CN112488235A - Elevator time sequence data abnormity diagnosis method based on deep learning - Google Patents

Elevator time sequence data abnormity diagnosis method based on deep learning Download PDF

Info

Publication number
CN112488235A
CN112488235A CN202011459448.1A CN202011459448A CN112488235A CN 112488235 A CN112488235 A CN 112488235A CN 202011459448 A CN202011459448 A CN 202011459448A CN 112488235 A CN112488235 A CN 112488235A
Authority
CN
China
Prior art keywords
sample
encoder
data
elevator
time
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011459448.1A
Other languages
Chinese (zh)
Inventor
陈华
曾亚辉
方建文
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Special Equipment Safety Supervision Inspection Institute of Jiangsu Province
Original Assignee
Special Equipment Safety Supervision Inspection Institute of Jiangsu Province
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Special Equipment Safety Supervision Inspection Institute of Jiangsu Province filed Critical Special Equipment Safety Supervision Inspection Institute of Jiangsu Province
Priority to CN202011459448.1A priority Critical patent/CN112488235A/en
Publication of CN112488235A publication Critical patent/CN112488235A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B66—HOISTING; LIFTING; HAULING
    • B66B—ELEVATORS; ESCALATORS OR MOVING WALKWAYS
    • B66B5/00—Applications of checking, fault-correcting, or safety devices in elevators
    • B66B5/0006—Monitoring devices or performance analysers
    • B66B5/0018—Devices monitoring the operating condition of the elevator system
    • B66B5/0031—Devices monitoring the operating condition of the elevator system for safety reasons
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B66—HOISTING; LIFTING; HAULING
    • B66B—ELEVATORS; ESCALATORS OR MOVING WALKWAYS
    • B66B5/00—Applications of checking, fault-correcting, or safety devices in elevators
    • B66B5/02—Applications of checking, fault-correcting, or safety devices in elevators responsive to abnormal operating conditions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Elevator Control (AREA)

Abstract

The invention provides an elevator time sequence data abnormity diagnosis method based on deep learning, which belongs to the field of special equipment operation process soft measurement modeling and application, and mainly comprises the following steps: determining auxiliary variables required by elevator soft measurement modeling based on the data, carrying out normalization processing on the auxiliary variables, and dividing the elevator data into time sequence data with fixed length by adopting a sliding window; carrying out trend prediction on the time sequence data based on a long-time memory network of an attention mechanism, and predicting a state value at the next moment; and inputting the state value into a normal range area reconstructed by a variational encoder, and judging whether the state value is in the normal range area or not to obtain the abnormal condition of the data. The method can improve the precision of the abnormity diagnosis in the elevator running process, can provide the effect of real-time detection, and can be effectively applied to the field of fault diagnosis of special detection equipment.

