CN112001482A

CN112001482A - Vibration prediction and model training method and device, computer equipment and storage medium

Info

Publication number: CN112001482A
Application number: CN202010819584.0A
Authority: CN
Inventors: 李逸帆; 刘文凯; 秦伟; 丁保剑
Original assignee: Guangzhou Xinke Jiadu Technology Co Ltd; PCI Suntek Technology Co Ltd
Current assignee: Guangzhou Xinke Jiadu Technology Co Ltd; PCI Suntek Technology Co Ltd
Priority date: 2020-08-14
Filing date: 2020-08-14
Publication date: 2020-11-27

Abstract

The embodiment of the invention discloses a vibration prediction and model training method, a device, computer equipment and a storage medium, wherein the vibration prediction method comprises the following steps: detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to open or close for n times through the belt to obtain n original vibration curves; inputting the n original vibration curves into a convolutional neural network respectively for processing so as to extract n original vibration characteristics respectively; sequentially inputting n original vibration characteristics into n chain-type dependent long-short term memory networks for processing so as to sequentially output n original implicit characteristics; and inputting the n original implicit characteristics into a deep neural network for processing to generate the amplitude of the belt when the motor drives the gate of the subway platform to be opened or closed next time through the belt, so as to obtain a target vibration curve. The state of the belt when the gate is opened or closed is monitored in real time, the control effect of opening or closing the gate is promoted, and therefore the efficiency of overhauling the belt is improved.

Description

Vibration prediction and model training method and device, computer equipment and storage medium

Technical Field

The embodiment of the invention relates to the technology of computer processing, in particular to a vibration prediction and model training method, a device, computer equipment and a storage medium.

Background

At present, the subway station arranges the shield door along the platform edge, keeps apart platform and driving tunnel region, reduces station air conditioning ventilation system's operation energy consumption, has reduced the influence of train noise and piston wind to the station simultaneously, prevents that personnel from falling the track and produce the accident, provides comfortable, safe environment of waiting for the passenger.

When the train arrives at the station, the gate in the shield door is opened, passengers can get on or off the train through the gate, then the gate is closed, and the train continues to run.

The gate of the subway platform and the motor run as a link through a belt, the belt jumps when the tension of the belt is too small, the control is not accurate enough, the friction is large when the tension of the belt is too large, and the loss of the belt, a belt pulley and the like is large.

In order to ensure the normal operation of the gate, the belt is usually repaired by adopting the maintenance modes such as fault repair, scheduled repair and the like at present.

The repair with failure means an after-repair action performed to restore the belt to a predetermined technical state after the belt fails or is damaged, and the regular repair actions performed to maintain the belt in the predetermined state before the belt fails are all manual operations, which not only have high cost, but also have a certain period after the belt fails and before the belt fails, so that the probability of successfully preventing the belt from failing is low, and the repair efficiency is low.

Disclosure of Invention

The embodiment of the invention provides a vibration prediction and model training method, a device, computer equipment and a storage medium, and aims to solve the problems of high maintenance cost and low efficiency of a belt between a gate and a motor of a subway platform.

In a first aspect, an embodiment of the present invention provides a vibration prediction method, including:

detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to open or close for n times through the belt to obtain n original vibration curves;

inputting the n original vibration curves into a convolutional neural network respectively for processing so as to extract n original vibration characteristics respectively;

sequentially inputting the n original vibration characteristics into n chain-dependent long-short term memory networks for processing so as to sequentially output n original implicit characteristics;

inputting the n original implicit characteristics into a deep neural network for processing so as to generate the amplitude of the belt when the motor drives the gate of the subway platform to be opened or closed next time through the belt, and obtaining a target vibration curve.

In a second aspect, an embodiment of the present invention further provides a training method for a vibration prediction model, including:

detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to open or close for n times through the belt to obtain n sample local oscillation curves;

detecting the amplitude of the belt when the motor drives a gate of the subway platform to open or close through the belt next time, and obtaining a reference vibration curve;

inputting the n sample vibration curves into a convolutional neural network respectively for processing so as to extract n sample vibration characteristics respectively;

sequentially inputting the n sample vibration characteristics into n chain-dependent long-short term memory networks for processing according to a sequence so as to sequentially output n sample implicit characteristics;

inputting the implicit characteristics of the n samples into a deep neural network for processing so as to generate the amplitude of the belt when the motor drives a gate of the subway platform to be opened or closed next time through the belt, and obtaining a predicted vibration curve;

and training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve.

In a third aspect, an embodiment of the present invention further provides a vibration prediction apparatus, including:

the system comprises an original vibration curve detection module, a vibration detection module and a vibration detection module, wherein the original vibration curve detection module is used for detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to be opened or closed through the belt for n times to obtain n original vibration curves;

the original vibration feature extraction module is used for respectively inputting the n original vibration curves into a convolutional neural network for processing so as to respectively extract n original vibration features;

the original implicit feature extraction module is used for sequentially inputting the n original vibration features into n chain-dependent long-short term memory networks for processing so as to sequentially output the n original implicit features;

and the target vibration curve prediction module is used for inputting the n original implicit characteristics into a deep neural network for processing so as to generate the amplitude of the belt when the motor drives the gate of the subway platform to be opened or closed next time through the belt, and a target vibration curve is obtained.

In a fourth aspect, an embodiment of the present invention further provides a training apparatus for a vibration prediction model, including:

the system comprises a sample vibration curve detection module, a signal processing module and a signal processing module, wherein the sample vibration curve detection module is used for detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to be opened or closed through the belt for n times to obtain n sample vibration curves;

the reference vibration curve detection module is used for detecting the amplitude of the belt when the motor drives the gate of the subway platform to be opened or closed through the belt next time, so as to obtain a reference vibration curve;

the sample vibration feature extraction module is used for respectively inputting the n sample vibration curves into a convolutional neural network for processing so as to respectively extract n sample vibration features;

the sample implicit feature extraction module is used for sequentially inputting the n sample vibration features into n chain-dependent long-short term memory networks for processing so as to sequentially output the n sample implicit features;

the predicted vibration curve prediction module is used for inputting the implicit characteristics of the n samples into a deep neural network for processing so as to generate the amplitude of the belt when the motor drives the gate of the subway platform to be opened or closed next time through the belt, and a predicted vibration curve is obtained;

and the vibration prediction model training module is used for training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve.

In a fifth aspect, an embodiment of the present invention further provides a computer device, where the computer device includes:

one or more processors;

a memory for storing one or more programs;

when executed by the one or more processors, cause the one or more processors to implement a vibration prediction method as described in the first aspect or a training method of a vibration prediction model as described in the second aspect.

In a sixth aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the vibration prediction method according to the first aspect or the training method of the vibration prediction model according to the second aspect.

In this embodiment, the amplitude of the belt is detected when the motor continuously drives the gate of the subway platform to open or close for n times through the belt, n original vibration curves are obtained, the n original vibration curves are respectively input into the convolutional neural network for processing, so as to respectively extract n original vibration features, the n original vibration features are sequentially input into n chain-dependent long-short term memory networks for processing according to a sequence, so as to sequentially output n original implicit features, the n original implicit features are input into the deep neural network for processing, so as to generate the amplitude of the belt when the motor drives the gate of the subway platform to open or close next time through the belt, and a target vibration curve is obtained, the state of the belt is stable within a certain time range, namely the vibration of the belt is correlated within a certain time range, and the original vibration curve belongs to a position correlated in a time dimension, The data of long sequence, through the processing of long sequence data of the dimensionality reduction realization of convolution neural network, the characteristic through long short-term memory network extraction time dimension realizes the processing of data under the time dimension, when this maps out the gate that the motor next time passes through belt drive subway platform and opens or closes, the amplitude of belt, can guarantee the accuracy of amplitude, thereby realize the state of real-time supervision belt when the gate is opened or is closed, help promoting the control effect of opening or closing the gate, and can provide the unusual early warning of belt state when the gate is opened or is closed, reduce the gate and take place the probability of trouble when opening or closing the belt, thereby promote the efficiency of overhauing the belt.

Drawings

Fig. 1 is a flowchart of a vibration prediction method according to an embodiment of the present invention;

fig. 2 is an exemplary diagram of an original vibration curve when a gate is opened according to an embodiment of the present invention;

FIG. 3 is a diagram illustrating an example of an abnormal amplitude affected by disturbance according to an embodiment of the present invention;

fig. 4 is a schematic structural diagram of a vibration prediction model according to an embodiment of the present invention;

fig. 5 is a flowchart of a training method of a vibration prediction model according to a second embodiment of the present invention;

fig. 6 is a schematic diagram illustrating a comparison between a predicted vibration curve and a reference vibration curve according to a second embodiment of the present invention;

FIG. 7 is a diagram illustrating an example of calculating a loss value according to a second embodiment of the present invention;

fig. 8 is a flowchart of a vibration prediction method according to a third embodiment of the present invention;

fig. 9 is an exemplary diagram of an original vibration curve when a gate is closed according to a third embodiment of the present invention;

FIG. 10 is a flowchart of a training method of a vibration prediction model according to a fourth embodiment of the present invention;

fig. 11 is a schematic structural diagram of a vibration prediction apparatus according to a fifth embodiment of the present invention;

fig. 12 is a schematic structural diagram of a training apparatus for a vibration prediction model according to a sixth embodiment of the present invention;

fig. 13 is a schematic structural diagram of a computer device according to a seventh embodiment of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.

