CN112381308A

CN112381308A - Training of current prediction model, current prediction method and related device

Info

Publication number: CN112381308A
Application number: CN202011314509.5A
Authority: CN
Inventors: 刘文凯; 李鸿飞; 秦伟; 贾建平
Original assignee: Guangdong Huazhiyuan Information Engineering Co ltd; Guangzhou Huajia Software Co ltd; Guangzhou Jiadu Urban Rail Intelligent Operation And Maintenance Service Co ltd; Guangzhou Xinke Jiadu Technology Co Ltd
Current assignee: Guangdong Huazhiyuan Information Engineering Co ltd; Guangzhou Huajia Software Co ltd; Guangzhou Jiadu Urban Rail Intelligent Operation And Maintenance Service Co ltd; Guangzhou Xinke Jiadu Technology Co Ltd
Priority date: 2020-11-20
Filing date: 2020-11-20
Publication date: 2021-02-19

Abstract

The invention provides a training method of a current prediction model and a current prediction method and a related device thereof, wherein the training method of the current prediction model comprises the following steps: the method comprises the steps of detecting a current value of an electromagnet when the electromagnet is continuously electrified for multiple times to unlock a platform door, obtaining a sample current curve, inputting the sample current curve into a nonlinear network to extract a nonlinear characteristic under a time dimension as a sample nonlinear characteristic, inputting the sample current curve into a linear network to extract a linear characteristic under the time dimension as a sample linear characteristic, generating a current value of the electromagnet when the electromagnet is electrified for unlocking the platform door next time according to the sample nonlinear characteristic and the sample linear characteristic to obtain a predicted current curve, training the nonlinear network and the linear network into a current prediction model according to the predicted current curve, extracting the nonlinear sample nonlinear characteristic under the time dimension through the nonlinear network, extracting the linear sample characteristic under the time dimension through the linear network, and ensuring the performance of the current prediction model.

Description

Training of current prediction model, current prediction method and related device

Technical Field

The embodiment of the invention relates to the technical field of traffic, in particular to training of a current prediction model, a current prediction method and a related device.

Background

At present, in traffic systems such as subways and Bus Rapid Transit (BRT), a station is usually provided with a platform door along the edge of the platform, so that the platform is isolated from a driving area, the operation energy consumption of an air-conditioning ventilation system of the station is reduced, meanwhile, the influence of train operation noise and piston wind on the station is reduced, personnel are prevented from falling off a track to generate accidents, and a comfortable and safe waiting environment is provided for passengers.

When the train arrives at the station, the platform door is opened, passengers can get on or off the train through the gate, then the platform door is closed, and the train continues to run. When the platform door is closed, the locking switch fastens the platform door, when the platform door is started, an electromagnet is electrified, and the electromagnet executes armature motion to drive the locking switch to be opened, so that the platform door can slide.

When the electromagnet breaks down, the platform door can be caused to break down, the door can be closed, the safety circuit can break down, the manual door opening and closing device can not be unlocked, the locking mechanism can not be locked, and the like, so that the sliding door of the platform door can not be opened and closed, the passengers can normally take on and off the subway train, and the train operation can be influenced.

In order to ensure the normal operation of the electromagnet, the electromagnet is generally overhauled by adopting maintenance modes such as fault repair, scheduled repair and the like at present.

The repair with failure is a post-repair action performed to restore the electromagnet belt to a specified technical state after the electromagnet belt fails or is damaged, while the repair with due date is a regular repair action performed to keep the electromagnet in the specified state before the electromagnet belt fails, and is manual work, so that the cost is high, and the repair with failure is performed after the electromagnet belt fails and after the electromagnet belt fails, the repair with due date has a certain period, so that the probability of successfully preventing the electromagnet from failing is low, and the repair efficiency is low.

Disclosure of Invention

The embodiment of the invention provides a training method of a current prediction model, a current prediction method and a related device thereof, and aims to solve the problems of untimely maintenance of electromagnets in platform doors, high maintenance cost and low efficiency.

In a first aspect, an embodiment of the present invention provides a method for training a current prediction model, including:

detecting the current value of the electromagnet when the electromagnet is continuously electrified for multiple times to unlock the platform door, and obtaining a sample current curve;

inputting the sample current curve into a nonlinear network to extract nonlinear characteristics under a time dimension, and taking the nonlinear characteristics as sample nonlinear characteristics;

inputting the sample current curve into a linear network to extract linear characteristics under a time dimension, and taking the linear characteristics as sample linear characteristics;

generating a current value of the electromagnet when the platform door is unlocked by next electrification according to the sample nonlinear characteristic and the sample linear characteristic to obtain a predicted current curve;

and training the nonlinear network and the linear network into a current prediction model according to the predicted current curve.

In a second aspect, an embodiment of the present invention further provides a current prediction method, including:

detecting the current value of the electromagnet when the electromagnet is continuously electrified for multiple times to unlock the platform door, and obtaining an original current curve;

inputting the original current curve into a nonlinear network to extract nonlinear characteristics under a time dimension as original nonlinear characteristics;

inputting the original current curve into a linear network to extract linear characteristics under time dimension as original linear characteristics;

and generating a current value when the electromagnet is electrified next time to unlock the platform door according to the original nonlinear characteristic and the original linear characteristic to obtain a target current curve.

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

the sample current curve detection module is used for detecting the current value when the electromagnet is continuously electrified for multiple times to unlock the platform door, and obtaining a sample current curve;

the sample nonlinear feature extraction module is used for inputting the sample current curve into a nonlinear network to extract nonlinear features under time dimension as sample nonlinear features;

the sample linear feature extraction module is used for inputting the sample current curve into a linear network to extract linear features under time dimension as sample linear features;

the predicted current curve generating module is used for generating a current value when the electromagnet is electrified next time to unlock the platform door according to the sample nonlinear characteristic and the sample linear characteristic to obtain a predicted current curve;

and the current prediction model training module is used for training the nonlinear network and the linear network into a current prediction model according to the predicted current curve.

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

the original current curve detection module is used for detecting the current value when the electromagnet is continuously electrified for multiple times to unlock the platform door, and obtaining an original current curve;

the original nonlinear feature extraction module is used for inputting the original current curve into a nonlinear network to extract nonlinear features under time dimension as original nonlinear features;

the original linear feature extraction module is used for inputting the original current curve into a linear network to extract linear features under time dimension as original linear features;

and the target current curve generating module is used for generating a current value when the electromagnet is electrified next time to unlock the platform door according to the original nonlinear characteristic and the original linear characteristic to obtain a target current 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 method of training a current prediction model as described in the first aspect, or a method of current prediction 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 method for training a current prediction model according to the first aspect, or the method for current prediction according to the second aspect.

