Cold and heat exchange system of liquid cooling cabinet and use method of cold and heat exchange system

Abstract

The invention relates to the technical field of manufacturing of electrical element assemblies, and discloses a cold and heat exchange system of a liquid cooling cabinet and a use method of the cold and heat exchange system. The cold and heat exchange method of the liquid cooling cabinet comprises the steps that a complex thermodynamic model of the liquid cooling cabinet is processed through a hierarchical model distillation algorithm, and an ultra-lightweight thermodynamic digital twin model is obtained; processing the operation state of the liquid cooling cabinet through a multi-mode micro-sensor network to obtain a lightweight environment and load feature vector; processing the lightweight environment and load feature vector and the ultra-lightweight thermodynamic digital twin model through a micro predictive controller to obtain an adaptive liquid cooling control strategy; processing the self-adaptive liquid cooling control strategy through an ultra-low power consumption execution unit to obtain an accurately executed physical cooling effect; according to the invention, through an optimization algorithm, efficient control of the liquid cooling cabinet system is realized in a resource limited environment, the precision, the response speed and the energy efficiency are improved, the environmental adaptability and reliability are enhanced, and the bottleneck of a traditional system is broken through.

Description

Technical Field

[0001] The present invention relates to the technical field of manufacturing electrical component assemblies, and more particularly to a heat exchange system for a liquid cooling cabinet and a method for using the system. Background Art

[0002] With the development of technology, traditional liquid-cooled cabinet systems have many shortcomings in scenarios with limited resources, changing environments, and difficult maintenance. They face computing power and deployment challenges, and the thermodynamic model is too large to run on edge controllers. They have slow response speeds, PID controller load response lags, and large temperature fluctuations. They have poor environmental adaptability, only adapting to -5°C to 40°C and -12% to +12% voltage fluctuations. They are expensive, require the installation of a large number of sensors, and are complex to maintain. They also have high energy consumption, with the execution unit power consumption accounting for 8% to 12% of the system and the control system power consumption reaching 10.5%, resulting in low energy efficiency. Summary of the Invention

[0003] The present invention provides a heat exchange system for a liquid-cooled cabinet and a method for using the same, to solve technical problems in related technologies such as insufficient computing power, delayed response, poor environmental adaptability, and high energy consumption of traditional liquid-cooled control systems under resource-constrained and harsh environmental scenarios.

[0004] The present invention provides a heat exchange method for a liquid cooling cabinet, comprising: A hierarchical model distillation algorithm is used to process the complex thermodynamic model of the liquid cooling cabinet to obtain an ultra-lightweight thermodynamic digital twin model. The operating status of the liquid cooling cabinet is processed through a multimodal micro-sensor network to obtain lightweight environment and load feature vectors; An adaptive liquid cooling control strategy is obtained by processing lightweight environment and load feature vectors and ultra-lightweight thermodynamic digital twin models through a micro-predictive controller. The ultra-low power execution unit processes the adaptive liquid cooling control strategy to achieve precise physical cooling effects.

[0005] Furthermore, the calculation formula of the feature space dimensionality reduction algorithm is: ; in, represents the original thermodynamic feature space, is the number of sampling points, is the original feature dimension; represents the feature space after dimensionality reduction, is the feature dimension after dimensionality reduction; represents a nonlinear dimensionality reduction function; represents the locally linear embedding algorithm; represents kernel principal component analysis; Indicates the threshold for retaining information.

[0006] Furthermore, the calculation formula of the knowledge distillation compression network is: ; in, represents the knowledge distillation loss function; represents the cross entropy loss function; represents the KL divergence loss function; represents the true label; represents the output of the teacher network; represents the output of the student network; represents the softmax function; represents the temperature parameter; Represents the balance factor.

[0007] Furthermore, the calculation formula of the load feature extraction algorithm is: ; in, Indicates the time at which the device Power consumption data; represents the extracted load feature vector; represents the time series feature extraction function; represents the statistical feature extraction function; represents the frequency domain feature extraction function; represents the trend feature extraction function; Indicates the time window parameter for trend extraction; is the time window.

[0008] Furthermore, the micro-predictive controller includes: a feedforward neural network predictor, used to process lightweight environment and load characteristic vectors to obtain a thermal load prediction sequence; a lightweight model predictive control algorithm, used to process the thermal load prediction sequence and the ultra-lightweight thermodynamic digital twin model to obtain a multi-step control sequence; a control strategy optimization algorithm, used to process the multi-step control sequence to obtain an adaptive liquid cooling control strategy.

[0009] Furthermore, the calculation formula of the lightweight model predictive control algorithm is:

[0010] ; ; ; in, represents the optimized control sequence; Indicates at time Predicted Temperature at all times; Indicates the target temperature; Indicates at time planned Control input at all times; Indicates at time Predicted Thermal load at all times; represents the environmental feature vector; Represents an ultra-lightweight thermodynamic digital twin model; and Represent the weighted matrices and The quadratic form of represents the length of the predictive control time domain; and Represents the upper and lower limit constraints of the control input; and Indicates the upper and lower limit constraints of temperature; It means that under the conditions of satisfying the upper and lower limit constraints of temperature and control input, as well as the temperature prediction model, a suitable control input sequence is found. , so that the sum of the deviation between the predicted temperature and the target temperature and the change amplitude of the control input is minimized, thereby obtaining the optimized control sequence .

