A fault classification method based on a BMC management system

Abstract

The application provides a fault classification method based on a BMC management system, a management system framework comprises a data acquisition module and a BMC management controller, the BMC management controller is integrated with a data preprocessing module, a data storage module and a PHM algorithm module; the method comprises the following steps: acquiring power consumption data in a corresponding system mode in real time; preprocessing the data, denoising the obtained original power consumption data, converting the original power consumption data into two-dimensional data through empirical decomposition, and aligning the data by zero padding; the preprocessed data enters the PHM algorithm module to obtain a classification result of the original data. The application proposes a data space conversion idea for the power consumption data scene, constructs a convolutional neural network based on a three-dimensional self-attention mechanism, classifies the converted data, can bring independence to a health management system, and the preprocessing method and the self-attention mechanism can greatly enhance the unstable power consumption data of an electronic system.

Description

Technical Field

[0001] This invention belongs to the field of electronic system fault monitoring technology based on machine learning, and particularly relates to a fault classification method based on a BMC management system. Background Technology

[0002] With the rapid development of science and technology, modern systems, from aerospace and autonomous driving to smart factories and cloud computing data centers, are becoming increasingly complex, posing serious challenges to maintaining system health. Traditionally, the maintenance methods for electronic systems are still based on periodic maintenance (PM). PM not only suffers from problems such as failing to fully utilize the lifespan of components, inducing "emerging" failures, and high maintenance costs, but also lacks the ability to detect, predict, and handle sudden failures. Therefore, condition-based maintenance (CBM) has received widespread attention. The concept of condition-based maintenance arose with the improvement of monitoring capabilities and processing levels of electronic systems. By acquiring various relevant parameters of the system, its health status can be inferred, and anomalies can be identified, thus determining whether equipment needs maintenance based on the health status. Compared to PM, CBM can significantly reduce maintenance costs, improve system reliability, and extend system lifespan.

[0003] BMC, short for Board Management Controller, is an embedded management microcontroller that operates independently of the main electronic system module. The BMC is a crucial support for CBM implementation and a core component of the server management system defined by the Intelligent Platform Management Interface (IPMI) protocol. It acts as a hardware manager integrated into servers, network devices, and other computer systems. The main functions of the BMC are monitoring device hardware status, performing remote management operations, and providing monitoring and control functions. The BMC can collect parameters such as voltage, current, fan speed, and CPU temperature from the electronic system platform through various types of sensors, and can also acquire parameter information from electronic system modules through serial communication. The acquired data can be processed locally and then transmitted via communication protocols, or it can be directly transmitted to a remote platform for processing. In current technology, the system uses distributed BMCs to collect system-wide health data and monitor status, while fault prediction, fault diagnosis, and health prediction are all completed in a central data center. This centralized processing approach not only poses a severe challenge to the system's computing power, but also suffers from problems such as large data transmission volumes and diagnostic latency. With the rise of edge computing and the enhancement of BMC's computing power, it is of practical significance to move fault prediction and health management functions forward. This can not only reduce the processing load of the central data center and increase the system's response speed, but also allow each submodule to deploy targeted fault prediction and fault diagnosis algorithms according to its characteristics, thereby improving the system's fault detection capabilities and increasing the accuracy of fault prediction.

[0004] Traditional BMC (Battery Management System) monitoring and alarm algorithms are insufficient. Current mainstream BMC technology is limited to providing threshold alarm functions for individual components of electronic modules. When a component, such as a fan, exceeds a threshold, remote management generates an alarm message. This does not meet the intelligent requirements of modern complex systems. Most of the alarm information generated in this way is redundant, which may lead to alarm failure—the real fault information is hidden among tens of thousands of false alarms.

[0005] With the rise of machine learning, PHM (Prognostic and Health Management) is also developing towards intelligence. Data shows that in the past decade, the number of papers on machine learning algorithms for PHM has gradually increased. PHM aims to assess the health of a system and provide functions such as fault detection and fault prediction. In electronic systems, system parameter information can reflect the system's state, and the failure of system components can be reflected in abnormalities in system parameters. Therefore, by analyzing system parameters, the health status of an electronic system can be assessed to a certain extent.

