Component failure prediction method, device, computer program product, and storage medium

Abstract

The application discloses a kind of component failure prediction method, equipment, computer program product and storage medium, belong to server field, for predicting target component failure based on multiple associated sensing features, solve the problem of poor component failure prediction accuracy.Considering that the component failure of server can be reflected through multiple associated sensing features, therefore, the application responds to the prediction instruction of the target firmware failure of the future target time, determines its prediction value and the numerical level of prediction value by synthesizing multiple associated sensing features, then obtains the failure probability prediction result according to the weighted result of numerical level, avoids predicting component failure through single sensing data, fully utilizes multidimensional data to realize the prediction of target component failure, so as to improve the prediction accuracy of component failure, and ensure the stable operation of server.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of servers, in particular to a component failure prediction method, device, computer program product and storage medium. BACKGROUND

[0002] The server has multiple components, and the reliability of the components also affects the reliability of the server, so it is necessary to predict the failure of the components. However, there is a lack of a mature component failure prediction scheme in the related art, so that the component failure prediction of the server has the problem of poor accuracy, which reduces the reliability of the server.

[0003] Therefore, how to provide a scheme to solve the above technical problems is a problem that those skilled in the art need to solve at present. SUMMARY

[0004] The purpose of the present application is to provide a component failure prediction method, device, computer program product and storage medium. In response to a prediction instruction for a target firmware failure at a future target time, the present application determines the prediction value and the numerical level of the prediction value by comprehensively considering multiple associated sensor features, and then obtains a failure probability prediction result according to the weighted result of the numerical level, thereby avoiding component failure prediction based on single sensor data, fully utilizing multi-dimensional data to predict target component failure, and improving the prediction accuracy of component failure and ensuring stable operation of the server.

[0005] To solve the above technical problems, the present application provides a component failure prediction method, comprising:

[0006] In response to a prediction instruction for a target component failure at a future target time, for any associated sensor feature of the target component failure, a prediction value of the associated sensor feature at the target time is determined according to historical data, wherein the target component failure has multiple associated sensor features;

[0007] For any prediction value, the numerical level of the prediction value within the reference value range of the associated sensor feature to which the prediction value belongs is determined;

[0008] According to the weighted result of the numerical level of each prediction value, a failure probability prediction result of the target component failure is determined.

[0009] On the other hand, for any prediction value, determining the numerical level of the prediction value within the reference value range of the associated sensor feature to which the prediction value belongs comprises:

[0010] For any prediction value, the numerical span of the associated sensor feature to which the prediction value belongs at the first preset number of sampling points in the past is taken as the reference value range of the associated sensor feature to which the prediction value belongs;

[0011] The reference value range corresponding to the predicted value is divided into a second preset number of continuous numerical intervals, wherein each numerical interval is sequentially sorted in ascending order of numerical value;

[0012] The sequence number of the numerical interval closest to the predicted value is taken as the numerical level of the predicted value.

[0013] On the other hand, in response to a prediction instruction for a target component fault at a future target time, for any associated sensor feature of the target component fault, determining a predicted value of the associated sensor feature at the target time according to historical data includes:

[0014] In response to a prediction instruction for a target component fault at a future target time, for any associated sensor feature of the target component fault, determining a rate of change of the associated sensor feature according to historical data;

[0015] According to the rate of change of the associated sensor feature and the time difference between the target time and the current time, the predicted value of the associated sensor feature at the target time is determined.

[0016] On the other hand, in response to a prediction instruction for a target component fault at a future target time, for any associated sensor feature of the target component fault, determining a rate of change of the associated sensor feature according to historical data includes:

[0017] In response to a prediction instruction for a target component fault at a future target time, for any associated sensor feature of the target component fault, determining a rate of change of the associated sensor feature according to sampling values of the associated sensor feature at a first preset number of past sampling points.

[0018] On the other hand, dividing the reference value range corresponding to the predicted value into a second preset number of continuous numerical intervals includes:

[0019] According to the binning type, the second preset number, and the reference value range corresponding to the predicted value, the boundaries of each numerical interval are determined;

[0020] According to the boundaries of each numerical interval, the reference value range corresponding to the predicted value is divided into a second preset number of continuous numerical intervals.

[0021] On the other hand, according to the binning type, the second preset number, and the reference value range corresponding to the predicted value, the boundaries of each numerical interval are determined, which includes:

[0022] When the binning type is equal-frequency binning, the total number of data points in the reference value range is counted;

[0023] divide the total number by the second preset number to obtain a number of data points expected to be contained in each value interval;

[0024] sort the data points in the reference value range according to the value size, divide the data points in sequence from the minimum value, so that the actual number of data points contained in each value interval is equal to the expected number of data points contained, thereby determining the boundary of each value interval.

[0025] On the other hand, according to the weighted result of the value level of each prediction value, the fault probability prediction result of the target component fault is determined, including:

[0026] weighting the weighted result of each prediction value, and taking the weighted result as the correlation degree;

[0027] According to the correlation degree and the prediction value of each associated sensing feature, index in the preset decision tree model of the target component fault to determine the fault probability prediction result of the target component fault.

[0028] On the other hand, the construction of the decision tree model satisfies the following conditions:

[0029] The root node is the correlation degree feature, and the leaf node is the fault probability;

[0030] When the correlation degree is greater than a preset threshold, a branch path is selected according to the prediction value of at least one associated sensing feature;

[0031] The fault probability value of the leaf node is generated by a hierarchical weighted voting mechanism.

