CN108415810B - A kind of hard disk state monitoring method and device - Google Patents

Abstract

The invention discloses a hard disk state monitoring method and device. The method includes: periodically acquiring different SMART attribute values of a monitored hard disk; and performing the following operations after acquiring the SMART attribute values of the hard disk each time: according to the current acquisition The SMART attribute value of the described hard disk, obtain the attribute integration of this hard disk; Input the attribute integration into the recurrent neural network introduced into the gated recursive unit, and use the hidden layer state of the recurrent neural network as the output; The information obtained from the output of the recurrent neural network is used to monitor whether the hard disk is about to fail. By applying the present invention, the deviation of the normal state of the hard disk drive can be traced back to an earlier period, so as to improve the failure detection rate or the failure prediction ability.

Description

A kind of hard disk state monitoring method and device

technical field

The invention relates to the technical field of hard disk failure monitoring, in particular to a hard disk state monitoring method and device.

Background technique

In the era of big data, data centers equipped with large storage systems play an important role in storing and processing data. However, complex systems have caused serious problems of IT equipment failure, with hard disks being the most common failed component. While a single hard drive failure may be relatively rare, stacking thousands of hard drives together magnifies the probability of failure, making failure events the norm rather than the exception in data center storage systems. Considering the huge financial losses caused by data loss and service interruption, the issue of hard drive reliability is one of the top concerns for data center administrators.

Today, almost all hard drives are equipped with Self-Monitoring, Analysis and Reporting Technology (SMART) to detect and report various drive reliability metrics. Studies have shown that hard drives can predict impending failure through SMART data. It is even used by some HDD manufacturers as a failure prediction model built into their products. However, the built-in model only provides a basic threshold-based assessment and its failure prediction capability is rather weak. In order to improve the failure prediction ability, the researchers proposed statistical and machine learning methods based on SMART data. While these methods show good performance in terms of fault detection rate and false positive rate, there are still some key unsolved challenges:

Failed hard drives often go through a deteriorating process from healthy to failure. But most of the existing methods predict failures based on a timestamp of SMART (Self-Monitoring Analysis and Reporting Technology) data, ignoring the temporal degradation process. Some methods based on Markov models and simple recurrent neural networks (RNNs) try to capture temporal dependencies in SMART data. However, limited by the inherent problems of these models, such as vanishing and exploding gradients in RNNs, these methods can only capture short-term temporal dependencies over several days.

However, according to the observation and analysis of the inventors of the present invention, the deviation from the normal state of some hard disk drives can be traced back to tens of days or even months, which greatly exceeds the capabilities of these methods.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide a hard disk state monitoring method and device, which can trace the deviation of the normal state of the hard disk drive to an earlier period, so as to improve the failure detection rate or the failure prediction ability.

Based on the above purpose, the present invention provides a hard disk state monitoring method, comprising:

Periodically obtain different SMART attribute values of the monitored hard disk; perform the following operations after obtaining the SMART attribute value of the hard disk each time:

According to the currently acquired SMART attribute value of the hard disk, the attribute integration of the hard disk is obtained;

The attribute integration is input into the recurrent neural network introduced into the gated recurrent unit, and the hidden layer state of the recurrent neural network is used as the output;

Based on the information obtained from the output of the recurrent neural network, monitor whether the hard disk is about to fail.

Wherein, according to the currently acquired SMART attribute value of the hard disk, the attribute integration of the hard disk is obtained, specifically:

得到的属性集成表示为

根据如下表达式一计算得到：A SMART vector composed of the currently acquired SMART attribute values of the hard disk on day t

The resulting property ensemble is expressed as

Calculated according to the following expression:

v _t =ReLU(W _V s _t +b _v ) (Expression 1)

表示SMART属性值的权重矩阵，

是偏置向量。ReLU是定义为ReLU(x)＝x+＝max(0,x)的激活函数，其中，max是逐元素操作；W_V和b_v是预先在训练过程中进行学习得到的向量；

表示维度为d_s的实数向量组；

表示维度为d_v的实数向量组；d_s为SMART属性值个数、d_v为属性集成值个数。in,

weight matrix representing SMART attribute values,

is the bias vector. ReLU is an activation function defined as ReLU(x)=x+=max(0,x), where max is an element-wise operation; W _V and b _v are vectors learned in advance during the training process;

represents a real vector group of dimension d _s ;

Represents a real vector group with dimension d _v ; d _s is the number of SMART attribute values, and d _v is the number of attribute integration values.