[0006] In existing technologies, there are precedents for combining the two for system health management. For example, in application number CN201911145557.3, entitled "A Fault Prediction Method and System Based on BMC Health Management Module," the BMC is used as an independent module for data acquisition. However, the PHM algorithm is not independent. As the health management unit of the electronic system, the BMC's data collection, processing, and transmission should be independent of the main operating module. This patented technology does not achieve complete independence between the BMC and the main module. Instead, it chooses to transmit the data received by the BMC to the database, while the data processing is left to the CPU of the main module. Such a fault detection algorithm cannot run when the electronic system actually malfunctions. The non-independent design of the BMC and CPU participating in the PHM algorithm calculation also violates the original intention of health management.

[0007] Existing solutions have failed to simultaneously meet the requirements of intelligence and independence, meaning that while satisfying intelligence, the independence of the computing platform must also be guaranteed to meet the needs of edge computing. Summary of the Invention

[0008] To address the above problems, this invention provides a fault classification method based on a BMC management system. The BMC management system framework includes a data acquisition module and a BMC management controller. The BMC management controller integrates a data preprocessing module, a data storage module, and a PHM algorithm module. The fault classification method includes the following steps:

[0009] S1, based on the data acquisition module, acquires power consumption data of multiple sub-modules of the electronic system in real time and reads it through the BMC management controller;

[0010] S2, based on the data preprocessing module, preprocesses the data. First, the original power consumption data is denoised, then it is transformed into two-dimensional data through empirical decomposition, and zero-padding is performed on the data to align all the data.

[0011] S3, the preprocessed data enters the PHM algorithm module to obtain the fault classification results of the original data; the PHM algorithm module is a three-dimensional self-attention convolutional neural network, including a three-dimensional convolutional layer, a three-dimensional self-attention TAM layer, a convolutional layer, a fully connected layer and a softmax layer, and is trained based on the dataset collected by the BMC management system.

[0012] S4 saves the raw data and fault classification results to the data storage module for data management and remote access.

[0013] Preferably, the data preprocessing in S2 includes data denoising, empirical decomposition, and zero-padding;

[0014] The denoised data 𝑥(i, 𝑡) is obtained based on data denoising. Then, according to the EMD decomposition method, it is transformed into the intrinsic mode space. The one-dimensional time-domain data of 𝑥(i, 𝑡) is converted into two-dimensional time-frequency domain data using the EMD method. The two-dimensional data is represented by E(i, t, m).

[0015] E(i,t,m) = EMD(𝑐(i,ə)) = {𝑐 𝑗 (i,falcon)∣𝑗∈{1, 2, …,𝑚}}∪{𝑟(i,falcon)}

[0016] Where EMD represents the empirical decomposition algorithm; j represents the index of the signal after EMD decomposition, 𝑐 𝑗 (𝑡) represents the j-th signal after EMD decomposition; m represents the number of signals obtained after EMD decomposition; 𝑟(𝑡) is the residual term after EMD signal decomposition, and i represents the current system mode;

[0017] Zero-padding is applied to the decomposed E(i,t,m):

[0018] First, find the maximum value of m for all E(i,t,m):

[0019] M = max(m i 0 <= i <= n

[0020] Where m i It is the number of decomposed signals obtained after EMD decomposition of the i-th fault;

[0021] For all E(i,t,m), after zero-padding, the data is represented as follows:

[0022] E(i,t,M) = E(i,t) = {𝑐 𝑗 (i,falcon)∣𝑗∈{1, 2,…, M}}∪{𝑟(i,falcon)}

[0023] i∈{0,1,2,…,n}

[0024] After processing, the original one-dimensional data of length L is transformed into two-dimensional data of length M x L.

