Management system and method for intelligent computing platform

Abstract

The present application discloses a management system and method for an intelligent computing platform, which collects the real-time temperature and fan speed of the intelligent computing platform through real-time monitoring, and introduces data processing and analysis algorithms based on artificial intelligence and deep learning to perform time series dynamic analysis and feature interactive response association coding on the real-time temperature and fan speed, so as to learn and capture the implicit correlation and interactive synergy characteristics between the real-time temperature and fan speed of the computing platform, so as to adaptively control the fan speed based on the dynamic changes of the real-time temperature of the intelligent computing platform. In this way, intelligent temperature management of the computing platform can be realized, and the instability problems caused by traditional threshold and linear control methods can be avoided. At the same time, the influence of hysteresis problems caused by dynamic loads can also be reduced, so as to ensure that the intelligent computing platform is within a safe operating temperature range, and provide a more reliable and stable intelligent computing platform.

Description

Technical Field

[0001] The present application relates to the field of intelligent management, and more specifically, to a management system and method for an intelligent computing platform. Background Art

[0002] With the rapid development of information technology, the application of intelligent computing platforms is becoming more and more widespread. These platforms play an important role in data processing, cloud computing and other fields. However, with the improvement of computing power, the heating problem of hardware has become increasingly prominent. Excessive temperature not only affects system performance, but also may cause hardware failure, data loss and system crash. Therefore, temperature management has become a key factor to ensure the stable operation of intelligent computing platforms.

[0003] However, traditional temperature management methods often rely on preset temperature thresholds, that is, they rely on simple threshold judgments and linear control algorithms. When the temperature exceeds a certain limit, the system will passively take corresponding measures, such as increasing the fan speed to control the temperature. This method cannot effectively cope with the complexity and dynamics of temperature changes. When faced with nonlinear and dynamic loads, it may cause response lag or over-adjustment, resulting in inaccurate temperature management. In addition, when the threshold method is actually used for heat dissipation control, the cooling fan may start and stop frequently due to instantaneous temperature fluctuations, affecting the stability of the equipment and the service life of the fan.

[0004] Therefore, an optimized management system for the intelligent computing platform is desired. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a management system and method for an intelligent computing platform, which collects the real-time temperature and fan speed of the intelligent computing platform through real-time monitoring, and introduces data processing and analysis algorithms based on artificial intelligence and deep learning to perform time series dynamic analysis and feature interactive response association coding for the real-time temperature and fan speed, so as to learn and capture the implicit correlation and interactive synergy characteristics between the real-time temperature and fan speed of the computing platform, so as to adaptively control the fan speed based on the dynamic changes of the real-time temperature of the intelligent computing platform. In this way, the intelligent temperature management of the computing platform can be realized, avoiding the instability problems caused by traditional threshold and linear control methods, and at the same time, it can also reduce the impact of the hysteresis problems caused by dynamic loads, so as to ensure that the intelligent computing platform is within a safe operating temperature range, providing a more reliable and stable intelligent computing platform.

[0006] According to one aspect of the present application, a management system for an intelligent computing platform is provided, which includes:

[0007] A real-time temperature data acquisition module is used to obtain the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor;

[0008] The fan speed data acquisition module is used to obtain the time series of the speed value of the cooling fan collected by the speed sensor;

[0009] A data parameter time series encoding module, used for performing time series encoding on the time series of the real-time temperature and the time series of the rotational speed value respectively to obtain a sequence of real-time temperature local time series associated implicit feature vectors and a sequence of rotational speed local time series associated implicit feature vectors;

[0010] A time series node feature propagation aggregation module, used to input the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors into a node feature propagation network based on a node energy decay mechanism to obtain a real-time temperature time series propagation aggregation representation vector and a speed time series propagation aggregation representation vector;

[0011] A real-time temperature-rotation speed timing interaction response module, used for performing timing feature interaction response processing on the real-time temperature timing propagation aggregation representation vector and the rotation speed timing propagation aggregation representation vector to obtain a real-time temperature-rotation speed timing interaction response representation vector;

[0012] The speed real-time control module is used to determine whether the speed value at the current time point should increase, decrease or remain unchanged based on the real-time temperature-speed time series interaction response representation vector.

[0013] According to another aspect of the present application, a management method for an intelligent computing platform is provided, which includes:

[0014] Obtain the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor;

[0015] Obtaining a time series of speed values ​​of the cooling fan collected by a speed sensor;

[0016] Performing time series coding on the time series of the real-time temperature and the time series of the rotational speed value respectively to obtain a sequence of local time series associated implicit feature vectors of the real-time temperature and a sequence of local time series associated implicit feature vectors of the rotational speed;

[0017] Inputting the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors into a node feature propagation network based on a node energy decay mechanism to obtain a real-time temperature time series propagation aggregate representation vector and a speed time series propagation aggregate representation vector;

[0018] Performing time series feature interactive response processing on the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector to obtain a real-time temperature-speed time series interactive response representation vector;

[0019] Based on the real-time temperature-rotation speed time series interaction response representation vector, it is determined whether the rotation speed value at the current time point should increase, decrease or remain unchanged.

