AI cluster server running at low temperature and use method thereof

Abstract

The invention relates to the technical field of AI cluster servers, in particular to a low-temperature running AI cluster server and a using method thereof.The method comprises the following steps that S1, the AI cluster server running at the low temperature through a cold plate is obtained, the cold plate is internally provided with a multi-parameter sensor unit for detecting temperature T, pH value, conductivity EC, suspended particle concentration SC and oxidation reduction potential ORP data influencing the performance of the cooling liquid; s2, carrying out noise removal and abnormal value processing on the obtained data of the temperature T, the pH value, the conductivity EC, the suspended particle concentration SC and the oxidation reduction potential ORP; s3, dividing the data set preprocessed in the step 2 into a training set and a test set, and constructing an LSTM model.According to the method, the health state of the cooling liquid in the AI cluster server running at the low temperature can be accurately and timely monitored and predicted, the time node of liquid cooling replacement can be accurately judged to the maximum extent, and the accuracy of liquid cooling replacement is improved. The stable operation of the server is further ensured; and the maintenance cost and risk are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of AI cluster servers, and specifically provides an AI cluster server with low-temperature operation and its usage method. Background Art

[0002] An AI cluster server with low-temperature operation is a high-performance computing platform designed specifically for artificial intelligence applications, and usually adopts liquid cooling technology to achieve efficient heat dissipation and stable operation; This server quickly exports the heat of the main heat sources (such as CPUs and GPUs) through a liquid cooling system (such as cold plates and cooling circuits), ensuring that key components operate in a low-temperature environment, thereby improving computing performance and reliability; For the existing AI cluster servers with low-temperature operation cooled by cold plates, during actual application, it is necessary to regularly replace the coolant in the cold plates to prevent the coolant from failing and affecting the normal operation of the cooling system. During the replacement process, if the replacement is carried out too early before the coolant fails, it will increase the frequency of coolant replacement and the cost of coolant replacement. If the replacement is carried out after the coolant fails, during the failure period, the cooling effect of the cold plate will be affected, and there is a possibility of system failure. Therefore, an AI cluster server with low-temperature operation and its usage method are proposed for the above problems. Summary of the Invention

[0003] The purpose of the present invention is to provide an AI cluster server with low-temperature operation and its usage method to solve the problem that for the existing AI cluster servers with low-temperature operation cooled by cold plates, during actual application, it is necessary to regularly replace the coolant in the cold plates to prevent the coolant from failing and affecting the normal operation of the cooling system. During the replacement process, if the replacement is carried out too early before the coolant fails, it will increase the frequency of coolant replacement and the cost of coolant replacement. If the replacement is carried out after the coolant fails, during the failure period, the cooling effect of the cold plate will be affected, and there is a possibility of system failure.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: An AI cluster server with low-temperature operation and its usage method, including the following steps: S1: Obtain an AI cluster server with low-temperature operation through a cold plate. The cold plate is internally provided with a multi-parameter sensor unit for temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP data that affect the performance of the coolant; S2: Perform noise and outlier removal processing on the obtained temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP data; S3: Divide the preprocessed dataset in step 2 into a training set and a test set, and construct an LSTM model to output the coolant health state index HSI1. The mathematical expression of the coolant health state index HSI1 is:

[0005] Where, is the forget gate, is the input gate, is the candidate memory cell, is the memory cell at the previous moment; S4: Select the preprocessed temperature, pH, conductivity, suspended particle concentration, and redox potential in step 2 as features, train a linear regression model, and output the coolant health state index HSI2. The mathematical expression of the coolant health state index HSI2 is:

[0006] Where, is the weight coefficient, is the bias term; S5: Perform weighted processing on the output results of the LSTM model and the linear regression model to obtain the final coolant health state index HSI0. The calculation formula of the final coolant health state index HSI0 is:

[0007] Where, and are the regression coefficients.

[0008] As a further optimization content of the present invention, it includes the following steps: S6: According to experimental data and historical experience, set the dynamic threshold HSI4 of the "coolant health state index", compare the set dynamic threshold HSI4 with the final coolant health state index HSI0 to determine whether the coolant needs to be replaced. The determination principle is: If HSI0 < HSI4, it is determined that the coolant needs to be replaced; If HSI0 ≥ HSI4, continue to monitor and predict the performance of the coolant.

[0009] As a further optimization content of the present invention, in the S2, removing noise from the temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP includes filtering processing and smoothing processing; The filtering processing uses a low-pass filter to remove high-frequency noise, and the mathematical expression is:

[0010] Where, where the filtering coefficient x[n] is the current input data, and y[n−1] is the output data at the previous moment; The smoothing process uses the moving average method to smooth the data, and the mathematical expression is:

[0011] In the formula, k is the radius of the smoothing window.

