A hard disk fault detection method, device, equipment and medium

Abstract

The application discloses a hard disk fault detection method, device and equipment and a medium, and relates to the field of hard disk detection.The method comprises the following steps: acquiring real-time vibration data and real-time noise data of a hard disk in operation; performing feature extraction and fusion on the real-time vibration data and the real-time noise data to obtain fused features; performing feature extraction on the real-time vibration data by using target statistical features to obtain statistical feature values; performing multi-layer wavelet analysis on the real-time noise data, and performing statistical analysis on the obtained wavelet coefficients to obtain a skewness value; splicing the fused features, the statistical feature values and the skewness value, and inputting the spliced features into a pre-trained long short-term memory network for multi-classification fault detection to determine the working state of the hard disk.The application can detect the hard disk fault at any time, realize early warning of the hard disk fault, ensure the real-time performance and accuracy of the hard disk fault detection, and ensure the data safety of a storage server with the hard disk as a main component.

Description

Technical Field

[0001] This invention relates to the field of hard disk testing, and in particular to a method, apparatus, device, and medium for hard disk fault detection. Background Technology

[0002] In storage servers that primarily utilize hard disk drives (HDDs), the demands for storage density and integration are constantly increasing to meet the requirements of high data storage stability and high signal transmission reliability. However, with the continuous increase in magnetic recording density and disk platter rotation speed, hard drive fault detection has become a significant bottleneck affecting storage performance.

[0003] During the operation of a hard disk drive (HDD), abnormal operating conditions can cause abnormal vibrations and noise in the internal components. These abnormalities have a relatively small impact on data transmission in the early stages of operation and are difficult to detect using the Self-Monitoring, Analysis and Reporting Technology (SMART) integrated into the HDD firmware. Furthermore, SMART can only monitor static parameters such as the number of remapped sectors and seek error rate, comparing these parameters with preset thresholds to detect hard drive failure. In reality, hard drive failure involves a progressive failure process: "parameter abnormality → performance degradation → functional failure." Currently, using a simple "normal / faulty" binary classification based on comparing static parameters with fixed thresholds makes it difficult to detect early-stage abnormalities. In the later stages, this limitation restricts the real-time performance and accuracy of hard drive failure detection and can directly lead to hard drive failure, ultimately threatening the data security of the storage server.

[0004] It is evident that how to perform hard drive failure detection at any time to achieve early warning of hard drive failure is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a hard disk fault detection method, apparatus, device, and medium that can perform hard disk fault detection at any time, thereby achieving early warning of hard disk faults, ensuring the real-time performance and accuracy of hard disk fault detection, and guaranteeing the data security of storage servers with hard disks as their main component. The specific solution is as follows:

[0006] In a first aspect, the present invention provides a hard disk fault detection method, comprising:

[0007] Acquire real-time vibration and noise data of the hard drive during operation;

[0008] Feature extraction and fusion are performed on real-time vibration data and real-time noise data to obtain fused features;

[0009] Target statistical features are used to extract features from real-time vibration data to obtain corresponding statistical feature values;

[0010] Multi-level wavelet analysis was performed on the real-time noise data, and statistical analysis was conducted on the wavelet coefficients obtained from the analysis to obtain the corresponding skewness values.

[0011] Features, statistical feature values, and skewness values ​​are spliced ​​and fused, and the spliced ​​features are input into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard drive.

[0012] Optionally, feature extraction and fusion are performed on real-time vibration data and real-time noise data to obtain fused features, including:

[0013] The real-time vibration data and real-time noise data are normalized to obtain normalized vibration data and noise data;

[0014] Multiple preset operations are performed on the normalized vibration and noise data to extract and fuse features, so as to obtain fused features;

[0015] Each preset operation includes a convolution operation, an activation operation, and a pooling operation, and the output of each preset operation becomes the input of the next preset operation.

[0016] Optionally, the real-time vibration data includes the first real-time vibration data of the hard drive's center of mass in the left-right direction, the second real-time vibration data in the front-back direction, and the third real-time vibration data in the vertical direction, wherein the left-right direction, the vertical direction, and the front-back direction are perpendicular to each other.

[0017] Accordingly, the target statistical features are used to extract features from the real-time vibration data to obtain the corresponding statistical feature values, including:

[0018] By utilizing the target statistical features corresponding to each real-time vibration data, feature extraction is performed on the corresponding real-time vibration data to obtain the statistical feature values ​​corresponding to each real-time vibration data.

[0019] Among them, the target statistical feature corresponding to the first real-time vibration data and the second real-time vibration data is the root square amplitude, and the target statistical feature corresponding to the third real-time vibration data is the root mean square value.

[0020] Optional, the process of determining the target statistical characteristics includes:

[0021] Collect several historical vibration data of the hard drive under different working conditions to obtain vibration datasets under different working conditions, and obtain a statistical feature set containing several different types of statistical features;

[0022] Based on the different types of statistical characteristics of each sample in the vibration dataset under the same working condition, the initial average distance and average characteristic value for different types of statistical characteristics under the same working condition are determined.

[0023] Based on the initial average distance, determine the first average distance and the first variance factor for different types of statistical characteristics under different working conditions, and determine the second average distance and the second variance factor for different types of statistical characteristics under different working conditions based on the average characteristic value;

[0024] From the first variance factor and the second variance factor, determine the first target variance factor and the second target variance factor with the largest values, respectively. Based on the first variance factor, the first target variance factor, the second variance factor and the second target variance factor, determine the compensation factors for different types of statistical characteristics under different working conditions.

[0025] The compensation factor is corrected using the first and second average distances to obtain correction factors for different types of statistical characteristics under different working conditions;

[0026] The maximum correction factor is determined from the correction factors, and the target statistical features are determined based on the statistical features corresponding to the maximum correction factor.

[0027] Optionally, when the multi-level wavelet analysis is reduced to a two-level wavelet analysis, multi-level wavelet analysis is performed on the real-time noise data, and statistical analysis is conducted on the obtained wavelet coefficients to obtain the corresponding skewness values, including:

[0028] The real-time noise data is processed using low-pass and high-pass filters corresponding to the preset wavelet basis functions to obtain the first-layer low-frequency approximation coefficients and the first-layer high-frequency detail coefficients.

[0029] The first-level low-frequency approximation coefficients are decomposed to obtain the second-level low-frequency approximation coefficients and the second-level high-frequency detail coefficients;

[0030] Based on the low-frequency approximation coefficients of the second layer and the high-frequency detail coefficients of each layer, the mean and standard deviation of the coefficients are determined, and the skewness value is determined based on the mean and standard deviation of the coefficients.