[0025] Preferably, the specific data processing procedure of the PHM algorithm module is as follows:

[0026] S31, the preprocessed data E(i,t,M) is input into a 3x3 three-dimensional convolutional layer with W channels, transforming the M * L two-dimensional data into W * M * L three-dimensional data, which becomes the feature map F:

[0027] F = cov(E)

[0028] S32, input the obtained feature map into the three-dimensional self-attention TAM layer to obtain its output F';

[0029] S33, then another convolutional layer, reduces the number of channels to 1;

[0030] S34, return to S31 and repeat 5 times. After the loop ends, enter a fully connected layer.

[0031] S35 Finally, through the softmax layer, the probability of each fault mode is obtained, thereby achieving the fault classification goal. The dimension of softmax is equal to the size of the fault mode n plus a normal mode, i.e., n+1.

[0032] Preferably, the three-dimensional self-attention TAM layer specifically comprises:

[0033] For a feature map, let it be F(C,H,W);

[0034] F represents the feature map, and C, H, and W represent the lengths of the three dimensions of the feature map, respectively.

[0035] S321, perform average pooling on F(C,H,W) along the three directions C, H, and W respectively to obtain three feature vectors of length C, H, and W respectively:

[0036] V c = avepool(P (hxw) )

[0037] V c This represents the feature vector obtained by average pooling along the C direction; avepool means average pooling, which means that for each two-dimensional data of size HxW along the C-dimensional direction, the average value is calculated to obtain a scalar. Finally, after the pooling operation is completed along the C-dimensional direction, a vector of length C is obtained.

[0038] Similarly, perform the same operation on the other two dimensions to obtain the result V. h V w ;

[0039] S322, for the obtained V x (x∈C,H,W) enters two fully connected layers. The first fully connected layer has a dimension that is half the length of the vector, and the second layer has the same dimension as the vector, resulting in V'. x :

[0040] V' x = MLP(V x )

[0041] Get V' xThen, perform matrix multiplication on the original feature map F:

[0042] F x = V' x ⊗ F

[0043] Where V' x ⊗ represents the scalar result obtained after two fully connected layers, ⊗ represents matrix multiplication, and F represents the original feature map;

[0044] S323, the obtained F c F h F w Add them together to get the output F' of TAM.

[0045] F'=F c + F h + F w .

[0046] Preferably, the PHM algorithm module is trained based on the dataset collected by the BMC management system. The dataset creation process is as follows:

[0047] For sub-modules of the electronic system under test that fall under different fault categories, their corresponding power consumption data is collected, and their fault category is used as a label. Power consumption data of multiple sub-modules of the same system are collected, assigned different fault numbers, and trained together under the same model.

[0048] First, define different modes i for the original data: for i=0, it means the system has no fault and is in normal mode; for integers i>0, it means the system is in the i-th fault category. Assuming there are n fault categories, then the value of i is:

[0049] i∈{0,1, 2, …, n}

[0050] Let the original data obtained be: origin(𝑡)

[0051] origin(i, 𝑡) = sample(i)

[0052] Where origin (i, 𝑡) represents the sampled power consumption data that changes over time, sample represents the sampling of the power consumption data of the electronic system, and i represents the current state of the system;

[0053] For each original data origin (i, 𝑡), i represents the label of the original data, indicating the fault category corresponding to the original data;

[0054] origin(i, 𝑡) is a one-dimensional time series of length L, where the length L is determined by the sampling rate and the sampling time:

[0055] L = sample_rate x period

[0056] Where L represents the number of points sampled, i.e. the length of the time series, sample_rate represents the sampling rate, and period represents the sampling time.

[0057] Preferably, the raw data and the calculated results are input into the data storage module, which is in the form of a stack queue and is monitored and managed by a remote control unit.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] This invention combines Board-level Health Management System (BMC) and Power Consumption Management System (PHM) to assess the health status of electronic systems using power consumption data (current and voltage). The addition of BMC enhances the independence of the health management system, and a board-level health management system architecture is proposed based on this. A data space transformation approach for power consumption data scenarios is also proposed. Furthermore, a three-dimensional self-attention mechanism is proposed, upon which a neural network is constructed, and fault classification is achieved based on the transformed space.

