Network device fault prediction method and system based on multi-source data fusion

Abstract

The application discloses a network equipment fault prediction method and system based on multi-source data fusion, relates to the technical field of network equipment fault prediction, and comprises a disk vibration prediction module, a vibration transmission analysis module and a fault disk positioning module. The disk vibration prediction module analyzes historical vibration signal data generated when each disk unit in a disk array reads and writes data, trains a vibration signal prediction neural network model, and predicts vibration signals generated when each disk unit reads and writes data according to real-time monitored data access requests. The vibration transmission analysis module constructs a three-dimensional material model for the disk array, analyzes amplitude and phase changes of each frequency component of the vibration signal from the positions of each disk unit to each vibration monitoring sensor, and the fault disk positioning module compares and analyzes predicted vibration signals and actually measured vibration signal data of the vibration monitoring sensor, constructs a fault positioning model, and positions a fault disk unit.

Description

Technical Field

[0001] This invention relates to the field of network equipment fault prediction technology, specifically to a network equipment fault prediction method and system based on multi-source data fusion. Background Technology

[0002] Disk units are prone to performance degradation or even failure during long-term operation due to mechanical failures, overload, aging, and other problems.

[0003] However, traditional disk array fault detection methods have the following technical shortcomings: On the one hand, traditional disk array fault prediction is mainly based on SMART (Self-Monitoring, Analysis and Reporting Technology). During fault analysis, the data layer and physical structure layer parameters of disk operation are relatively independent, lacking correlation analysis between read / write data and disk vibration, and cannot accurately determine whether the vibration abnormality originates from disk load characteristics or potential faults. On the other hand, using multiple sensors to independently monitor and analyze the faults of all disk arrays in the disk array, and simultaneously updating the fault analysis components when expanding the disks, severely limits the scalability of the disk array.

[0004] Therefore, there is a need for network device fault prediction methods and systems based on multi-source data fusion to address the aforementioned technical deficiencies. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting network device faults based on multi-source data fusion, so as to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A network device fault prediction method based on multi-source data fusion includes the following steps:

[0008] Step S100: Obtain historical data of vibration signals generated when reading and writing data in each disk unit of the disk array, use Fourier transform to obtain the spectral data of the vibration signals, and label the vibration signal feature parameters to construct a set of vibration signal feature parameters and train a vibration signal prediction neural network model; the vibration signal feature parameters include the read and write data feature parameters of the disk unit and the vibration signal feature parameters.

[0009] Step S200: Obtain physical structure data of disk array, construct three-dimensional material model, analyze the phase offset and amplitude attenuation coefficient of each frequency component of vibration signal from the location of each disk unit to each vibration monitoring sensor, and fit and obtain frequency-amplitude attenuation coefficient curve and frequency-phase offset curve.

[0010] Step S300: Monitor data access requests in the server transmission link, extract read and write data feature parameters, predict the read and write data feature parameters of each disk unit, and use the vibration signal prediction neural network model to predict the feature parameters of the vibration signal generated by the read and write data of each disk unit.

[0011] Step S400: Set the frequency step size, divide the frequency level range, divide the frequency components of the vibration signal into frequency levels, and calculate the amplitude attenuation coefficient and phase offset of each frequency component of the disk unit vibration signal based on the frequency-amplitude attenuation coefficient curve and frequency-phase offset curve transmitted from each disk unit vibration signal to each vibration monitoring sensor.

[0012] Step S500: Acquire vibration signal monitoring data from each vibration monitoring sensor in real time, compare and analyze the data with the predicted vibration signal characteristic parameters at each vibration monitoring sensor, construct a fault location model, and locate the fault in each disk unit.

[0013] In the above technical solution, step S100 is divided into the following steps:

[0014] Step S101: Obtain historical vibration signal data generated when each disk unit in the disk array reads and writes data; the historical vibration signal data is the vibration signal data generated when each disk unit independently reads and writes data.

[0015] Step S102: Use Fourier transform to obtain the spectral data of the vibration signal and perform vibration signal feature annotation;

[0016] For any historical vibration signal data s of any disk unit x in the disk array, the feature annotation result is x_s[D_x_s,S_x_s]; where D_x_s is the set of read and write data feature parameters of the historical vibration signal data s of disk unit x, and S_x_s is the vibration signal feature parameter of the historical vibration signal data s of disk unit x.

[0017] Step S103: Construct a vibration signal feature dataset and train a vibration signal prediction neural network model;

[0018] By combining the read / write operation characteristics of disks with the spectral characteristics of vibration signals, a multi-dimensional input neural network model is constructed. By monitoring data access requests in real time, the vibration characteristics under different loads are dynamically predicted, overcoming the limitations of isolated analysis of multi-source data in traditional methods.

[0019] In the above technical solution, step S200 is divided into the following steps:

[0020] Step S201: Set up vibration monitoring sensors to monitor the vibration signal of the disk array in real time;

[0021] Step S202: Obtain physical structure data of the disk array and construct a three-dimensional material model for the physical structure of the disk array;

[0022] Step S203: Monitor the changes in amplitude and phase of vibration signals of different frequencies transmitted from each disk unit to each vibration monitoring sensor, and then calculate the phase offset and amplitude attenuation coefficient of vibration signal transmission of each disk unit.

[0023] Among them, the phase difference between the vibration signals at the disk unit and the vibration monitoring sensor is used as the phase offset of the vibration signal transmission, and the ratio of the vibration signal amplitude at the disk unit and the vibration monitoring sensor is used as the amplitude attenuation coefficient of the vibration signal transmission.

