Fault classification method based on BMC management system

Abstract

The invention 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, and a data preprocessing module, a data storage module and a PHM algorithm module are integrated on the BMC management controller; the method comprises the following steps: acquiring power consumption data in a corresponding system mode in real time; the method comprises the steps of preprocessing data, performing noise reduction on the obtained original power consumption data, performing empirical decomposition to obtain two-dimensional data, performing zero filling on the data to enable the data to be aligned, enabling the preprocessed data to enter a PHM algorithm module, and obtaining a classification result of the original data. The invention provides a data space conversion thought aiming at a power consumption data scene, constructs a convolutional neural network based on a three-dimensional self-attention mechanism at the same time, performs fault classification on the converted data, can bring independence to a health management system, and improves the reliability of the health management system. The preprocessing method and the self-attention mechanism can greatly enhance unstable power consumption data of the electronic system.

Description

Technical Field

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

[0002] With the vigorous development of science and technology, nowadays, from aerospace, autonomous driving, intelligent factories to cloud computing data centers, modern systems have become increasingly complex, posing severe challenges to the maintenance of system health status. Traditionally, the maintenance method for electronic systems is still based on periodic maintenance (PM). PM not only has problems such as being unable to fully utilize the service life of components, giving rise to "nascent" faults, and high maintenance costs, but also lacks the capabilities of detecting, predicting, and handling sudden faults. Therefore, condition-based maintenance (CBM) has received extensive attention, and the concept of CBM was proposed along with the improvement of the monitoring ability and processing level of electronic systems. By obtaining various relevant parameters of the system, the health status of the system can be inferred therefrom, and the abnormalities of the system can be found, and then it is decided whether the equipment needs to be maintained according to the health status. Compared with PM, CBM can greatly reduce the maintenance cost, improve the system reliability, and extend the service life of the system.

[0003] The BMC, short for Board Management Controller, is an embedded management microcontroller that operates independently of the main module of the electronic system. The BMC is an important support for realizing CBM and is a core component in the server management system defined by the Intelligent Platform Management Interface (IPMI) protocol. It is a hardware manager integrated in servers, network devices, and other computer systems. The main functions of the BMC are to monitor the hardware status of devices, execute remote management operations, and provide monitoring and control functions for devices. The BMC can collect parameters such as voltage, current, fan speed, and CPU temperature of the electronic system platform through various types of sensors, and can also obtain parameter information of the electronic system module through communication such as serial ports. The acquired data can be processed locally and then sent through a communication protocol, or directly sent to a remote platform for processing through a communication protocol. In current technologies, the system collects and monitors the health data of the entire system through distributed BMCs, while fault prediction, fault diagnosis, and health prediction are all completed in the central data center. This centralized processing method not only poses a severe challenge to the computing power of the system, but also has problems such as large data transmission volume and diagnostic delay. With the rise of edge computing and the enhancement of the computing power of BMCs, it is of practical significance to move the fault prediction and health management functions forward. This can not only reduce the processing load of the central data center, increase the reaction speed of the system, but also deploy targeted fault prediction and fault diagnosis algorithms according to the characteristics of each sub-module, improve the fault detection ability of the system, and increase the accuracy of fault prediction.

[0004] The monitoring and alarm algorithm of traditional BMCs is insufficient. The current mainstream BMC technology is only limited to providing threshold alarm functions for individual components of electronic modules. When a component, such as a fan, when the speed of the fan exceeds the threshold, a warning message is sent for remote management, which does not meet the intelligent requirements of modern complex systems. Most of the warning messages in this way are 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 is also developing in the direction of intelligence. Data shows that in the past decade, the number of papers on machine learning algorithms for PHM has gradually increased; PHM, short for Prognostics and Health Management, aims to evaluate the health of the system, provide functions such as fault detection and fault prediction. In an electronic system, the parameter information of the system can reflect the state of the system, and the faults of the components of the system can be reflected in the abnormalities of the system parameters. Therefore, through the analysis of the system parameters, the health state of the electronic system can be evaluated to a certain extent.

