A hardware fault analysis method and application server based on power system network

Abstract

The present invention discloses a hardware fault analysis method and application server based on an electric power system network, the method comprising: performing regular inspections on hardware equipment in the electric power system and collecting relevant data; extracting time series data and non-time series data at different time points according to the acquired relevant data; constructing different feature matrices for the time series data and non-time series data according to the type of hardware equipment, inputting the time series feature matrix into a first fitting model, and inputting the non-time series feature matrix into a second fitting model; fusing the time series fitting data and the non-time series fitting data to output predicted fault classification data for the hardware equipment in the electric power system. The present invention fuses the fitted data, and the fusion processing method fully exploits the correlation between the first fitting data and the second fitting data, thereby obtaining a fitting feature with higher accuracy, and also obtaining hardware equipment fault type prediction data with high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring, and in particular to a hardware fault analysis method and an application server based on a power system network. Background Art

[0002] With the rapid development of smart grid construction, the power dispatching automation system is becoming more and more complex. More and more equipment is invested, the equipment operation time is getting longer and longer, and the equipment failure rate is increasing. The traditional manual operation and maintenance mode faces many challenges in data collection, inspection monitoring, fault alarm, asset management, etc. due to the different brands, models, types and security zones of the equipment. Each hardware device has an independent management platform. These management platforms are independent and scattered. The inspection work depends on the management platform of the hardware device. For different hardware devices, the system administrator needs to log in to the management platform of the hardware device to conduct inspections, which consumes a lot of time and energy of the system administrator. Not only is the inspection efficiency low, but the frequency and scope are also very limited. There is a lack of a unified and centralized operation and maintenance management interface for cross-type, cross-brand, and cross-model hardware devices as the daily operation entrance. Relying on manual collection, entry, and update of hardware equipment information, analysis of inspection results, and judgment of equipment failures is a large workload, and timeliness and accuracy are difficult to guarantee. Summary of the invention

[0003] Purpose of the invention: In order to overcome the deficiencies of the prior art, the present invention provides a method for analyzing hardware faults based on a power system network, which solves the problem of poor timeliness and accuracy in judging equipment faults. The present invention also provides an application server for hardware fault analysis based on a power system network.

[0004] Technical solution: According to a first aspect of the present invention, a method for analyzing hardware failures based on a power system network is provided, the method comprising:

[0005] Conduct regular inspections of hardware equipment in the power system and collect relevant data;

[0006] Extracting time series data and non-time series data at different time points t according to the acquired relevant data, wherein the time series data is used to indicate time series data related to the hardware equipment in the power system, and the non-time series data is used to indicate non-time series data related to the hardware equipment in the power system;

[0007] Construct different feature matrices for the time series data and the non-time series data according to the type of hardware devices, and obtain corresponding time series feature matrices and non-time series feature matrices respectively;

[0008] Inputting the time series feature matrix into a first fitting model to obtain time series fitting data, and inputting the non-time series feature matrix into a second fitting model to obtain non-time series fitting data;

[0009] The time series fitting data and the non-time series fitting data are fused and processed to output predicted fault classification data of the hardware equipment in the power system;

[0010] Wherein, the first fitting model includes sequentially connected convolution layer 1, contraction layer 1, convolution layer 2, pooling layer 1, contraction layer 2, convolution layer 3, pooling layer 2, contraction layer 3, convolution layer 4, and feedback layer, wherein the convolution layer 1 receives the time series feature matrix as input data, and the four convolution layers are respectively provided with four convolution kernels of different sizes to extract multidimensional features, a feedback matrix is ​​provided in the feedback layer, and the feedback layer after the convolution layer 4 is used to feed back the output connection weight matrix to the input data in reverse, and it uses the tanh activation function to feed back the input data, and the feedback matrix is ​​the matrix after the corresponding values ​​of the last two dimensions of the output connection weight matrix are interchanged, that is, if the output connection weight matrix is ​​expressed as: [X1, X2,…, X n-1 ,X n ], the feedback matrix is ​​expressed as: [X1,X2,…,X n ,X n-1 ], wherein n is the dimension corresponding to the output connection weight matrix, and the second fitting model adopts a long short-term memory network model.

[0011] Further, including:

[0012] The shrinkage layer includes 7 convolutional layers connected in sequence, and after the output result of the last convolutional layer is transmitted to the attention module, the output of the convolutional layer and the threshold obtained from the attention module are input into the soft thresholding layer to achieve noise reduction, and a fully connected layer is arranged between the third convolutional layer and the fourth convolutional layer, and a fully connected layer is also arranged between the sixth convolutional layer and the seventh convolutional layer.