In this embodiment, the current value of the electromagnet when the electromagnet is continuously electrified for a plurality of times to unlock the platform door is detected, a sample current curve is obtained, the sample current curve is input into the nonlinear network to extract the nonlinear characteristic in the time dimension as the sample nonlinear characteristic, the sample current curve is input into the linear network to extract the linear characteristic in the time dimension as the sample linear characteristic, the current value of the electromagnet when the electromagnet is electrified for unlocking the platform door is generated according to the sample nonlinear characteristic and the sample linear characteristic to obtain a predicted current curve, the nonlinear network and the linear network are trained into a current prediction model according to the predicted current curve, the electromagnet is in a stable state in a certain time range, that is, the current value when the electromagnet is electrified has correlation in a certain time range, and the sample current curve belongs to long-sequence data related in the time dimension, the nonlinear characteristic of the nonlinear sample under the time dimension is extracted through the nonlinear network, the richness of the characteristic is guaranteed, the linear characteristic of the nonlinear sample under the time dimension is extracted through the linear network, the problem of local scaling is concerned, the nonlinear characteristic of the sample and the linear characteristic of the sample are complementary, the current prediction model is trained according to the result, the performance of the current prediction model can be guaranteed, namely the accuracy of the predicted current value is guaranteed, the state of the electromagnet when the platform door is opened is monitored in real time, the control effect of opening the platform door is improved, the abnormal early warning of the state of the electromagnet when the platform door is started is provided, the probability of the electromagnet failing when the platform door is started is reduced, and the efficiency of overhauling the electromagnet is improved.

Drawings

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

fig. 2 is a schematic diagram illustrating a current of an electromagnet according to an embodiment of the present invention;

fig. 3 is a schematic structural diagram of a current prediction model according to an embodiment of the present invention;

fig. 4 is a schematic structural diagram of a convolutional neural network according to an embodiment of the present invention;

FIG. 5 is a diagram illustrating an exemplary process of a recursive jump component according to an embodiment of the present invention;

FIG. 6 is a diagram illustrating an exemplary calculation of a loss value according to an embodiment of the present invention;

FIG. 7 is a flowchart of a current prediction method according to a second embodiment of the present invention;

fig. 8 is a schematic structural diagram of a training apparatus for a current prediction model according to a third embodiment of the present invention;

fig. 9 is a schematic structural diagram of a current prediction apparatus according to a fourth embodiment of the present invention;

fig. 10 is a schematic structural diagram of a computer device according to a fifth 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 training method for a current prediction model according to an embodiment of the present invention, where the embodiment is applicable to a case where a current of an electromagnet is segmented in time sequence as a sample Tag (Tag) to train the current prediction model, and the method may be performed by a training apparatus for the current prediction model, where the training apparatus for the current prediction model 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 specifically includes the following steps:

step 101, detecting a current value when the electromagnet is continuously electrified for multiple times to unlock the platform door, and obtaining a sample current curve.

The platform door of the traffic system such as subway, BRT and the like is provided with components such as an electromagnet, a motor, a belt and the like.

The electromagnet is a device which generates electromagnetism by electrifying, in the electromagnet, a conductive winding matched with the power of the iron core is wound outside the iron core, and the iron core is magnetized after the conductive winding is electrified, so that the iron core has magnetism.

The conductive windings are typically formed in a bar or shoe shape to make the core easier to magnetize. In addition, the conductive winding is usually made of soft iron or silicon steel material with fast demagnetization so as to realize fast demagnetization after power failure.

In the concrete implementation, after receiving a door opening command, a door controller DCU of the traffic system energizes an electromagnet, and the electromagnet executes armature action to drive a locking switch to act, so that the locking of a platform door is released.

One end of the belt is sleeved on the motor, the other end of the belt is sleeved on the platform door, the door controller DCU controls the motor to rotate to drive the belt to rotate, and the belt drives the platform door to move towards two sides so as to drive the platform door to be opened.

In this embodiment, the electromagnets are connected in series with a current meter, which detects the current value of the electromagnet when the platform door is unlocked by energizing in real time at a predetermined frequency (for example, once every 5 ms), and a specified number (for example, 300) of current values are acquired each time.

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 sample current curve is generated by establishing coordinate axes with time (time) as the horizontal axis and a numerical value (value) as the vertical axis, and marking the current values of the electromagnets at the time of unlocking the station door by energization in the form of points on the coordinate axes.

As shown in fig. 2, from the view of the operation mechanism of the device, when the platform door is opened, in order to quickly release the locking of the platform door, the current of the electromagnet is usually quickly raised to a rated value and continues for a short time, when the electromagnet armature locking switch is finished, the energization of the electromagnet is stopped, and the current value quickly drops to 0.

When the current prediction model is trained, the current value of the electromagnet which is continuously electrified for many times within a period of history and unlocks the platform door can be extracted from the database to be used as a sample current curve.

So-called continuous, the operation of unlocking the station door by electromagnet energization may be performed one after another, and assuming that the number of the sample current curves is n, the sample current curve of unlocking the station door by electromagnet energization 1 st time, the sample current curve of unlocking the station door by electromagnet energization 2 nd time, and the sample current curve of unlocking the station door by electromagnet energization 3 rd time may be extracted.

And 102, inputting the sample current curve into a nonlinear network to extract nonlinear characteristics under the time dimension as sample nonlinear characteristics.

In a specific implementation, the current prediction model includes a non-linear network, where the non-linear network has a function of extracting a characteristic of non-linearity in a time dimension, and the non-linearity may refer to a situation where an output and an input are neither directly proportional nor inversely proportional.

In one embodiment of the invention, a nonlinear network comprises:

1. convolutional Neural Network (CNN)

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.

2. Recurrent Neural Network (RNN)

The recurrent neural network is a deep learning model which takes sequence data as input for modeling, and can accept wider time sequence structure input for describing dynamic time behavior because the recurrent neural network circularly transmits states in the self network.

In this embodiment, step 102 includes the following steps:

and step 1021, inputting the sample current curve into the convolutional neural network to extract nonlinear characteristics in the time dimension, and taking the nonlinear characteristics as short-term time characteristics of the sample.

In the present embodiment, as shown in fig. 3, time-series data is formed from the sample current curve 301, and the time-series data is input to the convolutional neural network 311 to be processed, and short-term features in the time dimension are extracted as sample short-term time features by the local dependency relationship between variables.

Assuming 100 sample current curves as input, a batch size of 200, each sample current curve comprising 300 points, the data has dimensions [200, 100, 300 ].

