A fault diagnosis method of an HBA card and an electronic device

Abstract

The application discloses a fault diagnosis method and electronic equipment of an HBA card, and relates to the technical field of servers. A time sequence feature set is extracted from time sequence data of the HBA card. Time domain statistical features of the time sequence feature set include one or more of I / O delay variance, temperature slope and link retry number. Frequency domain energy features include error frequency and / or main frequency energy ratio. Event burst features include signal strength fluctuation. According to the contribution degree of each feature in the time sequence feature set, multiple time sequence features used for fault diagnosis are selected from the time sequence feature set. According to the relevance of the multiple time sequence features and a set fault category, a fault label is determined. The fault label contains the fault category and the fault probability of the HBA card. Based on the fault label and the fault warning rule matched by the multiple time sequence features, fault response information is generated. The fault diagnosis accuracy is ensured, and the fault processing efficiency is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of servers, and in particular to a fault diagnosis method for an HBA card and electronic equipment. BACKGROUND

[0002] A host bus adapter (HBA) is a core connection component of a server and a storage network, and is responsible for protocol conversion and high-speed data transmission. Its failure will cause a series of problems such as data loss, business stagnation, and a sharp increase in operation and maintenance costs.

[0003] A traditional fault diagnosis method realizes fault diagnosis of an HBA by presetting a fixed threshold, such as an alarm when the temperature is greater than 85 DEG C. This implementation mode cannot distinguish between transient fluctuations and real faults, and has a high false alarm rate. For gradual faults such as fiber aging, the threshold cannot be triggered, and the detection is seriously missed. Moreover, this fault diagnosis method can only realize an alarm, and cannot realize timely processing of the fault, and can only realize fault diagnosis of the HBA by an engineer by analyzing a log file, so as to process the fault problem. The diagnosis process is inefficient and has poor real-time performance.

[0004] It can be seen that how to ensure the accuracy of fault diagnosis while improving the efficiency of fault processing is a problem to be solved by those skilled in the art. SUMMARY

[0005] The application provides a fault diagnosis method for an HBA card and electronic equipment to at least solve the problems of poor fault diagnosis accuracy and low fault processing efficiency in the related art.

[0006] The application provides a fault diagnosis method for an HBA card, comprising:

[0007] obtaining timing data of the HBA card;

[0008] extracting a timing feature set from the timing data; wherein the timing feature set comprises time domain statistical features, frequency domain energy features and event burst features; the time domain statistical features comprise one or more of I / O delay variance, temperature slope and link retry number; the frequency domain energy features comprise error frequency and / or main frequency energy ratio; and the event burst features comprise signal intensity fluctuation;

[0009] According to the contribution degrees of the features in the timing feature set, a plurality of timing features for fault diagnosis are selected from the timing feature set;

[0010] determining a fault label according to the relevance of the plurality of timing features and a set fault category;

[0011] generating fault response information based on the fault label and a fault warning rule matched by the plurality of timing features.

[0012] The application further provides a fault diagnosis device of an HBA card, comprising an acquisition unit, an extraction unit, a screening unit, a determination unit and a generation unit.

[0013] The acquisition unit is configured to acquire timing data of the HBA card.

[0014] The extraction unit is configured to extract a timing feature set from the timing data, wherein the timing feature set comprises time domain statistical features, frequency domain energy features and event burst features; the time domain statistical features comprise one or more of I / O delay variance, temperature slope and link retry number; the frequency domain energy features comprise error frequency and / or main frequency energy ratio; and the event burst features comprise signal strength fluctuation.

[0015] The screening unit is configured to screen a plurality of timing features for fault diagnosis from the timing feature set according to contribution degrees of features in the timing feature set.

[0016] The determination unit is configured to determine a fault label according to relevance of the plurality of timing features to a set fault category.

[0017] The generation unit is configured to generate fault response information based on the fault label and a fault early warning rule matched by the plurality of timing features.

[0018] The application further provides an electronic device, comprising a memory configured to store a computer program and a processor configured to execute the computer program to implement steps of the fault diagnosis method of the HBA card.

[0019] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement steps of the fault diagnosis method of the HBA card.

[0020] The application further provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement steps of the fault diagnosis method of the HBA card.

[0021] By the present application, the timing data of the HBA card is acquired; the timing feature set is extracted from the timing data; wherein the timing feature set includes multi-modal features such as time domain statistical features, frequency domain energy features and event burst features. The time domain statistical features can capture the overall distribution and trend of the timing data, the frequency domain energy features can capture the periodic faults in the timing data, and the event burst features can capture the millisecond-level burst faults. The error frequency and I / O delay variance of the HBA card are the most important features for fault detection, the frequency energy ratio and temperature slope have important reference value for HBA card fault reason analysis, and the signal strength fluctuation and link retry number are important parameters for HBA card link quality evaluation, so the time domain statistical features include one or more of the I / O delay variance, the temperature slope and the link retry number; the frequency domain energy features include the error frequency and / or the frequency energy ratio; the event burst features include the signal strength fluctuation. By extracting multi-modal timing features, the detection sensitivity is improved, and reliable data support is provided for subsequent fault detection. The timing feature set contains a large number of timing features, and some timing features have little contribution to fault detection, so it is not necessary to analyze them. Therefore, in order to improve the fault detection efficiency, the multiple timing features used for fault diagnosis can be selected from the timing feature set according to the contribution of each feature in the timing feature set. The fault label is determined according to the relevance of the multiple timing features and the set fault category. The fault label contains the fault category and the fault probability of the HBA card. The timing features associated with different fault categories are different. In order to realize the automatic processing of faults, all possible timing feature combination modes under different fault categories can be summarized to set corresponding fault warning rules. The fault warning rules can include corresponding fault handling methods. After the timing features are selected and the fault label is determined, the fault response information can be generated based on the fault warning rules matched by the fault label and the multiple timing features. The fault response information contains the response action required to solve the current HBA card fault. In this technical solution, the multi-modal timing features are extracted to improve the detection sensitivity. Based on the fault warning rules, the multiple timing features and the fault label are comprehensively analyzed to avoid false operation caused by single feature fluctuation. Based on the fault response information, the fault can be automatically processed in time to improve the fault handling efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any creative labor.

[0023] Figure 1 A flowchart of a fault diagnosis method of an HBA card is provided for the embodiments of the present application.

[0024] Figure 2 A flowchart of a method for obtaining multiple window time series data based on a dynamic window segmentation manner is provided for the embodiments of the present application.

[0025] Figure 3 A method flow of data cleaning is provided for the embodiments of the present application.

[0026] Figure 4 A method flowchart of screening time series features is provided for the embodiments of the present application.

[0027] Figure 5 A structural schematic diagram of a fault diagnosis device of an HBA card is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0029] It should be noted that, in the description of the present application, the terms “include”, “contain” or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. The terms “first”, “second” and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence.

[0030] In order for those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0031] Figure 1 A flowchart of a fault diagnosis method of an HBA card is provided for the embodiments of the present application, and the method comprises:

[0032] S101: Obtain time series data of the HBA card.

[0033] In the embodiments of the present application, the fault diagnosis of the HBA card is taken as an example for introduction. The fault diagnosis flow of the HBA card can be executed by management software or by a baseboard management controller (BMC), which is not limited herein.

[0034] The original time series data of the HBA card is collected in a distributed data collection manner, wherein the original time series data can include hardware sensor data, performance indicators and log information.