Description

Elevator time sequence data abnormity diagnosis method based on deep learning
Technical Field
The invention belongs to the field of soft measurement modeling and application of a special device in an operation process, and particularly relates to an abnormity diagnosis method of elevator time sequence data based on deep learning.
Background
With the steady advance of the modernization process, special equipment such as elevators, hoisting machinery, passenger ropeways and the like are widely applied, and the convenience and the efficiency of production and life of people are greatly improved due to the appearance of the special equipment. Therefore, special equipment occupies a very important position in human life, and is closely related to human life. However, with the popularization of special equipment, the safety problem of the special equipment is also greatly concerned, because the safety of human life and property is directly endangered once the special equipment fails, and the result is not imaginable. Aiming at the safety problem of the special equipment, the measures are taken to periodically inspect the special equipment and evaluate the running condition and the running life of the special equipment. However, this measure has four problems: 1. firstly, the professional requirement on quality inspectors is high, and the quality inspectors need a large amount of special equipment fault diagnosis experience, so that the evaluation result has reliability; 2. the system structure of the special equipment is complex, so that the fault caused by the complex system structure has too much uncertainty and cannot be accurately evaluated; 3. the probability of the equipment abnormality is low, the occurrence time is not fixed, and the equipment cannot be effectively controlled through regular inspection; 4. the regular inspection has serious hysteresis, and special equipment cannot be monitored in real time so as to ensure the life and property safety of people.
With the arrival of the artificial intelligence era, a plurality of manual operations are gradually replaced by intelligent equipment, and the intelligent equipment is used for monitoring a specific device in real time, so that the detection efficiency is higher, and the cost is low; and the artificial intelligence technology is mature and perfect, so that a plurality of effective achievements are obtained in the aspect of abnormity diagnosis, and the development of the abnormity diagnosis technology is greatly promoted. The abnormal diagnosis technology based on artificial intelligence mainly utilizes a sensor to collect data of key positions of special equipment, and determines the running condition and the running life of special equipment by observing the change of overall data and combining an expert diagnosis system. According to the method, the safety condition of the elevator can be effectively evaluated only by recording data display of the sensor at different moments in real time and transmitting the data to a computer of a user for data analysis through the Internet of things technology. Once a special device is easy to generate a large or small abnormality, the phenomena of waveform jumping, faults and the like of a sensor on the special device can be caused to form abnormal data. At the moment, the abnormal condition of the special equipment in the operation process can be diagnosed only by effectively analyzing the real-time data generated by the special equipment to obtain the abnormal points in the data, so that early warning measures are given in real time.
Disclosure of Invention
The purpose of the invention is as follows: the invention provides an elevator time sequence data abnormity diagnosis method based on deep learning, which aims at the problem that the traditional abnormity diagnosis method of special equipment needs to spend huge manpower and material resources to carry out manual inspection, adopts a data driving mode to carry out comprehensive evaluation on the running condition of the special equipment, and provides an elevator time sequence data abnormity diagnosis method based on deep learning.
The technical scheme is as follows: an abnormity diagnosis method of elevator time sequence data based on deep learning comprises the following steps:
step 1, collecting data, wherein the time sequence data of the elevator is obtained by sampling a door ball bearing, a humidity sensor and a vibration sensor on a key position on the elevator and is stored in a time sequence form;
step 2, preprocessing the time sequence data to obtain a modeling sample, and dividing the modeling sample into a training sample and a test sample;
step 3, establishing a prediction model of the long and short memory networks based on the attention mechanism by adopting the training samples, predicting by adopting the test samples, and predicting the state value at the next moment;
step 4, inputting the training sample into a variational encoder to obtain a target optimal loss value alpha;
step 5, inputting the state value into a range area reconstructed by a variational encoder, and determining the upper limit and the lower limit of a normal range interval;
step 6, collecting timing sequence data measured by a sensor on the elevator in real time on line, and preprocessing the timing sequence data;
and 7, inputting the preprocessed new sample into the prediction model of the long and short memory network based on the attention mechanism in the step 3 to obtain a corresponding state value, then inputting the state value into the step 5, judging whether the state value is at the upper limit and the lower limit of the normal range interval, and judging the abnormal condition of the data in real time.
In the step 2, the specific preprocessing process of preprocessing the time series data to obtain the modeling sample is as follows:
Figure BDA0002830868830000021
wherein the time-series data D { (x'1，y1)，(x′2，y2)，…，(x′m，ym)}，
Figure BDA0002830868830000022
The representation of a single piece of data,
Figure BDA0002830868830000023
representing the d attribute value in the ith data, preprocessing the time sequence data to obtain a modeling sample, and representing as follows: d { (x)1，y1)，(x2，y2)，…，(xm，ym)}。