Example one

Fig. 1 is a flowchart of a vibration prediction method according to an embodiment of the present invention, where the embodiment is applicable to a situation where a motor drives a gate of a subway platform to open through a belt, and a following vibration curve is predicted according to a preceding vibration curve of the belt, the method may be executed by a vibration prediction apparatus, the vibration prediction apparatus may be implemented by software and/or hardware, and may be configured in a computer device, such as a server, a workstation, a personal computer, and the like, and the method specifically includes the following steps:

s101, detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to be opened for n times through the belt, and obtaining n original vibration curves.

The platform comprises a platform, a shielding door, a motor and a gate, wherein the shielding door is arranged along the edge of the platform in a subway station, the motor and the gate are arranged in the shielding door, one end of a belt is sleeved on the motor, the other end of the belt is sleeved on the gate, the motor receives control signals of a control system under scenes such as arrival of a train and the like, the motor responds to the control signals and rotates to drive the belt to rotate, and the belt drives the gate to move towards two sides so as to drive the gate to be.

In this embodiment, for the belt layout sensor, the sensor detects, in real time, the amplitude (i.e., vibration amplitude) of the belt when the gate of the subway platform is driven by the belt to open by the motor at a preset frequency, and when the motor stops running, the belt is naturally vertical and is spaced from the sensor by a fixed distance, and at this time, the amplitude is constant.

Each detection forms a data structure which is stored in a database, wherein the data structure comprises parameters such as time (time) of collection, value (value) and the like.

As shown in fig. 2, a coordinate axis is established with time (time) as a horizontal axis and a value (value) as a vertical axis, and the amplitude of the belt when the gate of the subway station platform is driven by the belt is marked in the form of a point on the coordinate axis, thereby generating an original vibration curve 200.

As shown in fig. 2, from the operation mechanism of the equipment, when the gate of the subway platform is opened, in order to make the gate obtain an initial speed, the current of the motor is usually and rapidly increased to a rated value, at this time, the amplitude of the belt is large, then the value of the current is slowly reduced and is continued for a short time, so as to control the speed of opening the gate, at this time, the amplitude of the belt slowly decreases and tends to be gentle, at last, the gate is opened until the current disappears, the belt is in a tightened state, and the amplitude is constant.

As shown in fig. 2, the distance from the sensor when the belt is in a natural vertical state is 15000, and the distance from the sensor when the gate is opened is 15000.

If the amplitude of the belt when the motor drives the gate of the subway platform to be opened through the belt next time is predicted, the amplitude of the belt when the motor continuously passes through the belt for n times (n is a positive integer) to drive the gate of the subway platform to be opened and the belt in a period of time before the current time can be extracted from the database to be used as an original vibration curve.

The operation that the motors drive the gates of the subway platform to open through the belts is performed continuously, namely, before the current time is extracted, the amplitude of the belts is obtained when the 1 st motor drives the gates of the subway platform to open through the belts, the amplitude of the belts is obtained when the 2 nd motor drives the gates of the subway platform to open through the belts, and the amplitude of the belts is obtained when the 3 rd motor drives the gates of the subway platform to open through the belts and the amplitude of the belts is obtained when the … … nth motor drives the gates of the subway platform to open through the belts.

Because the amplitude collected from the belt can generate abnormal data due to field environment disturbance or communication signal loss and the like, the original vibration curve can be preprocessed when the motor is predicted to open the gate of the subway platform driven by the belt next time and before the amplitude of the belt is predicted, so that the original vibration curve meets the calculation specification.

In one example of preprocessing, the amplitude of the vibration that conforms to the traffic anomaly is looked up in the raw vibration curve as a first raw anomaly amplitude, and the first raw anomaly amplitude is replaced with the amplitude of the belt when not running, i.e., the distance of the belt from the sensor in a natural vertical position.

In this example, if the communication signal is lost during the process of acquiring the amplitude of the belt when the gate of the subway platform driven by the motor through the belt is opened, the system will return an abnormal value (i.e. the first original abnormal amplitude), such as 65535, and the abnormal value (i.e. the first original abnormal amplitude) can be replaced by the amplitude of the belt when the belt is not running, such as 15000.

In another example of preprocessing, the amplitude corresponding to the disturbance anomaly is searched for in the original vibration curve as a second original anomaly amplitude, and the second original anomaly amplitude is replaced by an amplitude adjacent to the second original anomaly amplitude.

In this example, during the process of collecting the amplitude of the belt when the gate of the subway platform driven by the motor through the belt is opened, if there is disturbance influence in the field environment, the belt generates abnormal values (i.e. the second original abnormal amplitude) in a reasonable value area, i.e. the value range of the abnormal values is within 0 to 30000, but the values obviously do not conform to the curve trend, such as the

points

301, 302, 303 shown by the circles shown in fig. 3, and at this time, the abnormal values (i.e. the second original abnormal amplitude) can be replaced by adjacent amplitudes.

Further, in this example, the abnormal value may be found by using a 3 σ rule, specifically, the original vibration curve is divided into a plurality of intervals to obtain an original vibration interval, for example, assuming that the original vibration curve has 1000 points, the original vibration curve may be divided into 5 intervals by 200 points in time sequence.

The mean value u and the standard deviation σ of the values of the amplitudes in each of the original vibration intervals are calculated.

A triple value of the standard deviation σ is added to the average value u to obtain a first target value u +3 σ.

Three times the standard deviation sigma is subtracted from the mean u to obtain a second target value u-3 sigma.

And if the value of the amplitude in the original vibration interval is larger than the first target value u +3 sigma or smaller than the second target value u-3 sigma, determining that the amplitude accords with the disturbance abnormality and is the second original abnormal amplitude.

In addition to the 3 σ rule, when implementing the embodiment of the present invention, other ways of detecting the second original abnormal amplitude may be set according to practical situations, for example, exceeding n times of the average value of all amplitudes, that is, the second original abnormal amplitude, and so on, which is not limited by the embodiment of the present invention.

Of course, the foregoing pretreatment is only an example, and when the embodiment of the present invention is implemented, other pretreatment may be set according to actual situations, which is not limited in the embodiment of the present invention. In addition, besides the above judgment processing method, a person skilled in the art may also adopt other preprocessing according to actual needs, and the embodiment of the present invention is not limited thereto.

S102, inputting the n original vibration curves into a convolutional neural network respectively for processing so as to extract n original vibration characteristics respectively.

In this embodiment, a vibration prediction model may be trained in advance, and the vibration prediction model may be used to predict the amplitude of the belt when the motor drives the gate of the subway platform to open through the belt next time according to the amplitude of the belt when the motor drives the gate of the subway platform to open through the belt n times continuously.

As shown in fig. 4, the vibration prediction model may include three layers, which are respectively the following in the forward propagation direction:

1. convolutional Neural Network (CNN) 410

2. Long Short-Term Memory network (LSTM) 420

3. Deep Neural Networks (DNN) 430

The convolutional neural network is a neural network with a convolutional structure, that is, the convolutional neural network includes a feature extractor composed of convolutional layers and sub-sampling layers (pooling layers). In the convolutional layer of the convolutional neural network, one neuron is connected to only part of the neighbor neurons. In a convolutional layer of CNN, there are usually several feature maps (featuremaps), each feature map is composed of some neurons arranged in a rectangle, and the neurons of the same feature map share a weight, where the shared weight is a convolution kernel. The convolution kernel is generally initialized in the form of a random decimal matrix, and the convolution kernel learns to obtain a reasonable weight in the training process of the network. Sharing weights (convolution kernels) brings the immediate benefit of reducing the connections between layers of the network, while reducing the risk of over-fitting. Subsampling is also called pooling (posing) and is typically in the form of mean subsampling (mean posing), maximum subsampling (max posing), minimum subsampling (min posing), and the like. Sub-sampling can be viewed as a special convolution process. Convolution and sub-sampling greatly simplify the complexity of the model and reduce the parameters of the model.

It should be noted that the convolutional neural network is a general network, that is, for n motors continuously driving the gate of the subway platform to open through the belt n times and the original vibration curve of the belt, the same convolutional neural network is used to extract n original vibration features.

In one example, as shown in fig. 4, convolutional neural network 410 includes a first convolutional layer 411, a second convolutional layer 412, a third convolutional layer 413, a Pooling layer (Pooling layer) 414.