[0035] In actual application, time series data related to the performance and health status of the HBA card is collected from three levels of hardware sensors, performance monitoring agents and log analysis engines. The hardware sensor data collected in real time from the hardware sensor layer can include temperature, voltage, signal strength and other information of the HBA card. The performance indicators obtained from the operating system at the performance monitoring agent layer can include input / output (I / O) throughput, delay, queue depth and other information of the HBA card. The log analysis engine layer analyzes the HBA card driver log in real time to extract error count (CRC), link retry count, timeout event and other information.

[0036] In order to balance real-time performance and feature completeness, the original time series data can be sampled according to a dynamic window segmentation manner to obtain a plurality of window time series data. In order to ensure the quality of the time series data, the plurality of window time series data can be data cleaned to obtain the time series data of the HBA card.

[0037] The dynamic window segmentation manner is to adaptively adjust the window length according to the fluctuation characteristics of the time series data, so as to ensure short window fast response and long window capturing long-term trend, thereby achieving the balance between real-time performance and feature completeness.

[0038] Data cleaning aims to improve the quality of the time series data of the HBA card, and the data cleaning can include operations such as noise elimination, missing value filling and dimension standardization, thereby providing reliable input for subsequent feature extraction and fault identification.

[0039] S102: Extracting a time series feature set from the time series data.

[0040] In order to facilitate subsequent fault detection and analysis, a time series feature set with high information content and low redundancy can be extracted from the time series data of the HBA card.

[0041] The time series feature set can include time domain statistical features, frequency domain energy features and event burst features.

[0042] For extraction of the time domain statistical features, the time domain statistical features can be determined based on the distribution trend of the time series data in each window.

[0043] The time domain statistical features are used to capture the overall distribution and trend of the time series data, and the time domain statistical features can include any one or any combination of mean, variance, skewness, kurtosis and slope.

[0044] The mean represents the average value of the data in the window, and the calculation formula is as follows: ; where n represents the number of sampled data within a window, μ represents the mean, x i represents the i-th time series data.

[0045] The variance represents the degree of dispersion of data, and the calculation formula is as follows: ; where represents the variance.

[0046] The skewness represents the asymmetry of data distribution, and the calculation formula is as follows: ; where Skewness represents the skewness.

[0047] The kurtosis represents the sharpness of data distribution, and the calculation formula is as follows: ; where Kurtosis represents the kurtosis.

[0048] The slope represents the trend of data calculated by linear regression, and the calculation formula is as follows: ; where s represents the slope, t i represents the i-th time point.

[0049] The purpose of extracting frequency domain energy features is to capture periodic patterns in time series data, such as link jitter, periodic errors caused by firmware defects, etc. In practical applications, time series data can be converted to frequency domain signals: ; where X k represents the frequency domain signal; k represents a constant, which is a pre-set fixed value.

[0050] Perform principal frequency component extraction on the frequency domain signal to determine the principal frequency energy ratio: ; where represents the principal frequency energy, E i represents the energy of the i-th frequency domain signal, and Energy Ratio represents the principal frequency energy ratio.

[0051] Extracting event burst features is to capture the burst patterns of abnormal events, such as sudden increase in CRC errors, sharp rise in link retry times, etc.

[0052] The sudden increase in CRC errors refers to the growth rate of error count per unit time exceeding the threshold, such as CRC errors growing ≥ 50 times in 10 seconds.

[0053] Different burst events have their own associated performance indicators, so event burst features can be extracted according to the correlation between performance indicators and fault events in time series data.

[0054] In practical applications, the correlation coefficient between burst events and performance indicators can be calculated as follows: ; where r represents the correlation coefficient, x i represents the value of the i-th burst event, and yi performance indicator of the ith event, μ x denotes the mean of the events, μ y denotes the mean of the performance indicators.

[0055] For example: the correlation coefficient r of the CRC error burst event and the I / O delay is 0.82, indicating that the two are strongly correlated.

[0056] Considering that the CRC error frequency and the I / O delay variance of the HBA card are the most important features for fault detection, the main frequency energy ratio and the temperature slope have important reference value for HBA card fault cause analysis, and the signal strength fluctuation and the link retry number are important parameters for realizing HBA card link quality evaluation, therefore the time domain statistical features can include the I / O delay variance, the temperature slope and the link retry number; the frequency domain energy features can include the error frequency and the main frequency energy ratio; and the event burst features can include the signal strength fluctuation.

[0057] S103: According to the contribution degree of each feature in the time sequence feature set, multiple time sequence features used for fault diagnosis are selected from the time sequence feature set.

[0058] The time sequence feature set contains a large number of time sequence features, and some time sequence features have very small contribution degree to fault detection, and it is unnecessary to analyze, therefore in order to improve the fault detection efficiency, multiple time sequence features used for fault diagnosis can be selected from the time sequence feature set according to the contribution degree of each feature in the time sequence feature set.

[0059] The SHAP value (SHapley Additive exPlanations) represents the influence degree of the model on the prediction of the sample deviating from the benchmark value, in actual application, the SHAP value can be used as the contribution degree, and multiple time sequence features with the largest contribution to the model prediction can be selected from the time sequence feature set.

[0060] As shown in Table 1, the list of time sequence features with larger contribution to the model prediction includes the CRC error frequency, the I / O delay variance, the main frequency energy ratio, the temperature slope, the signal fluctuation intensity, the queue depth mean, the link retry number, the voltage fluctuation, the throughput decline rate and the error code burst frequency, therefore the selected time sequence features can include the 10 time sequence features.

[0061] Table 1

[0062]

[0063] From Table 1, it can be known that the CRC error frequency and the I / O delay variance are the most important time sequence features for fault detection, and the total contribution degree is 43%. The main frequency energy ratio and the temperature slope have important reference value for root cause analysis. The signal strength fluctuation and the link retry number are mainly used for link quality evaluation.

[0064] S104: Determine the fault label according to the relevance of the plurality of time sequence features and the set fault category.

[0065] In the embodiments of the present application, the plurality of time sequence features can be input into the fault classification model to obtain the fault probability corresponding to each fault category; wherein the fault classification model is obtained based on historical sample data, and the historical sample data includes a set of historical time sequence features and the corresponding fault category; the fault category can include hardware failure, software failure and link failure.

[0066] After determining the fault probability corresponding to each fault category, the fault category with the highest fault probability and the fault probability thereof can be taken as the fault label. The fault label can include the fault category of the HBA card and the fault probability.

[0067] In actual application, the fault classification model can use LightGBM (Light Gradient Boosting Machine). The probability value output by the LightGBM model ranges between 0 and 1, indicating the possibility of HBA card failure in the current time window. LightGBM can map the model output to the interval [0, 1] through an activation function (Sigmoid). The formula corresponding to the activation function is as follows:

[0068] ;

[0069] wherein, represents the original output of the fault classification model for the input feature x, represents the final output fault probability.

[0070] In order to evaluate whether a fault occurs, a fault threshold can be set. For example, the fault threshold can be set to 0.5. When the fault probability is greater than the fault threshold, it means that a fault occurs.

[0071] In the embodiments of the present application, the fault threshold can be dynamically adjusted according to business needs to balance the false positive rate and the false negative rate. For high sensitivity scenarios, in order to reduce the false negative rate, the fault threshold can be adjusted to 0.3. For high precision scenarios, in order to reduce the false positive rate, the fault threshold can be adjusted to 0.7.

[0072] The fault category corresponding to the HBA card can be divided into three categories, namely hardware failure, software failure and link failure. LightGBM can output the fault probability corresponding to each fault category through a multi-classification task.

[0073] For ease of description, P 硬件 represents the fault probability of hardware failure, P 链路 represents the fault probability of link failure, and P 软件 represents the fault probability of software failure.