[0032] On the other hand, in response to a prediction instruction for a target component fault at a future target time, for any associated sensing feature of the target component fault, the prediction value of the associated sensing feature at the target time is determined according to historical data, including:

[0033] In response to a prediction instruction for a target component fault at a future target time, each associated sensing feature corresponding to the target component fault is determined from a preset correlation analysis matrix, wherein the preset correlation analysis matrix includes the correspondence between the component fault and the associated sensing feature;

[0034] For any associated sensing feature of the target component fault, the prediction value of the associated sensing feature at the target time is determined according to historical data.

[0035] On the other hand, the correspondence between the component fault and the associated sensing feature includes:

[0036] The associated sensing features of the memory fault include the error check and error correction rate of the memory, the memory voltage fluctuation amplitude, and the memory particle temperature;

[0037] The associated sensing features of the hard disk failure include a cyclic redundancy check error rate, a power supply ripple amplitude, a fan vibration frequency and a hard disk temperature.

[0038] On the other hand, the failure probability prediction result of the target component failure is determined according to a weighted result of the numerical level of each prediction value.

[0039] A weight matrix of the target component failure is determined from a preset associated analysis matrix, wherein the preset associated analysis matrix includes a corresponding relationship between component failures and weight matrices, and the weight matrix includes weights of each associated sensing feature of the target component failure.

[0040] A weighted result of the numerical level of each prediction value according to the weight matrix of the target component failure is taken as the failure probability prediction result of the target component failure.

[0041] On the other hand, the component failure prediction method further includes:

[0042] The weight matrix of the target component failure is adjusted according to a third preset number of groups of past verification data groups of the target component failure in combination with a gradient descent method.

[0043] The verification data group is a failure probability prediction result and a corresponding actual failure occurrence result, and the actual failure occurrence result includes occurrence or non-occurrence.

[0044] On the other hand, adjusting the weight matrix of the target component failure according to a third preset number of groups of past verification data groups of the target component failure in combination with a gradient descent method includes:

[0045] For each group of data in the third preset number of groups of verification data groups, the failure probability prediction result is substituted into a preset loss function to calculate a loss value corresponding to the group of data, wherein the loss function is used to measure the difference between the failure probability prediction result and the actual failure occurrence result.

[0046] The sum of the loss values corresponding to all the verification data groups is taken as a total loss, and the gradient of each weight in the weight matrix of the target component failure is calculated based on the total loss.

[0047] According to a preset learning rate, each weight in the weight matrix of the target component failure is updated in the opposite direction of the gradient to complete the adjustment of the weight matrix of the target component failure.

[0048] On the other hand, the application is applied to an auxiliary processor.

[0049] The auxiliary processor is arranged on a board card where a baseboard management controller is located, and the auxiliary processor is connected to a high extensible interface bus on the board card where the baseboard management controller is located.

[0050] The component failure prediction method further comprises:

[0051] The sensor data is collected by each peripheral controller over the Advanced Extensible Interface Bus.

[0052] In another aspect, collecting the sensor data by each peripheral controller over the Advanced Extensible Interface Bus comprises:

[0053] The error checking and correction error rate, operating voltage fluctuation, and grain temperature of the memory are collected by the integrated circuit bus controller and / or the enhanced integrated circuit bus controller.

[0054] The fan vibration frequency is collected by the fan speed control interface controller.

[0055] The voltage ripple amplitude of the power supply is collected by the power management bus controller.

[0056] The cyclic redundancy check error rate and the hard disk temperature of the hard disk are collected by the hard disk interface controller.

[0057] In another aspect, the component failure prediction method further comprises:

[0058] The failure probability prediction result of the target component failure is pushed to the baseboard management controller, so that the baseboard management controller alarms according to the failure probability prediction result and / or pushes the failure probability prediction result to a remote control terminal.

[0059] To solve the above technical problems, the application further provides a component failure prediction device, comprising:

[0060] A memory for storing a computer program.

[0061] A processor for executing the computer program to realize the steps of the component failure prediction method as described above.

[0062] In another aspect, the processor is an auxiliary processor arranged on a board card where the baseboard management controller is located.

[0063] The sensor data is collected by each peripheral controller over the Advanced Extensible Interface Bus.

[0064] To solve the above technical problems, the application further provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, realizes the steps of the component failure prediction method as described above.

[0065] To solve the above technical problems, 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 the steps of the component failure prediction method.

[0066] Beneficial effects: The application provides a component failure prediction method, which considers that the component failure of a server can be reflected through multiple associated sensing features, so that the application, in response to a prediction instruction of a target firmware failure at a future target time, determines the prediction value and the numerical level of the prediction value by comprehensively considering the multiple associated sensing features, and then obtains a failure probability prediction result according to the weighted result of the numerical level, thereby avoiding component failure prediction through single sensing data, fully utilizing multi-dimensional data to realize prediction of the target component failure, and thus improving the prediction accuracy of the component failure and ensuring stable operation of the server.

[0067] The application further provides a component failure prediction device, a computer program product and a storage medium, which have the same beneficial effects as the component failure prediction method. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the application, the related art and the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative effort.

[0069] Figure 1 A flowchart of a component failure prediction method provided by the application is shown in the figure.

[0070] Figure 2 A structure diagram of a baseboard management controller provided by the application is shown in the figure.

[0071] Figure 3 A structure diagram of a component failure prediction system provided by the application is shown in the figure.

[0072] Figure 4 A flowchart of another component failure prediction method provided by the application is shown in the figure.

[0073] Figure 5 A structure diagram of a component failure prediction device provided by the application is shown in the figure.