The first Convolutional layer 411, the second Convolutional layer 412, and the third Convolutional layer 413 are Convolutional layers (Convolutional layers), each Convolutional layer in the Convolutional neural network is composed of a plurality of Convolutional units, and parameters of each Convolutional unit are obtained through optimization of a back propagation algorithm. The convolution operation aims to extract different input features, the first layer of convolution layer can extract some low-level features such as edges, lines, angles and the like, and more layers of networks can iteratively extract more complex features from the low-level features.

Pooling layers typically result in very large-dimensional features after the layers are packed, and the features are cut into regions, averaged, maximized, or minimized to obtain new, smaller-dimensional features.

In this example, n raw vibration curves (x)₁，x₂，……，x_n-1，x_n)401 are respectively input into the first convolution layers 411 for convolution operation to respectively output n first original candidate features 402.

Assuming that the target vibration curve 401 is a vector with dimension 1000, the convolution kernel size of the first convolution layer 411 is 4, stride is 1, and after the convolution operation, the first original candidate feature 402 is a vector with dimension 997.

The n first original candidate features 402 are respectively input into the second convolution layer 412 for convolution operation to respectively output n second original candidate features 403.

Assuming that the convolution kernel size of the second convolution layer 412 is 8 and stride is 2, after the convolution operation, the second original candidate feature 403 is a vector with dimension 495.

The n second original candidate features 403 are respectively input into the third convolution layer 413 for convolution operation to respectively output n third original candidate features 404.

Assuming that the convolution kernel size of the third convolution layer 413 is 16 and stride is 2, after the convolution operation, the third original candidate feature 404 is a vector with dimension 240.

The n third original candidate features 404 are respectively input into the pooling layer 414 for pooling operation, so as to respectively output n original vibration features 405.

Assuming a kernel size of 4 and stride of 1 for pooling layer 414, after the maximum pooling operation, the original vibration signature 405 is a vector of dimension 237.

Of course, the structure of the convolutional neural network is only an example, and when the embodiment of the present invention is implemented, other structures of the convolutional neural network may be set according to practical situations, for example, two convolutional layers are used, or four convolutional layers are used, and the embodiment of the present invention is not limited thereto. In addition, besides the above structure of the convolutional neural network, a person skilled in the art may also adopt other structures of the convolutional neural network according to actual needs, and the embodiment of the present invention is not limited thereto.

S103, sequentially inputting the n original vibration characteristics into the n chain-dependent long-short term memory networks for processing according to the sequence, so as to sequentially output the n original implicit characteristics.

The long-short term memory Network belongs to a special type of RNN (Recurrent Neural Network), can learn long-term dependence information, and is suitable for processing and predicting important events with relatively long intervals and delays in a time sequence.

It should be noted that the long-term and short-term memory network is a dedicated network, and the original vibration curves and the long-term and short-term memory network are in a one-to-one correspondence relationship, that is, for the original vibration curves of the belt when each motor drives the gate of the subway platform to open through the belt, the independent long-term and short-term memory network is used to extract the original implicit features.

Furthermore, so-called chain dependency, i.e. the input of a following long-short term memory network depends on the output of a preceding long-short term memory network.

In a specific implementation, as shown in fig. 4, n original vibration features may be sequentially traversed in sequence, an original reserved feature output by a previous long-short term memory network is determined for a current original vibration feature, the current original vibration feature and the previous original reserved feature are input into the current long-short term memory network for processing, so as to output an original reserved feature and an original implicit feature, the current original reserved feature is output to a next long-short term memory network, and the process is repeated until n LSTMs complete processing the n original vibration features.

The hidden layer of the original RNN has only one state, h, which is very sensitive to short-term input. A state, c, is added to save a long-term state called cell state.

The LSTM uses the concept of gate (gate) in forward calculations. The gate is actually a fully connected layer, with the input being a vector and the output being a real vector between 0 and 1.

The LSTM controls the content of the cell state c with two gates, one being a forgetting gate (forget gate) which determines the cell state c at the last time t-1_t-1How much to keep the cell state c to the current time t_t(ii) a The other is an input gate (input gate), which determines the input x of the network at the present moment_tHow many cells to save to cell state c_t. LSTM uses output gate (output gate) to control cell state c_tHow much current output value h is output to LSTM_t。

For forgetting door f_t：

f_t＝σ(W_f·[h_t-1，x_t]+b_f)

Wherein, W_fIs the weight matrix of the forgetting gate, h_t-1Represents the output, x, of the last time instant t-1_tIndicates the input of the current time, [ h ]_t-1，x_t]Representing the concatenation of two vectors into a longer vector, b_fIs the bias term for the forgetting gate, σ is the sigmoid function.

For input gate i_t：

i_t＝σ(W_i·[h_t-1，x_t]+b_i)

Wherein, W_iIs a weight matrix of the input gate, b_iIs the offset term of the input gate.

For describing the cell state at the current time t

Wherein tanh () represents an activation function, W_cIs a weight matrix of cell states, b_cIs the bias term for the cell state.

From the cell state c at the last instant t-1_t-1Multiplication by element of forget gate f_tReuse the cell state at the current time t

Multiplying input Gate i by element_tThen, the two products are added to generate the unit state c of the current time_t：

Wherein, the symbol

Meaning multiplication by element.

In this way, the LSTM is remembered with respect to the current

And long term memory c_t-1Combine to form a new cell state c_t. It can keep information long before because of forgetting the control of the gate, and it can avoid the current irrelevant content to enter into the memory because of inputting the control of the gate.

Then, the output gate o_tThe influence of long-term memory on the current output is controlled:

o_t＝σ(W_o·[h_t-1，x_t]+b_o)

wherein, W_oIs a weight matrix of output gates, b_oIs an offset term of the output gate

Then, for the output gate o_tOutput h of_t：

In this embodiment, the current original vibration characteristic and the last original reserved characteristic may be input into the current long-short term memory network for processing, and the characteristic of forgotten gate output in the current long-short term memory network (i.e. the unit state c at the current time) may be determined_t) Determining output characteristics of output gates (i.e. output gate o) in current long-short term memory network for original reserved characteristics_tOutput h of_t) Is the original implicit feature.

And S104, inputting the n original implicit characteristics into a deep neural network for processing to generate the amplitude of the belt when the motor drives the gate of the subway platform to be opened next time through the belt, and obtaining a target vibration curve.

While neural networks are based on extensions of the perceptron, DNN can be understood as neural networks with many hidden layers. Multi-Layer neural networks and deep neural networks DNN are also really one thing to be referred to, and DNN is sometimes called a Multi-Layer perceptron (MLP).

From the DNN, which is divided according to the positions of different layers, the neural network layers inside the DNN can be divided into three types, an input layer, a hidden layer and an output layer, generally speaking, the first layer is the input layer, the last layer is the output layer, and the middle layers are all hidden layers.

In a specific implementation, as shown in FIG. 4, n original implicit features (h) are combined₁，h₂，……，h_n-1，h_n) Inputting the vibration amplitude into DNN, mapping the vibration amplitude to the amplitude of a belt when a motor drives a gate of the subway platform to be opened through the belt next time, and taking the amplitude as a target vibration curve x_n+1Target vibration curve x_n+1Dimension number of (d) and original vibration curve (x)₁，x₂，……，x_n-1，x_n) Are the same in number.

In this embodiment, when the motor continuously passes through the gate of the belt-driven subway platform for n times and the amplitude of the belt is detected, n original vibration curves are obtained, the n original vibration curves are respectively input into the convolutional neural network for processing, so as to respectively extract n original vibration features, the n original vibration features are sequentially input into n chain-dependent long and short term memory networks for processing according to the sequence, so as to sequentially output n original implicit features, the n original implicit features are input into the deep neural network for processing, so as to generate the amplitude of the belt when the motor next passes through the gate of the belt-driven subway platform and is opened, a target vibration curve is obtained, the state of the belt is stable within a certain time range, namely, the vibration of the belt is correlated within a certain time range, and the original vibration curves belong to a position, and a position, Data of long sequence, reduce the processing that the dimension realized long sequence data through the convolutional neural network, the characteristic through long short-term memory network extraction time dimension realizes the processing of data under the time dimension, when this maps out the gate of motor next time through belt drive subway platform and opens, the amplitude of belt, can guarantee the accuracy of amplitude, thereby realize the state of real-time supervision belt when the gate is opened, help promoting the control effect of opening the gate, and can provide the unusual early warning of belt state when the gate is opened, reduce the probability that the belt broke down when opening, thereby promote the efficiency of overhauing the belt.

Example two

Fig. 5 is a flowchart of a training method of a vibration prediction model according to a second embodiment of the present invention, where the method is applicable to training the vibration prediction model to predict a subsequent vibration curve according to a previous vibration curve of a belt for a motor driving a gate of a subway platform to open through the belt, and the training device of the vibration prediction model may be implemented by software and/or hardware, and may be configured in computer equipment, such as a server, a workstation, a personal computer, and the like, and the method specifically includes the following steps:

s501, detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to be opened for n times through the belt, and obtaining n sample local oscillation curves.