[0031] Optional, the training process for the Long Short-Term Memory network includes:

[0032] An initial population is constructed based on a first preset number of individuals, and the initial population is determined as the current population; wherein, each individual in the current population represents a combination of hyperparameters of the Long Short-Term Memory network, and different individuals represent different combinations of hyperparameters;

[0033] Perform a second preset number of iterations on the current population to obtain the target population;

[0034] The fitness of each individual in the target population is determined based on a preset fitness function, and the target individual with the highest fitness is selected from the target population; wherein, the preset fitness function is a function determined based on the fault detection accuracy of a long short-term memory network;

[0035] Determine the target hyperparameter combination for the representation of the target individual;

[0036] The long short-term memory network with the target hyperparameter combination is trained to obtain a pre-trained long short-term memory network.

[0037] Optionally, perform the current round of iteration operations on the current population, including:

[0038] The fitness of each individual in the current population is determined based on a preset fitness function;

[0039] The individuals in the current population are divided into discoverer set, follower set and vigilant set based on fitness and a preset ratio. The fitness of each individual in the discoverer set is greater than that of each individual in the follower set, and the fitness of each individual in the follower set is greater than that of each individual in the vigilant set.

[0040] The position of each individual in the discoverer set, follower set, and vigilant set is updated using different preset position update logics to perform the current round of iteration update for the current population.

[0041] Secondly, the present invention provides a hard disk fault detection device, comprising:

[0042] The data acquisition module is used to acquire real-time vibration data and real-time noise data of the hard drive during operation;

[0043] The feature fusion module is used to extract and fuse features from real-time vibration data and real-time noise data to obtain fused features;

[0044] The feature value determination module is used to extract features from real-time vibration data using target statistical features to obtain corresponding statistical feature values;

[0045] The skewness value determination module is used to perform multi-level wavelet analysis on real-time noise data and perform statistical analysis on the wavelet coefficients obtained from the analysis to obtain the corresponding skewness value.

[0046] The fault detection module is used to splice and fuse features, statistical feature values, and skewness values, and then input the spliced ​​features into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard drive.

[0047] Thirdly, the present invention provides an electronic device, comprising:

[0048] Memory, used to store computer programs;

[0049] A processor is used to execute computer programs to implement the steps of the aforementioned hard disk fault detection method.

[0050] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned hard disk fault detection method.

[0051] In this embodiment of the invention, real-time vibration data and real-time noise data during hard disk operation are acquired; features are extracted and fused from the real-time vibration data and real-time noise data to obtain fused features; features are extracted from the real-time vibration data using target statistical features to obtain corresponding statistical feature values; multi-layer wavelet analysis is performed on the real-time noise data, and statistical analysis is performed on the wavelet coefficients obtained from the analysis to obtain corresponding skewness values; the fused features, statistical feature values, and skewness values ​​are spliced ​​together, and the spliced ​​features are input into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard disk.

[0052] Beneficial Effects: This invention effectively solves the problem of low fault detection accuracy caused by using static parameters of the hard drive by collecting real-time vibration and noise data during hard drive operation and extracting and fusing features from these data. Furthermore, this invention utilizes statistical features to extract features from real-time vibration data and performs multi-layer wavelet analysis and statistical analysis of wavelet coefficients on real-time noise data, thereby deeply analyzing the vibration and noise data of the hard drive. By extracting vibration statistical feature values ​​and noise skewness values ​​highly correlated with the hard drive's operating state, it facilitates subsequent improvement in the accuracy of hard drive fault detection. Further, this invention concatenates the fusion features, vibration statistical feature values, and noise skewness values ​​of vibration and noise data, and uses a Long Short-Term Memory (LSTM) network to perform multi-class hard drive fault detection, thereby determining the hard drive's operating state. Compared to simple binary classification, this invention introduces multi-class hard drive fault detection, which can more accurately and precisely reflect the hard drive's operating state. Moreover, by using an LSM network, this invention can perform hard drive fault detection at any time, enabling early warning of hard drive faults, thus ensuring the real-time nature of hard drive fault detection and the data security of storage servers with hard drives as their main component. Attached Figure Description

[0053] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of a hard disk fault detection method provided in an embodiment of the present invention;

[0055] Figure 2 A flowchart for determining target statistical features provided in an embodiment of the present invention;

[0056] Figure 3 A structural diagram of a target neural network provided in an embodiment of the present invention;

[0057] Figure 4 A flowchart illustrating the training process of a target neural network, as provided in an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of a hard disk fault detection device provided in an embodiment of the present invention;

[0059] Figure 6 This is a structural diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0060] 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 of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0061] The terms "comprising" and "having," and any variations thereof, in the specification and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may include steps or units not listed.

[0062] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Currently, hard drives are simply classified as "normal / faulty" by comparing their static parameters with fixed thresholds. This approach struggles to detect early-stage anomalies and, in later stages, limits the real-time performance and accuracy of fault detection, potentially leading to hard drive failure and threatening data security on storage servers. Therefore, this invention provides a hard drive fault detection method capable of performing fault detection at any time, enabling early warning of hard drive failures. This ensures real-time performance and accuracy of fault detection and guarantees data security for storage servers where hard drives are a primary component.

[0064] See Figure 1 As shown, this embodiment of the invention provides a hard disk fault detection method, including:

[0065] Step S11: Obtain real-time vibration data and real-time noise data of the hard drive during operation.

[0066] In this embodiment of the invention, real-time vibration data collected by the vibration sensor during hard disk operation and real-time noise data collected by the audio sensor during hard disk operation are obtained; wherein, the vibration sensor is installed at the center of the hard disk and the audio sensor is fixed to the hard disk frame.

[0067] Furthermore, the hard disk fault detection method proposed in this invention can be applied to a Baseboard Management Controller (BMC). Accordingly, the Baseboard Management Controller can acquire data collected by vibration sensors and audio sensors via an SPI (Serial Peripheral Interface) bus.

[0068] Furthermore, the vibration sensor and the audio sensor need to maintain the same sampling rate to achieve synchronous acquisition of real-time vibration data and real-time noise data during hard drive operation, so that the real-time vibration data and real-time noise data are precisely aligned in time.

[0069] It should be noted that real-time vibration data and real-time noise data can be data at a single moment or a sequence containing data at multiple consecutive moments; and when the real-time vibration data and real-time noise data are sequences containing data at multiple consecutive moments, the number of data points contained in the real-time vibration data and real-time noise data is the same.

[0070] Furthermore, the real-time vibration data can specifically include the first real-time vibration data of the hard drive's center of mass in the left-right direction, the second real-time vibration data in the front-back direction, and the third real-time vibration data in the vertical direction, wherein the left-right direction, the vertical direction, and the front-back direction are perpendicular to each other.