[0020] Compared with the prior art, the present application provides a management system and method for an intelligent computing platform, which collects the real-time temperature and fan speed of the intelligent computing platform through real-time monitoring, and introduces data processing and analysis algorithms based on artificial intelligence and deep learning to perform time series dynamic analysis and feature interactive response association coding on the real-time temperature and fan speed, so as to learn and capture the implicit correlation and interactive synergy characteristics between the real-time temperature and fan speed of the computing platform, so as to adaptively control the fan speed based on the dynamic changes of the real-time temperature of the intelligent computing platform. In this way, intelligent temperature management of the computing platform can be realized, avoiding the instability problems caused by traditional threshold and linear control methods, and at the same time reducing the impact of hysteresis problems caused by dynamic loads, so as to ensure that the intelligent computing platform is within a safe operating temperature range, providing a more reliable and stable intelligent computing platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 A block diagram of a management system of an intelligent computing platform according to an embodiment of the present application;

[0023] Figure 2 A data flow diagram of a management system of an intelligent computing platform according to an embodiment of the present application;

[0024] Figure 3 The present invention is a flowchart of a method for managing an intelligent computing platform according to an embodiment of the present application.

[0025] Figure 4 Design a schematic diagram for the AI ​​server structure. DETAILED DESCRIPTION

[0026] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0027] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0028] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0029] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0030] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0031] Traditional temperature management methods often rely on preset temperature thresholds, that is, they rely on simple threshold judgments and linear control algorithms. When the temperature exceeds a certain limit, the system will passively take corresponding measures, such as increasing the fan speed to control the temperature. This method cannot effectively cope with the complexity and dynamics of temperature changes. When faced with nonlinear and dynamic loads, it may cause response lag or over-adjustment, resulting in inaccurate temperature management. In addition, when the threshold method is actually used for heat dissipation control, the cooling fan may be frequently started and stopped due to instantaneous temperature fluctuations, affecting the stability of the equipment and the service life of the fan. Therefore, an optimized management system for the intelligent computing platform is desired.

[0032] In the technical solution of the present application, a management system for an intelligent computing platform is proposed. Figure 1 A block diagram of a management system of an intelligent computing platform according to an embodiment of the present application. Figure 2FIG. 1 is a data flow diagram of a management system of an intelligent computing platform according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the management system 300 of the intelligent computing platform according to the embodiment of the present application includes: a real-time temperature data acquisition module 310, which is used to obtain the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor; a fan speed data acquisition module 320, which is used to obtain the time series of the speed value of the cooling fan collected by the speed sensor; a data parameter timing encoding module 330, which is used to respectively perform timing encoding on the time series of the real-time temperature and the time series of the speed value to obtain a sequence of local timing-related implicit feature vectors of the real-time temperature and a sequence of local timing-related implicit feature vectors of the speed; a timing node feature propagation aggregation module 340, which is used to convert the local timing of the real-time temperature into a sequence of implicit feature vectors of the real-time temperature; The sequence of sequence-associated implicit feature vectors and the sequence of speed local time-series associated implicit feature vectors are respectively input into the node feature propagation network based on the node energy decay mechanism to obtain the real-time temperature time-series propagation aggregation representation vector and the speed time-series propagation aggregation representation vector; the real-time temperature-speed time-series interactive response module 350 is used to perform time-series feature interactive response processing on the real-time temperature time-series propagation aggregation representation vector and the speed time-series propagation aggregation representation vector to obtain the real-time temperature-speed time-series interactive response representation vector; the speed real-time control module 360 ​​is used to determine whether the speed value at the current time point should increase, decrease or remain unchanged based on the real-time temperature-speed time-series interactive response representation vector.

[0033] In particular, the real-time temperature data acquisition module 310 and the fan speed data acquisition module 320 are used to obtain the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor; and to obtain the time series of the speed value of the cooling fan collected by the speed sensor. Among them, the temperature sensor is a device that converts temperature into an electrical signal. The speed sensor is a device that measures the speed of a rotating machine.

[0034] In particular, the data parameter timing encoding module 330 is used to perform timing encoding on the time series of the real-time temperature and the time series of the speed value, respectively, to obtain a sequence of local timing-related implicit feature vectors of real-time temperature and a sequence of local timing-related implicit feature vectors of speed. In a specific example of the present application, the time series of the real-time temperature and the time series of the speed value are input into a sequence encoder based on a bidirectional gated cyclic unit to obtain a sequence of local timing-related implicit feature vectors of real-time temperature and a sequence of local timing-related implicit feature vectors of speed. Considering that both temperature and speed data are time series data and have a time dependency, that is, the real-time temperature of the intelligent computing platform and the speed value of the cooling fan have a dynamic change law of time series and time series correlation characteristics in the time dimension. Based on this, in the technical solution of the present application, the time series of the real-time temperature and the time series of the speed value are input into a sequence encoder based on a bidirectional gated recurrent unit for encoding, so as to respectively extract the timing correlation feature information of the real-time temperature of the intelligent computing platform and the cooling fan speed in the time dimension, thereby obtaining a sequence of local timing correlation implicit feature vectors of the real-time temperature and a sequence of local timing correlation implicit feature vectors of the speed.