[0006] In the prior art, there have been precedents for combining the two for system health management. For example, in the solution with the application number CN201911145557.3 and the application name "A Fault Prediction Method and System Based on BMC Health Management Module", BMC is used as an independent module for data collection, but the PHM algorithm is not independent. As the health management unit of the electronic system, BMC should be independent from the main operation module in terms of collecting data, processing data, and transmitting data. However, this patented technology does not achieve complete independence between BMC and the main module. Instead, it chooses to transmit the data received by BMC to the database, and the data processing is handed over to the CPU of the main module. Such a fault detection algorithm cannot run when the electronic system actually fails. The non-independent design of BMC and CPU participating in the PHM algorithm operation together also violates the original intention of health management.

[0007] In the existing solution attempts, there is no one that can meet the requirements of both intelligence and independence, that is, while meeting intelligence, it is also necessary to ensure the independence of the computing carrier to meet the needs of edge computing. Summary of the Invention

[0008] In view of the above problems, the present invention provides a fault classification method based on the BMC management system. The BMC management system framework includes 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 fault classification method includes the following steps: S1, based on the data acquisition module, real-time acquire the power consumption data of multiple sub-modules of the electronic system, and read it through the BMC management controller; S2, preprocess the data based on the data preprocessing module. First, denoise the original power consumption data, then decompose it into two-dimensional data through empirical decomposition, and zero-fill the data to align all data; S3, the preprocessed data enters the PHM algorithm module to obtain the fault classification result of the original data; the PHM algorithm module is a convolutional neural network based on three-dimensional self-attention, 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; S4, save the original data and the fault classification result to the data storage module for data management and remote call.

[0009] Preferably, the data preprocessing in S2 includes data denoising, empirical decomposition, and zero filling; Noise reduction data 𝑥(i,𝑡) is obtained based on data noise reduction. According to the EMD decomposition method, it is transformed into the intrinsic mode space. Through the EMD method, the one-dimensional time-domain data of 𝑥(i,𝑡) is transformed into two-dimensional time-domain frequency-domain data, and the two-dimensional data is represented by E(i,t,m): E(i,t,m) = EMD(𝑥(i,𝑡)) = {𝑐 𝑗 (i,𝑡)∣𝑗∈{1, 2, …,𝑚}}∪{𝑟(i,𝑡)} where EMD represents the empirical decomposition algorithm; j represents the subscript 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 mode of the system; Zero-padding is performed on the decomposed E(i,t,m): First, find the maximum value of m for all E(i,t,m): M = max(m i ) 0<=i<=n where m i is the number of decomposed signals obtained after EMD decomposition of the i-th fault; For all E(i,t,m), after zero-padding, the data is represented as: E(i,t,M) = E(i,t) = {𝑐 𝑗 (i,𝑡)∣𝑗∈{1, 2,…, M}}∪{𝑟(i,𝑡)} i∈{0,1, 2,…, n} After processing, the original one-dimensional data with length L becomes two-dimensional data of M x L.

[0010] Preferably, the specific data processing process of the PHM algorithm module is as follows: S31, Input the preprocessed data E(i,t,M) into a 3x3 three-dimensional convolutional layer with W channels, and transform the two-dimensional data of M * L into three-dimensional data of W * M * L, which becomes the feature map F: F = cov(E) S32, Input the obtained feature map into the three-dimensional self-attention TAM layer to obtain its output F’; S33, Then pass through another convolutional layer to change the number of channels to 1; S34, Return to S31 and repeat 5 times. After the loop ends, enter a fully connected layer; Finally, through the softmax layer, the probabilities of each fault mode are obtained, thus achieving the fault classification goal. The dimension of softmax is equal to the size n of the fault mode plus one normal mode, that is, n + 1.

[0011] Preferably, the three-dimensional self-attention TAM layer is specifically: For a feature map, denoted as 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, perform average pooling operations on F(C, H, W) along the C, H, and W directions respectively, and obtain three feature vectors with lengths of C, H, and W: V c = avepool(P (hxw) ) V c represents the feature vector obtained by performing average pooling on the C direction; avepool represents average pooling, that is, for each two-dimensional data of size HxW along the C dimension direction, perform an average value operation to obtain a scalar. Finally, after all pooling operations on the C dimension direction are completed, a vector with a length of C is obtained; Similarly, perform the same operations on the other two dimensions to obtain the results V h 、V w ; S322, for the obtained V x (x ∈ C, H, W) enter two fully connected layers. The dimension size of the first fully connected layer is half of the vector length, and the dimension of the second layer is the same as the vector length, obtaining V’ x : V’ x = MLP(V x ) After obtaining V’ x , perform a matrix multiplication operation on the original feature map F: F x = V’ x ⊗ F where V’ x represents the result scalar obtained after passing through two fully connected layers, ⊗ represents matrix multiplication, and F represents the original feature map; S323, add the obtained F c 、F h 、F w to obtain the output F’ of TAM F’ = F c + F h + F w 。

