Fault diagnosis method, device, medium and program product

Abstract

The invention discloses a fault diagnosis method and device, a medium and a program product, and relates to the technical field of computers, and the method comprises the steps: obtaining time sequence data of a computer component; the time sequence data refers to a series of ordered data points generated and / or collected by the computer component within a set time interval; extracting at least one feature reflecting the fault state of the computer component from the time sequence data to obtain a time sequence feature; performing fault diagnosis based on the time sequence characteristics through a plurality of pre-constructed models to obtain a plurality of output results; the output result comprises a diagnosis result and a diagnosis value corresponding to the diagnosis result, and the diagnosis value is a quantitative representation of the diagnosis result generated by the model; and performing fusion processing on the plurality of output results to determine a final diagnosis result. The technical problems of long diagnosis time and low accuracy are solved by extracting the time sequence characteristics representing different fault states and comprehensively diagnosing decisions by multiple models, and the technical effect of efficiently and accurately diagnosing faults is achieved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a fault diagnosis method, device, medium and program product. Background Art

[0002] With the widespread application of artificial intelligence in various fields, the demand for servers has increased significantly. Servers include multiple components, such as central processing units (CPUs), graphics processing units (GPUs), and hard disks, and the demand for each component in the server has also increased dramatically. However, taking GPUs as an example, there may be manufacturing defects in the production process of GPUs, or the electronic components inside may age due to long-term use, slow heat dissipation, unstable power supply, incompatible drivers, firmware vulnerabilities, resource scheduling errors, and long-term high-load operation. These problems may cause GPU failures.

[0003] GPU failure can cause image anomalies, system crashes, calculation errors, and instant downtime. Failure to effectively identify and diagnose the faulty GPU will lead to serious server failures, which will have a significant impact on enterprise production and user use. Therefore, there is an urgent need for an efficient and accurate fault diagnosis method for computer components to ensure the stability and reliability of the server and guarantee a good user experience. Summary of the invention

[0004] The present application provides a fault diagnosis method, device, medium and program product to at least solve the problem of long time consumption and low accuracy of fault diagnosis in related technologies.

[0005] The present application provides a method for diagnosing a fault of a computer component, comprising: Acquire time series data of the computer component; wherein time series data refers to a series of ordered data points generated and / or collected by the computer component within a set time interval; Extracting at least one feature reflecting the fault state of the computer component from the time series data to obtain a time series feature; Fault diagnosis is performed based on time series features using multiple pre-built models to obtain multiple output results; wherein the output results include diagnostic results and diagnostic values ​​corresponding to the diagnostic results, and the diagnostic values ​​are quantitative representations of the diagnostic results generated by the model; Multiple output results are fused to determine the final diagnosis result.

[0006] The present application also provides a computer component fault diagnosis device, comprising: An acquisition module, used to acquire time series data of a computer component; wherein time series data refers to a series of ordered data points generated and / or collected by the computer component within a set time interval; A time series feature extraction module, used to extract at least one feature reflecting the fault state of the computer component from the time series data to obtain a time series feature; A fault diagnosis module is used to perform fault diagnosis based on time series features through multiple pre-built models to obtain multiple output results; wherein the output results include diagnostic results and diagnostic values ​​corresponding to the diagnostic results, and the diagnostic values ​​are quantitative representations of the diagnostic results generated by the model; The fusion output module is used to fuse multiple output results to determine the final diagnosis result.

[0007] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of any of the above-mentioned computer component fault diagnosis methods when executing the computer program.

[0008] The present application also provides a computer-readable storage medium, in which a computer program is stored, wherein when the computer program is executed by a processor, the steps of any of the above-mentioned computer component fault diagnosis methods are implemented.

[0009] The present application also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned computer component fault diagnosis methods when the computer program is executed by a processor.

[0010] Through this application, timing features that can represent different fault states of components are extracted, and comprehensive diagnostic decisions are made based on the timing features through multiple models to obtain accurate fault diagnosis results. This can solve the technical problems of long diagnosis time and low accuracy, achieve the technical effect of efficient and accurate fault diagnosis, and further ensure the stability and reliability of the server, guaranteeing a good user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0012] Figure 1 A schematic diagram of a flow chart of a fault diagnosis method provided in an embodiment of the present application; Figure 2 A technical framework diagram of a fault diagnosis method provided in an embodiment of the present application; Figure 3A technical framework diagram of a time domain feature extraction method provided in an embodiment of the present application; Figure 4 A technical framework diagram of a frequency domain feature extraction method provided in an embodiment of the present application; Figure 5 A schematic diagram of a flow chart of a fault diagnosis method provided in an embodiment of the present application; Figure 6 A schematic diagram of the structure of a residual module provided in an embodiment of the present application; Figure 7 A schematic diagram of the structure of a temporal convolutional network provided in an embodiment of the present application; Figure 8 A schematic diagram of the structure of a long short-term memory neural network provided in an embodiment of the present application; Fig. 9 A schematic diagram of the structure of a fault diagnosis device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0014] It should be noted that, in the description of this application, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence.

[0015] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0016] In conjunction with the specific application environment architecture or the specific hardware architecture on which the execution of the fault diagnosis method of the computer component depends, the specific application environment architecture or the specific hardware architecture is described herein.

[0017] Currently, GPUs may have the following problems during use due to factors such as hardware design, software driver, thermal management, power supply stability, and the complexity of computing loads: 1) Since GPUs are based on advanced processes (such as 5nm and 4nm), high-density transistor integration is susceptible to microscopic defects. Electromigration may occur after long-term use, causing circuit breakage or short circuit.

[0018] 2) Aging of MOSFETs or capacitors in multi-phase power supply designs can cause voltage ripple to exceed the standard, leading to unstable power supply to the GPU core or video memory, resulting in calculation errors or instantaneous downtime.

[0019] 3) The thermal design power (TDP) of high-performance GPUs can reach over 400W, and the heat flux density exceeds the limit of traditional cooling solutions. Hot spots (such as SM clusters) are prone to local overheating, triggering thermal throttling or even hardware fuses.

[0020] 4) Resource competition in parallel computing tasks (such as register bank conflicts and shared memory overflows) may cause undefined behavior. Insufficient exception handling mechanisms at the driver layer may lead to GPU context loss (GPU hang).

[0021] 5) An extremely long instruction stream (such as Monte Carlo simulation) may trigger the GPU watchdog timer (WatchdogTimer) to misjudge it as a hardware hang.

[0022] 6) Improper PCIe (Peripheral Component Interconnect Express) channel splitting in multi-GPU systems (such as incorrect configuration of x8 / x8 mode as x16 / x0) will cause bandwidth contention and transaction layer packet (TLP) transmission errors.

[0023] 7) GPUs that support ECC correct single-bit errors through Hamming codes and record multi-bit error rates, but uncorrected error accumulation may indicate problems such as memory module failure.

[0024] In summary, with the successive upgrades of GPUs, new models of GPUs may have other unknown faults and the diagnostic coverage of complex GPU faults is not high. In order to avoid abnormal operation of servers due to GPU faults, accurate and efficient diagnosis of GPU faults will be an urgent problem to be solved.

[0025] An embodiment of the present application provides a computer component fault diagnosis method, and the method is described in detail in conjunction with the execution flow of the computer component fault diagnosis method. Figure 1 A schematic diagram of a fault diagnosis method provided in an embodiment of the present application, specifically comprising: Figure 1 The following steps are shown: For example, see Figure 2 , Figure 2 A technical framework diagram of a fault diagnosis method provided in an embodiment of the present application, wherein the fault diagnosis method is executed by a fault diagnosis system. Figure 2 It can also be understood as a framework diagram of a fault diagnosis system, which includes a log collection module, a time series feature extraction module, a multi-model decision engine and a fusion decision module. Among them, the log collection module is used to collect log data of computer components and components associated with them. The computer components can be GPUs and / or hard disks. The following embodiments are described in detail using the GPU fault diagnosis scenario as an example. In the hard disk fault diagnosis scenario, time series features can be established based on the Smart parameters, temperature, voltage, and other information of the hard disk, and a multi-model decision algorithm and a multi-model decision engine are used to comprehensively diagnose whether the hard disk has a fault. The time series feature extraction module is used to extract the features of the collected log data to obtain time series features, wherein the time series features include time domain features, frequency domain features, and event sequence features. For the specific feature extraction process, see the following embodiments. The multi-model decision engine is used to generate multiple diagnostic results based on the same time series feature through multiple models, wherein the multiple models include short-term feature models, long-term dependency models, and knowledge rule engines (also known as knowledge rule reasoning models). The fusion decision module is used to fuse the diagnostic results of multiple models and output the final diagnostic result. It is described in detail in the following embodiments.