[0035] In particular, the time series node feature propagation aggregation module 340 is used to input the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors into the node feature propagation network based on the node energy decay mechanism to obtain the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector. It should be understood that for the real-time temperature, its change in the time dimension is a time series process with temporal continuity and correlation. Each of the real-time temperature local time series associated implicit feature vectors captures the local pattern and trend of temperature change in a local time period during the operation of the intelligent computing platform. There are correlations and features based on the global time series between the real-time temperature time series features in different local time periods. Therefore, in order to be able to effectively aggregate these local time series features of real-time temperature on the intelligent computing platform to form a more comprehensive representation of the real-time temperature time series changes, in the technical solution of the present application, the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors are further input into the node feature propagation network based on the node energy decay mechanism to obtain the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector. Through the processing of the node feature propagation network based on the node energy decay mechanism, the node energy decay process between the real-time temperature local time series features in each local time period can be simulated, and the time-space aggregation representation of the node features can be realized, thereby improving the model's ability to understand the full time domain characteristics and change patterns of the real-time temperature.

[0036] In an embodiment of the present application, the sequence of the real-time temperature local timing associated implicit feature vectors and the sequence of the speed local timing associated implicit feature vectors are respectively input into a node feature propagation network based on a node energy decay mechanism to obtain a real-time temperature timing propagation aggregate representation vector and a speed timing propagation aggregate representation vector, including: first, based on the maximum value, average value and variance of each real-time temperature local timing associated implicit feature vector in the sequence of the real-time temperature local timing associated implicit feature vectors, the node energy statistical paradigm value of each real-time temperature local timing associated implicit feature vector is calculated to obtain a sequence of real-time temperature local timing associated node energy statistical paradigm values, wherein the real-time temperature local timing node energy statistical paradigm value corresponding to the current real-time temperature local timing associated implicit feature vector in the sequence of the real-time temperature local timing node energy statistical paradigm value is used as the current node energy statistical paradigm value, and other real-time temperature local timing node energy statistical paradigm values ​​are used as historical node energy statistical paradigm values ​​to obtain the current real-time temperature local timing node point energy statistical paradigm values ​​and the sequence of historical real-time temperature local time series node energy statistical paradigm values; then, counting the node propagation space span values ​​between each other real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vector and the current real-time temperature local time series associated implicit feature vector to obtain a sequence of real-time temperature local time series node propagation space span values; and based on the sequence of real-time temperature local time series node propagation space span values ​​and the sequence of historical real-time temperature local time series node energy statistical paradigm values, determining the node energy propagation attenuation coefficient values ​​of each other real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vector to obtain a sequence of real-time temperature local time series node energy propagation attenuation coefficient values; that is, by counting the propagation space span values ​​between different temperature local time series node feature vectors, the topological structure and time series space relationship between nodes, that is, the time series node association characteristics between each real-time temperature local time series associated implicit feature vector. These span values ​​are combined with the node energy statistical paradigm values ​​of the historical real-time temperature local time series to determine the node energy propagation attenuation coefficient of the real-time temperature local time series. This coefficient is inversely correlated with the historical node energy of the real-time temperature local time series, simulating the attenuation characteristics of each real-time temperature local time series node energy over time and space.Furthermore, using the sequence of the energy propagation attenuation coefficient values ​​of the real-time temperature local time series nodes as the weight sequence, the weighted sum of all other real-time temperature local time series associated implicit feature vectors in the sequence of the real-time temperature local time series associated implicit feature vectors is calculated to obtain the historical real-time temperature local time series node energy decay time series aggregate feature vector; that is, using these real-time temperature local time series decay coefficients as weights, that is, considering the energy decay effect, the weighted sum of the node feature vector sequence is further calculated based on the energy statistical paradigm value of the real-time temperature local time series to generate the historical node energy decay time series aggregate feature vector of the real-time temperature local time series, so as to aggregate the feature information of the real-time temperature local time series historical nodes. Finally, based on the energy statistical paradigm value of the current real-time temperature local time series node, the weighted sum of the historical real-time temperature local time series node energy decay time series aggregate feature vector and the current real-time temperature local time series associated implicit feature vector is calculated to obtain the real-time temperature time series propagation aggregate representation vector. In the technical solution of the present application, by fusing the energy statistical paradigm value of the current node of the real-time temperature local time series, the weighted sum of the historical real-time temperature local time series aggregated feature vector and the current real-time temperature local time series node feature vector is calculated to generate a real-time temperature time series propagation aggregated representation vector, which integrates the information of time series dynamics and spatial structure. In summary, the node feature propagation network based on the node energy decay mechanism dynamically evaluates the energy level of each real-time temperature local time series node feature, captures the topological structure between nodes, simulates energy propagation decay, and aggregates time series space features. This method has significant advantages in improving the model's sensitivity to the time series dynamics of real-time temperature local time series node features, enhancing feature expression, and adapting to complex network structures. In addition, by dynamically adjusting the node energy weights and aggregated time series space features, the model can more accurately identify and process the real-time temperature local time series key information in the data, thereby achieving better performance in the subsequent cooling fan speed adaptive control task.