When the vibration prediction model is trained, the amplitude of a belt when a motor continuously passes through the belt for n times to drive a gate of a subway platform to be opened and in a historical period of time can be extracted from a database to be used as n sample vibration curves.

Because the amplitude collected from the belt can generate abnormal data due to field environment disturbance or communication signal loss and the like, a sample vibration curve can be preprocessed before the vibration prediction model is trained, so that in a preprocessing example, the vibration curve of the sample meets the calculation specification.

In one example of preprocessing, the amplitude of the sample vibration curve is looked up for compliance with the traffic anomaly, and as a first sample anomaly amplitude, the first sample anomaly amplitude is replaced with the amplitude of the belt when not in operation.

In another example of preprocessing, the amplitude conforming to the disturbance anomaly is found in the sample vibration curve, and as the second sample anomaly amplitude, the amplitude adjacent to (on the left side or on the right side) the second sample anomaly amplitude is substituted for the second sample anomaly amplitude.

In a specific implementation, a sample vibration curve can be divided into a plurality of intervals to obtain a sample vibration interval; calculating the average value and the standard deviation of the amplitude in the sample vibration interval; adding a triple value of the standard deviation on the basis of the average value to obtain a first sample value; subtracting a triple value of the standard deviation on the basis of the average value to obtain a second sample value; and if the amplitude in the sample vibration interval is larger than the first sample value or smaller than the second sample value, determining that the amplitude accords with the disturbance abnormality and determining that the amplitude is the second sample abnormal amplitude.

S502, detecting the amplitude of the belt when the motor drives the gate of the subway platform to be opened through the belt next time, and obtaining a reference vibration curve.

In this embodiment, the amplitude of the belt when the motor drives the gate of the subway platform to be opened next time (i.e., the (n + 1) th time) through the belt may be detected as a reference vibration curve, and the reference vibration curve may be regarded as a Tag (Tag) for the vibration prediction model.

And S503, inputting the n sample vibration curves into a convolutional neural network respectively for processing so as to extract n sample vibration characteristics respectively.

In this embodiment, the convolutional neural network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a pooling layer.

And respectively inputting the n sample local oscillation curves into the first convolution layers for convolution operation so as to respectively output n first sample candidate characteristics.

And inputting the n first sample candidate features into the second convolution layers respectively for convolution operation so as to output n second sample candidate features respectively.

And inputting the n second sample candidate features into the third convolution layers respectively for convolution operation so as to output n third sample candidate features respectively.

And inputting the n third sample candidate features into the pooling layer respectively for pooling operation so as to output n sample vibration features respectively.

S504, sequentially inputting the n sample vibration characteristics into the n chain-dependent long-short term memory networks for processing according to the sequence, so as to sequentially output the n sample implicit characteristics.

In this embodiment, the n sample vibration features may be sequentially traversed in order, and the sample retention feature output by the previous long-term and short-term memory network may be determined for the current sample vibration feature; inputting the vibration characteristic of the current sample and the retention characteristic of the previous sample into the current long-short term memory network for processing so as to output the retention characteristic and the implicit characteristic of the sample.

In a specific implementation, the vibration characteristic of the current sample and the retention characteristic of the previous sample are input into the current long-short term memory network for processing; determining the characteristics output by a forgetting gate in the current long-short term memory network as sample retention characteristics; and determining the characteristic of the output gate in the current long-short term memory network as the sample implicit characteristic.

In one embodiment of the invention, a dropout mechanism is added in each long-short term memory network to prevent overfitting.

For the dropout mechanism, in the training (forward propagation), a part of neural network units in the current long-short term memory network can be ignored in a mask mode and the like according to a certain probability (for example, when dropout is 0.5, the probability is 50%);

and inputting the vibration characteristics of the current sample into the current long-short term memory network in sequence, and extracting the implicit characteristics of the sample by using the non-ignored neural network units.

And S505, inputting the implicit characteristics of the n samples into a deep neural network for processing so as to generate the amplitude of the belt when the motor drives the gate of the subway platform to be opened next time through the belt, and obtaining a predicted vibration curve.

In this embodiment, the implicit characteristics of n samples are input into the DNN, and are mapped to the amplitude of the belt when the motor drives the gate of the subway platform to open through the belt next time, and the amplitude is used as a predicted vibration curve, and the dimension number of the predicted vibration curve is the same as the dimension number of the sample vibration curve.

S506, training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve.

In this embodiment, the vibration prediction model includes a convolutional neural network, n chain-dependent long-short term memory networks, and a deep neural network, and there are 3 parameters for training the vibration prediction model, which are the modeling time length, the batch training size, and the training iteration number, respectively.

The modeling time length means that the vibration curve is predicted to be related to the vibration curve in the previous n samples, wherein n is the modeling time length, such as 300.

The batch training size refers to the size of the number of samples of the vibration prediction model during each training, i.e., the number of sample vibration curves, such as 200.

The number of training iterations refers to the number of training iterations of the vibration prediction model, e.g., 300.

As shown in fig. 6, the reference vibration curve 601 is an actual result, the predicted vibration curve 602 is an estimated result, and the quality of the vibration prediction model can be evaluated by comparing the predicted vibration curve 601 with the reference vibration curve 602, so as to guide the training of the vibration prediction model.

In one embodiment of the present invention, S506 may include the steps of:

s5061, a difference between the predicted vibration curve and the reference vibration curve is calculated as a loss value.

In this embodiment, the predicted vibration curve may be compared with the reference vibration curve, and the difference between the predicted vibration curve and the reference vibration curve may be obtained as the LOSS value LOSS of the amplitude of the belt when the motor drives the gate of the subway platform to open next time through the belt.

In one example, as shown in fig. 7, a difference between an amplitude 701 (darker portion) in the predicted vibration curve and an amplitude 702 (lighter portion) in the reference vibration curve may be calculated as point deviation values, respectively, for each same time (i.e., time at which vibration is collected), and an average of all the point deviation values may be calculated as a LOSS value LOSS.

Of course, the LOSS value LOSS may be calculated in other manners besides the average value of the point deviation values, for example, a sum of the point deviation values is calculated, and the like, which is not limited in this embodiment.

S5062, judging whether stop conditions are met; if so, then S5063 is performed, otherwise, S5064 is performed.

S5063, outputting a convolutional neural network, and taking n chain-dependent long-short term memory networks and a deep neural network as vibration prediction models.

In this embodiment, a stop condition, such as a loss value being smaller than a preset threshold, reaching of the number of training iterations, or the like, may be set in advance.

When the stopping condition is met, the training vibration prediction model is determined to be completed, the structures and parameters (including weights) of the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network are stored, and when the parameters (including weights) are loaded, the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network can be used for predicting the amplitude of the belt when the motor drives the gate of the subway platform to be opened through the belt for n times continuously according to the amplitude of the belt when the motor drives the gate of the subway platform to be opened through the belt for the next time.

And when the stopping condition is not met, continuously training the vibration prediction model, and updating the parameters of the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network.

And S5064, updating the deep neural network, the n chain-dependent long-short term memory networks and the convolutional neural network based on the loss value, returning to execute S503, S504 and S505, and entering next training.

In this embodiment, the parameters (including weights) of the deep neural network, the n chain-dependent long-short term memory networks, and the convolutional neural network may be updated in a back-propagation manner based on the loss values.

For the dropout mechanism, in the training (back propagation), the non-ignored neural network units in the current long-short term memory network can be determined, and the weights in the non-ignored neural network units are updated in a mode of masking and the like.

In another embodiment of the invention, a mechanism for recharging the hidden layer is added, and the connection with the last iteration is disconnected, because the hidden layer is not reset by the Batch every time, the LOSS value LOSS may be trapped in the local optimum, and the random disturbance can be added to the feasible solution when the hidden layer is trapped in the local optimum by adding the mechanism, so that the local optimum can be more easily jumped out.

In the specific implementation, in the training, data of a hidden layer in the first long-short term memory network is randomly set and input aiming at the first long-short term memory network, and when the data are reversely transmitted, the first (namely, the first) long-short term memory network is trained on the basis of the data.