[0071] It should also be noted that the vibration sensor can specifically be a vibration acceleration sensor, and correspondingly, real-time vibration data can specifically be real-time vibration acceleration data. Vibration acceleration data describes the numerical information of the acceleration change of an object during vibration, and analyzing vibration acceleration data can help identify faults in mechanical equipment.

[0072] Step S12: Extract and fuse features from real-time vibration data and real-time noise data to obtain fused features.

[0073] After acquiring real-time vibration and noise data during hard drive operation, in order to suppress the interference of background noise and outliers from vibration and audio sensors on hard drive fault detection results under the complex operating conditions of storage servers, this embodiment of the invention utilizes a convolutional neural network (CNN) to extract and fuse features from the real-time vibration and noise data to obtain corresponding fused features.

[0074] Specifically, the real-time vibration and noise data are first normalized to obtain normalized vibration and noise data. Then, multiple preset operations are performed on the normalized vibration and noise data for feature extraction and fusion to obtain corresponding fused features. Each preset operation includes convolution, activation, and pooling operations, and the output of each preset operation serves as the input for the next. In this way, through multiple "convolution-activation-pooling" operations, this invention can gradually learn the deep feature representation of the data and obtain fused features from multi-source sensor data, thereby improving the accuracy of hard drive fault detection.

[0075] In the process of normalizing real-time vibration data and real-time noise data, the real-time vibration data and real-time noise data are input into the input layer of the convolutional neural network to perform Min-Max Normalization (Min-Max Normalization, deviation normalization) on the real-time vibration data, thereby mapping the real-time vibration data to the specified interval [-1,1] to obtain normalized vibration data. The real-time noise data is then normalized by energy normalization to adjust the energy of the real-time noise data to a uniform scale for easy comparison and analysis, thus obtaining normalized noise data.

[0076] In the process of performing multiple preset operations on normalized vibration data and noise data for feature extraction and fusion, for each preset operation, convolution, activation and pooling operations are performed sequentially on the normalized vibration data and noise data or the output of the previous preset operation to obtain the output of the current preset operation. After all preset operations have been completed, the fusion features are determined based on the output of the last preset operation.

[0077] For the first preset operation, convolution, activation and pooling operations are performed sequentially on the normalized vibration data and noise data to obtain the output of the first preset operation. The output of the first preset operation is then used as the input for the next preset operation.

[0078] For non-first preset operations, the output of the current preset operation is obtained by sequentially performing convolution, activation, and pooling operations on the output of the previous preset operation.

[0079] It should be noted that convolution and activation operations are performed through convolutional layers in a convolutional neural network. Convolution is used for feature mapping, while activation is used to enhance the non-linear expressive power of features using activation functions (such as ReLU). Pooling is performed through pooling layers in a convolutional neural network to reduce the dimensionality of features using max pooling.

[0080] It should also be noted that one preset operation corresponds to a set of network layers containing convolutional and pooling layers. Correspondingly, multiple preset operations correspond to multiple sets of network layers, and the multiple sets of network layers are connected in sequence according to the order of the multiple preset operations.

[0081] Step S13: Use the target statistical features to extract features from the real-time vibration data to obtain the corresponding statistical feature values.

[0082] In this embodiment of the invention, after acquiring real-time vibration data during hard drive operation, it is also necessary to extract features from the real-time vibration data using target statistical features to obtain corresponding statistical feature values. These statistical features include, but are not limited to, mean, variance, root mean square amplitude, root mean square value, peak value, skewness, kurtosis, margin, and waveform indices. The target statistical features are feature parameters that are strongly correlated with hard drive fault detection and are pre-selected from these statistical features.

[0083] Furthermore, the real-time vibration data includes the first real-time vibration data of the hard disk's center of mass in the left-right direction, the second real-time vibration data in the front-back direction, and the third real-time vibration data in the vertical direction, and each real-time vibration data corresponds to a target statistical feature.

[0084] Specifically, by utilizing the target statistical features corresponding to each real-time vibration data, feature extraction is performed on the corresponding real-time vibration data to obtain the statistical feature values ​​corresponding to each real-time vibration data; among them, the target statistical features corresponding to the first and second real-time vibration data are both root square amplitude values, and the target statistical feature corresponding to the third real-time vibration data is the root mean square value.

[0085] The formulas for calculating the root square magnitude and root mean square value are as follows:

[0086] ;

[0087] ;

[0088] Represents the root mean square value. This represents the nth data point in the real-time vibration data, where N represents the total number of data points in the real-time vibration data. This represents the square root magnitude.

[0089] Step S14: Perform multi-level wavelet analysis on the real-time noise data, and perform statistical analysis on the wavelet coefficients obtained from the analysis to obtain the corresponding skewness values.

[0090] In this embodiment of the invention, after acquiring real-time noise data during hard disk operation, in order to reflect the energy changes of noise data in different frequency bands, the invention performs multi-level wavelet analysis on the real-time noise data to obtain wavelet coefficients, and performs statistical analysis on the wavelet coefficients to obtain the corresponding skewness values, thereby reflecting the distribution characteristics of noise data under different frequency components through the skewness values.

[0091] When multi-level wavelet analysis is converted to two-level wavelet analysis, for the first-level wavelet analysis, low-pass and high-pass filters corresponding to the preset wavelet basis functions are used to process the real-time noise data to obtain the first-level low-frequency approximation coefficients and the first-level high-frequency detail coefficients. The preset wavelet basis functions can be the Daubechies wavelet basis functions. For the second-level wavelet analysis, the first-level low-frequency approximation coefficients (i.e., the previous level low-frequency approximation coefficients) are decomposed to obtain the second-level low-frequency approximation coefficients and the second-level high-frequency detail coefficients. Then, based on the second-level low-frequency approximation coefficients and the high-frequency detail coefficients of each level, the mean and standard deviation of the coefficients are determined, and the skewness value is determined based on the mean and standard deviation of the coefficients.

[0092] For example, in the wavelet analysis of the j-th layer (j>1 and is an integer), let the low-frequency approximation coefficient of the j-th layer be denoted as The high-frequency detail coefficients of the j-th layer are At this point, the low-frequency approximation coefficients of the j-th layer and the high-frequency detail coefficients of each layer (i.e., the high-frequency detail coefficients from the first layer to the j-th layer) are merged into a one-dimensional array, denoted as . The formula for calculating the mean coefficient under the j-th layer wavelet analysis is:

[0093] ;

[0094] This represents the mean coefficient value under the j-th level wavelet analysis. This represents the total number of wavelet coefficients in the one-dimensional array w under the j-th level wavelet analysis. Let represent the k-th wavelet coefficient in the one-dimensional array w under the j-th layer wavelet analysis.

