Intelligent hardware temperature prediction method

Abstract

The invention discloses an intelligent hardware temperature prediction method. The method comprises the following steps: S1, data acquisition; s2, feature processing: determining key features influencing the hardware temperature; s3, a hybrid model is constructed, a first model is a time sequence statistical model, and a second model is a regression model based on machine learning; s4, hybrid model prediction: predicting the first model and the second model respectively, and fusing the results of the first model and the second model through a linear regression model; and S5, prediction result integration: integrating the fine-grained prediction result into a coarse-grained comprehensive prediction result. The invention provides a comprehensive, accurate and reliable method for predicting and managing the temperature of the single-board hardware, so that the operation efficiency and the maintenance effect of the hardware are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hardware temperature monitoring and prediction, and in particular to a method for predicting hardware high temperature by data collection and analysis and model building, so as to optimize hardware operation and maintenance management. Background Art

[0002] In modern communication equipment, data centers, and other high-performance computing environments, temperature monitoring and management of single-board hardware is crucial. High temperature problems may cause damage to the equipment itself and serious damage to the business. Taking the baseband unit (BBU) in the base station operation and maintenance scenario as an example, the BBU is the core component responsible for processing signals from user equipment. The stable operation of the BBU is extremely important for maintaining the quality and performance of the network. The BBU is connected to the remote radio unit (RRU) through an optical fiber interface to control the RRU and process the data received by the RRU. These processes include channel encoding and decoding, multiplexing and demultiplexing, spread spectrum modulation and demodulation, network measurement and reporting, power control, and synchronous clock. However, the BBU often faces high temperature and overheating problems during operation, which may be caused by continuous high-load operation, insufficient heat dissipation measures or heat dissipation equipment failure, and excessively high ambient temperature. High temperature and overheating may not only cause BBU performance degradation, such as slow processing speed and increased response delay, but may also cause automatic restart or complete shutdown, seriously affecting the continuity of network services and user experience.

[0003] Existing hardware temperature management technologies can determine whether the current temperature is overheated by monitoring relevant hardware indicators, such as hardware temperature, ambient temperature, fan speed, etc. However, most existing hardware temperature management technologies only intervene after the temperature reaches the critical value. In fact, the maintenance time required for single-board hardware is relatively long. For example, the BBU single board has a long processing cycle for the station. Serious impacts caused by high temperature may have occurred during the maintenance process.

[0004] Therefore, the existing technology can only monitor temperature indicators in real time, but cannot provide early warning. The maintenance cycle of single-board hardware is relatively long, and it is easy for the hardware to be seriously affected by high temperature during the maintenance process. Therefore, developing a technology that can predict high temperature of single-board hardware in advance is of great significance to ensure the reliability of network communication and service quality. Summary of the invention

[0005] The present invention uses a highly customized data collection solution, combined with a hybrid prediction method of data analysis, machine learning, and time series modeling, to predict the temperature trend of single-board hardware, so as to achieve early warning of high temperature risks and proactive operation and maintenance management, significantly improving hardware maintenance efficiency and network service stability.

[0006] Technical solution:

[0007] An intelligent hardware temperature prediction method includes the following steps:

[0008] S1, data collection;

[0009] S2. Feature processing to determine the key features that affect the hardware temperature; the key features here are the ambient temperature and historical temperature data, ambient humidity, hardware location (indoor or outdoor), business volume, hardware system age, cooling system data, and the hardware's own temperature and historical data. Customized features can also be added according to different hardware, such as chassis temperature, fan speed, power consumption, etc.

[0010] S3. Construct a hybrid model, where the first model is a time series statistical model and the second model uses a regression model based on machine learning;

[0011] S4, hybrid model prediction, the first model and the second model make predictions respectively, and the results of the two are fused through a linear regression model;

[0012] S5: Integrate the prediction results, integrating the fine-grained prediction results into coarse-grained comprehensive prediction results.

[0013] Preferably, the data collection includes collecting ambient temperature, humidity, hardware location, historical temperature data, business volume, hardware system age, cooling system information and current and historical temperature data of the hardware itself.

[0014] Preferably, the feature processing includes:

[0015] S201, normalize the data and unify the dimensions of each indicator;

[0016] S202, selecting features in the normalized data that have a strong causal relationship with the prediction target as key features for identifying the impact on the hardware temperature;

[0017] S203, time series feature enhancement, decomposing the time series as follows:

[0018] Y t =T t +S t +R t

[0019] Among them, Y t is the time series to be decomposed at time t; T t is the trend decomposition sequence at time t, showing the long-term change trend of the sequence; S t is the seasonal decomposition sequence at time t, showing the short-term periodic changes of the sequence; R t It is the residual decomposition sequence at time t, showing the irregular or noise components of the sequence.

