A control method of a humanoid robot low-noise liquid cooling heat dissipation system

Abstract

This invention discloses a control method for a low-noise liquid cooling system for a humanoid robot, comprising the following steps: Step 1: Multiphysics data acquisition; Step 2: Multiphysics spatial partitioning; Step 3: Construction of a five-dimensional spatiotemporal tensor; Step 4: Dominant coupling relationship; Step 5: Noise multi-scale decomposition; Step 6: Matrix generation; Step 7: Differentiated control; Step 8: Cooperative scheduling; Step 9: Model update. This invention also includes a low-noise liquid cooling system for a humanoid robot, comprising the following modules: a multiphysics monitoring module, a spatial partitioning module, a coupling analysis module, a noise prediction module, a silence tag generation module, a parameter configuration module, a cooperative scheduling module, and an online learning module. By constructing a multi-dimensional spatiotemporal tensor, this invention can clearly identify the nonlinear correlations and influence laws between various physical quantities, resulting in faster liquid cooling control response, smaller steady-state error, and stronger system stability.

Description

Technical Field

[0001] This invention relates to the field of low-noise liquid cooling heat dissipation system for humanoid robots, and particularly to a low-noise liquid cooling heat dissipation system and control method for humanoid robots. Background Technology

[0002] As embodied intelligence and humanoid robots move from laboratory research and development to mass production, robot core controllers are rapidly evolving towards high computing power, high integration, low power consumption, and small size. Liquid cooling, due to its high heat dissipation efficiency and strong space adaptability, has become the core heat dissipation solution for embodied intelligent equipment such as humanoid robots, L4 level unmanned vehicles, and quadruped robot dogs.

[0003] Current liquid cooling systems rely solely on chip temperature as feedback, employing simple PID algorithms for closed-loop regulation. This neglects the strong coupling between temperature, flow rate, vibration, noise, and computing load, making them ill-suited for handling nonlinear disturbances under complex operating conditions. Consequently, issues such as temperature oscillations, flow rate fluctuations, and control lag arise, leading to frequent chip frequency reduction and incomplete utilization of computing power. Therefore, this invention proposes a low-noise liquid cooling system and control method for humanoid robots. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a low-noise liquid cooling system and control method for humanoid robots.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A control method for a low-noise liquid cooling system for a humanoid robot includes the following steps:

[0007] Step 1: Multiphysics Data Acquisition and 3D Field Construction. Acquire data on the temperature, coolant inlet and outlet temperatures, real-time flow rate, vibration amplitude and spectrum of pumps, valves and pipelines, noise sound pressure level, and chip load power consumption of the heterogeneous computing chip, drive module, pumps, valves, pipelines and radiator of the humanoid robot's chest cavity and whole body. Simultaneously collect the spatial coordinate information corresponding to each sensor. Use the Kriging spatial interpolation algorithm to perform spatial fitting on discrete monitoring points to generate a 3D spatial distribution map of temperature field, flow field, vibration field and noise field, which completely restores the heat, flow, vibration and noise distribution state of the robot's internal whole domain.

[0008] Step 2: Multiphysics Spatial Partitioning and Topology Generation. Based on the three-dimensional distribution of the temperature field and the intensity level of the heat source, the partition boundaries are identified by isotherm extraction and temperature gradient calculation. The robot's interior is divided into three physical field partitions: a core high-heat zone, a medium-heat coupling zone, and a low-temperature steady-state zone, using a region growing clustering algorithm. The average temperature, average flow demand, average vibration intensity, and noise baseline value of each partition are calculated. The partition number and spatial range are associated to generate a multiphysics heterogeneous partition topology, providing a spatial basis for subsequent differentiated control.

[0009] Step 3: Construction of the five-dimensional spatiotemporal tensor and extraction of coupling factors. The spatiotemporal sequence data of temperature field, flow field, vibration field, noise field, and computing load field are organized into a fifth-order spatiotemporal tensor. The range normalization of each physical field data is performed, and the data is scaled to the 0-1 interval to eliminate the difference in dimensions. The CP tensor decomposition algorithm is used to iteratively solve the fifth-order tensor to obtain the spatial mode factor matrix, the temporal mode factor matrix, and the multi-physics coupling factor matrix. The nonlinear correlation weights between each physical field are extracted.