[0095] Then, based on the mean coefficients, low-frequency approximation coefficients, and high-frequency detail coefficients of each layer under the j-th layer wavelet analysis, the standard deviation of the coefficients under the j-th layer wavelet analysis is determined. The formula for calculating the standard deviation of the coefficients under the j-th layer wavelet analysis is as follows:

[0096] ;

[0097] Let represent the standard deviation of the coefficients under the j-th level wavelet analysis.

[0098] After obtaining the mean and standard deviation of the coefficients under the j-th level wavelet analysis, the skewness value under the j-th level wavelet analysis is determined based on these values. The formula for calculating the skewness value under the j-th level wavelet analysis is as follows:

[0099] ;

[0100] This represents the skewness value under the wavelet analysis of the j-th layer.

[0101] It should be noted that the embodiments of the present invention are only illustrated by two-layer wavelet analysis. However, three-layer or even more-layer wavelet analysis can be used according to actual needs. The analysis process is similar to that of two-layer wavelet analysis, and will not be illustrated in detail here.

[0102] It should also be noted that steps S12, S13 and S14 are not in any particular order of execution; they can be executed simultaneously or sequentially, and no specific restrictions are made here.

[0103] Step S15: Concatenate and fuse features, statistical feature values, and skewness values, and input the concatenated features into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard drive.

[0104] In this embodiment of the invention, fused features, statistical feature values, and skewness values ​​are concatenated to obtain concatenated features. These concatenated features are then input into a pre-trained Long Short-Term Memory (LSTM) network to perform multi-class fault detection on the hard drive based on the concatenated features, thereby determining the hard drive's operating status. The hard drive's operating status includes, but is not limited to, normal status, warning status, alarm status, and fault status.

[0105] It should be noted that the Long Short-Term Memory (LSTM) network mainly consists of an input gate, a forgetting gate, an output gate, and cell states. Its core processes are as follows:

[0106] Forgetting gate: determines the cell state at time t-1. The specific formula for calculating how much information needs to be forgotten includes:

[0107] ;

[0108] This represents the forget gate information at time t. This represents the Sigmoid function. This represents the first weight matrix. This represents the first bias vector. This represents the input at time t, that is, the concatenated features at time t. This represents the hidden state at time t-1. This represents the hidden state at time t-1. Features after splicing at time t Then, the parts are assembled.

[0109] Input gate: determines the concatenated features at time t. The specific formulas for calculating how much information needs to be added to the cell state include:

[0110] ;

[0111] ;

[0112] This represents the input gate information at time t. This represents the second weight matrix. Let tanh denote the second bias vector, and let tanh denote the hyperbolic tangent function. This represents the third weight matrix. This represents the third bias vector. This represents the predicted cell state at time t.

[0113] Cell state update: The cell state at time t is updated by combining information from the forget gate and the input gate. The specific calculation formula includes:

[0114] ;

[0115] This represents the cell state at time t. This represents the cell state at time t-1. This indicates element-wise multiplication.

[0116] Output gate: determines the cell state at time t. The specific formula for calculating how much information needs to be output includes:

[0117] ;

[0118] ;

[0119] This represents the output information at time t. This represents the fourth weight matrix. This represents the fourth bias vector. This represents the hidden state at time t.

[0120] To enhance the recognition accuracy of the Long Short-Term Memory (LSTM) network, this embodiment of the invention utilizes the ISSA (Improved Sparrow Search Algorithm) algorithm to optimize the hyperparameters of the LSM network and trains the LSM network with the optimal hyperparameters to obtain a pre-trained LSM network. The hyperparameters of the LSM network include, but are not limited to, the number of hidden layers and the number of hidden layer neurons.

[0121] In the training process of the Long Short-Term Memory (LSTM) network, an initial population is first constructed based on a pre-generated number of individuals, and this initial population is designated as the current population. Each individual in the current population represents a hyperparameter combination of the LSM network, with different individuals representing different hyperparameter combinations. Then, the current population undergoes a second set number of iterations to obtain the target population. The fitness of each individual in the target population is determined based on a pre-defined fitness function, and the target individual with the highest fitness is selected. Next, the target hyperparameter combination represented by the target individual is determined; this target hyperparameter combination is the optimal hyperparameter combination. Finally, the LSM network using the target hyperparameter combination is trained to obtain a pre-trained LSM network.

[0122] The preset fitness function is a function used to determine the fault detection accuracy based on the Long Short-Term Memory (LSTM) network; the fault detection accuracy of the LTM network = number of correctly classified samples / total number of samples. It should be noted that the fitness of each individual in the population indirectly reflects the fault detection accuracy of the LTM network when using the hyperparameter combinations represented by each individual. Since the fault detection accuracy of the LTM network may differ with different hyperparameter combinations, the fitness of each individual in the population may also differ accordingly.

[0123] During the current round of iteration operations on the current population, the fitness of each individual in the current population is determined based on a preset fitness function. Then, the individuals in the current population are divided into a discoverer set, a follower set, and a vigilant set based on the fitness and a preset ratio. The fitness of each individual in the discoverer set is greater than that of each individual in the follower set, and the fitness of each individual in the follower set is greater than that of each individual in the vigilant set. Finally, the positions of each individual in the discoverer set, follower set, and vigilant set are updated using different preset position update logics to perform the current round of iteration update on the current population.

[0124] Preferably, the preset position update logic corresponding to the discoverer set is as follows:

[0125] ;

[0126] in, Let represent the position of the i-th individual in the j-th dimension at the (t+1)-th iteration, where j = 1, 2, ..., D, and D represents the dimension of the hyperparameters, that is, the total number of hyperparameters contained in a set of hyperparameter combinations. This represents the position of the i-th individual in the j-th dimension during the t-th iteration. This indicates the step size control parameter. This indicates the maximum preset total number of iterations, which is also the second preset number; This represents the warning value at the t-th iteration. This represents the preset safety threshold, and Q represents a random number that follows a standard normal distribution. This represents the random step size generated by the Levy flight.

[0127] It should be noted that, by introducing exponential decay and Levy flight into the individual position update of the discoverer set, the embodiments of the present invention can achieve dynamic balance exploration and development of individual positions.

[0128] It should also be noted that, in one specific implementation, the warning value at each iteration... The warning value can be randomly generated, ranging from (0,1). In another specific implementation, the initial warning value for the current iteration is generated based on the warning value from the previous iteration. Specifically, if the warning value from the previous iteration is greater than 0 and less than a preset value (e.g., 0.5), the initial warning value for the current iteration is equal to the warning value from the previous iteration × 2. If the warning value from the previous iteration is not less than the preset value and not greater than 1, the initial warning value for the current iteration is equal to (1 - the warning value from the previous iteration) × 2. Then, it is determined whether the initial warning value for the current iteration is the same as the warning value from the iteration before that. If they are the same, a new value is randomly generated within the range (0,1) that is different from the initial warning value for the current iteration to obtain the final warning value for the current iteration. If they are different, the initial warning value for the current iteration is determined as the final warning value for the current iteration. In this way, by randomly generating warning values ​​in segments according to different situations or generating warning values ​​for the current iteration based on the warning values ​​from the previous iteration, the distribution of warning values ​​can be random and uniform, thereby improving the stability of the ISSA algorithm.