[0013] Further, including:

[0014] The soft thresholding layer comprises:

[0015] Determine whether the input time series fitting data is logic judgment data. If so, use the hard threshold function for processing, which is expressed as:

[0016] Otherwise, an improved soft threshold function is used for processing, and the improved soft threshold function is expressed as:

[0017]

[0018] Among them, f s,t,k is the kth characteristic index element of the current device numbered s at time t, is the kth feature index element of the current device at time t after threshold processing, α is the threshold, which is obtained by the attention module, and λ is the adjustment coefficient, which is calculated according to f s,t,k and α are adjusted appropriately.

[0019] Further, including:

[0020] The time series fitting data and the non-time series fitting data are fused and processed to output predicted fault classification data of the hardware equipment in the power system, including:

[0021] Inputting the time series fitting data and non-time series fitting data into the merging layer to obtain the target fitting data output by the merging layer;

[0022] The target fitting data is input into the classification layer, so as to perform fault classification diagnosis through the classification layer.

[0023] Further, including:

[0024] The merging layer includes a deconvolution layer and a splicing layer. The deconvolution layer is used to upsample the matrix corresponding to the time series fitting data to the same dimension as the matrix corresponding to the non-time series fitting data, and then the splicing layer is used to splice and fuse the time series fitting data and the non-time series fitting data of the same dimension.

[0025] Further, including:

[0026] The time series data includes: processor, memory, hard disk usage information, device resource usage, device alarms, network equipment, power supply status data, and fan, voltage sensor, temperature sensor operation data; non-time series data includes: basic information of manufacturer, brand, model and serial number, as well as array card logs and system logs.

[0027] On the other hand, the present invention also provides an application server based on an electric power system network, in which the hardware fault analysis method as described above is placed in the application server, and the application server is placed in a safety zone 3 of the electric power system network, and the electric power system network also includes a safety zone 1 and a safety zone 2. The safety zone 1 and the safety zone 2 pre-process the collected relevant data and send them to the application server in the safety zone 3 through a power-dedicated forward isolation device.

[0028] Further, including:

[0029] The application server is also used to parse all relevant data files using a parsing method and store them in a corresponding database.

[0030] Further, including:

[0031] The data collection process includes:

[0032] Use SSH protocol to log in to the specified server;

[0033] Collect operating system status, processor status, memory status, and hard disk space usage information through system commands;

[0034] Use smartmontools to collect hard disk status and hard disk SMART information;

[0035] Use megacli or arcconf to collect RAID status and array card logs based on the server array card model;

[0036] Use dmidecode command to collect server hardware information;

[0037] Use the dmesg command to collect server log information;

[0038] Collect server sensor status through IPMI, including power status, fan status, processor temperature, system temperature, power consumption and other hardware status and operation records.

[0039] Further, including:

[0040] The data preprocessing includes: using SHELL to filter the collected relevant data.

[0041] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0042] (1) The fault prediction model of the present invention uses an improved threshold function to handle fault conditions of different equipment characteristics, thereby avoiding the problem of poor fault judgment accuracy caused by the application of existing functions in power systems. In addition, the fault prediction model can also ensure stability and accuracy when facing different types of data input and noise, thereby improving the robustness of the model.

[0043] (2) The present invention divides the collected power system hardware related data into time series data and non-time series data for separate processing, and adopts fitting models to perform fitting processing on each of them, and finally fuses the fitted data. The fusion processing method can fully explore the correlation between the first fitting data and the second fitting data, so as to obtain a more accurate fitting feature, so as to facilitate the fault type prediction based on the fitting feature, and also obtain the hardware equipment fault type prediction data with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is an overall flow chart of the fault analysis algorithm described in Example 1 of the present invention;

[0045] Figure 2 This is a schematic diagram of the fault prediction model structure according to Embodiment 1 of the present invention;

[0046] Figure 3 This is a schematic diagram of the first fitting model structure described in Example 1 of the present invention;

[0047] Figure 4 This is a schematic diagram of the shrinkage layer structure described in Example 1 of the present invention;

[0048] Figure 5 This is a schematic diagram of the application server structure described in Example 2 of the present invention;

[0049] Figure 6 This is a schematic diagram of data collection before filtering according to Embodiment 2 of the present invention;

[0050] Figure 7 This is a schematic diagram of data collection and filtering according to Embodiment 2 of the present invention;

[0051] Figure 8 This is a schematic diagram of the file format described in Example 2 of the present invention;

[0052] Fig. 9 This is a flow chart of the data collection method described in Example 2 of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Example 1

[0055] like Figure 1 As shown, the present invention provides a hardware fault analysis method based on a power system network, the method comprising the following steps:

[0056] S101 performs regular inspections on hardware equipment in the power system and collects relevant raw data;

[0057] The hardware devices of the power system in the embodiment of the present application are various hardware devices predicted in a power system network.