[0010] Step 4: Identification of Dominant Coupling Relationships. Based on the weight distribution of each component in the coupling factor matrix, the coupling mode is determined. A single field weight ratio > 50% indicates a single physical field dominant mode, a sum of two field weights > 80% indicates a dual physical field coupling dominant mode, and a balance of weights in three or more fields indicates a multi-field collaborative coupling mode. The frequency of occurrence of each mode is statistically analyzed, and the mode with the highest frequency is determined as the overall dominant coupling relationship of the system, which serves as the core basis for formulating control strategies.

[0011] Step 5: Noise Multi-Scale Decomposition and Evolution Prediction. A time series decomposition algorithm is used to separate noise and vibration signals into slow time-varying components and fast time-varying components. The slow time-varying components are caused by slow changes in temperature, load, and flow rate, while the fast time-varying components are caused by pump pulsation, fluid turbulence, and structural resonance. A linear regression mapping function is established between the physical field coupling factor, the mean temperature, the mean flow rate, and the slow time-varying noise components. At the same time, an autoregressive correlation model is established between the vibration spectrum characteristics and the fast time-varying noise components. Based on the above model, the noise evolution sequence of each partition within the next 10-30 second time window is predicted, thus completing the channel and noise state prediction.

[0012] Step Six: Generating a Quiet Quality Tag Matrix. Using a preset noise threshold as the criterion, each zone and time period is marked as a shallow quiet period and a deep high noise period. This generates a quiet quality tag matrix for each zone and time period, forming a global quiet quality map covering both spatial and temporal dimensions, providing a scheduling basis for data transmission and command issuance.

[0013] Step 7: Configure differentiated liquid cooling control parameters for each zone. With the optimization goals of achieving chip temperature control standards, minimizing operating noise, and optimizing system energy efficiency, we configure dedicated flow rates, pump speeds, valve openings, and cooling fan strategies for the core high-heat zone, medium-heat coupling zone, and low-temperature steady-state zone based on the heat load level, predicted channel status, and noise constraints of each zone. For extreme temperature zones below -20℃ or above 60℃, we introduce temperature correction factors to improve the heat dissipation power margin. At the same time, we constrain the pump speed within the quiet range to avoid structural resonance frequencies.

[0014] Step 8: Data and instruction collaborative scheduling and execution are scheduled based on the noise quality tag matrix. Multi-physics field monitoring data is uploaded first during shallow noise periods. Control instructions such as liquid cooling flow regulation, water pump speed regulation, valve opening and closing, and computing load limitation are issued in advance before the arrival of deep high noise periods. Non-critical adjustment actions are delayed during deep high noise periods, the speed of the actuator is reduced and the redundancy of instruction transmission is increased. Emergency temperature control instructions adopt multi-path redundant transmission to ensure highly reliable real-time issuance.

[0015] Step Nine: Online learning and model updates involve real-time collection of actual operating data for each zone, such as temperature control error, flow tracking deviation, noise pressure, data transmission success rate, and system energy consumption. The recursive least squares algorithm is used to update the physical field coupling and noise mapping function coefficients, and the gradient descent algorithm is used to update the vibration and noise correlation model parameters. A forgetting factor is introduced to achieve adaptive weighting of historical data. A small forgetting factor is used for the fast load change zone, and a large forgetting factor is used for the steady-state zone to ensure the long-term accuracy and stability of the model.

[0016] Preferably, in step two, the partition boundaries and thresholds are set as follows: the temperature gradient threshold is the global temperature gradient mean plus 1 standard deviation, and the temperature difference threshold within the partition is 2–5 times the sensor measurement accuracy; the core high-heat zone is the heterogeneous computing chip area, the medium thermal coupling zone is the power supply and drive module area, and the low-temperature steady-state zone is the pump, valve, pipeline, and heat sink area.

[0017] Preferably, the dimensions of the fifth-order tensor in step three are, in order, the number of spatial grid points, the number of time sampling points, and the number of physical field types. The physical field types include five types: temperature, flow rate, vibration, noise, and computing load. Normalization is used to eliminate dimensional differences and improve the accuracy of tensor decomposition and the reliability of coupling relationship identification.

[0018] Preferably, the coupling mode determination rule in step four is as follows: a single field weight ratio of >50% indicates a single physical field dominance; a sum of two field weights of >80% indicates dual-field coupling dominance; and a balance of three or more field weights indicates multi-field collaborative coupling. The mode with the highest frequency of occurrence is determined as the overall dominant coupling relationship.

[0019] Preferably, the slow time-varying component in step five represents the long-term impact of slow changes in temperature, load, and flow rate on noise, while the fast time-varying component represents the rapid fluctuations in noise caused by pump and valve pulsation, fluid turbulence, and structural resonance. The prediction time window is 10–30 seconds, which meets the requirements of real-time control and command scheduling.