[0129] Preferably, the preset position update logic corresponding to the follower set is as follows:

[0130] ;

[0131] in, This represents the position of the i-th individual in the j-th dimension during the (t+1)-th iteration. This represents the position of the i-th individual in the j-th dimension during the t-th iteration. This represents the position of the individual with the lowest fitness in the current population on the j-th dimension at the t-th iteration. This represents the position of the individual with the highest fitness in the discoverer set at the t-th iteration in the j-th dimension. Let represent the fitness of the i-th individual in the t-th iteration. Let A represent the fitness of the individual with the highest fitness in the discoverer set at iteration t, and let A be a 1×D matrix whose elements are randomly selected from ±1. L represents a 1×D vector of all 1s.

[0132] Preferably, the preset position update logic corresponding to the set of vigilants is as follows:

[0133] ;

[0134] in, This represents the position of the i-th individual in the j-th dimension during the (t+1)-th iteration. This represents the position of the r-th individual randomly selected in the t-th iteration on the j-th dimension. This represents the position of the i-th individual in the j-th dimension during the t-th iteration. This represents the position of the individual with the highest fitness in the current population on the j-th dimension at the t-th iteration. Indicates obedience Random numbers.

[0135] That is, during the second preset number of iterations of the current population, the first iteration is performed by dividing the current population into a set of discoverers, followers, and vigilants, and updating the position of each individual in each set using different preset position update logics. This achieves the first iteration update of the current population. At this time, the current iteration count is recorded and it is determined whether the current iteration count has reached the second preset number. If not, the next iteration is performed on the current population after the first iteration update until the current iteration count reaches the second preset number. At this time, the iteration operation ends, and the current population after the last iteration update is determined as the target population. Then, the target individual with the highest fitness is found from the target population. The target hyperparameter combination represented by the target individual is the optimal hyperparameter combination.

[0136] It should be noted that the hard disk fault detection method proposed in this embodiment of the invention can also work in conjunction with SMART technology. Furthermore, the technical concept of the hard disk fault detection method proposed in this embodiment of the invention can be extended to systems containing mechanical hard disks, and even to systems containing mechanical vibration components, for fault detection.

[0137] Beneficial Effects: This invention effectively solves the problem of low fault detection accuracy caused by using static parameters of the hard drive by collecting real-time vibration and noise data during hard drive operation and extracting and fusing features from these data. Furthermore, this invention utilizes statistical features to extract features from real-time vibration data and performs multi-layer wavelet analysis and statistical analysis of wavelet coefficients on real-time noise data, thereby deeply analyzing the vibration and noise data of the hard drive. By extracting vibration statistical feature values ​​and noise skewness values ​​highly correlated with the hard drive's operating state, it facilitates subsequent improvement in the accuracy of hard drive fault detection. Further, this invention concatenates the fusion features, vibration statistical feature values, and noise skewness values ​​of vibration and noise data, and uses a Long Short-Term Memory (LSTM) network to perform multi-class hard drive fault detection, thereby determining the hard drive's operating state. Compared to simple binary classification, this invention introduces multi-class hard drive fault detection, which can more accurately and precisely reflect the hard drive's operating state. Moreover, by using an LSM network, this invention can perform hard drive fault detection at any time, enabling early warning of hard drive faults, thus ensuring the real-time nature of hard drive fault detection and the data security of storage servers with hard drives as their main component.

[0138] Since the target statistical features are pre-selected features strongly correlated with hard drive fault detection from statistical features such as mean, variance, root mean square amplitude, root mean square value, peak value, skewness, kurtosis, margin, and waveform indicators, this embodiment of the invention uses an improved distance evaluation algorithm to perform correlation quantification analysis on each statistical feature in order to select target statistical features strongly correlated with hard drive fault detection. Specifically, see... Figure 2 As shown, the process of determining the target statistical characteristics includes:

[0139] Step S21: Collect several historical vibration data of the hard drive under different working conditions to obtain vibration datasets under different working conditions, and obtain a statistical feature set containing several different types of statistical features.

[0140] In this embodiment of the invention, several historical vibration data points (one historical vibration data point refers to a data sequence containing multiple data points) are collected under each type of working state to obtain a corresponding vibration dataset. That is, each type of working state corresponds to a vibration dataset, and each vibration dataset contains the same number of vibration data points. Simultaneously, this embodiment of the invention also needs to obtain a statistical feature set, which includes several different types of statistical features, such as mean, variance, root mean square amplitude, root mean square value, peak value, skewness, kurtosis, margin, waveform indices, etc.

[0141] With the total number of operating state categories as C, each vibration dataset contains Taking a set of vibration data containing J different types of statistical features as an example, let the feature value of the m-th vibration data in the c-th working state vibration dataset for the j-th statistical feature be denoted as... .

[0142] Step S22: Based on the different types of statistical characteristics of each sample in the vibration dataset under the same working condition, determine the initial average distance and average characteristic value for different types of statistical characteristics under the same working condition.

[0143] First, based on the different statistical characteristics of any two different samples in the vibration dataset under the same working condition, the initial average distance for different statistical characteristics under the same working condition is determined, and the formulas involved are as follows:

[0144] ;

[0145] Let represent the initial average distance for the j-th statistical feature under the c-th working state, where There are a total of C×J possible values; This represents the feature value of the l-th vibration data in the vibration dataset under the c-th working condition for the j-th statistical feature.

[0146] Secondly, based on the different types of statistical characteristics of each sample in the vibration dataset under the same working condition, the average characteristic value for different types of statistical characteristics under the same working condition is determined. The formulas involved are as follows:

[0147] ;

[0148] Let represent the average characteristic value of the j-th statistical feature under the c-th working state, where There are a total of C×J possible values.

[0149] Step S23: Determine the first average distance and the first variance factor for different types of statistical features under different working conditions based on the initial average distance, and determine the second average distance and the second variance factor for different types of statistical features under different working conditions based on the average characteristic value.

[0150] Based on the initial average distance for different types of statistical features under the same working conditions, determine the first average distance for different types of statistical features under different working conditions.

[0151] Specifically, based on the initial average distances for different types of statistical features under the same working conditions, the initial average distances for the same type of statistical features under different working conditions are obtained. The average of these initial average distances is then calculated to obtain the first average distance for the same type of statistical features under different working conditions. This process is repeated to obtain the first average distances for different types of statistical features under different working conditions. The formulas involved are as follows:

[0152] ;

[0153] Let represent the first average distance for the j-th statistical feature under different working conditions, where There are a total of J possible values.