[0020] Preferably, in S202, for a simple system, a key feature affecting the hardware temperature is selected through causal testing, and the specific steps are:

[0021] Fit model (1) and model (2) using time series X and time series Y:

[0022] Y t =a0+a1Y t-1 +a2Y t-2 +…+∈ t (1)

[0023] Y t =β0+β1Y t-1 +β2Y t-2 +…+γ1X t-1 +γ2X t-2 +…+∈ t (2)

[0024] Among them, Y t represents the value of time series Y at time t, a0 represents the baseline level of time series Y in model (1), that is, the value of Y at time t when the past values ​​of the Y series itself are all zero, β0 represents the baseline level of time series Y in model (2) after considering the influence of the past values ​​of time series X, γ1 represents the influence coefficient of time series X at time t-1 on the value of Y at time t, and X t-1 Indicates the value of the X sequence at time t-1. The time step is set according to the specific hardware operation and maintenance scenario;

[0025] For the fitted models (1) and (2), the minimum mean square error (MSE) of the fitting is compared to determine whether model (2) is better than model (1); if the MSE of model (2) is less than the MSE of model (1), it means that feature X is the key feature of feature Y to be predicted.

[0026] Preferably, in S202, for a complex system, the key features that affect the hardware temperature are selected by transferring entropy, and the specific steps are:

[0027] For time series X and time series Y, the transfer entropy T from X to Y X→Y The calculation method is:

[0028]

[0029] in, represents the past k states of the time series Y; represents the past l states of the time series X; p() represents the joint probability distribution;

[0030] If the transfer entropy TX→Y > 0, it means that the past state of time series X provides information for the future state of time series Y, that is, X is the key feature.

[0031] Preferably, in S203, specifically:

[0032] 1) Trend decomposition sequence T t Calculated by the following formula:

[0033]

[0034] Among them, 2m+1 represents the window length of the moving average;

[0035] 2) Seasonal decomposition sequence S t Calculated by the following formula:

[0036] S t =LOESS(Y t -T t ,λ)

[0037] where λ is the time span parameter used by the LOESS smoother to capture the seasonal component;

[0038] 3) Residual decomposition sequence R t Calculated by the following formula:

[0039] R t =Y t -T t -S t .

[0041] Preferably, in 2), the specific calculation steps of LOESS smoothing are as follows:

[0042] 2-1) Calculate the local weighted fit:

[0043] Y t =β0+β1(x j -x i )+β2(x j -x i ) 2 +∈ j

[0044] Among them, β0 means when x i 、x j When they are equal, the response variable Y t The expected value of j -x i )’s first-order linear coefficient and second-order linear coefficient; x i 、x j Represents the fitting target points and local data points of local fitting;

[0045] 2-2) Calculate the minimum mean square error:

[0046] β mse =argmin β ∑ j w ij (y-β0-β1(x j -x i )-β2(x j -x i ) 2 ) 2

[0047] Among them, β mse represents a set of β0, β1, β2 found by minimizing the mean square error; weight Distance d ij =|x i -x j |,d i Represents the maximum distance within the local fitting range and defines the calculation weight w ij The size of the neighborhood used when ;

[0048] 2-3) Get i The fitted value of is: Indicates that in x i The fitted value y at i ;

[0049] Repeat LOESS smoothing for each time instant t, and the seasonal decomposition series S t .

[0050] Preferably, the first model is: an autoregressive integrated moving average model ARIMA, an ARIMA model with regression term ARIMAX or a seasonal autoregressive integrated moving average model SARIMA;

[0051] The second model is: random forest regression model or gradient boosted tree regression model.

[0052] Preferably, the first model is an ARIMA model ARIMAX with regression terms, and the second model is a random forest regression model; the specific steps of S4 are:

[0053] 1) Get the dual model prediction result at time t, Z t =[Y t,ARIMAX ,Y t,RF ],in:

[0054] Y t,ARIMAX =ARIMAX(X t )

[0055] Y t,RF=RF(X t )

[0056] Where, X t represents the time feature set obtained after S2 processing, Y t,ARIMAX Represents the ARIMAX model for X t The prediction result, Y t,RF Represents the random forest model for X t The prediction results,

[0057] 2) Use the linear regression model to fuse the prediction results of the two models:

[0058] P t =σ(β0+β1Y t,ARIMAX +β2Y t,RF )

[0059] Among them, P t is the fused prediction result at time t, β0, β1, β2 are logistic regression coefficients,

[0060] Preferably, in S5, the fine-grained prediction results are integrated into a coarse-grained comprehensive prediction result, for example, the prediction results of the collected hourly granularity data are integrated at the daily granularity.