Wherein, the gated recurrent unit in the recurrent neural network includes a gated unit and a recursive unit; wherein,

In a gated recursive unit, the gated unit is used to control the information flow of the recursive unit, so that the recursive unit captures long-term dependencies; where,

The gating unit includes a reset gate and an update gate to allow the recursive unit to maintain existing content or update content based on existing content.

Wherein, the relationship between the input and output of the recurrent neural network is specifically realized by the recursive algorithm of the following four expressions:

r _t =sigmoid(W _r v _t +U _r h _t-1 ) (Expression 2)

z _t =sigmoid(W _z v _t +U _z h _t-1 ) (Expression 3)

h _t '=tanh(Wv _t +U(r _t ⊙h _t-1 )) (Expression 4)

h _t =z _t ⊙h _t-1 +(1-z _t )⊙h' _t (Expression 5)

Among them, ⊙ is the element-wise multiplication operation; r _t represents the reset gate, h' _t represents the alternative state, z _t represents the update gate, and h _t represents the current hidden layer state of the recurrent neural network, that is, the current output of the recurrent neural network; h _t-1 represents the state of the hidden layer obtained in the last time step of the recurrent neural network _; the parameters W _r , Ur , W _z , U _z , W, and U are the weight vectors learned in advance during the training process.

Wherein, according to the information obtained from the output of the recurrent neural network, monitoring whether the hard disk will be in a fault state specifically includes:

According to the hidden layer state output by the recurrent neural network, an attention distribution vector is generated; wherein, the attention distribution vector is used as the weight vector of the current hidden layer state, reflecting the current hidden layer state and the health state of the hard disk difference between;

By monitoring the size of the weight value in the weight vector, it is determined whether the hard disk will be in a fault state.

Further, before periodically acquiring different SMART attribute values of the monitored hard disk, it also includes:

The recurrent neural network is trained using SMART data from healthy and failed hard drives.

The present invention also provides a hard disk state monitoring device, comprising:

The feature integration module is used to periodically obtain the different SMART attribute values of the monitored hard disk; after obtaining the SMART attribute value of the hard disk each time, according to the SMART attribute value of the currently obtained hard disk, the attribute integration of the hard disk is obtained. ;

a time-dependent extraction module for integrating the attributes into a recurrent neural network introduced into a gated recurrent unit, and using the hidden layer state of the recurrent neural network as an output;

A fault information monitoring module, configured to monitor whether the hard disk will be in a fault state according to the information obtained from the output of the recurrent neural network.

Further, the device also includes:

A training module for training the recurrent neural network using the SMART data of the healthy hard disk and the faulty hard disk.

In the technical solution of the present invention, in order to capture the long-term time dependence in the SMART data, a gated recursive unit (GRU) is introduced on the basis of the existing simple RNN, which avoids the disappearance and explosion of gradients when processing long-term sequences. question. As a result, the deviation from the normal state of the hard disk drive can be traced back to an earlier period, so as to improve the failure detection rate or the failure prediction ability.

Further, in the technical solution of the present invention, an attention mechanism is also designed to generate an attention distribution vector for the hidden layer state output by the recurrent neural network to reflect the difference between the current hidden layer state and the health state of the hard disk. By analyzing the attention distribution, where a higher attention weight means a more important role, it is possible to gain a deep understanding of which days in the past have the greatest impact on the current state of the hard disk; so that the degradation process of the hard disk can be automatically revealed, there are Helps track down the cause of hard drive failures.