[0011] Furthermore, the ultra-low power execution unit includes: a control signal quantization algorithm for processing the adaptive liquid cooling control strategy to obtain a digital control signal; a pulse width modulation conversion algorithm for processing the digital control signal to obtain a PWM drive signal; a multi-layer safety protector for processing the PWM control signal and system status information to obtain a safety protection control signal; and a low-power drive circuit for processing the safety protection control signal to obtain a precise physical cooling effect.

[0012] The present invention provides a heat exchange system for a liquid cooling cabinet, comprising: A hierarchical model distillation module is used to process complex thermodynamic models of liquid-cooled cabinets and obtain ultra-lightweight thermodynamic digital twin models; A multimodal micro-sensor network module is used to process the operating status of the liquid cooling cabinet and obtain lightweight environment and load feature vectors; A micro-predictive controller module is used to process lightweight environmental and load characteristic vectors and ultra-lightweight thermodynamic digital twin models to obtain adaptive liquid cooling control strategies; Ultra-low power execution unit module, used to process adaptive liquid cooling control strategy to obtain precise execution of physical cooling effect.

[0013] Furthermore, the multimodal micro-sensor network module includes: a sparse sensor sampling optimization unit for processing temperature field data to obtain key point temperature distribution characteristics; a load feature extraction unit for processing computing device power consumption data to obtain computing load feature vectors; an environmental adaptability perception unit for processing external environmental parameters to obtain environmental feature vectors; and a feature fusion compression unit for processing multi-source sensor data to obtain lightweight environment and load feature vectors.

[0014] Furthermore, the multi-step control sequence optimization of the micro-predictive controller module adopts a control strategy optimization algorithm, and its calculation formula is: ; ; ; in, Represents a multi-step control sequence; Represents the optimized control strategy at the current moment; Represents the first control input in the original multi-step control sequence; represents the control correction item based on energy efficiency considerations; represents the control strategy optimization function; represents the energy efficiency-cooling balance factor; Indicates the current system energy status; Indicates the current temperature; Indicates the temperature threshold; represents the energy efficiency evaluation function; Indicates safety margin; represents the energy status scoring function.

[0015] The beneficial effects of the present invention are: through a series of algorithms and optimization designs, the present invention realizes efficient control of the liquid cooling cabinet heat exchange system under extremely low computing resource conditions in scenarios with limited resources and harsh environments, significantly improves control accuracy and response speed, optimizes energy efficiency, enhances environmental adaptability and reliability, and solves the application bottleneck of traditional liquid cooling control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a heat exchange method for a liquid cooling cabinet of the present invention. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses a method for heat exchange of a liquid cooling cabinet, such as Figure 1 Shown, including: Step 100: Process the complex thermodynamic model of the liquid cooling cabinet using a hierarchical model distillation algorithm to obtain an ultra-lightweight thermodynamic digital twin model; This step aims to compress the original thermodynamic model of the traditional complex liquid-cooled cabinet that requires a server to run into an ultra-lightweight thermodynamic digital twin model that can run on the cabinet edge control unit while maintaining high prediction accuracy.

[0019] Sub-step 101: Processing the complete thermodynamic simulation data set using a feature space dimensionality reduction algorithm to obtain a reduced-dimensionality thermal feature map; This sub-step uses a method that combines principal component analysis with nonlinear manifold learning to compress the high-dimensional thermodynamic feature space to an extremely low dimension. The calculation formula is: ; in: represents the original thermodynamic feature space, is the number of sampling points, is the original feature dimension (usually hundreds to thousands of dimensions), is the field of real numbers; represents the feature space after dimensionality reduction, is the feature dimension after dimensionality reduction (usually 5-10 dimensions); represents a nonlinear dimensionality reduction function; represents a locally linear embedding algorithm used to capture the local topological structure of the data; represents kernel principal component analysis, which is used to further extract nonlinear features; Indicates the threshold for retaining information, usually set to 0.95; Through this algorithm, the thousands of dimensions of features of the original thermodynamic model can be compressed into a low-dimensional representation of no more than 10 dimensions, while retaining more than 95% of the information.

[0020] Sub-step 102: Processing the dimension-reduced hot feature map through the knowledge distillation compression network to obtain an edge device deployment model; This sub-step uses the knowledge distillation technology of the teacher-student network architecture to transfer the knowledge of the large and complex thermodynamic model (teacher network) to a very small neural network (student network). The calculation formula is: ; in: is the knowledge distillation loss function; is the cross entropy loss function, which is used to measure the true label Output with student network (ultra-lightweight model) After the softmax function Differences after treatment; is the KL divergence loss function, which is used to measure the output of the teacher network (complete thermodynamic model) Network output with students After a specific temperature parameter Differences in probability distributions after softening the probability distributions; represents the true label; represents the output of the teacher network (full thermodynamic model); represents the output of the student network (super lightweight model); represents the softmax function; represents the temperature parameter, which is used to soften the probability distribution; Represents the balance factor, which controls the weight of hard labels and soft labels.