[0060] Because power consumption data from electronic systems is inherently non-stationary, traditional machine learning methods struggle to achieve good results. This invention proposes a data spatial transformation method. It uses EMD (Electronic Data Modulation) to convert one-dimensional data into two-dimensional data, transforming the collected one-dimensional power consumption data into a two-dimensional time-frequency domain structure. Simultaneously, a three-dimensional self-attention mechanism module based on a convolutional neural network is constructed to classify faults in the spatially transformed data. Experiments show that this invention can bring independence to the health management system, and the preprocessing method and self-attention mechanism can significantly enhance the processing of non-stationary power consumption data from electronic systems. Attached Figure Description

[0061] Figure 1 This is the architecture diagram of the BMC management system of this invention.

[0062] Figure 2 This is a schematic diagram of the network structure of the PHM algorithm module.

[0063] Figure 3 This is a schematic diagram of a three-dimensional self-attention TAM layer network structure. Detailed Implementation

[0064] The specific implementation process of the present invention will be further described below with reference to specific embodiments.

[0065] The background of this invention is the fault classification of electronic systems based on power consumption data. Power consumption data refers to the general term for voltage and current data, and one of them can be used to replace power consumption data.

[0066] The management system of this invention includes a data acquisition module and a BMC management controller. The BMC management controller integrates a data preprocessing module, a data storage module, and a PHM algorithm module. The overall structure is as follows: Figure 1 As shown.

[0067] This invention uses the OpenBmc framework to build the BMC firmware operating system. The CPU is a Broadcom BCM2711, a quad-core Cortex-A72 (ARM v8) CPU, which belongs to the ARM architecture and is a 64-bit CPU.

[0068] Furthermore, in small subsystems, reducing the computational burden on the BMC's CPU should also be considered. In small electronic systems, the algorithm model should not be too large; development using C / C++ is optimal, and data should undergo preprocessing such as dimensionality reduction and feature selection to reduce the algorithm's computational load. Simultaneously, during the preprocessing stage, hardware structures such as FPGAs / DSPs can be incorporated to accelerate computationally intensive parts like neural networks and Fourier transforms, thus speeding up algorithm execution while reducing the load on the BMC's CPU.

[0069] BMC firmware is a Linux distribution that runs on a specific CPU. OpenBmc is an open-source framework for building this distribution. OpenBmc supports a range of CPUs to build the corresponding BMC firmware. By modifying the corresponding parameters using OpenBmc, you can obtain the corresponding customized version of BMC. Furthermore, the Redfish part needs to be integrated with the algorithm part, using the RESTful API specification to write corresponding programs to remotely transmit the algorithm's data and raw data to a remote data center.

[0070] 1. Training dataset

[0071] There are several ways to acquire data. The first is to acquire the data to be tested in the electronic system by reading the sensors of the electronic system sub-module. The second way is to acquire the power consumption data of the corresponding module of the electronic system by transmitting signals through serial port, IIC, SPI, bus, etc. after the electronic system under test reads the information.

[0072] Electronic systems can be divided into multiple sub-modules, including CPU modules, memory modules, and power supply modules. The data collected is for a specific module of the system under test, specifically the values ​​of its current and voltage sensors.

[0073] Training data: For each submodule of the electronic system under test, which falls under different fault categories, power consumption data is collected. The fault category is used as a label and as training data. Although faults and data collection are specific to submodules, the scheme collects power consumption data from multiple submodules of the same system, assigns them different fault numbers, and trains them together under the same model.

[0074] Define the following for the collected training data:

[0075] First, define different modes *i* for the original data: *i*=0 indicates the system has no fault and is in normal mode; integers *i*>0 indicate the system is in the *i*th fault category. Assuming we have *n* fault categories, then *i* takes values...

[0076] i∈{0,1, 2, … , n}

[0077] Let the original data obtained be: origin(𝑡)

[0078] origin(i, 𝑡) = sample(i)

[0079] Where origin(i, 𝑡) represents the sampled power consumption data that varies over time, sample indicates that the power consumption data of the electronic system is sampled, and i represents the current state of the system.