Description of drawings

1 is a flowchart of a method for monitoring a hard disk state according to an embodiment of the present invention;

2 is an internal connection structure diagram of a gated recursive unit provided by an embodiment of the present invention;

FIG. 3 is a schematic diagram of an internal structure of a hard disk state monitoring device according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

The following describes in detail the embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals refer to the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, but not to be construed as a limitation of the present invention.

It will be understood by those skilled in the art that the singular forms "a", "an", "the" and "the" as used herein can include the plural forms as well, unless expressly stated otherwise. It will further be understood that when we refer to an element as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Furthermore, "connected" or "coupled" as used herein may include wirelessly connected or wirelessly coupled. As used herein, the term "and/or" includes all or any element and all combination of one or more of the associated listed items.

It should be noted that all expressions using "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities with the same name but not the same or non-identical parameters. It can be seen that "first" and "second" It is only for the convenience of expression and should not be construed as a limitation to the embodiments of the present invention, and subsequent embodiments will not describe them one by one.

The inventor of the present invention observes and analyzes the change of the SMART attribute value of the hard disk drive, and finds that the deterioration process of the SMART attribute value can usually be traced back to 15 days or even 40 days. For example, for SMART_197_RAW, the change point usually occurs within 20 days before the failure; however, for SMART_7_RAW and SMART_242_RAW, it usually takes 40 days back to find the change point. This is clearly beyond the capabilities of Markov models or simple RNNs. High-performance predictive models require a method that can extract long-term temporal dependencies.

In order to solve the above problems, in the technical scheme of the present invention, in order to capture the long-term time dependence in the SMART data, a gated recursive unit (GRU) is introduced on the basis of the existing simple RNN, which avoids the occurrence of long-term sequences. The gradient vanishing and exploding problem. As a result, the deviation from the normal state of the hard disk drive can be traced back to an earlier period, so as to improve the failure detection rate or the failure prediction ability.

The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

A method for monitoring the status of a hard disk provided by the embodiment of the present invention can periodically monitor the failure status of the hard disk: periodically obtain different SMART attribute values of the monitored hard disk, for example, obtain different SMART attribute values of the monitored hard disk every day, Then, carry out the following operations after obtaining the SMART attribute value of the hard disk each time: monitor whether the hard disk will have a fault state according to the SMART attribute value obtained;

The specific method of monitoring the hard disk status after obtaining different SMART attribute values of the monitored hard disk at one time, the process is shown in Figure 1, including the following steps:

Step S101: Acquire different SMART attribute values of the monitored hard disk, and obtain an integrated representation of the current attributes of the hard disk.

Specifically, according to the currently acquired SMART attribute value of the hard disk, an integrated representation of the attributes of the hard disk is obtained; for example, the t-th day of the hard disk can be obtained according to the currently acquired different SMART attribute values of the monitored hard disk on the t day Attribute integrated representation of .

得到的属性集成表示为

可根据如下表达式一计算得到：For example, the SMART vector composed of the currently acquired SMART attribute values on day t

The resulting property ensemble is expressed as

It can be calculated according to the following expression:

v _t =ReLU(W _V s _t +b _v ) (Expression 1)

表示SMART属性值的权重矩阵，

是偏置向量。ReLU是定义为ReLU(x)＝x+＝max(0,x)的激活函数，其中，max是逐元素操作；W_V和b_v是预先在训练过程中进行学习得到的向量；

表示维度为d_s的实数向量组；

表示维度为d_v的实数向量组；d_s为SMART属性值个数、d_v为属性集成值个数。in,

weight matrix representing SMART attribute values,

is the bias vector. ReLU is an activation function defined as ReLU(x)=x+=max(0,x), where max is an element-wise operation; W _V and b _v are vectors learned in advance during the training process;

represents a real vector group of dimension d _s ;

Represents a real vector group with dimension d _v ; d _s is the number of SMART attribute values, and d _v is the number of attribute integration values.