[0021] The distilled student network structure uses a special "MicroThermalNet" that includes: Input layer: receives the reduced hot feature vector (5-10 dimensions); Hidden layer: up to 2 layers, each layer has no more than 16 neurons, and uses ReLU activation function; Output layer: predicts temperature distribution and fluid parameters (5-8 dimensional output).

[0022] The total number of parameters after model quantization is controlled within 5KB and can be directly deployed on edge computing devices such as MCU.

[0023] Sub-step 103: Processing the edge device deployment model through parameter quantization and sparsification algorithms to obtain an ultra-lightweight thermodynamic digital twin model; This sub-step uses a method that combines mixed precision quantization and weight pruning to further compress the model volume and computational complexity. The calculation formula is: ; is a constraint condition; in: Represents a variable and Find the minimum value; represents the original floating-point weight matrix; Represents the quantized three-value weight matrix (values are -1, 0, 1); represents the weight scaling factor; is the square of the Euclidean norm, which is used to measure the original weight matrix and the scaled quantized weight matrix The purpose of this part is to keep the numerical similarity between the quantized weight matrix and the original weight matrix as much as possible; Represents the regularization coefficient, which is used to control the sparsity of the model; express The L0 norm, that is, the number of non-zero elements, is minimized, which causes more weights to become 0, thereby achieving sparseness of the model; Indicates the quantized weight matrix, is a The matrix element value can only be -1, 0 or 1; Represents the weight scaling factor is a positive real number.

[0024] In summary, this formula satisfies and Under the constraints of and The value of is minimized to minimize the difference between the original weight matrix and the quantized weight matrix and the number of non-zero elements in the quantized weight matrix, so as to achieve the purpose of compressing the model volume and computational complexity.

[0025] The resulting ultra-lightweight thermodynamic digital twin model has the following features: weight parameters are represented using 2 bits (-1, 0, 1), which is 16 times more compressed than 32-bit floating-point numbers; the model sparsity exceeds 80%, meaning that over 80% of the weights are 0; activation functions are implemented using a lookup table (LUT) to avoid complex mathematical operations; and all computational operations are optimized to fixed-point integer arithmetic, eliminating the need for floating-point units. This model utilizes knowledge distillation and quantization compression techniques to condense the complex original thermodynamic model into an ultra-lightweight version that can run efficiently on resource-constrained edge devices.

[0026] Step 200: Processing the edge device operating status through a multimodal micro-sensor network to obtain lightweight environment and load feature vectors; This step builds an efficient micro-sensor network to collect multi-dimensional operating data from edge devices and perform lightweight processing to provide input for subsequent control decisions.

[0027] Sub-step 201: Processing temperature field data using a sparse sensor sampling optimization algorithm to obtain key point temperature distribution characteristics; This sub-step uses the sensor optimization layout technology based on the heat conduction model to achieve the maximum temperature field reconstruction accuracy with the least number of sensor collection points. The calculation formula is: ; in: represents the set of all possible sensor locations; represents the optimized sensor layout scheme; Indicates the upper limit of the number of sensors allowed (usually 3-8); represents the complete temperature field; Represents a location-based collection Temperature field reconstructed at sampling points; represents the Frobenius norm; Indicates satisfaction yes A subset of The number of elements is Under the condition of , find the solution that minimizes the subsequent function value. ; It's about Here is the function of , that is, the complete temperature field With location-based collections Temperature field reconstructed at sampling points The square of the Frobenius norm between is used to measure the difference between the two. The formula refers to the set of all possible sensor positions In , the number of elements selected is Subset of , so that the complete temperature field Based on this subset Temperature field reconstructed at sampling points The difference between them (measured by the square of the Frobenius norm) is the smallest, and the subset obtained at this time This is the optimized sensor layout solution .

[0028] In actual deployment, a "layered temperature monitoring strategy" is adopted, including: Active area: near the computing unit, with high sampling density and sampling frequency of 10-50Hz; Secondary area: near the power supply and storage unit, with medium sampling density and sampling frequency of 1-5Hz; Passive area: edge area, low sampling density, sampling frequency is 0.1-1Hz; Through adaptive sampling frequency control, system power consumption is minimized while ensuring temperature measurement accuracy.

[0029] Sub-step 202: Processing computing device power consumption data using a load feature extraction algorithm to obtain a computing load feature vector; This sub-step uses lightweight time series feature extraction technology to extract the spatiotemporal features of the computing load from the power consumption data of the edge device. The calculation formula is: ; in: represents the extracted load feature vector; represents the time series feature extraction function; Indicates the time at which the device Power consumption data; Indicates the time window parameter for trend extraction; Represents a statistical feature extraction function, used to calculate mean, variance, peak, etc.; Represents the frequency domain feature extraction function, which is based on a simplified version of the Fast Fourier Transform (FFT); Represents the trend feature extraction function, which is used to capture the load change trend.

[0030] Extracting functions through time series features , for the device at time Power consumption data and the time window parameters for trend extraction Processing, the statistical feature extraction function right The calculation results and frequency domain feature extraction function right The calculation results and trend feature extraction function right and The calculation results are combined into a column vector to obtain the extracted load feature vector .