In the embodiment, the amplitude of the belt when the motor continuously drives the gate of the subway platform to open through the belt for n times is detected to obtain n sample local oscillation curves; detecting the amplitude of a belt when a motor drives a gate of a subway platform to open through the belt next time, obtaining a reference vibration curve, respectively inputting n sample vibration curves into a convolutional neural network for processing, respectively extracting n sample vibration characteristics, sequentially inputting n sample vibration characteristics into n chain-dependent long-short term memory networks for processing according to the sequence, sequentially outputting n sample hidden characteristics, inputting the n sample hidden characteristics into a deep neural network for processing, generating the amplitude of the belt when the motor drives the gate of the subway platform to open through the belt next time, obtaining a predicted vibration curve, training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve, and training the belt in a certain time range, the state is relatively stable, i.e. the vibration is correlated in a certain time range, the sample vibration curve belongs to a long sequence of data correlated in the time dimension, the dimension reduction of the convolutional neural network is used for realizing the processing of long sequence data, the long and short term memory network is used for extracting the characteristics of the time dimension for realizing the processing of data under the time dimension, therefore, when the gate of the subway platform is driven to be opened by the motor through the belt next time, the amplitude of the belt is predicted, the vibration prediction model is trained, the performance of the vibration prediction model can be ensured, namely, the accuracy of the predicted amplitude is ensured, thereby realizing the real-time monitoring of the state of the belt when the gate is opened, being beneficial to improving the control effect of opening the gate, and can provide the unusual early warning of belt state when the gate is opened, reduce the gate probability that the belt broke down when opening to promote the efficiency of overhauing the belt.

EXAMPLE III

Fig. 8 is a flowchart of a vibration prediction method according to a third embodiment of the present invention, where the present embodiment is applicable to a situation where a motor drives a gate of a subway platform to close through a belt, and a following vibration curve is predicted according to a preceding vibration curve of the belt, and the method may be executed by a vibration prediction apparatus, where the vibration prediction apparatus may be implemented by software and/or hardware, and may be configured in computer equipment, such as a server, a workstation, a personal computer, and the like, and the method specifically includes the following steps:

s801, detecting the amplitude of the belt when the motor continuously drives the gate of the subway platform to close through the belt for n times, and obtaining n original vibration curves.

In this embodiment, for the belt layout sensor, the sensor detects, in real time, the amplitude (i.e., vibration amplitude) of the belt when the gate of the subway platform is driven to close by the belt by the motor at a preset frequency, and when the motor stops running, the belt is naturally vertical and is spaced from the sensor by a fixed distance, and at this time, the amplitude is constant.

As shown in fig. 9, a coordinate axis is established with time (time) as a horizontal axis and a value (value) as a vertical axis, and the amplitude of the belt when the gate of the subway platform is driven by the belt to be closed is marked in the form of a point on the coordinate axis, thereby generating an original vibration curve 900.

As shown in fig. 9, from the operation mechanism of the equipment, when the gate of the subway platform is closed, the current of the motor is at the rated value, the fluctuation of the amplitude of the belt is large, then the value of the current is slowly reduced, the amplitude of the belt slowly rises, the fluctuation is reduced and tends to be gentle until the current disappears when the gate is closed, the belt is in a tight state, and the amplitude is constant.

As shown in fig. 9, the distance from the sensor when the belt is naturally vertical was 15000, the distance from the sensor when the gate was closed was 15000, and the amplitude stopped at a position close to 15000 when the motor was stopped.

Compared with the original vibration curve of the motor for driving the gate of the subway platform to be opened shown in fig. 2, when the motor drives the gate of the subway platform to be opened and closed, the original vibration curve has different characteristics, such as opposite vibration directions, so that the two states of driving the gate of the subway platform to be opened and closed by the motor are usually processed independently, namely a vibration prediction model is trained independently, and the amplitude of a belt when the motor drives the gate of the subway platform to be opened and closed next time is predicted independently to be used as the original vibration curve.

If the amplitude of the belt when the gate of the subway platform is driven to close by the motor through the belt next time is predicted, the amplitude of the belt when the gate of the subway platform is driven to close by the motor for n (n is a positive integer) times continuously in a period of time before the current time is extracted from the database and used as an original vibration curve.

The operation that the motors drive the gates of the subway platform to close through the belts is performed continuously, namely, before the current time is extracted, the amplitude of the belts is obtained when the 1 st motor drives the gates of the subway platform to close through the belts, the amplitude of the belts is obtained when the 2 nd motor drives the gates of the subway platform to close through the belts, and the amplitude of the belts is obtained when the 3 rd motor drives the gates of the subway platform to close through the belts, and the amplitude of the belts is obtained when the … … nth motor drives the gates of the subway platform to close through the belts.

Because the amplitude collected from the belt can generate abnormal data due to field environment disturbance or communication signal loss and the like, the original vibration curve can be preprocessed when the motor is predicted to close the gate of the subway platform driven by the belt next time and before the amplitude of the belt is predicted, so that the original vibration curve meets the calculation specification.

In this example, if the communication signal is lost during the process of acquiring the amplitude of the belt when the gate of the subway platform driven by the motor through the belt is closed, the system will return an abnormal value (i.e. the first original abnormal amplitude), such as 65535, and the abnormal value (i.e. the first original abnormal amplitude) can be replaced by the amplitude of the belt when the belt is not running, such as 15000.

In another example of preprocessing, the amplitude corresponding to the disturbance anomaly is searched for in the original vibration curve, and as the second original anomaly amplitude, the amplitude adjacent to the second original anomaly amplitude (on the left side or on the right side) is substituted for the second original anomaly amplitude.

In this example, when the gate of the subway platform is driven by the motor through the belt to close and the amplitude of the belt is acquired, if the field environment has disturbance influence, the belt generates abnormal values (i.e. the second original abnormal amplitude) in a reasonable value-taking area, that is, the value-taking range of the abnormal values is within 0 to 30000, but the abnormal values obviously do not conform to the curve trend.

S802, inputting the n original vibration curves into a convolutional neural network respectively for processing so as to extract n original vibration characteristics respectively.

In this embodiment, a vibration prediction model may be trained in advance, and the vibration prediction model may be used to predict the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time according to the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt n times continuously.

The vibration prediction model may comprise a three-layer structure, and the directions of forward propagation are respectively:

1. convolutional Neural Network (CNN)

2. Long Short Term Memory network (Long Short-Term Memory, LSTM)

3. Deep Neural network (Deep Neural Networks, DNN)

It should be noted that the convolutional neural network is a general network, that is, for n motors continuously closing the gate of the subway platform through the belt n times and the original vibration curve of the belt, the same convolutional neural network is used to extract n original vibration features.

In one example, the convolutional neural network includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a Pooling layer (Pooling layer).

Wherein, the first convolution layer, the second convolution layer and the third convolution layer are convolution layers (Convolutional layers).

In this example, n raw vibration curves (x)₁，x₂，……，x_n-1，x_n) Inputting the first convolution layers respectively for convolution operation to output n first original candidate features respectively.

And inputting the n first original candidate features into the second convolution layer respectively for convolution operation so as to output n second original candidate features respectively.

And inputting the n second original candidate features into the third convolution layer respectively for convolution operation so as to output n third original candidate features respectively.

And inputting the n third original candidate features into the pooling layer respectively for pooling operation so as to output n original vibration features respectively.

And S803, sequentially inputting the n original vibration characteristics into n chain-dependent long-short term memory networks for processing so as to sequentially output the n original implicit characteristics.

The long-short term memory network is a special network, the original vibration curve and the long-short term memory network are in one-to-one correspondence, namely, the independent long-short term memory network is used for extracting the original implicit characteristics aiming at the original vibration curve of the belt when each motor drives the gate of the subway platform to close through the belt.

In a specific implementation, n original vibration features can be sequentially traversed according to a sequence, an original reserved feature output by a last long-short term memory network is determined according to the current original vibration feature, the current original vibration feature and the last original reserved feature are input into the current long-short term memory network for processing, so that the original reserved feature and an original implicit feature are output, the current original reserved feature is output to a next long-short term memory network, and the process is repeated until n LSTMs finish processing the n original vibration features.

LSTM uses two gates to control the contents of cell state c, oneA forgetting gate (forget gate) which determines the state c of the cell at the previous time t-1_t-1How much to keep the cell state c to the current time t_t(ii) a The other is an input gate (input gate), which determines the input x of the network at the present moment_tHow many cells to save to cell state c_t. LSTM uses output gate (output gate) to control cell state c_tHow much current output value h is output to LSTM_t。

For forgetting door f_t：

f_t＝σ(W_f·[h_t-1，x_t]+b_f)

For input gate i_t：

i_t＝σ(W_i·[h_t-1，x_t]+b_i)

For describing the cell state at the current time t

Wherein, the symbol

Meaning multiplication by element.

In this way, the LSTM is remembered with respect to the current

o_t＝σ(W_o·[h_t-1，x_t]+b_o)

Then, for the output gate o_tOutput h of_t：

In this embodiment, the current original vibration characteristic and the last original reserved characteristic may be input into the current long-short term memory network for processing, and the characteristic of forgotten gate output in the current long-short term memory network (i.e. the unit state c at the current time) may be determined_t) Determining output characteristics of output gates (i.e. output gate o) in current long-short term memory network for original reserved characteristics_tOutput h of_t) Is originally implicitAnd (5) characterizing.

S804, inputting the n original implicit characteristics into a deep neural network for processing so as to generate the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time, and obtaining a target vibration curve.