Step S102: Integrate and input the attributes obtained above into the recurrent neural network introduced into the gated recurrent unit, and use the hidden layer state of the recurrent neural network as the output.

Specifically, the recurrent neural network in the embodiment of the present invention includes a plurality of gated recursive units, and when the attribute integration obtained above is input into the recurrent neural network, each vector element of the attribute integration is used as the enter.

The gated recursive unit in the recurrent neural network according to the embodiment of the present invention includes a gated unit and a recursive unit; wherein, in a gated recursive unit, the gated unit is used to control the information flow of the recursive unit, so that the recursive unit can capture long Time scale dependence; wherein a gating unit includes reset gates and update gates to allow the recursive unit to maintain existing content or update content based on existing content. Figure 2 shows the internal connection structure of the gated recursive unit.

The recurrent neural network (RNN) maintains a recursive hidden layer state, which is updated at each time step from the current input and previous hidden layer state, while the input v _t and output h _t of the recurrent neural network introducing the gated recurrent unit The relationship between can be achieved by a recursive algorithm of the following four expressions:

r _t =sigmoid(W _r v _t +U _r h _t-1 ) (Expression 2)

z _t =sigmoid(W _z v _t +U _z h _t-1 ) (Expression 3)

h _t '=tanh(Wv _t +U(r _t ⊙h _t-1 )) (Expression 4)

h _t =z _t ⊙h _t-1 +(1-z _t )⊙h' _t (Expression 5)

Among them, ⊙ is the element-wise multiplication operation; the parameters W _r , U _r , W _z , U _z , W, U are the weight vectors learned in advance during the training process; the sigmoid function can convert any real value to [0, 1]; the Tanh function converts any real value to the range [-1,1].

r _t represents the reset gate, h' _t represents the alternative state, z _t represents the update gate, h _t represents the current hidden layer state of the recurrent neural network (hidden layer state of the recurrent neural network on the t day), that is, the current state of the recurrent neural network The output of ; h _t-1 represents the hidden layer state obtained in the last time step of the recurrent neural network (the hidden layer state of the recurrent neural network on day t-1);

When the reset gate _rt is close to 0, the alternative state _h't can forget the previous hidden state and reset to the current input; the update gate _zt controls the hidden state from the previous time step _{ht- 1} and the amount of information flowing in the alternative state _h't .

Step S103: According to the information obtained from the output of the recurrent neural network, monitor whether the hard disk will be in a fault state.

In this step, the long-term information of the SMART data of the hard disk can be obtained from the output of the recurrent neural network, and according to the long-term information of the acquired SMART data, the deviation of the normal state of the hard disk drive can be traced back to an earlier period , so as to facilitate early monitoring of whether the hard disk is about to fail, thereby improving the failure detection rate or the failure prediction capability.

In addition, in this step, when monitoring whether the hard disk will fail according to the information obtained from the output of the recurrent neural network, a more optimal technical solution can be adopted: after obtaining the output of the recurrent neural network Based on the long-term information of the SMART data of the hard disk reflected by the hidden layer state, an attention mechanism is designed, which can automatically focus on the degradation process of the hard disk. This attention mechanism can reveal which information has the greatest impact on fault prediction, providing fault tracing diagnosis. It helps managers trace back to specific days and identify the cause of failures.

Specifically, an attention distribution vector can be generated according to the hidden layer state output by the recurrent neural network; wherein, the attention distribution vector is used as the weight vector of the current hidden layer state, reflecting the difference between the current hidden layer state and the The difference between the health states of the hard disks; by monitoring the size of the weight value in the weight vector, it is determined whether the hard disk will be in a fault state.