[0024] Step S204: Use curve fitting to obtain the frequency-amplitude attenuation coefficient curve and frequency-phase offset curve of the vibration signal of each disk unit at different frequencies transmitted to each vibration monitoring sensor;

[0025] For any disk unit x, the frequency-amplitude attenuation coefficient curve of the vibration signal transmitted to the vibration monitoring sensor s is denoted as L_a(x,s), and the frequency-phase offset curve is denoted as...

[0026] A three-dimensional material model of the disk array is constructed, and the phase shift and amplitude attenuation coefficient of the vibration signal are analyzed to accurately calculate the propagation characteristics of the vibration signal and improve the prediction accuracy.

[0027] In the above technical solution, step S400 is divided into the following steps:

[0028] Step S401: Set the frequency step size f_step for the vibration signal spectrum, divide all frequency components in the vibration signal belonging to the frequency level interval [(n-1)×f_step, n×f_step] into frequency level n, and add the amplitudes of all frequency components in the same frequency level and divide by the frequency step size f_step to obtain the amplitude of the corresponding frequency level; where n is the frequency level number and the value is a positive integer;

[0029] Step S402: For any disk unit x, the amplitude attenuation coefficient k(x,s,f) of any frequency f component of the vibration signal generated by reading and writing data transmitted to the vibration monitoring sensor s is calculated as follows:

[0030] Extract the frequency level interval to which frequency f belongs, calculate the integral average value of L_a(x,s) with respect to frequency within the interval, and divide the calculation result by the frequency step size to obtain k(x,s,f);

[0031] For any disk unit x, the phase shift of any frequency f component of the vibration signal generated by reading and writing data transmitted to the vibration monitoring sensor s is... The calculation method is as follows:

[0032] Extract the frequency class range to which frequency f belongs, and calculate the range within the frequency class range. The integral average of the frequencies is calculated by dividing the result by the frequency step size.

[0033] The vibration signal is divided into multiple frequency levels by using a frequency step size division method. The amplitude attenuation and phase shift of different frequency components are calculated by integral averaging. This reduces the amount of calculation and improves the real-time performance while improving the accuracy of the equalization algorithm analysis.

[0034] In the above technical solution, step S500 is divided into the following steps:

[0035] Step S501: Using the characteristic parameters of the vibration signal generated by the predicted read and write data of each disk unit, and based on the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal generated by the read and write data of the disk unit to each vibration monitoring sensor, simulate the spectrum data of the vibration signal of each disk unit transmitted to the vibration monitoring sensor.

[0036] Step S502: Perform a complex summation of the vibration signals transmitted from each disk unit to the vibration monitoring sensor in the frequency domain to obtain the amplitude and phase of each frequency component of the predicted vibration signal from each vibration monitoring sensor;

[0037] Step S503: The vibration signal of the disk array is monitored in real time by a vibration monitoring sensor set on the outer contour of the disk array, and the amplitude and phase of each frequency component of the measured vibration signal are obtained by Fourier transform.

[0038] Step S504: Calculate the deviation between the measured vibration signal and the predicted vibration signal of the disk array vibration signal and perform a weighted summation, and set a vibration deviation threshold; when the weighted summation result is greater than or equal to the vibration deviation threshold, it is determined that there is a faulty disk unit in the disk array, and a fault location model is further constructed to locate the faulty disk unit in the disk array.

[0039] A fault location model is constructed, and a sparse optimization method is adopted. By iteratively solving the spectral deviation of disk units, the fault probability of each disk unit is accurately calculated, and a fault priority ranking is provided to help managers conduct accurate troubleshooting.

[0040] In the above technical solution, the step S504 of constructing a fault location model to locate the faulty disk unit in the disk array includes the following:

[0041] Construct a fault location model M, as follows:

[0042]

[0043] M = min_ΔS x (f){∑ s,f ||ΔS s (f)-∑ x [k x,f ×ΔS x (f)]|| 2 +λ×∑ x,f (ΔS x (f))};

[0044] Where, ΔS x (f) represents the spectral deviation of the frequency f component of the vibration signal of disk unit x, ΔS s (f) represents the spectral deviation of the frequency f component of the vibration signal from the vibration monitoring sensor s, k x,f Let be the transfer function of a vibration signal of frequency f transmitted from disk unit x to vibration monitoring sensor s, where λ is the regularization parameter and j is the complex unit; ||p|| 2 Let p be the square of the L2 norm of parameter p;

[0045] The spectral deviation of the frequency f component of the vibration signal of each disk unit is initialized to 0. The spectral deviation ΔS of the frequency f component of the vibration signal of any disk unit x is selected. x (f) is used as a variable, and the spectral deviation of the frequency f component of the vibration signal of the remaining disk units is fixed. The solution is then used to minimize the value of the fault location model M. x (f);

[0046] Set a threshold for the number of disks to be identified as faulty, using the aforementioned ΔS. x (f) The solution method iterates the spectral deviation of the vibration signal frequency f component of each disk unit multiple times until the number of disk units in all disk units whose spectral deviation of the vibration signal frequency f component is greater than the vibration abnormality threshold S' is less than or equal to the threshold for the number of faulty disks to be determined, and marks the disk units whose spectral deviation of the vibration signal frequency f component is greater than the vibration abnormality threshold S' as faulty disk units.