[0154] Based on the initial average distance for different types of statistical features under the same working conditions, the first variance factor for different types of statistical features under different working conditions is determined, wherein the first variance factor is the factor corresponding to the first average distance.

[0155] Specifically, based on the initial average distances for different types of statistical features under the same working conditions, the initial average distances for the same type of statistical features under different working conditions are obtained. Then, the maximum and minimum initial average distances are determined from these initial average distances. Finally, the first variance factor for the same type of statistical feature under different working conditions is determined based on the ratio of the maximum to the minimum initial average distance. This process is repeated to obtain the first variance factors for different types of statistical features under different working conditions. The formulas involved are as follows:

[0156] ;

[0157] Denotes the first variance factor for the j-th statistical feature under different working conditions, where, There are a total of J possible values.

[0158] Based on the average characteristic values ​​of different types of statistical features under the same working conditions, determine the second average distance for different types of statistical features under different working conditions.

[0159] Specifically, based on the average characteristic values ​​of different types of statistical features under the same working conditions, the average characteristic values ​​of the same type of statistical features under different working conditions are obtained. Then, based on the average characteristic values ​​of the same type of statistical features under different working conditions, the second average distance for the same type of statistical features under different working conditions is determined. This process is repeated to obtain the second average distance for different types of statistical features under different working conditions. The formulas involved are as follows:

[0160] ;

[0161] This represents the second average distance for the j-th statistical feature under different working conditions, where... There are a total of J possible values. It represents the average characteristic value of the j-th statistical feature under the e-th working state.

[0162] Based on the average eigenvalues ​​of different types of statistical features under the same working conditions, the second variance factor for different types of statistical features under different working conditions is determined, wherein the second variance factor is the factor corresponding to the second average distance.

[0163] Specifically, based on the average eigenvalues ​​of different types of statistical characteristics under the same working conditions, the average eigenvalues ​​of the same type of statistical characteristics under different working conditions are obtained. Then, the maximum and minimum average eigenvalues ​​are determined from these average eigenvalues. Finally, the second variance factor for the same type of statistical characteristic under different working conditions is determined based on the ratio of the maximum to the minimum average eigenvalues. This process is repeated to obtain the second variance factors for different types of statistical characteristics under different working conditions. The formulas involved are as follows:

[0164] ;

[0165] Denotes the second variance factor for the j-th statistical feature under different working conditions, where, There are a total of J possible values.

[0166] Step S24: From the first variance factor and the second variance factor, determine the first target variance factor and the second target variance factor with the largest values ​​respectively, and based on the first variance factor, the first target variance factor, the second variance factor and the second target variance factor, determine the compensation factors for different types of statistical characteristics under different working conditions.

[0167] From the first variance factors for different types of statistical characteristics under different working conditions, determine the first target variance factor with the largest value. Furthermore, from the second variance factors for different types of statistical characteristics under different working conditions, the second objective variance factor with the largest value is determined. .

[0168] Using the first and second objective variance factors, and based on these factors under different operating conditions for the same type of statistical characteristic, compensation factors for the same type of statistical characteristic under different operating conditions are determined. This process can be repeated to obtain compensation factors for different types of statistical characteristics under different operating conditions. The formulas involved are as follows:

[0169] ;

[0170] Let represent the compensation factor for the j-th statistical feature under different working conditions, where There are a total of J possible values.

[0171] Step S25: Use the first average distance and the second average distance to correct the compensation factor to obtain correction factors for different types of statistical characteristics under different working conditions.

[0172] By utilizing the first and second average distances for the same type of statistical characteristic under different working conditions, the compensation factors for the same type of statistical characteristic under different working conditions are corrected to obtain the correction factors for the same type of statistical characteristic under different working conditions. This process can be repeated to obtain the correction factors for different types of statistical characteristics under different working conditions. The formulas involved are as follows:

[0173] ;

[0174] Let represent the correction factor for the j-th statistical feature under different working conditions, where There are a total of J possible values.

[0175] Step S26: Determine the maximum correction factor from the correction factors, and determine the target statistical features based on the statistical features corresponding to the maximum correction factor.

[0176] According to one specific implementation method, the maximum correction factor is determined directly from the correction factors for different types of statistical features under different working conditions, and the target statistical feature is determined based on the statistical feature corresponding to the maximum correction factor.

[0177] According to another specific implementation, the correction factors for the same type of statistical feature under different working conditions are first normalized to obtain normalized correction factors for the same type of statistical feature under different working conditions. This process is repeated to obtain normalized correction factors for different types of statistical features under different working conditions. Then, the largest correction factor is determined from the normalized correction factors for different types of statistical features under different working conditions, and the target statistical feature is determined based on the statistical feature corresponding to the largest correction factor. The formulas involved in the normalization process are as follows:

[0178] ;

[0179] Let represent the normalization correction factor for the j-th statistical feature under different working conditions, where There are a total of J possible values.

[0180] It should be noted that the larger the correction factor, the stronger the correlation between the statistical characteristics corresponding to the correction factor and hard drive fault detection, and the better it can distinguish different working states of the hard drive; correspondingly, the smaller the correction factor, the weaker the correlation between the statistical characteristics corresponding to the correction factor and hard drive fault detection.

[0181] It should also be noted that, since the vibration data includes first vibration data of the hard drive's center of mass in the left-right direction, second vibration data in the front-back direction, and third vibration data in the vertical direction, steps S21-S26 described above need to be performed on the first vibration data, second vibration data, and third vibration data respectively to obtain the target statistical features corresponding to the first vibration data, second vibration data, and third vibration data respectively. The target statistical features corresponding to the first vibration data, second vibration data, and third vibration data may be partially the same or completely identical, or partially different or completely different.

[0182] Beneficial effects: This embodiment of the invention uses an improved distance evaluation algorithm to perform correlation quantification analysis on each statistical feature in the statistical feature set, in order to screen target statistical features that are strongly correlated with hard disk fault detection. Then, the target statistical features are used to extract features from real-time vibration data to analyze the vibration data of the hard disk in depth. The extracted vibration statistical feature values ​​that are highly correlated with the working state of the hard disk are used to predict the working state of the hard disk, which facilitates the subsequent improvement of the accuracy of hard disk fault detection.

[0183] As can be seen from the two embodiments described above, the target neural network used to perform multi-class fault detection on the hard drive to determine the working state of the hard drive in the embodiments of the present invention is a neural network composed of a convolutional neural network, an improved distance evaluation module, a wavelet analysis module, and a long short-term memory network.