[0012] As a further optimization content of the present invention, wherein: in the S2, the Z - score method and the box - plot method are used to remove outliers for the temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP. The mathematical expression of the Z - score method is:

[0013] In the formula, x is the data point, μ is the mean value, and σ is the standard deviation; The box - plot method is used to identify and remove data points that exceed 1.5 times the inter - quartile range.

[0014] As a further optimization content of the present invention, wherein: the construction process of the LSTM model in the S3 is as follows: S31: Data preparation: Divide the pre - processed data in step two into a training set and a test set; S32: Model structure: Build an LSTM network, including an input layer, a hidden layer, and an output layer. The input layer receives the pre - processed data, the hidden layer processes the data through a forget gate, an input gate, and an output gate, and the output layer outputs the health status index of the coolant.

[0015] S33: Loss function: Use the mean squared error as the loss function to measure the difference between the predicted value and the actual value; S34: Optimization method: Use the Adam optimizer for model training, and continuously optimize the model parameters through the backpropagation algorithm to minimize the loss function; S35: Training process: Optimize the model parameters through multiple iterative trainings; S36: Prediction output: The LSTM network outputs the health status index HSI1 of the coolant.

[0016] As a further optimization content of the present invention, wherein: the training process of the linear regression model in the S4 is as follows: S41: Feature selection: Select the temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP pre - processed in step two as features; S42: Model training: Use historical data to train the linear regression model, and estimate the model parameters by the least squares method; S43: Loss function: The mean square error is used as the loss function to measure the difference between the predicted value and the actual value; S44: Optimization method: Use gradient descent method to optimize model parameters and minimize loss function; S45: Prediction output: The linear regression model outputs the coolant health status index HSI2.

[0017] As a further optimized content of the present invention, wherein: in said S5, the process of determining the regression coefficient is: S51: Data partitioning: Divide historical data into training set and validation set; S52: Model training: Use the training set to train the weighted model and determine the regression coefficient; S53: Loss function: Mean square error (MSE) is used as the loss function; S54: Optimization method: Use gradient descent method to optimize regression coefficients; S55: Validation Evaluation: Use the validation set to evaluate the accuracy of the model and adjust the regression coefficients to optimize model performance.

[0018] As further optimized content of the present invention, it includes: an AI cluster server running at low temperature; A liquid-cooled cold plate that fits the AI ​​cluster server running at low temperature. The cold plate has a built-in multi-parameter sensor unit for real-time monitoring of the temperature T, pH, conductivity EC, suspended particle concentration SC, and redox potential ORP data that affect the performance of the coolant; A central data platform for processing the data collected by the multi-parameter sensor unit, including noise removal and outlier processing; An LSTM model and a linear regression model, wherein the LSTM model and the linear regression model are used to output health status indexes HSI1 and HSI2 of the coolant, respectively; The weighted processing module is used to perform weighted processing on the output results of the LSTM model and the linear regression model to obtain the final coolant health status index HSI0; The determination module is used to compare the final coolant health status index HSI0 with the set dynamic threshold HSI4 to determine whether the coolant needs to be replaced.

[0019] As further optimized content of the present invention, wherein: the system also includes a memory, a processor and an electronic program stored in the memory and capable of running on the processor, wherein the processor can implement the steps of the method described in any one of claims 1-7 when running the electronic program.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. In the present invention, the health status of the coolant in the AI cluster server operating at low temperature can be accurately and timely monitored and predicted, and the time node for liquid cooling replacement can be accurately judged to the greatest extent, further ensuring the stable operation of the server and reducing the maintenance cost and risk. 2. In the present invention, methods such as filtering, smoothing, and outlier removal are adopted in the data processing stage to effectively remove noise and outliers, ensuring the quality of the data input into the model. By using the LSTM model to capture the long-term dependence relationship of time series data and combining with the linear regression model to fit the linear relationship of the data, the prediction result of the coolant health status index is further optimized. Through multi-model weighted processing, the advantages of different models are fully utilized, improving the comprehensive performance of the prediction result. 3. In the present invention, by setting a dynamic threshold and making an automatic judgment, it is possible to timely and accurately determine whether the coolant needs to be replaced, avoiding the subjectivity and error of manual judgment. The automatic judgment mechanism not only improves the timeliness and accuracy of coolant maintenance but also reduces the cost of manual intervention and maintenance risk. In addition, the modular design and intelligent processing of the system make the monitoring and maintenance of the coolant health status more efficient and convenient. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flowchart for detecting the health index of the coolant in an AI cluster server operating at low temperature according to the present invention. Figure 2 It is a system block diagram of an AI cluster server operating at low temperature according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0022] Please refer to Figure 1-2 , the present invention provides a technical solution: An AI cluster server operating at low temperature and its usage method, including the following steps: S1: Obtain an AI cluster server operating at low temperature through a cold plate. The cold plate is internally equipped with a multi-parameter sensor unit to collect data on temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP that affect the performance of the coolant. S2: Perform noise and outlier removal processing on the obtained data of temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP. S3: Divide the preprocessed data set in step two into a training set and a test set, and construct an LSTM model to output the coolant health status index HSI1. The mathematical expression of the coolant health status index HSI1 is:

[0023] In the formula, is the forget gate, is the input gate, is the candidate memory cell, is the memory cell at the previous moment; S4: Select the preprocessed temperature, pH, conductivity, suspended particle concentration, and redox potential in Step 2 as features, train a linear regression model, and output the coolant health state index HSI2. The mathematical expression of the coolant health state index HSI2 is:

[0024] In the formula, is the weight coefficient, is the bias term; S5: Perform weighted processing on the output results of the LSTM model and the linear regression model to obtain the final coolant health state index HSI0. The calculation formula of the final coolant health state index HSI0 is:

[0025] In the formula, and are regression coefficients. By real-time monitoring the performance indicators of the coolant through the multi-parameter sensor unit and combining the LSTM model and the linear regression model for comprehensive evaluation, the health state of the coolant can be accurately predicted, potential problems can be discovered in time, the operation stability and reliability of the AI cluster server can be improved, and the maintenance cost and risk can be reduced.

[0026] As a further implementation technical solution of this scheme, the following steps are also included: S6: According to the experimental data and historical experience, set the dynamic threshold HSI4 of the "coolant health state index", compare the set dynamic threshold HSI4 with the final coolant health state index HSI0 to determine whether the coolant needs to be replaced. The determination principle is: If HSI0 < HSI4, it is determined that the coolant needs to be replaced; If HSI0 ≥ HSI4, continue to monitor and predict the performance of the coolant. By setting the dynamic threshold and making a comparison judgment, it is possible to timely and accurately determine whether the coolant needs to be replaced, avoiding the subjectivity and error of manual judgment, and improving the timeliness and accuracy of coolant maintenance; As a further implementation technical solution of this scheme, in S2, removing noise from the temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP includes filtering processing and smoothing processing; The filtering processing uses a low-pass filter to remove high-frequency noise. The mathematical expression is:

[0027] In the formula, where the filtering coefficient \(x[n]\) is the current input data, and \(y[n - 1]\) is the output data at the previous moment; Smoothing processing uses the moving average method to smooth the data, and its mathematical expression is:

[0028] In the formula, \(k\) is the radius of the smoothing window. In S2, for the temperature \(T\), pH value \(pH\), conductivity \(EC\), suspended particle concentration \(SC\), and oxidation-reduction potential \(ORP\), the Z-score method and the box plot method are used to remove outliers. The mathematical expression of the Z-score method is:

[0029] where \(x\) is the data point, \(\mu\) is the mean, and \(\sigma\) is the standard deviation; The box plot method is used to identify and remove data points that exceed 1.5 times the interquartile range. Through filtering and smoothing processing, the noise in the data can be effectively removed, improving the accuracy and reliability of the data, providing a higher-quality data basis for subsequent model training and prediction, helping to improve the prediction accuracy of the coolant health status index. At the same time, using the Z-score method and the box plot method to remove outliers can effectively identify and remove outliers in the data, further improving the quality and reliability of the data, ensuring the accuracy of model training and prediction, and avoiding the negative impact of outliers on the model performance; As a further implementation technical solution of this scheme, the LSTM model construction process in S3 is as follows: S31: Data preparation: Divide the preprocessed data in step two into a training set and a test set; S32: Model structure: Construct an LSTM network, including an input layer, a hidden layer, and an output layer. The input layer receives the preprocessed data, and the hidden layer processes the data through the forget gate, input gate, and output gate. The output layer outputs the health status index of the coolant.