[0020] Preferably, in step six, the noise quality label matrix uses 40dB or 50dB as the noise threshold. Noise ≤ the threshold is marked as a shallow quiet period and assigned a value of 1, and noise > the threshold is marked as a deep high-noise period and assigned a value of 0, thus forming a spatiotemporal noise quality map.

[0021] Preferably, the differentiated configuration in step seven includes: a priority heat dissipation strategy is adopted for the core high-heat zone and the pump speed is constrained to the quiet range; an on-demand heat dissipation strategy is adopted for the medium-heat coupling zone; a strategy of maintaining flow and intermittent operation is adopted for the low-temperature steady-state zone; and the pump speed avoids the resonant frequency point to suppress resonant noise.

[0022] Preferably, the emergency temperature control command in step eight adopts multi-path redundant transmission to ensure zero-delay and high-reliability command issuance under extreme working conditions, delay non-critical adjustment actions during periods of high noise, and reduce the speed of water pumps and valves to achieve silent operation.

[0023] A low-noise liquid cooling system for humanoid robots includes the following modules:

[0024] The multiphysics monitoring module is used to collect temperature, flow rate, vibration, noise, computing load and spatial coordinate data in real time, providing a precise data source for full-domain control and ensuring comprehensive and real-time monitoring.

[0025] The spatial partitioning module is used to generate a three-dimensional physical field distribution map and complete the heat source partitioning, realize spatial differential control, avoid global high power consumption and high noise, and improve temperature control and noise reduction efficiency.

[0026] The coupling analysis module is used to complete the fifth-order tensor decomposition and coupling factor extraction, accurately identify the nonlinear coupling relationship of multi-physics fields, and provide core basis for prediction and scheduling.

[0027] The noise prediction module is used for multi-scale noise decomposition and future noise state prediction, which can identify high noise and quiet periods in advance and achieve active noise avoidance.

[0028] The silent quality label generation module is used to generate a spatiotemporal silent quality label matrix to form a unified scheduling map, ensuring orderly and reliable data and instruction transmission.

[0029] The differentiated parameter configuration module is used to generate exclusive liquid cooling control parameters for different zones, achieving a balance between on-demand heat dissipation, noise constraint, and optimal energy efficiency.

[0030] The collaborative scheduling module is used to schedule data and instructions according to the noise quality label, avoiding high-noise periods and significantly reducing the overall operating noise of the machine.

[0031] The online learning module is used to collect operational data in real time and adaptively update model parameters, thereby improving the long-term stability, control accuracy, and environmental adaptability of the system.

[0032] Preferably, the liquid-cooled actuator adopts an integrated cold head with flux-free vacuum welding, molecular-level self-healing coolant, and microchannel heat dissipation structure, meeting the requirements of ≥250W heat dissipation capacity, IP68 full sealing, operating noise <40dB, zero frequency reduction at 60℃, and compatibility with NVIDIA Orin / Thor, Ambarella CV3 and domestic heterogeneous computing chips.

[0033] The present invention has the following beneficial effects:

[0034] 1. By constructing multidimensional spatiotemporal tensors of temperature field, flow field, vibration field, noise field, and computing load field, and using tensor decomposition algorithm to accurately extract multi-physics coupling factors, it is possible to clearly identify the nonlinear correlation and influence law between various physical quantities. From the perspective of control mechanism, it breaks through the limitation of traditional single PID feedback control that only relies on temperature signal, effectively suppresses the fluctuations and interference caused by multi-physics coupling, and makes the liquid cooling heat dissipation control response faster, the steady-state error smaller, and the system stability stronger.

[0035] 2. By using regional growth clustering and temperature gradient partitioning, the robot's interior is divided into a core high-heat zone, a medium-heat coupling zone, and a low-temperature steady-state zone. Flow rate, speed, and valve strategy control parameters are independently configured for different zones based on their heat source intensity, noise constraints, and heat dissipation requirements. This avoids resource waste, local overheating, or excessive noise caused by global unified control. While ensuring efficient heat dissipation of the core chip, it significantly reduces overall energy consumption and ineffective operation, achieving optimal synergy between heat dissipation, noise reduction, and energy saving.