[0074] Figure 6 A structure diagram of a computer readable storage medium provided by the application is shown in the figure. DETAILED DESCRIPTION

[0075] The core of the present application is to provide a component failure prediction method, device, computer program product and storage medium, in response to a prediction instruction of a target firmware failure at a future target time, the prediction value and the numerical level of the prediction value are determined by comprehensively considering multiple associated sensing features, and then the failure probability prediction result is obtained according to the weighted result of the numerical level, so that the component failure prediction is avoided by using single sensing data, and the target component failure is predicted by fully utilizing multi-dimensional data, so that the prediction accuracy of the component failure can be improved, and the stable operation of the server is ensured.

[0076] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, 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 some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0077] For the present application Figure 1 , Figure 1 A flowchart of a component failure prediction method provided by the present application is shown in the figure, and the component failure prediction method comprises the following steps.

[0078] S101: In response to a prediction instruction of a target component failure at a future target time, for any associated sensing feature of the target component failure, a prediction value of the associated sensing feature at the target time is determined according to historical data, wherein the target component failure has multiple associated sensing features.

[0079] Specifically, considering the technical problems in the background art and the fact that the component failure of the server can be reflected by multiple associated sensing features, in the embodiments of the present application, the failure probability of the target component failure at the future target time is predicted by using the multiple associated sensing features related to the target component failure, and the prediction value of each associated sensing feature can be used to accurately predict the component failure probability based on multi-dimensional data. Therefore, in this step, in response to the prediction instruction of the target component failure at the future target time, for any associated sensing feature of the target component failure, the prediction value of the associated sensing feature at the target time is determined according to historical data.

[0080] S102: For any prediction value, a numerical level of the prediction value in a reference value range of the associated sensing feature is determined.

[0081] Specifically, in order to simplify the calculation and improve the prediction efficiency, in the embodiment of the present application, for any prediction value, the numerical level of the prediction value in the reference value range of the associated sensing feature can be determined, so as to take the numerical level as the data basis for predicting the component failure probability in the subsequent step.

[0082] S103: According to the weighted result of the numerical level of each prediction value, the failure probability prediction result of the target component failure is determined.

[0083] Specifically, after determining the numerical level of each prediction value in the reference value range of the associated sensing feature, the weighted result of the numerical level of each prediction value can be obtained by simple weighting calculation, and the failure probability prediction result of the target component failure is determined. By grading and weighting the prediction values of multiple associated sensing features, a cross-component associated failure mapping mode is constructed to adapt to the dynamic changes of the server running environment.

[0084] Specifically, the component failure prediction method can be applied to an auxiliary processor (such as a neural processing unit (NPU, Neural Processing Unit)), in a heterogeneous hardware architecture with a baseboard management controller (BMC, Baseboard Management Controller) embedded neural processing unit (NPU, Neural Processing Unit), the NPU controls peripheral controllers such as advanced extensible interface (AXI, Advanced eXtensible Interface) bus and inter-integrated circuit (I2C, Inter-Integrated Circuit) bus controllers, and real-time collects running parameters such as error checking and correction (ECC, Error Checking and Correction) error rate of memory. For memory failure prediction at a future target time, calculate the change rate of memory particle temperature according to historical data, combine the time difference between the target time and the current time to obtain the temperature prediction value at the target time; take the numerical span of the prediction value at the first preset number of sampling points in the past as the reference value range, divide it into the second preset number of continuous numerical intervals, and determine the serial number of the interval to which the prediction value belongs as the numerical level; based on the weight matrix corresponding to the memory failure in the preset correlation analysis matrix, weight the numerical level of each associated sensing feature (such as ECC error rate, memory voltage fluctuation amplitude, memory particle temperature) to obtain the failure probability prediction result.

[0085] Specifically, in order to better illustrate the embodiment of the present application, please refer to Figure 2 , Figure 2 A structural diagram of a baseboard management controller provided by the present application, Figure 2The server processor in the server can collect sensing data through the first to fourth peripheral controllers on the high-order scalable interface bus and execute the component failure prediction method without the central processor of the baseboard management controller, thereby reducing the burden of the central processor.

[0086] The present application provides a component failure prediction method, which considers that the component failure of a server can be reflected through multiple associated sensing features, and thus, in response to a prediction instruction of a target firmware failure at a future target time, the present application determines the prediction value and the numerical level of the prediction value by comprehensively considering the multiple associated sensing features, and then obtains the failure probability prediction result according to the weighted result of the numerical level, thereby avoiding component failure prediction through single sensing data, fully utilizing multi-dimensional data to realize prediction of the target component failure, and thus, the prediction accuracy of the component failure can be improved, and the stable operation of the server can be ensured.

[0087] On the basis of the above-mentioned embodiments:

[0088] As an optional embodiment, for any prediction value, the numerical level of the prediction value in the reference value range of the associated sensing feature to which the prediction value belongs includes:

[0089] For any prediction value, the numerical span of the associated sensing feature to which the prediction value belongs at the first preset number of sampling points in the past is taken as the reference value range of the associated sensing feature to which the prediction value belongs.

[0090] The reference value range corresponding to the prediction value is divided into continuous second preset number of numerical intervals, wherein each numerical interval is sequentially sorted in ascending order.

[0091] The sequence number of the nearest numerical interval of the prediction value is taken as the numerical level of the prediction value.

[0092] Specifically, in order to convert the prediction value into a quantifiable feature level and improve the accuracy of failure prediction, the embodiments of the present application can dynamically determine the reference value range and divide the numerical intervals, so that the numerical level can more accurately reflect the position of the prediction value in the historical data distribution.