[0030] S33: Loss function: Use the mean squared error as the loss function to measure the difference between the predicted value and the actual value; S34: Optimization method: Use the Adam optimizer for model training, and continuously optimize the model parameters through the backpropagation algorithm to minimize the loss function; S35: Training process: Optimize the model parameters through multiple iterations of training; S36: Prediction output: The LSTM network outputs the health status index HSI1 of the coolant, which can systematically perform model training and optimization, give full play to the advantages of the LSTM network in processing time series data, improve the prediction accuracy and reliability of the coolant health status index, and provide strong support for the maintenance decision of the coolant; As a technical solution for further implementation of this solution, the training process of the linear regression model in S4 is: S41: feature selection: selecting the temperature T, pH, conductivity EC, suspended particle concentration SC and redox potential ORP after pretreatment in step 2 as features; S42: Model training: Use historical data to train the linear regression model and estimate the model parameters by the least squares method; S43: Loss function: The mean square error is used as the loss function to measure the difference between the predicted value and the actual value; S44: Optimization method: Use gradient descent method to optimize model parameters and minimize loss function; S45: Prediction output: The linear regression model outputs the coolant health status index HSI2, which can make full use of historical data for model training and optimization, give full play to the advantages of the linear regression model in fitting data relationships, improve the prediction accuracy and reliability of the coolant health status index, and provide a multi-angle evaluation basis for coolant maintenance decisions; As a technical solution for further implementing this solution, in S5, the process of determining the regression coefficient is: S51: Data partitioning: Divide historical data into training set and validation set; S52: Model training: Use the training set to train the weighted model and determine the regression coefficient; S53: Loss function: Mean square error (MSE) is used as the loss function; S54: Optimization method: Use gradient descent method to optimize regression coefficients; S55: Validation and evaluation: Use the validation set to evaluate the accuracy of the model and adjust the regression coefficient to optimize the model performance. This can scientifically determine the regression coefficient of the weighted model, give full play to the advantages of the weighted model in comprehensive evaluation, improve the accuracy and reliability of the final coolant health status index, and provide a more accurate basis for coolant maintenance decisions; As a technical solution for further implementation of this plan, it includes AI cluster servers running at low temperature; Liquid cooling cold plate that fits the low-temperature AI cluster server. The cold plate has a built-in multi-parameter sensor unit for real-time monitoring of the temperature T, pH, conductivity EC, suspended particle concentration SC, and redox potential ORP data that affect the performance of the coolant. A central data platform for processing the data collected by the multi-parameter sensor units, including noise removal and outlier processing; LSTM model and linear regression model, LSTM model and linear regression model are used to output coolant health status index HSI1 and HSI2 respectively; A weighted processing module, which is used to perform weighted processing on the output results of the LSTM model and the linear regression model to obtain the final coolant health state index HSI0; A determination module, which is used to compare and judge the final coolant health state index HSI0 with the set dynamic threshold HSI4 to determine whether the coolant needs to be replaced. The AI cluster server operating at low temperature can realize real-time monitoring, comprehensive evaluation and automatic judgment of the coolant health state by integrating a multi-parameter sensor unit, a central data platform, an LSTM model, a linear regression model, a weighted processing module and a determination module, improve the intelligence level and automation degree of the system, reduce the manual intervention and maintenance cost, and ensure the long-term stable operation of the AI cluster server; As a further technical solution of this solution, the system further includes a memory, a processor, and an electronic program stored in the memory and capable of running on the processor. Among them, when the processor runs the electronic program, it can implement the steps of any one of the methods in claims 1-7, further improve the intelligence level and automation degree of the system, realize the full-process automation of coolant health state monitoring and maintenance, improve the operation efficiency and reliability of the system, reduce the manual intervention and maintenance cost, and provide a strong guarantee for the stable operation of the AI cluster server.

[0031] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above examples is only used to help understand the method and its core idea of the present invention. The above is only the preferred implementation manner of the present invention. It should be noted that due to the limited nature of written expression and the objectively infinite specific structures, for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements, embellishments or changes can be made, or the above technical features can be combined in an appropriate manner; these improvements, embellishments, changes or combinations, or directly applying the concept and technical solution of the invention to other occasions without improvement, shall all be regarded as the protection scope of the present invention.

Claims

1. A method for using an AI cluster server operating at low temperature, characterized in that, It includes the following steps: S1: Obtain an AI cluster server operating at low temperature through a cold plate. The cold plate is built-in with a multi-parameter sensor unit to measure data on temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP that affect the performance of the coolant; S2: Remove noise and outliers from the obtained data on temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP; S3: Divide the preprocessed data set in step two into a training set and a test set, and build an LSTM model to output the coolant health state index HSI1. The mathematical expression of the coolant health state index HSI1 is: ； Wherein, is the forget gate, is the input gate, is the candidate memory cell, is the memory cell at the previous moment; S4: Select the preprocessed temperature, pH, conductivity, suspended particle concentration, and redox potential in step two as features, train a linear regression model, and output the coolant health state index HSI2. The mathematical expression of the coolant health state index HSI2 is: ； In the formula, is the weight coefficient, is the bias term; S5: Perform weighted processing on the output results of the LSTM model and the linear regression model to obtain the final coolant health state index HSI0. The calculation formula of the final coolant health state index HSI0 is: ； Wherein, and are regression coefficients.