[0047] Further, based on the iteration results, the vibration signal deviation of each disk unit is calculated. For any disk unit x, the vibration signal deviation Res(x) is calculated using the following formula:

[0048] Res(x)=∑ s,f ||ΔS s (f)-∑ x [k x,f ×ΔS x (f)]||2;

[0049] For any disk units a and b, if Res(a) > Res(b), it is determined that the failure probability of disk unit a is less than that of disk unit b; if Res(a) < Res(b), it is determined that the failure probability of disk unit a is greater than that of disk unit b.

[0050] Further, each disk unit is sequentially fed back to the administrator for disk unit fault troubleshooting in ascending order of the vibration signal deviation.

[0051] A network device fault prediction system based on multi-source data fusion in the above technical solution, the system includes: a disk vibration prediction module, a vibration transmission analysis module, and a faulty disk location module;

[0052] The disk vibration prediction module analyzes the historical data of the vibration signals generated when each disk unit in the disk array reads and writes data, trains a vibration signal prediction neural network model, and predicts the vibration signals generated when each disk unit reads and writes data according to the real-time monitored data access requests; the vibration transmission analysis module analyzes the amplitude and phase changes of each frequency component of the vibration signal transmitted from the azimuth of each disk unit to each vibration monitoring sensor by constructing a three-dimensional material model of the disk array; the faulty disk location module constructs a fault location model by comparing and analyzing the predicted vibration signals and the measured vibration signal data of the vibration monitoring sensors, and locates faults for each disk unit.

[0053] In the above technical solution, the disk vibration prediction module includes: a vibration data processing unit, a vibration feature analysis unit, and a vibration signal prediction unit;

[0054] The vibration data processing unit is used to process the historical data of the vibration signals generated when each disk unit in the disk array reads and writes data; the vibration feature analysis unit uses Fourier transform to obtain the spectrum data of the vibration signals and performs vibration signal feature parameter annotation; the vibration signal prediction unit trains a vibration signal prediction neural network model by constructing a vibration signal feature parameter set.

[0055] In the above technical solution, the vibration transmission analysis module includes: a material model construction unit and a signal transmission analysis unit;

[0056] The material model construction unit obtains the physical structure data of the disk array and constructs a three-dimensional material model; the signal transmission analysis unit analyzes the amplitude and phase changes of each frequency component of the vibration signal transmitted from the azimuth of each disk unit to each vibration monitoring sensor, and fits to obtain a frequency-amplitude attenuation coefficient curve and a frequency-phase offset curve.

[0057] In the above technical solution, the fault disk location module includes: a vibration deviation analysis unit, a location model construction unit, and a fault disk screening unit;

[0058] The vibration deviation analysis unit compares and analyzes the vibration signal predicted by the vibration monitoring sensor with the measured vibration signal data; the positioning model construction unit constructs a fault positioning model based on the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal; the fault disk screening unit iterates the spectral deviation of the frequency f component of the vibration signal of each disk unit multiple times, and filters the fault disk units according to the threshold of the number of fault disks to be determined, and performs disk unit fault investigation feedback in order of increasing vibration signal deviation of each fault disk unit.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] In this invention, by integrating disk read / write data features and vibration signal features, a neural network model is trained to fully consider the correlation between disk data read / write operations and the generated vibration signals, thereby enhancing fault prediction capabilities and improving the accuracy of fault analysis.

[0061] In this invention, fault source is traced through external vibration signals of the disk array, a fault location model is constructed using sparse optimization, an L1 regularization term is introduced to force sparse solutions, and a non-negative amplitude constraint is added to ensure the rationality of fault location. At the same time, it effectively solves the problem of disk array expansion obstacles caused by monitoring each disk unit separately.

[0062] Furthermore, this invention combines disk read / write data characteristics and disk array material characteristics to comprehensively predict and analyze the effect of vibration data generated by each disk unit in the disk array during read / write operations after being collected by sensors. This allows for comparative analysis to locate faulty disk units, overcoming the isolation inherent in multi-source data analysis in traditional methods and significantly improving the accuracy and reliability of disk fault analysis and location results. Attached Figure Description

[0063] Figure 1 This is a flowchart of the network device fault prediction method based on multi-source data fusion according to the present invention;

[0064] Figure 2 This is an organizational structure diagram of the network device fault prediction system based on multi-source data fusion according to the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example: Please refer to Figures 1-2 The present invention provides the following technical solution:

[0067] like Figure 1 As shown, this invention provides a network device fault prediction method based on multi-source data fusion, the method comprising the following steps:

[0068] Step S100: Obtain historical data of vibration signals generated when reading and writing data in each disk unit of the disk array, use Fourier transform to obtain the spectral data of the vibration signals, and label the vibration signal feature parameters to construct a set of vibration signal feature parameters and train a vibration signal prediction neural network model; the vibration signal feature parameters include the read and write data feature parameters of the disk unit and the vibration signal feature parameters.

[0069] Step S200: Obtain physical structure data of disk array, construct three-dimensional material model, analyze the phase offset and amplitude attenuation coefficient of each frequency component of vibration signal from the location of each disk unit to each vibration monitoring sensor, and fit and obtain frequency-amplitude attenuation coefficient curve and frequency-phase offset curve.

[0070] Step S300: Monitor data access requests in the server transmission link, extract read and write data feature parameters, predict the read and write data feature parameters of each disk unit, and use the vibration signal prediction neural network model to predict the feature parameters of the vibration signal generated by the read and write data of each disk unit.

