solution, this application achieves a rapid response to abnormal coolant flow distribution and particulate contamination. By dynamically adjusting the injection frequency and coordinating the self-cleaning operation, it effectively eliminates local flow channel blockage and restores flow balance, avoiding the decline in heat dissipation performance caused by viscosity drift or contaminant accumulation, and ensuring the thermal stability of the server under complex operating conditions.

[0099] In some of the solutions described above in this application, when flow deviation or particle abnormality is detected, although adjusting the microvalve opening and bypass micropump speed through model predictive control algorithm can partially alleviate the problem, sudden particle accumulation or sudden change in local flow resistance can easily lead to instantaneous blockage of the injection orifice plate. Traditional linear adjustment methods are difficult to quickly remove the blockage and restore flow balance.

[0100] This application further proposes that the injection spectrum adjustment operation is carried out by a microvalve array performing periodic opening and closing oscillations on the microvalve opening setpoint of each injection zone in a pulse manner. The oscillation opening and closing period is 0.5 to 3 seconds and the duty cycle is 20% to 60%. During the oscillation, the corresponding bypass micropump speed setpoint is increased by 5% to 20%. Within the recovery time of 2 to 10 seconds after the oscillation ends, the control unit linearly returns the microvalve opening setpoint and bypass micropump speed setpoint to the setpoint calculated by the model predictive control algorithm. At the same time, the duty cycle is positively correlated with the oil viscosity drift rate to compensate for the increase in flow resistance caused by the increase in viscosity.

[0101] The periodic opening and closing oscillation employs a combination of a fixed period and a variable duty cycle. The lower limit of the duty cycle is set at 20% to prevent localized overheating caused by complete closure of the microvalve, while the upper limit is set at 60% to prevent flow fluctuations from exceeding the system's tolerance. The increase in bypass micropump speed is negatively correlated with the oscillation duty cycle: a 20% increase in speed at a 20% duty cycle and a 5% increase at a 60% duty cycle. The recovery time is dynamically adjusted based on the oscillation duration: 10 seconds for 30 seconds of oscillation and 2 seconds for 5 seconds. The duty cycle fine-tuning coefficient is calculated based on the viscosity drift rate; the duty cycle increases by 3% for every 10% increase in viscosity.

[0102] Specifically, when the change rate of particle concentration or flow deviation triggers the injection spectrum adjustment, the microvalve array alternately opens and closes at a cycle of 0.5 to 3 seconds, forming a pulsed jet that impacts the particles deposited in the nozzle. The duty cycle is controlled within the range of 20% to 60%, ensuring that the instantaneous flow velocity impact force effectively clears blockages while avoiding drastic flow fluctuations that could affect heat dissipation stability. Simultaneously increasing the bypass micropump speed by 5% to 20% compensates for the flow loss during the pulse shutdown phase, maintaining the overall flow balance in the loop. After the oscillation ends, the microvalve opening and pump speed linearly revert to the model predictive control setpoint within 2 to 10 seconds, preventing abrupt changes from causing secondary oscillations. When the oil viscosity increases, leading to increased flow resistance, the duty cycle is adjusted positively; for example, when the viscosity increases by 20%, the duty cycle is increased by 6%, enhancing the pulsed jet intensity to overcome viscous resistance and ensuring that the clearing effect is not affected by oil deterioration.

[0103] As a preferred embodiment, the solution of this application is implemented as follows: When the rate of change of particle concentration data exceeds a preset threshold and the differential pressure signal shows an upward trend, the control unit activates the injection spectrum adjustment operation. The microvalve array generates an oscillation control signal according to the preset pulse period parameters, and periodically adjusts the microvalve opening of the target injection area with a period of 1.2 seconds, wherein the opening time accounts for 45% of the single cycle duration. During the oscillation adjustment, the bypass micropump speed setting value of the corresponding area is increased to 115% of the original reference value. After the oscillation operation is executed for 18 seconds, the recovery phase begins, and the control unit gradually restores the microvalve opening and pump speed to the set value output by the model predictive control algorithm within 6 seconds using linear interpolation. At the same time, based on the real-time calculated oil viscosity drift rate, when the viscosity increases by more than 3% per hour, the duty cycle is increased by 0.8% to enhance the flow channel scouring intensity.

[0104] Through the above technical solution, this application effectively solves the problem that traditional fixed injection modes cannot dynamically adapt to changes in coolant viscosity and particle deposition. By periodically oscillating and adjusting the opening and closing state of the micro-valve in conjunction with pump speed compensation, the removal of deposited particles in the flow channel can be achieved without interrupting the cooling process. At the same time, through the dynamic correlation between viscosity drift rate and duty cycle, the increase in flow resistance caused by oil deterioration is automatically compensated, ensuring the balance of flow distribution and the stability of system operation.