In the step 3, establishing a prediction model of the long and short memory network based on the attention mechanism is divided into an encoding stage and a decoding stage. The encoder stage is implemented as follows:
step 1, equally dividing training samples into time sequence data with n batches, and inputting current time t based on the input of a long and short memory network of an attention mechanism
Figure BDA0002830868830000024
Step 2, given time t, the implicit state of the last time is expressed as
Figure BDA0002830868830000025
Outputting the activation function Sigmoid (. sigma.) is calculated together with the fully connected layer, and the Sigmoid (. sigma.) function is defined by the following equation:
Figure BDA0002830868830000031
step 3, setting the number of the hidden units as h, and inputting X in small batches at given time ttAnd a last hidden state Ht-1Input gate of time t
Figure BDA0002830868830000032
Forgetting door
Figure BDA0002830868830000033
And output gate
Figure BDA0002830868830000034
Respectively as follows:
It＝σ(XtWxi+Ht-1Whi+bi)
Ft＝σ(XtWxf+Ht-1Whf+bf)
Ot＝σ(XtWxo+Ht-1Who+bo)
Wxi，Wxf，
Figure BDA0002830868830000035
and Whi，Whf，
Figure BDA0002830868830000036
Is a weight parameter, bi，bf，
Figure BDA0002830868830000037
Is a deviation parameter;
step 4, candidate memory cells
Figure BDA0002830868830000038
Using Tanh (-) functionAs an activation function, the Tanh (-) function is defined as:
Figure BDA0002830868830000039
step 5, CtIs calculated as:
Figure BDA00028308688300000310
wherein the content of the first and second substances,
Figure BDA00028308688300000311
and
Figure BDA00028308688300000312
is a weight parameter that is a function of,
Figure BDA00028308688300000313
is a deviation parameter;
step 6, memory cell CtThe memory cell C of the current time, information in the implicit state, is obtained by controlling the input gate, the forgetting gate and the output gate using the multiplication by element, symbol-tThe formula of (1) is:
Figure BDA00028308688300000314
step 7, controlling the memory cell to the hidden state H through the output gatetThe expression of (a) is:
Ht＝Ot⊙Tanh(Ct)；
the decoder stage is implemented as follows:
step 1, for the index t of the encoder, the implicit state of the decoder time t' can be represented by the implicit state H of the previous momentt′-1By way of representation, the current implicit state may be represented as:
Ht′＝g(ct′，Ht′-1)
wherein g is a matrix multiplication operation;
step 2, weighted average is carried out on all encoder hidden states of the decoder in the background of time t':
Figure BDA0002830868830000041
step 3, obtaining probability distribution by using softmax operation:
Figure BDA0002830868830000042
step 4, for et′tThe implicit variables of time t and t' are input by the f-function:
ett＝f(Ht′-1，Ht)
wherein the f function is inner product operation, and perception transformation is carried out through an activation function tanh:
f(s，h)＝tanh(WsHt′-1+WhHt)vT
wherein
Figure BDA0002830868830000043
All the model parameters can be learned, and all the matrix parameters are initialized through Gaussian distribution;
step 5, the function of the final training error is:
Figure BDA0002830868830000044
the method comprises the following specific steps of:
step 1, initializing encoder H0，Wxi，Wxf，Wxo、Whi，Whf，Who、bi，bf，boWxc、Whc、bc、v，Ws，WhThe matrix which satisfies corresponding dimensionality and has a numerical value of 0 to 1, wherein the sample batch size n is 8, T is m/n, and epoach of neural network training is 1000;
step 2, setting epoach to be 1;
step 3, setting t to be 1;
step 4, letting Ht-1＝H0And sample XtSubstituting the formula in the step 2 to obtain H1；
Step 5, changing T to T +1, and turning to step 64 when T is less than or equal to T, or turning to step 66;
step 6, obtaining HTLet t' be 1;
step 7, entering a decoder part and enabling Ht′-1＝HTIs substituted into Ht′＝g(ct′，Ht′-1) To obtain Ht′Then make it possible to
Figure BDA0002830868830000051
Wherein
Figure BDA0002830868830000052
Step 8, changing T 'to T' +1, and turning to step 67 when T is less than or equal to T, or turning to step 69;
step 9, finally according to
Figure BDA0002830868830000053
Obtaining an error value error, and training an initialization parameter through back propagation; the parameter iterative optimization is carried out by adopting a random gradient descent algorithm, and the learning rate lr is 0.0003;
step 10, when the epoach is equal to or less than 1000, turning to step 63, otherwise, finishing training, establishing a long and short memory network prediction model based on an attention mechanism, and storing the model;
step 11, placing a test sample into the trained model for prediction, and predicting a state value at the next moment;
in the step 4, the variational encoder comprises two parts: the encoder and the decoder are all full-connection layer neural networks, the encoder uses a four-layer full-connection network, the input dimensionality is d, the output dimensionality is 4, and the dimensionalities of two middle hidden layers are 8 and 6 respectively; the dimension of the decoder is opposite to that of the encoder, the input and output dimensions are respectively 4 and d, and the dimension of the middle hidden layer is 6 and 8; the encoder mainly maps the extracted data features into low-dimensional potential feature vectors, and the decoder decodes and reconstructs the low-dimensional potential feature vectors to be as close as possible to the original data features. The method comprises the following concrete steps:
step 1, k represents that the original sample dimension is 12, subscript j represents the dimension of the potential feature vector, namely output dimension 4, the encoder part of the training sample encoder obtains the low-dimensional potential feature vector, and the sample is
Figure BDA0002830868830000054
d is 12, then the output is
Figure BDA0002830868830000055
Obtaining the latent variable zi：
Figure BDA0002830868830000056
Figure BDA0002830868830000057
Figure BDA0002830868830000058
Where randn _ like () denotes the return of a random number satisfying a normal distribution, and then the latent variable ziSample data x' with the same dimension as the training sample is restored through a decoderi；
Step 2, finding the correct distribution of the latent variables by using the lower bound of the marginal likelihood estimation of the input data as a target loss function:
Figure BDA0002830868830000061
comparing errors of the output sample data and the training sample, calculating a target loss value through the formula, wherein the target to be achieved is to make the difference between the sample data and the training sample as small as possible, and when a variational encoder is trained, a random gradient descent algorithm is used for training, the initialization parameter of a neural network is optimized, and the obtained target loss value is alpha;
in the step 5, the state value is input into the range region reconstructed by the variational encoder, and the upper and lower limits of the normal range region are determined, wherein the specific process is as follows:
step 1, inputting a test sample, and obtaining a state value of the test sample through model training based on a reconstruction probability threshold value alpha obtained by a training sample;
and 2, inputting the state values into a variational encoder model for carrying out abnormity diagnosis, and recording the maximum value and the minimum value of the state values of the time series data with the target loss value lower than the target loss value to obtain [ delta max, delta min ], thereby obtaining the upper limit and the lower limit [ delta min, delta max ] of the normal range interval.
Has the advantages that: according to the invention, the abnormal condition of the special equipment in the operation process is diagnosed by analyzing the real-time data generated by the special equipment, so that the special equipment is effectively monitored, and the detection hysteresis is avoided; the method is provided for monitoring the fault of the special equipment in real time in the operation process.
Drawings
Fig. 1 is an overall architecture diagram of an abnormality diagnosis method for elevator time series data based on deep learning.
Fig. 2 is a prediction result after elevator time series data soft measurement modeling based on deep learning.
Fig. 3 shows an abnormality diagnosis effect of the abnormality diagnosis method for elevator time series data based on deep learning.
Detailed Description
The following further describes the embodiments of the present invention with reference to the drawings. Referring to fig. 1, an abnormality diagnosis method for elevator time series data based on deep learning includes the following steps:
step 1, collecting data, wherein the time sequence data of the elevator is obtained by sampling a door ball bearing, a humidity sensor and a vibration sensor on a key position on the elevator and is stored in a time sequence form;
step 2, preprocessing the time sequence data to obtain a modeling sample, and dividing the modeling sample into a training sample and a test sample;
step 3, establishing a prediction model of the long and short memory networks based on the attention mechanism by adopting the training samples, predicting by adopting the test samples, and predicting the state value at the next moment;
step 4, inputting the training sample into a variational encoder to obtain a target optimal loss value alpha;
step 5, inputting the state value into a range area reconstructed by a variational encoder, and determining the upper limit and the lower limit of a normal range interval;
step 6, collecting timing sequence data measured by a sensor on the elevator in real time on line, and preprocessing the timing sequence data;
and 7, inputting the preprocessed new sample into the prediction model of the long and short memory network based on the attention mechanism in the step 3 to obtain a corresponding state value, then inputting the state value into the step 5, judging whether the state value is at the upper limit and the lower limit of the normal range interval, and judging the abnormal condition of the data in real time.
In a further embodiment, the specific preprocessing process of preprocessing the time series data in step 2 to obtain the modeling sample is as follows:
Figure BDA0002830868830000071
wherein the time-series data D { (x'1，y1)，(x′2，y2)，…，(x′m，ym)}，
Figure BDA0002830868830000072
Representing a single numberAccording to the above-mentioned technical scheme,
Figure BDA0002830868830000073
representing the d attribute value in the ith data, preprocessing the time sequence data to obtain a modeling sample, and representing as follows: d { (x)1，y1)，(x2，y2)，…，(xm，ym)}。
In a further embodiment, the step 3 of establishing the prediction model of the long and short memory network based on the attention mechanism is divided into an encoding stage and a decoding stage. The encoder stage is implemented as follows:
step 1, equally dividing training samples into time sequence data with n batches, and inputting current time t based on the input of a long and short memory network of an attention mechanism
Figure BDA0002830868830000074
Step 2, given time t, the implicit state of the last time is expressed as
Figure BDA0002830868830000075
The output is obtained by jointly calculating an activation function Sigmoid (·) function (sigma) and a full connection layer, wherein the Sigmoid (·) function is defined by the following formula:
Figure BDA0002830868830000076
step 3, setting the number of the hidden units as h, and inputting X in small batches at given time ttAnd a last hidden state Ht-1Input gate of time t
Figure BDA0002830868830000077
Forgetting door
Figure BDA0002830868830000078
And output gate
Figure BDA0002830868830000079
Respectively as follows:
It＝σ(XtWxi+Ht-1Whi+bi)
Ft＝σ(XtWxf+Ht-1Whf+bf)
Ot＝σ(XtWxo+Ht-1Who+bo)
Wxi，Wxf，
Figure BDA0002830868830000081
and Whi，Whf，
Figure BDA0002830868830000082
Is a weight parameter, bi，bf，
Figure BDA0002830868830000083
Is a deviation parameter;
step 4, candidate memory cells
Figure BDA0002830868830000084
Using a Tanh (-) function as the activation function, the Tanh (-) function is defined as:
Figure BDA0002830868830000085
step 5, CtIs calculated as:
Figure BDA0002830868830000086
wherein the content of the first and second substances,
Figure BDA0002830868830000087
and
Figure BDA0002830868830000088
is a weight parameter that is a function of,
Figure BDA0002830868830000089
is a deviation parameter;