[0026] Troubleshooting methods include: S101. Obtain timing data of computer components.

[0027] Here, time series data refers to a series of ordered data points generated and / or collected by computer components within a set time interval.

[0028] Understandably, time series data can be log data, which contains information such as GPU temperature, voltage, power consumption, video memory load rate, ECC (Error Correction Code) error count, PCIe (Peripheral Component Interconnect Express) transmission error rate, PCIe bandwidth, etc. Time series data can be understood as all log data generated within a certain period of time. Among them, the set time interval refers to the time unit of data collection, which can be seconds, minutes, hours, days, etc. The collected data points are arranged in the order of the time they were collected, and each data point contains a timestamp and the corresponding measurement value, such as the temperature collected at a certain moment.

[0029] It is understandable that the sampling frequency can also be set according to the GPU load. For example, when the GPU is under low load, the sampling frequency is 1HZ. In this case, due to the low GPU activity, the lower sampling frequency is sufficient to capture the main performance characteristics without losing important information. At the same time, it also helps to reduce the occupation of system resources, such as reducing CPU usage and reducing memory consumption; when the GPU is under high load, the sampling frequency is 10HZ, which means that there are a lot of computing tasks going on, its utilization, temperature changes, and possible bottlenecks, so the sampling frequency can be increased.

[0030] It is understandable that time series data is collected by sliding time windows within a set time interval, and a series of data points are collected in each window. For example, if the set time interval is 10s, the first window is 0-5s, the second window is 1-6s, the third window is 2-7s, and so on, to complete the collection of time series data.

[0031] S102: extract at least one feature reflecting the failure state of the computer component from the time series data to obtain a time series feature.

[0032] It is understandable that, based on the above S101, the information in the log data that can reflect the GPU fault status is continuous, periodic and discrete. In order to make full use of the log data to diagnose GPU faults more comprehensively, the useful information in the log data needs to be converted into time series features. Among them, the time series features include time domain features, frequency domain features and event sequence features, etc. Extracting multiple features can more comprehensively and accurately characterize the fault information. Specifically, the time domain features are in a sliding time window of a fixed or dynamic length. The arithmetic mean, variance, kurtosis and adjacent window differences of all data points in the sliding time window are calculated. The frequency domain features are to extract the energy ratio of the main frequency component through FFT to identify periodic anomalies, such as the regular temperature rise caused by the vibration of the cooling fan. The event sequence feature is to encode discrete events (such as ECC error bursts) as timestamp sequences. It is understandable that the extraction order of the three features is not limited, and they can be extracted simultaneously or sequentially.

[0033] Among them, the time series features include time domain features.

[0034] Optionally, extracting at least one feature reflecting the fault state of a computer component from the time series data to obtain a time series feature can be specifically implemented by the following steps: Calculate the mean and variance of the time series data; calculate the kurtosis based on the time series data, mean and variance; wherein the kurtosis reflects the data distribution of the time series data; obtain the mean and variance of the previous data; calculate the mean difference and variance ratio based on the mean and variance of the time series data and the mean and variance of the previous data; wherein the time domain features include the mean, variance, kurtosis, mean difference and variance ratio of the time series data.

[0035] It is understandable that the mean and variance of the data points in each sliding time window (hereinafter referred to as the window) are calculated, where the mean reflects the average level of the data points in the window. For example, the mean of the GPU core temperature in the window continues to increase, which may indicate a decrease in heat dissipation efficiency or load abnormality. The mean calculation formula is shown in formula (1). The variance measures the degree of dispersion of the data points in the window from the mean and can capture the volatility of the data points. For example, a sudden increase in voltage variance may be caused by an unstable power module or a momentary short circuit. The contrast calculation formula is shown in formula (2). Kurtosis can describe the steepness of the distribution of data points in the window, reflect the probability of extreme values ​​and identify extreme events. For example, the kurtosis abnormality of the video memory error rate may indicate a hardware failure (such as damage to the video memory particles). The kurtosis calculation formula is shown in formula (3). The mean and variance of the previous data refer to the mean and variance corresponding to the previous window, or the mean and variance of the time series data collected in the previous set time period. The following embodiment takes the mean and variance corresponding to the previous window and the mean and variance corresponding to the current window in the same time series data as an example for detailed description. The mean and variance of the current window and the previous window can be calculated by formula (1) and formula (2). The mean difference and variance ratio calculated according to the mean and variance of adjacent windows (current window and previous window) can be understood as the adjacent window difference (hereinafter referred to as difference). The difference can quantify the difference in statistical features between adjacent windows, which is used to detect state mutation or gradual change trend and capture state migration. The difference is mainly reflected in two aspects: mean difference and variance ratio. The mean difference represents the absolute difference or relative change rate of the adjacent window means. For example, if the voltage mean difference of adjacent windows exceeds a threshold (such as 10%), it may trigger a power module fault warning. The variance ratio represents the ratio of the variances of adjacent windows, which detects volatility changes.

[0036] Formula (1) In the formula, is the mean, N is the window size, is the i-th data point in the window.

[0037] Formula (2) In the formula, is the variance.

[0038] Formula (3) In the formula, is the kurtosis, where It means that the distribution of data points is steeper than the normal distribution (peaked and thick-tailed), and there may be sudden anomalies. It means that the data points are distributed relatively flatly and the fluctuations are gentle.

[0039] , Formula (4) In the formula, is the mean difference, VR is the variance ratio, and are the mean and variance of the current time window, and are the mean and variance of the previous window.

[0040] For example, see Figure 3 , Figure 3 A technical framework diagram of a time domain feature extraction method provided in an embodiment of the present application. After completing the collection of log data, time domain features of the log data are extracted. The time domain features include the mean, variance, kurtosis and adjacent window difference calculated for each window.

[0041] Among them, the time series features include frequency domain features, the set time interval includes at least one time window, and the time series data includes multiple data acquired in at least one time window.

[0042] Optionally, extracting at least one feature reflecting the fault state of a computer component from the time series data to obtain a time series feature can be specifically implemented by the following steps: Perform frequency domain conversion on multiple data obtained in the time window to obtain a complex spectrum corresponding to the time window; extract the main frequency component of the complex spectrum to obtain the main frequency energy ratio corresponding to the time window; count the mean and standard deviation of the main frequency energy ratio corresponding to at least one time window, and calculate the abnormal threshold range based on the mean, standard deviation and preset threshold; determine the target main frequency energy ratio within the abnormal threshold range as the frequency domain feature.