[0105] In the above embodiments, the coordinated use of a zoned adjustable injection component and a bypass micropump enables dynamic distribution of coolant within the liquid pool of the liquid-cooled cabinet according to the local heat load of the server. The microvalve array and the injection orifice plate work together to achieve real-time adjustment of the equivalent diameter and opening of the nozzles, improving the uniformity of the zoned flow and suppressing hot spots from the source. The oil condition monitoring unit in the cooling distribution unit continuously outputs kinematic viscosity data, water content data, and particle concentration data. The control unit fuses these data with temperature, flow, and differential pressure signals, and then uses a model predictive control algorithm combined with a spatiotemporal Transformer model to predict and compensate for oil viscosity drift and zoned resistance changes. This allows the flow distribution and heat exchange capacity to adaptively optimize with oil properties and load fluctuations, thereby maintaining stable chip junction temperature and reducing pump power and overcooling redundancy during long-term operation. The self-cleaning filter unit employs a dual-stage filter element and a backwashing solenoid valve. It uses the change rate of particle concentration data and differential pressure signal as trigger criteria, combined with jet spectrum adjustment operations to perform pulse oscillation flushing on the jet orifice plate. This effectively prevents scale and micro-clogging in the jet orifices, shortens maintenance downtime, and extends component lifespan. The high-temperature side-coupled energy recovery unit consists of a thermoelectric power generation module and a DC-DC conversion module. It converts waste heat from the coolant into electrical energy to feed back into the sensing and control components, improving system energy efficiency and independence.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A single-phase immersion liquid cooling based server heat dissipation system, characterized in that, The application relates to a liquid cooling cabinet, a cooling distribution unit, a self-cleaning filter unit and a control unit. The liquid cooling cabinet is provided with a liquid pool, the bottom of the liquid pool is provided with a partition-adjustable spraying assembly, the partition-adjustable spraying assembly is composed of a spraying hole plate and a micro valve array, a plurality of independent spraying areas are formed, the equivalent diameter of the spraying hole of each spraying area is 2-10 mm and the opening degree is adjustable, and a bypass micro pump is arranged corresponding to each spraying area. The cooling distribution unit comprises a heat exchanger, a circulating main pump and an oil product state monitoring unit, the oil product state monitoring unit is used for detecting the kinetic viscosity data, the water content data and the particle concentration data of the cooling liquid in real time. The backflow outlet of the liquid pool is provided with the self-cleaning filter unit, the self-cleaning filter unit comprises a double-stage filter core and a backwashing electromagnetic valve and is used for executing reverse flushing under a control instruction. The cooling distribution unit further comprises a high-temperature side coupled energy recovery unit, the high-temperature side coupled energy recovery unit is composed of a thermoelectric power generation module and a direct current conversion module and is used for converting the waste heat of the cooling liquid into electric energy. The control unit is based on the kinetic viscosity data, the water content data and the particle concentration data output by the oil product state monitoring unit, combines temperature signals, flow signals and pressure difference signals, runs a model predictive control algorithm, predicts oil product viscosity drift by combining a space-time Transformer model, outputs the opening degree of the micro valve of each spraying area, the rotating speed of the bypass micro pump and the timing control amount of the reverse flushing and triggers the self-cleaning and spraying spectrum adjustment operation when detecting flow deviation or particle abnormality. When the oil product state monitoring unit detects the kinetic viscosity, the water content and the particle concentration of the cooling liquid in real time, the following steps are included.

2. The single-phase immersion liquid cooling based server heat dissipation system according to claim 1, wherein, The oil product state monitoring unit comprises a sampling branch connected in parallel with a backflow pipeline, a constant-temperature measuring cavity, a micro vibration beam viscosity sensor, a capacitive water content sensor, a laser scattering particle counter, a micropore deaerator and a data fusion module. The sampling branch is continuously sampled by a micro sampling pump at a flow rate of 0.1-1.0 L / min, the sample liquid enters the constant-temperature measuring cavity after passing through the micropore deaerator and is measured under the condition that the temperature is 35-45 DEG C and the temperature fluctuation is less than or equal to plus or minus 0.5 DEG C, the kinetic viscosity data are obtained by the micro vibration beam viscosity sensor, the particle counter counts the channels with a particle size of 4-100 mu m and outputs the particle concentration data, no filter material with a pore size less than 4 mu m is arranged at the front end of the laser scattering particle counter, the capacitive water content sensor measures the dielectric constant change of the cooling liquid in the constant-temperature measuring cavity and obtains the water content data according to a pre-stored calibration curve, and the data fusion module outputs the kinetic viscosity data at a period of 1-5 s, the water content data and the particle concentration data at a period of 10-30 s. Before the control unit executes the model predictive control algorithm, the following steps are further included.