Specifically, the attention distribution vector a _t can be generated according to the hidden layer state h _t output by the recurrent neural network and according to the following expression 6:

是健康状态向量，可以被视作健康硬盘的特征的高阶表示；

可以预先在训练过程中进行学习得到。上式被用来比较健康状态向量与当前隐层状态之间的差异，并得到该差异的权重。Wherein, i is a natural number, and its summation range in Expression 6 is [t-T+1, t]; T is the size of the time window of the sequence input to the recurrent neural network. u _t is the hidden layer state h _t is transformed into a position-based representation through the tanh activation function, as shown in Expression 7 below;

is the health state vector, which can be regarded as a high-order representation of the characteristics of a healthy hard disk;

It can be learned in advance during the training process. The above formula is used to compare the difference between the healthy state vector and the current hidden layer state, and get the weight of this difference.

u _t =tanh(W _a h _t +b _a ) (Expression 7)

表示维度为d_r×d_r的实数矩阵，

表示维度为d_r的实数向量，均是预先在训练过程中进行学习得到的参数；d_r为所述递归神经网络的门控递归单元个数。in,

represents a real matrix of dimension d _r ×d _r ,

The real number vectors representing the dimension d _r are all parameters learned in advance in the training process; d _r is the number of gated recursive units of the recurrent neural network.

After obtaining the attention distribution vector a _t , the hidden layer state with attention weight can be obtained according to the following expression 8

With the help of the attention mechanism, the most informative part of the failure can be focused, so better evaluation and prediction can often be made. What's more, by analyzing the attention distribution, where higher attention weights mean more important roles, it is possible to gain insight into which days in the past have had the greatest impact on the current state of the hard drive; it can automatically reveal the degradation process of the hard drive , and help us track down the cause of a hard drive failure.

In fact, before step S101 is performed, that is, before the different SMART attribute values of the monitored hard disk are periodically acquired, a training process may be performed first. The training process includes using the SMART data of the healthy hard disk and the faulty hard disk to train the recurrent neural network, that is, the parameters in the above-mentioned recurrent neural network are obtained by training; the training process can also include learning other parameters:

W_a、b_a。而训练方法可以采用本领域技术人员所熟知的梯度下降法等，此处不赘述。Specifically, in the training process, the SMART data of the healthy hard disk and the faulty hard disk can be used for calculation and verification to determine the parameters W _r , _Ur , W _z , U _z , W, and U in the recurrent neural network. Of course, W _V and b _v can also be obtained at the same time during the training process, as well as the parameters in the attention mechanism

W _a , b _a . The training method can be a gradient descent method well known to those skilled in the art, etc., which will not be repeated here.

Based on the above method, a hard disk state monitoring device provided by an embodiment of the present invention has an internal structure as shown in FIG.

可根据如下表达式一计算得到的属性集成。Wherein, the feature integration module 301 is used to periodically acquire different SMART attribute values of the monitored hard disk; after each acquisition of the SMART attribute value of the hard disk, obtain the SMART attribute value of the hard disk according to the currently acquired SMART attribute value of the hard disk. Attribute integration; Specifically, the feature integration module 301 periodically obtains different SMART attribute values of the monitored hard disk; for the SMART vector formed by the SMART attribute values of the T-th day of the currently acquired hard disk

The attribute integration can be calculated according to the following expression 1.

The time-dependent extraction module 302 is used to integrate the attributes obtained by the feature integration module 301 into the recurrent neural network introduced into the gated recurrent unit, and use the hidden layer state of the recurrent neural network as the output; and the input and output of the recurrent neural network The relationship between them is specifically realized by the recursive algorithm of the above expressions 2, 3, 4, and 5.

The fault information monitoring module 303 is configured to monitor whether the hard disk will be in a fault state according to the information obtained from the output of the recurrent neural network. Specifically, the fault information monitoring module 303 generates an attention distribution vector according to the hidden layer state output by the recurrent neural network; wherein, the attention distribution vector is used as the weight vector of the current hidden layer state, reflecting the current hidden layer state. The difference between the state and the health state of the hard disk; by monitoring the size of the weight value in the weight vector, it is determined whether the hard disk will be in a fault state. The fault information monitoring module 303 can calculate the attention distribution vector according to the above expressions 6 and 7.

Further, the device for monitoring the state of a hard disk provided by the embodiment of the present invention may further include: a training module 304 .