[0043] It can be understood that both time domain features and frequency domain features are extracted for the data points in each window. The steps for extracting frequency domain features for a certain window are as follows: preprocess the log data, which includes signal acquisition, de-operation processing and segmented windowing processing. If the log data is obtained, then obtain the time series data from the log data, such as temperature, voltage, load rate, etc., and ensure the sampling frequency to obtain the sampling signal. Subsequently, use a sliding average filter or wavelet denoising to eliminate the high-frequency noise in the sampled signal to avoid interference with the frequency domain analysis. Divide the continuous signal into fixed lengths and apply a Hamming window to reduce spectrum leakage. After completing the preprocessing of the log data, perform a fast Fourier transform (FFT) on the preprocessed data, that is, perform FFT on the data points in the window, perform frequency domain conversion to obtain a complex spectrum, and the frequency domain conversion is shown in formula (5). Subsequently, extract the main frequency component of the complex spectrum after FFT conversion to obtain the main frequency energy ratio corresponding to the window, and extract the main frequency component as shown in formula (6). Subsequently, the mean and standard deviation of the main frequency energy ratios corresponding to all time windows are counted, that is, the mean and variance of all main frequency energy ratios are calculated. The abnormal threshold range is calculated based on the mean, standard deviation and preset threshold, and the abnormal threshold range is specifically shown in formula (7). Subsequently, the target main frequency energy ratio within the abnormal threshold range is determined as a frequency domain feature, that is, if the value of the main frequency energy ratio calculated in real time exceeds the abnormal threshold range, an alarm is triggered. For example, the periodic freezing of the GPU cooling fan can be detected or periodic abnormal phenomena caused by mechanical wear, circuit aging, etc. can be identified.

[0044] Formula (5) In the formula, is the complex spectrum, is the sampling signal after data preprocessing.

[0045] , Formula (6) In the formula, is a single frequency component, Recorded as amplitude , is the total energy, is the main frequency energy ratio, and M represents the window.

[0046] Formula (7) In the formula, T is the abnormal threshold range, and are the mean and variance of the main frequency energy ratio, 3 is the preset threshold, and the preset threshold can be determined according to user needs.

[0047] For example, see Figure 4 , Figure 4 A technical framework diagram of a frequency domain feature extraction method provided in an embodiment of the present application. After completing the log data collection, the frequency domain features of the log data are extracted. The specific implementation steps are as follows: the log data is preprocessed, and the data preprocessing includes at least one processing method such as signal sampling, denoising, and segmented windowing; the preprocessed sampled signal is fast Fourier transformed, the main frequency component is extracted, and the abnormal judgment logic is processed. The specific implementation steps refer to the above embodiment and are not repeated here.

[0048] Among them, time series data includes discrete data and continuous data reflecting discrete events, and time series features include event sequence features.

[0049] Optionally, extracting at least one feature reflecting the fault state of a computer component from the time series data to obtain a time series feature can be specifically implemented by the following steps: Determine the event type of discrete events; wherein discrete events refer to non-continuous events with timestamps that occur at a certain time point within a set time interval; according to the event type, the discrete events are encoded into timestamp sequences to obtain discrete features; the discrete features and continuous features are concatenated on the feature dimension to obtain event sequence features; wherein continuous features are extracted from continuous data.

[0050] It can be understood that discrete data is data that can only take specific values, and these values ​​are usually countable, such as the HyperText Transfer Protocol (HTTP) status code in log data (such as 200, 404, 500, etc.). Continuous data is a data type that can take any value, usually involving measurement or counting, such as response time, CPU usage, etc. Event sequence characteristics are to encode discrete events (such as ECC error bursts) as timestamp sequences, and combine continuous data to construct multimodal (multi-model) inputs. Among them, discrete events refer to non-continuous events that occur suddenly at a specific time point and have a clear timestamp, such as ECC errors, PCIe retransmission requests, and driver crash logs. ECC errors refer to burst errors in video memory error correction codes (such as single-particle upset events), PCIe retransmission requests refer to retransmission events triggered by data transmission errors, and driver crash logs refer to the time point records of GPU driver abnormal exits. Determine the event type of discrete events based on the nature and purpose of discrete events, or the type set by users according to their needs. Subsequently, different encoding methods are used according to the event type to encode discrete events into timestamp sequences and obtain event sequence features. Finally, the encoded discrete event sequence (event sequence features) is fused with continuous features as a unified input. Continuous features are extracted from continuous data (temperature, voltage, etc.) collected by continuous sensors. Event sequence features can also be understood as discrete features. In the process of fusing discrete features and continuous features, time alignment and feature splicing are also required. Among them, time alignment refers to the division of continuous data and discrete data in the same window. For example, the mean and variance of temperature and voltage sampled in real time in the same window (continuous features) and the statistical ECC error count (discrete features). Feature splicing refers to splicing continuous features with discrete features in the feature dimension. For example, continuous features include temperature mean = 75 , variance = 2.1, power consumption = 150W (dimension = 3), discrete features include ECC error count = 4, time difference mean = 10s (dimension = 2), continuous features and discrete features are concatenated into an input vector of dimension 5, that is, 75,2.1,150,4,10.

[0051] Optionally, according to the event type, the discrete events are encoded into a timestamp sequence to obtain discrete features, which can be implemented by the following steps: If the event type is a single event, binary coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; wherein binary coding refers to marking whether the event occurs at each sampling time point; or, if the event type is a high-frequency event, count coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; wherein a high-frequency event refers to an event that occurs more than a first set threshold in a first time period, and count coding refers to counting the number of occurrences of the event in each time window; or, if the event type is an intermittent event, time difference coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; wherein an intermittent event refers to an event that occurs discontinuously and has periodicity, and time difference coding refers to recording the time interval between adjacent events; or, if the event type is a complex event, time embedded coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; wherein a complex event refers to an event that occurs with a period greater than a second set threshold and relies on multiple data for trend analysis, and time embedded coding refers to mapping the timestamp of the event into a continuous vector.

[0052] It can be understood that the event types of discrete events include single events, high-frequency events, intermittent events and complex events, where a single event refers to a single type of event, such as a driver crash, a high-frequency event refers to an event that occurs more than the first set threshold in a short period of time, that is, an event that is not sensitive to time intervals and has a high density of events, such as multiple ECC errors in a short period of time, an intermittent event refers to an event that occurs discontinuously, is periodic or intermittent, that is, a sudden change in pattern, such as intermittent failures from sparse to dense, such as periodic freezes of cooling fans, and a complex event refers to an event that has an occurrence period greater than the second set threshold and requires trend analysis based on multiple data, that is, an event that is long-term dependent on analysis, such as a periodic increase in error rate caused by aging. After determining the event type, binary coding is used for coarse-grained detection of a single event. Binary coding refers to marking whether an event occurs at each sampling time point. Count coding is used for high-frequency events. Count coding refers to counting the number of times an event occurs within a sliding window. Time difference coding is used for intermittent events. Time difference coding refers to recording the time interval between adjacent events to reflect the rhythm of event occurrence. Time embedding coding is used for complex events. Time embedding coding refers to mapping timestamps into continuous vectors to capture periodicity / trend characteristics.

[0053] S103, performing fault diagnosis based on time series characteristics using a plurality of pre-built models to obtain a plurality of output results.

[0054] The output results include diagnostic results and diagnostic values ​​corresponding to the diagnostic results. The diagnostic values ​​are quantitative representations of the diagnostic results generated by the model.

[0055] It is understandable that, based on the above S103, multiple pre-built models (hereinafter referred to as multi-models) are obtained, and different models are used to diagnose whether there is a fault in the timing features for different types of timing features. Each model will obtain an output result, and the input of the multi-model is the timing features. Among them, each of the multiple output results output by the multi-model includes a diagnostic result and a diagnostic value corresponding to the diagnostic result. The diagnostic value is a quantitative representation of the diagnostic result generated by the model, such as a probability value and a confidence level.

[0056] Among them, the various models include short-term feature models, long-term dependency models, and knowledge rule reasoning models.

[0057] Understandably, the model that detects sudden anomalies in a short period of time is called a short-term feature model, such as the sudden anomaly of voltage drop; the model that captures the performance degradation trend across hours is called a long-term dependency model, such as the temperature drift caused by silicone grease aging; the model that needs to inject prior rules to improve interpretability is called a knowledge rule engine (knowledge rule reasoning model), for example, "video memory error rate > 1e-5 and accompanied by PCIe retransmission" is judged as a video memory failure.

[0058] S104: Fusing multiple output results to determine a final diagnosis result.

[0059] It can be understood that, based on the above S103, multiple diagnosis results output by multiple models such as short-term feature models, long-term dependency models, and knowledge rule reasoning models are further integrated to output the final diagnosis result.