[0058] S102 extracts time series data and non-time series data at different time points t according to the acquired relevant data.

[0059] The time series data in the embodiment of the present application is used to indicate time series data related to the hardware equipment in the power system, and the non-time series data is used to indicate non-time series data related to the hardware equipment in the power system.

[0060] Exemplarily, the time series data and non-time series data involved in the embodiments of the present application may include but are not limited to: time series data includes: processor, memory, hard disk usage information, device resource usage, device alarms, network equipment, power supply status data, and fan, voltage sensor, temperature sensor operation data; non-time series data includes: manufacturer, brand, model and serial number basic information as well as array card logs and system logs.

[0061] In a possible implementation, the corresponding server device may obtain the time series data and non-time series data related to the hardware device at preset time intervals, that is, collect the time series data and non-time series data at different time points t.

[0062] Exemplarily, through the above data extraction script, data from various aspects are summarized and finally output as a csv file in dataframe format. The present invention is implemented using technologies such as Java, Vue, and Dameng database, and Python programming language can be used to process data into csv format, use pandas for data processing, and finally deploy through streamlit components to achieve the above functions.

[0063] S103 constructs different feature matrices for the time series data and the non-time series data according to the types of hardware devices, and obtains corresponding time series feature matrices and non-time series feature matrices.

[0064] In this embodiment, the data needs to be further cleaned, mainly dealing with missing values ​​and outliers. Missing values ​​are inserted using linear interpolation. Outliers need to be analyzed in combination with the actual data collection situation. If the state is abnormal, it is retained. If the abnormality is caused during the data extraction process, the data is considered to be deleted.

[0065] Furthermore, feature extraction is performed on the relevant data, and features related to hardware device failures are extracted from the data. Features other than numerical features are coded, and static configuration data including basic information (manufacturer, brand, model, serial number, etc.) and information of various components (such as motherboard, processor, memory, hard disk, etc.) are coded discretely. Other dynamic operation data including overall operation data of the equipment, such as equipment resource usage, equipment alarms, network connectivity, equipment power, fan, voltage, temperature, array card status, etc., are normalized and standardized.

[0066] Exemplarily, in this embodiment, different timing feature matrices and non-timing feature matrices are constructed according to the type of hardware devices. Exemplarily, the hardware devices may include but are not limited to: motherboard, processor, memory, hard disk, and non-timing feature matrices and timing feature matrices are constructed for the above hardware devices respectively. The matrix can be represented as the device number and the state data corresponding to a certain feature indicator at a certain time point; Exemplarily, the occupancy rate k of processor No. 2 at 8 o'clock in the morning is 50%, which is expressed as: 0.5 2,8,k , the corresponding processor timing feature matrix is ​​expressed as:

[0067]

[0068] Among them, ID is the number of the current device in the power system.

[0069] Specifically, numerical features such as occupancy rate within a day and within an hour can be constructed according to thresholds. Each column of the dataframe data can be constructed according to failure rate / 24h and failure rate / h. Two different dataframes can be constructed from the source dataframe to separate time series features and non-time series features.

[0070] S104: Input the time series feature matrix into the first fitting model to obtain time series fitting data, and input the non-time series feature matrix into the second fitting model to obtain non-time series fitting data.

[0071] That is Figure 2 The figure shows a schematic diagram of the fault prediction model structure, which uses feature weighting and threshold softening for time series and non-time series features to ensure the rationality of the fused data. The non-time series features of the previous step are fitted using a deep residual shrinkage network with a feedback mechanism. Multiple BatchNorm Relu convolutional layers are superimposed, and the output connection weight matrix of the last layer is reversely fed back to the input data through the tanh activation function. The matrix shape of the feedback matrix is ​​opposite to the last two dimensions of the final output matrix.

[0072] Finally, through the threshold softening output, the time series features are fitted using LSTM. The deep residual shrinkage network combines the deep residual network, SENet and soft threshold function, replaces the reweighting with threshold softening, and obtains the weight coefficient in the small network as the weight of data fusion. Finally, the matrix connection of the fully connected layer is used to achieve the effect of fusion of feature data in different formats.