In one example, as shown in fig. 4, the convolutional neural network includes a plurality of convolutional layers (ConV)401 and a plurality of active layers 402, and the number of convolutional layers 401 and active layers 402 may be set by those skilled in the art according to actual situations, for example, the number of convolutional layers 401 is 100, the number of active layers 402 is 100, and so on.

The convolutional layer 401 and the active layer 402 are connected in an interlaced manner, and in general, if the first layer of the convolutional neural network is the convolutional layer 401 and the last layer is the active layer 402, the convolutional neural network has the structure of the convolutional layer 401, the active layer 402, the convolutional layer 401, the active layers 402, … …, the convolutional layer 401, and the active layer 402.

The parameters of convolutional layer 401 are kept consistent, for example, the size of convolutional kernel is 5 × 5, the step size is 1, and the output dimension is 100.

The activation functions applied by the activation layer 402 are kept consistent, a reduce (Linear rectification function) function, a tanh function and the like can be applied as the activation functions, and the expression capability of the convolutional neural network is improved by adding a nonlinear factor.

In this example, the sample current curve may be sequentially input to the convolutional layer 401 to perform convolution processing, input to the active layer 402 to perform activation processing, and when all convolution processing and all activation processing are completed, a characteristic that is nonlinear in the time dimension may be output as a sample short-term time characteristic.

In this example, the convolutional neural network includes a plurality of convolutional layers and a plurality of active layers, and after convolutional processing of the convolutional layers and activation processing of the active layers, rich information in the sample current curve can be retained, so that damage to the information in the sample current curve caused by using a pooling layer and the like is avoided, accuracy of short-term time characteristics of the sample is improved, and performance of the current prediction model is improved.

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, a full link layer is set after an activation layer, and the like, which is not limited in this embodiment of the present invention. 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.

And step 1022, inputting the short-term time characteristics of the sample into the recurrent neural network to extract nonlinear characteristics under the time dimension, and taking the nonlinear characteristics as the nonlinear characteristics of the sample.

In this embodiment, the short-term temporal features of the samples are configured into time series data, which are input to a recurrent neural network for processing, and the long-term features in the time dimension are periodically extracted as the nonlinear features of the samples.

In one example, as shown in fig. 3, the recurrent neural network includes a recursion component 312, a first recursion jump component 313, and a second recursion jump component 314.

Where recursive component 312 is a recursive layer with Gated Recursive Units (GRUs), and updates the activation function using Relu, etc. as the hidden layer.

The GRU uses a gating mechanism to control the input, memory etc. information to make a prediction at the current time step, the GRU having two gates, a reset gate (reset gate) and an update gate (update gate) which determines how to combine the new input information with the previous memory, the update gate defining the amount of the previous memory saved to the current time step.

The recursive layer with GRUs can remember historical information to learn the relatively long-term dependencies, however, due to the gradient vanishing problem, GRUs often cannot capture very long-term dependencies.

While the first and second

recursive hop components

313, 314 are each a recursive hop component (RNN-skip), i.e. a recursive structure with time-hopping connections, i.e. hop links (skip) are added between the current hidden unit and the hidden unit of the same phase of the adjacent cycle to extend the time span of the information stream to capture very long-term dependencies.

Specifically, Skip gates are added to the neurons of the recurrent neural network, and the output of the Skip gates determines whether the neurons of the recurrent neural network perform state updating at the moment. When the output of the Skip gate is 0, the neuron does not update the state at that time, and retains the state at the previous time. When the output of the Skip gate is 1, the state of the neuron is updated at that time, and a new neuron state is generated. Therefore, after the Skip gate is added, the state updating of the neuron at certain time is skipped, the updating frequency of the neuron is reduced, and higher prediction accuracy is obtained.

In this example, the model parameters of the first recursive skip component 313 and the second recursive skip component 314 are different, for example, the parameter period length skip is 25 and the period number pt is 3 in the first recursive skip component, and the parameter period length skip is 10 and the period number pt is 9 in the second recursive skip component.

Wherein the period number pt is used to indicate how many data points with skip as a period will exist in the time window.

As shown in fig. 5, for example, pt is 2, skip is 7, and for a sequence vector 501 with a length of 2 × 7 is 14, where 7 is a period, and 2 is period data, that is, the sequence vector represents a state at the same time every 7 points, passes through a reshape (function of changing the structure of an array) 502 and a permute (function of rearranging an array according to a specified vector) 503, and is converted into a matrix 504 of dimension (7, 2), where data representing the state at the same time is located in the same row, each row represents a long period sample, the time step input to the recurrent neural network is 2, that is, the data at the same period is used as input, and the number of periods is used as the time step.

The input of each of the first recursive jump component 313 and the second recursive jump component 314 may be 100-dimensional data output by the convolutional neural network 311, and the output is 10-dimensional.

In this example, the sample short-term temporal features are input into the recursion component 312 to extract features that are non-linear in the time dimension as the first sample long-term temporal features.

The sample short-term temporal features are input into the first recursive skip component 313 to extract features that are non-linear in the time dimension as second sample long-term temporal features.

The sample short-term temporal features are input into the second recursive skip component 314 to extract features that are non-linear in the time dimension as third sample long-term temporal features.

The outputs of the recursive component, the first recursive jump component, and the second recursive jump component are combined by the network such as the fully connected layer, that is, the long-term time characteristic of the first sample, the long-term time characteristic of the second sample, and the long-term time characteristic of the third sample are input to the network such as the fully connected layer, and are fused into the sample nonlinear characteristic 315.

The input to the fully-connected layer includes the hidden state of the recursive component at time t of the timestamp

Recursively jumping p hidden states of a component from timestamp t-p +1 to time t

Marking as

The output of the full connection layer

Can be used forExpressed as:

wherein, W^RTo a hidden state

Set weight, W^STo a hidden state

The weight set, b is a constant.

It should be noted that, if the dimension of the non-linear network output (i.e., the sample non-linear feature) is not consistent with the dimension of the linear network output (i.e., the sample linear feature), the non-linear network output (i.e., the sample non-linear feature) may be mapped by the linear neurons, so that the dimension of the non-linear network output (i.e., the sample non-linear feature) is consistent with the dimension of the linear network output (i.e., the sample linear feature), which facilitates the fusion of the two.

For example, the dimension of the nonlinear network output (i.e., the sample nonlinear feature) is [350, 1000], and the dimension of the linear network output (i.e., the sample linear feature) is [200, 1000], then the nonlinear network output (i.e., the sample nonlinear feature) can be mapped to [200, 1000] using linear neurons.