[0060] Optionally, multiple output results are fused to determine the final diagnosis result, which can be achieved through the following steps: Determine whether the multiple output results include only one diagnostic result; if the multiple output results include only one first diagnostic result, determine the first diagnostic result as the final diagnostic result; or, if the multiple output results include multiple diagnostic results, compare and analyze the second diagnostic result in the third output result, the third diagnostic result in the first output result, and / or the fourth diagnostic result in the second output result to determine the final diagnostic result.

[0061] It is understandable that it is necessary to determine whether multiple diagnostic results are consistent, that is, whether there is only one diagnostic result among the multiple diagnostic results. For example, multiple diagnostic results all indicate that the GPU temperature is too high. In this case, the first diagnostic result is directly used as the final diagnostic result, wherein the first diagnostic result is the only diagnostic result that exists among the multiple diagnostic results. If the multiple output results include at least two diagnostic results, the diagnostic results output by each model are compared and analyzed to determine the final diagnostic result, wherein the third output result output by the knowledge rule reasoning model includes the second diagnostic result. The first output result output by the short-term feature model includes the third diagnostic result. The second output result output by the long-term dependency model includes the fourth diagnostic result for comparison and analysis, and there are at least two diagnostic results among the second diagnostic result, the third diagnostic result and the fourth diagnostic result.

[0062] Optionally, the second diagnostic result in the third output result, the third diagnostic result in the first output result, and / or the fourth diagnostic result in the second output result are compared and analyzed to determine a final diagnostic result, which can be specifically achieved by the following steps: When the second diagnostic result is determined based on a rule case, the second diagnostic result is determined as the final diagnostic result; or, when the second diagnostic result is not determined based on a rule case and the second diagnostic result is different from the third diagnostic result and / or the fourth diagnostic result, the target confidence in the third output result is weighted averaged with the first probability value in the first output result and / or the second probability value in the second output result, and the second diagnostic result, the third diagnostic result and the fourth diagnostic result are determined as the final diagnostic result; or, when the second diagnostic result is not determined based on a rule case and the second diagnostic result is the same as the third diagnostic result and the fourth diagnostic result, the target confidence, the first probability value and the second probability value are compared, and the diagnostic result corresponding to the maximum value is determined as the final diagnostic result.

[0063] It is understandable that in the process of diagnosis result fusion, if there is a clear rule case in the knowledge rule reasoning model, that is, the second diagnosis result is determined based on a real case, then the second diagnosis result is directly output, and the diagnosis results of other models are no longer output; if the second diagnosis result is inconsistent with the third diagnosis result and / or the fourth diagnosis result, the target confidence and the first probability value in the first output result and / or the second probability value in the second output result are weighted averaged, and then the diagnosis results are output separately, that is, for the same fault diagnosis result, each model must perform a weighted average of the confidence and probability values ​​when outputting the diagnosis value. The final diagnosis result includes the second diagnosis result and the third diagnosis result and / or the fourth diagnosis result, as well as the diagnosis value after weighted average of various diagnosis results. For example, the second diagnosis result is diagnosis A, the confidence is a, the third diagnosis result is diagnosis B, the probability value is b, and the third diagnosis result is C, the probability value is c. In this case, the three diagnosis results are different. For the diagnosis result of diagnosis A, the probability values ​​of the other two models are obtained, and the confidence a and the probability values ​​of the other two models are weighted averaged to obtain the diagnosis value A corresponding to diagnosis A, and the diagnosis A and the diagnosis value A are output. Similarly, the diagnosis values ​​of other diagnosis results are calculated. If the second diagnosis result is consistent with the diagnosis result in the model, the diagnosis result with higher confidence (probability value) is output. For example, the second diagnosis result is diagnosis A1, the confidence is a, the third diagnosis result is diagnosis B, the probability value is b, and the third diagnosis result is A2, the probability value is c. In this case, if the confidence a is greater than the probability value c, diagnosis A1 is determined as the final diagnosis result. If the confidence a is less than the probability value c, diagnosis A2 is determined as the final diagnosis result. Other possible fusion decision methods are not described in detail.

[0064] The fault diagnosis method provided by the present application converts useful information in log data into time series features such as time domain features, frequency domain features, and event sequence features, which can not only diagnose short time series faults of milliseconds, but also diagnose long time series faults such as those across months. Secondly, a sliding time window is used when extracting time domain features to calculate the arithmetic mean, variance, kurtosis, and adjacent window differences of all data points in the sliding window, so as to perform more accurate fault diagnosis through continuous data. In the frequency domain feature extraction, the energy ratio of the main frequency component is extracted by FFT to identify periodic anomalies. In the event sequence feature extraction, discrete event coding and multimodal input fusion strategies are used to clarify non-continuous events. Through the extraction of multiple time series features, it has high versatility and can perform fault diagnosis on GPUs of different manufacturers and models. In addition, when making fusion decisions on multiple output results, the final diagnostic results are collaboratively decided by using learning models and knowledge rule reasoning models, achieving high-precision, explainable fault diagnosis and component-level positioning, with strong practicality.

[0065] Based on the above embodiments, Figure 5 A flowchart of a fault diagnosis method provided in an embodiment of the present application is provided. Optionally, fault diagnosis is performed based on time series features using a plurality of pre-built models to obtain a plurality of output results, including the following: Figure 5 The following steps are shown: S501: Using the time series feature as the input of the short-term feature model to diagnose the sudden abnormality through the short-term feature model to obtain a first output result.

[0066] Among them, the short-term feature model adopts a temporal convolutional network.

[0067] Understandably, the short-term feature model uses a lightweight temporal convolutional network (TCN) to detect sudden abnormalities, because the temporal convolutional network has the advantages of millisecond-level response requirements (such as voltage sags must be alarmed within 10ms), lightweight (model parameter volume <1M, inference memory usage <50MB), accurate capture of short-term mutations (such as temperature spikes, ECC error rate surges), and the ability to suppress noise interference. The input of the short-term feature model is the temporal feature, and the output is the diagnostic value and the diagnostic result corresponding to the diagnostic value, which is used to diagnose sudden abnormal faults reflected by the temporal feature. Among them, the temporal convolutional network is shown in formula (8).

[0068] Formula (8) In the formula, is the diagnostic value output by the temporal convolutional network, that is, the probability value. is the filter, is the time series feature, d is The expansion factor at .

[0069] Optionally, the time series feature is used as the input of the short-term feature model to diagnose the sudden abnormality through the short-term feature model to obtain a first output result, which can be specifically implemented by the following steps: The sudden abnormality is diagnosed by adjusting the value of the dilation factor in the time series convolutional network to obtain a first output result; wherein the dilation factor is used to enable the time series convolutional network to capture pattern changes in time series features at different time scales.

[0070] Understandably, in the process of diagnosing sudden abnormal faults through short-term feature models, it is also possible to diagnose whether the GPU has a fault by adjusting the value of the dilation factor or by detecting long time series features, where the dilation factor is a parameter in the dilated convolution, which determines the size of the interval between elements in the convolution kernel. By adjusting the dilation factor, the receptive field of the convolution operation can be controlled, that is, the time span of the input data that the network can "see", that is, the dilation factor is mainly used to adjust the time period of the time series features, such as detecting sudden abnormalities from the first 1 second to the first 10 minutes in the time series features. For example, in the server monitoring log, if the CPU usage suddenly soars to 90% at a certain moment, while it usually remains at around 20%, this phenomenon may be regarded as a sudden abnormality.

[0071] Among them, the temporal convolutional network includes a residual module, which includes multiple convolutional layers, convolution kernel weights, activation functions, and connection methods between network layers.

[0072] Optionally, before diagnosing the sudden abnormality by using the short-term feature model to obtain the first output result, the method further includes: Set the relationship between the expansion factor and the number of convolutional layers; the association means that the expansion factor increases with the number of network layers; decompose the weight into two independent parameters, direction and amplitude, to constrain the weight distribution; set the gated activation function; set the connection mode to skip connection; skip connection refers to a network structure that adds the input to the output.