[0073] In the embodiments of the present application, Figure 3As shown, the first fitting model includes sequentially connected convolution layer 1, contraction layer 1, convolution layer 2, pooling layer 1, contraction layer 2, convolution layer 3, pooling layer 2, contraction layer 3, convolution layer 4, and feedback layer 1, wherein four convolution layers are respectively provided with four convolution kernels of different sizes to extract multidimensional features. The convolution kernels in the embodiment of the present application can be appropriately adjusted according to experiments or requirements, and the feedback layer is the last convolution layer, and the output connection weight matrix is ​​fed back to the input data in reverse, and the tanh activation function is used to feed back to the input data, and the corresponding weight matrix is ​​opposite to the latter two dimensions of the output matrix. For example, the output matrix dimension is [1, a, b], and the feedback matrix is ​​[1, b, a]. The second fitting model adopts the long short-term memory network model.

[0074] In the embodiments of the present application, Figure 4 As shown, the structures of the above three contraction layers all include 7 convolutional layers connected in sequence, and after the output result of the last convolutional layer is transmitted to the attention module, the output of the convolutional layer and the threshold obtained from the attention module are input into the improved soft thresholding layer to achieve noise reduction, and a fully connected layer is arranged between the third convolutional layer and the fourth convolutional layer, and a fully connected layer is also arranged between the sixth convolutional layer and the seventh convolutional layer.

[0075] In an embodiment of the present application, the implementation of the soft thresholding layer includes the following steps:

[0076] First, determine whether the input time series fitting data is logic judgment data. If so, use the hard threshold function for processing, which is expressed as:

[0077] Otherwise, an improved soft threshold function is used for processing, and the improved soft threshold function is expressed as:

[0078]

[0079] Among them, f s,t,k is the kth characteristic index element of the current device numbered s at time t, is the kth feature index element of the current device at time t after threshold processing, α is the threshold, which is obtained by the attention module, and λ is the adjustment coefficient, which is calculated according to f s,t,k and the threshold α value to adjust.

[0080] Exemplarily, logical judgment data refers to non-numerical data, such as the physical and logical status of the hard disk is Failed, offline, faulty at a certain moment, or the fan module or power module status is Failed or offline, the network card interface status is down, and the network connection status is no. Other situations are data involving numerical values, such as the controller resource utilization rate is greater than 80%, the motherboard and hard disk temperature is greater than 50°C, etc.

[0081] The present invention uses different functions to judge different types of data, so as to obtain corresponding fitting data according to different data types. This firstly avoids program errors or eliminating some valid data as redundant data, and secondly, ensures the pertinence, effectiveness and accuracy of subsequent data fusion.

[0082] In this step, the improved soft threshold function introduces an adjustment coefficient, and its size can be flexibly set according to the needs, and the existing soft threshold function ignores the and f s,t,k There is a fixed error between them, and the mutation characteristics of the original data matrix may be lost, which will lead to a decrease in the signal-to-noise ratio of the reconstructed matrix and a large mean square error. Therefore, the present invention improves the soft threshold function representation method to make parameter adjustment more flexible and more in line with the predicted data of the power system.

[0083] S105 fuses the time series fitting data and the non-time series fitting data to output predicted fault classification data of the hardware equipment in the power system.

[0084] That is, data fusion. The present invention adopts a data merging method of matrix fusion to make data of different categories more reasonable and interpretable when merged. The implementation method is to connect the output att of the two-part model to the weight matrix, and use the weight of the weight matrix to replace the weight of the feature, which enhances the fitting strength. The output connected by the weight matrix is ​​used as the final output to indicate and analyze the current operating status of different types of hardware devices. When the matrix is ​​fused, the differences in the features and dimensions of the two matrices will affect the results after matrix fusion. The reverse convolution method is used to upsample the time series matrix to the same dimension as the non-time series matrix to obtain a more accurate fusion effect. The size relationship after transformation is as follows:

[0085] i s =I+(s–1)(I-1)

[0086] Among them, I represents the input size, s represents the stride, and the output size after stride transformation is as follows:

[0087] O=(i s +2p-k) / s+1

[0088] Among them, O represents the output size, p represents the padding size, and the final deconvolution output size is as follows:

[0089] O=s(I-1)+2p-k+2.

[0090] In an embodiment of the present application, the timing fitting data and the non-timing fitting data are specifically input into a merging layer to obtain target fitting data output by the merging layer;

[0091] The target fitting data is input into the classification layer, so as to perform fault classification diagnosis through the classification layer.

[0092] In this embodiment, the merging layer includes a deconvolution layer and a splicing layer. The deconvolution layer is used to upsample the matrix corresponding to the timing fitting data to the same dimension as the matrix corresponding to the non-timing fitting data, and then the splicing layer is used to splice and fuse the timing fitting data and the non-timing fitting data of the same dimension.