[0031] The algorithm is optimized for the characteristics of edge computing: it uses sliding window technology to reduce memory usage and dynamically adjusts the window size; it adopts a fixed-point FFT algorithm to avoid complex floating-point operations; and it compresses and stores historical data through differential coding.

[0032] Sub-step 203: Processing external environmental parameters using an environmental adaptability perception algorithm to obtain an environmental feature vector; This sub-step uses environmental factor correlation analysis technology to extract key environmental parameters related to cooling efficiency. The calculation formula is: ; in: Represents the environmental parameter vector, including ambient temperature, humidity, air pressure, wind speed, etc.; represents the extracted environmental feature vector; represents the environmental adaptability feature extraction function; Indicates the Nonlinear mapping function of environmental parameters; Indicates the The importance weight of each environmental parameter.

[0033] The algorithm automatically adjusts the weights of various environmental factors through online learning: ; in: Indicates the Moment The weight of each environmental parameter; Indicates the Moment The weight of each environmental parameter, that is, the weight updated by this formula; represents the learning rate, which controls the step size of the weight update; represents the cooling efficiency loss function; represents the cooling efficiency loss function For the first Moment Environmental parameter weights The partial derivative of .

[0034] In the Moment The weight of the environmental parameters Based on this, plus the learning rate And the partial derivative of the cooling efficiency loss function with respect to this weight The product of , thus getting the Weight after updating , so as to achieve online learning adjustment of the weights of various environmental factors.

[0035] Sub-step 204: Processing multi-source sensor data through a feature fusion compression algorithm to obtain lightweight environment and load feature vectors; This sub-step uses heterogeneous feature adaptive fusion technology to integrate multi-source heterogeneous sensor data into a unified low-dimensional feature vector. The calculation formula is: ; in: Represents the unified feature vector after fusion, which is obtained by the feature fusion function Temperature characteristics , load characteristics , environmental characteristics and fusion hyperparameters The result obtained after processing; It is a feature fusion function, which realizes the function of fusing different types of features; 、 、 Represent temperature characteristics, load characteristics and environmental characteristics respectively, and are the original feature data involved in fusion; It is a fusion hyperparameter used to adjust the feature weights and control the fusion method during the fusion process; Represents the activation function, which The calculation results are transformed nonlinearly, so that the fused feature vector has better expression ability; 、 、 Represents the fusion weight matrix of each feature, which is used to adjust the importance of different features in the fusion process. Different weight matrices will affect the contribution ratio of each original feature in the final fusion feature vector; Represents the bias term, which can introduce a fixed offset to the fused result, helping the model to better fit the data.

[0036] Through the feature fusion function , for temperature characteristics , load characteristics , environmental characteristics Multiply by the corresponding weight matrix 、 、 Then add the bias term , and finally through the activation function Perform nonlinear transformation to obtain the unified eigenvector after fusion , which is the lightweight environment and load characteristic vector.

[0037] To adapt to the edge computing environment, the feature fusion algorithm has been optimized as follows: a low-rank matrix approximation is used instead of a fully connected weight matrix to reduce computational complexity; a shift operation is used instead of a multiplication operation to speed up the calculation process; and an incremental feature update is implemented to avoid repeated calculation of unchanged features. The final generated features are all represented by 8-bit fixed-point numbers, the total dimension of the vector is controlled within 16 dimensions, and the amount of feature data in a single time step does not exceed 16 bytes to meet the storage and computing resource limitations of edge computing devices.

[0038] Step 300: Processing lightweight environment and load feature vectors and ultra-lightweight thermodynamic digital twin models through a micro-predictive controller to obtain an adaptive liquid cooling control strategy; This step utilizes the outputs of step 100 and step 200 to build a resource-efficient predictive control system and generate an optimized liquid cooling control strategy.

[0039] Sub-step 301: Processing lightweight environment and load feature vectors through a feedforward neural network predictor to obtain a thermal load prediction sequence; This sub-step uses a minimalist feedforward neural network structure to predict the heat load changes in the future time window. The calculation formula is:

[0040] in: represents the lightweight environment and load feature vector obtained in step 204; Indicates the past Historical heat load data for time steps; Indicates the future The predicted heat load sequence of time steps; represents the heat load prediction function; Indicates that the parameter is Multilayer Perceptron; Represents the vector concatenation operation, that is, and Perform splicing, and then input the spliced result into the parameter Multilayer Perceptron In the thermal load prediction function Get the future time-step predicted heat load series .

[0041] To adapt to the limited resources of edge devices, the predictor adopts the following optimized design: the network depth does not exceed 2 layers, and the number of neurons in each layer does not exceed 8; linear activation functions are used instead of nonlinear functions to reduce computational complexity; the prediction time window length is Dynamically adjusted based on available computing resources (usually 3-10 steps); historical window length Fixed to 4, compressing earlier historical data through exponentially weighted average; Sub-step 302: Processing the heat load prediction sequence and the ultra-lightweight thermodynamic digital twin model through a lightweight model predictive control algorithm to obtain a multi-step control sequence; This sub-step is based on the principle of model predictive control (MPC) and combines the ultra-lightweight thermodynamic digital twin model obtained in step 100 to calculate the optimal control sequence in the future control time domain. The calculation formula is: ; ; ; ; in: represents the optimized control sequence; Indicates at time Predicted Temperature at all times; Indicates the target temperature; Indicates at time planned Control input at all times; Indicates at time Predicted Thermal load at all times; represents the environmental feature vector; Represents an ultra-lightweight thermodynamic digital twin model; and Represent the weighted matrices and The quadratic form of Indicates the length of the predictive control time domain (usually 3-5 steps); and Represents the upper and lower limit constraints of the control input; and Represents the upper and lower temperature constraints.