[0036] 3. By decomposing noise at multiple scales and predicting trends, we can identify periods of high noise and periods of shallow silence in advance. Based on the silence quality label matrix, we can schedule and control the data transmission and command execution during the silence period. We can complete key adjustments before the high noise period arrives and actively avoid pump and valve resonance and fluid turbulence noise peaks from the control timing. Compared with traditional passive noise reduction methods, the control is smoother, the noise fluctuation is smaller, and the silence effect is more stable. Attached Figure Description

[0037] Figure 1 This is a module logic flowchart of a low-noise liquid cooling system for a humanoid robot proposed in this invention.

[0038] Figure 2 This is a diagram showing the operating data of the heat dissipation system in this invention. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0040] Reference Figure 1 A low-noise liquid cooling system for humanoid robots includes a multiphysics monitoring module, a spatial partitioning module, a coupling analysis module, a noise prediction module, a silent quality label generation module, a differentiated parameter configuration module, a collaborative scheduling module, an online learning module, and a liquid cooling actuator.

[0041] Multiphysics monitoring module: In addition to collecting basic physical quantities, it also performs filtering, noise reduction, and outlier removal on the signals to ensure stable and reliable input data and provide a high-quality data source for subsequent algorithms.

[0042] Spatial partitioning module: After completing the three-dimensional field interpolation, it will automatically align with the boundary of the robot's thoracic cavity structure to avoid partition overstepping and ensure that each heat-generating component belongs to a unique physical partition, thereby improving the targeting of control.

[0043] The coupling analysis module is responsible for unifying and structuring multi-source heterogeneous data. Through tensor decomposition, it simplifies complex coupling relationships into computable and schedulable coupling factors, enabling the system to understand the intrinsic influence patterns between "temperature, flow rate, vibration, noise, and load".

[0044] Noise prediction module: The prediction results are continuously updated using a rolling time window to ensure that the noise prediction always follows the current operating conditions and will not lag due to sudden load changes.

[0045] The noise quality label generation module refreshes the label matrix in real time, ensuring that scheduling instructions are always based on the latest noise status, thus achieving dynamic noise scheduling.

[0046] Differentiated parameter configuration module: Supports online adaptive fine-tuning of parameters. Without changing the core strategy, it slightly adjusts the flow rate and speed according to the real-time temperature, balancing stability and sensitivity.

[0047] Cooperative scheduling module: Built-in priority mechanism to ensure that emergency control commands are executed first, and ordinary monitoring data are queued in an orderly manner to avoid channel conflicts.

[0048] Online learning module: It has an error accumulation judgment function. When the prediction error continuously exceeds the threshold, it automatically accelerates the update to ensure that the model converges quickly.

[0049] Liquid-cooled actuator: It adopts an integrated fluxless vacuum-welded cold head and microchannel structure, combined with molecular-level self-healing coolant, to achieve a heat dissipation capacity of ≥250W in a small space of ≤2L. It is IP68 fully sealed, with an operating noise of less than 40dB. The chip maintains zero frequency reduction under a high temperature of 60℃, and is compatible with NVIDIA Orin / Thor, Ambarella CV3 and domestic high-computing power chips.

[0050] Control methods:

[0051] Reference Figure 1 and Figure 2 Step 1: Multiphysics Data Acquisition and 3D Field Construction

[0052] After the system is powered on and initialized, the multiphysics monitoring module collects real-time operating data of the entire internal cavity of the humanoid robot at a fixed sampling rate of 10Hz. This includes the junction temperature of the computing chip, the surface temperature of the cold plate, the inlet and outlet temperatures of the coolant, the real-time circulation flow rate, the vibration amplitude and spectral characteristics of the water pump and pipeline, the sound pressure level of the whole machine, the real-time power consumption and load rate of the heterogeneous computing chip, and simultaneously records the spatial coordinates of each sensor.

[0053] To avoid interference signals affecting subsequent calculations, the module first performs moving average filtering and outlier removal on the raw data. Then, the spatial partitioning module calls the Kriging spatial interpolation algorithm to perform global spatial fitting on the discrete monitoring points.

[0054]

[0055] The estimated values ​​of the physical quantities at the points to be interpolated. The interpolation weights are strictly equal to 1. Through the interpolation algorithm, scattered measurement points are expanded into a continuously distributed global data field, ultimately generating a three-dimensional spatial distribution map of temperature field, flow field, vibration field, and noise field, which fully restores the real spatial distribution of heat, flow, vibration, and noise inside the robot.

[0056] Step 2: Multiphysics Spatial Partitioning and Topology Generation

[0057] The spatial partitioning module automatically extracts isotherms based on the three-dimensional temperature field distribution and calculates the spatial temperature gradient point by point to determine the degree of temperature change. The system sets the temperature gradient threshold to the global temperature gradient mean plus one standard deviation, and the temperature difference threshold within a partition to be 2–5 times the measurement accuracy of the temperature sensor.