[0093] Specifically, taking hard disk failure prediction as an example, for the prediction value of the hard disk cyclic redundancy check (CRC) error rate, the numerical span of the first preset number (such as 5) of sampling points in the past is taken as the reference value range. Assuming that the range is [10, 50], the second preset number is 2, the equal frequency bin type is adopted, the total number of data points in the range is counted and divided into 2 intervals, and the interval boundary is determined to be 30. If the prediction value is 35, the nearest interval is [30, 50], the sequence number is 2, and the numerical level is 2.

[0094] Of course, in addition to this specific form, "for any predicted value, determine the numerical level of the predicted value within the reference value range of the associated sensing feature to which the predicted value belongs" can also be in other forms, which are not limited herein by embodiments of the present application.

[0095] As an optional embodiment, in response to the prediction instruction for the target component failure at the future target time, for any associated sensing feature of the target component failure, determining the predicted value of the associated sensing feature at the target time according to the historical data comprises:

[0096] In response to the prediction instruction for the target component failure at the future target time, for any associated sensing feature of the target component failure, determining the rate of change of the associated sensing feature according to the historical data;

[0097] According to the rate of change of the associated sensing feature and the time difference between the target time and the current time, the predicted value of the associated sensing feature at the target time is determined.

[0098] Specifically, to predict the future parameter value based on the historical data trend and solve the problem that the traditional static threshold cannot adapt to the change of the hardware state, in embodiments of the present application, the predicted value can be made more consistent with the dynamic trend of the component operation through the calculation of the rate of change and the time difference.

[0099] For example, when predicting the target time value of the memory particle temperature, the rate of change thereof is calculated according to the sampling values of the temperature at the first preset number (such as 5) of sampling points in the past. For example, the temperature rises from 72℃ to 90℃ in the past 20 seconds, and the rate of change is (90-72) / 20=0.9℃ / second. The target time is 30 seconds away from the current time, and the predicted value is 90+0.9x30=117℃ (which can be adjusted in combination with the hardware safety threshold in actual application).

[0100] Of course, in addition to this specific form, "in response to the prediction instruction for the target component failure at the future target time, for any associated sensing feature of the target component failure, determining the predicted value of the associated sensing feature at the target time according to the historical data" can also be in other forms, which are not limited herein by embodiments of the present application.

[0101] As an optional embodiment, in response to the prediction instruction for the target component failure at the future target time, for any associated sensing feature of the target component failure, determining the rate of change of the associated sensing feature according to the historical data comprises:

[0102] In response to the prediction instruction for the target component failure at the future target time, for any associated sensing feature of the target component failure, the rate of change of the associated sensing feature is determined according to the sampling values of the associated sensing feature at the first preset number of sampling points in the past.

[0103] Specifically, to ensure the accuracy and reliability of the change rate calculation, the statistical analysis can be performed based on sufficient historical sampling data in the embodiments of the present application, so that the change rate can truly reflect the dynamic change trend of the associated sensing features.

[0104] Specifically, for example, for the change rate calculation of the power supply voltage ripple amplitude, the values of the first preset number (such as 10) sampling points in the past are obtained, and a linear regression method is used to fit the data trend to obtain the change rate per unit time. For example, the ripple amplitude of 10 sampling points increases from 20mV to 40mV in 50 seconds, and the change rate is (40-20) / 50=0.4mV / s.

[0105] Of course, in addition to this specific form, "determining the change rate of the associated sensing feature according to historical data for any associated sensing feature of the target component failure in response to the prediction instruction for the target component failure of the future target moment" can also be in other forms, which are not limited in the embodiments of the present application.

[0106] As an optional embodiment, dividing the reference value range corresponding to the prediction value into a continuous second preset number of value intervals comprises:

[0107] According to the binning type, the second preset number, and the reference value range corresponding to the prediction value, the boundaries of each value interval are determined;

[0108] According to the boundaries of each value interval, the reference value range corresponding to the prediction value is divided into a continuous second preset number of value intervals.

[0109] Specifically, to adapt to different data distribution characteristics, the embodiments of the present application can convert continuous sensing data into discrete interval features through flexible configuration of the binning type, number and range, which is convenient for subsequent association calculation and model reasoning.

[0110] Specifically, for example, for the prediction value of the fan vibration frequency, the reference value range is determined to be [20, 80] Hz, the binning type is equal frequency binning, and the second preset number is 3. The total number of data points in this range is 60, and each interval is expected to contain 20 data points. The data is sorted by value, and the interval [20, 35], [35, 55], and [55, 80] is obtained by dividing from the minimum value, and the interval boundaries are 35 and 55, respectively.

[0111] Of course, in addition to this specific form, "dividing the reference value range corresponding to the prediction value into a continuous second preset number of value intervals" can also be in other forms, which are not limited in the embodiments of the present application.

[0112] As an optional embodiment, the method for determining the boundaries of each numerical interval comprises the following steps:

[0113] When the binning type is equal-frequency binning, the total number of data points in the reference value range is counted;

[0114] The total number is divided by the second preset number to obtain the number of data points expected to be contained in each numerical interval;

[0115] The data points in the reference value range are sorted according to the numerical value, and the data points are divided in turn from the smallest value, so that the actual number of data points contained in each numerical interval is equal to the expected number of data points, thereby determining the boundaries of each numerical interval.

[0116] Specifically, in order to make the binning result more uniformly reflect the data distribution density, the equal-frequency binning method is used in the embodiment of the application to ensure that each interval contains a similar number of data points, thereby avoiding feature quantization deviation caused by uneven data distribution.