[0037] Among them, the process of calculating the node energy statistical normal form value of each real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vector based on the maximum value, average value and variance of each real-time temperature local time series associated implicit feature vector to obtain a sequence of real-time temperature local time series node energy statistical normal form values ​​includes: calculating the maximum value, average value and variance of the real-time temperature local time series associated implicit feature vector to obtain the real-time temperature local time series maximum value, the real-time temperature local time series average value and the real-time temperature local time series variance; calculating the sum of the real-time temperature local time series variance and the regularization term hyperparameter and then multiplying it with the constant 4 to obtain the first real-time temperature local time series node energy statistical factor; respectively calculate the product of the real-time temperature local timing variance and the regularization term hyperparameter and the constant 2 to obtain the double-modulated real-time temperature local timing variance and the double-modulated regularization term hyperparameter; calculate the square of the difference between the real-time temperature local timing maximum value and the real-time temperature local timing average value, and then add it with the double-modulated regularization term hyperparameter and the double-modulated real-time temperature local timing variance to obtain a second real-time temperature local timing node energy statistical factor; calculate the division between the first real-time temperature local timing node energy statistical factor and the second real-time temperature local timing node energy statistical factor to obtain the real-time temperature local timing node energy statistical paradigm value. More specifically, based on the sequence of the real-time temperature local time series node propagation space span values ​​and the sequence of the historical real-time temperature local time series node energy statistical paradigm values, the process of determining the node energy propagation attenuation coefficient values ​​of each other real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vector to obtain the sequence of real-time temperature local time series node energy propagation attenuation coefficient values ​​includes: using each real-time temperature local time series node propagation space span value in the sequence of the real-time temperature local time series node propagation space span value as an exponential power, calculating the natural constant e as the base The exponential function value of is used to obtain the sequence of spatial span values ​​supported by the real-time temperature local timing node propagation class; the positional sum of the sequence of spatial span values ​​supported by the real-time temperature local timing node propagation class and the sequence of spatial span values ​​of the real-time temperature local timing node propagation is calculated to obtain the sequence of spatial span modulation coefficients of the real-time temperature local timing node propagation; the positional division between the sequence of historical real-time temperature local timing node energy statistical paradigm values ​​and the sequence of spatial span modulation coefficients of the real-time temperature local timing node propagation is calculated to obtain the sequence of energy propagation attenuation coefficient values ​​of the real-time temperature local timing node.

[0038] In summary, in the above embodiment, the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors are respectively input into the node feature propagation network based on the node energy decay mechanism to obtain the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector, including: inputting the sequence of the real-time temperature local time series associated implicit feature vectors into the node feature propagation network based on the node energy decay mechanism and processing them with the following node feature propagation formula to obtain the real-time temperature time series propagation aggregate representation vector; wherein, the node feature propagation formula is:

[0039] X={x1,x2,...x k-1 ,x k}

[0040]

[0041]

[0042] Where X is the sequence of the real-time temperature local time series associated implicit feature vectors, x k-1 and x k are respectively the (k-1)th and kth real-time temperature local time series associated implicit feature vectors in the sequence of the real-time temperature local time series associated implicit feature vectors, x i is the i-th real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vectors, is the sth position eigenvalue in the i-th real-time temperature local time series associated implicit eigenvector, L is the length of the i-th real-time temperature local time series associated implicit eigenvector, μ i is the mean of the implicit eigenvector associated with the local temporal sequence of the i-th real-time temperature, σ 2 i is the variance of the implicit feature vector associated with the local time series of the i-th real-time temperature, ∈ is the regularization hyperparameter, max(x i ) represents the maximum value of the implicit feature vector of the local temporal association of the i-th real-time temperature. is the node energy statistical norm value of the i-th real-time temperature local time series associated implicit feature vector, Count(·) represents the node propagation space span value, α, β, γ and δ are trainable hyperparameters, and v1 is the real-time temperature time series propagation aggregation representation vector.

[0043] In particular, the real-time temperature-speed timing interaction response module 350 is used to perform timing feature interaction response processing on the real-time temperature timing propagation aggregation representation vector and the speed timing propagation aggregation representation vector to obtain the real-time temperature-speed timing interaction response representation vector. In the process of temperature management for the intelligent computing platform, temperature and speed are two important dynamic indicators, which reflect the thermal state of the system and the heat dissipation capacity of the cooling fan, respectively. However, there are complex timing correlation relationships and implicit interactions between the real-time temperature timing characteristics of the intelligent computing platform and the cooling fan speed timing characteristics. Therefore, in order to learn and capture the dynamic interaction information and implicit correlation characteristics between the system thermal state timing characteristics and the cooling fan heat dissipation capacity timing characteristics, help the temperature management system identify the mutual influence of the two under different conditions, and provide a basis for the subsequent cooling fan speed control, in the technical solution of the present application, the real-time temperature timing propagation aggregation representation vector and the speed timing propagation aggregation representation vector are further input into the feature interaction response module based on the adaptive distinguishable mechanism to obtain the real-time temperature-speed timing interaction response representation vector. It is worth mentioning that the feature interaction response module based on the adaptive distinguishable mechanism can more deeply and comprehensively understand the potential correlation and implicit interaction between the real-time temperature time series propagation aggregation features and the speed time series propagation aggregation features, providing a more in-depth and detailed control basis for the temperature management system.