[0058] After the boundary conditions are determined, a region growing clustering algorithm is used to gradually expand from the seed point, automatically dividing the internal space of the thoracic cavity into three non-overlapping regions:

[0059] Core high-heat zone: This corresponds to the location of the heterogeneous computing chip, where the heat flux density is the highest and the temperature changes most drastically.

[0060] Mid-thermal coupling zone: corresponds to secondary heat-generating components such as power management chips, driver modules, and interface circuits;

[0061] Low-temperature steady-state region: corresponds to actuators such as water pumps, proportional valves, radiators, and cooling pipes.

[0062] After partitioning is completed, the module automatically calculates the average temperature, average flow demand, average vibration intensity, and average noise baseline value for each partition, and associates the partition number, spatial range, and physical field statistics to generate a stable multi-physics heterogeneous partition topology, providing an accurate spatial basis for subsequent partition-based differentiated control.

[0063] Step 3: Construction of Multidimensional Spatiotemporal Tensors and Extraction of Coupling Factors

[0064] The coupling analysis module organizes temperature, flow rate, vibration, noise, and computing load data over a continuous 15-second period into a structured multidimensional spatiotemporal tensor, enabling a unified expression of the temporal and spatial information of multiphysics fields.

[0065] To eliminate calculation biases caused by differences in dimensions such as temperature, flow rate, power consumption, and vibration amplitude, the module first performs range normalization on all data:

[0066]

[0067] After normalization, all data are scaled to the 0–1 range. Subsequently, the CP tensor decomposition algorithm is used to iteratively solve the tensor, and the alternating least squares method is used to continuously optimize it. Finally, the spatial mode factor matrix, the temporal mode factor matrix, and the multiphysics coupling factor matrix are obtained, which accurately extracts the nonlinear correlation weights between temperature and flow rate, flow rate and vibration, vibration and noise, and load and temperature.

[0068] Step 4: Identification of Multiphysics Dominant Coupling Relationships

[0069] The coupling analysis module performs weight percentage statistics and pattern determination for each column of the coupling factor matrix: if the weight percentage of a certain type of physical field is >50%, it is determined to be dominated by a single physical field; if the sum of the weights of any two types of physical fields is >80%, it is determined to be dominated by dual-field coupling; if the weight distribution of three or more types is balanced, it is determined to be multi-field synergistic coupling.

[0070] The module counts the frequency of coupling patterns for all decomposed components and identifies the most frequent pattern as the dominant coupling relationship in the system. In this embodiment, temperature, flow rate, and computing load exhibit the strongest correlation and are thus identified as the dominant coupling relationship in the system. Subsequent optimizations of heat dissipation flow rate, pump speed, and scheduling strategies are all based on this dominant relationship.

[0071] Step 5: Noise Multi-scale Decomposition and Spatiotemporal Evolution Prediction

[0072] The noise prediction module decomposes the noise time-series signal into slow time-varying components and fast time-varying components:

[0073]

[0074] Among them, the slow time-varying component is driven by the slow changes in temperature, load, and flow rate, with gentle changes and clear trends; the fast time-varying component is caused by pump pulsation, fluid turbulence, and structural resonance, with high fluctuation frequency and drastic changes.

[0075] The module establishes prediction models: on the one hand, it uses the physical field coupling factor, average temperature of each zone, and average flow rate as inputs to establish a mapping relationship with the slow time-varying components of noise; on the other hand, it uses vibration spectrum characteristics as inputs to establish an autoregressive correlation model with the fast time-varying components of noise. Based on a rolling time window, the module continuously predicts the noise change trend of each zone within the next 10–30 seconds, identifying periods of deep high noise and shallow quiet periods in advance, providing predictive support for quiet scheduling.

[0076] Step Six: Generating the Silent Quality Tag Matrix

[0077] The noise quality label generation module uses 40dB as the noise judgment threshold and performs binary labeling on each partition and each time window: noise ≤ 40dB is marked as a shallow quiet period and assigned a value of 1; noise > 40dB is marked as a deep high noise period and assigned a value of 0.

[0078] The system generates a partition and time period quiet quality label matrix based on the number of partitions and time windows, and updates it in real time at a fixed frequency to form a full-domain quiet scheduling map covering spatial and temporal dimensions, providing unified, clear and executable scheduling rules for monitoring data uploading and control command issuance.