[0117] For example, taking the memory voltage fluctuation amplitude as an example, there are 100 data points in the reference value range, the second preset number is 4, and each interval is expected to contain 25 data points. The data is sorted from small to large according to the numerical value, and the maximum value of the first 25 data is 15mV, which is used as the upper limit of the first interval; the maximum value of the next 25 data is 25mV, which is used as the upper limit of the second interval, and so on, to determine the interval boundaries as 15, 25, 35mV, forming the interval division of [0, 15], [15, 25], [25, 35], [35, +∞].

[0118] Of course, in addition to this specific form, "determining the boundaries of each numerical interval according to the binning type, the second preset number, and the reference value range corresponding to the prediction value" can also be other forms, which are not limited in the embodiment of the application.

[0119] As an optional embodiment, the method for determining the boundaries of each numerical interval comprises the following steps:

[0120] The weighted results of each prediction value are weighted, and the weighted results are used as the correlation degree;

[0121] According to the correlation degree and the prediction value of each associated sensing feature, the decision tree model of the target component fault is indexed to determine the fault probability prediction result of the target component fault.

[0122] Specifically, to combine the correlation degree and the decision tree model for fault probability reasoning, the correlation degree can be used to reflect the comprehensive influence of multiple features in the embodiment of the application, and the accuracy and interpretability of fault prediction are improved through hierarchical judgment of the decision tree model.

[0123] In the memory fault prediction, the numerical levels of the associated sensing features (ECC error rate, memory voltage fluctuation amplitude, memory particle temperature) are 2, 2, and 2 respectively, the weight matrix is [0.4, 0.3, 0.3], and the correlation degree is 2 x 0.4 + 2 x 0.3 + 2 x 0.3 = 2.0. In the preset decision tree model, the root node is the correlation degree, and when the correlation degree > 1.5, it is judged whether the temperature is > 80℃. If the temperature prediction value is 97℃, the fault probability is inferred to be 0.9.

[0124] Of course, in addition to this specific form, "determining the fault probability prediction result of the target component fault according to the weighted result of the numerical level of each prediction value" can also be other forms, which are not limited in the embodiment of the application.

[0125] As an optional embodiment, the construction of the decision tree model satisfies the following conditions:

[0126] The root node is the correlation degree feature, and the leaf node is the fault probability;

[0127] When the correlation degree is greater than a preset threshold, a branch path is selected according to the prediction value of at least one associated sensing feature;

[0128] The fault probability value of the leaf node is generated by a hierarchical weighted voting mechanism.

[0129] Specifically, to construct an efficient fault reasoning model, the correlation degree is taken as the root node in the embodiment of the application, and the fault probability is generated by a hierarchical weighted voting mechanism, so that the model can quickly traverse the decision path and adapt to online real-time reasoning requirements.

[0130] Specifically, in the decision tree model corresponding to the memory fault, the root node is the correlation degree feature, when the correlation degree ≤ 1.5, the leaf node fault probability is 0.1; when the correlation degree > 1.5, the memory particle temperature is further judged, if ≤ 80℃, the fault probability is 0.4, if > 80℃, it is 0.9. The model is realized by LightGBM (Light Gradient Boosting Machine) reasoning core to realize hardware acceleration traversal.

[0131] Of course, in addition to this specific form, "the construction condition of the decision tree model" can also be other forms, which are not limited in the embodiment of the application.

[0132] As an optional embodiment, in response to the prediction instruction of the target component fault at the target time point in the future, the determining of the predicted value of the associated sensing feature of the target component fault at the target time point according to the historical data comprises:

[0133] In response to the prediction instruction of the target component fault at the target time point in the future, the corresponding associated sensing features of the target component fault are determined from the preset associated analysis matrix, wherein the preset associated analysis matrix comprises the corresponding relationship between the component fault and the associated sensing feature.

[0134] The predicted value of the associated sensing feature of the target component fault at the target time point is determined according to the historical data.

[0135] Specifically, in order to determine the associated sensing features of the target component, the corresponding relationship between the component fault and the feature can be established by the preset associated analysis matrix in the embodiment of the application, the associated modeling analysis of the multi-component state is realized, and the limitation of single component detection is solved.

[0136] For example, the associated sensing features corresponding to the memory fault are obtained from the preset associated analysis matrix, including the memory ECC error rate, the memory voltage fluctuation amplitude and the memory particle temperature. The memory ECC error rate data is collected by the I2C bus controller, the sampling period is 5 seconds each time, and 5 time point data are continuously collected for subsequent prediction calculation.

[0137] As an optional embodiment, the corresponding relationship between the component fault and the associated sensing feature comprises:

[0138] The associated sensing features of the memory fault include the error check and correction error rate of the memory, the memory voltage fluctuation amplitude and the memory particle temperature.

[0139] The associated sensing features of the hard disk fault include the cyclic redundancy check error rate of the hard disk, the power supply ripple amplitude, the fan vibration frequency and the hard disk temperature.

[0140] Specifically, in order to accurately predict the fault features of different components, the specific associated sensing features are defined according to the hardware physical characteristics in the embodiment of the application, so that the fault prediction is more targeted and accurate.

[0141] Specifically, the associated sensing features of the hard disk fault include the hard disk CRC error rate, the power supply ripple amplitude, the fan vibration frequency and the hard disk temperature. The hard disk CRC error rate is collected by the hard disk interface controller, the power supply ripple amplitude is collected by the power management bus (PMBus, Power Management Bus) controller, and the fan vibration frequency is collected by the fan speed control interface (FANTACHO) controller.