[0074] The failure probability can be normalized by a probability distribution function (Softmax) as follows: ; wherein, .

[0075] The failure label is set by selecting the failure category with the maximum failure probability:

[0076] .

[0077] The time sequence features associated with different failure categories are different. Table 2 records the association list between the failure categories and the time sequence features.

[0078] Table 2

[0079]

[0080] For example, the input features include CRC error frequency = 150 times per minute, I / O delay variance = 50 ms2, and frequency energy ratio = 40%, and the output features include P 硬件 = 0.1, P 链路 = 0.85, and P 软件 = 0.05. Then, the failure probability of the failure label of the link failure is 0.85.

[0081] S105: Based on the failure label and the failure warning rule matched by the plurality of time sequence features, failure response information is generated.

[0082] The time sequence features associated with different failure categories are different. The same failure category can be further divided into different failure problems. The distribution of different time sequence features under the same failure category reflects different failure problems, and each failure problem has a corresponding failure handling method.

[0083] In the embodiments of the present application, in order to realize the automatic processing of the failure, all possible time sequence feature combination modes under different failure categories can be summarized to set corresponding failure warning rules. The failure warning rules can include corresponding failure handling methods. After the time sequence features are screened and the failure label is determined, the failure response information can be generated based on the failure label and the failure warning rule matched by the plurality of time sequence features.

[0084] In a specific implementation, a target failure warning rule matched with the failure label and the target time sequence feature can be screened from a set rule library. Each failure warning rule included in the rule library has a corresponding threshold range, which includes a threshold range of the failure probability and a threshold range of each time sequence feature. According to the threshold range corresponding to each failure warning rule in the rule library, the failure warning rule matched with the failure label and the target time sequence feature can be screened from the rule library.

[0085] The fault label and the target timing feature are analyzed according to the target fault early warning rule to determine a fault priority. According to a response action mapping table, fault response information matched with the fault priority is determined, wherein the fault response information includes an automatic response mode and a notification mode.

[0086] The syntax structure of the fault early warning rule is as follows:

[0087] IF <condition expression> ;

[0088] AND / OR <condition expression> ;

[0089] THEN <action> ;

[0090] WITH PRIORITY <level>.

[0091] Taking a link fault as an example, IF fault probability ≥ 0.9; AND CRC error frequency ≥ 100 times / minute; AND delay variance ≥ 20 ms 2 ; THEN trigger “link abnormal early warning”; WITH PRIORITY 1# highest priority.

[0092] Taking a heat dissipation fault in a hardware fault as an example, IF fault probability ≥ 0.8; AND temperature slope ≥ 0.5℃ / minute; AND throughput decline rate ≥ 20%; THEN trigger “heat dissipation abnormal early warning”; WITH PRIORITY 2.

[0093] In the embodiment of the application, response time requirements can be set for different priorities, and the higher the priority level is, the shorter the corresponding response time is, so as to ensure that a high-priority fault can be processed faster. Table 3 below is a priority level list.

[0094] Table 3

[0095]

[0096] Different automatic response actions can be set for different fault priorities. In the embodiment of the application, in order to realize timely processing of faults, a response action mapping table can be constructed in advance, and the automatic response mode and the notification mode corresponding to different priorities are recorded in the response action mapping table. In order to ensure the timeliness of fault processing, an escalation mechanism can also be set in the response action mapping table. Table 4 below is a response action mapping table.

[0097] Table 4

[0098]

[0099] Take the HBA card burst failure in a data center as an example. The event sequence is: time=00:00: CRC error frequency breaks the threshold (120 times per minute); time=00:05: I / O delay variance exceeds the limit (25 ms2); time=00:06: failure probability rises to 0.92.

[0100] Rule trigger: IF failure probability >= 0.9; AND CRC error >= 100 times per minute; AND delay variance >= 20 ms 2 ; THEN trigger "link abnormality"; WITH PRIORITY 1.

[0101] Response flow: time=00:06: automatically isolate the failed card and switch to the redundant link; time=00:07: send a short message to the on-duty engineer; time=00:10: the engineer confirms the failure and prompts "check the fiber interface"; time=00:15: replace the small form-factor pluggable optical module (SFP), and the failure is resolved.

[0102] It can be seen from the above technical solution that the timing data of the HBA card is obtained; the timing feature set is extracted from the timing data; wherein the timing feature set includes multi-modal features such as time domain statistical features, frequency domain energy features, and event burst features. The time domain statistical features can capture the overall distribution and trend of the timing data, the frequency domain energy features can capture the periodic faults in the timing data, and the event burst features can capture the millisecond-level burst faults. The error frequency and I / O delay variance of the HBA card are the most important features for fault detection, the frequency energy ratio and temperature slope have important reference value for HBA card fault reason analysis, and the signal strength fluctuation and link retry number are important parameters for HBA card link quality evaluation, so the time domain statistical features include one or more of the I / O delay variance, the temperature slope, and the link retry number; the frequency domain energy features include the error frequency and / or the frequency energy ratio; and the event burst features include the signal strength fluctuation. By extracting multi-modal timing features, the detection sensitivity is improved, and reliable data support is provided for subsequent fault detection. The timing feature set contains a large number of timing features, and some timing features have little contribution to fault detection and do not need to be analyzed, so in order to improve the fault detection efficiency, a plurality of timing features used for fault diagnosis can be selected from the timing feature set according to the contribution of each feature in the timing feature set. The fault label is determined according to the relevance of the plurality of timing features and the set fault category. The fault label contains the fault category and the fault probability of the HBA card. The timing features associated with different fault categories are different, in order to realize the automatic processing of the fault, all possible timing feature combination modes under different fault categories can be summarized to set the corresponding fault warning rules. The fault warning rules can include the corresponding fault handling mode. After the timing features are selected and the fault label is determined, the fault response information can be generated based on the fault warning rules matched by the fault label and the plurality of timing features. The fault response information contains the response action required to solve the current HBA card fault. In this technical solution, multi-modal timing features are extracted to improve the detection sensitivity. Based on the fault warning rules, the plurality of timing features and the fault label are comprehensively analyzed to avoid false operation caused by single feature fluctuation, and based on the fault response information, the fault can be automatically processed in time to improve the fault handling efficiency.

[0103] In the embodiments of the present application, in order to better adapt to the state change of the HBA card and ensure the rationality of the threshold range setting, the threshold range can be adjusted.

[0104] The adjustment method can include baseline adaptive adjustment, that is, dynamically calculating the threshold according to the device historical data. Taking the adjustment of the error threshold as an example, the average value and the standard deviation of the error count in a set time period can be counted; based on the average value and the standard deviation, the error threshold is determined.

[0105] For example, base = moving average (historical CRC errors, window = 24h); threshold = base + 3*standard deviation (historical CRC errors); wherein, base represents the average value of error count in 24h; threshold represents the adjusted error threshold.

[0106] The adjustment manner can comprise environmental factor compensation: considering the environmental parameters such as room temperature and load rate of the machine room to adjust the threshold. Taking the adjustment of the temperature slope threshold as an example, a compensation coefficient can be determined according to the difference between the current room temperature and the standard room temperature of the machine room; and the quotient of the temperature slope threshold and the compensation coefficient is taken as the adjusted temperature slope threshold.

[0107] For example, compensation coefficient = 1 + 0.1*(current room temperature - standard room temperature 25℃); effective slope = measured slope / compensation coefficient. Wherein, measured slope represents the current temperature slope threshold, and effective slope represents the adjusted temperature slope threshold.