[0071] Step S400: Set the frequency step size, divide the frequency level range, divide the frequency components of the vibration signal into frequency levels, and calculate the amplitude attenuation coefficient and phase offset of each frequency component of the disk unit vibration signal based on the frequency-amplitude attenuation coefficient curve and frequency-phase offset curve transmitted from each disk unit vibration signal to each vibration monitoring sensor.

[0072] Step S500: Acquire vibration signal monitoring data from each vibration monitoring sensor in real time, compare and analyze the data with the predicted vibration signal characteristic parameters at each vibration monitoring sensor, construct a fault location model, and locate the fault in each disk unit.

[0073] Step S100 consists of the following steps:

[0074] Step S101: Obtain historical vibration signal data generated when each disk unit in the disk array reads and writes data; the historical vibration signal data is the vibration signal data generated when each disk unit independently reads and writes data.

[0075] Step S102: Use Fourier transform to obtain the spectral data of the vibration signal and perform vibration signal feature annotation;

[0076] For any historical vibration signal data s of any disk unit x in the disk array, the feature annotation result is x_s[D_x_s,S_x_s]; where D_x_s is the set of read and write data feature parameters of the historical vibration signal data s of disk unit x, and S_x_s is the vibration signal feature parameter of the historical vibration signal data s of disk unit x.

[0077] Step S103: Construct a vibration signal feature dataset and train a vibration signal prediction neural network model;

[0078] In practice, as parameters such as the data access mode of disk read / write, the data read range of a single IO operation, and the concurrency of data access requests change, the vibration frequency and amplitude of the read / write head inside the mechanical disk also vary significantly. In a disk array, each disk stores different data, and the vibration conditions generated by each disk during read / write operations are also different. Therefore, when analyzing disk failures, it is necessary to conduct a comprehensive analysis based on the data read / write operations being performed on the disk. For disk units with high concurrency and high operation rate, the basis for vibration signal analysis should be significantly different from that of disks with low data operation rate, so that the results of the failure analysis are reliable and reasonable.

[0079] Furthermore, in specific implementation, an index is constructed for the data in each disk unit. By analyzing the index, it is determined which data each disk unit needs to read and write in each data access request of the disk array, thereby determining the data read and write behavior characteristics of each disk unit. The vibration signal prediction neural network model is then used to predict the vibration data generated by data read and write in each disk unit.

[0080] Step S200 consists of the following steps:

[0081] Step S201: Set up vibration monitoring sensors to monitor the vibration signal of the disk array in real time;

[0082] Step S202: Obtain physical structure data of the disk array and construct a three-dimensional material model for the physical structure of the disk array;

[0083] Step S203: Monitor the changes in amplitude and phase of vibration signals of different frequencies transmitted from each disk unit to each vibration monitoring sensor, and then calculate the phase offset and amplitude attenuation coefficient of vibration signal transmission of each disk unit.

[0084] Among them, the phase difference between the vibration signals at the disk unit and the vibration monitoring sensor is used as the phase offset of the vibration signal transmission, and the ratio of the vibration signal amplitude at the disk unit and the vibration monitoring sensor is used as the amplitude attenuation coefficient of the vibration signal transmission.

[0085] Step S204: Use curve fitting to obtain the frequency-amplitude attenuation coefficient curve and frequency-phase offset curve of the vibration signal of each disk unit at different frequencies transmitted to each vibration monitoring sensor;

[0086] For any disk unit x, the frequency-amplitude attenuation coefficient curve of the vibration signal transmitted to the vibration monitoring sensor s is denoted as L_a(x,s), and the frequency-phase offset curve is denoted as...

[0087] In practice, a 3D scanner is used to acquire physical structure data of the disk array, which is then imported into COMSOL Multiphysics to construct a three-dimensional material model. Material properties are defined, and the orientation of each disk unit in its respective three-dimensional material model is determined. Furthermore, a 10Hz to 10kHz sinusoidal sweep frequency signal is applied to each disk unit using an exciter, and the sensor response is recorded. Based on the input and response of the sweep frequency signal, the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal of each disk unit are calculated, and the calculation results of different frequency components are curve fitted.

[0088] Step S400 consists of the following steps:

[0089] Step S401: Set the frequency step size f_step for the vibration signal spectrum, divide all frequency components in the vibration signal belonging to the frequency level interval [(n-1)×f_step, n×f_step] into frequency level n, and add the amplitudes of all frequency components in the same frequency level and divide by the frequency step size f_step to obtain the amplitude of the corresponding frequency level; where n is the frequency level number and the value is a positive integer;

[0090] Step S402: For any disk unit x, the amplitude attenuation coefficient k(x,s,f) of any frequency f component of the vibration signal generated by reading and writing data transmitted to the vibration monitoring sensor s is calculated as follows:

[0091] Extract the frequency level interval to which frequency f belongs, calculate the integral average value of L_a(x,s) with respect to frequency within the interval, and divide the calculation result by the frequency step size to obtain k(x,s,f);

[0092] For any disk unit x, the phase shift of any frequency f component of the vibration signal generated by reading and writing data transmitted to the vibration monitoring sensor s is... The calculation method is as follows:

[0093] Extract the frequency class range to which frequency f belongs, and calculate the range within the frequency class range. The integral average of the frequencies is calculated by dividing the result by the frequency step size.

[0094] In practical implementation, the frequency step size f_step = 100 Hz is set, and the frequency level ranges are divided as follows: Level 1 (0-100 Hz), Level 2 (100-200 Hz) ... Level 100 (9.9-10 kHz). Considering that the high-frequency components of the vibration signal have a strong interference on the working condition and lifespan of the disk, the frequency step size is reduced to 50 Hz when the frequency range is higher than 5 kHz, and the frequency level ranges are corrected to Level 1 (0-100 Hz), Level 2 (100-200 Hz) ... Level 50 (4.9 kHz-5 kHz), Level 51 (5 kHz-5.05 kHz) ... Level 150 (9.95-10 kHz).