[0142] Of course, in addition to the specific form, the "correspondence between component failure and associated sensing features" can also be in other forms, which are not limited herein.

[0143] As an optional embodiment, the determination of the failure probability prediction result of the target component failure comprises:

[0144] The weight matrix of the target component failure is determined from the preset correlation analysis matrix, wherein the preset correlation analysis matrix comprises a correspondence between component failures and weight matrices, and the weight matrix comprises weights of each associated sensing feature of the target component failure.

[0145] The weighted result of the numerical level of each prediction value according to the weight matrix of the target component failure is taken as the failure probability prediction result of the target component failure.

[0146] Specifically, to realize the weighted fusion of multiple features, the weight matrix in the preset correlation analysis matrix can be used to weight the numerical level of each feature, so as to reflect the difference in the influence degree of different features on the failure and improve the prediction accuracy.

[0147] Specifically, for example, the weight matrix [0.3, 0.2, 0.2, 0.3] corresponding to the hard disk failure is obtained from the preset correlation analysis matrix, which corresponds to the weights of the hard disk CRC error rate, power supply ripple amplitude, fan vibration frequency and hard disk temperature. The numerical levels of each feature are 3, 2, 2 and 3 respectively, and the weighted result is 3x0.3+2x0.2+2x0.2+3x0.3=2.6, which is taken as the probability prediction result of the hard disk failure.

[0148] As an optional embodiment, the component failure prediction method further comprises:

[0149] According to the third preset number of past verification data groups of the target component failure, the weight matrix of the target component failure is adjusted in combination with the gradient descent method.

[0150] The verification data group is the failure probability prediction result and the corresponding actual failure occurrence result, and the actual failure occurrence result includes occurrence or non-occurrence.

[0151] Specifically, to enable the weight matrix to adapt to the real-time running environment of the server, the historical verification data and the gradient descent method are used for dynamic adjustment, so as to improve the adaptability of the model to the change of the hardware state and the prediction accuracy.

[0152] For example, for the weight matrix [0.4, 0.3, 0.3] of memory failure, the past third preset number (such as 100 groups) of verification data sets are collected, each group containing the failure probability prediction result and the actual failure occurrence result. The prediction result is substituted into the cross-entropy loss function, the total loss is calculated, and the weight matrix is updated based on the gradient descent method with a learning rate of 0.05. If the prediction probability of a certain group of data is 0.9 and the actual failure occurs, the adjusted weight may become [0.42, 0.32, 0.26].

[0153] As an optional embodiment, adjusting the weight matrix of the target component failure according to the past third preset number of groups of verification data of the target component failure and combining the gradient descent method comprises:

[0154] For each group of data in the third preset number of groups of verification data, the failure probability prediction result is substituted into the preset loss function to calculate the loss value corresponding to the group of data, wherein the loss function is used to measure the difference between the failure probability prediction result and the actual failure occurrence result.

[0155] The sum of the loss values corresponding to all the verification data groups is taken as the total loss, and the gradient of each weight in the weight matrix of the target component failure is calculated based on the total loss.

[0156] According to the preset learning rate, each weight in the weight matrix of the target component failure is updated in the opposite direction of the gradient to complete the adjustment of the weight matrix of the target component failure.

[0157] Specifically, to ensure the scientificity and effectiveness of the weight matrix adjustment, the embodiments of the present application measure the difference between the prediction and the actual by the loss function, update the weights based on the gradient calculation and the learning rate, and realize the optimization iteration of the weight matrix.

[0158] For example, for the third preset number (such as 50 groups) of verification data, the loss value (such as using the mean square error loss function) of each group is calculated, and the total loss is the sum of all group losses. The gradient of each weight in the weight matrix is calculated, for example, the gradient of the weight w1 is The weight is updated in the opposite direction of the gradient: Wherein η is the learning rate 0.01, and the iterative adjustment of the weight matrix is completed.

[0159] As an optional embodiment, it is applied to an auxiliary processor;

[0160] The auxiliary processor is arranged on a board card where the baseboard management controller is located, and the auxiliary processor is connected with the advanced extensible interface bus on the board card where the baseboard management controller is located.

[0161] The component failure prediction method further comprises:

[0162] Collecting the sensing data through each peripheral controller on the advanced extensible interface bus.

[0163] Specifically, to realize efficient data collection and fault prediction, the component fault prediction method can be applied to an auxiliary processor in the embodiment of the application, peripheral controllers are connected through an AXI bus, BMC CPU resource occupation is reduced, and system operation efficiency is improved.

[0164] Specifically, the auxiliary processor can be arranged on a board card where the BMC is located and connected with peripheral controllers such as an I2C bus controller and a PMBus controller through an AXI bus. The memory ECC error rate, working voltage fluctuation and particle temperature are collected through the I2C bus controller, and the fan vibration frequency is collected through the FANTACHO controller, so as to realize real-time collection of sensing data.

[0165] As an optional embodiment, collecting the sensing data through each peripheral controller on the advanced extensible interface bus includes:

[0166] Collecting the memory error checking and correction error rate, working voltage fluctuation and particle temperature through an integrated circuit bus controller and / or an enhanced integrated circuit bus controller;

[0167] Collecting the fan vibration frequency through a fan speed control interface controller;

[0168] Collecting the power supply voltage ripple amplitude through a power management bus controller;

[0169] Collecting the hard disk cyclic redundancy check error rate and hard disk temperature through a hard disk interface controller.