In the embodiment, the recursive component is used for extracting the relatively long-term dependency relationship from the sample short-term time characteristic, the first recursive jump component and the second recursive jump component are respectively used for extracting the very long-term correlation from the sample short-term time characteristic, and the diversity of the long-term characteristic in the time dimension can be enriched in a complementary form, so that the accuracy of the sample nonlinear characteristic is improved, and the performance of the current prediction model is improved.

Of course, the structure of the recurrent neural network is only an example, and when the embodiment of the present invention is implemented, other structures of the recurrent neural network may be set according to actual situations, for example, LSTM (Long Short Term Memory) is used instead of GRU, and the like, which is not limited by the embodiment of the present invention. In addition, besides the above structure of the recurrent neural network, a person skilled in the art may also adopt other structures of the recurrent neural network according to actual needs, and the embodiment of the present invention is not limited thereto.

In one embodiment of the invention, a dropout mechanism is added to the recurrent neural network to prevent overfitting.

For the dropout mechanism, for other layers except the last layer in the recurrent neural network, such as the other layers except the last layer in the first recursive jump component and the second recursive jump component, in the training (forward propagation), a part of neural network units in the current recurrent neural network can be ignored through a mask or the like with a specified probability (for example, when dropout is 0.5, the probability is 50%).

Inputting the short-term time characteristics of the sample into a recurrent neural network, and extracting the nonlinear characteristics under the time dimension by using the non-ignored neural network units as the nonlinear characteristics of the sample.

And 103, inputting the sample current curve into a linear network to extract linear characteristics under the time dimension, and taking the linear characteristics as sample linear characteristics.

Due to the nonlinear characteristics of nonlinear networks such as a convolutional neural network and a cyclic neural network (such as a recursive component, a first recursive jump component and a second recursive jump component), the nonlinear networks are insensitive to the scale of input in prediction, that is, small-range fluctuation cannot be identified, so that a linear network can be arranged in a current prediction model, the problem of local scaling is concerned, and the prediction accuracy of the current prediction model is improved.

And inputting the sample current curve into a linear network for processing, and extracting linear characteristics under the time dimension to be used as sample linear characteristics.

In an embodiment of the present invention, as shown in fig. 3, the linear network includes an Autoregressive Model (AR) 316, and the Autoregressive Model 316 is a process using itself as a regression variable, that is, a linear regression Model using linear combinations of random variables at a plurality of previous moments to describe random variables at a later moment, and can be used for processing time series.

In this embodiment, the sample current curve 301 may be input into the autoregressive model 316 to linearly predict the current value when the electromagnet is next energized to release the platform door lock, and the linear characteristic of the sample is expressed as follows:

wherein, X_tThe current value X when the electromagnet is electrified for the t time to release the locking of the platform door_t-tC is a constant term, which is the current value when the electromagnet is electrified for the t-1 th time to unlock the platform door,

is a natural correlation number, epsilon_tTo assume that the mean is equal to 0, the standard deviation is equal to the random error value of σ, which is invariant for any t.

At this time, the current value (namely the expected value) when the electromagnet is electrified next time to unlock the platform door is equal to the linear combination of the current values when the electromagnet is electrified for unlocking the platform door for a plurality of times continuously, a constant term and a random error are added, and because the data required by autoregressive is less, the current value can be predicted by using a self sample current curve, the calculation amount is low, and the operation is simple and convenient.

In the present embodiment, the number of output nodes of the AR channel is 25.

Of course, the above linear networks are only examples, and when implementing the embodiment of the present invention, other linear networks may be set according to practical situations, for example, a moving average Model (MA), an autoregressive moving average model (ARMA), a differential autoregressive moving average model (ARIMA), and the like, which is not limited in this embodiment of the present invention. In addition, besides the linear network, those skilled in the art may also adopt other linear networks according to actual needs, and the embodiment of the present invention is not limited to this.

And 104, generating a current value when the electromagnet is electrified next time to unlock the platform door according to the sample nonlinear characteristic and the sample linear characteristic, and obtaining a predicted current curve.

In the present embodiment, as shown in fig. 3, a fusion operation 317 may be performed on the sample nonlinear characteristic and the sample linear characteristic, so as to generate a current value when the electromagnet is next energized to unlock the platform door, as a predicted current curve.

In specific implementation, the dimension of the sample nonlinear feature is generally the same as that of the sample linear feature, the sample nonlinear feature and the sample linear feature can be spliced according to the dimension to serve as a sample comprehensive feature, and the sample comprehensive feature is input into a preset mapping network (such as a full connection layer), so that the sample comprehensive feature is mapped to be a current value when the electromagnet is electrified next time to unlock a platform door, and a predicted current curve is obtained.

And 105, training the nonlinear network and the linear network into a current prediction model according to the predicted current curve.

In a specific implementation, the predicted current curve is an estimated result, and may be used to evaluate the quality of the current prediction model (i.e., the nonlinear network and the linear network), so as to guide the training of the current prediction model (i.e., the nonlinear network and the linear network).

In one embodiment of the present invention, step 105 comprises the steps of:

and 1051, detecting the current value of the electromagnet when the electromagnet is electrified next time to unlock the platform door, and obtaining a reference current curve.

In this embodiment, the current value of the electromagnet at the next (i.e. the (n + 1) th time) of energization unlocking of the platform door can be detected as a reference current curve, and the reference current curve can be regarded as a Tag (Tag) of a sample (i.e. a sample current curve) for the current prediction model.

Step 1052, calculating the difference between the predicted current profile and the reference current profile as the loss value.

The reference current curve is an actual result, the predicted current curve is an estimated result, the predicted current curve is compared with the reference current curve, and the difference between the predicted current curve and the reference current curve can be obtained and used as a LOSS value LOSS of the current value of the electromagnet when the platform door is unlocked in the next electrifying process.

In one example, as shown in fig. 6, a difference between a current value 601 (a darker portion) in the predicted current curve and a current value 602 (a lighter portion) in the reference current curve may be calculated as point deviation values, respectively, for each same time (i.e., a time when the current is collected), and an average value of all the point deviation values may be calculated as a loss value.

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.

Step 1053, judging whether a stop condition is satisfied; if yes, go to step 1054, otherwise go to step 1055.

And 1054, outputting the nonlinear network and the linear network as a current prediction model.

And 1055, updating the nonlinear network and the linear network, returning to execute the step 102, the step 103 and the step 104, and entering the next iterative training.

In this embodiment, a stop condition, such as the LOSS value LOSS being less than a preset threshold, the number of training iterations being reached, etc., may be set in advance.