[0108] In the embodiments of the present application, the threshold range in the fault early warning rule is adjusted in the manner of dynamic threshold adjustment mechanism, so that the threshold range can better adapt to the changes in device performance and environment, ensuring the rationality of the threshold range setting, thereby the fault early warning rule matching the current time sequence characteristics can be screened out.

[0109] Considering that in actual application, there can be multiple fault early warning rules matching the fault label and the target time sequence characteristics, and the fault handling manners determined by different fault early warning rules are different, which can cause diagnostic conflicts. Therefore, a conflict arbitration mechanism can be adopted to eliminate the conflict problem. The conflict arbitration mechanism can comprise a level priority and a contribution weight mechanism.

[0110] Each fault early warning rule has its corresponding priority, and in the case that there are multiple matched fault early warning rules, the target fault early warning rule can be determined in the manner of level priority, that is, the fault early warning rule with the highest rule priority is selected as the target fault early warning rule. In the case that there are multiple fault early warning rules with the highest rule priority, the target fault early warning rule can be determined in the manner of contribution weight, that is, the fault early warning rule with the highest contribution is selected as the target fault early warning rule.

[0111] In the embodiments of the present application, the rule priorities corresponding to the fault early warning rules can be pre-set. In actual application, the rule priority of the target fault early warning rule can be adjusted according to the number of times of triggering of the target fault early warning rule in a period of time. The period of time can be flexibly set, for example, it can be set to 24 hours (h).

[0112] The more the number of times of triggering of the target fault early warning rule, the higher the rule priority of the target fault early warning rule can be adjusted.

[0113] The adjustment manner of the rule priority is as follows:

[0114] adjusted_priority = base_priority * (1 + 0.2 * log (trigger times + 1));

[0115] Wherein, adjusted_priority represents the adjusted rule priority, and base_priority represents the rule priority before adjustment.

[0116] In the embodiment of the present application, by introducing the conflict arbitration mechanism, the fault warning rule with the highest priority is preferentially selected as the target fault warning rule. When there are multiple fault warning rules with the highest priority, the fault warning rule with the highest contribution degree is selected as the target fault warning rule, which effectively eliminates the diagnostic conflict. And based on the trigger frequency of the fault warning rule, the rule priority of the fault warning rule is adjusted, so that the rule priority of the fault warning rule is more in line with the actual demand.

[0117] In order to map the warning fault to a specific root cause and realize accurate positioning of the fault, root cause analysis can be performed on the fault.

[0118] Considering that the Bi-directional Long Short-Term Memory (Bi-LSTM) model can capture multi-scale time sequence dependence, realize feature enhancement, and support explainable diagnosis. Therefore, in the embodiment of the present application, the Bi-directional Long Short-Term Memory model can be used to perform root cause analysis.

[0119] In specific implementation, the time sequence data and its corresponding time sequence feature set can be input into the Bi-directional Long Short-Term Memory model to obtain an output result; wherein, the output result can include the prediction probability of each type of fault and the feature vector used for fault mode matching. Based on the feature similarity between the output result and each root cause rule in the fault mode library, the root cause analysis result is determined.

[0120] Bi-LSTM can capture multi-scale time sequence dependence, analyze future impact through forward LSTM, and trace historical reasons through reverse LSTM. Table 5 is a list of Bi-LSTM processing manners under different time scales.

[0121] Table 5

[0122]

[0123] Based on the feature enhancement of Bi-LSTM, for the input 50-dimensional time sequence feature, a 128-dimensional high-value feature vector can be output, which improves the retention rate of key features and the accuracy of pattern matching.

[0124] Bi-LSTM supports explainable diagnosis.

[0125] Attention weight example (t=12:30 weight 0.92):

[0126] Timeline: [12:00, 12:15, 12:30, 12:45];

[0127] Attention: [0.02, 0.05, 0.92, 0.01] → Indicates that the engineer focuses on checking the logs at 12:30.

[0128] Feature contribution analysis, quantifying feature impact:

[0129] Main frequency energy ratio (1-5Hz) contribution: +37% (promote "firmware defect" determination);

[0130] Voltage fluctuation contribution: -5% (suppress "power failure" misjudgment).

[0131] When using Bi-LSTM to perform root cause analysis, historical data needs to be extracted, which can include historical time series data within the last 24 hours and time series features with a feature dimension of 50. These historical data are used as input data for Bi-LSTM, and the output results include fault type probability distribution and 128-dimensional feature vectors.

[0132] Among them, the fault type probability distribution represents the prediction probability of each type of fault, and the dimension is equal to the number of fault types. The 128-dimensional feature vector (feature_vector) is the final hidden layer state, representing the feature vector used for fault pattern matching.

[0133] Output result example as follows:

[0134] "fault_probabilities":{

[0135] "firmware defect": 0.91,

[0136] "link instability": 0.07,

[0137] "heat dissipation failure": 0.02

[0138] },

[0139] "feature_vector": [0.24, -1.32, 0.57,...] / / 128-dimensional compressed features

[0140] }。

[0141] The output of the Bi-LSTM model is matched with rules in the fault mode library using feature similarity, and the rule with the highest similarity is taken as the final root cause analysis result. The feature similarity calculation formula is as follows:

[0142] ;

[0143] Among them, F i T represents the eigenvector; i This represents the predefined feature template values ​​in the fault mode library, which are standardized feature representations belonging to known fault types; w i The feature weights can be determined by their contribution. Table 6 below lists the failure mode library.

[0144] Table 6

[0145]

[0146] The typical pattern is as follows:

[0147] {

[0148] "Pattern_ID":"F007",

[0149] "Name": "Firmware Defect"

[0150] "Key_Features":[

[0151] "FFT detected a periodic signal of 0.0033Hz".

[0152] "Voltage fluctuation <2%"

[0153] Error code burst interval 300±10 seconds

[0154] ],

[0155] "Repair_Guide":[

[0156] "Upgrade firmware to v3.2.1",

[0157] "Rollback to stable version v2.8.4"

[0158] ],

[0159] "Confidence": 0.96

[0160] }

[0161] By collecting new fault samples, extracting features in the samples and standardizing the features. After training the standardized feature information through the Bi-LSTM model, a feature template is generated, the newly generated template is added to the fault mode library, and the confidence of the template is initialized to 0.8. Through the continuous improvement of the fault module library, the accuracy of the root cause analysis result is ensured.

[0162] In the embodiments of the application, the time series data can be obtained by dynamic window segmentation, Figure 2 A flowchart of a method for obtaining multiple window time series data based on dynamic window segmentation provided by the embodiments of the application, the method comprising:

[0163] S201: In the initial state, the first window time series data is obtained from the original time series data according to the set window length.

[0164] Before obtaining the time series data according to the dynamic window segmentation, the parameters of the dynamic window segmentation algorithm are defined as follows:

[0165] Input parameter: original time series data Such as I / O delay, CRC error count, etc. Wherein, X represents the original time series data, x t Represents the value of the time series at the t-th sliding window, i.e. the time point t.

[0166] Output parameter: dynamic window sequence Each window W i Contains a set of consecutive data points.

[0167] Key parameter: initial window length W int The default value can be set to 300 seconds.

[0168] Standard deviation threshold σ threshold It can be set according to the distribution of historical data, such as 10% of the data range, to avoid being too sensitive to noise.

[0169] Overlap rate R overlap , indicating the overlap rate between windows, to avoid feature omission, and to balance the calculation efficiency and feature continuity, usually set to a value between 30%-50%.

[0170] Divide the data by the initial window length W int Generate the first window W1, for example: if the data sampling interval is 1 second, W int = 300 seconds, then W1 contains 300 data points.