2. The usage method of a low-temperature operating AI cluster server according to claim 1, characterized in that: It also includes the following steps: S6: According to experimental data and historical experience, set the dynamic threshold HSI4 of the "coolant health state index", compare the set dynamic threshold HSI4 with the final coolant health state index HSI0 to determine whether the coolant needs to be replaced. The determination principle is: If HSI0 < HSI4, it is determined that the coolant needs to be replaced; If HSI0 ≥ HSI4, continue to monitor and predict the performance of the coolant.

3. The usage method of a low-temperature operating AI cluster server according to claim 1, characterized in that: In S2, removing noise from temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP includes filtering processing and smoothing processing; The filtering processing uses a low-pass filter to remove high-frequency noise. The mathematical expression is: ； In the formula, is the filtering coefficient, x[n] is the current input data, and y[n−1] is the output data at the previous moment; The smoothing processing uses the moving average method to smooth the data. The mathematical expression is: ； In the formula, k is the radius of the smoothing window.

4. The usage method of a low-temperature operating AI cluster server according to claim 1, characterized in that: In S2, removing outliers from temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP uses the Z-score method and the box plot method. The mathematical expression of the Z-score method is: ； In the formula, x is the data point, μ is the mean, and σ is the standard deviation; The box plot method is used to identify and remove data points exceeding 1.5 times the interquartile range.

5. The usage method of a low-temperature operating AI cluster server according to claim 1, characterized in that: The LSTM model construction process in S3 is: S31: Data preparation: Divide the data preprocessed in step two into a training set and a test set; S32: Model structure: Build an LSTM network, including an input layer, a hidden layer, and an output layer. The input layer receives the preprocessed data, and the hidden layer processes the data through the forget gate, input gate, and output gate. The output layer outputs the coolant health state index; S33: Loss function: Use the mean square error as the loss function to measure the difference between the predicted value and the actual value; S34: Optimization method: Use the Adam optimizer to train the model, and continuously optimize the model parameters through the backpropagation algorithm to minimize the loss function; S35: Training process: Optimize the model parameters through multiple iterative trainings; S36: Prediction output: The LSTM network outputs the health state index HSI1 of the coolant.

6. The usage method of a low-temperature operating AI cluster server according to claim 1, characterized in that: The training process of the linear regression model in S4 is as follows: S41: Feature selection: Select the preprocessed temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP in step 2 as features; S42: Model training: Use historical data to train the linear regression model and estimate the model parameters by the least squares method; S43: Loss function: Adopt the mean square error as the loss function to measure the difference between the predicted value and the actual value; S44: Optimization method: Use the gradient descent method to optimize the model parameters and minimize the loss function; S45: Prediction output: The linear regression model outputs the health state index HSI2 of the coolant.

7. The usage method of an AI cluster server operating at low temperature according to claim 1, characterized in that: In S5, the process of determining the regression coefficient is as follows: S51: Data division: Divide the historical data into a training set and a validation set; S52: Model training: Use the training set to train the weighted model and determine the regression coefficient; S53: Loss function: Adopt the mean square error (MSE) as the loss function; S54: Optimization method: Use the gradient descent method to optimize the regression coefficient; S55: Validation and evaluation: Use the validation set to evaluate the accuracy of the model and adjust the regression coefficient to optimize the model performance.

8. An AI cluster server operating at low temperature according to any one of claims 1-7, characterized in that: Including an AI cluster server for low-temperature operation; A liquid cooling cold plate attached to the AI cluster server for low-temperature operation. The cold plate is built with a multi-parameter sensor unit for real-time monitoring of data such as temperature T, pH value, conductivity EC, suspended particle concentration SC, and redox potential ORP that affect the performance of the coolant; A central data platform for processing the data collected by the multi-parameter sensor unit, including noise removal and outlier processing; An LSTM model and a linear regression model, where the LSTM model and the linear regression model are respectively used to output the health state index HSI1 and HSI2 of the coolant; A weighted processing module for weighting the output results of the LSTM model and the linear regression model to obtain the final health state index HSI0 of the coolant; A determination module for comparing and judging the final health state index HSI0 of the coolant with the set dynamic threshold HSI4 to determine whether the coolant needs to be replaced.

9. An AI cluster server operating at low temperature according to claim 8, characterized in that: The system further includes a memory, a processor, and an electronic program stored in the memory and capable of running on the processor. Wherein, when the processor runs the electronic program, it can implement the steps of the method according to any one of claims 1-7.

Images