[0042] Under the conditions of satisfying the upper and lower limit constraints of temperature and control input, as well as the temperature prediction model, by finding a suitable control input sequence , so that the sum of the deviation between the predicted temperature and the target temperature and the change amplitude of the control input is minimized, thereby obtaining the optimized control sequence .

[0043] To reduce computational complexity, the standard MPC algorithm is simplified as follows: the nonlinear optimization problem is approximated as a linear quadratic problem (by local linearization); the prediction time domain Dynamically adjust according to computing resources; adopt the early stopping strategy of gradient descent to limit the maximum number of iterations.

[0044] Sub-step 303: Processing the multi-step control sequence through a control strategy optimization algorithm to obtain an adaptive liquid cooling control strategy; This sub-step further optimizes the control sequence generated by MPC, that is, the multi-step control sequence obtained in step 302 is processed by the control strategy optimization algorithm. , and its calculation formula is: ; in: represents the optimized control strategy at the current moment, i.e., the adaptive liquid cooling control strategy; Represents the first control input in the original multi-step control sequence; represents the control correction item based on energy efficiency considerations; represents the control strategy optimization function; represents the energy efficiency-cooling balance factor; The control correction term is calculated using the "Energy Efficiency Sensitive Thermal Management Algorithm": ; in: Indicates the current system energy status; Indicates the current temperature; Indicates the temperature threshold; Represents the energy efficiency evaluation function, with a value range of [0,1]; The specific form of the function is: ; in: Indicates safety margin; represents the energy status scoring function, with a value range of [0,1]; The resulting adaptive liquid cooling control strategy Includes: coolant flow control instructions; fan speed control instructions (if any); thermal management working mode selection; energy efficiency optimization parameters.

[0045] Step 400: Processing an adaptive liquid cooling control strategy through an ultra-low power execution unit to obtain a precise physical cooling effect; This step converts the control strategy generated in step 300 into a physical control signal and implements precise control through an efficient actuator: Sub-step 401: Processing the adaptive liquid cooling control strategy through a control signal quantization algorithm to obtain a digital control signal; This sub-step converts the continuous control strategy into a discrete digital control signal to meet the input requirements of the digital execution unit. The calculation formula is: ; in: represents the adaptive liquid cooling control strategy obtained in step 303; Represents the quantized digital control signal; represents the control signal quantization function; Indicates the quantization step size; Represents the rounding function.

[0046] The algorithm is specially optimized for edge devices: quantization step size It can be dynamically adjusted to strike a balance between control accuracy and energy consumption; different quantization accuracies are used for key control parameters and secondary control parameters; and fast quantization based on table lookup method is achieved to avoid floating-point operations.

[0047] Sub-step 402: Processing the digital control signal through a pulse width modulation conversion algorithm to obtain a PWM drive signal; This sub-step converts the digital control signal into a pulse width modulation (PWM) signal to efficiently drive the actuator. The calculation formula is: ; in: represents the digital control code generated in sub-step 401; Indicates the duty cycle value of the PWM signal, ranging from 0% to 100%; represents the PWM transfer function; Indicates the carrier frequency, which determines the basic period of the PWM signal; Indicates the minimum value of the control range; Indicates the maximum value of the control range.

[0048] The efficient PWM generator adopts the following optimized design: Dynamically adjust the carrier frequency according to the characteristics of different actuators Utilize hardware timers to generate PWM signals, reducing CPU processing burden; use pre-calculated duty cycle tables to avoid real-time division operations; implement phase staggering for multiple PWM channels to reduce peak current; The output PWM control signal is a set of pulse width modulation waveforms with variable duty cycle, which is used to accurately control the cooling actuator.

[0049] Sub-step 403: Processing the PWM control signal and the system status information through the multi-layer safety protector to obtain a safety protection control signal; This sub-step uses a multi-layer safety protector to add a safety protection mechanism based on the PWM control signal generated in sub-step 402 to ensure that the system can still operate safely under abnormal conditions. The multi-layer safety protector consists of three parts: a status monitoring module, an abnormality diagnosis module, and a safety response module. The status monitoring module collects key system status information, the abnormality diagnosis module detects potential abnormalities, and the safety response module modifies the control signal under abnormal conditions. The calculation formula of the multi-layer safety protector is: ; in: represents the PWM control signal generated in sub-step 402; Represents the system status information vector, including sensor data, power status, etc. Indicates the control signal for adding security protection; Represents a security protection function; Represents the emergency state detection function, which returns 1 when a dangerous state is detected; Represents a warning status detection function, which returns 1 when a potential risk is detected; Indicates emergency control signal, used for safe operation in emergency situations; A signal adjustment function representing a warning state modifies the control signal according to the risk level; The multi-layer safety protector includes three levels of protection mechanisms: hardware monitoring layer: including a watchdog circuit to regularly detect the operating status of the controller; software monitoring layer: real-time monitoring of sensor data and actuator feedback signals; fail-safe layer: defines multiple fail-safe modes to ensure that the system automatically switches to a safe state in the event of control failure; The output safety protection control signal is a set of safety-verified PWM signals that can automatically adjust or switch to safety mode under abnormal circumstances.