[0170] Specifically, to the sensing data characteristics of different components, the corresponding peripheral controllers are used for accurate collection in the embodiment of the application, so as to ensure the accuracy and real-time performance of the data and provide reliable data support for fault prediction.

[0171] Among them, the memory ECC error rate (times / million), working voltage fluctuation amplitude (mV) and particle temperature (℃) are collected through the I2C bus controller and the enhanced I2C (I3C, Inter-Integrated Circuit 3) bus controller; the fan vibration frequency (Hz) is collected through the FANTACHO controller, and the sampling period is 10 seconds; the power supply voltage ripple amplitude (mV) is collected through the PMBus controller; and the hard disk CRC error rate (times / million) and hard disk temperature (℃) are collected through the hard disk interface controller.

[0172] As an optional embodiment, the component fault prediction method further includes:

[0173] The failure probability prediction result of the target component failure is pushed to the baseboard management controller, so that the baseboard management controller alarms according to the failure probability prediction result and / or pushes the failure probability prediction result to a remote control terminal.

[0174] Specifically, to realize timely feedback and remote management of the failure prediction result, the failure probability prediction result is pushed to the BMC in the embodiment of the application, so as to alarm and report remotely, and the timeliness and convenience of the server hardware reliability management are improved.

[0175] Specifically, for example, after the NPU calculates and generates the memory failure probability prediction result 0.9, the result is pushed to the BMC firmware. The monitoring program in the BMC firmware determines to issue a failure warning according to a preset threshold (such as 0.7), and reports to the remote control terminal through a Web (webpage), Redfish (Redfish interface), Intelligent Platform Management Interface (IPMI, Intelligent Platform Management Interface), Simple Network Management Protocol (SNMP, Simple Network Management Protocol) and the like.

[0176] Specifically, in order to better illustrate the embodiment of the application, please refer to Figure 3 and Figure 4 , Figure 3 a structure diagram of a component failure prediction system provided by the application, Figure 4 a flowchart of another component failure prediction method provided by the application, Figure 3 In the embodiment, the remote control terminal is connected with the baseboard management controller through a network interface, and in the baseboard management controller, a central processing unit, an auxiliary processing unit and a peripheral controller are connected in sequence, and the peripheral controller is connected with a memory, a hard disk, a case, a fan and a power supply, thereby constituting a complete system architecture. Figure 4 In the embodiment, component operation data is acquired through data acquisition first, then data preprocessing is performed to regularize the data, and then feature generation is performed to divide the numerical interval and determine the numerical level, then correlation analysis is performed to determine the correlation degree, and then model reasoning outputs the result, and the model reasoning result can be used for weight matrix updating and reverse assistance correlation analysis.

[0177] Please refer to Figure 5 , Figure 5 a structure diagram of a component failure prediction device provided by the application, which comprises:

[0178] a memory 51 for storing a computer program;

[0179] a processor 52 for executing the computer program to realize the steps of the component failure prediction method in the foregoing embodiments.

[0180] As an optional embodiment, the processor is an auxiliary processor arranged on a board card where the baseboard management controller is located.

[0181] The auxiliary processor collects the sensing data through each peripheral controller on the advanced extensible interface bus.

[0182] For the component failure prediction device provided by the embodiment of the present application, please refer to the foregoing embodiment of the component failure prediction method, and the embodiment of the present application will not be described here again.

[0183] The present application also provides a computer program product, comprising computer programs / instructions, which, when executed by a processor, implement the steps of the component failure prediction method in the foregoing embodiments.

[0184] For the computer program product provided by the embodiment of the present application, please refer to the foregoing embodiment of the component failure prediction method, and the embodiment of the present application will not be described here again.

[0185] Please refer to Figure 6 , Figure 6 A structural schematic diagram of a computer readable storage medium provided by the present application is shown, and the computer readable storage medium 61 stores a computer program 62, which, when executed by a processor, implements the steps of the component failure prediction method in the foregoing embodiments.

[0186] For the computer readable storage medium provided by the embodiment of the present application, please refer to the foregoing embodiment of the component failure prediction method, and the embodiment of the present application will not be described here again.

[0187] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other. It should be further explained that, in the specification, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or sequence between the entities or operations. Moreover, 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 the process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another same element in the process, method, article or device including the element.

[0188] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A component failure prediction method characterized by, The method comprises: in response to a prediction instruction for a target component fault at a target time point in the future, determining a predicted value of any associated sensing feature of the target component fault at the target time point according to historical data, wherein the target component fault has a plurality of associated sensing features; for any predicted value, taking a numerical span of the associated sensing feature to which the predicted value belongs at a first preset number of past sampling points as a reference value range of the associated sensing feature to which the predicted value belongs; when the binning type is equal-frequency binning, counting a total number of data points in the reference value range; dividing the total number by a second preset number to obtain a number of data points expected to be contained in each numerical interval; sorting the data points in the reference value range according to numerical size, dividing the data points in sequence from the minimum value, so that the actual number of data points contained in each numerical interval is equal to the number of data points expected to be contained, thereby determining the boundaries of each numerical interval; according to the boundaries of each numerical interval, dividing the reference value range corresponding to the predicted value into a continuous second preset number of numerical intervals, wherein each numerical interval is sequentially sorted in ascending order of numerical value; taking the serial number of the numerical interval closest to the predicted value as the numerical level of the predicted value; weighting the weighted results of each predicted value, and taking the weighted result as the correlation degree; according to the correlation degree and the predicted values of each associated sensing feature, indexing in a preset decision tree model of the target component fault to determine a fault probability prediction result of the target component fault; the construction of the decision tree model satisfies the following conditions: the root node is the correlation degree feature, and the leaf node is the fault probability; when the correlation degree is greater than a preset threshold, a branch path is selected according to the predicted value of at least one associated sensing feature; the fault probability value of the leaf node is generated by a hierarchical weighted voting mechanism.