[0080] For each original data origin (i, 𝑡), i represents the label of the original data, indicating the fault category corresponding to that original data.

[0081] origin (i, 𝑡) is a one-dimensional time series of length L. The length L is determined by the sampling rate and the sampling time.

[0082] L = sample_rate x period

[0083] Where L represents the number of points sampled, i.e. the length of the time series, sample_rate represents the sampling rate, and period represents the sampling time.

[0084] Under different system states, the corresponding raw data is collected and preprocessed according to the preprocessing method.

[0085] For data of different fault categories, the preprocessed two-dimensional data is input into the model for training.

[0086] During training, the cross-entropy function was used as the loss function, and the Adam optimizer was used as the optimization function. Training continued until the model accuracy converged stationarily.

[0087] 2. Data Preprocessing Module

[0088] (1) Data noise reduction

[0089] The obtained raw data can be denoised. Noise refers to random fluctuations and outliers in the data. Noise can be reduced and data quality improved through moving averages, median filtering, Gaussian filtering, wavelet transforms, and machine learning methods such as denoising autoencoders. Noise reduction is not the focus of this invention, and various methods can be used. This invention uses the mean denoising method:

[0090] igh(i,falcon) = mean_denoise(origin(i,falcon),m)

[0091] mean_denoise represents the mean denoising method, m represents the sliding time window, mean_denoise calculates the mean of the data within a time window of size m and obtains the calculation result.

[0092] (2) Empirical decomposition

[0093] Based on the obtained 𝑥(i, 𝑡), it can be transformed into the intrinsic mode space using the EMD decomposition method. The one-dimensional time-domain data of 𝑥(i, 𝑡) is then converted into two-dimensional time-frequency domain data using the EMD method. The two-dimensional data is represented by E(i, t, m):

[0094] E(i,t,m) = EMD(𝑐(i,ə)) = {𝑐 𝑗 (i,falcon)∣𝑗∈{1, 2,…, 𝑚}}∪{𝑟(i,falcon)}

[0095] Where EMD represents the empirical decomposition algorithm; j represents the index of the signal after EMD decomposition, 𝑐 𝑗 (𝑡) represents the j-th signal after EMD decomposition. m represents the number of signals obtained after EMD decomposition. 𝑟(𝑡) is the residual term after EMD signal decomposition. i represents the current system mode.

[0096] (3) Zero-fill

[0097] Since different original data result in different numbers of signals m, for original data with fewer decomposed signals m, zero-padding is required for the decomposed E(i,t,m).

[0098] First, find the maximum value of m for all E(i,t,m):

[0099] M = max(m i 0 <= i <= n

[0100] Where m i It is the number of decomposed signals obtained after the i-th fault is decomposed in EMD.

[0101] For all E(i,t,m), after zero-padding, the data can be represented as

[0102] E(i,t,M) = E(i,t) = {𝑐 𝑗 (i,falcon)∣𝑗∈{1, 2,…, M}}∪{𝑟(i,falcon)}

[0103] i∈{0,1,2,…,n}

[0104] After processing, the original one-dimensional data of length L is transformed into two-dimensional data of length M x L.

[0105] 3. PHM Algorithm Module

[0106] A convolutional neural network based on a 3D self-attention module (TAM) is constructed, including a 3D convolutional layer, a TAM layer, convolutional layers, a fully connected layer, and a softmax layer, finally outputting the fault classification result. Two-dimensional data first passes through a 3D convolutional layer to increase channel dimension, then enters the 3D self-attention TAM layer, followed by convolutional layers, and finally reaches the fully connected layer. The fault classification result is then output through a softmax layer. The obtained algorithm result is transmitted to a remote control unit.

[0107] (1) Input E(i,t,M) into a 3x3 convolutional layer with W channels to transform the M * L two-dimensional data into W * M * L three-dimensional data, which becomes the feature map F.