[0012] Preferably, the PHM algorithm module is trained based on the dataset collected by the BMC management system. The process of making the dataset is as follows: For sub-modules of the electronic system to be detected under different fault categories, collect their corresponding power consumption data. The fault category it is in serves as a label. Collect the power consumption data of multiple sub-modules of the same system, assign different fault numbers, and participate in training together under the same model; First, set different modes i of the original data: for i = 0, it means the system has no fault and is in the normal mode. For integers i > 0, it means the system is in the i-th fault category. Suppose there are n fault categories, then the value range of i is: i ∈ {0, 1, 2, …, n} Let the obtained original data be: origin(𝑡) origin(i, 𝑡) = sample(i) Where origin (i, 𝑡) represents the sampled power consumption data that changes with time, sample represents sampling the power consumption data of the electronic system, and i represents the current state of the system; For each original data origin (i, 𝑡), its i represents the label of the original data, indicating which fault category the original data corresponds to; origin (i, 𝑡) is a one-dimensional time series with a length of L. The length L is 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.

[0013] Preferably, the original data and the calculated results are input into the data storage module. The form of the data storage module is a stack queue, and it is monitored and managed by the remote control unit.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention combines BMC and PHM to evaluate the health status of the electronic system through power consumption data (current and voltage) on the electronic system, increasing the independence of BMC for the health management system, and based on this, a board-level health management system architecture is proposed. At the same time, a data space conversion idea for the power consumption data scenario is also proposed. And based on this, a three-dimensional self-attention mechanism is proposed. Based on this mechanism, a neural network is constructed, and fault classification is realized according to the neural network for the converted space.

[0015] Due to the non - stationary characteristics of the power consumption data of electronic systems, traditional machine learning methods are difficult to achieve good results on it. The present invention proposes a method for data space conversion. By using EMD, one - dimensional data is converted into two - dimensional data, and the one - dimensional power consumption data collected is subjected to space conversion to make it into a two - dimensional time - frequency domain structure. At the same time, a three - dimensional self - attention mechanism module based on a convolutional neural network is constructed to classify faults for the data after space conversion. Experiments show that the present invention can bring independence to the health management system, and the pre - processing method and the self - attention mechanism can greatly enhance the non - stationary power consumption data of electronic systems. Brief Description of the Drawings

[0016] Figure 1 It is the architecture diagram of the BMC management system of the present invention.

[0017] Figure 2 It is the schematic diagram of the network structure of the PHM algorithm module.

[0018] Figure 3 It is the schematic diagram of the network structure of the three - dimensional self - attention TAM layer. Detailed Embodiments

[0019] The following further explains the specific implementation process of the present invention in combination with specific embodiments.

[0020] The background of the present invention is the fault classification for power consumption data of electronic systems. Power consumption data refers to the general term of voltage and current data, and one of them can be taken to replace the power consumption data.

[0021] The management system of the present invention includes a data acquisition module and a BMC management controller. The BMC management controller is integrated with a data pre - processing module, a data storage module, and a PHM algorithm module. The overall structure is as Figure 1 shown.

[0022] The present invention uses the OpenBmc framework to build the BMC firmware operating system. The CPU is Broadcom BCM2711, a quad - core Cortex - A72 (ARM v8), which is a 64 - bit CPU belonging to the ARM architecture.

[0023] In addition, if in a small subsystem, considering the lightweight calculation of the CPU as the BMC, it should also be taken into account. In a small - scale electronic system, the algorithm model should not be too large. It is best to develop using C / C++ language, and the data should also be subjected to certain pre - processing such as dimensionality reduction and feature selection to reduce the computational amount of the algorithm. At the same time, in the pre - processing stage, hardware structures such as FPGA / DSP can be added to perform hardware acceleration on fixed parts with high computational amounts such as neural networks and Fourier transform parts, reducing the burden on the BMC's CPU while accelerating the operation of the algorithm.