2. The component failure prediction method according to claim 1, characterized by, in response to a prediction instruction for a target component fault at a target time point in the future, determining a predicted value of any associated sensing feature of the target component fault at the target time point according to historical data comprises: in response to a prediction instruction for a target component fault at a target time point in the future, determining a change rate of any associated sensing feature of the target component fault according to historical data; determining the predicted value of the associated sensing feature at the target time point according to the change rate of the associated sensing feature and a time difference between the target time point and the current time.

3. The component failure prediction method according to claim 2, characterized by, in response to a prediction instruction for a target component fault at a target time point in the future, determining a change rate of any associated sensing feature of the target component fault according to historical data comprises: in response to a prediction instruction for a target component fault at a target time point in the future, determining a change rate of any associated sensing feature of the target component fault according to sampling values of the associated sensing feature at a first preset number of past sampling points.

4. The component failure prediction method according to claim 1, characterized by, In response to the prediction instruction of the target component failure at the future target time, determining a predicted value of each associated sensing feature of the target component failure at the target time according to historical data comprises: In response to the prediction instruction of the target component failure at the future target time, determining each associated sensing feature corresponding to the target component failure from a preset associated analysis matrix, wherein the preset associated analysis matrix comprises a corresponding relationship between component failures and associated sensing features; For each associated sensing feature of the target component failure, determining a predicted value of the associated sensing feature at the target time according to historical data.

5. The component failure prediction method according to claim 4, characterized by, The corresponding relationship between the component failure and the associated sensing feature comprises: The associated sensing features of the memory failure comprise error checking and error correction rate of the memory, memory voltage fluctuation amplitude, and memory particle temperature; The associated sensing features of the hard disk failure comprise cyclic redundancy check error rate of the hard disk, power supply ripple amplitude, fan vibration frequency, and hard disk temperature.

6. The component failure prediction method according to claim 1, characterized by, According to the weighted results of the numerical levels of each predicted value, determining the failure probability prediction result of the target component failure comprises: Determining a weight matrix of the target component failure from the preset associated analysis matrix, wherein the preset associated analysis matrix comprises a corresponding relationship between component failures and the weight matrix, and the weight matrix comprises weights of each associated sensing feature of the target component failure; Taking the weighted results of the numerical levels of each predicted value according to the weight matrix of the target component failure as the failure probability prediction result of the target component failure.

7. The component failure prediction method according to claim 6, characterized by, The component failure prediction method further comprises: Adjusting the weight matrix of the target component failure according to a third preset number of past verification data groups of the target component failure in combination with the gradient descent method; Wherein, the verification data group is a failure probability prediction result and its corresponding actual failure occurrence result, and the actual failure occurrence result comprises occurrence or non-occurrence.

8. The component failure prediction method according to claim 7, characterized by, Adjusting the weight matrix of the target component failure according to a third preset number of past verification data groups of the target component failure in combination with the gradient descent method comprises: For each data in the third preset number of verification data groups, the failure probability prediction result is substituted into a preset loss function to calculate the loss value corresponding to the data, wherein the loss function is used to measure the difference between the failure probability prediction result and the actual failure occurrence result; Taking the sum of the loss values corresponding to all verification data groups as the total loss, calculating the gradient of each weight in the weight matrix of the target component failure based on the total loss; According to the preset learning rate, updating each weight in the weight matrix of the target component failure in the opposite direction of the gradient to complete the adjustment of the weight matrix of the target component failure.

9. The component failure prediction method according to any one of claims 1 to 8, characterized by, Applied to an auxiliary processor; The auxiliary processor is arranged on a board card where a baseboard management controller is located, and the auxiliary processor is connected with a high extensible interface bus on the board card where the baseboard management controller is located. The component failure prediction method further comprises: Collecting sensing data through each peripheral controller on the high extensible interface bus.

10. The component failure prediction method according to claim 9, characterized by, Collecting sensing data through each peripheral controller on the high extensible interface bus comprises: The error check and correction rate, working voltage fluctuation and particle temperature of the memory are collected through the integrated circuit bus controller and / or enhanced integrated circuit bus controller; The fan vibration frequency is collected through the fan speed control interface controller; The voltage ripple amplitude of the power supply is collected through the power management bus controller; The cyclic redundancy check error rate and hard disk temperature of the hard disk are collected through the hard disk interface controller.

11. The component failure prediction method of claim 9, wherein, The component failure prediction method further comprises: The failure probability prediction result of the target component failure is pushed to the baseboard management controller, so that the baseboard management controller alarms according to the failure probability prediction result and / or pushes the failure probability prediction result to a remote control terminal.

12. A component failure prediction device characterized by comprising: It comprises: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the component failure prediction method according to any one of claims 1 to 11.

13. The component failure prediction apparatus according to claim 12, characterized by, The processor is an auxiliary processor arranged on a board card where the baseboard management controller is located; The auxiliary processor collects sensing data through each peripheral controller on the advanced extensible interface bus.

14. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to implement the steps of the component failure prediction method according to any one of claims 1 to 11.

15. A computer-readable storage medium, characterized in that, The computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to implement the steps of the component failure prediction method according to any one of claims 1 to 11.

Images