[0184] by Figure 3 Taking a specific target neural network structure diagram as an example, the convolutional neural network contains an input layer and two sets of network layers. Each set of network layers contains convolutional layers and pooling layers, and the two sets of network layers are connected sequentially. The long short-term memory network contains an input layer, a dropout layer, a fully connected layer, and a softmax layer. The dropout layer determines how much information from the cell state at the previous time step needs to be discarded. The fully connected layer is used to output the raw output values ​​of each category of hard disk working state. The softmax layer is used to convert the raw output values ​​of each category of hard disk working state into output probabilities.

[0185] like Figure 4 As shown, the training process of the target neural network is as follows:

[0186] Vibration and noise data of the hard drive under different operating conditions are collected using vibration acceleration sensors and audio sensors, and the collected vibration and noise data are filtered. The vibration data includes first vibration data in the left-right direction, second vibration data in the front-back direction, and third vibration data in the vertical direction, all originating from the hard drive's center of mass. A sliding window approach (window size n, sliding step size s) is then used to segment the vibration and noise data into several vibration and noise data sequences, each containing n data points. These sequences are normalized, and a hard drive fault detection dataset is built based on the normalized sequences. The dataset is then divided into training and validation sets according to a preset ratio. Each sample in the dataset contains three vibration data sequences and one noise data sequence under the same operating condition, and the three vibration data sequences and one noise data sequence are synchronized in time.

[0187] The target neural network is trained using a training set to obtain the trained neural network. During the training process, the ISSA algorithm is used to optimize the hyperparameters of the Long Short-Term Memory (LSTM) network, and these hyperparameters are fixed after finding the optimal ones. Then, a convolutional neural network is used to extract and fuse features from vibration and noise data in each sample to obtain the fused features for each sample. An improved distance evaluation module is used to extract vibration features from each sample using target statistical features to obtain statistical feature values ​​for each sample. A wavelet analysis module is then used to perform multi-level wavelet analysis on the noise data in each sample, and the obtained wavelet coefficients are statistically analyzed to obtain the skewness value for each sample. The fused features, statistical feature values, and skewness values ​​of each sample are concatenated to obtain the concatenated features for each sample. These concatenated features are then input into the LSM network for multi-class fault detection to obtain multi-class probability distributions for different operating states. The final operating state of the hard drive is predicted based on the operating state corresponding to the highest probability in the probability distribution. Furthermore, the cross-entropy loss function is used to evaluate the loss value between the predicted output and the actual output of the target neural network, and the loss value is used to optimize the target neural network.

[0188] For loss calculation, after the concatenated features of each sample are input into the Long Short-Term Memory (LSTM) network, the fully connected layers in the LTM network output raw output values ​​for different operating states. The Softmax layers in the LTM network then convert these raw output values ​​for different operating states into probability values ​​for those states. Finally, the loss value is calculated based on these probability values ​​and the actual operating state of each sample. The formulas involved are as follows:

[0189] ;

[0190] ;

[0191] This represents the original output value for the i-th working state, where k represents the total number of different working state categories. Let L represent the probability value for the i-th working state, and L represent the loss value. This represents the one-hot encoded vector for the i-th operating state, and the one-hot encoded vector for different operating states is a vector determined based on the actual operating state of each sample data.

[0192] The trained neural network is validated using a validation set. When the validation passes, for example, when the validation accuracy reaches a preset accuracy, the trained neural network is determined to be a well-trained neural network. At this point, the well-trained neural network can be applied to the real-time online fault detection of hard drives.

[0193] It should be noted that, in this embodiment of the invention, the hard drive's operating status is divided into four categories: normal state, warning state, alarm state, and fault state. Furthermore, in real-time online fault detection, the detected real-time operating status of the hard drive needs to be uploaded to a pre-set operation and maintenance system so that maintenance personnel can monitor the hard drive status in a timely manner and intervene and maintain the hard drive promptly when abnormalities occur.

[0194] Beneficial Effects: This invention effectively solves the problem of low fault detection accuracy caused by using static parameters of the hard drive by collecting real-time vibration and noise data during hard drive operation and extracting and fusing features from these data. Furthermore, this invention utilizes statistical features to extract features from real-time vibration data and performs multi-layer wavelet analysis and statistical analysis of wavelet coefficients on real-time noise data, thereby deeply analyzing the vibration and noise data of the hard drive. By extracting vibration statistical feature values ​​and noise skewness values ​​highly correlated with the hard drive's operating state, it facilitates subsequent improvement in the accuracy of hard drive fault detection. Further, this invention concatenates the fusion features, vibration statistical feature values, and noise skewness values ​​of vibration and noise data, and uses a Long Short-Term Memory (LSTM) network to perform multi-class hard drive fault detection, thereby determining the hard drive's operating state. Compared to simple binary classification, this invention introduces multi-class hard drive fault detection, which can more accurately and precisely reflect the hard drive's operating state. Moreover, by using an LSM network, this invention can perform hard drive fault detection at any time, enabling early warning of hard drive faults, thus ensuring the real-time nature of hard drive fault detection and the data security of storage servers with hard drives as their main component.

[0195] See Figure 5 As shown, an embodiment of the present invention provides a hard disk fault detection device, comprising:

[0196] Data acquisition module 11 is used to acquire real-time vibration data and real-time noise data of the hard drive during operation;

[0197] The feature fusion module 12 is used to extract and fuse features from real-time vibration data and real-time noise data to obtain fused features;

[0198] The feature value determination module 13 is used to extract features from real-time vibration data using target statistical features to obtain corresponding statistical feature values;

[0199] The skewness value determination module 14 is used to perform multi-level wavelet analysis on real-time noise data and perform statistical analysis on the wavelet coefficients obtained from the analysis to obtain the corresponding skewness value.

[0200] The fault detection module 15 is used to splice and fuse features, statistical feature values ​​and skewness values, and input the spliced ​​features into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard disk.

[0201] Since the embodiments of the device part correspond to the embodiments described above, please refer to the embodiments described in the method part for the embodiments of the device part, and will not be repeated here.

[0202] Beneficial Effects: This invention effectively solves the problem of low fault detection accuracy caused by using static parameters of the hard drive by collecting real-time vibration and noise data during hard drive operation and extracting and fusing features from these data. Furthermore, this invention utilizes statistical features to extract features from real-time vibration data and performs multi-layer wavelet analysis and statistical analysis of wavelet coefficients on real-time noise data, thereby deeply analyzing the vibration and noise data of the hard drive. By extracting vibration statistical feature values ​​and noise skewness values ​​highly correlated with the hard drive's operating state, it facilitates subsequent improvement in the accuracy of hard drive fault detection. Further, this invention concatenates the fusion features, vibration statistical feature values, and noise skewness values ​​of vibration and noise data, and uses a Long Short-Term Memory (LSTM) network to perform multi-class hard drive fault detection, thereby determining the hard drive's operating state. Compared to simple binary classification, this invention introduces multi-class hard drive fault detection, which can more accurately and precisely reflect the hard drive's operating state. Moreover, by using an LSM network, this invention can perform hard drive fault detection at any time, enabling early warning of hard drive faults, thus ensuring the real-time nature of hard drive fault detection and the data security of storage servers with hard drives as their main component.