step 6, memory cell CtThe memory cell C of the current time, information in the implicit state, is obtained by controlling the input gate, the forgetting gate and the output gate using the multiplication by element, symbol-tThe formula of (1) is:
Figure BDA00028308688300000810
step 7, controlling the memory cell to the hidden state H through the output gatetThe expression of (a) is:
Ht＝Ot⊙Tanh(Ct)；
the decoder stage is implemented as follows:
step 1, for the index t of the encoder, the implicit state of the decoder time t' can be represented by the implicit state H of the previous momentt′-1By way of representation, the current implicit state may be represented as:
Ht′＝g(ct′，Ht′-1)
wherein g is a matrix multiplication operation;
step 2, weighted average is carried out on all encoder hidden states of the decoder in the background of time t':
Figure BDA00028308688300000811
step 3, obtaining probability distribution by using softmax operation:
Figure BDA0002830868830000091
step 4, for et′tThe implicit variables of time t and t' are input by the f-function:
ett＝f(Ht′-1，Ht)
wherein the f function is inner product operation, and perception transformation is carried out through an activation function tanh:
f(s，h)＝tanh(WsHt′-1+WhHt)vT
wherein
Figure BDA0002830868830000092
All the model parameters can be learned, and all the matrix parameters are initialized through Gaussian distribution;
step 5, the function of the final training error is:
Figure BDA0002830868830000093
the method comprises the following specific steps of:
step 1, initializing encoder H0，Wxi，Wxf，Wxo、Whi，Whf，Who、bi，bf，boWxc、Whc、bc、v，Ws，WhThe matrix which satisfies corresponding dimensionality and has a numerical value of 0 to 1, wherein the sample batch size n is 8, T is m/n, and epoach of neural network training is 1000;
step 2, setting epoach to be 1;
step 3, setting t to be 1;
step 4, letting Ht-1＝H0And sample XtSubstituting the formula in the step 2 to obtain H1；
Step 5, changing T to T +1, and turning to step 64 when T is less than or equal to T, or turning to step 66;
step 6, obtaining HTLet t' be 1:
step 7, entering a decoder part and enabling Ht′-1＝HTIs substituted into Ht′＝g(ct′，Ht′-1) To obtain Ht′Then make it possible to
Figure BDA0002830868830000094
Wherein
Figure BDA0002830868830000095
Step 8, changing T 'to T' +1, and turning to step 67 when T is less than or equal to T, or turning to step 69;
step 9, finally according to
Figure BDA0002830868830000101
Obtaining an error value error, and training an initialization parameter through back propagation; the parameter iterative optimization is carried out by adopting a random gradient descent algorithm, and the learning rate lr is 0.0003;
step 10, when the epoach is equal to or less than 1000, turning to step 63, otherwise, finishing training, establishing a long and short memory network prediction model based on an attention mechanism, and storing the model;
step 11, placing a test sample into the trained model for prediction, and predicting a state value at the next moment;
in a further embodiment, in the step 4, the variational encoder comprises two parts: the encoder and the decoder are all full-connection layer neural networks, the encoder uses a four-layer full-connection network, the input dimensionality is d, the output dimensionality is 4, and the dimensionalities of two middle hidden layers are 8 and 6 respectively; the dimension of the decoder is opposite to that of the encoder, the input and output dimensions are respectively 4 and d, and the dimension of the middle hidden layer is 6 and 8; the encoder mainly maps the extracted data features into low-dimensional potential feature vectors, and the decoder decodes and reconstructs the low-dimensional potential feature vectors to be as close as possible to the original data features. The method comprises the following concrete steps:
step 1, k represents that the original sample dimension is 12, subscript j represents the dimension of the potential feature vector, namely output dimension 4, the encoder part of the training sample encoder obtains the low-dimensional potential feature vector, and the sample is
Figure BDA0002830868830000102
d is 12, then the output is
Figure BDA0002830868830000103
Obtaining the latent variable zi：
Figure BDA0002830868830000104
Figure BDA0002830868830000105
Figure BDA0002830868830000106
Where randn _ like () denotes the return of a random number satisfying a normal distribution, and then the latent variable ziSample data x' with the same dimension as the training sample is restored through a decoderi；
Step 2, finding the correct distribution of the latent variables by using the lower bound of the marginal likelihood estimation of the input data as a target loss function:
Figure BDA0002830868830000107
comparing errors of the output sample data and the training sample, calculating a target loss value through the formula, wherein the target to be achieved is to make the difference between the sample data and the training sample as small as possible, and when a variational encoder is trained, a random gradient descent algorithm is used for training, the initialization parameter of a neural network is optimized, and the obtained target loss value is alpha;
in a further embodiment, in step 5, the state value is input into the range region reconstructed by the variational encoder, and the upper and lower limits of the normal range interval are determined, where the specific process is as follows:
step 1, inputting a test sample, and obtaining a state value of the test sample through model training based on a reconstruction probability threshold value alpha obtained by a training sample;
and 2, inputting the state values into a variational encoder model for carrying out abnormity diagnosis, and recording the maximum value and the minimum value of the state values of the time series data with the target loss value lower than the target loss value to obtain [ delta max, delta min ], thereby obtaining the upper limit and the lower limit [ delta min, delta max ] of the normal range interval.
Although the preferred embodiments of the present invention have been described in detail, the present invention is not limited to the details of the embodiments, and various equivalent modifications can be made within the technical spirit of the present invention, and the scope of the present invention is also within the scope of the present invention.