[0093] In the embodiment of the present application, the fault classification includes faults of servers, storage devices, and network devices. For example, taking the characteristics of servers and storage devices as an example, when the server characteristics appear:

[0094] The CPU, memory and hard disk usage are greater than 90%;

[0095] The CPU, memory, and motherboard voltages are lower or higher than normal, and the CPU core count or memory capacity decreases;

[0096] The temperature of the motherboard and hard disk is greater than 50°C, and the temperature of the processor CPU and GPU is greater than 85°C;

[0097] The fan speed is reduced or 0RPM, the power input voltage and output power are lower than or higher than normal values;

[0098] The hard disk status and hard disk SMART information show Failed, bad, and offline, and the array card RAID status is degraded;

[0099] The network card interface status is down, and the network connection status is no;

[0100] I / O error appears in the system log, data reconstruction appears in the array card log, and Asserted information appears in the hardware sensor record.

[0101] Storage device characteristics appear

[0102] The controller resource usage is greater than 80%;

[0103] The hard disk expansion box is in offline state and the temperature exceeds 50°C;

[0104] The BBU voltage is lower than the normal value and the remaining life is less than 90 days;

[0105] The fan module or power module status is Failed or offline.

[0106] The RAID and LUN status is Degraded or Fault;

[0107] The physical and logical status of the hard disk shows Failed, Offline, or Faulty;

[0108] The above is the abnormal status information of the server and storage device.

[0109] S106 model training: model training is performed on the processed data, and two parts of the model are trained separately. The features are the processed operation data of the CPU, memory, power supply, fan, array card, hard disk, network card and other components. This information is integrated and converted into a feature vector suitable for model training. The label is the operation status of the device, and the binary cross entropy loss is used for training. The parameters of the model are adjusted intermittently during the training process to improve the accuracy of the model's fault analysis.

[0110] L(y,y^)=-[ylog(y^)+(1-y)log(1-y^)]

[0111] S108 model evaluation, using cross validation to divide the cleaned data into a training set and a test set in a ratio of 7:3, and verify the estimated effect of the model on the test set. The F1 value is used as an evaluation indicator to evaluate the fitting effect of the non-time series model. The Segmental F1-Score is used as an evaluation indicator to evaluate the fitting effect of the time series model.

[0112] F1=(precision+recall) / (2×precision×recall)

[0113] S109 fault analysis collects time series for a period of time for post-launch model evaluation. The accuracy, efficiency and stability of the model are evaluated based on the model effect after launch, so as to conduct targeted follow-up improvements.

[0114] S110 optimization and adjustment tracks the effect of the model after it goes online, adjusts the model's hyperparameters through grid search, adjusts the hyperparameters by comparing the performance on the validation set, and subsequently optimizes the model's effect.

[0115] Example 2

[0116] The present invention also provides an application server based on the power system network, such as Figure 5As shown, the hardware fault analysis method as described above is placed in the application server, and the application server is placed in the safety zone 3 of the power system network. The power system network also includes a safety zone 1 and a safety zone 2. The safety zone 1 and the safety zone 2 pre-process the collected relevant data and send them to the application server in the safety zone 3 through a power-dedicated forward isolation device.

[0117] In the embodiments of the present application, the power system partitions the equipment according to its own information security requirements: real-time control area (security zone 1), non-control area (security zone 2), production management area (security zone 3) and management information area (security zone 4). Different security zones are isolated using physical isolation devices. The network communication between security zone 1 and security zone 3 is physically isolated, meeting the requirements of the "Overall Plan for the Safety Protection of the National Power Secondary System", but the system administrator must access the corresponding security partition to inspect the hardware equipment and analyze the equipment operation status, which causes inconvenience to the operation and maintenance work.

[0118] Safety Zone 1 is the real-time zone and controls the production area.

[0119] The traditional typical business systems in the control area include power data acquisition and monitoring systems, energy management systems, wide-area phasor measurement systems, distribution network automation systems, substation automation systems, power plant automatic monitoring systems, etc. Their main users are dispatchers and operation personnel. The real-time data transmission is in milliseconds or seconds, and their data communication uses the real-time subnet or dedicated channels of the power dispatching data network for transmission. The area also includes control systems that use dedicated channels, such as: relay protection, safety automatic control systems, low-frequency (or low-voltage) automatic load reduction systems, load control management systems, etc. These systems require real-time data transmission in milliseconds or seconds, among which the load control management system requires real-time data transmission in minutes.

[0120] Safety Zone 2 is a non-real-time zone and a non-controlled production zone.