[0079] Step 7: Configure Zoned Differentiated Liquid Cooling Control Parameters

[0080] The differentiated parameter configuration module aims to achieve stable chip temperature control, the lowest overall noise, and the best system energy efficiency. It combines the thermal load levels of each zone, predicted noise levels, and heat dissipation requirements to configure dedicated control strategies for the three types of zones:

[0081] Core high-heat zone: Prioritize heat dissipation capacity, set the flow rate to the upper limit of demand, strictly limit the water pump speed to the quiet range, and actively avoid structural resonance frequency points;

[0082] Medium thermal coupling zone: Adopts on-demand flow distribution strategy to operate stably at medium and low speeds, balancing heat dissipation efficiency and noise control;

[0083] Low-temperature steady-state zone: Only the minimum circulation flow is maintained, and the water pump and valve can operate in an intermittent mode to further reduce noise and energy consumption.

[0084] For extreme operating conditions where the chip temperature is above 60℃ or the ambient temperature is below -20℃, the system automatically introduces a temperature correction factor to appropriately increase the heat dissipation power margin, ensuring that the chip operates stably across the entire temperature range without triggering frequency reduction or thermal throttling.

[0085] Step 8: Coordinated scheduling and execution of monitoring data and control commands

[0086] The collaborative scheduling module executes strict scheduling logic based on the silence quality label matrix:

[0087] Routine monitoring data such as temperature, flow rate, vibration, and noise are uploaded in an orderly manner only during periods of slight silence (marked as 1);

[0088] Control commands such as flow regulation, pump speed adjustment, valve switching, and computing load limiting must be issued before the arrival of the period of deep high noise.

[0089] Once the system enters a period of high noise (marked as 0), it automatically delays non-critical adjustment actions, reduces the speed of water pumps and valves, and increases the redundancy of command transmission.

[0090] For emergency temperature control commands such as chip overheating, abnormal flow, and coolant leakage, the system immediately activates multi-path redundant transmission, skipping the regular scheduling queue to ensure that the command is issued with zero delay and high reliability.

[0091] Step Nine: Collection of Operational Indicators and Online Model Updates

[0092] The online learning module collects key system operation indicators in real time, including temperature control error, flow tracking deviation, real-time noise pressure, data transmission success rate, total system energy consumption, and cumulative pump and valve runtime.

[0093] The module employs a recursive least squares algorithm to update the coefficients of the physical field coupling and noise mapping models, and a gradient descent algorithm to update the parameters of the vibration-noise correlation model, continuously reducing the prediction error. Simultaneously, a forgetting factor is introduced to adaptively adjust the weights of historical data: a smaller forgetting factor is used in regions of rapid load change to improve state tracking speed; a larger forgetting factor is used in steady-state operation regions to improve control stability and smoothness.

[0094] Through continuous online learning and iterative updates, the system can automatically adapt to various scenarios such as sudden changes in computing load, component aging, and environmental migration, maintaining high-precision temperature control, low-noise operation, and highly reliable heat dissipation throughout the humanoid robot's 5-10 year lifespan.

[0095] In summary, the system supports stable heat dissipation for high-power computing chips of ≥250W, maintaining zero frequency throttling even at 60℃. The overall steady-state operating noise is ≤38dB, meeting the requirements for quiet operation in home, medical, and service settings. Chip temperature fluctuations are controlled within ±2℃, with no temperature oscillations or frequent protection restarts. The system has IP68 fully sealed protection capabilities, enabling long-term stable operation in dusty, high-humidity, and corrosive environments. The same hardware platform is seamlessly compatible with multiple mainstream heterogeneous computing chips, significantly shortening the product development cycle for customers. The fluxless vacuum liquid cooling structure has an extremely low leakage rate, enabling 5–10 years of maintenance-free operation.