When the stopping condition is met, the training current prediction model is determined to be completed, and the structures and parameters (including the weights in the neural network unit) of the nonlinear network and the linear network are stored, and when the parameters (including the weights in the neural network unit) are loaded, the nonlinear network and the linear network can be used for predicting the current value when the electromagnet is electrified next time to unlock the platform door according to the current value when the electromagnet is electrified for multiple times continuously to unlock the platform door.

When the stopping condition is not met, the current prediction model continues to be trained, and in this embodiment, back propagation may be performed based on the LOSS value LOSS, and parameters of the nonlinear network and the linear network (including weights in the neural network unit) may be updated in a random gradient descent (SGD) manner or the like.

In one embodiment of the present invention, the nonlinear network includes a convolutional neural network and a cyclic neural network, and a dropout mechanism may be added to the cyclic neural network (e.g., the first recursive skip component and the second recursive skip component) to prevent overfitting.

For the dropout mechanism, in the training (back propagation), for other layers except the last layer in the recurrent neural network (such as the first recurrent skip component and the second recurrent skip component), non-ignored neural network units in the recurrent neural network are determined, and the weights in the non-ignored neural network units are updated in a mask mode and the like.

In this embodiment, if the current prediction model (the nonlinear network and the linear network) is trained, the current prediction model (the nonlinear network and the linear network) may be tested, and the test index is the coefficient R.

The coefficient of determinability, also called the measure coefficient, the coefficient of determinability, and the index of determinability, represents a numerical characteristic of the relationship between a random variable and a plurality of random variables, and a statistical index for reflecting the reliability of the regression model to account for the variation of the dependent variable, which can be defined as the ratio of the variation of the independent variable to the total variation of the independent variable.

The coefficients of determination R tested were as follows:

96.02％、95.66％、96.75％、95.37％、96.60％、91.74％、96.98％、95.61％、96.78％、91.76％、96.53％、89.86％、97.40％、93.02％、97.37％、97.04％、95.90％、96.01％、96.53％、96.88％、95.84％、95.61％、97.51％。

Example two

Fig. 7 is a flowchart of a current prediction method according to a second embodiment of the present invention, which is applicable to a case where a future current is predicted based on a previous current of an electromagnet for a platform door opening, and the method may be executed by a current prediction apparatus, which 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 specifically includes the following steps:

step 701, detecting a current value when the electromagnet is continuously electrified for multiple times to unlock the platform door, and obtaining an original current curve.

In practical application, if the current value of the electromagnet when the electromagnet is electrified next time to unlock the platform door is predicted in real time, the current value of the electromagnet when the electromagnet is continuously electrified for multiple times in a period of time before the current time to unlock the platform door can be extracted from the database and used as an original current curve.

Step 702, inputting the original current curve into a nonlinear network to extract nonlinear characteristics under the time dimension as original nonlinear characteristics.

In a particular implementation, a current prediction model has been trained in advance, the current prediction model including a non-linear network and a linear network.

The original current curve is input into a nonlinear network for processing, and nonlinear features under the time dimension can be extracted and used as original nonlinear features.

In an embodiment of the present invention, the non-linear network includes a convolutional neural network and a cyclic neural network, and then in this embodiment, the step 702 includes the following steps:

step 7021, the original current curve is input into a convolutional neural network to extract nonlinear features in the time dimension, and the nonlinear features are used as original short-term time features.

In the present embodiment, time series data is composed of raw current curves, and is input to a convolutional neural network for processing, and short-term features in a time dimension are extracted as raw short-term time features through local dependency relationships between variables.

In one example, the convolutional neural network comprises a plurality of convolutional layers and a plurality of active layers, and the convolutional layers and the active layers are connected in an interlaced manner, so in this example, the original current curve can be sequentially input into the convolutional layers to perform convolutional processing and input into the active layers to perform active processing; when all convolution processing and all activation processing are completed, the nonlinear characteristic in the time dimension is output as the original short-term time characteristic.

In this example, the convolutional neural network includes a plurality of convolutional layers and a plurality of active layers, and after convolutional processing of the convolutional layers and activation processing of the active layers, rich information in the original current curve can be retained, so that damage to information in the original current curve caused by using a pooling layer and the like is avoided, accuracy of original short-term time characteristics is improved, and accuracy of current prediction is improved.

Step 7022, inputting the original short-term time feature into a recurrent neural network to extract a nonlinear feature in a time dimension, and taking the nonlinear feature as an original nonlinear feature.

In this embodiment, the original short-term temporal features are configured into time series data, which are input into a recurrent neural network for processing, and long-term features in the time dimension are periodically extracted as original nonlinear features.

In one embodiment of the invention, a recurrent neural network includes a recurrent component, a first recurrent skip component, a second recurrent skip component; in this embodiment, the original short-term time features are input into the recursive component to extract nonlinear features in the time dimension, which are used as first original long-term time features; inputting the original short-term time characteristic into a first recursive jump component to extract a nonlinear characteristic under a time dimension as a second original long-term time characteristic; inputting the original short-term time characteristic into a second recursive jump component to extract a nonlinear characteristic under the time dimension as a third original long-term time characteristic; and fusing the first original long-term time characteristic, the second original long-term time characteristic and the third original long-term time characteristic into an original nonlinear characteristic.

In the embodiment, the recursion component is used for extracting the relatively long-term dependency relationship from the original short-term time characteristic, the first recursion jump component and the second recursion jump component are respectively used for extracting the very long-term correlation from the original short-term time characteristic, and the diversity of the long-term characteristic in the time dimension can be enriched in a complementary mode, so that the accuracy of the original nonlinear characteristic is improved, and the accuracy of current prediction is improved.

And 703, inputting the original current curve into a linear network to extract linear characteristics under the time dimension as original linear characteristics.

In this embodiment, the original current curve is input into a linear network for processing, and a linear feature in a time dimension can be extracted as an original linear feature.

In one example, the linear network comprises an autoregressive model, and the original current curve is input into the autoregressive model to linearly predict the current value when the electromagnet is electrified next time to unlock the platform door, so that the current value is used as the original linear characteristic.

Step 704, generating a current value when the electromagnet is powered on next time to unlock the platform door according to the original nonlinear characteristic and the original linear characteristic, and obtaining a target current curve.

In this embodiment, a fusion operation may be performed on the original nonlinear characteristic and the original linear characteristic, so as to generate a current value when the electromagnet is next powered on to unlock the platform door, as the target current curve.

In specific implementation, the dimensionality of the original nonlinear features is generally the same as the dimensionality of the original linear features, and the original nonlinear features and the original linear features can be spliced according to the dimensionality to serve as original comprehensive features; and mapping the original comprehensive characteristics into a current value when the electromagnet is electrified next time to unlock the platform door, and obtaining a target current curve.