[0108] F = cov(E)

[0109] (2) Input the obtained feature map into the three-dimensional self-attention module to obtain its output F';

[0110] (3) After another convolutional layer, the number of channels is reduced to 1;

[0111] (4) Return to (1) and repeat 5 times;

[0112] (5) After the loop ends, a fully connected layer is entered; through softmax, the probability of each fault mode is obtained, thereby achieving the fault classification goal. The dimension of softmax is equal to the size of the fault mode n plus a normal mode, i.e., n+1.

[0113] For the 3D self-attention TAM layer, the details are as follows:

[0114] For a feature map, let it be F(C,H,W);

[0115] F represents the feature map, and C, H, and W represent the lengths of the three dimensions of the feature map, respectively.

[0116] 1) Perform average pooling on F(C,H,W) along the C, H, and W directions respectively to obtain three feature vectors of length C, H, and W:

[0117] V c = avepool(P (hxw) )

[0118] V c This represents the feature vector obtained by average pooling along the C direction; avepool means average pooling, which means that for each two-dimensional data of size HxW along the C-dimensional direction, the average value is calculated to obtain a scalar. Finally, after the pooling operation is completed along the C-dimensional direction, a vector of length C is obtained.

[0119] Similarly, perform the same operation on the other two dimensions to obtain the result V. h V w ;

[0120] 2) For the obtained V x (x∈C,H,W) enters two fully connected layers. The first fully connected layer has a dimension that is half the length of the vector, and the second layer has the same dimension as the vector, resulting in V'. x :

[0121] V' x = MLP(V x )

[0122] Get V' x Then, perform matrix multiplication on the original feature map F:

[0123] F x = V' x ⊗ F

[0124] Where V' x ⊗ represents the scalar result obtained after two fully connected layers, ⊗ represents matrix multiplication, and F represents the original feature map;

[0125] 3) Obtain F c F h F w Add them together to get the output F' of TAM.

[0126] F'=F c + F h + F w .

[0127] 4. Data storage and remote control

[0128] The raw data and the calculated results are input into the data storage module, which is in the form of a stack queue, thus satisfying both the sequential and real-time requirements of the data.

[0129] The classification results obtained after calculation can be monitored and controlled remotely through various communication methods, such as HTTPS and sockets. This article uses Redfish and IPMI to establish communication with the remote control end. The remote control end obtains continuous and real-time data through different requests.

[0130] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0131] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A BMC management system based fault classification method, characterized in that: The BMC management system framework comprises a data acquisition module and a BMC management controller, and the BMC management controller is integrated with a data preprocessing module, a data storage module and a PHM algorithm module; the fault classification method comprises the following steps: S1, based on the data acquisition module, real-time acquisition of power consumption data of multiple sub-modules of the electronic system is performed, and the power consumption data is read through the BMC management controller; S2, based on the data preprocessing module, the data is preprocessed, first, the original power consumption data is denoised, then the data is decomposed by empirical decomposition into two-dimensional data, and the data is zero-padded to align all the data; S3, the preprocessed data enters the PHM algorithm module to obtain a fault classification result of the original data; the PHM algorithm module is a convolutional neural network based on three-dimensional self-attention, comprising a three-dimensional convolutional layer, a three-dimensional self-attention TAM layer, a convolutional layer, a fully connected layer and a softmax layer, and is trained based on a data set collected by the BMC management system; The three-dimensional self-attention TAM layer is specifically: For a feature map, F(C, H, W); F represents the feature map, and C, H and W represent the lengths of the three dimensions of the feature map in sequence; S321, average pooling operations are performed on F(C, H, W) along the C, H and W directions respectively, and three feature vectors with lengths of C, H and W are obtained respectively: V c = avepool(P (hxw) ) V c avepool represents average pooling, i.e. for each two-dimensional data with size HxW along the C dimension, an average value operation is performed to obtain a scalar, and finally, after the pooling operation is completed for all C dimension directions, a vector with length C is obtained; The same operation is done for the other two dimensions, resulting in V h , w ; S322, for the V x , x e C, H, W, into two layers of fully connected layers, the first layer of fully connected layer dimension size is half of the vector length, the second layer dimension and the vector length are the same, V' x : V’ x = MLP(V x ) V' = V + F x After that, matrix multiplication operation is performed on the original feature map F: F x = V’ x F where V x represents the resulting scalar after passing through two fully connected layers, represents matrix multiplication, and F represents the original feature map; S323, the resulting F c , F h , F w are added to obtain the output F' of the TAM F' = F c + F h + F w ; S4, the original data and the fault classification result are saved to the data storage module for data management and remote calling.