[0073] Understandably, in TCN, since deep networks are prone to gradient vanishing or exploding, in order to eliminate this problem, the residual module needs to be designed. Specifically: set the relationship between the expansion factor and the number of convolutional layers. For example, set the expansion factor to increase exponentially with the number of network layers, d×2, and set the convolution kernel size to K=3. This setting expands the causal convolution layer. By decomposing the weights into two independent parameters, direction and amplitude, the weight distribution is directly constrained to achieve the effect of accelerating model convergence and avoiding small batch noise. Set the gated activation function to enhance the nonlinear expression ability and suppress irrelevant features at the same time, that is, you can set the features to be retained, amplified or suppressed. Set the connection method of the jump connection to solve the problem of small gradients. The jump connection refers to a network structure that adds the input to the output.

[0074] For example, Figure 6 A schematic diagram of the structure of a residual module provided in an embodiment of the present application. The residual module involves configurations such as multiple convolutional layers, weights, activation functions, and connection methods between network layers. Specifically, the residual module is configured with convolutional layer expansion, weight normalization, gated activation, and jump connections to eliminate the problem of gradient disappearance or explosion. For specific setting instructions, please refer to the above embodiment and will not be repeated here.

[0075] For example, Figure 7 A schematic diagram of the structure of a temporal convolutional network provided in an embodiment of the present application, in which the temporal features pass through the input layer and enter the 3-layer residual module ( Figure 7 The hidden layer in the residual module is the residual module), and finally reaches the output layer. In the first residual module, d=2, K=3 are set, in the second residual module, d=4, K=3 are set, and in the third residual module, d=8, K=3 are set. The convolution layer is expanded by increasing the expansion factor. The residual module can be added according to user needs later. The setting method of the residual module is not described in detail. The final fault diagnosis result is determined in the output layer. Figure 7 It can be inferred that after being processed by the time series convolutional network, irrelevant feature information in the time series features can be filtered out layer by layer, and the feature information that can reflect GPU related information can be amplified, and finally the diagnosis results and the probability of failure can be output.

[0076] S502: Using the time series feature as an input of a long-term dependency model to diagnose abnormal performance trends through the long-term dependency model to obtain a second output result.

[0077] Among them, the long-term dependency model adopts long short-term memory neural network.

[0078] Understandably, the long-term dependency model uses the Long Short-Term Memory network (LSTM), because the LSTM network can capture time series features with performance decay trends across hours, such as the problem that the GPU core temperature slowly rises over time due to the decline in thermal conductivity of silicone grease due to aging; the problem that the voltage fluctuation is aggravated due to the decay of the capacitance value of the power supply module due to the aging of the capacitor; the problem that the fan bearing wears out and the heat dissipation efficiency is reduced, and the fan speed needs to be increased by 5%~10% under the same load. LSTM determines whether the data is updated or discarded through the logic control of the gate unit, and overcomes the shortcomings of excessive weight influence and easy gradient disappearance and explosion, so that the network can converge better and faster, and can effectively improve the prediction accuracy. The input of the long-term dependency model is also a time series feature, and the output is the diagnosis value and the diagnosis result corresponding to the diagnosis value. The diagnosis value is also a probability value.

[0079] It is understandable that the order of reasoning of the three models is not limited.

[0080] For example, Figure 8A schematic diagram of the structure of a long short-term memory neural network provided in an embodiment of the present application. LSTM has three gates, namely the forget gate, input gate, and output gate, which determine whether information is remembered or forgotten at each moment. The input gate determines how much new information is added at each moment, the forget gate controls whether the information will be forgotten at each moment, and the output gate determines whether there is information output at each moment. The calculation formula of the forget gate is shown in formula (9).

[0081] Formula (9) In the formula, is the hidden state at the previous moment, is the current input, that is, the time series feature, and are the weight and bias of the forget gate.

[0082] The calculation formula of the input gate is shown in formula (10).

[0083] , Formula (10) In the formula, is a candidate memory, and are the weights and biases of the input gate, and are the weights and biases of the candidate memories.

[0084] The calculation formula for memory update is shown in formula (11).

[0085] Formula (11) In the formula, is the candidate memory of the previous moment, It is the candidate memory updated at the current moment.

[0086] The calculation formula of the output gate is shown in formula (12).

[0087] , Formula (12) In the formula, is the final hidden state.

[0088] Optionally, in order to more accurately predict the fault trend over a long period of time in the time series characteristics and output reasonable fault handling opinions in the fault diagnosis results, it is necessary to add potential fitting loss to the predicted diagnosis results to correct the predicted trend so that the given diagnosis results are more realistic and reliable. The calculation formula of the trend term is shown in formula (13).

[0089] Formula (13) In the formula, is a parameter that controls the smoothness of the trend.

[0090] Optionally, the time series feature is used as the input of the long-term dependency model to diagnose the abnormal performance trend through the long-term dependency model to obtain a second output result, which can be specifically implemented by the following steps: Taking the time series features as the input of the long-term dependency model, the hidden states at different moments calculated by the output gate in the long short-term memory neural network are obtained; wherein the output gate is used to determine whether there is information output at each moment; based on the hidden state at the current moment, the hidden state at the previous moment and the set parameters, the trend of the time series features is calculated to obtain the predicted state at the current moment; wherein the set parameters refer to the parameters that control the smoothness of the trend; the loss is calculated according to the hidden state at the current moment and the predicted state at the current moment; the loss is compared with the set range to determine the diagnosis result to obtain the second output result.

[0091] It is understandable that the time series features are used as the input of the long-term dependency model, and the hidden states at different times are obtained through the input gate, forget gate and output gate. The hidden state is the output of the output gate, that is, the output of formula (12) Then, the hidden state at the current moment, the hidden state at the previous moment, the time series features and the set parameters are obtained, the trend of the time series features is calculated, and the predicted state at the current moment is obtained. The predicted state is the predicted trend output by formula (13). , the setting parameter refers to the . Then, the loss is calculated according to the predicted state at the current moment and the hidden state at the current moment, as shown in formula (14). It is determined whether the loss at the current moment is within a set range, which can be understood as an acceptable range, and the diagnosis result is determined to obtain a second output result, which is the output result of the long-term dependency model.

[0092] Formula (14) In the formula, For loss, 720 means that one month consists of 720 hours, and the specific time can be determined according to user needs.

[0093] For example, assuming that the aging of silicone grease causes the GPU temperature to rise by 0.3 degrees per month, The value of is 14400, the real GPU temperature trend (hidden state) is [75.0, 75.1, 75.2, ..., 77.4], the predicted GPU temperature trend (predicted state) is [75.0, 75.05, 75.15, ..., 77.3], and the calculated loss is 0.02. If the loss is within an acceptable range, the diagnosis result of "the GPU temperature rises by 0.3 degrees per month due to the aging of silicone grease" can be obtained.

[0094] S503, using the time series feature as an input of the knowledge rule reasoning model to diagnose the anomaly through a plurality of rules predefined in the knowledge rule reasoning model to obtain a third output result; The multiple output results include a first output result, a second output result and a third output result.

[0095] Understandably, there are multiple rules predefined in the knowledge rule reasoning model, and the rules mainly come from expert experience, historical fault analysis, and manufacturer documents. Among them, expert experience refers to the failure mode summarized by hardware engineers (such as the correlation between video memory error rate and PCIe retransmission), historical fault analysis refers to mining frequently co-occurring event combinations from fault logs (such as high temperature and insufficient fan speed), that is, there are real examples, and manufacturer documents refer to the fault diagnosis guide provided by the GPU manufacturer (such as hardware problems corresponding to specific error codes). The input of the knowledge rule reasoning model is the time series feature, and the output is the third output result, and fault diagnosis is mainly performed through rules.