[0121] The traditional typical business systems in the non-control area include dispatcher training simulation system, reservoir dispatching automation system, fault recording information management system, electric energy metering system, real-time and next-day power market operation system, etc. Its main users are power dispatchers, hydropower dispatchers, relay protection personnel and power market traders, etc. At the plant and station end, it also includes electric energy remote terminals, fault recording devices and power plant quotation systems, etc. The data collection frequency of the non-control area is minute or hour level, and its data communication uses the non-real-time subnet of the power dispatching data network.

[0122] Security Zone 3 and Zone 4 are management information zones.

[0123] The management information area refers to the collection of power enterprise management business systems outside the production control area. The traditional typical business systems of the management information area include the dispatching production management system, administrative telephone network management system, power enterprise data network, etc. Power enterprises can divide security areas according to specific circumstances, but it should not affect the safety of the production control area.

[0124] To ensure the security of the power system network, data transmission from Security Zone 1 and Security Zone 2 to Security Zone 3 requires isolation by a forward isolation device. The forward isolation device is a safety protection device located between the dispatching data network and the public information network, and is used for one-way data transmission from Security Zone 1 and Security Zone 2 to Security Zone 3. It can identify illegal requests and prevent data access and operations beyond authority, thereby effectively resisting malicious destruction and attack activities on the power system network initiated by viruses, hackers, etc. in various forms, and protecting the security of the real-time closed-loop monitoring system and the dispatching data network; at the same time, it uses non-network transmission methods to realize information and resource sharing between the two networks, ensuring the safe and stable operation of the power system.

[0125] The main function of the forward isolation device is to prohibit the direct establishment of TCP connections between applications in two areas. It decomposes the TCP connection between applications in two areas into two TCP virtual connections from the applications in the two areas to the two network cards inside and outside the isolation device. The two network cards inside and outside the isolation device are non-network connections inside the device. The forward isolation device only allows one-way transmission of presentation layer and application layer data, that is, the response message from safety zone 3 to safety zone 1 and safety zone 2 is prohibited from carrying application data (only one byte of data can be passed at most, and it is either all 0 or all 1). After the forward isolation device is added to safety zone 1, safety zone 2 and safety zone 3 of the power secondary system, ordinary C / S and B / S programs cannot pass through the device, and the program needs to be modified according to the programming regulations of the isolation device before data communication can be carried out.

[0126] Different safety zones in the power system are isolated by isolation devices, and the network cannot communicate normally. System administrators must visit the corresponding security partition to inspect hardware equipment and analyze the operating status of equipment, which causes inconvenience to operation and maintenance work.

[0127] System administrators of different hardware devices need to log in to the management platform of the hardware devices separately to conduct inspections, which consumes a lot of time and energy of system administrators. Not only is the inspection efficiency low, but the frequency and scope are also very limited. There is a lack of a unified and centralized operation and maintenance management interface that can serve as a daily operation entry point for hardware devices across types, brands, and models.

[0128] Manual collection, entry, and update of hardware equipment information, analysis of inspection results, and judgment of equipment failures are labor-intensive, and timeliness and accuracy are difficult to guarantee.

[0129] Therefore, the present invention provides a system and method for implementing hardware fault analysis across security partitions, using a unified and centralized hardware operation and maintenance management interface as the daily operation entrance to ensure the timeliness and accuracy of operation and maintenance, and improve the operation and maintenance capabilities of hardware equipment at all stages of its life cycle.

[0130] In the embodiment of the present application, the purpose of the present invention is achieved through the following technical solution: a collection program is deployed in the security zone 1 and the security zone 2, which is responsible for regularly collecting static configuration data and dynamic operation data of hardware devices.

[0131] Static configuration data includes not only basic information (manufacturer, brand, model, serial number, etc.), but also information about each component (such as motherboard, processor, memory, hard disk, etc.). Dynamic operation data includes not only overall equipment operation data such as equipment resource usage, equipment alarms, network connectivity, equipment power status, but also component-level operation data such as fans, voltage, and temperature. These monitoring data can provide support for fault analysis and advance prediction.

[0132] Specifically, data collection is first performed, wherein the data collected from safety zone 1 and safety zone 2 are sent to the application server of safety zone 3 through the power-specific forward isolation device. The application server of safety zone 3 is also responsible for data collection of safety zone 3.

[0133] The security zone three application server uses a parsing algorithm to parse all data files, and the parsed data is stored in the database.

[0134] The safety zone three application server uses the above fault analysis algorithm to analyze the data in the database, determine the hardware fault information of the equipment, and generate alarm information of the corresponding level.