2. The method of claim 1, wherein the BMC management system is a baseboard management controller (BMC) management system. The data preprocessing in S2 comprises data denoising, empirical decomposition and zero padding; Obtaining denoising data based on data denoising (i, ) into intrinsic mode space according to an EMD decomposition method, and converting (i, ) one-dimensional time domain data into two-dimensional time domain frequency domain data, and the two-dimensional data is represented as E(i,t,m). where EMD represents the Empirical Mode Decomposition algorithm; j represents the index of the signal after EMD decomposition, j ( ) represents the jth signal after EMD decomposition; m represents the number of signals obtained after EMD decomposition; (i, ) is the residual term after EMD signal decomposition, where i represents the current mode of the system; Zero padding is performed on the decomposed E(i, t, m): First, the maximum value of m of all E(i, t, m) is found: M = max(m i ) 0<=i<=n wherein m i is the number of decomposition signals obtained after EMD decomposition of the ith fault; After zero padding, the data of all E(i, t, m) is represented as: E(i, t, M) = E(i, t) = { j (i, )∣ j ∈{1, 2,…, M}}∪{ (i, )} i∈{0,1, 2,…, n} After processing, the original one-dimensional data with a length of L becomes two-dimensional data with a length of M x L.

3. The method of claim 1, wherein the BMC management system is a baseboard management controller (BMC) management system. The specific data processing process of the PHM algorithm module is as follows: S31, the preprocessed data E(i, t, M) is input into a three-dimensional convolutional layer with W channels and a size of 3x3, so that the two-dimensional data with a length of M x L is changed into three-dimensional data with a length of W x M x L, becoming a feature map F: F = cov(E) S32, the obtained feature map is input into the three-dimensional self-attention TAM layer to obtain the output F'; S33, after passing through a convolutional layer, the number of channels is changed to 1; S34, return to S31 and repeat 5 times, and after the loop ends, enter a fully connected layer; S35, finally, pass through a softmax layer to obtain the probability of each fault mode, so as to realize the fault classification target, and the dimension of the softmax is equal to the size n of the fault mode plus one normal mode, i.e. n+1.

4. The method of claim 1, wherein: The PHM algorithm module is trained based on a data set collected by the BMC management system, and the data set is prepared as follows: The sub-modules of the electronic system under different fault categories are collected to obtain the corresponding power consumption data, and the fault category thereof is taken as a label; the power consumption data of multiple sub-modules of the same system are collected, and different fault numbers are assigned; and the sub-modules are trained together under the same model. First, set different modes i of the original data: for i = 0, it means that the system has no fault and is in normal mode, and for i > 0, it means that the system is in the i-th fault category, assuming that there are n fault categories, then i takes the value of: i∈{0,1, 2, …, n} Let the acquired original data be: origin( ) origin(i, ) = sample(i) Wherein, origin (i, ) represents the sampled power consumption data according to time variation, sample represents sampling power consumption data of the electronic system, i represents the current state of the system; For each original data origin (i, ), i represents the label of the original data, representing which fault category the original data corresponds to; origin (i, ) is a one-dimensional time series of length L, the length L being determined by the sampling rate and the sampling time: L = sample_rate x period Where L represents how many points are sampled, that is, the length of the time series, sample_rate represents the sampling rate, and period represents the sampling time.

5. The method for fault classification based on BMC management system of claim 1, wherein: For the original data and the calculated results, they are input into the data storage module, which is in the form of a stack queue, and is monitored and managed through the remote control unit.

Images