[0096] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A control method for a low-noise liquid cooling system for a humanoid robot, characterized in that, Includes the following steps: Step 1: Multiphysics Data Acquisition and 3D Field Construction. Acquire data on the temperature, coolant inlet and outlet temperatures, real-time flow rate, vibration amplitude and spectrum of pumps, valves and pipelines, noise sound pressure level, and chip load power consumption of the heterogeneous computing chip, drive module, pumps, valves, pipelines and radiator of the humanoid robot's chest cavity and whole body. Simultaneously collect the spatial coordinate information corresponding to each sensor. Use the Kriging spatial interpolation algorithm to perform spatial fitting on discrete monitoring points to generate a 3D spatial distribution map of temperature field, flow field, vibration field and noise field, which completely restores the heat, flow, vibration and noise distribution state of the robot's internal whole domain. Step 2: Multiphysics Spatial Partitioning and Topology Generation. Based on the three-dimensional distribution of the temperature field and the intensity level of the heat source, the partition boundaries are identified by isotherm extraction and temperature gradient calculation. The robot's interior is divided into three physical field partitions: a core high-heat zone, a medium-heat coupling zone, and a low-temperature steady-state zone, using a region growing clustering algorithm. The average temperature, average flow demand, average vibration intensity, and noise baseline value of each partition are calculated. The partition number and spatial range are associated to generate a multiphysics heterogeneous partition topology, providing a spatial basis for subsequent differentiated control. Step 3: Construction of the five-dimensional spatiotemporal tensor and extraction of coupling factors. The spatiotemporal sequence data of temperature field, flow field, vibration field, noise field, and computing load field are organized into a fifth-order spatiotemporal tensor. The range normalization of each physical field data is performed, and the data is scaled to the 0-1 interval to eliminate the difference in dimensions. The CP tensor decomposition algorithm is used to iteratively solve the fifth-order tensor to obtain the spatial mode factor matrix, the temporal mode factor matrix, and the multi-physics coupling factor matrix. The nonlinear correlation weights between each physical field are extracted. Step 4: Identification of Dominant Coupling Relationships. Based on the weight distribution of each component in the coupling factor matrix, the coupling mode is determined. A single field weight ratio > 50% indicates a single physical field dominant mode, a sum of two field weights > 80% indicates a dual physical field coupling dominant mode, and a balance of weights in three or more fields indicates a multi-field collaborative coupling mode. The frequency of occurrence of each mode is statistically analyzed, and the mode with the highest frequency is determined as the overall dominant coupling relationship of the system, which serves as the core basis for formulating control strategies. Step 5: Noise Multi-Scale Decomposition and Evolution Prediction. A time series decomposition algorithm is used to separate noise and vibration signals into slow time-varying components and fast time-varying components. The slow time-varying components are caused by slow changes in temperature, load, and flow rate, while the fast time-varying components are caused by pump pulsation, fluid turbulence, and structural resonance. A linear regression mapping function is established between the physical field coupling factor, the mean temperature, the mean flow rate, and the slow time-varying noise components. At the same time, an autoregressive correlation model is established between the vibration spectrum characteristics and the fast time-varying noise components. Based on the above model, the noise evolution sequence of each partition within the next 10-30 second time window is predicted, thus completing the channel and noise state prediction. Step Six: Generating a Quiet Quality Tag Matrix. Using a preset noise threshold as the criterion, each zone and time period is marked as a shallow quiet period and a deep high noise period. This generates a quiet quality tag matrix for each zone and time period, forming a global quiet quality map covering both spatial and temporal dimensions, providing a scheduling basis for data transmission and command issuance. Step 7: Configure differentiated liquid cooling control parameters for each zone. With the optimization goals of achieving chip temperature control standards, minimizing operating noise, and optimizing system energy efficiency, we configure dedicated flow rates, pump speeds, valve openings, and cooling fan strategies for the core high-heat zone, medium-heat coupling zone, and low-temperature steady-state zone based on the heat load level, predicted channel status, and noise constraints of each zone. For extreme temperature zones below -20℃ or above 60℃, we introduce temperature correction factors to improve the heat dissipation power margin. At the same time, we constrain the pump speed within the quiet range to avoid structural resonance frequencies. Step 8: Data and instruction collaborative scheduling and execution are scheduled based on the noise quality tag matrix. Multi-physics field monitoring data is uploaded first during shallow noise periods. Control instructions such as liquid cooling flow regulation, water pump speed regulation, valve opening and closing, and computing load limitation are issued in advance before the arrival of deep high noise periods. Non-critical adjustment actions are delayed during deep high noise periods, the speed of the actuator is reduced and the redundancy of instruction transmission is increased. Emergency temperature control instructions adopt multi-path redundant transmission to ensure highly reliable real-time issuance. Step Nine: Online learning and model updates involve real-time collection of actual operating data for each zone, such as temperature control error, flow tracking deviation, noise pressure, data transmission success rate, and system energy consumption. The recursive least squares algorithm is used to update the physical field coupling and noise mapping function coefficients, and the gradient descent algorithm is used to update the vibration and noise correlation model parameters. A forgetting factor is introduced to achieve adaptive weighting of historical data. A small forgetting factor is used for the fast load change zone, and a large forgetting factor is used for the steady-state zone to ensure the long-term accuracy and stability of the model.

2. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, In step two, the partition boundaries and thresholds are set as follows: the temperature gradient threshold is the global temperature gradient mean plus 1 standard deviation, and the temperature difference threshold within the partition is 2–5 times the sensor measurement accuracy; the core high-heat zone is the heterogeneous computing chip area, the medium thermal coupling zone is the power supply and drive module area, and the low-temperature steady-state zone is the pump, valve, pipeline, and heat sink area.

3. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, In step three, the dimensions of the fifth-order tensor are, in order, the number of spatial grid points, the number of time sampling points, and the number of physical field types. The physical field types include five types: temperature, flow rate, vibration, noise, and computing load. Normalization is used to eliminate dimensional differences and improve the accuracy of tensor decomposition and the reliability of coupling relationship identification.

4. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, The coupling mode determination rule in step four is as follows: a single field weight ratio of >50% indicates a single physical field dominance; the sum of the weights of two fields >80% indicates dual-field coupling dominance; and a balance of weights in three or more fields indicates multi-field collaborative coupling. The mode with the highest frequency of occurrence is determined as the overall dominant coupling relationship.

5. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, The slow time-varying component in step five represents the long-term impact of slow changes in temperature, load, and flow rate on noise, while the fast time-varying component represents the rapid fluctuations in noise caused by pump and valve pulsation, fluid turbulence, and structural resonance. The prediction time window is 10–30 seconds, which meets the requirements of real-time control and command scheduling.

6. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, In step six, the noise quality label matrix uses 40dB or 50dB as the noise threshold. Noise ≤ the threshold is marked as a shallow quiet period and assigned a value of 1, and noise > the threshold is marked as a deep high noise period and assigned a value of 0, thus forming a spatiotemporal noise quality map.

7. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, The differentiated configuration in step seven includes: a priority heat dissipation strategy is adopted for the core high-heat zone and the pump speed is constrained to the quiet range; an on-demand heat dissipation strategy is adopted for the medium thermal coupling zone; a strategy of maintaining flow and intermittent operation is adopted for the low-temperature steady-state zone; and the pump speed avoids the resonant frequency point to suppress resonant noise.

8. The control method for a low-noise liquid cooling system for a humanoid robot according to claim 1, characterized in that, The emergency temperature control command in step eight adopts multi-path redundant transmission to ensure zero-delay and high-reliability command issuance under extreme working conditions, delays non-critical adjustment actions during periods of high noise, and reduces the speed of water pumps and valves to achieve silent operation.

9. A liquid cooling system for executing the control method of a low-noise liquid cooling system for a humanoid robot according to any one of claims 1-8, characterized in that, Includes the following modules: The multiphysics monitoring module is used to collect temperature, flow rate, vibration, noise, computing load and spatial coordinate data in real time, providing a precise data source for full-domain control and ensuring comprehensive and real-time monitoring. The spatial partitioning module is used to generate a three-dimensional physical field distribution map and complete the heat source partitioning, realize spatial differential control, avoid global high power consumption and high noise, and improve temperature control and noise reduction efficiency. The coupling analysis module is used to complete the fifth-order tensor decomposition and coupling factor extraction, accurately identify the nonlinear coupling relationship of multi-physics fields, and provide core basis for prediction and scheduling. The noise prediction module is used for multi-scale noise decomposition and future noise state prediction, which can identify high noise and quiet periods in advance and achieve active noise avoidance. The silent quality label generation module is used to generate a spatiotemporal silent quality label matrix to form a unified scheduling map, ensuring orderly and reliable data and instruction transmission. The differentiated parameter configuration module is used to generate exclusive liquid cooling control parameters for different zones, achieving a balance between on-demand heat dissipation, quiet operation, and optimal energy efficiency. The collaborative scheduling module is used to schedule data and instructions according to the noise quality label, avoiding high-noise periods and significantly reducing the overall operating noise of the machine. The online learning module is used to collect operational data in real time and adaptively update model parameters, thereby improving the long-term stability, control accuracy, and environmental adaptability of the system.

10. A low-noise liquid cooling system for a humanoid robot according to claim 9, characterized in that, The liquid-cooled actuator adopts an integrated cold head with flux-free vacuum welding, molecular-level self-healing coolant, and microchannel heat dissipation structure, meeting the requirements of ≥250W heat dissipation capacity, IP68 full sealing, operating noise <40dB, zero frequency reduction at 60℃, and is compatible with NVIDIA Orin / Thor, Ambarella CV3 and domestic heterogeneous computing chips.

Images