[0044] In an embodiment of the present application, the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector are subjected to time series feature interaction response processing to obtain a real-time temperature-speed time series interaction response representation vector, including: firstly calculating the position-by-position response between the speed time series propagation aggregation representation vector and the real-time temperature time series propagation aggregation representation vector to obtain a real-time temperature-speed time series position-by-position response feature vector; in this way, the local time series responsiveness characteristics and interaction between the real-time temperature time series aggregation information and the speed time series aggregation information can be effectively captured. Subsequently, the real-time temperature-speed time series position-by-position response feature vector is normalized using the Softmax function to obtain a normalized real-time temperature-speed time series position-by-position response feature vector; that is, the real-time temperature-speed time series position-by-position responses are normalized using the Softmax function to form a probability distribution, thereby ensuring the standardization of the feature response and providing preparation for the subsequent gating function weight screening. The normalization process not only balances the scale of the feature response, but also makes the expression of the real-time temperature-speed time series features in the model more balanced. Furthermore, the normalized real-time temperature-speed timing position-by-position response feature vector is input into a learnable gating function to obtain a real-time temperature-speed timing response screening weight mask vector; it should be understood that the model can adaptively learn and output the real-time temperature-speed timing response screening weight mask vector, and this dynamic weighting process realizes the selection and reinforcement of the timing features between the real-time temperature and the speed through the gating mechanism, and dynamically adjusts the importance of the feature position. This step is the core of feature optimization because it directly affects the sensitivity and distinguishing ability of the model to the real-time temperature-speed timing feature response. Then, the position point multiplication between the real-time temperature-speed timing response screening weight mask vector and the normalized real-time temperature-speed timing position-by-position response feature vector is calculated to obtain the real-time temperature-speed timing position-by-position response distinguishable weight mask vector; in this way, the weight distribution is further refined, and precise weight control is provided for the final real-time temperature-speed timing interaction fusion saliency representation and optimization. This position-by-position weight refinement ensures that the model can assign the most appropriate weight to each position of the real-time temperature-speed timing responsiveness feature, thereby more effectively capturing useful information in the data. Finally, the position-by-position dot multiplication between the real-time temperature-speed timing position-by-position response distinguishable weight mask vector and the real-time temperature-speed timing position-by-position response feature vector is calculated to obtain the real-time temperature-speed timing interactive response representation vector. That is, by performing a dot multiplication of the real-time temperature-speed timing position-by-position response distinguishable weight mask vector and the real-time temperature-speed timing position-by-position response feature vector, a real-time temperature-speed timing interactive response representation vector is generated.This step integrates the feature response after weight adjustment and generates the final real-time temperature-speed time series interaction response representation vector after significant expression enhancement, which significantly improves the model's ability to capture key real-time temperature-speed time series interaction responsiveness information and provides a basis for the subsequent adaptive control of the cooling fan speed value. Through the processing of the feature interaction response module based on the adaptive distinguishable mechanism, the expression ability of the real-time temperature-speed time series correlation feature and the overall performance of the model can be significantly improved through fine responsiveness analysis and adaptive weight adjustment at each position.

[0045] Among them, the process of using the Softmax function to normalize the real-time temperature-speed timing position-by-position response feature vector to obtain the normalized real-time temperature-speed timing position-by-position response feature vector includes: using the negative number of the feature value of each position in the normalized real-time temperature-speed timing position-by-position response feature vector as a power, calculating the natural exponential function value with the natural constant e as the base to obtain the normalized real-time temperature-speed timing position-by-position response class support feature vector; calculating the inverse of the sum of the feature value of each position in the normalized real-time temperature-speed timing position-by-position response class support feature vector and the constant one to obtain the real-time temperature-speed timing response screening weight mask vector.

[0046] In summary, in the above embodiment, the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector are subjected to time series feature interaction response processing to obtain the real-time temperature-speed time series interaction response representation vector, including: inputting the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector into the feature interaction response module based on the adaptive distinguishable mechanism to process them with the following feature interaction response formula to obtain the real-time temperature-speed time series interaction response representation vector; wherein the feature interaction response formula is:

[0047] v r =v2 / v1

[0048] v n =softmax(v r )

[0049]

[0050] v s =v n ⊙v a

[0051]

[0052] Wherein, v1 and v2 represent the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector respectively, v ris the real-time temperature-speed time series position-by-position response feature vector, softmax(·) represents the softmax function, v n is the normalized real-time temperature-speed time series position-by-position response feature vector, exp(·) is the value of the natural exponential function, and v a The weight mask vector is used to filter the real-time temperature-speed timing response. ⊙ represents the point multiplication by position, v s The real-time temperature-speed timing position-by-position response can be distinguished by the weight mask vector, The real-time temperature-rotation speed timing interaction response represents a vector.

[0053] In particular, the speed real-time control module 360 ​​is used to determine whether the speed value at the current time point should increase, decrease or remain unchanged based on the real-time temperature-speed time series interactive response representation vector. In a specific example of the present application, the real-time temperature-speed time series interactive response representation vector is input into a temperature management result generator based on a classifier to obtain a temperature management result, and the temperature management result is used to indicate whether the speed value at the current time point should increase, decrease or remain unchanged. That is, the interactive response information between the real-time temperature time series aggregation feature of the intelligent computing platform and the speed time series aggregation feature of the cooling fan is used for classification processing, so as to adaptively control the fan speed based on the real-time temperature dynamic change of the intelligent computing platform. In this way, the intelligent temperature management of the computing platform can be realized, the instability problem caused by the traditional threshold and linear control method can be avoided, and the influence of the hysteresis problem caused by the dynamic load can also be reduced, so as to ensure that the intelligent computing platform is within a safe operating temperature range and provide a more reliable and stable intelligent computing platform.