[0024] BMC firmware is a Linux distribution running on a specific CPU. OpenBmc is an open-source framework for building this distribution. OpenBmc supports a series of CPUs to build the corresponding BMC firmware for that CPU. By modifying the corresponding parameters using OpenBmc, the corresponding customized version of BMC can be obtained. In addition, for the redfish part, it needs to be combined with the algorithm part, and corresponding programs are written using the Restful Api specification to remotely transmit the data of the algorithm and the original data to a remote data center.

[0025] 1. Training dataset There are various ways to obtain data. The first way is to obtain the data to be detected of the electronic system by reading the sensors of the electronic system sub-module. The second way is that after the electronic system to be detected reads the information, through serial port, IIC, SPI, bus and other ways, the power consumption data of the corresponding module of the electronic system is obtained by signal transmission.

[0026] The electronic system can be divided into multiple sub-modules, including the cpu module, memory module, power supply module. The data is for a certain module of the system to be detected, and the values of its current and voltage sensors are collected.

[0027] Training data: For the sub-modules of the electronic system to be detected in different fault categories, collect their corresponding power consumption data. The fault category it is in is used as a label and used as training data. Although the faults and data collection are for sub-modules, the scheme collects the power consumption data of multiple sub-modules of the same system, assigns different fault numbers, and participates in training together under the same model.

[0028] Definition for the collected training data: First, set different modes i for the original data: For i = 0, it means the system has no fault and is in the normal mode. For integers i > 0, it means the system is in the i-th fault category. Assume we have n fault categories, then the value range of i is i ∈ {0, 1, 2, …, n} Let the obtained original data be: origin(𝑡) origin(i, 𝑡) = sample(i) Among them, origin (i, 𝑡) represents the sampled power consumption data changing with time, and sample represents sampling the power consumption data of the electronic system. i represents the current state of the system.

[0029] For each original data origin (i, 𝑡), its i represents the label of the original data, indicating which fault category the original data corresponds to origin(i, 𝑡) is a one-dimensional time series with a length of L. The length L is determined by the sampling rate and the sampling time.

[0030] L = sample_rate x period Where L represents the number of sampled points, that is, the length of the time series, sample_rate represents the sampling rate, and period represents the sampling time.

[0031] Under different system states, collect the corresponding original data and perform preprocessing according to the preprocessing method.

[0032] For data of different fault categories, use the two-dimensional data obtained after preprocessing and input it into the model for training During training, the loss function used is the cross-entropy function, and the optimization function used is the adam optimizer. Train until the model accuracy converges stably. 2. Data Preprocessing Module (1) Data Denoising For the obtained original data, denoising can be performed. Noise refers to random fluctuations and outliers in the data. It can reduce noise and improve the quality of the data through methods such as moving average, median filtering, Gaussian filtering, wavelet transform, and machine learning methods such as denoising autoencoders. Denoising is not the focus of this invention, and multiple methods can be used. The method used in this invention is mean denoising: 𝑥(i, 𝑡) = mean_denoise(origin(i, 𝑡), m) mean_denoise represents the mean denoising method, m represents the sliding time window, and mean_denoise calculates the mean value of the data within the time window of size m to obtain the calculation result.

[0033] (2) Empirical Mode Decomposition According to the obtained 𝑥(i, 𝑡), it can be transformed into the intrinsic mode space according to the EMD decomposition method. Through the EMD method, the one-dimensional time-domain data of 𝑥(i, 𝑡) is transformed into two-dimensional time-domain frequency-domain data. The two-dimensional data is represented by E(i,t,m): E(i,t,m) = EMD(𝑥(i, 𝑡)) = {𝑐 𝑗 (i, 𝑡)∣𝑗∈{1, 2,…, 𝑚}}∪{𝑟(i, 𝑡)} Where EMD represents the empirical mode decomposition algorithm; j represents the subscript 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 mode of the system.

[0034] (3)Zero-padding Due to different original data, the number of decomposed signals m is different. Therefore, for the original data with a relatively small number of decomposed signals m, zero-padding needs to be performed on the decomposed E(i,t,m).

[0035] First, find the maximum value of m for all E(i,t,m): M = max(m i ) 0 <= i <= n where m i is the number of decomposed signals obtained after EMD decomposition of the i-th fault.