[0095] Furthermore, based on the fitting curves of amplitude attenuation coefficient and phase shift, the amplitude attenuation coefficient k(x,s,f) and phase shift for each frequency range are calculated. For any frequency level n, the following formula is used for calculation:

[0096] k(x,s,f)=[1 / (f_step)]×∫ n [L_a(x,s)]df;

[0097]

[0098] Among them, ∫ n [L_a(x,s)] is the integral of the frequency-amplitude attenuation coefficient curve over the frequency level n interval. The frequency-phase offset curve is integrated over the frequency level range of frequency level n.

[0099] In practical implementation, considering that the amplitude attenuation coefficient and phase offset may change nonlinearly with frequency, using the value at the center point of the interval or the average value at both boundaries of the interval as the final calculation result cannot reflect the overall relationship between the amplitude attenuation coefficient and phase offset and frequency in each frequency level interval. Therefore, the method of integral averaging is adopted to ensure that the overall amplitude attenuation coefficient and phase offset in each frequency level interval are accurately reflected.

[0100] Step S500 consists of the following steps:

[0101] Step S501: Using the characteristic parameters of the vibration signal generated by the predicted read and write data of each disk unit, and based on the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal generated by the read and write data of the disk unit to each vibration monitoring sensor, simulate the spectrum data of the vibration signal of each disk unit transmitted to the vibration monitoring sensor.

[0102] Step S502: Perform a complex summation of the vibration signals transmitted from each disk unit to the vibration monitoring sensor in the frequency domain to obtain the amplitude and phase of each frequency component of the predicted vibration signal from each vibration monitoring sensor;

[0103] Step S503: The vibration signal of the disk array is monitored in real time by a vibration monitoring sensor set on the outer contour of the disk array, and the amplitude and phase of each frequency component of the measured vibration signal are obtained by Fourier transform.

[0104] Step S504: Calculate the deviation between the measured vibration signal and the predicted vibration signal of the disk array vibration signal and perform a weighted summation, and set a vibration deviation threshold; when the weighted summation result is greater than or equal to the vibration deviation threshold, it is determined that there is a faulty disk unit in the disk array, and a fault location model is further constructed to locate the faulty disk unit in the disk array.

[0105] In step S504, the fault location model is constructed, and the faulty disk unit in the disk array is located by the following:

[0106] Construct a fault location model M, as follows:

[0107]

[0108] M = min_ΔS x (f){∑ s,f ||ΔS s (f)-∑ x [k x,f ×ΔS x (f)]|| 2 +λ×∑ x,f (ΔS x (f))};

[0109] Where, ΔS x (f) represents the spectral deviation of the frequency f component of the vibration signal of disk unit x, ΔS s (f) represents the spectral deviation of the frequency f component of the vibration signal from the vibration monitoring sensor s, k x,f Let be the transfer function of a vibration signal of frequency f transmitted from disk unit x to vibration monitoring sensor s, where λ is the regularization parameter and j is the complex unit; ||p|| 2 Let p be the square of the L2 norm of parameter p;

[0110] In the fault location model M formula, ∑ s,f ||ΔS s (f)-∑ x [k x,f ×ΔS x (f)]|| 2 The residual term is used to ensure that the model's predictions are as close as possible to the actual observed data. By minimizing the prediction error, it guarantees the vibration anomaly quantity ΔS output by the model iteration. x (f) It can reasonably explain the bias observed by the sensor, λ×∑ x,f (ΔS x (f) is a regularization term. It accurately locates disk units with a high probability of abnormal vibration signals by using sparse constraints. It also filters false abnormal signals caused by sensor noise or cross-interference by screening disk units, thus avoiding the location of fault sources caused by minor vibration deviations or noise interference in disk units.

[0111] The spectral deviation of the frequency f component of the vibration signal of each disk unit is initialized to 0. The spectral deviation ΔS of the frequency f component of the vibration signal of any disk unit x is selected. x (f) is used as a variable, and the spectral deviation of the frequency f component of the vibration signal of the remaining disk units is fixed. The solution is then used to minimize the value of the fault location model M. x (f);

[0112] In practice, each frequency component is iterated separately. The spectral deviation of this frequency component for each disk unit is preset to 0 and stored in the deviation matrix of the corresponding frequency component. The frequency f component deviation ΔS of the vibration signal of any disk unit x is selected. x (f) is used as a variable in the calculation, simplifying the fault location model iteratively to a function of ΔS. x (f) is a one-dimensional optimization problem;

[0113] For ΔS x (f) Take the derivative and set it to zero, then calculate ΔS. x The solution to (f) is used as ΔS x The iteration result of (f) is used to iterate over the variables ΔS corresponding to the remaining different values ​​of x using the above iteration method. x (f) After iterative updates, ΔS becomes the variable corresponding to all x values. x The initial value of (f);

[0114] Set a threshold for the number of disks to be identified as faulty, using the aforementioned ΔS. x(f) The solution method iterates multiple times on the spectral deviation of the frequency f component of the vibration signal of each disk unit until the number of disk units with the spectral deviation of the frequency f component of the vibration signal greater than the vibration anomaly threshold S' among all disk units is less than or equal to the threshold of the number of faulty disks to be determined, and marks all disk units with the spectral deviation of the frequency f component of the vibration signal greater than the vibration anomaly threshold S' as faulty disk units; when setting the vibration anomaly threshold S', a dynamic threshold can be set according to the analyzed frequency f to retain more high-frequency anomaly components, so as to ensure more sensitive detection and identification of the high-frequency abnormal vibration signals of each disk unit;