[0061] Beneficial Effects of the Invention

[0062] The present invention provides a comprehensive, accurate and reliable method to predict and manage the temperature of single-board hardware, thereby greatly improving the operating efficiency and maintenance effect of the hardware. Specifically, it includes:

[0063] 1. Provides a unique data collection method: including ambient temperature, ambient humidity, hardware location (indoor or outdoor), historical temperature data, business volume, hardware system age, cooling system data, and hardware temperature and its historical data. Collecting these indicators helps to fully understand the various factors that affect hardware temperature.

[0064] 2. A method is provided to use causal analysis to select features that have strong causal effects on the prediction target from the normalized indicators. By calculating the causal value between each feature and the prediction target, the factors that have the greatest impact on the hardware temperature can be effectively determined.

[0065] 3. It provides a method to enhance the characteristics of time series. The time series is decomposed into trend series, seasonal series and residual series. This decomposition helps the subsequent model to better identify the trend and seasonal characteristics of the time series, thereby improving the accuracy of the prediction.

[0066] 4. A method for prediction using a hybrid model is provided. The method includes three sub-models: the first one uses a time series statistical model to capture the trend and seasonality of temperature data; the second model uses a trainable regression model to capture the complex relationship between indicators and hardware temperature. Finally, the third model is used to fuse the prediction results of the two models to obtain the hardware temperature prediction results of the hybrid model.

[0067] 5. Provided is a method for collecting fine-grained data training models and predicting results, and then integrating the fine-grained prediction results into coarse-grained comprehensive results. The prediction results of fine-grained data are obtained through the first four points of the present invention, and then the prediction results of fine-grained data are integrated into coarse-grained prediction results through mathematical operations. This method is more accurate and reliable than directly using fine-grained prediction results or using coarse-grained data for training and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is the overall flow chart of the intelligent single board hardware temperature prediction of the present invention

[0069] Figure 2 This is a flow chart of the intelligent single board hardware temperature prediction feature processing of the present invention

[0070] Figure 3 A hybrid model prediction flow chart for intelligent single board hardware temperature prediction of the present invention DETAILED DESCRIPTION

[0071] To make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the following will describe the implementation of the present invention in detail with reference to the accompanying drawings. The following implementation should not constitute any limitation on the specific implementation of the present invention, and each implementation part can be combined with each other without contradiction.

[0072] It should be noted that the device in the embodiment of the present invention may be various base station hardware devices. To facilitate the description of the principle, the embodiment of the present invention uses a common BBU board as an example device, which does not limit the technical solution of the present invention.

[0073] Combine the following Figure 1 The overall processing flow of the first embodiment is given.

[0074] Step 1: Collect data. In this step, the system automatically collects hourly granularity hardware data indicators related to the BBU board, including ambient temperature and humidity, the specific location of the board (indoor or outdoor), board power consumption, board fan speed, and board CPU temperature. These data are the basis of the prediction model, and the key is to provide sufficient input to reflect the operating status of the board.

[0075] Step 2: Combine Figure 2, feature processing, including processing data and time series decomposition.

[0076] Processing data: First, normalize the data collected in step 1 to unify the data dimension for subsequent analysis and processing. Then, use the Granger causality test to screen data features that have a strong causal relationship with the prediction target (i.e., the chip temperature of the BBU board). The key to this step is to identify and focus on factors that have a significant impact on chip temperature prediction.

[0077] Time series decomposition: For the data features with strong causality screened out in the data processing step, time series decomposition is performed to divide each data series into trend series, seasonal series and residual series. The purpose of this method is to enable the model to more accurately capture the cyclical and non-cyclical changes of the data and improve the accuracy of the prediction.

[0078] Step 3: Combine Figure 3 , build a hybrid forecasting model. Use the Autoregressive Integrated Moving Average (ARIMA) model to process and predict data characteristics with trend or seasonality. At the same time, deploy a random forest regression model to handle more complex nonlinear relationships and interactions between data. The outputs of these two models are fused through a logistic regression model to produce the final comprehensive forecast results.