Claims (9)

1. An abnormity diagnosis method of elevator time sequence data based on deep learning is characterized by comprising the following steps:
step 1, collecting data, wherein the time sequence data of the elevator is obtained by sampling a door ball bearing, a humidity sensor and a vibration sensor on a key position on the elevator and is stored in a time sequence form;
step 2, preprocessing the time sequence data to obtain a modeling sample, and dividing the modeling sample into a training sample and a test sample;
step 3, establishing a prediction model of the long and short memory networks based on the attention mechanism by adopting the training samples, predicting by adopting the test samples, and predicting the state value at the next moment;
step 4, inputting the training sample into a variational encoder to obtain a target optimal loss value alpha;
step 5, inputting the state value into a range area reconstructed by a variational encoder, and determining the upper limit and the lower limit of a normal range interval;
step 6, collecting timing sequence data measured by a sensor on the elevator in real time on line, and preprocessing the timing sequence data;
and 7, inputting the preprocessed new sample into the prediction model of the long and short memory network based on the attention mechanism in the step 3 to obtain a corresponding state value, then inputting the state value into the step 5, judging whether the state value is at the upper limit and the lower limit of the normal range interval, and judging the abnormal condition of the data in real time.
2. The method for diagnosing the abnormality of the elevator time series data based on the deep learning of claim 1, wherein in the step 2, the specific preprocessing process for preprocessing the time series data to obtain the modeling sample comprises:
Figure FDA0002830868820000011
wherein the time-series data D { (x'1,y1),(x′2,y2),…,(x′m,ym)}，
Figure FDA0002830868820000012
The representation of a single piece of data,
Figure FDA0002830868820000013
representing the d attribute value in the ith data, preprocessing the time sequence data to obtain a modeling sample, and representing as follows: d { (x)1,y1),(x2,y2),…,(xm,ym)}。
3. The method for diagnosing the abnormality of the elevator timing data based on the deep learning of claim 1, wherein in the step 3, the prediction model of the long and short memory networks based on the attention mechanism is established and divided into an encoding stage and a decoding stage.
4. The method for diagnosing the abnormality of the elevator time series data based on the deep learning as claimed in claim 3, wherein the encoder stage is implemented by the following steps:
step 41, equally dividing the training sample into n batches of time sequence data, and inputting the current time t based on the input of the long and short memory network of the attention mechanism
Figure FDA0002830868820000021
Step 42, given time t, the implicit state of the last time is represented as
Figure FDA0002830868820000022
The output is obtained by jointly calculating an activation function Sigmoid (·) function (sigma) and a full connection layer, wherein the Sigmoid (·) function is defined by the following formula:
Figure FDA0002830868820000023
step 43, setting the number of the hidden units as h, and inputting X in small batches at given time ttAnd a last hidden state Ht-1Input gate of time t
Figure FDA0002830868820000024
Forgetting door
Figure FDA0002830868820000025
And output gate
Figure FDA0002830868820000026
Respectively as follows:
It＝σ(XtWxi+Ht-1Whi+bi)
Ft＝σ(XtWxf+Ht-1Whf+bf)
Ot＝σ(XtWxo+Ht-1Who+bo)
Figure FDA0002830868820000027
and
Figure FDA0002830868820000028
is a weight parameter that is a function of,
Figure FDA0002830868820000029
is a deviation parameter;
step 44, candidate memory cells
Figure FDA00028308688200000210
Using a Tanh (-) function as the activation function, the Tanh (-) function is defined as:
Figure FDA00028308688200000211
step 45, CtIs calculated as:
Figure FDA00028308688200000212
wherein the content of the first and second substances,
Figure FDA00028308688200000213
and
Figure FDA00028308688200000214
is a weight parameter that is a function of,
Figure FDA00028308688200000215
is a deviation parameter;
step 46, memory cell CtThe memory cell C of the current time, information in the implicit state, is obtained by controlling the input gate, the forgetting gate and the output gate using the multiplication by element, symbol-tThe formula of (1) is:
Figure FDA00028308688200000216
step 47, controlling the memory cell to the hidden state H by the output gatetThe expression of (a) is:
Ht＝Ot⊙Tanh(Ct)。
5. the method for diagnosing the abnormality of the elevator time series data based on the deep learning as claimed in claim 3, wherein the decoder stage is implemented by the following steps:
step 51, for the index t of the encoder, the implicit state of the decoder time t' can be represented by the implicit state H of the previous momentt′-1By way of representation, the current implicit state may be represented as:
Ht′＝g(ct′,Ht′-1)
wherein g is a matrix multiplication operation;
step 52, weighted average is performed on all encoder implicit states in the background of the decoder at time t':
Figure FDA0002830868820000031
and 53, obtaining probability distribution by using softmax operation:
Figure FDA0002830868820000032
step 54, for et′tThe implicit variables of time t and t' are input by the f-function:
ett＝f(Ηt′-1,Ht)
wherein the f function is inner product operation, and perception transformation is carried out through an activation function tanh:
f(s,h)＝tanh(WsΗt′-1+WhHt)vT
wherein
Figure FDA0002830868820000033
All the model parameters can be learned, and all the matrix parameters are initialized through Gaussian distribution;
step 55, the function of the last training error is:
Figure FDA0002830868820000034
6. the method for diagnosing the abnormality of the elevator time series data based on the deep learning as claimed in claim 3, wherein the prediction model of the long and short memory network based on the attention mechanism comprises the following specific steps:
step 61, initialize encoder H0，Wxi,Wxf,Wxo、Whi,Whf,Who、bi,bf,bo Wxc、Whc、bc、v,Ws,WhThe matrix which satisfies corresponding dimensionality and has a numerical value of 0 to 1, wherein the sample batch size n is 8, T is m/n, and epoach of neural network training is 1000;
step 62, setting epoach to be 1;
step 63, setting t to be 1;
step 64, let Ht-1＝H0And sample XtSubstituting the formula in the step 2 to obtain H1；
Step 65, changing T to T +1, and turning to step 64 when T is less than or equal to T, or turning to step 66;
step 66, obtaining HTLet t' be 1;
step 67, enter the decoder part, let Ht′-1＝HTIs substituted into Ht′＝g(ct′,Ht′-1) To obtain Ht′Then make it possible to
Figure FDA0002830868820000041
Wherein
Figure FDA0002830868820000042
Step 68, changing T 'to T' +1, and turning to step 67 when T is less than or equal to T, or turning to step 69;
step 69, finally according to
Figure FDA0002830868820000043
Obtaining an error value error, and training an initialization parameter through back propagation; the parameter iterative optimization is carried out by adopting a random gradient descent algorithm, and the learning rate lr is 0.0003;
step 610, when the epoach is equal to or less than 1000, turning to step 63, otherwise, finishing training, so as to establish a prediction model of the long and short memory network based on the attention mechanism, and storing the model;
and 611, putting the test sample into the trained model for prediction, and predicting the state value at the next moment.
7. The method for diagnosing the abnormality of the elevator time series data based on the deep learning of claim 1, wherein the variation encoder in the step 4 comprises two parts: the encoder and the decoder are all full-connection layer neural networks, the encoder uses a four-layer full-connection network, the input dimensionality is d, the output dimensionality is 4, and the dimensionalities of two middle hidden layers are 8 and 6 respectively; the dimension of the decoder is opposite to that of the encoder, the input and output dimensions are respectively 4 and d, and the dimension of the middle hidden layer is 6 and 8; the encoder mainly maps the extracted data features into low-dimensional potential feature vectors, and the decoder decodes and reconstructs the low-dimensional potential feature vectors to be as close as possible to the original data features.
8. The method for diagnosing the abnormality of the elevator time series data based on the deep learning of claim 7 is characterized by comprising the following steps:
steps 81 and k indicate that the original sample dimension is 12, subscript j indicates the dimension of the potential feature vector, namely output dimension 4, the encoder part of the training sample encoder obtains the low-dimensional potential feature vector, and the sample is
Figure FDA0002830868820000044
Then output through the encoder as
Figure FDA0002830868820000045
Obtaining the latent variable zi：
Figure FDA0002830868820000051
Figure FDA0002830868820000052
Figure FDA0002830868820000053
Where randn _ like () denotes the return of a random number satisfying a normal distribution, and then the latent variable ziSample data x' with the same dimension as the training sample is restored through a decoderi；
And step 82, finding the correct distribution of the latent variables by using the lower bound of the marginal likelihood estimation of the input data as a target loss function:
Figure FDA0002830868820000054
and comparing errors of the output sample data and the training sample, calculating a target loss value by the above formula, wherein the target to be achieved is to make the difference between the sample data and the training sample as small as possible, and when the variational encoder is trained, the random gradient descent algorithm is used for training, the initialization parameter of the neural network is optimized, and the obtained target loss value is alpha.
9. The method for diagnosing the abnormality of the elevator time series data based on the deep learning according to the claim 1, characterized in that the step 5 further comprises the following specific processes:
step 91, inputting a test sample, and obtaining a state value of the test sample through model training based on a reconstruction probability threshold value alpha obtained by a training sample;
and step 92, inputting the state values into a variational encoder model for carrying out abnormity diagnosis, and recording the maximum value and the minimum value of the state values of the time series data with the target loss value lower than the target loss value to obtain [ delta max, delta min ], and obtaining the upper limit and the lower limit [ delta min, delta max ] of the normal range interval.
CN202011459448.1A 2020-12-11 2020-12-11 Elevator time sequence data abnormity diagnosis method based on deep learning Pending CN112488235A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011459448.1A CN112488235A (en) 2020-12-11 2020-12-11 Elevator time sequence data abnormity diagnosis method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011459448.1A CN112488235A (en) 2020-12-11 2020-12-11 Elevator time sequence data abnormity diagnosis method based on deep learning