In the embodiment of the present invention, since the detection of the original current curve (step 701), the extraction of the original nonlinear feature (step 702), the extraction of the original linear feature (step 703), and the calculation of the target current curve (step 704) are basically similar to the application of the first embodiment, the description is relatively simple, and the relevant points can be referred to the partial description of the first embodiment, which is not described in detail herein.

In this embodiment, the current value of the electromagnet when the locking of the platform door is released by continuously energizing the electromagnet for a plurality of times is detected to obtain an original current curve, the original current curve is input into a nonlinear network to extract a nonlinear characteristic in a time dimension as an original nonlinear characteristic, the original current curve is input into a linear network to extract a linear characteristic in the time dimension as an original linear characteristic, the current value of the electromagnet when the platform door is released by energizing the electromagnet for the next time is generated according to the original nonlinear characteristic and the original linear characteristic to obtain a target current curve, the state of the electromagnet is stable in a certain time range, that is, the current value of the electromagnet when the electromagnet is energized has correlation in a certain time range, the original current curve belongs to long-sequence data related in the time dimension, the nonlinear original nonlinear characteristic in the time dimension is extracted through the nonlinear network, the method has the advantages that the richness of the characteristics is guaranteed, the linear original linear characteristics under the time dimension are extracted through the linear network, the problem of local scaling is concerned, the original nonlinear characteristics and the original linear characteristics are complementary, a target current curve is generated, the accuracy of the predicted current value can be guaranteed, the state of the electromagnet when the platform door is opened is monitored in real time, the control effect of opening the platform door is improved, the abnormal early warning of the state of the electromagnet when the platform door is started is provided, the probability of the electromagnet breaking down when the platform door is started is reduced, and the efficiency of overhauling the electromagnet is improved.

It should be noted that, for simplicity of description, the method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present invention is not limited by the illustrated order of acts, as some steps may occur in other orders or concurrently in accordance with the embodiments of the present invention. Further, those skilled in the art will appreciate that the embodiments described in the specification are presently preferred and that no particular act is required to implement the invention.

EXAMPLE III

Fig. 8 is a block diagram of a structure of a training apparatus for a current prediction model according to a third embodiment of the present invention, which may specifically include the following modules:

a sample current curve detection module 801, configured to detect a current value when the electromagnet is continuously powered on for multiple times to unlock the platform door, and obtain a sample current curve;

a sample nonlinear feature extraction module 802, configured to input the sample current curve into a nonlinear network to extract a nonlinear feature in a time dimension, which is used as a sample nonlinear feature;

a sample linear feature extraction module 803, configured to input the sample current curve into a linear network to extract a linear feature in a time dimension, which is used as a sample linear feature;

a predicted current curve generating module 804, configured to generate a current value when the electromagnet is next powered on to unlock the platform door according to the sample nonlinear characteristic and the sample linear characteristic, so as to obtain a predicted current curve;

and a current prediction model training module 805 configured to train the nonlinear network and the linear network into a current prediction model according to the predicted current curve.

In one embodiment of the invention, the nonlinear network comprises a convolutional neural network, a cyclic neural network;

the sample nonlinear feature extraction module 802 includes:

the sample short-term time characteristic generation submodule is used for inputting the sample current curve into a convolutional neural network to extract nonlinear characteristics under the time dimension as sample short-term time characteristics;

and the sample nonlinear feature generation submodule is used for inputting the short-term time features of the samples into the recurrent neural network to extract nonlinear features under the time dimension, and the nonlinear features are used as sample nonlinear features.

In one embodiment of the invention, the convolutional neural network comprises a plurality of convolutional layers, a plurality of active layers, and the convolutional layers are connected with the active layers in an interleaving manner;

the sample short-term temporal feature generation submodule comprises:

the convolution activation processing unit is used for sequentially inputting the sample current curve into the convolution layer to execute convolution processing and inputting the sample current curve into the activation layer to execute activation processing;

and the sample short-term time characteristic output unit is used for outputting a nonlinear characteristic in a time dimension as a sample short-term time characteristic when all the convolution processing and all the activation processing are executed.

In one embodiment of the invention, the recurrent neural network comprises a recurrent component, a first recurrent skip component, a second recurrent skip component;

the sample nonlinear feature generation submodule includes:

the first sample long-term time characteristic extraction unit is used for inputting the sample short-term time characteristic into the recursive component to extract nonlinear characteristics under the time dimension as a first sample long-term time characteristic;

the second sample long-term time characteristic extraction unit is used for inputting the sample short-term time characteristic into the first recursive jump component to extract nonlinear characteristics under the time dimension as a second sample long-term time characteristic;

a third sample long-term time feature extraction unit, configured to input the sample short-term time feature into a second recursive skip component to extract a nonlinear feature in a time dimension, which is used as a third sample long-term time feature;

and the sample long-term time characteristic fusion unit is used for fusing the first sample long-term time characteristic, the second sample long-term time characteristic and the third sample long-term time characteristic into a sample nonlinear characteristic.

In one embodiment of the present invention, the sample nonlinear feature generation submodule includes:

the neural network unit neglecting unit is used for neglecting part of the neural network units with a specified probability for other layers except the last layer in the recurrent neural network;

and the neural network unit extraction unit is used for inputting the sample short-term time characteristics into the recurrent neural network and extracting the characteristics of nonlinearity under the time dimension by using the non-ignored neural network units as sample nonlinear characteristics.

In one embodiment of the invention, the linear network comprises an autoregressive model;

the sample linear feature extraction module 803 includes:

and the sample autoregressive submodule is used for inputting the sample current curve into an autoregressive model to linearly predict the current value when the electromagnet is electrified next time to unlock the platform door, and the current value is used as the sample linear characteristic.

In one embodiment of the present invention, the predicted current curve generating module 804 includes:

the sample comprehensive characteristic generating submodule is used for splicing the sample nonlinear characteristic and the sample linear characteristic to be used as a sample comprehensive characteristic;

and the sample comprehensive characteristic mapping submodule is used for mapping the sample comprehensive characteristic to a current value when the electromagnet is electrified next time to unlock the platform door, so as to obtain a predicted current curve.

In one embodiment of the present invention, the current prediction model training module 805 comprises:

the reference current curve detection submodule is used for detecting the current value of the electromagnet when the platform door is unlocked after the electromagnet is electrified next time, and obtaining a reference current curve;

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

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

the current prediction model output submodule is used for outputting the nonlinear network and the linear network as a current prediction model;

and the current prediction model updating submodule is used for updating the nonlinear network and the linear network and returning to call the sample nonlinear feature extraction module 802.