3. The single-phase immersion liquid cooling based server heat dissipation system of claim 2, wherein, A state space model containing cooling liquid flow resistance coefficient, zone-level pressure drop and temperature gradient is constructed based on the kinetic viscosity data, the particle concentration data, the temperature signals, the flow signals and the pressure difference signals, and a state vector is calculated at a sampling period of 1-2 s, the state vector is used for real-time representing the partition flow field and the thermal load coupling relationship. ​ 4. The single-phase immersion liquid cooling based server heat dissipation system of claim 3, wherein, The model predictive control algorithm of the control unit takes the state vector as input, predicts within a 30-120s window, and forms a target function composed of chip junction temperature deviation, partition flow unevenness, and pump power weight term. The constraint conditions include micro-valve opening range 0%-100%, bypass micro-pump speed change rate ≤20% / s, and pressure difference upper limit. The model predictive control algorithm adds a kinematic viscosity correction parameter, which is used to correct the influence of cooling liquid viscosity change on flow distribution in the prediction model.

5. The single-phase immersion liquid cooling based server heat dissipation system of claim 4, wherein, The control unit combines the Spacetime Transformer model to perform multi-head self-attention calculation on the collected temperature signal, flow signal, pressure difference signal, and kinematic viscosity data, extracts spatiotemporal correlation features to predict the oil viscosity drift rate and partition resistance coefficient trend within the next 30-120s, and updates the kinematic viscosity correction parameter and flow resistance term in the state space model according to the prediction results.

6. The single-phase immersion liquid cooling based server heat dissipation system of claim 5, wherein, The control unit sets the first control quantity of the model predictive control algorithm as the micro-valve opening degree set value and the bypass micro-pump speed set value of each injection area, applies acceleration and deceleration limits and a minimum holding time of 0.5-2s to the set values, and calculates a trim bias based on the partition flow unevenness to correct the micro-valve opening degree.

7. The single-phase immersion liquid cooling based server heat dissipation system of claim 6, wherein, The control unit generates a backflush timing control quantity based on the particle concentration data change rate, the pressure difference signal change rate, and the partition resistance coefficient trend. When the trigger condition is met, the micro-valve opening degree set value of the corresponding injection area is lowered to no more than 10%, the corresponding bypass micro-pump speed set value is increased by 10%-30%, and the backflush solenoid valve is controlled to open for 0.5-1.5s and cycle 2-5 times with an interval of 30-120s.

8. The single-phase immersion liquid cooling based server heat dissipation system of claim 7, wherein, The control unit also generates a pre-adjustment control quantity based on the oil viscosity drift rate and the partition resistance coefficient trend, which includes a baseline bias for the micro-valve opening degree set value of each injection area and a baseline bias for the bypass micro-pump speed set value, and prioritizes the backflush timing control quantity; within a predetermined recovery time of 2-10s, the set values are returned to the set values calculated by the model predictive control algorithm in a linear manner.

9. The single-phase immersion liquid cooling based server heat dissipation system of claim 1, wherein, When the control unit detects a flow deviation or particle anomaly and triggers the self-cleaning and injection spectrum adjustment operations, it includes: When the control unit detects that the partition flow unevenness exceeds a set threshold or either the particle concentration data change rate or the pressure difference signal change rate exceeds a set threshold, it generates a self-cleaning trigger instruction and an injection spectrum adjustment instruction: when the anomaly is mild, the injection spectrum adjustment operation is performed and maintained for 5-30s, and if the partition flow unevenness does not fall within the threshold, the backflush timing control quantity is additionally executed; when the anomaly is severe, the injection spectrum adjustment instruction and the backflush timing control quantity are issued simultaneously.

10. The single-phase immersion liquid cooling based server heat dissipation system of claim 9, wherein, The injection spectrum adjustment operation is implemented by the micro valve array to periodically open and close oscillation of the micro valve opening degree set value of each injection area, the oscillation open and close period is 0.5-3s, the duty cycle is 20%-60%, and the corresponding bypass micro pump speed set value is increased by 5%-20% during the oscillation, and the control unit returns the micro valve opening degree set value and the bypass micro pump speed set value to the set value calculated by the model predictive control algorithm within the recovery time of 2-10s after the oscillation ends, and the duty cycle is positively adjusted according to the oil viscosity drift rate to compensate for the increase in flow resistance caused by the increase in viscosity.

Images