[0171] S202: According to the fluctuation index of the time series data in the previous window, adjust the window length of the next window.

[0172] In the embodiments of the present application, the fluctuation indicators can include a standard deviation and a slope. i for measuring the degree of data dispersion; the slope s i The data trend, such as the temperature rising rate, is calculated by linear regression.

[0173] The standard deviation calculation formula is as follows: ;

[0174] The slope calculation formula is as follows: ;

[0175] wherein x j represents the jth sampling data in the sliding window; , μ i represents the average value of the sampling data in the sliding window; t j represents the time stamp; n represents the number of sampling data in the sliding window; σ i represents the standard deviation, s i represents the slope.

[0176] The adjustment principle of the window length can be that when the data fluctuation is intense, the window length is reduced to improve the sensitivity; when the data is stable, the window length is extended to capture periodic faults; and when the data fluctuation is within a reasonable range, the window length is maintained unchanged.

[0177] In a specific implementation, when the fluctuation indicators of the time series data in the previous window meet the data fluctuation condition, the previous window length is adjusted according to the set reduction rule to serve as the window length of the next window. When the fluctuation indicators of the time series data in the previous window meet the data stability condition, the previous window length is adjusted according to the set extension rule to serve as the window length of the next window. When the fluctuation indicators of the time series data in the previous window do not meet the data fluctuation condition and do not meet the data stability condition, it indicates that the fluctuation of the time series data is within a reasonable range, and at this time, the window length can be maintained unchanged, that is, the previous window length is taken as the window length of the next window.

[0178] The data fluctuation condition can include judging whether the standard deviation of the time series data in the previous window is greater than a first standard threshold and whether the absolute value of the slope of the time series data in the previous window is greater than a first slope threshold.

[0179] When the standard deviation of the time series data in the previous window is greater than the first standard threshold and the absolute value of the slope of the time series data in the previous window is greater than the first slope threshold, it indicates that the fluctuation indicators of the time series data in the previous window meet the data fluctuation condition, and at this time, the window length of the next window can be determined according to the set window minimum value, the previous window length and the corresponding reduction ratio.

[0180] The data stationary condition can include judging whether the standard deviation of the time series data in the previous window is less than or equal to a second standard threshold value and whether the absolute value of the slope of the time series data in the previous window is less than or equal to a second slope threshold value.

[0181] In a case where the standard deviation of the time series data in the previous window is less than or equal to the second standard threshold value and the absolute value of the slope of the time series data in the previous window is less than or equal to the second slope threshold value, it is indicated that the fluctuation index of the time series data in the previous window satisfies the data stationary condition, and the window length of the next window can be determined according to the set window maximum value, the length of the previous window and the corresponding extension ratio.

[0182] For the setting of the threshold value, the half value of the first standard threshold value can be taken as the second standard threshold value, and the half value of the first slope threshold value can be taken as the second slope threshold value.

[0183] In actual application, the window length can be adjusted according to the following formula:

[0184] ;

[0185] wherein, W min represents the window minimum value, W max represents the window maximum value, W i represents the length of the i th window, W i+1 represents the length of the i+1 th window, σ threshold represents the first standard threshold value, s threshold represents the first slope threshold value, 0.5σ threshold represents the second standard threshold value, 0.5s threshold represents the second slope threshold value.

[0186] S203: According to the window length of the next window and the overlap rate, the window time series data of the next window is cut from the original time series data until all the original time series data is cut.

[0187] According to the adjusted W i+1 and the overlap rate R overlap , the next window time series data can be generated.

[0188] The starting position of the next window = the starting position of the previous window + (1-R overlap ) * W i .

[0189] For example, if W i = 300 seconds and R overlap = 50%, the starting position of the next window, i.e., the new window, is 150 seconds of the previous window.

[0190] In the embodiment of the present application, by adjusting the window length of the next window based on the fluctuation of the timing data in the previous window, the problem that the traditional fixed window cannot capture the rapid change of transient faults and there is a problem of missing detection of sudden abnormalities is solved. The redundant calculation caused by using a short window for smooth data is solved, and resource waste is avoided.

[0191] In the embodiment of the present application, in order to improve the quality of timing data, data cleaning can be performed on the multiple window timing data, Figure 3 A data cleaning method flowchart is provided for the embodiment of the present application, and the method comprises:

[0192] S301: Filtering noise of the multiple window timing data according to a set noise filtering mode to obtain filtered multiple window timing data.

[0193] The HBA card sensor may generate transient noise (such as temperature spikes, I / O delay outliers) due to electromagnetic interference and signal jitter, so a noise filtering mode is used to filter the noise, and a median filtering mode is used to process the noise here.

[0194] Median filtering: replace the current point with the median of the data in the sliding window, which effectively suppresses impulse noise. The median filtering formula is as follows: ;

[0195] Where y t represents the data corresponding to time point t after median filtering, x t represents the value of the time series at time point t before median filtering, median represents a function of taking the median, and the window size k is dynamically adjusted according to the noise duration.

[0196] For example, transient noise (duration <1 second) selects k=3, i.e. 3 sampling points. If multiple abnormal points are continuously detected, such as voltage fluctuation >10% in 5 windows, the window is expanded to k=5.

[0197] For example: original temperature data: [61, 63, 62, 119, 61]→filtered: [61, 63, 62, 61, 61]. The 119℃ spike is removed.

[0198] S302: supplementing the first type of missing data in the filtered multiple window timing data based on a linear interpolation mode to obtain multiple window timing data after initial supplementation.

[0199] Where the first type of missing data is data with a number of consecutive missing sampling points less than a set missing value.

[0200] In practical applications, data points can be missing due to transmission interruption or sensor failure, and therefore missing value processing is needed. The missing data can be classified into two categories according to the amount of missing data, i.e., first category of missing data (short-time missing) and second category of missing data (long-time missing).

[0201] The value of the missing value can be set in advance. For example, if the missing value is set to 3, the number of consecutive missing sampling points < 3 consecutive points indicates that the missing data belongs to the first category of missing data, and the number of consecutive missing sampling points > 3 consecutive points indicates that the missing data belongs to the second category of missing data.

[0202] The processing method for short-time missing is linear interpolation: filling based on the data points before and after. The calculation formula of linear interpolation is as follows: ; wherein, x t represents the value of the time series at time point t.

[0203] For example: the original delay sequence: [5, NaN, NaN, 8] → after interpolation: [5, 6, 7, 8].

[0204] S303: According to the influence weight of the missing data on the historical data, the second category of missing data in the multiple window time series data after the initial filling is filled to obtain the multiple window time series data after the final filling.

[0205] The second category of missing data is the data with a number of consecutive missing sampling points greater than or equal to the set missing value.

[0206] For long-time missing, a prediction method based on historical patterns can be used, and an ARIMA model is used to predict the missing section, which is suitable for periodic indicators, such as daily fixed period I / O load.

[0207] The formula corresponding to the ARIMA model is as follows: ;

[0208] ; wherein, x t represents the value of the time series at time point t, such as the I / O delay of the HBA card, temperature, etc.; c represents a constant term, representing the mean or baseline of the time series; represents the autoregressive coefficient, which is used to represent the influence weight of the past p time points on the current value; represents the moving average coefficient, which is used to represent the influence weight of the random error of the past q time points on the current value; represents the random error at time point t, which belongs to white noise, and is usually assumed to follow a normal distribution with mean 0 and variance σ 2 ; p represents the autoregressive order, which is used to represent the use of the values of the past p time points to predict the current value; q represents the moving average order, which is used to represent the use of the random errors of the past q time points to predict the current value.