[0050] Sub-step 404: Processing the safety protection control signal through a low-power drive circuit to obtain a precise physical cooling effect; This sub-step uses a low-power drive circuit to convert the safety protection control signal generated in sub-step 403 into an actual physical control effect. The low-power drive circuit consists of three parts: a signal amplification module, a power regulation module, and a feedback monitoring module. The signal amplification module amplifies the control signal to a level sufficient to drive the actuator. The power regulation module adjusts the output power according to the load demand. The feedback monitoring module monitors the execution effect and provides feedback. The efficiency calculation formula of the low-power drive circuit is: ; in: Indicates the efficiency of the drive circuit, which measures the ratio of input power converted into effective output power; Indicates the effective output power, used to drive the cooling actuator; Indicates input power, the total power consumed by the driving circuit; Represents the power loss, including switching loss, conduction loss and static loss.

[0051] The low-power drive circuit design adheres to the following key principles: monostable design: automatically returns to a safe state when the control signal is interrupted to avoid the risk of loss of control; gate drive optimization: optimizes the MOSFET drive waveform to minimize switching losses; power cascade control: automatically enables or disables some power units according to load demand; temperature compensation: built-in temperature compensation mechanism ensures stable operation over a wide temperature range.

[0052] Through this optimized design, the low-power drive circuit achieves over 95% efficiency, reducing static power consumption to microwatts, and keeping the overall control system power consumption within 3% of the total cooling system power. The physical control signals processed by the low-power drive circuit, known as safety protection control signals, represent the actual operating status of the cooling system's actuators (such as pumps and fans). These signals include quantitative indicators such as coolant flow rate, temperature control range, and system response time. By processing these signals, precise physical cooling effects are achieved.

[0053] In one embodiment of the present invention, an example of the aforementioned heat exchange system of a liquid cooling cabinet and a method of using the same is provided: Application scenario description: A liquid-cooled cabinet system in an edge data center has the following characteristics: Number of liquid-cooled cabinets: 20; Number of servers per cabinet: 48; Maximum power per cabinet: 30kW; Deployment environment: Edge computing site in mountainous areas; Ambient temperature range: -15°C to 55°C; Power grid conditions: Unstable, with voltage fluctuations ranging from -25% to +25%; Computing load characteristics: AI edge inference and data preprocessing, with large load fluctuations.

[0054] In this scenario, traditional liquid cooling control systems cannot meet actual needs due to insufficient computing power, delayed response, and poor environmental adaptability. Therefore, this data center applied the liquid cooling cabinet heat exchange system of this embodiment to meet the challenges of limited resources and harsh environments.

[0055] Step 100 Implementation Example: Practical Application of Hierarchical Model Distillation Algorithm: Traditional liquid-cooled cabinet thermodynamic models need to be run on high-performance servers, resulting in a model size of 500MB, making them impractical for deployment on edge controllers. By applying the hierarchical model distillation algorithm, we achieved extreme model lightweighting, as shown in Table 1: Table 1: Thermodynamic model compression process data:

[0056] Feature dimensionality reduction implementation details: A two-stage dimensionality reduction strategy was adopted. First, local linear embedding (LLE) was used to reduce the original 3072-dimensional features to 64 dimensions, retaining 97.8% of the information. Then, kernel principal component analysis (KPCA) was used to further reduce the dimensions to 8 dimensions, with an overall information retention rate of 95.3%. Knowledge distillation implementation details: Soft labels were generated using a temperature parameter τ = 2.5, a balance factor α = 0.3, and 10,000 training iterations. The final student network structure was: 8-dimensional input → 12 neurons → 12 neurons → 6-dimensional output, with ReLU as the activation function. Parameter quantization example: Weight quantization is performed using the three values {-1, 0, 1}. After quantization, the model sparsity reaches 83.7%, and the weight scaling factors of each layer are [0.125, 0.0625, 0.03125].