The training module 304 is used to train the above-mentioned recurrent neural network using the SMART data of the healthy hard disk and the faulty hard disk, that is, use the SMART data of the healthy hard disk and the faulty hard disk to train the above-mentioned recurrent neural network, and determine the parameters W _r , _{Ur, Wz, Uz , W} , U.

W_a、b_a。The training module 304 can also train to obtain parameters W _v and b _v , and parameters in the attention mechanism while training the recurrent neural network.

W _a , b _a .

In the technical solution of the present invention, in order to capture the long-term time dependence in the SMART data, a gated recursive unit (GRU) is introduced on the basis of the existing simple RNN, which avoids the disappearance and explosion of gradients when processing long-term sequences. question. As a result, the deviation from the normal state of the hard disk drive can be traced back to an earlier period, so as to improve the failure detection rate or the failure prediction ability.

Further, in the technical solution of the present invention, an attention mechanism is also designed to generate an attention distribution vector for the hidden layer state output by the recurrent neural network to reflect the difference between the current hidden layer state and the health state of the hard disk. By analyzing the attention distribution, where a higher attention weight means a more important role, it is possible to gain a deep understanding of which days in the past have the greatest impact on the current state of the hard disk; thus, the degradation process of the hard disk can be automatically revealed. Helps track down the cause of hard drive failures.

As will be appreciated by those skilled in the art, the present invention includes apparatuses for performing one or more of the operations described in this application. These devices may be specially designed and manufactured for the required purposes, or they may include those known in general purpose computers. These devices have computer programs stored in them that are selectively activated or reconfigured. Such a computer program may be stored in a device (eg, computer) readable medium including, but not limited to, any type of medium suitable for storing electronic instructions and coupled to a bus, respectively Types of disks (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROM (Read-Only Memory, read-only memory), RAM (Random Access Memory, random access memory), EPROM (Erasable Programmable Read-Only Memory, Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory, Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic card or optical card. That is, a readable medium includes any medium that stores or transmits information in a form that can be read by a device (eg, a computer).

Those skilled in the art will understand that computer program instructions can be used to implement each block of these structural diagrams and/or block diagrams and/or flow diagrams, and combinations of blocks in these structural diagrams and/or block diagrams and/or flow diagrams . Those skilled in the art can understand that these computer program instructions can be provided to a general-purpose computer, a professional computer or a processor of other programmable data processing methods to implement, so that the present invention can be executed by a processor of a computer or other programmable data processing method. The block or blocks specified in the block or blocks of the block diagrams and/or block diagrams and/or flow diagrams of the invention are disclosed.

Those skilled in the art can understand that the various operations, methods, steps, measures, and solutions discussed in the present invention may be alternated, modified, combined or deleted. Further, other steps, measures, and solutions in the various operations, methods, and processes that have been discussed in the present invention may also be alternated, modified, rearranged, decomposed, combined, or deleted. Further, steps, measures and solutions in the prior art with various operations, methods, and processes disclosed in the present invention may also be alternated, modified, rearranged, decomposed, combined or deleted.

Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary, and is not intended to imply that the scope of the present disclosure (including the claims) is limited to these examples; under the spirit of the present invention, the above embodiments or There may also be combinations between technical features in different embodiments, steps may be carried out in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity. Therefore, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (8)

1. A hard disk state monitoring method comprises the following steps:

periodically acquiring different automatic detection analysis and reporting technology SMART attribute values of the monitored hard disk; after the SMART attribute value of the hard disk is obtained each time, the following operations are carried out:

obtaining attribute integration of the hard disk according to the SMART attribute value of the hard disk which is obtained currently;

integrating and inputting the attribute into a recurrent neural network with a gated recurrent unit introduced, and taking the hidden layer state of the recurrent neural network as output;

monitoring whether the hard disk will fail according to information obtained from the output of the recurrent neural network: generating an attention distribution vector according to the hidden layer state output by the recurrent neural network; wherein, the attention distribution vector is used as a weight vector of the current hidden layer state, and reflects the difference between the current hidden layer state and the health state of the hard disk; and determining whether the hard disk is in a fault state or not by monitoring the weight value in the weight vector.