[0054] Preferably, inputting the real-time temperature-rotation speed time series interaction response representation vector into a classifier-based temperature management result generator to obtain a temperature management result comprises:

[0055] Calculate the first norm and second norm of the mean vector of the real-time temperature time series propagation aggregation representation vector and the rotation speed time series propagation aggregation representation vector;

[0056] Calculate the weighted sum of the inverse of the square root of the second norm and the first norm, and perform point multiplication with the point-added sum vector of the real-time temperature timing propagation aggregation representation vector and the rotation speed timing propagation aggregation representation vector to obtain a first real-time temperature-rotation speed timing interaction response syndrome vector;

[0057] Perform a dot multiplication of the dot product vector of the real-time temperature time series propagation aggregation representation vector and the rotation speed time series propagation aggregation representation vector and the square root of the length of the real-time temperature-rotation speed time series interaction response representation vector to obtain a second real-time temperature-rotation speed time series interaction response syndrome vector;

[0058] Calculating a weighted sum of the first real-time temperature-rotation speed timing interaction response correction subvector and the second real-time temperature-rotation speed timing interaction response correction subvector to obtain a real-time temperature-rotation speed timing interaction response correction vector;

[0059] Performing a dot multiplication of the real-time temperature-rotation speed timing interaction response correction vector and the real-time temperature-rotation speed timing interaction response representation vector to obtain a corrected real-time temperature-rotation speed timing interaction response representation vector; and

[0060] The corrected real-time temperature-rotation speed time series interaction response representation vector is input into a temperature management result generator based on a classifier to obtain a temperature management result.

[0061] The real-time temperature-speed timing interaction response correction vector is expressed as:

[0062]

[0063] V1 and V2 are respectively the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector, V μ is its mean vector, L is the length of the real-time temperature-speed time series interaction response representation vector, and α and β are weighted sums of the hyperparameters. (Complete)

[0064] Here, considering that the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector respectively represent the energy attenuation time series propagation characteristics of the real-time temperature and speed values ​​based on the bidirectional gated cyclic time series encoding on the local time series nodes, when they perform feature interaction response based on the adaptive distinguishable mechanism, it is expected to further compensate for the imbalance of response domain interaction offset caused by the imbalance of correspondence ratio under the fine-grained distribution of time series, so as to improve the expression effect of the real-time temperature-speed time series interaction response representation vector, thereby improving the accuracy of the classification results.

[0065] Therefore, based on the constrained representation of the structured foreground and background distinction of the superfluid of the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector based on the mean vector norm between the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector, the feature-level key correspondence between the feature vectors of the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector is modeled, and the global correlation relationship between the corresponding features is adjusted, so as to make positive fine-grained correspondence suggestions through the unbalanced proportion control between the corresponding eigenvalues ​​of the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector, so as to avoid the imbalance of response domain offset between the feature vectors of the real-time temperature time series propagation aggregate representation vector and the speed time series propagation aggregate representation vector by focusing on the focus.

[0066] In this way, by correcting the real-time temperature-speed time series interaction response representation vector through the real-time temperature-speed time series interaction response correction vector, the expression effect of the real-time temperature-speed time series interaction response representation vector can be improved, thereby improving the accuracy of the temperature management result obtained by the temperature management result generator based on the classifier input. In this way, the fan speed can be more accurately adaptively controlled based on the real-time temperature dynamic changes of the intelligent computing platform, thereby realizing the intelligent temperature management of the computing platform.

[0067] As described above, the management system 300 of the intelligent computing platform according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a management algorithm of the intelligent computing platform. In a possible implementation, the management system 300 of the intelligent computing platform according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the management system 300 of the intelligent computing platform can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the management system 300 of the intelligent computing platform can also be one of the many hardware modules of the wireless terminal.

[0068] Alternatively, in another example, the management system 300 of the intelligent computing platform and the wireless terminal may also be separate devices, and the management system 300 of the intelligent computing platform may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0069] Furthermore, a management method for an intelligent computing platform is also provided.

[0070] Figure 3 FIG. 1 is a flow chart of a management method of an intelligent computing platform according to an embodiment of the present application. Figure 3As shown, according to the management method of the intelligent computing platform of the embodiment of the present application, the steps include: S1, obtaining the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor; S2, obtaining the time series of the speed value of the cooling fan collected by the speed sensor; S3, respectively performing time encoding on the time series of the real-time temperature and the time series of the speed value to obtain a sequence of local time series associated implicit feature vectors of real-time temperature and a sequence of local time series associated implicit feature vectors of speed; S4, respectively inputting the sequence of local time series associated implicit feature vectors of real-time temperature and the sequence of local time series associated implicit feature vectors of speed into the node feature propagation network based on the node energy decay mechanism to obtain a real-time temperature time series propagation aggregation representation vector and a speed time series propagation aggregation representation vector; S5, performing time series feature interactive response processing on the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector to obtain a real-time temperature-speed time series interactive response representation vector; S6, based on the real-time temperature-speed time series interactive response representation vector, determining whether the speed value at the current time point should increase, decrease or remain unchanged.