[0036] For all E(i,t,m), after zero-padding, the data can be expressed as E(i,t,M) = E(i,t) = {c 𝑗 (i,t) | j ∈ {1, 2,…, M}} ∪ {r(i,t)} i ∈ {0,1, 2,…, n} After processing, the original one-dimensional data with length L becomes two-dimensional data of M x L.

[0037] 3. PHM algorithm module Construct a convolutional neural network based on a three-dimensional self-attention module (TAM), including a three-dimensional convolutional layer, a TAM layer, a convolutional layer, a fully connected layer, and a softmax layer, and finally output the fault classification result. The two-dimensional data first passes through the three-dimensional convolutional layer to increase the channel dimension and become three-dimensional data, then enters the three-dimensional self-attention TAM layer, then passes through the convolutional layer, and finally reaches the fully connected layer. And the fault classification result is output through softmax. The obtained algorithm result is transmitted to the remote control unit.

[0038] (1)Input E(i,t,M) into the 3x3 convolutional layer with W channels, and turn the two-dimensional data of M * L into three-dimensional data of W * M * L, becoming the feature map F F = cov(E) (2)Input the obtained feature map into the three-dimensional self-attention module to obtain its output F’; (3)Then pass through another convolutional layer to change the number of channels to 1; (4)Go back to (1) and repeat 5 times; After the loop ends, enter a fully connected layer; through softmax, the probabilities of each fault mode are obtained, thus achieving the fault classification goal. The dimension of softmax is equal to the size n of the fault mode plus one normal mode, that is, n + 1; For the three-dimensional self-attention TAM layer, it is as follows: For a feature map, denoted as 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; 1) Perform average pooling operations on F(C, H, W) along the C, H, and W directions respectively to obtain three feature vectors with lengths of C, H, and W: V c = avepool(P (hxw) ) V c represents the feature vector obtained by performing average pooling on the C direction; avepool represents average pooling, that is, for each two-dimensional data of size HxW along the C dimension direction, an average value operation is performed to obtain a scalar. Finally, after all pooling operations are completed in the C dimension direction, a vector with a length of C is obtained; Similarly, the same operations are performed on the other two dimensions to obtain the results V h 、V w ; 2) For the obtained V x (x ∈ C, H, W) enter two fully connected layers. The dimension size of the first fully connected layer is half of the vector length, and the dimension of the second layer is the same as the vector length, obtaining V' x : V' x = MLP(V x ) After obtaining V' x , perform a matrix multiplication operation on the original feature map F: F x = V' x ⊗ F where V' x represents the result scalar obtained after passing through two fully connected layers, ⊗ represents matrix multiplication, and F represents the original feature map; 3) Add the obtained F c , F h , F w to obtain the output F' of TAM F' = F c + F h + F w .

[0039] 4. Data storage and remote control For the original data and the calculated results, they are input into the data storage module, and the form of the data storage module is a stack queue, which can satisfy the sequentiality and real-time performance of the data at the same time.

[0040] For the classification results obtained after calculation, they can be monitored and controlled by remote control through various communication methods, such as https, socket. In this paper, the redfish and ipmi methods are used to establish communication with the remote control end. The remote control end obtains continuous data and real-time data through different requests.

[0041] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0042] Although the specific implementation manners of the present invention are described above, they are not limitations on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A fault classification method based on a BMC management system, characterized in that: The BMC management system framework includes 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 includes 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, and reading through the BMC management controller; S2, preprocessing the data based on the data preprocessing module, firstly reducing the noise of the original power consumption data, then converting it into two-dimensional data through empirical decomposition, and zero-filling the data to align all the data; S3, the preprocessed data enters the PHM algorithm module to obtain the fault classification result of the original data; the PHM algorithm module is a convolutional neural network based on three-dimensional self-attention, 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 data set collected by the BMC management system; S4, saves the original data and fault classification results into the data storage module for data management and remote calling.