[0096] Understandably, the knowledge rule reasoning model supports logical judgments between multiple rules: AND, OR, NOT, etc. For example, during the fault diagnosis process of multiple pairs of timing features, if Rule 1 "Video memory error rate>1e-5" and Rule 2 "PCIe retransmission rate>0.1" are triggered at the same time, the diagnosis result of video memory hardware fault can be generated. In the knowledge rule reasoning model, each rule also includes information such as rule change history, rule source, effective time, verification status, priority, confidence, etc. The specific content of each rule is not limited.

[0097] Optionally, the time series feature is used as an input of the knowledge rule reasoning model to diagnose the anomaly through a plurality of rules predefined in the knowledge rule reasoning model to obtain a third output result, which can be specifically implemented by the following steps: Match the time series features with multiple rules predefined in the knowledge rule reasoning model to determine at least one target rule; obtain the rule relationship between the multiple rules; wherein the rule relationship reflects the degree of correlation between different rules; determine the calculation strategy of at least one target rule based on the rule relationship, and calculate the target confidence based on the calculation strategy; wherein the third output result includes the confidence and the diagnostic result corresponding to the target confidence.

[0098] It is understandable that the extracted time series features include information such as the mean, variance, kurtosis, and difference between adjacent windows in the window. The time series features are matched with multiple rules to determine at least one target rule. In order to avoid the defects of overestimation of rule redundancy (if the rules are strongly correlated, the confidence will be repeatedly calculated after integration, resulting in false confidence) and underestimation of complementary rules (if the rules describe the same fault from different dimensions, the independent assumption will underestimate the true confidence) when multiple target rules are matched to point to the same conclusion, a classification fusion strategy based on rule relations will be adopted in the knowledge rule reasoning model to solve these two problems. Specifically, the rule relations between multiple predefined rules are obtained. The rule relations reflect the degree of correlation between different rules. For example, the first rule and the second rule are independent of each other, the first rule and the third rule are completely correlated, and the first rule and the fourth rule are partially correlated. Subsequently, according to the rule relations, the degree of correlation between the matched target rules is determined, and then the confidence is calculated based on the calculation strategy corresponding to the degree of correlation. The calculation strategies corresponding to different degrees of correlation are different.

[0099] Wherein, at least one target rule includes a first rule and a second rule, a calculation strategy of at least one target rule is determined according to the rule relationship, and the target confidence is calculated according to the calculation strategy, which can be specifically implemented by the following steps: If the rule relationship between the first rule and the second rule is independent of each other, the target confidence is calculated based on the first confidence of the first rule and the second confidence of the second rule; or, if the rule relationship is strongly correlated, the correlation coefficient between the first rule and the second rule is determined; and the target confidence is calculated based on the correlation coefficient, the first confidence and the second confidence, wherein the strong correlation refers to a relationship in which the correlation coefficient between the rules is greater than a third set threshold; or, if the rule relationship is a complementary relationship, the complementary weight between the first rule and the second rule is determined; and the target confidence is calculated based on the complementary weight, the correlation coefficient, the first confidence and the second confidence.

[0100] It is understandable that the confidence calculation is described in detail by taking the first rule and the second rule in multiple target rules as an example. Specifically, if the rule relationship between the first rule and the second rule is independent of each other, that is, the two rules are completely unrelated, then the target confidence is calculated based on the confidence of each rule, where the confidence of the independent rule is shown in formula (15). If the rule relationship between the first rule and the second rule is strongly correlated, then the correlation coefficient of each rule is determined based on the rule relationship, mainly by adding the correlation coefficient to the rule. ,in, , if the first rule and the second rule are independent, then ; If the first rule and the second rule are completely related, then ; If the first rule and the second rule have a certain correlation, then , where strong correlation means that the correlation coefficient between the rules is greater than the third set threshold. The third set threshold can be set according to user needs. For example, the third set threshold is 0.5. If the correlation coefficient between the first rule and the second rule is If the correlation coefficient is greater than 0.5, it means that the relationship between the first rule and the second rule is strongly correlated. Greater than 0.5, it also indicates that the rule relationship between the first rule and the second rule is strongly correlated. If the correlation coefficient If it is less than 0.5, it means that the rule relationship between the first rule and the second rule is not strongly correlated and can be defined as weakly correlated. After determining the correlation coefficient, the target confidence is calculated based on the correlation coefficient, the first confidence and the second confidence. The calculation formula is shown in formula (16). If the rule relationship between the first rule and the second rule is a complementary relationship, that is, the two rules are complementary in nature, in this case, the complementary weight between the first rule and the second rule is determined according to the rule relationship, where the complementary weight is used to reduce its effective weight when the first rule is complementary to other rules to avoid repeated calculation of confidence. After determining the correlation coefficient and the complementary weight, the target confidence is calculated based on the complementary weight, the correlation coefficient, the first confidence and the second confidence. The specific calculation formula is shown in formula (17).

[0101] Formula (15) In the formula, is the target confidence, is the confidence of the rule, where is the first confidence level, is the second confidence level.

[0102] Formula (16) In the formula, is the correlation coefficient between rule i and rule j, where rule i is the first rule and rule j is the second rule.

[0103] Formula (17) In the formula, is the complementary weight of the first rule.

[0104] It can be understood that in the knowledge rule reasoning model, the confidence of the final diagnosis result will be output by selecting an appropriate calculation strategy based on the association relationship of the diagnosis rules.

[0105] The fault diagnosis method provided by the embodiment of the present disclosure introduces weight normalization, gated activation, and skip connection in the residual module of the short-term feature model to solve the problem of gradient vanishing or exploding in TCN due to deep networks. In order to more accurately predict the trend of time series features across long periods of time and output reasonable fault handling opinions in the fault diagnosis results, potential fitting loss is added to the prediction results output by the long-term dependent model to correct the prediction trend. In the knowledge rule reasoning model, confidence, priority, correlation coefficient, and weight are added to the diagnostic rules. If the diagnostic results of multiple diagnostic rules point to the same conclusion, the confidence of the diagnostic results is calculated using a classification fusion strategy based on rule relationships.

[0106] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method.

[0107] The embodiment of the present application also provides a computer component fault diagnosis device, Fig. 9 A schematic diagram of the structure of a fault diagnosis device provided in an embodiment of the present application. The fault diagnosis device includes an acquisition module 901, a time series feature extraction module 902, a fault diagnosis module 903 and a fusion output module 904, wherein: An acquisition module 901 is used to acquire time series data of a computer component; wherein time series data refers to a series of ordered data points generated and / or collected by a computer component within a set time interval; A time series feature extraction module 902 is used to extract at least one feature reflecting the fault state of a computer component from the time series data to obtain a time series feature; The fault diagnosis module 903 is used to perform fault diagnosis based on time series features through multiple pre-built models to obtain multiple output results; wherein the output results include diagnostic results and diagnostic values ​​corresponding to the diagnostic results, and the diagnostic values ​​are quantitative representations of the diagnostic results generated by the model; The fusion output module 904 is used to fuse multiple output results to determine the final diagnosis result.

[0108] Among them, the time series features include time domain features.

[0109] Optionally, the time series feature extraction module 902 is used to: Calculate the mean and variance of time series data; Calculate the kurtosis based on the time series data, mean and variance; the kurtosis reflects the data distribution of the time series data; Obtaining the mean and variance of previous data; wherein the previous data refers to the time series data of the computer component obtained before a set time interval; Calculate mean differences and variance ratios based on the mean and variance of the time series data and the mean and variance of the previous data; Among them, the time domain features include the mean, variance, kurtosis, mean difference and variance ratio of the time series data.

[0110] The time series features include frequency domain features, the set time interval includes at least one time window, and the time series data includes multiple data acquired in each time window.

[0111] Optionally, the time series feature extraction module 902 is used to: Perform frequency domain conversion on multiple data obtained in each time window to obtain the complex spectrum corresponding to each time window; Extract the main frequency component of the complex spectrum to obtain the main frequency energy ratio corresponding to each time window; Count the mean and standard deviation of the main frequency energy ratio corresponding to all time windows, and calculate the abnormal threshold range based on the mean, standard deviation and preset threshold; The target main frequency energy ratio within the abnormal threshold range is determined as a frequency domain feature.