[0057] Step 200 implementation example: practical application of multimodal micro-sensor network: In field deployment, traditional methods require installing more than 30 temperature sensors and multiple load monitoring devices in each cabinet, which is costly and complex to maintain. By applying this implementation, we achieved efficient sensor network deployment, as shown in Table 2: Table 2: Sensor deployment optimization data:

[0058] Load feature extraction implementation details: 32-point FFT is used to analyze the power consumption fluctuation spectrum, extracting three main frequency components, four statistical features (mean, standard deviation, crest factor, and skewness), and three trend features (short-term slope, medium-term slope, and fluctuation index), which together form a 10-dimensional load feature vector. Environmental adaptability implementation details: The weights of environmental factors are dynamically adjusted through online learning. Initially, the weights of ambient temperature, humidity, air pressure, and wind speed were set to 0.5, 0.2, 0.1, and 0.2, respectively. After five days of adaptive learning, the weights converged to [0.65, 0.15, 0.05, 0.15], significantly improving environmental adaptability. Feature fusion implementation details: The original 26-dimensional features (8-dimensional temperature + 10-dimensional load + 8-dimensional environment) are fused and compressed into 14 dimensions using a low-rank approximation method. The features are represented using 8-bit fixed-point numbers, resulting in a total data size of only 14 bytes. Step 300 implementation example: practical application of micro-predictive controller: In this project, the traditional PID controller was unable to predict load fluctuations, and the response lag resulted in large temperature fluctuations. By deploying the micro-predictive controller of this embodiment, early response and optimized control were achieved, as shown in Table 3: Table 3: Predictive control performance data:

[0059] Thermal load prediction implementation details: A two-layer neural network structure with eight neurons per layer is used to predict load changes four steps into the future using data from four historical time steps. The prediction window length can be adaptively adjusted from 3 to 5 steps. The prediction accuracy reaches 89.7% in high-load change scenarios and 96.8% in low-load change scenarios. MPC control implementation details: The optimization problem is simplified to a linear quadratic problem, with a maximum number of iterations of 8. Early termination conditions are when the gradient norm is less than 0.01 or when the objective function improves by less than 0.5%. In typical scenarios, convergence is achieved within an average of 4.3 iterations, and computational latency is kept within 7ms. Control strategy optimization implementation details: The energy efficiency-cooling balance factor Φ is dynamically adjusted within the range of [0.1, 0.8], the safety margin ε is set to 2°C, and the energy state scoring function μ (E_sys) is implemented using a piecewise linear function. In battery-powered mode, the value of μ is reduced to 0.3 to reduce cooling power consumption. Step 400 implementation example: practical application of ultra-low power execution unit: In this project, traditional execution units consume high power, accounting for 8% to 12% of the total system power. The application of the ultra-low power execution unit in this implementation significantly reduces the energy consumption of the control system itself, as shown in Table 4: Table 4: Execution unit performance data:

[0060] Control signal quantization implementation details: 8-bit quantization is used for coolant flow control, with a quantization step size of 0.5%, achieving 200 levels of fine control within the range of 0% to 100%; 6-bit quantization is used for fan speed, achieving 64 levels of speed regulation; PWM drive implementation details: The carrier frequency is set to 10.5kHz, the duty cycle resolution is 0.5%, and the phase staggering technology reduces the peak current fluctuation by 28%, reducing electromagnetic interference; Safety protection implementation details: An integrated three-level safety protection mechanism provides real-time monitoring for 15 common fault types, with an average safety protection response time of 3.2ms and a maximum of no more than 9.7ms. When a potential overheating risk is detected, the system can switch to safe mode within 8.5ms. Low-power drive implementation details: Using synchronous rectification technology and adaptive gate drive optimization, the efficiency reaches 86.7% at 10% load, 92.3% at 50% load, and 94.8% at full load, improving overall energy efficiency by 9.8% compared to traditional solutions. After the edge data center applied this implementation for 6 months, the technical effectiveness of the system was verified by collecting operational data, as shown in Table 5: Table 5: Technical effect verification data:

[0061] Comprehensive effect analysis: Resource efficiency: This implementation achieves efficient control with extremely low computing resources, reducing ROM usage by 95.2% and RAM usage by 94.4%. It can run smoothly on a 100MHz MCU, significantly reducing hardware costs. Control accuracy and response speed: Temperature control accuracy has been improved from ±5.8°C (of traditional solutions) to ±2.1°C, a 63.8% improvement. Load response time has been reduced from 850ms to 16.5ms, a 51.5-fold improvement, meeting the requirements of highly variable load scenarios such as AI inference. Energy efficiency optimization: Control system power consumption was reduced from 10.5% to 2.3%, a 78.1% reduction. Annual energy consumption was also reduced by 30%, saving 2,628 kWh per cabinet per year, and a total of 52,560 kWh for 20 cabinets. Adaptability and reliability: The ambient temperature adaptability range has been expanded by 2.4 times, and the voltage fluctuation adaptability has been improved by 2.3 times, enabling the system to operate stably in harsh environments. The fault detection rate has been increased by 35%, and the system availability has been improved to 99.996%, with annual downtime not exceeding 21 minutes.

[0062] Through practical application verification, this implementation method has demonstrated significant technical advantages in the actual operating environment of the edge data center, solving the application bottleneck of traditional liquid cooling control systems in resource-constrained environments, and providing an efficient and reliable solution for the liquid cooling cabinet heat exchange system.

[0063] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for heat exchange of a liquid cooling cabinet, characterized in that: include: A hierarchical model distillation algorithm is used to process the complex thermodynamic model of the liquid cooling cabinet to obtain an ultra-lightweight thermodynamic digital twin model. The operating status of the liquid cooling cabinet is processed through a multimodal micro-sensor network to obtain lightweight environment and load feature vectors; An adaptive liquid cooling control strategy is obtained by processing lightweight environment and load feature vectors and ultra-lightweight thermodynamic digital twin models through a micro-predictive controller. The ultra-low power execution unit processes the adaptive liquid cooling control strategy to achieve precise physical cooling effects.