[0079] Step 4: Make predictions using the hybrid prediction model.

[0080] Step 5: Integrate forecast results. The hourly forecast results obtained in step 4 are integrated into daily forecast results using mathematical methods (such as weighted average). This step provides a broader perspective to help the operation and maintenance team conduct long-term resource planning and temperature monitoring.

[0081] Combine the following Figure 1 The overall processing flow of the second embodiment is given.

[0082] Step 1: Collect data. This step involves automatically collecting hourly granularity hardware data indicators related to the BBU board, such as ambient temperature, ambient humidity, BBU board chip historical temperature, board CPU temperature, board location, power consumption, fan speed, etc. These data provide the necessary basic input for the prediction model, reflecting the operation of the board under different environmental conditions and operating states.

[0083] Step 2: Combine Figure 2 , feature processing, including: processing data and computing causality and time series decomposition.

[0084] Processing data: After the data is normalized and dimensionalized, transfer entropy is used to calculate the causality between different data indicators. Transfer entropy is a concept in information theory that is used to quantify the degree of influence of a system's state sequence on another system's state sequence. Through this method, features with high causal contribution to the prediction target (i.e., BBU board CPU temperature) are screened out.

[0085] Time series decomposition: Decompose the selected data features with high causality into time series, splitting each data series into trend series, seasonal series and residual series. This decomposition helps the subsequent model to more accurately capture the cyclical and non-cyclical changes of the data and provide support for complex predictions.

[0086] Step 3: Combine Figure 3 , constructing prediction models. The first model uses the Seasonal Autoregressive Integrated Moving Average (SARIMA) model, which is a widely used time series prediction model that can handle data with seasonal variations. The second model uses the Gradient Boosting Tree regression model, which is a powerful regression and classification technique in machine learning that can handle complex nonlinear relationships. The results of these two models are fused through a linear regression model to improve the accuracy and robustness of the prediction.

[0087] Step 4: Use the prediction model to make predictions.

[0088] Step 5: Integrate the forecast results. The hourly forecast results obtained in step 4 are integrated into a comprehensive forecast result at the daily granularity using mathematical methods (such as weighted average).

[0089] In two embodiments of the present invention, we show how to predict the CPU temperature of a BBU board through different methods and techniques. Although both embodiments revolve around temperature prediction, they differ significantly in the choice of data processing methods and prediction models. Embodiment 1 mainly uses the Granger causality test to screen features related to temperature prediction and uses ARIMA and random forest models for temperature prediction. This combination focuses on exploring the linear and nonlinear patterns of data from a time series perspective, as well as optimizing prediction results through model fusion. In contrast, Embodiment 2 introduces transfer entropy to calculate causality, which can more deeply reveal the dynamic transfer of information between variables, and is particularly suitable for processing complex system dynamics. In addition, Embodiment 2 uses SARIMA and gradient boosting tree models, which demonstrates the consideration of seasonal factors and emphasizes the processing of more complex data relationships. The final prediction results are also fused through a linear regression model to achieve more accurate predictions.

[0090] In summary, the first and second embodiments respectively demonstrate two different technical approaches to achieve the goal of predicting the temperature trend of hardware. The choice of which implementation method to choose depends on the specific application scenario, the available data type, and the required prediction accuracy.

[0091] Two examples are used to explain in detail how to perform temperature trend prediction from data collection to the integration of prediction results. It also shows the in-depth analysis and processing of data at different levels to improve the accuracy and reliability of predictions.

Claims

1. An intelligent hardware temperature prediction method, characterized in that It includes the following steps: S1, data collection; S2, feature processing, determining the key features that affect hardware temperature; S3. Construct a hybrid model, where the first model is a time series statistical model and the second model uses a regression model based on machine learning; S4, hybrid model prediction, the first model and the second model make predictions respectively, and the results of the two are fused through a linear regression model; S5: Integrate the prediction results, integrating the fine-grained prediction results into coarse-grained comprehensive prediction results.

2. The method according to claim 1, characterized in that The data collection includes collecting ambient temperature, humidity, hardware location, historical temperature data, business volume, hardware system age, cooling system information and the current and historical temperature data of the hardware itself.