[0115] Further, in the subsequent iteration variable ΔS x (f), set the vibration anomaly threshold S'. When ΔS x (f) < S', correct the corresponding item value of ΔS x (f) in the deviation matrix to 0. When ΔS x (f) ≥ S', continue the iteration to achieve the selection and rejection of the abnormal amounts of different frequency components of each disk unit during the iteration process, and then correct most of the elements in the deviation matrix to 0 through multiple iterations;

[0116] In the iteration result of any frequency f component, the vibration signals of the disk units corresponding to the remaining elements all satisfy: If the abnormal vibration of the frequency f component of disk unit x is transmitted to vibration monitoring sensor s as a predicted vibration signal, and when the deviation between the predicted vibration signal of vibration monitoring sensor s and the measured vibration signal is the smallest, the abnormal vibration of the frequency f component of disk unit x is stronger and is recognized as abnormal vibration in actual operation;

[0117] Further calculate the vibration signal deviation of each disk unit according to the iteration result. For any disk unit x, the calculation formula of the vibration signal deviation Res(x) is as follows:

[0118] Res(x) = ∑ s,f ||ΔS s (f) - ∑ x [k x,f ×ΔS x (f)]||2;

[0119] For any disk units a and b, if Res(a) > Res(b), it is determined that the failure probability of disk unit a is less than that of disk unit b. If Res(a) < Res(b), it is determined that the failure probability of disk unit a is greater than that of disk unit b;

[0120] Further, arrange each disk unit in ascending order of the vibration signal deviation and feedback the disk unit fault troubleshooting to the management personnel in turn;

[0121] In practical implementation, when the frequency f component deviates by ΔS x (f) When the number of disk units greater than the vibration abnormality threshold is less than or equal to the threshold for the number of faulty disks to be determined, it is determined that the above iterative algorithm has converged. For any frequency component f, if the vibration signal deviation calculated by any disk unit x is small, in actual operation, it means that when the abnormal vibration occurs by this frequency component, it is more consistent with the actual vibration signal of the vibration monitoring sensor s. Therefore, it can be considered that the probability of abnormal vibration of the frequency component f of disk unit x is greater, and thus each disk unit involved is regarded as a faulty disk unit.

[0122] Furthermore, when assuming a single disk failure, the vibration signal deviation between the model-predicted signal spectrum deviation and the actual signal spectrum deviation is calculated. The smaller the vibration signal deviation, the more consistent the disk failure is with the analysis results, i.e., the higher the probability of the disk failure. Therefore, each disk unit is sorted in ascending order of vibration signal deviation calculation results as the order of disk unit failure probability from large to small.

[0123] like Figure 2 As shown, the present invention also provides a network device fault prediction system based on multi-source data fusion, the system comprising: a disk vibration prediction module, a vibration transmission analysis module, and a fault disk location module;

[0124] The disk vibration prediction module analyzes historical data of vibration signals generated when each disk unit in the disk array reads and writes data, trains a vibration signal prediction neural network model, and predicts the vibration signals generated by each disk unit's read and write data based on real-time monitored data access requests. The vibration transmission analysis module constructs a three-dimensional material model of the disk array and analyzes the amplitude and phase changes of each frequency component of the vibration signal transmitted from the location of each disk unit to each vibration monitoring sensor. The fault disk location module constructs a fault location model by comparing and analyzing the vibration signals predicted by the vibration monitoring sensors with the measured vibration signal data, and locates the fault in each disk unit.

[0125] The disk vibration prediction module includes: a vibration data processing unit, a vibration feature analysis unit, and a vibration signal prediction unit;

[0126] The vibration data processing unit is used to process historical vibration signal data generated when each disk unit in the disk array reads and writes data; the vibration feature analysis unit uses Fourier transform to obtain the spectral data of the vibration signal and annotates the vibration signal feature parameters; the vibration signal prediction unit trains a vibration signal prediction neural network model by constructing a set of vibration signal feature parameters.

[0127] The vibration transmission analysis module includes: a material model construction unit and a signal transmission analysis unit;

[0128] The material model construction unit acquires physical structure data of the disk array and constructs a three-dimensional material model; the signal transmission analysis unit analyzes the amplitude and phase changes of each frequency component of the vibration signal from the location of each disk unit to each vibration monitoring sensor, and fits to obtain the frequency-amplitude attenuation coefficient curve and the frequency-phase offset curve.

[0129] The fault disk location module includes: a vibration deviation analysis unit, a location model construction unit, and a fault disk screening unit;

[0130] The vibration deviation analysis unit compares and analyzes the vibration signal predicted by the vibration monitoring sensor with the measured vibration signal data; the positioning model construction unit constructs a fault positioning model based on the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal; the fault disk screening unit iterates the spectral deviation of the frequency f component of the vibration signal of each disk unit multiple times, and filters the fault disk units according to the threshold of the number of fault disks to be determined, and performs disk unit fault investigation feedback in order of increasing vibration signal deviation of each fault disk unit.