[0209] S304: Normalize the plurality of windowed time series data after the final supplement to obtain time series data of the HBA card.

[0210] The multi-source data has large dimension differences, such as temperature unit ℃ and delay unit ms, which will affect the convergence speed and accuracy of the model, and therefore, the data needs to be normalized. In this embodiment, the Min-Max normalization method is used for data processing. The normalization formula is as follows:

[0211] ;

[0212] wherein, x represents the normalized time series data at the time point t, x max x represents the maximum value of the time series data, x min x represents the minimum value of the time series data.

[0213] In order to adapt to the baseline drift caused by equipment aging, x max and x min may be reset every 24 hours.

[0214] In the embodiment of the present application, the short-time missing data is supplemented by using median filtering and linear interpolation, and the long-time missing data is supplemented according to the influence weight of the historical data on the missing data. The denoising effect and information retention are effectively balanced, and the loss of key abnormal patterns caused by excessive cleaning is avoided.

[0215] Considering that the time series features included in the time series feature set are relatively large, some time series features do not have actual help for fault diagnosis, and therefore, the time series features included in the time series feature set can be screened. Figure 4 A method flowchart for screening time series features is provided in the embodiment of the present application, and the method comprises:

[0216] S401: The remaining features except the target feature in the time series feature set are taken as a target subset.

[0217] wherein, the target feature is any one of all the features included in the time series feature set.

[0218] The time series feature set includes time domain statistical features, frequency domain energy features, and event burst features, and the SHAP value can be used to screen a plurality of key features with the largest contribution to the model prediction from the time series feature set.

[0219] S402: The contribution degree of the target feature is determined according to the time series feature set, the target subset, and the respective prediction values thereof.

[0220] In a specific implementation, the time sequence feature set and the target subset can be respectively taken as input features of the fault classification model to determine the prediction values corresponding to the time sequence feature set and the target subset respectively; the time sequence feature set and the target subset are processed according to a set weight calculation rule to obtain the weight; the difference between the prediction value corresponding to the time sequence feature set and the prediction value corresponding to the target subset is multiplied by the weight to obtain the contribution degree corresponding to the target feature.

[0221] For the time sequence feature j, the contribution degree calculation formula is: ;

[0222] Wherein, represents the contribution degree of the time sequence feature j, F represents the set of all features, that is, the time sequence feature set; S represents the target subset not containing the feature j; f(S) represents the prediction value of the model for the target variable when only using the features in the target subset S; represents the weight.

[0223] When f(S) is calculated, the features not contained in S are shielded or filled with default values such as mean, median or zero. The fairness of contribution allocation is ensured by setting the weight.

[0224] S403: Select a set number of time sequence features from the time sequence feature set in the order of contribution degree from high to low.

[0225] The set number can be 10, and 10 time sequence features can be selected from the time sequence feature set in the order of contribution degree from high to low. The 10 time sequence features can be referred to the introduction of Table 1, and will not be described here.

[0226] In the embodiments of the present application, a comprehensive solution for fault diagnosis of the HBA card is provided by integrating multi-dimensional sensor data, the fault diagnosis accuracy is improved, and the false positive rate is reduced. When performing fault analysis, the set number of time sequence features with the highest contribution degree are first selected from the time sequence feature set, the calculation amount is greatly reduced while ensuring the analysis accuracy, and the fault analysis efficiency is improved.

[0227] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better implementation.

[0228] The above introduction is all taken as an example of fault diagnosis of the HBA card. The fault diagnosis method of the HBA card provided in the embodiments of the present application can also be widely applied to electrocardiogram (ECG) abnormality detection, such as: converting the CRC error surge detection of the HBA card into ST segment elevation identification of the electrocardiogram signal; converting the self-adaptive window adjustment mechanism into heart rate variability for different patients. It can also be applied to the field of industrial Internet of Things, such as converting the I / O delay variance of the HBA card into spindle vibration frequency spectrum analysis of the machine tool in the vibration monitoring of the numerical control machine tool; and using the frequency domain energy ratio to predict tool wear (1-5Hz component anomaly indicating wear) and the like.

[0229] Figure 5 A structure diagram of an HBA card fault diagnosis device provided in the embodiments of the present application includes an acquisition unit 51, an extraction unit 52, a screening unit 53, a determination unit 54, and a generation unit 55.

[0230] The acquisition unit 51 is configured to acquire time sequence data of the HBA card.

[0231] The extraction unit 52 is configured to extract a time sequence feature set from the time sequence data; wherein the time sequence feature set includes time domain statistical features, frequency domain energy features, and event burst features.

[0232] The screening unit 53 is configured to screen a plurality of time sequence features for fault diagnosis from the time sequence feature set according to the contribution degrees of the features in the time sequence feature set.

[0233] The determination unit 54 is configured to determine a fault label according to the relevance of the plurality of time sequence features to the set fault categories.

[0234] The generation unit 55 is configured to generate fault response information based on the fault label and a fault warning rule matched by the plurality of time sequence features.

[0235] In some embodiments, the acquisition unit includes a collection subunit, a sampling subunit, and a cleaning subunit.

[0236] The collection subunit is configured to collect original time sequence data of the HBA card; wherein the original time sequence data includes hardware sensor data, performance indicators, and log information.

[0237] The sampling subunit is configured to sample the original time sequence data in a dynamic window segmentation manner to obtain a plurality of window time sequence data.

[0238] The cleaning subunit is configured to perform data cleaning on the plurality of window time sequence data to obtain the time sequence data of the HBA card.

[0239] In some embodiments, the sampling subunit is configured to, in an initial state, obtain first window time sequence data from the original time sequence data according to a set window length.

[0240] adjusting a window length of a next window according to the fluctuation index of the time series data in the previous window;

[0241] extracting window time series data of the next window from the original time series data according to the window length of the next window and the overlap rate, until all of the original time series data is extracted.

[0242] In some embodiments, the sampling subunit is configured to, in a case where the fluctuation index of the time series data in the previous window satisfies a data fluctuation condition, adjust the length of the previous window to be the length of the next window according to a set reduction rule;

[0243] In a case where the fluctuation index of the time series data in the previous window satisfies a data fluctuation condition, the length of the previous window is adjusted to be the length of the next window according to a set extension rule;

[0244] In a case where the fluctuation index of the time series data in the previous window does not satisfy the data fluctuation condition and does not satisfy the data fluctuation condition, the length of the previous window is taken as the length of the next window.

[0245] In some embodiments, the sampling subunit is configured to, in a case where the standard deviation of the time series data in the previous window is greater than a first standard threshold and the absolute value of the slope of the time series data in the previous window is greater than a first slope threshold, determine the length of the next window according to a set minimum window value, the length of the previous window, and a corresponding reduction ratio thereof.

[0246] In some embodiments, the sampling subunit is configured to, in a case where the standard deviation of the time series data in the previous window is less than or equal to a second standard threshold and the absolute value of the slope of the time series data in the previous window is less than or equal to a second slope threshold, determine the length of the next window according to a set maximum window value, the length of the previous window, and a corresponding extension ratio thereof; wherein the second standard threshold is less than the first standard threshold, and the second slope threshold is less than the first slope threshold.

[0247] In some embodiments, the cleaning subunit is configured to perform noise filtering on the plurality of window time series data according to a set noise filtering manner to obtain filtered plurality of window time series data;

[0248] Based on a linear interpolation manner, the first type of missing data in the filtered plurality of window time series data is supplemented to obtain a plurality of initially supplemented window time series data; wherein the first type of missing data is data with a number of consecutive missing sampling points less than a set missing value;

[0249] According to the influence weight of the missing data on the historical data, the second type of missing data in the plurality of window time series data after the initial supplement is supplemented to obtain the plurality of window time series data after the final supplement; wherein the second type of missing data is data with a number of continuous missing sampling points greater than or equal to a set missing value;

[0250] The plurality of window time series data after the final supplement is normalized to obtain the time series data of the HBA card.