In a specific implementation, n original implicit features (h)₁，h₂，……，h_n-1，h_n) Inputting the vibration amplitude into DNN, mapping the vibration amplitude into the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time, and taking the vibration amplitude as a target vibration curve x_n+1Target vibration curve x_n+1Dimension number of (d) and original vibration curve (x)₁，x₂，……，x_n-1，x_n) Are the same in number.

In this embodiment, when the motor continuously closes the gate of the subway platform through the belt for n times, the amplitude of the belt is detected to obtain n original vibration curves, the n original vibration curves are respectively input into the convolutional neural network for processing to respectively extract n original vibration features, the n original vibration features are sequentially input into n chain-dependent long and short term memory networks for processing according to the sequence, so as to sequentially output n original implicit features, the n original implicit features are input into the deep neural network for processing to generate the amplitude of the belt when the motor closes the gate of the subway platform through the belt next time, a target vibration curve is obtained, the state of the belt is stable within a certain time range, namely, the vibration of the belt is correlated within a certain time range, and the original vibration curves belong to a position, a position, Data of long sequence, through the processing of long sequence data of the dimensionality reduction realization of convolution neural network, the characteristic through long short-term memory network extraction time dimension realizes the processing of data under the time dimension, when this maps out the gate that the motor next time passes through belt drive subway platform and closes, the amplitude of belt, can guarantee the accuracy of amplitude, thereby realize the state of real-time supervision belt when the gate is closed, help promoting the control effect of closing the gate, and can provide the unusual early warning of belt state when the gate is closed, reduce the probability that the belt broke down when the gate is closed, thereby promote the efficiency of overhauing the belt.

Example four

Fig. 10 is a flowchart of a method for training a vibration prediction model according to a fourth embodiment of the present invention, where the present embodiment is applicable to training the vibration prediction model to predict a subsequent vibration curve according to a previous vibration curve of a belt for a motor driving a gate of a subway platform to close through the belt, and the method may be executed by a training device of the vibration prediction model, where the vibration prediction device may be implemented by software and/or hardware and may be configured in a computer device, such as a server, a workstation, a personal computer, and the like, and the method specifically includes the following steps:

s1001, detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to close through the belt for n times, and obtaining n sample local oscillation curves.

When the vibration prediction model is trained, the amplitude of a belt when a motor continuously drives a gate of a subway platform to close for n times through the belt in a historical period of time can be extracted from a database to be used as n sample vibration curves.

Because the amplitude collected from the belt can generate abnormal data due to field environment disturbance or communication signal loss and the like, a sample vibration curve can be preprocessed before training a vibration prediction model, so that in a preprocessing example, the sample vibration curve meets the calculation specification.

S1002, detecting the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time, and obtaining a reference vibration curve.

In this embodiment, the amplitude of the belt when the motor drives the gate of the subway platform to close by the belt the next time (i.e., the (n + 1) th time) can be detected as a reference vibration curve, and the reference vibration curve can be regarded as a Tag (Tag) for the vibration prediction model.

And S1003, inputting the n sample vibration curves into a convolutional neural network respectively for processing so as to extract n sample vibration characteristics respectively.

And S1004, sequentially inputting the n sample vibration characteristics into n chain-dependent long-short term memory networks for processing according to the sequence, so as to sequentially output the n sample implicit characteristics.

S1005, inputting the implicit characteristics of the n samples into the deep neural network for processing to generate the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time, and obtaining a predicted vibration curve.

In this embodiment, the implicit characteristics of n samples are input into the DNN, and are mapped to the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time, and the amplitude is used as a predicted vibration curve, and the dimension number of the predicted vibration curve is the same as the dimension number of the sample vibration curve.

And S1006, training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve.

The reference vibration curve is an actual result, the predicted vibration curve is an estimated result, and the predicted vibration curve is compared with the reference vibration curve, so that the quality of the vibration prediction model can be evaluated, and the training of the vibration prediction model is guided.

In one embodiment of the present invention, S1006 may include the steps of:

s10061, calculating a difference between the predicted vibration curve and the reference vibration curve as a loss value.

In this embodiment, the predicted vibration curve may be compared with the reference vibration curve, and the difference between the predicted vibration curve and the reference vibration curve may be obtained as the LOSS value LOSS of the amplitude of the belt when the motor is next closed by the gate of the belt-driven subway platform.

In one example, as shown, the difference between the amplitude in the predicted vibration curve and the amplitude in the reference vibration curve may be calculated separately for each same time (i.e., the time at which the vibration is collected) as a point deviation value, and the average of all the point deviation values may be calculated as a LOSS value LOSS.

S10062, judging whether a stop condition is met; if so, S10063 is executed, and if not, S10064 is executed.

S10063, outputting a convolutional neural network, and taking n chain-dependent long-short term memory networks and a deep neural network as vibration prediction models.

When the stopping condition is met, the training vibration prediction model is determined to be completed, the structures and parameters (including weights) of the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network are stored, and when the parameters (including weights) are loaded, the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network can be used for predicting the amplitude of the belt when the motor drives the gate of the subway platform to be closed through the belt for n times continuously according to the amplitude of the belt when the motor drives the gate of the subway platform to be closed through the belt for the next time.

And S10064, updating the deep neural network, the n chain-dependent long-short term memory networks and the convolutional neural network based on the loss value, returning to execute S1003, S1004 and S1005, and entering the next training.

In the embodiment, the amplitude of the belt is detected when the motor continuously drives the gate of the subway platform to close through the belt for n times, so that n sample local oscillation curves are obtained; detecting the amplitude of a belt when a motor drives a gate of a subway platform to close through the belt next time, obtaining a reference vibration curve, respectively inputting n sample vibration curves into a convolutional neural network for processing, respectively extracting n sample vibration characteristics, sequentially inputting n sample vibration characteristics into n chain-dependent long-short term memory networks for processing according to the sequence, sequentially outputting n sample hidden characteristics, inputting the n sample hidden characteristics into a deep neural network for processing, generating the amplitude of the belt when the motor drives the gate of the subway platform to close through the belt next time, obtaining a predicted vibration curve, training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve, and training the belt in a certain time range, the state is relatively stable, i.e. the vibration is correlated in a certain time range, the sample vibration curve belongs to a long sequence of data correlated in the time dimension, the dimension reduction of the convolutional neural network is used for realizing the processing of long sequence data, the long and short term memory network is used for extracting the characteristics of the time dimension for realizing the processing of data under the time dimension, therefore, when the motor drives the gate of the subway platform to close through the belt next time, the amplitude of the belt is predicted, the vibration prediction model is trained, the performance of the vibration prediction model can be ensured, namely, the accuracy of the predicted amplitude is ensured, thereby realizing the real-time monitoring of the state of the belt when the gate is closed, being beneficial to improving the control effect of closing the gate, and can provide the unusual early warning of belt state when the gate is closed, reduce the gate and take place the probability of trouble when closing the belt to promote the efficiency of overhauing the belt.

EXAMPLE five

Fig. 11 is a schematic structural diagram of a vibration prediction apparatus according to a fifth embodiment of the present invention, where the apparatus may specifically include the following modules:

the system comprises an original vibration curve detection module 1101, a vibration detection module and a control module, wherein the original vibration curve detection module 1101 is used for detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to be opened or closed through the belt for n times so as to obtain n original vibration curves;

an original vibration feature extraction module 1102, configured to input the n original vibration curves into a convolutional neural network respectively for processing, so as to extract n original vibration features respectively;

an original implicit feature extraction module 1103, configured to sequentially input the n original vibration features into n chain-dependent long-term and short-term memory networks for processing, so as to sequentially output the n original implicit features;

and a target vibration curve prediction module 1104, configured to input the n original implicit features into a deep neural network for processing, so as to generate an amplitude of the belt when the motor drives a gate of the subway platform to open or close next time through the belt, so as to obtain a target vibration curve.

In one embodiment of the present invention, further comprising:

the first original abnormal amplitude searching module is used for searching the amplitude which accords with the service abnormality in the original vibration curve and taking the amplitude as a first original abnormal amplitude;

a first original anomaly amplitude replacement module for replacing the first original anomaly amplitude with an amplitude of the belt when not in operation;

and/or the presence of a gas in the gas,

the second original abnormal amplitude searching module is used for searching the amplitude which accords with the disturbance abnormality in the original vibration curve to be used as a second original abnormal amplitude;

a second original anomaly amplitude replacement module to replace the second original anomaly amplitude with an amplitude adjacent to the second original anomaly amplitude.