Publications (1)

Publication Number Publication Date
CN112488235A true CN112488235A (en) 2021-03-12

Family

ID=74916525

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011459448.1A Pending CN112488235A (en) 2020-12-11 2020-12-11 Elevator time sequence data abnormity diagnosis method based on deep learning

Country Status (1)

Country Link
CN (1) CN112488235A (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113111096A (en) * 2021-04-08 2021-07-13 东方电气集团科学技术研究院有限公司 Abnormity detection method for high-dimensional time sequence working condition data of power generation equipment
CN113239614A (en) * 2021-04-22 2021-08-10 西北工业大学 Atmospheric turbulence phase space-time prediction algorithm
CN114266201A (en) * 2022-03-01 2022-04-01 杭州市特种设备检测研究院(杭州市特种设备应急处置中心) Self-attention elevator trapping prediction method based on deep learning
CN114538234A (en) * 2022-02-14 2022-05-27 深圳市爱丰达盛科技有限公司 Internet of things big data elevator safe operation standard AI self-building system and method
CN114626467A (en) * 2022-03-17 2022-06-14 湖南优湖科技有限公司 Feature cross elevator trapping time series prediction model construction method based on deep learning, obtained model and prediction method
CN114675118A (en) * 2022-05-30 2022-06-28 广东电网有限责任公司佛山供电局 Transformer winding abnormality detection method, device, equipment and storage medium
CN115078894A (en) * 2022-08-22 2022-09-20 广东电网有限责任公司肇庆供电局 Method, device and equipment for detecting abnormity of electric power machine room and readable storage medium
CN115797708A (en) * 2023-02-06 2023-03-14 南京博纳威电子科技有限公司 Power transmission and distribution synchronous data acquisition method
CN115859202A (en) * 2022-11-24 2023-03-28 浙江邦盛科技股份有限公司 Abnormal detection method and device under non-stationary time sequence data flow field scene
CN115983087A (en) * 2022-09-16 2023-04-18 山东财经大学 Method for detecting time sequence data abnormity by combining attention mechanism and LSTM and terminal

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108197736A (en) * 2017-12-29 2018-06-22 北京工业大学 A kind of Air Quality Forecast method based on variation self-encoding encoder and extreme learning machine
CN110689075A (en) * 2019-09-26 2020-01-14 北京工业大学 Fault prediction method of self-adaptive threshold of refrigeration equipment based on multi-algorithm fusion
US20200131003A1 (en) * 2017-10-20 2020-04-30 China University Of Mining And Technology Multiple-state health monitoring apparatus and monitoring method for critical components in hoisting system
CN111861272A (en) * 2020-07-31 2020-10-30 西安交通大学 Multi-source data-based complex electromechanical system abnormal state detection method
CN111914873A (en) * 2020-06-05 2020-11-10 华南理工大学 Two-stage cloud server unsupervised anomaly prediction method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200131003A1 (en) * 2017-10-20 2020-04-30 China University Of Mining And Technology Multiple-state health monitoring apparatus and monitoring method for critical components in hoisting system
CN108197736A (en) * 2017-12-29 2018-06-22 北京工业大学 A kind of Air Quality Forecast method based on variation self-encoding encoder and extreme learning machine
CN110689075A (en) * 2019-09-26 2020-01-14 北京工业大学 Fault prediction method of self-adaptive threshold of refrigeration equipment based on multi-algorithm fusion
CN111914873A (en) * 2020-06-05 2020-11-10 华南理工大学 Two-stage cloud server unsupervised anomaly prediction method
CN111861272A (en) * 2020-07-31 2020-10-30 西安交通大学 Multi-source data-based complex electromechanical system abnormal state detection method