3. The method according to claim 1, characterized in that The feature processing includes: S201, normalize the data and unify the dimensions of each indicator; S202, selecting features in the normalized data that have a strong causal relationship with the prediction target as key features for identifying the impact on the hardware temperature; S203, time series feature enhancement, decomposing the time series as follows: Y t =T t +S t +R t Among them, Y t is the time series to be decomposed at time t; T t is the trend decomposition sequence at time t, showing the long-term change trend of the sequence; S t is the seasonal decomposition sequence at time t, showing the short-term periodic changes of the sequence; R t It is the residual decomposition sequence at time t, showing the irregular or noise components of the sequence.

4. The method according to claim 3, characterized in that In S202, for a simple system, a key feature that affects the hardware temperature is selected through causal testing. The specific steps are: Fit model (1) and model (2) using time series X and time series γ: AND t =a0+a1Y t-1 +a2Y t-2 +…+∈ t (1) Y t =β0+β1Y t-1 +β2Y t-2 +…+γ1X t-1 +γ2X t-2 +…+∈ t (2) Among them, Y t represents the value of time series Y at time t, a0 represents the baseline level of time series Y in model (1), that is, the value of Y at time t when the past values ​​of the Y series itself are all zero, β0 represents the baseline level of time series Y in model (2) after considering the influence of the past values ​​of time series X, γ1 represents the influence coefficient of time series X at time t-1 on the value of Y at time t, X t-1 Indicates the value of the X sequence at time t-1. The time step is set according to the specific hardware operation and maintenance scenario; For the fitted models (1) and (2), the minimum mean square error (MSE) of the fitting is compared to determine whether model (2) is better than model (1); if the MSE of model (2) is less than the MSE of model (1), it means that feature X is the key feature of feature Y to be predicted.

5. The method according to claim 3, characterized in that In S202, for a complex system, the key features that affect the hardware temperature are selected by transferring entropy. The specific steps are: For time series X and time series Y, the transfer entropy T from X to Y X→Y The calculation method is: in, represents the past k states of the time series Y; represents the past I states of the time series ×; p() represents the joint probability distribution; If the transfer entropy T X→Y > 0, it means that the past state of time series X provides information for the future state of time series γ, that is, X is the key feature.

6. The method according to claim 3, characterized in that In S203, specifically: 1) Trend decomposition sequence T t Calculated by the following formula: Among them, 2m+1 represents the window length of the moving average; 2) Seasonal decomposition sequence S t Calculated by the following formula: S t =LOESS(Y t -T t ,λ) where λ is the time span parameter used by the LOESS smoother to capture the seasonal component; 3) Residual decomposition sequence R t Calculated by the following formula: R t =Y t -T t -S t 。 7. The method according to claim 6, characterized in that 2), the specific calculation steps of LOESS smoothing are as follows: 2-1) Calculate the local weighted fit: Y t =β0+β1(x j -x i )+β2(x j -x i ) 2 +∈ j Among them, β0 means when x i 、x j When they are equal, the response variable Y t The expected value of j -x i )’s first-order linear coefficient and second-order linear coefficient; x i 、x j Represents the fitting target points and local data points of local fitting; 2-2) Calculate the minimum mean square error: Among them, β mse represents a set of β0, β1, β2 found by minimizing the mean square error; weight Distance d ij =|x i -x j |,d i Represents the maximum distance within the local fitting range and defines the calculation weight w ij The size of the neighborhood used when ; 2-3) Get i The fitted value of : Indicates that in x i The fitted value y at i ; Repeat LOESS smoothing for each time instant t, and the seasonal decomposition series S t .

8. The method according to claim 1, characterized in that The first model is: autoregressive integrated moving average model ARIMA, ARIMA model with regression term ARIMAX or seasonal autoregressive integrated moving average model SARIMA; The second model is: random forest regression model or gradient boosted tree regression model.

9. The method according to claim 8, characterized in that The first model is the ARIMA model ARIMAX with regression terms, and the second model is the random forest regression model; the specific steps of S4 are: 1) Get the dual model prediction result at time t, Z t =[Y t,ARIMAX , Y t，RF ],in: Y t，ARIMAX =ARIMAX(X t ) Y t，RF =RF(X t ) Where, X t represents the time feature set obtained after S2 processing, Y t，ARIMAX Represents the ARIMAX model for X t The prediction result, Y t，RF Represents the random forest model for X t The prediction results, 2) Use the linear regression model to fuse the prediction results of the two models: P t =σ(β0+β1Y t，ARIMAX +β2Y t，RF ) Among them, P t is the fused prediction result at time t, β0, β1, β2 are logistic regression coefficients, 10. The method according to claim 1, characterized in that In S5, the fine-grained prediction results are integrated into a coarse-grained comprehensive prediction result.

Images