[0131] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A network device fault prediction method based on multi-source data fusion, characterized in that... The method includes the following steps: Step S100: Obtain historical data of vibration signals generated when reading and writing data in each disk unit of the disk array, use Fourier transform to obtain the spectral data of the vibration signals, and label the vibration signal feature parameters to construct a set of vibration signal feature parameters and train a vibration signal prediction neural network model; the vibration signal feature parameters include the read and write data feature parameters of the disk unit and the vibration signal feature parameters. Step S200: Obtain physical structure data of disk array, construct three-dimensional material model, analyze the phase offset and amplitude attenuation coefficient of each frequency component of vibration signal from the location of each disk unit to each vibration monitoring sensor, and fit and obtain frequency-amplitude attenuation coefficient curve and frequency-phase offset curve. Step S300: Monitor data access requests in the server transmission link, extract read and write data feature parameters, predict the read and write data feature parameters of each disk unit, and use the vibration signal prediction neural network model to predict the feature parameters of the vibration signal generated by the read and write data of each disk unit. Step S400: Set the frequency step size, divide the frequency level range, divide the frequency components of the vibration signal into frequency levels, and calculate the amplitude attenuation coefficient and phase offset of each frequency component of the disk unit vibration signal based on the frequency-amplitude attenuation coefficient curve and frequency-phase offset curve transmitted from each disk unit vibration signal to each vibration monitoring sensor. Step S500: Real-time acquisition of vibration signal monitoring data from each vibration monitoring sensor, comparison and analysis with the predicted vibration signal characteristic parameters at each vibration monitoring sensor, construction of a fault location model, and fault location of each disk unit; Step S501: Using the characteristic parameters of the vibration signal generated by the predicted read and write data of each disk unit, and based on the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal generated by the read and write data of the disk unit to each vibration monitoring sensor, simulate the spectrum data of the vibration signal of each disk unit transmitted to the vibration monitoring sensor. Step S502: Perform a complex summation of the vibration signals transmitted from each disk unit to the vibration monitoring sensor in the frequency domain to obtain the amplitude and phase of each frequency component of the predicted vibration signal from each vibration monitoring sensor; Step S503: The vibration signal of the disk array is monitored in real time by a vibration monitoring sensor set on the outer contour of the disk array, and the amplitude and phase of each frequency component of the measured vibration signal are obtained by Fourier transform. Step S504: Calculate the deviation between the measured vibration signal and the predicted vibration signal of the disk array vibration signal and perform a weighted summation, and set a vibration deviation threshold; when the weighted summation result is greater than or equal to the vibration deviation threshold, it is determined that there is a faulty disk unit in the disk array, and a fault location model is further constructed to locate the faulty disk unit in the disk array. Constructing a fault location model to locate faulty disk units in a disk array includes the following: Construct a fault location model M, as follows: k x,f =k(x,s,f)×[cos(δφ(x,s,f))+j×sin(δφ(x,s,f))]; M=min_ΔS x (f){∑ s,f ||ΔS s (f)−∑ x [k x,f ×ΔS x (f)]|| 2 +λ×∑ x,f (ΔS x (f))}; Where, ΔS x (f) represents the spectral deviation of the frequency f component of the vibration signal of disk unit x, ΔS s (f) represents the spectral deviation of the frequency f component of the vibration signal from the vibration monitoring sensor s, k x,f Let be the transfer function of a vibration signal of frequency f transmitted from disk unit x to vibration monitoring sensor s, where λ is the regularization parameter and j is the complex unit; ||p|| 2 δφ(x,s,f) represents the L2 norm squared of parameter p; k(x,s,f) represents the amplitude attenuation coefficient; δφ(x,s,f) represents the phase offset. The spectral deviation of the frequency f component of the vibration signal of each disk unit is initialized to 0. The spectral deviation ΔS of the frequency f component of the vibration signal of any disk unit x is selected. x (f) is used as a variable, and the spectral deviation of the frequency f component of the vibration signal of the remaining disk units is fixed. The solution is then used to minimize the value of the fault location model M. x (f); Set a threshold for the number of disks to be identified as faulty, using the aforementioned ΔS. x (f) The solution method iterates the spectral deviation of the vibration signal frequency f component of each disk unit multiple times until the number of disk units in all disk units whose spectral deviation of the vibration signal frequency f component is greater than the vibration abnormality threshold S' is less than or equal to the threshold for the number of faulty disks to be determined, and marks the disk units whose spectral deviation of the vibration signal frequency f component is greater than the vibration abnormality threshold S' as faulty disk units. Further, based on the iteration results, the vibration signal deviation of each disk unit is calculated. For any disk unit x, the vibration signal deviation Res(x) is calculated using the following formula: Res(x)=∑ x,f ||ΔS s (f)−∑ x [k x,f ×ΔS x (f)]||2; For any disk units a and b, if Res(a) > Res(b), it is determined that the failure probability of disk unit a is less than that of disk unit b; if Res(a) < Res(b), it is determined that the failure probability of disk unit a is greater than that of disk unit b. Further, each disk unit is sequentially fed back to the administrator for disk unit failure troubleshooting in ascending order of the vibration signal deviation.

2. The network device fault prediction method based on multi-source data fusion according to claim 1, characterized in that, The step S100 is divided into the following steps: Step S101: Obtain the historical vibration signal data generated when each disk unit in the disk array reads and writes data; the historical vibration signal data is the vibration signal data generated when each disk unit reads and writes data independently. Step S102: Use Fourier transform to obtain the spectrum data of the vibration signal and perform vibration signal feature annotation. For any item of historical vibration signal data s of any disk unit x in the disk array, the feature annotation result is x_s[D_x_s, S_x_s]; where D_x_s is the set of read / write data feature parameters of the historical vibration signal record s of disk unit x, and S_x_s is the vibration signal feature parameter of the historical vibration signal record s of disk unit x. Step S103: Construct a vibration signal feature data set and train a vibration signal prediction neural network model.