2. The heat exchange method of a liquid cooling cabinet according to claim 1, characterized in that: The calculation formula of the feature space dimensionality reduction algorithm is: ； in, represents the original thermodynamic feature space, is the number of sampling points, is the original feature dimension; represents the feature space after dimensionality reduction, is the feature dimension after dimensionality reduction; represents a nonlinear dimensionality reduction function; represents the locally linear embedding algorithm; represents kernel principal component analysis; Indicates the threshold for retaining information.

3. The heat exchange method for a liquid cooling cabinet according to claim 1, characterized in that: The calculation formula of the knowledge distillation compression network is: ； in, represents the knowledge distillation loss function; represents the cross entropy loss function; represents the KL divergence loss function; represents the true label; represents the output of the teacher network; represents the output of the student network; represents the softmax function; represents the temperature parameter; Represents the balance factor.

4. The heat exchange method for a liquid cooling cabinet according to claim 1, characterized in that: The calculation formula of the load feature extraction algorithm is: ； in, Indicates the time at which the device Power consumption data; represents the extracted load feature vector; represents the time series feature extraction function; represents the statistical feature extraction function; represents the frequency domain feature extraction function; represents the trend feature extraction function; Indicates the time window parameter for trend extraction; is the time window.

5. The heat exchange method for a liquid cooling cabinet according to claim 1, characterized in that: The micro-predictive controller includes: a feedforward neural network predictor, used to process lightweight environment and load feature vectors to obtain a thermal load prediction sequence; a lightweight model predictive control algorithm, used to process the thermal load prediction sequence and an ultra-lightweight thermodynamic digital twin model to obtain a multi-step control sequence; and a control strategy optimization algorithm, used to process the multi-step control sequence to obtain an adaptive liquid cooling control strategy.

6. The heat exchange method for a liquid cooling cabinet according to claim 5, characterized in that: The calculation formula of the lightweight model predictive control algorithm is: ； ； ； ； in, represents the optimized control sequence; Indicates at time Predicted Temperature at all times; Indicates the target temperature; Indicates at time planned Control input at all times; Indicates at time Predicted Thermal load at all times; represents the environmental feature vector; Represents an ultra-lightweight thermodynamic digital twin model; and Represent the weighted matrices and The quadratic form of represents the length of the predictive control time domain; , Represents the upper and lower limit constraints of the control input; , Indicates the upper and lower limit constraints of temperature; It means that under the conditions of satisfying the upper and lower limit constraints of temperature and control input, as well as the temperature prediction model, a suitable control input sequence is found. , so that the sum of the deviation between the predicted temperature and the target temperature and the change amplitude of the control input is minimized, thereby obtaining the optimized control sequence .

7. The heat exchange method for a liquid cooling cabinet according to claim 1, characterized in that: The ultra-low power execution unit includes: a control signal quantization algorithm for processing the adaptive liquid cooling control strategy to obtain a digital control signal; a pulse width modulation conversion algorithm for processing the digital control signal to obtain a PWM drive signal; a multi-layer safety protector for processing the PWM control signal and system status information to obtain a safety protection control signal; and a low-power drive circuit for processing the safety protection control signal to obtain a precise physical cooling effect.

8. A heat exchange system for a liquid cooling cabinet, characterized in that: include: A hierarchical model distillation module is used to process complex thermodynamic models of liquid-cooled cabinets and obtain ultra-lightweight thermodynamic digital twin models; A multimodal micro-sensor network module is used to process the operating status of the liquid cooling cabinet and obtain lightweight environment and load feature vectors; A micro-predictive controller module is used to process lightweight environmental and load characteristic vectors and ultra-lightweight thermodynamic digital twin models to obtain adaptive liquid cooling control strategies; Ultra-low power execution unit module, used to process adaptive liquid cooling control strategy to obtain precise execution of physical cooling effect.

9. The heat exchange system of the liquid cooling cabinet according to claim 8, characterized in that: The multimodal micro-sensor network module includes: a sparse sensor sampling optimization unit for processing temperature field data to obtain key point temperature distribution characteristics; a load feature extraction unit for processing computing device power consumption data to obtain computing load feature vectors; an environmental adaptability perception unit for processing external environmental parameters to obtain environmental feature vectors; and a feature fusion compression unit for processing multi-source sensor data to obtain lightweight environment and load feature vectors.

10. The heat exchange system of the liquid cooling cabinet according to claim 8, characterized in that: The multi-step control sequence optimization of the micro-predictive controller module adopts the control strategy optimization algorithm, and its calculation formula is: ； ； ； in, Represents a multi-step control sequence; Represents the optimized control strategy at the current moment; Represents the first control input in the original multi-step control sequence; represents the control correction item based on energy efficiency considerations; represents the control strategy optimization function; represents the energy efficiency-cooling balance factor; Indicates the current system energy status; Indicates the current temperature; Indicates the temperature threshold; represents the energy efficiency evaluation function; Indicates safety margin; represents the energy status scoring function.

Images