2. The method according to claim 1, wherein the obtaining of the attribute integration of the hard disk according to the SMART attribute value of the hard disk currently obtained specifically comprises:

according to the SMART vector composed of the currently acquired SMART attribute values of the t day of the hard disk

The resulting attribute integration is represented as

The calculation is carried out according to the following expression I:

v_t＝ReLU(W_Vs_t+b_v) (expression one)

Wherein,

a weight matrix representing SMART attribute values,

is a bias vector; ReLU is an activation function defined as ReLU (x) ═ x + ═ max (0, x), where max is an element-by-element operation; w_VAnd b_vIs a vector obtained by learning in the training process in advance;

the representation dimension is d_sThe set of real vectors of (2);

the representation dimension is d_vThe set of real vectors of (2); d_sFor the number of SMART attribute values, d_vThe number of values is integrated for the attribute.

3. The method of claim 2, wherein the gated recursion units in the recurrent neural network comprise a gated unit and a recursion unit; wherein,

the gate control unit is used for controlling the information flow of the recursion unit so that the recursion unit captures the dependence of a long time scale; wherein,

the gating cell includes a reset gate and an update gate to allow the recursion cell to hold existing content or update content on an existing content basis.

4. The method according to claim 3, characterized in that the relation between the inputs and outputs of the recurrent neural network is realized in particular by a recurrent algorithm of the following four expressions:

r_t＝sigmoid(W_rv_t+U_rh_t-1) (expression two)

z_t＝sigmoid(W_zv_t+U_zh_t-1) (expression III)

h_t'＝tanh(Wv_t+U(r_t⊙h_t-1) (expression four)

h_t＝z_t⊙h_t-1+(1-z_t)⊙h'_t(expression five)

Wherein, an element-by-element multiplication operation; r is_tDenotes a reset gate, h'_tRepresenting alternative states, z_tRepresents an update gate, h_tRepresenting the current hidden layer state of the recurrent neural network, namely the current output of the recurrent neural network; h is_t-1Representing the hidden layer state obtained in the last time step of the recurrent neural network; parameter W_r、U_r、W_z、U_zW, U are learned in advance during training.

5. The method of any of claims 1-4, further comprising, prior to periodically obtaining different SMART attribute values for the monitored hard disk:

training the recurrent neural network using SMART data of healthy and failed hard disks.

6. A hard disk state monitoring device, comprising:

the characteristic integration module is used for periodically acquiring different SMART attribute values of the monitored hard disk; after the SMART attribute value of the hard disk is obtained every time, obtaining the attribute integration of the hard disk according to the currently obtained SMART attribute value of the hard disk;

the time dependence extraction module is used for integrating and inputting the attribute into a recurrent neural network with a gating recurrent unit and taking the hidden layer state of the recurrent neural network as output;

a fault information monitoring module for monitoring whether the hard disk will have a fault state according to the information obtained from the output of the recurrent neural network: generating an attention distribution vector according to the hidden layer state output by the recurrent neural network; wherein, the attention distribution vector is used as a weight vector of the current hidden layer state, and reflects the difference between the current hidden layer state and the health state of the hard disk; and determining whether the hard disk is in a fault state or not by monitoring the weight value in the weight vector.

7. The apparatus of claim 6, wherein the gated recursion units in the recurrent neural network comprise a gating unit and a recursion unit; wherein,

gating a gate unit in a recursive unit to control the flow of information for the recursive unit in the gated recursive unit such that the recursive unit captures long time scale dependencies; wherein,

the gating cell includes a reset gate and an update gate to allow the recursion cell to hold existing content or update content on an existing content basis.

8. The apparatus of any of claims 6-7, further comprising:

and the training module is used for training the recurrent neural network by using SMART data of the healthy hard disk and the fault hard disk.

Images