[0071] In summary, the management method of the intelligent computing platform according to the embodiment of the present application is explained, which collects the real-time temperature and fan speed of the intelligent computing platform through real-time monitoring, and introduces data processing and analysis algorithms based on artificial intelligence and deep learning to perform time series dynamic analysis and feature interactive response association coding on the real-time temperature and fan speed, so as to learn and capture the implicit correlation and interactive synergy characteristics between the real-time temperature and fan speed of the computing platform, so as to adaptively control the fan speed based on the dynamic changes of the real-time temperature of the intelligent computing platform. In this way, the intelligent temperature management of the computing platform can be realized, avoiding the instability problems caused by traditional threshold and linear control methods, and at the same time reducing the impact of hysteresis problems caused by dynamic loads, so as to ensure that the intelligent computing platform is within a safe operating temperature range, providing a more reliable and stable intelligent computing platform.

[0072] In one example, Figure 4 This is a schematic diagram of the AI ​​server structure design. The AI ​​server has a total power consumption of 2700W, and a single chip can consume up to 500W. The server uses a dual-row eight-fan cooling structure, and the management system can flexibly adjust the speed to meet energy-saving and environmentally friendly operation requirements.

[0073] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A management system for an intelligent computing platform, characterized in that: include: A real-time temperature data acquisition module is used to obtain the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor; The fan speed data acquisition module is used to obtain the time series of the speed value of the cooling fan collected by the speed sensor; A data parameter time series encoding module, used for performing time series encoding on the time series of the real-time temperature and the time series of the rotational speed value respectively to obtain a sequence of real-time temperature local time series associated implicit feature vectors and a sequence of rotational speed local time series associated implicit feature vectors; A time series node feature propagation aggregation module, used to input the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors into a node feature propagation network based on a node energy decay mechanism to obtain a real-time temperature time series propagation aggregation representation vector and a speed time series propagation aggregation representation vector; A real-time temperature-rotation speed timing interaction response module, used for performing timing feature interaction response processing on the real-time temperature timing propagation aggregation representation vector and the rotation speed timing propagation aggregation representation vector to obtain a real-time temperature-rotation speed timing interaction response representation vector; A speed real-time control module, used to determine whether the speed value at the current time point should be increased, decreased or kept unchanged based on the real-time temperature-speed time series interaction response representation vector; Among them, the data parameter timing encoding module is used to: input the time series of the real-time temperature and the time series of the speed value into a sequence encoder based on a bidirectional gated cyclic unit to obtain a sequence of implicit feature vectors associated with the local timing of the real-time temperature and a sequence of implicit feature vectors associated with the local timing of the speed.

2. The management system of the intelligent computing platform according to claim 1, characterized in that: The time series node feature propagation aggregation module includes: A node energy statistical paradigm value calculation unit, used for calculating the node energy statistical paradigm value of each real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vector based on the maximum value, average value and variance of each real-time temperature local time series associated implicit feature vector to obtain a sequence of real-time temperature local time series node energy statistical paradigm values, wherein the real-time temperature local time series node energy statistical paradigm value corresponding to the current real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series node energy statistical paradigm value is used as the current node energy statistical paradigm value, and other real-time temperature local time series node energy statistical paradigm values ​​are used as historical node energy statistical paradigm values ​​to obtain a sequence of the current real-time temperature local time series node energy statistical paradigm value and historical real-time temperature local time series node energy statistical paradigm values; A node propagation space span value statistics unit is used to count the node propagation space span values ​​between each other real-time temperature local time series associated implicit feature vector in the sequence of the real-time temperature local time series associated implicit feature vector and the current real-time temperature local time series associated implicit feature vector to obtain a sequence of real-time temperature local time series node propagation space span values; A node energy propagation attenuation coefficient value calculation unit is used to determine the node energy propagation attenuation coefficient values ​​of other real-time temperature local time series associated implicit feature vectors in the sequence of real-time temperature local time series associated implicit feature vectors based on the sequence of real-time temperature local time series node propagation space span values ​​and the sequence of historical real-time temperature local time series node energy statistical paradigm values ​​to obtain a sequence of real-time temperature local time series node energy propagation attenuation coefficient values; A node energy attenuation time series aggregation feature calculation unit is used to calculate the weighted sum of all other real-time temperature local time series associated implicit feature vectors in the sequence of the real-time temperature local time series associated implicit feature vectors using the sequence of the real-time temperature local time series node energy propagation attenuation coefficient values ​​as a weight sequence to obtain a historical real-time temperature local time series node energy attenuation time series aggregation feature vector; The real-time temperature time series propagation aggregation representation unit is used to calculate the weighted sum of the historical real-time temperature local time series node energy attenuation time series aggregation feature vector and the current real-time temperature local time series associated implicit feature vector based on the current real-time temperature local time series node energy statistical norm value to obtain the real-time temperature time series propagation aggregation representation vector.