3. The network device fault prediction method based on multi-source data fusion according to claim 1, characterized in that, The step S200 is divided into the following steps: Step S201: Set vibration monitoring sensors to monitor the vibration signal of the disk array in real time. Step S202: Obtain the physical structure data of the disk array and construct a three-dimensional material model for the physical structure of the disk array. Step S203: Monitor the change data of the amplitude and phase of the vibration signals at different frequencies transmitted from each disk unit to each vibration monitoring sensor, and then calculate the phase offset and amplitude attenuation coefficient of the vibration signal transmission of each disk unit. Among them, the phase difference between the vibration signals at the disk unit and the vibration monitoring sensor is used as the phase offset of the vibration signal transmission, and the amplitude ratio of the vibration signals at the disk unit and the vibration monitoring sensor is used as the amplitude attenuation coefficient of the vibration signal transmission. Step S204: Use curve fitting to obtain the frequency-amplitude attenuation coefficient curve and frequency-phase offset curve of the vibration signals of each disk unit at different frequencies transmitted to each vibration monitoring sensor. For any disk unit x, the frequency-amplitude attenuation coefficient curve of the vibration signal transmitted to the vibration monitoring sensor s is denoted as L_a(x, s), and the frequency-phase offset curve is denoted as L_φ(x, s).

4. The network device fault prediction method based on multi-source data fusion according to claim 1, characterized in that, The step S400 is divided into the following steps: Step S401: Set the frequency step f_step for the vibration signal spectrum, divide all frequency components belonging to the frequency level interval [(n - 1)×f_step, n×f_step] in the vibration signal into frequency level n, and add the amplitudes of all frequency components in the same frequency level and divide by the frequency step f_step as the amplitude of the belonging frequency level; where n is the frequency level number and the value is a positive integer. Step S402: For any disk unit x, the amplitude attenuation coefficient k(x,s,f) of any frequency f component of the vibration signal generated by reading and writing data transmitted to the vibration monitoring sensor s is calculated as follows: Extract the frequency level interval to which frequency f belongs, calculate the integral average value of L_a(x,s) with respect to frequency within the interval, and divide the calculation result by the frequency step size to obtain k(x,s,f); For any disk unit x, the phase offset δφ(x,s,f) of any frequency f component of the vibration signal generated by reading and writing data transmitted to the vibration monitoring sensor s is calculated as follows: Extract the frequency class interval to which frequency f belongs, calculate the integral average value of L_φ(x,s) with respect to frequency within the interval, and divide the calculation result by the frequency step size to obtain δφ(x,s,f).

5. A network device fault prediction system based on multi-source data fusion, applying the network device fault prediction method based on multi-source data fusion as described in any one of claims 1-4, characterized in that, The system includes: a disk vibration prediction module, a vibration transmission analysis module, and a faulty disk location module; The disk vibration prediction module analyzes historical data of vibration signals generated when each disk unit in the disk array reads and writes data, trains a vibration signal prediction neural network model, and predicts the vibration signals generated by each disk unit's read and write data based on real-time monitored data access requests. The vibration transmission analysis module constructs a three-dimensional material model of the disk array and analyzes the amplitude and phase changes of each frequency component of the vibration signal transmitted from the location of each disk unit to each vibration monitoring sensor. The fault disk location module constructs a fault location model by comparing and analyzing the vibration signals predicted by the vibration monitoring sensors with the measured vibration signal data, and locates the fault in each disk unit.

6. The network device fault prediction system based on multi-source data fusion according to claim 5, characterized in that, The disk vibration prediction module includes: a vibration data processing unit, a vibration feature analysis unit, and a vibration signal prediction unit; The vibration data processing unit is used to process historical vibration signal data generated when each disk unit in the disk array reads and writes data; the vibration feature analysis unit uses Fourier transform to obtain the spectral data of the vibration signal and annotates the vibration signal feature parameters; the vibration signal prediction unit trains a vibration signal prediction neural network model by constructing a set of vibration signal feature parameters.

7. The network device fault prediction system based on multi-source data fusion according to claim 5, characterized in that, The vibration transmission analysis module includes: a material model construction unit and a signal transmission analysis unit; The material model construction unit acquires physical structure data of the disk array and constructs a three-dimensional material model; the signal transmission analysis unit analyzes the amplitude and phase changes of each frequency component of the vibration signal from the location of each disk unit to each vibration monitoring sensor, and fits to obtain the frequency-amplitude attenuation coefficient curve and the frequency-phase offset curve.

8. The network device fault prediction system based on multi-source data fusion according to claim 5, characterized in that, The fault disk location module includes: a vibration deviation analysis unit, a location model construction unit, and a fault disk screening unit; The vibration deviation analysis unit compares and analyzes the vibration signal predicted by the vibration monitoring sensor with the measured vibration signal data; the positioning model construction unit constructs a fault positioning model based on the amplitude attenuation coefficient and phase offset of each frequency component of the vibration signal; the fault disk screening unit iterates the spectral deviation of the frequency f component of the vibration signal of each disk unit multiple times, and filters the fault disk units according to the threshold of the number of fault disks to be determined, and performs disk unit fault investigation feedback in order of increasing vibration signal deviation of each fault disk unit.

Images