In one embodiment of the invention, the second original anomaly amplitude replacement module comprises:

the original vibration interval segmentation submodule is used for segmenting the original vibration curve into a plurality of intervals to obtain an original vibration interval;

the original statistical characteristic calculation submodule is used for calculating the average value and the standard deviation of the amplitude in the original vibration interval;

a first target value calculation submodule for adding a triple value of the standard deviation to the average value to obtain a first target value;

a second target value calculation submodule for subtracting a triple value of the standard deviation from the average value to obtain a second target value;

and the original amplitude comparison submodule is used for determining that the amplitude accords with the disturbance abnormity and is a second original abnormal amplitude if the amplitude in the original vibration interval is larger than the first target value or smaller than the second target value.

In one embodiment of the invention, the convolutional neural network comprises a first convolutional layer, a second convolutional layer, a third convolutional layer, a pooling layer;

the original vibration feature extraction module 1102 includes:

the first original candidate feature extraction submodule is used for inputting the n original vibration curves into a first convolution layer respectively to carry out convolution operation so as to output n first original candidate features respectively;

the second original candidate feature extraction submodule is used for inputting the n first original candidate features into a second convolution layer respectively to carry out convolution operation so as to output n second original candidate features respectively;

a third original candidate feature extraction sub-module, configured to input the n second original candidate features into a third convolution layer respectively for convolution operation, so as to output n third original candidate features respectively;

and the original vibration feature output sub-module is used for inputting the n third original candidate features into a pooling layer respectively for pooling operation so as to output the n original vibration features respectively.

In one embodiment of the present invention, the original implicit feature extraction module 1103 includes:

the original reserved characteristic determining submodule is used for sequentially traversing the n original vibration characteristics according to the sequence and determining the original reserved characteristics output by the previous long-term and short-term memory network aiming at the current original vibration characteristics;

and the original implicit characteristic output sub-module is used for inputting the current original vibration characteristic and the last original reserved characteristic into the current long-short term memory network for processing so as to output the original reserved characteristic and the original implicit characteristic.

In one embodiment of the present invention, the original implicit characteristic output sub-module comprises:

the original vibration characteristic input unit is used for inputting the current original vibration characteristic and the last original reserved characteristic into the current long-short term memory network for processing;

the original forgetting gate output unit is used for determining the characteristics output by the forgetting gate in the long-short term memory network at present as original reserved characteristics;

and the original output gate output unit is used for determining that the output characteristics of the output gates in the current long-short term memory network are original implicit characteristics.

The vibration prediction device provided by the embodiment of the invention can execute the vibration prediction method provided by any embodiment of the invention, and has corresponding functional modules and beneficial effects of the execution method.

EXAMPLE six

Fig. 12 is a schematic structural diagram of a training apparatus for a vibration prediction model according to a sixth embodiment of the present invention, where the apparatus may specifically include the following modules:

the system comprises a sample vibration curve detection module 1201, a signal processing module and a signal processing module, wherein the sample vibration curve detection module 1201 is used for detecting the amplitude of a belt when a motor continuously drives a gate of a subway platform to open or close through the belt for n times to obtain n sample vibration curves;

a reference vibration curve detection module 1202, configured to detect an amplitude of the belt when the motor drives a gate of the subway platform to open or close through the belt next time, so as to obtain a reference vibration curve;

a sample vibration feature extraction module 1203, configured to input the n sample vibration curves into a convolutional neural network respectively for processing, so as to extract n sample vibration features respectively;

a sample implicit feature extraction module 1204, configured to sequentially input the n sample vibration features into n chain-dependent long-term and short-term memory networks for processing, so as to sequentially output n sample implicit features;

a predicted vibration curve prediction module 1205, configured to input the n hidden sample features into a deep neural network for processing, so as to generate an amplitude of the belt when the motor drives a gate of the subway platform to open or close next time through the belt, and obtain a predicted vibration curve;

and the vibration prediction model training module 1206 is used for training the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network into a vibration prediction model according to the predicted vibration curve and the reference vibration curve.

In one embodiment of the present invention, further comprising:

the first sample abnormal amplitude searching module is used for searching the amplitude which accords with the service abnormality in the sample vibration curve and is used as the first sample abnormal amplitude;

a first sample anomaly amplitude replacement module for replacing the first sample anomaly amplitude with an amplitude of the belt when not in operation;

and/or the presence of a gas in the gas,

the second sample abnormal amplitude searching module is used for searching the amplitude which accords with the disturbance abnormality in the sample vibration curve to be used as the second sample abnormal amplitude;

a second sample anomaly amplitude replacement module to replace the second sample anomaly amplitude with an amplitude adjacent to the second sample anomaly amplitude.

In one embodiment of the invention, the second sample anomalous amplitude replacement module comprises:

the sample vibration interval segmentation submodule is used for segmenting the sample vibration curve into a plurality of intervals to obtain a sample vibration interval;

the sample statistical characteristic calculation submodule is used for calculating the average value and the standard deviation of the amplitude in the sample vibration interval;

a first sample value calculation submodule for adding a triple value of the standard deviation to the average value to obtain a first sample value;

a second sample value calculation submodule for subtracting a triple value of the standard deviation from the average value to obtain a second sample value;

and the sample amplitude comparison submodule is used for determining that the amplitude accords with the disturbance abnormity and is the second sample abnormity amplitude if the amplitude in the sample vibration interval is larger than the first sample value or smaller than the second sample value.

the sample vibration feature extraction module 1203 includes:

the first sample candidate feature extraction submodule is used for inputting the n sample vibration curves into a first convolution layer respectively for convolution operation so as to output n first sample candidate features respectively;

a second sample candidate feature extraction sub-module, configured to input the n first sample candidate features into a second convolution layer respectively for convolution operation, so as to output n second sample candidate features respectively;

a third sample candidate feature extraction sub-module, configured to input the n second sample candidate features into a third convolution layer respectively for convolution operation, so as to output n third sample candidate features respectively;

and the sample vibration characteristic output sub-module is used for inputting the n third sample candidate characteristics into a pooling layer respectively for pooling operation so as to output the n sample vibration characteristics respectively.

In one embodiment of the present invention, the sample implicit feature extraction module 1204 comprises:

the sample retention characteristic determination submodule is used for sequentially traversing the n sample vibration characteristics according to the sequence and determining the sample retention characteristics output by the last long-term and short-term memory network aiming at the current sample vibration characteristics;

and the sample implicit characteristic output submodule is used for inputting the current sample vibration characteristic and the last sample retention characteristic into the current long-short term memory network for processing so as to output the sample retention characteristic and the sample implicit characteristic.

In one embodiment of the present invention, the sample implicit characteristic output sub-module includes:

the sample vibration characteristic input unit is used for inputting the current sample vibration characteristic and the last sample retention characteristic into the current long-short term memory network for processing;

the sample forgetting gate output unit is used for determining the characteristics output by a forgetting gate in the long-short term memory network at present as sample retention characteristics;

and the sample output gate output unit is used for determining that the characteristic output by the output gate in the current long-term and short-term memory network is a sample implicit characteristic.

the neural network unit neglecting submodule is used for neglecting part of neural network units in the current long-term and short-term memory network in the training;

and the neural network unit extraction submodule is used for inputting the current sample vibration characteristics into the current long-term and short-term memory network and extracting the sample implicit characteristics by using the neural network units which are not ignored.

In one embodiment of the present invention, the vibration prediction model training module 1206 comprises:

a loss value calculation operator module for calculating a difference between the predicted vibration curve and the reference vibration curve as a loss value;

a stop condition judgment submodule for judging whether a stop condition is satisfied; if yes, calling a vibration prediction model output sub-module, and if not, calling a vibration prediction model update sub-module;

the vibration prediction model output submodule is used for outputting the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network as vibration prediction models;

and the vibration prediction model updating submodule is used for updating the deep neural network, the n long-term and short-term memory networks with chain dependency and the convolutional neural network based on the loss value, and returning and calling the sample vibration feature extraction module 1203, the sample implicit feature extraction module 1204 and the predicted vibration curve prediction module 1205.

In one embodiment of the invention, the loss value operator module comprises:

a point deviation value calculation unit for calculating, as a point deviation value, a difference value between the amplitude in the predicted vibration curve and the amplitude in the reference vibration curve, respectively, for each same time;

and the average value calculating unit is used for calculating the average value of all the point deviation values as a loss value.

In one embodiment of the present invention, the vibration prediction model update sub-module includes:

the non-neglect determining unit is used for determining the non-neglected neural network units in the long-term and short-term memory network in the training;

a non-ignored updating unit for updating the weights in the non-ignored neural network units.

a hidden layer random setting unit, configured to randomly set and input data of a hidden layer in the first long-short term memory network for the first long-short term memory network in the training;

and the network training unit is used for training the first long-short term memory network based on the data.

The training device for the vibration prediction model provided by the embodiment of the invention can execute the training method for the vibration prediction model provided by any embodiment of the invention, and has corresponding functional modules and beneficial effects of the execution method.