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

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

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 invention, the nonlinear network comprises a convolutional neural network, a cyclic neural network; the current prediction model update sub-module includes:

the neural network unit determining unit is used for determining the neural network units which are not ignored in the recurrent neural network;

a neural network unit updating unit for updating the weights in the neural network units that are not ignored.

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

Example four

Fig. 9 is a block diagram of a current prediction apparatus according to a fourth embodiment of the present invention, which may specifically include the following modules:

an original current curve detection module 901, configured to detect a current value when the electromagnet is continuously energized for multiple times to unlock the platform door, so as to obtain an original current curve;

an original nonlinear feature extraction module 902, configured to input the original current curve into a nonlinear network to extract a nonlinear feature in a time dimension, which is used as an original nonlinear feature;

an original linear feature extraction module 903, configured to input the original current curve into a linear network to extract a linear feature in a time dimension, where the linear feature is used as an original linear feature;

and a target current curve generating module 904, configured to generate a current value when the electromagnet is powered on next time to unlock the platform door according to the original nonlinear characteristic and the original linear characteristic, so as to obtain a target current curve.

the original nonlinear feature extraction module 902 includes:

the original short-term time characteristic generation submodule is used for inputting the original current curve into a convolutional neural network to extract nonlinear characteristics under the time dimension as original short-term time characteristics;

and the original nonlinear feature generation submodule is used for inputting the original short-term time features into the recurrent neural network to extract nonlinear features under the time dimension, and the nonlinear features are used as original nonlinear features.

the original short-term time characteristic generation submodule comprises:

the convolution activation processing unit is used for sequentially inputting the original current curve into the convolution layer to execute convolution processing and inputting the original current curve into the activation layer to execute activation processing;

and the original short-term time characteristic output unit is used for outputting a nonlinear characteristic in a time dimension as an original short-term time characteristic when all the convolution processing and all the activation processing are executed.

the original nonlinear feature generation submodule comprises:

the first original long-term time feature extraction unit is used for inputting the original short-term time features into a recursive component to extract nonlinear features under a time dimension as first original long-term time features;

the second original long-term time characteristic extraction unit is used for inputting the original short-term time characteristic into the first recursive jump component to extract nonlinear characteristics under the time dimension as a second original long-term time characteristic;

a third original long-term time feature extraction unit, configured to input the original short-term time feature into a second recursive skip component to extract a nonlinear feature in a time dimension, which is used as a third original long-term time feature;

and the original long-term time feature fusion unit is used for fusing the first original long-term time feature, the second original long-term time feature and the third original long-term time feature into an original nonlinear feature.

the original linear feature extraction module 903 includes:

and the original autoregressive submodule is used for inputting the original current curve into an autoregressive model to linearly predict the current value when the electromagnet is electrified next time to unlock the platform door, and the current value is used as an original linear characteristic.

In one embodiment of the present invention, the target current curve generating module 904 comprises:

an original comprehensive characteristic generation submodule, configured to splice the original nonlinear characteristic and the original linear characteristic to serve as an original comprehensive characteristic;

and the original comprehensive characteristic mapping submodule is used for mapping the original comprehensive characteristic to a current value when the electromagnet is electrified next time to unlock the platform door, so as to obtain a predicted current curve.

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

EXAMPLE five

Fig. 10 is a schematic structural diagram of a computer device according to a fifth embodiment of the present invention. FIG. 10 illustrates a block diagram of an exemplary computer device 12 suitable for use in implementing embodiments of the present invention. The computer device 12 shown in FIG. 10 is only an example and should not bring any limitations to the functionality or scope of use of embodiments of the present invention.

As shown in FIG. 10, computer device 12 is embodied in the form of a general purpose computing device. The components of computer device 12 may include, but are not limited to: one or more processors or processing units 16, a system memory 28, and a bus 18 that couples various system components including the system memory 28 and the processing unit 16.

Bus 18 represents one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures. By way of example, such architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, micro-channel architecture (MAC) bus, enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

Computer device 12 typically includes a variety of computer system readable media. Such media may be any available media that is accessible by computer device 12 and includes both volatile and nonvolatile media, removable and non-removable media.

The system memory 28 may include computer system readable media in the form of volatile memory, such as Random Access Memory (RAM)30 and/or cache memory 32. Computer device 12 may further include other removable/non-removable, volatile/nonvolatile computer system storage media. By way of example only, storage system 34 may be used to read from and write to non-removable, nonvolatile magnetic media (not shown in FIG. 10, and commonly referred to as a "hard drive"). Although not shown in FIG. 10, a magnetic disk drive for reading from and writing to a removable, nonvolatile magnetic disk (e.g., a "floppy disk") and an optical disk drive for reading from or writing to a removable, nonvolatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 by one or more data media interfaces. Memory 28 may include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the invention.

A program/utility 40 having a set (at least one) of program modules 42 may be stored, for example, in memory 28, such program modules 42 including, but not limited to, an operating system, one or more application programs, other program modules, and program data, each of which examples or some combination thereof may comprise an implementation of a network environment. Program modules 42 generally carry out the functions and/or methodologies of the described embodiments of the invention.

Computer device 12 may also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), with one or more devices that enable a user to interact with computer device 12, and/or with any devices (e.g., network card, modem, etc.) that enable computer device 12 to communicate with one or more other computing devices. Such communication may be through an input/output (I/O) interface 22. Also, computer device 12 may communicate with one or more networks (e.g., a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with the other modules of computer device 12 via bus 18. It should be understood that although not shown in the figures, other hardware and/or software modules may be used in conjunction with computer device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.

The processing unit 16 executes programs stored in the system memory 28 to perform various functional applications and data processing, such as a training method for implementing a current prediction model or a current prediction method provided by any of the embodiments of the present invention.

EXAMPLE six

A sixth embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method for training a current prediction model or each process of the current prediction method provided in any of the above embodiments is implemented, and the same technical effects can be achieved, and are not described herein again to avoid repetition.

A computer readable storage medium may include, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

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 method for training a current prediction model, comprising:

2. The method of claim 1, wherein the non-linear network comprises a convolutional neural network, a cyclic neural network;

inputting the sample current curve into a nonlinear network to extract nonlinear characteristics under a time dimension, wherein the nonlinear characteristics are taken as sample nonlinear characteristics, and the method comprises the following steps:

inputting the sample current curve into a convolutional neural network to extract nonlinear characteristics under a time dimension as sample short-term time characteristics;

and inputting the short-term time characteristics of the sample into a recurrent neural network to extract nonlinear characteristics under the time dimension, and taking the nonlinear characteristics as sample nonlinear characteristics.