[0112] Among them, time series data includes discrete data and continuous data reflecting discrete events, and time series features include event sequence features.

[0113] Optionally, the time series feature extraction module 902 is used to: Determine the event type of a discrete event; a discrete event is a non-continuous event with a timestamp that occurs at a certain point in time within a set time interval; Encode discrete events into time stamp sequences according to event types to obtain discrete features; The discrete features and continuous features are concatenated in the feature dimension to obtain event sequence features; among them, the continuous features are extracted from continuous data.

[0114] Optionally, the time series feature extraction module 902 is used to: If the event type is a single event, binary coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; binary coding means marking whether the event occurs at each sampling time point; or, If the event type is a high-frequency event, count coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; wherein a high-frequency event refers to an event that occurs more than a first set threshold number of times in a first time period, and count coding refers to counting the number of occurrences of the event in each time window; or, If the event type is an intermittent event, the discrete event is encoded into a timestamp sequence using time difference coding to obtain discrete features; intermittent events refer to discontinuous and periodic events, and time difference coding refers to recording the time interval between adjacent events; or, If the event type is a complex event, time embedding coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; where a complex event refers to an event whose occurrence period is greater than the second set threshold and relies on multiple data for trend analysis, and time embedding coding refers to mapping the timestamp of the event into a continuous vector.

[0115] Among them, the various models include short-term feature models, long-term dependency models, and knowledge rule reasoning models.

[0116] Optionally, the fault diagnosis module 903 is used to: Using the time series feature as an input of the short-term feature model to diagnose sudden abnormalities through the short-term feature model to obtain a first output result; Using the time series feature as an input of the long-term dependency model to diagnose abnormal performance trends through the long-term dependency model to obtain a second output result; Using the time series feature as an input of the knowledge rule reasoning model to diagnose the anomaly through a plurality of rules predefined in the knowledge rule reasoning model to obtain a third output result; The multiple output results include a first output result, a second output result and a third output result.

[0117] Among them, the short-term feature model adopts a temporal convolutional network.

[0118] Optionally, the fault diagnosis module 903 is used to: The sudden abnormality is diagnosed by adjusting the value of the dilation factor in the time series convolutional network to obtain a first output result; wherein the dilation factor is used to enable the time series convolutional network to capture pattern changes in time series features at different time scales.

[0119] Among them, the temporal convolutional network includes a residual module, which includes multiple convolutional layers, convolution kernel weights, activation functions, and connection methods between network layers.

[0120] Optionally, the fault diagnosis method 900 is further used for: Set the relationship between the expansion factor and the number of convolutional layers; the relationship means that the expansion factor increases with the number of network layers; Decompose the weight into two independent parameters, direction and magnitude, to constrain the weight distribution; Set the gated activation function; Set the connection mode to skip connection; a skip connection is a network structure that adds input to output.

[0121] Among them, the long-term dependency model adopts long short-term memory neural network.

[0122] Optionally, the fault diagnosis module 903 is used to: The time series features are used as the input of the long-term dependency model to obtain the hidden states at different times calculated by the output gate in the long short-term memory neural network; the output gate is used to determine whether there is information output at each moment; Based on the hidden state at the current moment, the hidden state at the previous moment and the set parameters, the trend of the time series feature is calculated to obtain the predicted state at the current moment; wherein the set parameters refer to the parameters that control the smoothness of the trend; Calculate the loss based on the hidden state at the current moment and the predicted state at the current moment; The loss is compared with a set range, and a diagnosis result is determined to obtain a second output result.

[0123] Optionally, the fault diagnosis module 903 is used to: Matching the temporal features with multiple rules predefined in the knowledge rule reasoning model to determine at least one target rule; Obtaining the rule relationship between multiple rules; wherein the rule relationship reflects the degree of correlation between different rules; A calculation strategy for at least one target rule is determined according to the rule relationship, and the target confidence is calculated according to the calculation strategy; wherein the third output result includes the confidence and the diagnosis result corresponding to the target confidence.

[0124] The at least one target rule includes a first rule and a second rule.

[0125] Optionally, the fault diagnosis module 903 is used to: If the rule relationship between the first rule and the second rule is independent of each other, the target confidence is calculated according to the first confidence of the first rule and the second confidence of the second rule; or, If the rule relationship is strongly correlated, determine the correlation coefficient between the first rule and the second rule; and calculate the target confidence according to the correlation coefficient, the first confidence and the second confidence, wherein the strong correlation refers to a relationship in which the correlation coefficient between the rules is greater than a third set threshold; or, If the rule relationship is a complementary relationship, then the complementary weight between the first rule and the second rule is determined; and the target confidence is calculated based on the complementary weight, the correlation coefficient, the first confidence and the second confidence.

[0126] Optionally, the fusion output module 904 is used to: Determining whether the multiple output results include only one diagnosis result; If the multiple output results include only one first diagnosis result, the first diagnosis result is determined as the final diagnosis result; or, If the multiple output results include multiple diagnosis results, the second diagnosis result in the third output result, the third diagnosis result in the first output result, and / or the fourth diagnosis result in the second output result are compared and analyzed to determine a final diagnosis result.

[0127] Optionally, the fusion output module 904 is used to: In the case where the second diagnosis result is determined based on a rule case, determining the second diagnosis result as the final diagnosis result; or, In the case where the second diagnostic result is not determined based on a rule case and the second diagnostic result is different from the third diagnostic result and / or the fourth diagnostic result, a weighted average is performed on the target confidence in the third output result and the first probability value in the first output result and / or the second probability value in the second output result, and the second diagnostic result, the third diagnostic result and the fourth diagnostic result are determined as the final diagnostic result; or, When the second diagnostic result is not determined based on a rule case and the second diagnostic result is the same as the third diagnostic result and the fourth diagnostic result, the target confidence, the first probability value and the second probability value are compared, and the diagnostic result corresponding to the maximum value is determined as the final diagnostic result.

[0128] For the description of the features in the embodiment corresponding to the fault diagnosis device for a computer component, reference may be made to the relevant description of the embodiment corresponding to the fault diagnosis method for a computer component, which will not be described in detail here.

[0129] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned computer component fault diagnosis method embodiments.

[0130] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above-mentioned computer component fault diagnosis method embodiments when running.

[0131] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0132] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any of the above-mentioned computer component fault diagnosis method embodiments are implemented.

[0133] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, the non-volatile computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps in any of the above-mentioned computer component fault diagnosis method embodiments are implemented.

[0134] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0135] The above is a detailed introduction to the fault diagnosis of a computer component provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and its core idea of ​​the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A method for diagnosing a fault of a computer component, characterized in that: include: Acquire time series data of a computer component; wherein the time series data refers to a series of ordered data points generated and / or collected by the computer component within a set time interval; Extracting at least one feature reflecting the fault state of the computer component from the time series data to obtain a time series feature; Perform fault diagnosis based on the time series characteristics through multiple pre-built models to obtain multiple output results; wherein the output results include diagnostic results and diagnostic values ​​corresponding to the diagnostic results, and the diagnostic values ​​are quantitative representations of the diagnostic results generated by the models; The multiple output results are fused to determine the final diagnosis result.

2. The method according to claim 1, characterized in that: The time series feature includes a time domain feature, and extracting at least one feature reflecting the fault state of the computer component from the time series data to obtain the time series feature includes: Calculate the mean and variance of the time series data; Calculate the kurtosis according to the time series data, the mean and the variance; wherein the kurtosis reflects the data distribution of the time series data; Obtaining the mean and variance of previous data; wherein the previous data refers to the time series data of the computer component obtained before the set time interval; Calculate the mean difference and variance ratio according to the mean and variance of the time series data and the mean and variance of the previous data; Among them, the time domain features include the mean, variance, kurtosis, mean difference and variance ratio of the time series data.