[0135] The three-zone application server uniformly displays alarm information through the alarm overview view. Administrators can view specific alarm information and handle it as needed, including current alarm display, historical alarm query, alarm setting and other functions. Current alarms can view real-time alarm information, support retrieval by alarm level (emergency, major, minor, etc.), device type, device name, IP address and other key conditions, and can query alarm details, confirm alarms, close alarms, export alarms and other operations. Administrators can query historical alarm information lists through historical alarms, such as historical alarm details, export alarms, etc.

[0136] The three-zone security application server provides a display page that supports the visualization of all-round operation and maintenance information on large screens and PCs. It can display the operating status of hardware devices. Through a two-dimensional view, it can provide managers with a unified perspective, assisting managers to quickly obtain equipment operating status and alarm information, and intuitively locate abnormal devices.

[0137] In the embodiment of the present application, the data collection implementation process is as follows:

[0138] The hardware equipment of the power dispatching automation system mainly includes servers, storage devices, network equipment, workstations, etc. Different data acquisition command templates are formulated according to different hardware equipment.

[0139] Collection methods such as Fig. 9 As shown, first use the SSH protocol to log in to the device;

[0140] Use SHELL to filter collected data;

[0141] Get hardware device sensor data through IPMI;

[0142] Use Linux Expect to realize the automated interactive process. Through the IP address table of the hardware devices, data can be collected from all the hardware devices in the system at the same time.

[0143] For example, take server data collection as an example:

[0144] Use SSH protocol to log in to the specified server;

[0145] Collect operating system status, CPU status, memory status, hard disk space usage, etc. through system commands;

[0146] Use smartmontools to collect hard disk status and hard disk SMART information;

[0147] Use megacli or arcconf to collect RAID status and array card logs based on the server array card model;

[0148] Collect server hardware information through dmidecode;

[0149] Collect server log information through dmesg;

[0150] Collect server sensor status through IPMI, including power status, fan status, CPU temperature, system temperature, power consumption, other hardware status and operation records.

[0151] In some embodiments, the collected raw data displays a lot of content and contains a lot of useless data, so the collected data needs to be filtered. Figure 6 Collect CPU quantity, core number, model, etc. Use SHELL to filter the collected data. The filtered data shows the physical number of CPUs, core number, and CPU model. Figure 7 shown.

[0152] Exemplarily, the collected data is generated into a data file in txt format, and the file name is machine name + IP address + collection time, such as Figure 8 shown.

[0153] All files collected from safety zone 1 and safety zone 2 are integrated into one data file, and then the integrated data file is transmitted to the designated directory of the safety zone 3 application server through the forward isolation device. The safety zone 3 application server completes the data collection task of the safety zone 3 and stores the integrated data file in the designated directory, and then uses the fault analysis method to analyze the fault type.

[0154] Embodiment 2 of the present invention provides a unified and centralized operation and maintenance management interface as a daily operation portal, and implements operation status display views of servers, storage devices, etc., to assist managers in quickly obtaining device operation status and alarm information, intuitively locate abnormal devices, ensure the timeliness and accuracy of operation and maintenance, and improve the operation and maintenance capabilities of hardware equipment at all stages of their life cycle.

[0155] The present invention has been verified and achieved the expected effect. The following steps are the verification process and result display:

[0156] After verification testing on the test set, the accuracy of the preliminary model training can reach 82%.

[0157] Using out-of-time sequence cross-validation, the final accuracy is stabilized at 87% on different training sets and validation sets, and the F1-Score can reach 0.81.

[0158] Finally, after the hyperparameter search and adjustment, the maximum number of features considered by max_features is set to "log2", the maximum depth of the max_depth decision tree is set to 10, the criterion node division standard is set to "entrop", and the maximum number of leaf nodes of max_leaf_nodes is set to 5.

[0159] Finally, the accuracy on the out-of-time sequence can reach 93%, and the F1-Score can reach 0.89.

[0160] Verified by the F1 index, the present invention can achieve considerable fault analysis effects.