EXAMPLE seven

Fig. 13 is a schematic structural diagram of a computer device according to a seventh embodiment of the present invention. As shown in fig. 13, the computer apparatus includes a processor 1300, a memory 1301, a communication module 1302, an input device 1303, and an output device 1304; the number of the processors 1300 in the computer device may be one or more, and one processor 1300 is taken as an example in fig. 13; the processor 1300, the memory 1301, the communication module 1302, the input device 1303 and the output device 1304 in the computer apparatus may be connected by a bus or other means, and fig. 13 illustrates an example of connection by a bus.

The memory 1301 may be used as a computer-readable storage medium to store software programs, computer-executable programs, and modules, such as modules corresponding to the vibration prediction method and the training method of the vibration prediction model in this embodiment (for example, an original vibration curve detection module 1101, an original vibration feature extraction module 1102, an original implicit feature extraction module 1103, and a target vibration curve prediction module 1104 in the vibration prediction apparatus shown in fig. 11; and a sample vibration curve detection module 1201, a reference vibration curve detection module 1202, a sample vibration feature extraction module 1203, a sample implicit feature extraction module 1204, a predicted vibration curve prediction module 1205, and a vibration prediction model training module 1206 in the training apparatus of the vibration prediction model shown in fig. 12). The processor 1300 executes various functional applications and data processing of the computer device by running software programs, instructions and modules stored in the memory 1301, namely, implements the vibration prediction method and the training method of the vibration prediction model described above.

The memory 1301 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created according to use of the computer device, and the like. Further, the memory 1301 may include high speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device. In some examples, the memory 1301 can further include memory located remotely from the processor 1300, which can be connected to a computer device through a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

A communication module 1302, which can be used to establish a connection with and enable data interaction with an external device (e.g., keyboard, pointing device, display screen, etc.), can also be used to communicate with one or more devices that enable a user to interact with the computing device, and/or can communicate with any device (e.g., network card, modem, etc.) that enables the computing device to communicate with one or more other computing devices.

The input means 1303 may be used to receive input numeric or character information and generate key signal inputs related to user settings and function control of the computer device, and may also be a camera for acquiring images and a sound pickup device for acquiring audio data.

The output device 1304 may include an audio device such as a speaker or a display device such as a display screen.

It should be noted that the specific composition of the input device 1303 and the output device 1304 may be set according to actual situations.

The processor 1300 executes various functional applications of the device and data processing, that is, implements the vibration prediction method and the training method of the vibration prediction model described above by running software programs, instructions, and modules stored in the memory 1301.

The computer device provided by the embodiment of the invention can execute the vibration prediction method and the training method of the vibration prediction model provided by any embodiment of the invention, and has corresponding functions and beneficial effects.

Example eight

An eighth embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements a vibration prediction method or a training method for a vibration prediction model.

The vibration prediction method includes:

The training method of the vibration prediction model comprises the following steps:

Of course, the computer readable storage medium provided in the embodiments of the present invention, the computer program thereof is not limited to the method operations described above, and may also perform related operations in the vibration prediction method and the training method of the vibration prediction model provided in any embodiments of the present invention.

From the above description of the embodiments, it is obvious for those skilled in the art that the present invention can be implemented by software and necessary general hardware, and certainly, can also be implemented by hardware, but the former is a better embodiment in many cases. Based on such understanding, the technical solutions of the present invention may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a floppy disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a FLASH Memory (FLASH), a hard disk or an optical disk of a computer, and includes several instructions for enabling a computer device (which may be a personal computer, a server, or a network device) to execute the methods according to the embodiments of the present invention.

It should be noted that, in the embodiments of the vibration prediction model and the training apparatus for the vibration prediction model, the included units and modules are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, specific names of the functional units are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present invention.

It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A vibration prediction method, comprising:

2. The method according to claim 1, wherein after obtaining n original vibration curves from the amplitude of the belt when the detection motor continuously opens or closes the gate of the subway platform through the belt n times, the method further comprises:

searching the amplitude which accords with the service abnormity on the original vibration curve to be used as a first original abnormal amplitude;

replacing the first original anomaly amplitude with an amplitude of the belt when not in operation;

and/or the presence of a gas in the gas,

searching the amplitude which accords with disturbance abnormity on the original vibration curve to be used as a second original abnormal amplitude;

replacing the second original anomaly amplitude with an amplitude adjacent to the second original anomaly amplitude.

3. The method of claim 2, wherein said finding an amplitude corresponding to a disturbance anomaly in said raw vibration curve as a second raw anomaly amplitude comprises:

dividing the original vibration curve into a plurality of intervals to obtain an original vibration interval;

calculating the average value and the standard deviation of the amplitude in the original vibration interval;

adding a triple value of the standard deviation on the basis of the average value to obtain a first target value;

subtracting a triple value of the standard deviation on the basis of the average value to obtain a second target value;

and if the amplitude in the original vibration interval is larger than the first target value or smaller than the second target value, determining that the amplitude meets disturbance abnormity and is a second original abnormal amplitude.

4. The method of any one of claims 1-3, wherein the convolutional neural network comprises a first convolutional layer, a second convolutional layer, a third convolutional layer, a pooling layer;

the respectively inputting the n original vibration curves into a convolutional neural network for processing so as to respectively extract n original vibration features, including:

inputting the n original vibration curves into a first convolution layer respectively for convolution operation so as to output n first original candidate characteristics respectively;

inputting the n first original candidate features into a second convolution layer respectively for convolution operation so as to output n second original candidate features respectively;

inputting the n second original candidate features into a third convolution layer respectively for convolution operation so as to output n third original candidate features respectively;

and inputting the n third original candidate features into a pooling layer respectively for pooling operation so as to output n original vibration features respectively.

5. The method according to any one of claims 1-3, wherein the sequentially inputting the n original vibration features into the n chain-dependent long-short term memory networks for processing so as to sequentially output n original implicit features comprises:

sequentially traversing the n original vibration characteristics according to the sequence, and determining original reserved characteristics output by the last long-term and short-term memory network aiming at the current original vibration characteristics;

inputting the current original vibration characteristic and the last original reserved characteristic into the current long-short term memory network for processing so as to output the original reserved characteristic and the original implicit characteristic.

6. The method of claim 5, wherein inputting the current original vibration feature and the last original preserved feature into the current long-short term memory network for processing to output an original preserved feature and an original implicit feature comprises:

inputting the current original vibration characteristic and the last original reserved characteristic into the current long-short term memory network for processing;

determining the characteristics output by a forgetting gate in the long and short term memory network as original reserved characteristics;

and determining the characteristics of the output gate in the long-short term memory network as original implicit characteristics.

7. A method for training a vibration prediction model, comprising:

8. The method according to claim 7, wherein the sequentially inputting the n sample vibration characteristics into the n chain-dependent long-short term memory networks for processing so as to sequentially output n sample implicit characteristics comprises:

in the training, neglecting part of neural network units in the current long-term and short-term memory network;

and extracting the implicit characteristics of the samples from the current sample vibration characteristics by using the non-ignored neural network units input into the current long-short term memory network.

9. The method according to claim 7 or 8, wherein the training of the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network as vibration prediction models according to the predicted vibration profile and the reference vibration profile comprises:

calculating a difference between the predicted vibration curve and the reference vibration curve as a loss value;

judging whether a stop condition is met;

if so, outputting the convolutional neural network, the n chain-dependent long-short term memory networks and the deep neural network as vibration prediction models;

and if not, updating the deep neural network, the n chain-dependent long-short term memory networks and the convolutional neural network based on the loss value, and returning to execute the step of respectively inputting the n sample vibration curves into the convolutional neural network for processing so as to respectively extract n sample vibration characteristics.

10. The method of claim 9, wherein said calculating a difference between said predicted vibration profile and said reference vibration profile as a loss value comprises:

calculating a difference value between the amplitude in the predicted vibration curve and the amplitude in the reference vibration curve as a point deviation value, respectively, for each identical time;

calculating the average value of all the point deviation values as a loss value.

11. The method of claim 9, wherein the updating the deep neural network, the n chain-dependent long-short term memory networks, and the convolutional neural network comprises:

in the training, determining the neural network units which are not ignored in the long-term and short-term memory network at present;

updating weights in the non-ignored neural network elements.

12. The method of claim 9, wherein the updating the deep neural network, the n chain-dependent long-short term memory networks, and the convolutional neural network comprises:

in the training, aiming at the first long-short term memory network, randomly setting and inputting data of a hidden layer in the first long-short term memory network;

training the first long-short term memory network based on the data.

13. A vibration prediction apparatus, comprising:

14. An apparatus for training a vibration prediction model, comprising:

15. A computer device, characterized in that the computer device comprises:

one or more processors;

a memory for storing one or more programs;

when executed by the one or more processors, cause the one or more processors to implement a vibration prediction method as claimed in any one of claims 1 to 6 or a training method of a vibration prediction model as claimed in any one of claims 7 to 12.

16. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a method for vibration prediction according to any one of claims 1 to 6 or a method for training a vibration prediction model according to any one of claims 7 to 12.