[0203] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. Specifically, the electronic device may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the hard disk fault detection method disclosed in any of the foregoing embodiments. Furthermore, the electronic device in this embodiment may specifically be an electronic computer.

[0204] In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0205] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0206] The operating system 221 is used to manage and control the various hardware devices on the electronic device and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the hard disk fault detection method executed by the electronic device as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0207] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed hard disk fault detection method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0208] Furthermore, this application also discloses a computer program product, including a computer program / instructions; wherein, when the computer program / instructions are executed by a processor, they implement the aforementioned disclosed hard disk fault detection method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0209] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0210] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0211] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0212] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0213] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A hard disk fault detection method, characterized in that, include: Acquire real-time vibration and noise data of the hard drive during operation; Feature extraction and fusion are performed on the real-time vibration data and the real-time noise data to obtain fused features. The process includes: normalizing the real-time vibration data and the real-time noise data to obtain normalized vibration data and noise data; performing multiple preset operations on the normalized vibration data and noise data for feature extraction and fusion to obtain fused features; each preset operation includes convolution, activation, and pooling operations, and the output of each preset operation serves as the input for the next preset operation. The real-time vibration data is subjected to feature extraction using target statistical features to obtain corresponding statistical feature values. The real-time vibration data includes first real-time vibration data of the hard disk's center of mass in the left-right direction, second real-time vibration data in the front-back direction, and third real-time vibration data in the vertical direction, wherein the left-right direction, the vertical direction, and the front-back direction are mutually perpendicular. Accordingly, the step of using target statistical features to extract features from the real-time vibration data to obtain corresponding statistical feature values ​​includes: using target statistical features corresponding to each of the real-time vibration data to extract features from the corresponding real-time vibration data to obtain statistical feature values ​​corresponding to each of the real-time vibration data; wherein the target statistical features corresponding to the first and second real-time vibration data are both root square amplitude values, and the target statistical feature corresponding to the third real-time vibration data is a root mean square value. Multi-level wavelet analysis is performed on the real-time noise data, and statistical analysis is performed on the obtained wavelet coefficients to obtain the corresponding skewness values. Specifically, when the multi-level wavelet analysis is a two-level wavelet analysis, the process of performing multi-level wavelet analysis on the real-time noise data and performing statistical analysis on the obtained wavelet coefficients to obtain the corresponding skewness values ​​includes: processing the real-time noise data using low-pass and high-pass filters corresponding to preset wavelet basis functions to obtain first-level low-frequency approximation coefficients and first-level high-frequency detail coefficients; decomposing the first-level low-frequency approximation coefficients to obtain second-level low-frequency approximation coefficients and second-level high-frequency detail coefficients; determining the mean and standard deviation of the coefficients based on the second-level low-frequency approximation coefficients and each level of high-frequency detail coefficients, and determining the skewness value based on the mean and standard deviation of the coefficients. The fused features, statistical feature values, and skewness values ​​are concatenated, and the concatenated features are input into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard drive.

2. The hard disk fault detection method according to claim 1, characterized in that, The process of determining the target statistical features includes: Collect several historical vibration data of the hard drive under different working conditions to obtain vibration datasets under different working conditions, and obtain a statistical feature set containing several different types of statistical features; Based on the different types of statistical characteristics of each sample in the vibration dataset under the same working condition, the initial average distance and average characteristic value for different types of statistical characteristics under the same working condition are determined. Based on the initial average distance, determine the first average distance and the first variance factor for different types of statistical features under different working conditions, and determine the second average distance and the second variance factor for different types of statistical features under different working conditions based on the average feature value; From the first variance factor and the second variance factor, determine the first target variance factor and the second target variance factor with the largest values, respectively, and based on the first variance factor and the first target variance factor, the second variance factor and the second target variance factor, determine the compensation factors for different types of statistical characteristics under different working conditions; The compensation factor is corrected using the first average distance and the second average distance to obtain correction factors for different types of statistical characteristics under different working conditions; The largest correction factor is determined from the correction factors, and the target statistical feature is determined based on the statistical feature corresponding to the largest correction factor.

3. The hard disk fault detection method according to claim 1 or 2, characterized in that, The training process of the Long Short-Term Memory network includes: An initial population is constructed based on a first preset number of individuals, and the initial population is determined as the current population; wherein, each individual in the current population represents a hyperparameter combination of the long short-term memory network, and different individuals represent different hyperparameter combinations; Perform a second preset number of iterations on the current population to obtain the target population; The fitness of each individual in the target population is determined based on a preset fitness function, and the target individual with the highest fitness is determined from the target population; wherein, the preset fitness function is a function determined based on the fault detection accuracy of the long short-term memory network; Determine the target hyperparameter combination representing the target individual; The long short-term memory network using the target hyperparameter combination is trained to obtain the pre-trained long short-term memory network.

4. The hard disk fault detection method according to claim 3, characterized in that, Performing the current round of iteration operation on the current population includes: The fitness of each individual in the current population is determined based on the preset fitness function; The individuals in the current population are divided using fitness and based on a preset ratio to obtain a discoverer set, a follower set, and a vigilant set; wherein the fitness of each individual in the discoverer set is greater than the fitness of each individual in the follower set, and the fitness of each individual in the follower set is greater than the fitness of each individual in the vigilant set. The position of each individual in the discoverer set, the follower set, and the watcher set is updated using different preset position update logics to perform the current round of iteration update for the current population.

5. A hard disk fault detection device, characterized in that, The hard disk fault detection device is used to implement the hard disk fault detection method as described in any one of claims 1 to 4, and the hard disk fault detection device comprises: The data acquisition module is used to acquire real-time vibration data and real-time noise data of the hard drive during operation; The feature fusion module is used to extract and fuse features from the real-time vibration data and the real-time noise data to obtain fused features; The feature value determination module is used to extract features from the real-time vibration data using target statistical features to obtain corresponding statistical feature values; The skewness value determination module is used to perform multi-layer wavelet analysis on the real-time noise data and perform statistical analysis on the wavelet coefficients obtained from the analysis to obtain the corresponding skewness value. The fault detection module is used to splice the fused features, statistical feature values ​​and the skewness values, and input the spliced ​​features into a pre-trained long short-term memory network for multi-class fault detection to determine the working status of the hard disk.

6. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the hard disk fault detection method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the hard disk fault detection method as described in any one of claims 1 to 4.

Images