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113111096A (en) * 2021-04-08 2021-07-13 东方电气集团科学技术研究院有限公司 Abnormity detection method for high-dimensional time sequence working condition data of power generation equipment
CN113111096B (en) * 2021-04-08 2023-09-05 东方电气集团科学技术研究院有限公司 Abnormality detection method for high-dimensional time sequence working condition data of power generation equipment
CN113239614A (en) * 2021-04-22 2021-08-10 西北工业大学 Atmospheric turbulence phase space-time prediction algorithm
CN114538234A (en) * 2022-02-14 2022-05-27 深圳市爱丰达盛科技有限公司 Internet of things big data elevator safe operation standard AI self-building system and method
CN114538234B (en) * 2022-02-14 2023-06-30 深圳市爱丰达盛科技有限公司 Automatic construction system and method for safe operation standard AI of big data elevator of Internet of things
CN114266201B (en) * 2022-03-01 2022-07-22 杭州市特种设备检测研究院(杭州市特种设备应急处置中心) Self-attention elevator trapping prediction method based on deep learning
CN114266201A (en) * 2022-03-01 2022-04-01 杭州市特种设备检测研究院(杭州市特种设备应急处置中心) Self-attention elevator trapping prediction method based on deep learning
CN114626467A (en) * 2022-03-17 2022-06-14 湖南优湖科技有限公司 Feature cross elevator trapping time series prediction model construction method based on deep learning, obtained model and prediction method
CN114675118A (en) * 2022-05-30 2022-06-28 广东电网有限责任公司佛山供电局 Transformer winding abnormality detection method, device, equipment and storage medium
CN115078894A (en) * 2022-08-22 2022-09-20 广东电网有限责任公司肇庆供电局 Method, device and equipment for detecting abnormity of electric power machine room and readable storage medium
CN115983087A (en) * 2022-09-16 2023-04-18 山东财经大学 Method for detecting time sequence data abnormity by combining attention mechanism and LSTM and terminal
CN115983087B (en) * 2022-09-16 2023-10-13 山东财经大学 Method for detecting time sequence data abnormality by combining attention mechanism with LSTM (link state machine) and terminal
CN115859202A (en) * 2022-11-24 2023-03-28 浙江邦盛科技股份有限公司 Abnormal detection method and device under non-stationary time sequence data flow field scene
CN115859202B (en) * 2022-11-24 2023-10-10 浙江邦盛科技股份有限公司 Abnormality detection method and device under non-stationary time sequence data stream scene
CN115797708A (en) * 2023-02-06 2023-03-14 南京博纳威电子科技有限公司 Power transmission and distribution synchronous data acquisition method

Similar Documents

Publication Publication Date Title
CN112488235A (en) Elevator time sequence data abnormity diagnosis method based on deep learning
Ma et al. Discriminative deep belief networks with ant colony optimization for health status assessment of machine
CN111813084B (en) Mechanical equipment fault diagnosis method based on deep learning
CN108095716B (en) Electrocardiosignal detection method based on confidence rule base and deep neural network
She et al. Wear indicator construction of rolling bearings based on multi-channel deep convolutional neural network with exponentially decaying learning rate
CN110555230B (en) Rotary machine residual life prediction method based on integrated GMDH framework
CN113762486B (en) Method and device for constructing fault diagnosis model of converter valve and computer equipment
CN108664690A (en) Long-life electron device reliability lifetime estimation method under more stress based on depth belief network
CN108398268A (en) A kind of bearing performance degradation assessment method based on stacking denoising self-encoding encoder and Self-organizing Maps
CN110705812A (en) Industrial fault analysis system based on fuzzy neural network
CN113096818B (en) Method for evaluating occurrence probability of acute diseases based on ODE and GRUD
CN115673596B (en) Welding abnormity real-time diagnosis method based on Actor-Critic reinforcement learning model
CN113917334A (en) Battery health state estimation method based on evolution LSTM self-encoder
CN114297918A (en) Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning
CN111222798B (en) Complex industrial process key index soft measurement method
CN112763967A (en) BiGRU-based intelligent electric meter metering module fault prediction and diagnosis method
CN114239397A (en) Soft measurement modeling method based on dynamic feature extraction and local weighted deep learning
CN113011102B (en) Multi-time-sequence-based Attention-LSTM penicillin fermentation process fault prediction method
CN117521512A (en) Bearing residual service life prediction method based on multi-scale Bayesian convolution transducer model
CN112381213A (en) Industrial equipment residual life prediction method based on bidirectional long-term and short-term memory network
CN117290726A (en) CAE-BiLSTM-based fault early warning method for mobile equipment
CN116430164A (en) Cable online monitoring method based on distributed temperature measurement and fault waveform analysis
CN115153549A (en) BP neural network-based man-machine interaction interface cognitive load prediction method
CN115828744A (en) White light LED fault on-line diagnosis and service life prediction method
CN115376638A (en) Physiological characteristic data analysis method based on multi-source health perception data fusion

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210312