3. The method according to claim 1, characterized in that The time series feature includes a frequency domain feature, the set time interval includes at least one time window, the time series data includes a plurality of data acquired in the time window, and extracting at least one feature reflecting the fault state of the computer component from the time series data to obtain the time series feature includes: Performing frequency domain conversion on the multiple data acquired in the time window to obtain a complex spectrum corresponding to the time window; Extracting the main frequency component of the complex spectrum to obtain the main frequency energy ratio corresponding to the time window; Counting the mean and standard deviation of the main frequency energy ratio corresponding to the at least one time window, and calculating the abnormal threshold range according to the mean, the standard deviation and a preset threshold; The target main frequency energy ratio within the abnormal threshold range is determined as the frequency domain feature.

4. The method according to claim 1, characterized in that: The time series data includes discrete data and continuous data reflecting discrete events, the time series features include event sequence features, and extracting at least one feature reflecting the fault state of the computer component from the time series data to obtain the time series features includes: Determine the event type of the discrete event; wherein the discrete event refers to a non-continuous event with a timestamp that occurs at a certain time point within the set time interval; Encoding the discrete event into a timestamp sequence according to the event type to obtain a discrete feature; The discrete features and the continuous features are concatenated in the feature dimension to obtain event sequence features; wherein the continuous features are extracted from the continuous data.

5. The method according to claim 4, characterized in that The step of encoding the discrete event into a timestamp sequence according to the event type to obtain discrete features includes: If the event type is a single event, the discrete event is encoded into a timestamp sequence using binary coding to obtain discrete features; wherein the binary coding means marking whether the event occurs at each sampling time point; or, If the event type is a high-frequency event, the discrete event is encoded into a timestamp sequence using counting coding to obtain discrete features; wherein the high-frequency event refers to an event that occurs more than a first set threshold number of times within a first time period, and the counting coding refers to the number of occurrences of the event within a statistical time window; or, If the event type is an intermittent event, the discrete event is encoded into a timestamp sequence using time difference coding to obtain discrete features; wherein the intermittent event refers to an event that occurs discontinuously and has periodicity, and the time difference coding refers to recording the time interval between adjacent events; or, If the event type is a complex event, time embedding coding is used to encode the discrete event into a timestamp sequence to obtain discrete features; wherein the complex event refers to an event whose occurrence period is greater than a second set threshold and relies on multiple data for trend analysis, and the time embedding coding refers to mapping the timestamp of the event into a continuous vector.

6. The method according to claim 1, characterized in that The multiple models include a short-term feature model, a long-term dependency model, and a knowledge rule reasoning model. The multiple models pre-built are used to perform fault diagnosis based on the time series features to obtain multiple output results, including: Using the time series feature as an input of the short-term feature model to diagnose a sudden abnormality through the short-term feature model to obtain a first output result; Using the time series feature as an input of the long-term dependency model to diagnose abnormal performance trends through the long-term dependency model to obtain a second output result; Using the time series feature as an input of the knowledge rule reasoning model to diagnose anomalies through a plurality of rules predefined in the knowledge rule reasoning model to obtain a third output result; The multiple output results include the first output result, the second output result and the third output result.

7. The method according to claim 6, characterized in that The short-term feature model adopts a time series convolutional network, and the time series feature is used as an input of the short-term feature model to diagnose sudden abnormalities through the short-term feature model to obtain a first output result, including: The sudden abnormality is diagnosed by adjusting the value of the dilation factor in the time series convolutional network to obtain a first output result; wherein the dilation factor is used to enable the time series convolutional network to capture the pattern changes in the time series features at different time scales.

8. The method according to claim 7, characterized in that The temporal convolutional network includes a residual module, and the residual module includes a plurality of convolutional layers, weights of convolutional kernels, activation functions, and connection modes between network layers. Before diagnosing the sudden abnormality through the short-term feature model and obtaining the first output result, the method further includes: Setting an association relationship between the expansion factor and the number of network layers of the plurality of convolutional layers; wherein the association relationship means that the expansion factor increases with the number of network layers; Decomposing the weight into two independent parameters, direction and magnitude, to constrain the weight distribution; Set the gated activation function; The connection mode is set to a skip connection; wherein the skip connection refers to a network structure that adds input to output.

9. The method according to claim 6, characterized in that The long-term dependency model adopts a long short-term memory neural network, and the time series feature is used as the input of the long-term dependency model to diagnose the abnormal performance trend through the long-term dependency model to obtain a second output result, including: The time series feature is used as the input of the long-term dependency model to obtain the hidden states at different moments calculated by the output gate in the long short-term memory neural network; wherein the output gate is used to determine whether there is information output at each moment; Based on the hidden state at the current moment, the hidden state at the previous moment and the set parameters, the trend of the time series feature is calculated to obtain the predicted state at the current moment; wherein the set parameters refer to parameters for controlling the smoothness of the trend; Calculating a loss based on the hidden state at the current moment and the predicted state at the current moment; The loss is compared with a set range, and a diagnosis result is determined to obtain a second output result.

10. The method according to claim 6, characterized in that The method of using the time series feature as the input of the knowledge rule reasoning model to diagnose anomalies through a plurality of rules predefined in the knowledge rule reasoning model to obtain a third output result includes: Matching the time series feature with a plurality of rules predefined in the knowledge rule reasoning model to determine at least one target rule; Obtaining the rule relationship between the multiple rules; wherein the rule relationship reflects the degree of correlation between different rules; A calculation strategy for the at least one target rule is determined according to the rule relationship, and a target confidence is calculated according to the calculation strategy; wherein the third output result includes the confidence and a diagnosis result corresponding to the target confidence.

11. The method according to claim 10, characterized in that The at least one target rule includes a first rule and a second rule, and determining a calculation strategy for the at least one target rule according to the rule relationship, and calculating the target confidence according to the calculation strategy, includes: If the rule relationship between the first rule and the second rule is independent of each other, the target confidence is calculated according to the first confidence of the first rule and the second confidence of the second rule; or, If the rule relationship is strongly correlated, determine the correlation coefficient between the first rule and the second rule; and calculate the target confidence based on the correlation coefficient, the first confidence and the second confidence, wherein the strong correlation refers to a relationship in which the correlation coefficient between the rules is greater than a third set threshold; or, If the rule relationship is a complementary relationship, a complementary weight between the first rule and the second rule is determined; and a target confidence is calculated based on the complementary weight, the correlation coefficient, the first confidence, and the second confidence.

12. The method according to claim 6, characterized in that The fusing the multiple output results to determine the final diagnosis result includes: Determining whether the multiple output results include only one diagnosis result; If the multiple output results include only one first diagnosis result, the first diagnosis result is determined as the final diagnosis result; or, If the multiple output results include multiple diagnosis results, the second diagnosis result in the third output result, the third diagnosis result in the first output result and / or the fourth diagnosis result in the second output result are compared and analyzed to determine a final diagnosis result.

13. The method according to claim 12, characterized in that The comparing and analyzing the second diagnosis result in the third output result, the third diagnosis result in the first output result, and / or the fourth diagnosis result in the second output result to determine a final diagnosis result includes: In the case where the second diagnosis result is determined based on a rule case, determining the second diagnosis result as the final diagnosis result; or, In the case where the second diagnostic result is not determined based on the rule case and the second diagnostic result is different from the third diagnostic result and / or the fourth diagnostic result, a weighted average is performed on the target confidence in the third output result and the first probability value in the first output result and / or the second probability value in the second output result, and the second diagnostic result, the third diagnostic result and the fourth diagnostic result are determined as the final diagnostic results; or, When the second diagnostic result is not determined based on the rule case and the second diagnostic result is the same as the third diagnostic result and the fourth diagnostic result, the target confidence, the first probability value and the second probability value are compared, and the diagnostic result corresponding to the maximum value is determined as the final diagnostic result.

14. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method for diagnosing a fault of a computer component as claimed in any one of claims 1 to 13 when executing the computer program.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the fault diagnosis method for a computer component according to any one of claims 1 to 13.

16. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the fault diagnosis method of a computer component as claimed in any one of claims 1 to 13 are implemented.

Images