[0251] In some embodiments, the extraction unit is configured to determine the time domain statistical features based on the distribution trend of the time series data in each window; wherein the time domain statistical features include any one or any combination of mean, variance, skewness, kurtosis, and slope.

[0252] The time series data is converted into a frequency domain signal, and a main frequency component extraction is performed on the frequency domain signal to determine the main frequency energy ratio.

[0253] According to the correlation between the performance indicators and the fault events in the time series data, the event burst features are extracted.

[0254] In some embodiments, the screening unit includes a contribution degree determination subunit and a feature screening subunit as subunits.

[0255] As a subunit, the remaining features in the time series feature set except the target feature are used as a target subset; wherein the target feature is any one of all the features included in the time series feature set.

[0256] The contribution degree determination subunit is configured to determine the contribution degree of the target feature according to the time series feature set, the target subset, and their respective prediction values.

[0257] The feature screening subunit is configured to screen a set number of time series features from the time series feature set in descending order of the contribution degree.

[0258] In some embodiments, the contribution degree determination subunit is configured to use the time series feature set and the target subset as input features of the fault classification model to determine the prediction values corresponding to the time series feature set and the target subset, respectively.

[0259] According to the set weight calculation rule, the time series feature set and the target subset are processed to obtain the weight.

[0260] The difference between the prediction value corresponding to the time series feature set and the prediction value corresponding to the target subset is multiplied by the weight to obtain the contribution degree of the target feature.

[0261] In some embodiments, the determining unit is configured to input the plurality of time sequence features into a fault classification model to obtain a fault probability corresponding to each fault category; the fault classification model is trained based on historical sample data, the historical sample data including a set of historical time sequence features and corresponding fault categories; the fault categories include hardware faults, software faults, and link faults;

[0262] The fault category with the highest fault probability and the fault probability are taken as the fault label.

[0263] In some embodiments, the generating unit includes a matching subunit, an analyzing subunit, and a responding subunit.

[0264] The matching subunit is configured to filter out a target fault warning rule matching the fault label and the target time sequence feature from a set rule library; each fault warning rule in the rule library has a corresponding threshold range.

[0265] The analyzing subunit is configured to analyze the fault label and the target time sequence feature according to the target fault warning rule to determine a fault priority.

[0266] The responding subunit is configured to determine fault response information matching the fault priority according to a response action mapping table; the fault response information includes an automatic response mode and a notification mode.

[0267] In some embodiments, the threshold range includes an error threshold and a temperature slope threshold; to adjust the threshold range, the device further includes a statistical unit, a first determining unit, a second determining unit, and an adjusting unit.

[0268] The statistical unit is configured to statistically determine an average value and a standard deviation of error counts in a set time period.

[0269] The first determining unit is configured to determine the error threshold based on the average value and the standard deviation.

[0270] The second determining unit is configured to determine a compensation coefficient according to a difference between a current room temperature and a standard room temperature.

[0271] The adjusting unit is configured to take a quotient of the temperature slope threshold and the compensation coefficient as an adjusted temperature slope threshold.

[0272] In some embodiments, the matching subunit is configured to filter out a fault warning rule matching the fault label and the target time sequence feature from a set rule library according to a threshold range corresponding to each fault warning rule in the rule library; each fault warning rule has a corresponding rule priority.

[0273] In the case where multiple fault warning rules are matched, the fault warning rule with the highest rule priority is selected as the target fault warning rule.

[0274] In the case that there are multiple failure warning rules with the highest rule priority, the failure warning rule with the highest contribution degree is selected as the target failure warning rule.

[0275] In some embodiments, further comprising a priority adjustment unit;

[0276] The priority adjustment unit is configured to adjust the rule priority of the target failure warning rule according to the number of times of triggering of the target failure warning rule within the period of time.

[0277] In some embodiments, after the failure response information is generated based on the failure warning rule matched based on the failure label and the plurality of time sequence features, further comprising an analysis unit and a root cause determination unit;

[0278] The analysis unit is configured to input the time sequence data and the corresponding time sequence feature set thereof into a bidirectional long short-term memory network model to obtain an output result; wherein the output result comprises a prediction probability of each type of failure and a feature vector for failure mode matching;

[0279] The root cause determination unit is configured to determine a root cause analysis result based on a feature similarity between the output result and each root cause rule in the failure mode library.

[0280] The features of the embodiments of the failure diagnosis device of the HBA card can be referred to the related descriptions of the embodiments of the failure diagnosis method of the HBA card, which will not be repeated here.

[0281] It can be seen from the above technical solution that the timing data of the HBA card is acquired; timing feature sets are extracted from the timing data; wherein the timing feature sets include multi-modal features such as time domain statistical features, frequency domain energy features, and event burst features. The time domain statistical features can capture the overall distribution and trend of the timing data, the frequency domain energy features can capture periodic faults in the timing data, and the event burst features can capture millisecond-level burst faults. The error frequency and I / O delay variance of the HBA card are the most important features for fault detection, the frequency energy ratio and temperature slope have important reference value for HBA card fault reason analysis, and the signal strength fluctuation and link retry number are important parameters for HBA card link quality evaluation, so the time domain statistical features include one or more of the I / O delay variance, the temperature slope, and the link retry number; the frequency domain energy features include the error frequency and / or the frequency energy ratio; and the event burst features include the signal strength fluctuation. By extracting multi-modal timing features, the detection sensitivity is improved, and reliable data support is provided for subsequent fault detection. The timing feature sets contain a large number of timing features, and some timing features have little contribution to fault detection and do not need to be analyzed, so in order to improve the fault detection efficiency, a plurality of timing features used for fault diagnosis can be selected from the timing feature sets according to the contribution of each feature in the timing feature sets. The fault label is determined according to the relevance of the plurality of timing features and the set fault category. The fault label contains the fault category and the fault probability of the HBA card. The timing features associated with different fault categories are different, and in order to realize the automation of the fault, all possible timing feature combination modes under different fault categories can be summarized to set corresponding fault warning rules. The fault warning rules can include corresponding fault handling methods. After the timing features are selected and the fault label is determined, the fault response information can be generated based on the fault label and the fault warning rules matched by the plurality of timing features. The fault response information contains the response action required to solve the current HBA card fault. In the technical solution, the multi-modal timing features are extracted to improve the detection sensitivity. The plurality of timing features and the fault label are comprehensively analyzed based on the fault warning rules, which avoids the misoperation caused by the fluctuation of a single feature, and the fault can be automatically processed in time based on the fault response information, which improves the fault handling efficiency.

[0282] Embodiments of the present application also provide an electronic device comprising a memory and a processor, the memory storing a computer program, and the processor being configured to run the computer program to perform the steps in any of the above-mentioned HBA card fault diagnosis method embodiments.

[0283] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is configured to execute the steps in any of the above-mentioned HBA card fault diagnosis method embodiments when running.

[0284] In an example embodiment, the above-mentioned computer readable storage medium can include, but is not limited to, a U disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0285] The embodiment of the present application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the steps in any of the above-mentioned HBA card fault diagnosis method embodiments.

[0286] The embodiment of the present application further provides another computer program product, which comprises a non-volatile computer readable storage medium, and the non-volatile computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps in any of the above-mentioned HBA card fault diagnosis method embodiments.