[0161] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0162] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for analyzing hardware failures based on a power system network, characterized in that: The method includes: Conduct regular inspections of hardware equipment in the power system and collect relevant data; Extracting time series data and non-time series data at different time points t according to the acquired relevant data, wherein the time series data is used to indicate time series data related to the hardware equipment in the power system, and the non-time series data is used to indicate non-time series data related to the hardware equipment in the power system; Construct different feature matrices for the time series data and the non-time series data according to the type of hardware devices, and obtain corresponding time series feature matrices and non-time series feature matrices respectively; Inputting the time series feature matrix into a first fitting model to obtain time series fitting data, and inputting the non-time series feature matrix into a second fitting model to obtain non-time series fitting data; The time series fitting data and the non-time series fitting data are fused and processed to output predicted fault classification data of the hardware equipment in the power system; Wherein, the first fitting model includes sequentially connected convolution layer 1, contraction layer 1, convolution layer 2, pooling layer 1, contraction layer 2, convolution layer 3, pooling layer 2, contraction layer 3, convolution layer 4, and feedback layer, wherein the convolution layer 1 receives the time series feature matrix as input data, and the four convolution layers are respectively provided with four convolution kernels of different sizes to extract multidimensional features, a feedback matrix is ​​provided in the feedback layer, and the feedback layer after the convolution layer 4 is used to feed back the output connection weight matrix to the input data in reverse, and it uses the tanh activation function to feed back the input data, and the feedback matrix is ​​the matrix after the corresponding values ​​of the last two dimensions of the output connection weight matrix are interchanged, that is, if the output connection weight matrix is ​​expressed as: [X1, X2,…, X n-1 ,X n ], the feedback matrix is ​​expressed as: [X1,X2,…,X n ,X n-1 ], wherein n is the dimension corresponding to the output connection weight matrix, and the second fitting model adopts a long short-term memory network model.

2. The hardware failure analysis method according to claim 1, characterized in that: The shrinkage layer includes 7 convolutional layers connected in sequence, and after the output result of the last convolutional layer is transmitted to the attention module, the output of the convolutional layer and the threshold obtained from the attention module are input into the soft thresholding layer to achieve noise reduction, and a fully connected layer is arranged between the third convolutional layer and the fourth convolutional layer, and a fully connected layer is also arranged between the sixth convolutional layer and the seventh convolutional layer.

3. The hardware failure analysis method according to claim 2, characterized in that: The soft thresholding layer comprises: Determine whether the input time series fitting data is logic judgment data. If so, use the hard threshold function for processing, which is expressed as: Otherwise, an improved soft threshold function is used for processing, and the improved soft threshold function is expressed as: Among them, f s,t,k is the kth characteristic index element of the current device numbered s at time t, is the kth feature index element of the current device at time t after threshold processing, α is the threshold, which is obtained by the attention module, and λ is the adjustment coefficient, which is calculated according to f s,t,k and α are adjusted appropriately.

4. The method for analyzing hardware failure according to any one of claims 1 to 3, characterized in that: The time series fitting data and the non-time series fitting data are fused and processed to output predicted fault classification data of the hardware equipment in the power system, including: Inputting the time series fitting data and non-time series fitting data into the merging layer to obtain the target fitting data output by the merging layer; The target fitting data is input into the classification layer, so as to perform fault classification diagnosis through the classification layer.

5. The hardware failure analysis method according to claim 4, characterized in that: The merging layer includes a deconvolution layer and a splicing layer. The deconvolution layer is used to upsample the matrix corresponding to the time series fitting data to the same dimension as the matrix corresponding to the non-time series fitting data, and then the splicing layer is used to splice and fuse the time series fitting data and the non-time series fitting data of the same dimension.

6. The hardware failure analysis method according to claim 1, characterized in that: The time series data includes: processor, memory, hard disk usage information, device resource usage, device alarms, network equipment, power supply status data, and fan, voltage sensor, temperature sensor operation data; non-time series data includes: basic information of manufacturer, brand, model and serial number, as well as array card logs and system logs.

7. An application server based on a power system network, characterized in that: The hardware fault analysis method described in any one of claims 1 to 6 is placed in the application server, and the application server is placed in the safety zone 3 of the power system network. The power system network also includes a safety zone 1 and a safety zone 2. The safety zone 1 and the safety zone 2 pre-process the collected relevant data and send them to the application server in the safety zone 3 through a power-specific forward isolation device.

8. The application server according to claim 7, characterized in that: The application server is also used to parse all relevant data files using a parsing method and store them in a corresponding database.

9. The application server according to claim 7, characterized in that: The data collection process includes: Use SSH protocol to log in to the specified server; Collect operating system status, processor status, memory status, and hard disk space usage information through system commands; Use smartmontools to collect hard disk status and hard disk SMART information; Use megacli or arcconf to collect RAID status and array card logs based on the server array card model; Use dmidecode command to collect server hardware information; Use the dmesg command to collect server log information; Collect server sensor status through IPMI, including power status, fan status, processor temperature, system temperature, power consumption and other hardware status and operation records.

10. The application server according to claim 7, characterized in that: The data preprocessing includes: using SHELL command to filter the collected relevant data.

Images