3. The method of claim 2, wherein the convolutional neural network comprises a plurality of convolutional layers, a plurality of active layers, the convolutional layers being interleaved with the active layers;

inputting the sample current curve into a convolutional neural network to extract nonlinear characteristics under a time dimension as sample short-term time characteristics, wherein the method comprises the following steps:

inputting the sample current curve into a convolution layer to execute convolution processing and inputting the sample current curve into an activation layer to execute activation processing in sequence;

and outputting a characteristic of nonlinearity in a time dimension as a sample short-term time characteristic when all the convolution processing and all the activation processing are completed.

4. The method of claim 2, wherein the recurrent neural network comprises a recurrent component, a first recurrent skip component, a second recurrent skip component;

inputting the sample short-term time features into a recurrent neural network to extract nonlinear features under a time dimension, wherein the sample short-term time features are used as sample nonlinear features, and the method comprises the following steps:

inputting the short-term time characteristic of the sample into a recursive component to extract a nonlinear characteristic under a time dimension as a long-term time characteristic of the first sample;

inputting the sample short-term time characteristic into a first recursive jump component to extract a nonlinear characteristic under a time dimension, and taking the nonlinear characteristic as a second sample long-term time characteristic;

inputting the short-term time characteristic of the sample into a second recursive jump component to extract a nonlinear characteristic under a time dimension, and taking the nonlinear characteristic as a long-term time characteristic of a third sample;

and fusing the first sample long-term time characteristic, the second sample long-term time characteristic and the third sample long-term time characteristic into a sample nonlinear characteristic.

5. The method of claim 2, wherein the inputting the sample short-term temporal features into a recurrent neural network to extract features that are nonlinear in a time dimension as sample nonlinear features comprises:

ignoring part of neural network units with a specified probability for other layers except the last layer in the recurrent neural network;

inputting the sample short-term time features into the recurrent neural network, and extracting features of nonlinearity in a time dimension by using non-ignored neural network units as sample nonlinear features.

6. The method of claim 1, wherein the linear network comprises an autoregressive model;

the inputting the sample current curve into a linear network to extract linear characteristics under a time dimension as sample linear characteristics comprises:

and inputting the sample current curve into an autoregressive model to linearly predict the current value when the electromagnet is electrified next time to unlock the platform door, and taking the current value as the sample linear characteristic.

7. The method of claim 1, wherein generating a predicted current curve based on the sample non-linear characteristic and the sample linear characteristic for a current value at which the electromagnet is next energized to unlock the platform door comprises:

splicing the sample nonlinear features and the sample linear features to obtain sample comprehensive features;

and mapping the comprehensive sample characteristics to a current value when the electromagnet is electrified next time to unlock the platform door, and obtaining a predicted current curve.

8. The method of any of claims 1-7, wherein training the non-linear network and the linear network according to the predicted current profile into a current prediction model comprises:

detecting the current value of the electromagnet when the platform door is unlocked by next electrification to obtain a reference current curve;

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

judging whether a stop condition is met;

if yes, outputting the nonlinear network and the linear network as a current prediction model;

if not, updating the nonlinear network and the linear network, and returning to execute the step of inputting the sample current curve into the nonlinear network to extract nonlinear characteristics under the time dimension as sample nonlinear characteristics.

9. The method of claim 8, wherein said calculating a difference between said predicted current profile and said reference current profile as a loss value comprises:

calculating a difference value between the current value in the predicted current curve and the current value in the reference current curve as a point deviation value for each same time;

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

10. The method of claim 8, wherein the non-linear network comprises a convolutional neural network, a cyclic neural network; the updating the non-linear network and the linear network comprises:

determining non-ignored neural network elements in the recurrent neural network;

updating weights in the neural network elements that are not ignored.

11. A method of predicting current, comprising:

12. The method of claim 11, wherein the non-linear network comprises a convolutional neural network, a cyclic neural network;

inputting the original current curve into a nonlinear network to extract nonlinear features under a time dimension, wherein the extracting is used as original nonlinear features and comprises the following steps:

inputting the original current curve into a convolutional neural network to extract nonlinear characteristics under a time dimension as original short-term time characteristics;

and inputting the original short-term time characteristics into a recurrent neural network to extract nonlinear characteristics under the time dimension as original nonlinear characteristics.

13. The method of claim 12, wherein the convolutional neural network comprises a plurality of convolutional layers, a plurality of active layers, the convolutional layers being interleaved with the active layers;

inputting the original current curve into a convolutional neural network to extract nonlinear features under a time dimension, wherein the extracting is used as original short-term time features and comprises the following steps:

inputting the original current curve into a convolution layer to execute convolution processing and inputting the original current curve into an activation layer to execute activation processing in sequence;

and outputting a nonlinear characteristic in a time dimension as an original short-term time characteristic when all the convolution processing and all the activation processing are executed.

14. The method of claim 12, wherein the recurrent neural network comprises a recurrent component, a first recurrent skip component, a second recurrent skip component;

inputting the original short-term time characteristic into a recurrent neural network to extract a nonlinear characteristic under a time dimension as an original nonlinear characteristic, wherein the method comprises the following steps:

inputting the original short-term time characteristic into a recursive component to extract a nonlinear characteristic under a time dimension as a first original long-term time characteristic;

inputting the original short-term time characteristic into a first recursive jump component to extract a nonlinear characteristic under a time dimension as a second original long-term time characteristic;

inputting the original short-term time characteristic into a second recursive jump component to extract a nonlinear characteristic under a time dimension as a third original long-term time characteristic;

fusing the first original long-term time feature, the second original long-term time feature, and the third original long-term time feature into an original nonlinear feature.

15. The method of claim 11, wherein the linear network comprises an autoregressive model;

the inputting the original current curve into a linear network to extract linear characteristics under a time dimension as original linear characteristics comprises:

and inputting the original current curve into an autoregressive model to linearly predict the current value when the electromagnet is electrified next time to unlock the platform door, wherein the current value is used as an original linear characteristic.

16. The method according to any one of claims 11 to 15, wherein generating a target current profile based on the original non-linear characteristic and the original linear characteristic to generate a current value for the electromagnet at the next energization to unlock the platform door comprises:

splicing the original nonlinear features and the original linear features to obtain original comprehensive features;

and mapping the original comprehensive characteristics to a current value when the electromagnet is electrified next time to unlock the platform door, and obtaining a target current curve.

17. An apparatus for training a current prediction model, comprising:

18. A current prediction apparatus, comprising:

19. 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 method of training a current prediction model according to any one of claims 1-10, or a method of current prediction according to any one of claims 11-16.

20. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a method of training a current prediction model according to any one of claims 1 to 10, or a method of current prediction according to any one of claims 11 to 16.