[0287] The skilled person can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0288] The above describes in detail the HBA card fault diagnosis method and electronic device provided by the present application. The principles and implementation manners of the present application are described by applying specific examples in this paper, and the above description of the examples is only applicable to help understand the method and core idea of the present application. It should be pointed out that for ordinary skilled person in the technical field, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A method of diagnosing a failure of an HBA card, characterized by, The method comprises the following steps: acquiring timing data of an HBA card; extracting a timing feature set from the timing data; wherein the timing feature set comprises time domain statistical features, frequency domain energy features, and event burst features; the time domain statistical features comprise one or more of I / O delay variance, temperature slope, and link retry count; the frequency domain energy features comprise error frequency and / or main frequency energy ratio; and the event burst features comprise signal strength fluctuation; screening a plurality of timing features for fault diagnosis from the timing feature set according to the contribution degrees of the features in the timing feature set; determining a fault label according to the relevance of the plurality of timing features to a set fault category; screening a target fault early warning rule matching the fault label and a target timing feature from a set rule base; wherein each fault early warning rule in the rule base has a corresponding threshold range; the threshold range comprises a threshold range of fault probability and a threshold range of each timing feature; analyzing the fault label and the target timing feature according to the target fault early warning rule to determine a fault priority; determining fault response information matching the fault priority according to a response action mapping table; wherein the fault response information comprises an automatic response mode and a notification mode; inputting the timing data and the corresponding timing feature set into a bidirectional long short-term memory network model to obtain an output result; wherein the output result comprises a prediction probability of each fault and a feature vector for fault pattern matching; determining a root cause analysis result based on the feature similarity between the output result and each root cause rule in a fault pattern library.

2. The method of claim 1, wherein, The method for acquiring timing data of an HBA card comprises the following steps: collecting original timing data of the HBA card; wherein the original timing data comprises hardware sensor data, performance indicators, and log information; sampling the original timing data according to a dynamic window segmentation method to obtain a plurality of window timing data; performing data cleaning on the plurality of window timing data to obtain the timing data of the HBA card.

3. The method of claim 2, wherein, The method for sampling the original timing data according to a dynamic window segmentation method to obtain a plurality of window timing data comprises the following steps: in an initial state, a first window timing data is obtained by cutting the original timing data according to a set window length; adjusting the window length of a next window according to the fluctuation index of the timing data in a previous window; cutting the window timing data of the next window from the original timing data according to the window length of the next window and an overlap rate until all of the original timing data is cut.

4. The method of claim 3, wherein, The method for adjusting the window length of a next window according to the fluctuation index of the timing data in a previous window comprises the following steps: in a case where the fluctuation index of the timing data in the previous window meets a data fluctuation condition, the length of the previous window is reduced as the window length of the next window according to a set reduction rule; in a case where the fluctuation index of the timing data in the previous window meets a data stability condition, the length of the previous window is increased as the window length of the next window according to a set extension rule; In a case where the fluctuation index of the time series data in the previous window does not satisfy the data fluctuation condition and does not satisfy the data stationary condition, the length of the previous window is taken as the window length of the next window.

5. The method of diagnosing a failure of an HBA card according to claim 4, wherein, In a case where the fluctuation index of the time series data in the previous window satisfies the data fluctuation condition, the length of the previous window is reduced according to a set reduction rule to be taken as the window length of the next window, including: In a case where the standard deviation of the time series data in the previous window is greater than a first standard threshold and the absolute value of the slope of the time series data in the previous window is greater than a first slope threshold, the window length of the next window is determined according to a set window minimum value, the length of the previous window and a corresponding reduction ratio thereof.

6. The method of diagnosing a failure of an HBA card according to claim 5, wherein, In a case where the fluctuation index of the time series data in the previous window satisfies the data stationary condition, the length of the previous window is increased according to a set extension rule to be taken as the window length of the next window, including: In a case where the standard deviation of the time series data in the previous window is less than or equal to a second standard threshold and the absolute value of the slope of the time series data in the previous window is less than or equal to a second slope threshold, the window length of the next window is determined according to a set window maximum value, the length of the previous window and a corresponding extension ratio thereof; wherein the second standard threshold is less than the first standard threshold, and the second slope threshold is less than the first slope threshold.

7. The method of diagnosing a failure of an HBA card according to claim 2, wherein, The plurality of window time series data is subjected to data cleaning to obtain the time series data of the HBA card, including: The plurality of window time series data is subjected to noise filtering according to a set noise filtering mode to obtain filtered plurality of window time series data; First type missing data in the filtered plurality of window time series data is supplemented based on a linear interpolation mode to obtain initially supplemented plurality of window time series data; wherein the first type missing data is data with a number of continuously missing sampling points less than a set missing value; Second type missing data in the initially supplemented plurality of window time series data is supplemented according to an influence weight of historical data on missing data to obtain finally supplemented plurality of window time series data; wherein the second type missing data is data with a number of continuously missing sampling points greater than or equal to a set missing value; The finally supplemented plurality of window time series data is subjected to normalization processing to obtain the time series data of the HBA card.

8. The method of claim 1, wherein, From the contribution degrees of each feature in the time series feature set, a plurality of time series features for fault diagnosis are screened out from the time series feature set, including: Remaining features except a target feature in the time series feature set are taken as a target subset; wherein the target feature is any one of all features contained in the time series feature set; The contribution degree corresponding to the target feature is determined according to the time series feature set, the target subset and their respective corresponding prediction values; A set number of time series features are screened out from the time series feature set in order from high to low according to the contribution degrees.

9. The method of diagnosing a failure of an HBA card according to claim 8, wherein, The contribution degree corresponding to the target feature is determined according to the time series feature set, the target subset and their respective corresponding prediction values, including: The time series feature set and the target subset are respectively taken as input features of a fault classification model to determine the prediction values corresponding to the time series feature set and the target subset respectively; According to a set weight calculation rule, the time sequence feature set and the target subset are processed to obtain a weight; A difference between a predicted value corresponding to the time sequence feature set and a predicted value corresponding to the target subset is multiplied by the weight to obtain a contribution degree corresponding to the target feature.

10. The method of claim 1, wherein, The threshold range includes an error threshold and a temperature slope threshold; For adjustment of the threshold range, the method further includes: Statistically obtaining an average value and a standard deviation of error counts in a set time period; Based on the average value and the standard deviation, an error threshold is determined; According to a difference between a current room temperature and a standard room temperature, a compensation coefficient is determined; A quotient of the temperature slope threshold and the compensation coefficient is taken as an adjusted temperature slope threshold.

11. The method of diagnosing a failure of an HBA card according to claim 1, wherein, From a set rule library, a target fault early warning rule matching the fault label and the target time sequence feature is screened out, including: According to a threshold range corresponding to each fault early warning rule in the rule library, a fault early warning rule matching the fault label and the target time sequence feature is screened out from the set rule library; each fault early warning rule has a corresponding rule priority; In a case where there are multiple matched fault early warning rules, a fault early warning rule with the highest rule priority is selected as the target fault early warning rule; In a case where there are multiple fault early warning rules with the highest rule priority, a fault early warning rule with the highest contribution degree is selected as the target fault early warning rule.

12. The method of claim 11, wherein, Further including: According to a number of times of triggering of the target fault early warning rule in a cycle time, a rule priority of the target fault early warning rule is adjusted.

13. An electronic device, comprising: Including: A memory for storing a computer program; A processor for executing the computer program to implement the steps of the fault diagnosis method of the HBA card according to any one of claims 1 to 12.

Images