3. The management system of the intelligent computing platform according to claim 2, characterized in that: The node energy statistical paradigm value calculation unit is used to: Calculating the maximum value, average value and variance of the real-time temperature local time series associated implicit feature vector to obtain the real-time temperature local time series maximum value, the real-time temperature local time series average value and the real-time temperature local time series variance; Calculate the sum of the real-time temperature local time series variance and the regularization term hyperparameter and then multiply it by a constant 4 to obtain a first real-time temperature local time series node energy statistical factor; Calculate the product of the local time series variance of the real-time temperature and the regularization term hyperparameter and a constant 2 respectively to obtain the double-modulated local time series variance of the real-time temperature and the double-modulated regularization term hyperparameter; After calculating the square of the difference between the maximum value of the real-time temperature local timing and the average value of the real-time temperature local timing, the square is added to the double modulation regularization term hyperparameter and the double modulation real-time temperature local timing variance to obtain a second real-time temperature local timing node energy statistical factor; The division between the first real-time temperature local timing node energy statistics factor and the second real-time temperature local timing node energy statistics factor is calculated to obtain a real-time temperature local timing node energy statistics normal value.

4. The management system of the intelligent computing platform according to claim 3, characterized in that: The node energy propagation attenuation coefficient value calculation unit is used to: Taking each real-time temperature local time series node propagation space span value in the sequence of the real-time temperature local time series node propagation space span value as an exponential power, calculating an exponential function value with a natural constant e as a base to obtain a sequence of real-time temperature local time series node propagation class support space span values; Calculate the positional sum of the sequence of the real-time temperature local time series node propagation class support space span values ​​and the sequence of the real-time temperature local time series node propagation space span values ​​to obtain a sequence of real-time temperature local time series node propagation space span modulation coefficients; The positional division between the sequence of the historical real-time temperature local timing node energy statistical norm values ​​and the sequence of the real-time temperature local timing node propagation space span modulation coefficients is calculated to obtain the sequence of the real-time temperature local timing node energy propagation attenuation coefficient values.

5. The management system of the intelligent computing platform according to claim 4, characterized in that: The real-time temperature-speed timing interactive response module includes: A real-time temperature-rotation speed timing position-by-position response unit, used for calculating the position-by-position response between the rotation speed timing propagation aggregate representation vector and the real-time temperature timing propagation aggregate representation vector to obtain a real-time temperature-rotation speed timing position-by-position response feature vector; A normalization unit, used for normalizing the real-time temperature-rotation speed time series position-by-position response feature vector using a Softmax function to obtain a normalized real-time temperature-rotation speed time series position-by-position response feature vector; A real-time temperature-rotation speed timing gating response unit, used for inputting the normalized real-time temperature-rotation speed timing position-by-position response feature vector into a learnable gating function to obtain a real-time temperature-rotation speed timing response screening weight mask vector; A real-time temperature-speed timing position-by-position response distinguishable weight mask unit, used for calculating the position-by-position point multiplication between the real-time temperature-speed timing response screening weight mask vector and the normalized real-time temperature-speed timing position-by-position response feature vector to obtain a real-time temperature-speed timing position-by-position response distinguishable weight mask vector; The real-time temperature-speed timing interaction response representation unit is used to calculate the position point multiplication between the real-time temperature-speed timing position-by-position response distinguishable weight mask vector and the real-time temperature-speed timing position-by-position response feature vector to obtain the real-time temperature-speed timing interaction response representation vector.

6. The management system of the intelligent computing platform according to claim 5, characterized in that: The normalization unit is used for: Taking the negative number of each position characteristic value in the normalized real-time temperature-speed time series position-by-position response characteristic vector as a power, calculating the value of a natural exponential function with a natural constant e as a base to obtain a normalized real-time temperature-speed time series position-by-position response class support characteristic vector; The inverse of the sum of the feature value of each position in the normalized real-time temperature-rotation speed timing position-by-position response class support feature vector and a constant one is calculated to obtain the real-time temperature-rotation speed timing response screening weight mask vector.

7. The management system of the intelligent computing platform according to claim 6, characterized in that: The real-time speed control module is used to: input the real-time temperature-speed timing interaction response representation vector into a temperature management result generator based on a classifier to obtain a temperature management result, and the temperature management result is used to indicate whether the speed value at the current time point should increase, decrease or remain unchanged.

8. A method for managing an intelligent computing platform, using the management system of the intelligent computing platform according to claim 1, characterized in that: include: Obtain the time series of the real-time temperature of the intelligent computing platform collected by the temperature sensor; Obtaining a time series of the speed values ​​of the cooling fan collected by the speed sensor; Performing time series coding on the time series of the real-time temperature and the time series of the rotational speed value respectively to obtain a sequence of local time series associated implicit feature vectors of the real-time temperature and a sequence of local time series associated implicit feature vectors of the rotational speed; Inputting the sequence of the real-time temperature local time series associated implicit feature vectors and the sequence of the speed local time series associated implicit feature vectors into a node feature propagation network based on a node energy decay mechanism to obtain a real-time temperature time series propagation aggregate representation vector and a speed time series propagation aggregate representation vector; Performing time series feature interactive response processing on the real-time temperature time series propagation aggregation representation vector and the speed time series propagation aggregation representation vector to obtain a real-time temperature-speed time series interactive response representation vector; Based on the real-time temperature-rotation speed time series interaction response representation vector, it is determined whether the rotation speed value at the current time point should increase, decrease or remain unchanged.

Images