2. A fault classification method based on a BMC management system as claimed in claim 1, characterized in that: The data preprocessing in S2 includes data denoising, empirical decomposition and zero filling; Based on data denoising, the denoised data 𝑥(i,𝑡) is obtained and converted to the intrinsic mode space according to the EMD decomposition method. The one-dimensional time domain data of 𝑥(i,𝑡) is converted into two-dimensional time domain and frequency domain data through the EMD method. The two-dimensional data is represented by E(i,t,m): E(i,t,m) = EMD(𝑥(i,𝑡)) = {𝑐 𝑗 (i,𝑡)∣𝑗∈{1, 2, …,𝑚}}∪{𝑟(i,𝑡)} Where EMD represents the empirical decomposition algorithm; j represents the subscript of the signal after EMD decomposition, 𝑐 𝑗 (𝑡) represents the jth 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; Zero-fill the decomposed E(i,t,m): First, find the maximum value of m for all E(i,t,m): M = max(m i ) 0<=i<=n Where m i is the number of decomposed signals obtained after EMD decomposition of the i-th fault; For all E(i,t,m), after zero padding, the data is represented as: E(i,t,M) = E(i,t) = {𝑐 𝑗 (i,𝑡)∣𝑗∈{1, 2,…, M}}∪{𝑟(i,𝑡)} i∈{0,1, 2,…, n} After processing, the original one-dimensional data with a length of L becomes two-dimensional data of M x L.

3. A fault classification method based on a BMC management system as claimed in claim 1, characterized in that: The specific data processing process of the PHM algorithm module is as follows: S31, input the preprocessed data E(i,t,M) into a 3x3 three-dimensional convolutional layer with W channels, and transform the M * L two-dimensional data into W * M * L three-dimensional data, which becomes the feature map F: F = cov(E) S32, input the obtained feature map into the three-dimensional self-attention TAM layer to obtain its output F'; S33, after another convolution layer, the number of channels becomes 1; S34, return to S31 and repeat 5 times. After the cycle is completed, enter a fully connected layer; S35, finally through the softmax layer, the probability of each fault mode is obtained, so as to achieve the fault classification goal. The dimension of softmax is equal to the size of the fault mode n plus a normal mode, that is, n+1.

4. A fault classification method based on a BMC management system as claimed in claim 1, characterized in that: The three-dimensional self-attention TAM layer is specifically: For a feature map, let it be F(C,H,W); F represents the feature map, C, H, and W represent the lengths of the three dimensions of the feature map respectively; S321, perform average pooling operations on F(C,H,W) along the three directions of C, H, and W respectively, and obtain three feature vectors of length C, H, and W respectively: V c = avepool(P (hxw) ) V c represents the feature vector obtained by average pooling in the C direction; avepool represents average pooling, that is, for each two-dimensional data of size HxW along the C dimension, the average value is calculated to obtain a scalar. Finally, after all the pooling operations are completed in the C dimension, a vector of length C is obtained; Similarly, do the same operation for the other two dimensions and get the result V h 、V w ; S322, for the obtained V x (x∈C,H,W) enters two fully connected layers. The dimension of the first fully connected layer is half of the vector length, and the dimension of the second layer is the same as the vector length, and V' is obtained. x : V’ x = MLP(V x ) Get V' x Finally, perform matrix multiplication on the original feature map F: F x = V’ x ⊗ F Where V' x It represents the result scalar after two fully connected layers, ⊗ represents matrix multiplication, and F represents the original feature map; S323, the obtained F c 、F h 、F w Add together to get the output F' of TAM F’=F c + F h + F w 。 5. A fault classification method based on a BMC management system as claimed in claim 1, characterized in that: The PHM algorithm module is trained based on the data set collected by the BMC management system. The data set production process is as follows: The sub-modules of the electronic system to be tested that are in different fault categories collect their corresponding power consumption data, and use their fault categories as labels. The power consumption data of multiple sub-modules of the same system are collected, assigned different fault numbers, and 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. For integers i>0, it means that the system is in the i-th fault category. Assuming there are n fault categories, the value of i is: i∈{0,1, 2, …, n} Assume the original data obtained is: origin(𝑡) origin(i,𝑡) = sample(i) Among them, origin (i,𝑡) represents the sampled power consumption data that changes according to time, sample represents the sampling of the power consumption data of the electronic system, and i represents the current state of the system; For each original data origin (i,𝑡), i represents the label of the original data, which indicates the fault category to which the original data corresponds; origin (i,𝑡) is a one-dimensional time series with a length of L, and the length L is determined by the sampling rate and sampling time: L = sample_rate x period Where L represents the number of sampled points, i.e. the length of the time series, sample_rate represents the sampling rate, and period represents the sampling time.

6. A fault classification method based on a BMC management system as claimed in claim 1, characterized in that: The original data and the calculated results are input into the data storage module in the form of a stack queue and are monitored and managed by a remote control unit.

Images