CN110596506A - Converter Fault Diagnosis Method Based on Time Convolutional Network - Google Patents

Abstract

The present invention relates to a fault diagnosis method for a power electronic converter based on a time convolution network technology, comprising the following steps: Step S1: collect electrical signals at measurement points and perform noise reduction processing to obtain sample data with fault information; Step S2: Use normalization to reduce the dimension of the sample data with fault information, and establish a data sample database by one-to-one correspondence between the obtained fault features and fault types; Step S3: construct a fault classifier based on a time convolutional network, and according to the data samples The database is trained and tested to obtain the optimal network structure parameters; Step S4: reconstruct the fault classifier based on the time convolution network according to the optimal network structure parameters, and obtain the fault classifier with the optimal parameters; Step S5: put The fault classifier network with optimal parameters is written into simulink, and the real-time fault diagnosis and location of the power electronic converter in actual operation are performed. The invention can judge the health status of the converter more accurately and reliably.

Description

Converter Fault Diagnosis Method Based on Time Convolutional Network

technical field

The invention relates to the technical field of power electronics, in particular to a converter fault diagnosis method based on a time convolution network.

Background technique

With the advent of the era of Industry 4.0, power electronic technology has been more widely used in various fields of production and life. Accordingly, power electronic fault diagnosis technology is also indispensable.

First of all, power electronic converters are mostly used as control equipment or core power sources. If the type of fault cannot be scientifically diagnosed, it will only raise the problem of carbuncle and affect the self-isolation and self-recovery of the device fault. Moreover, with the expansion of the fault range and the increase of functional failures, there is a great potential safety hazard.

Second, as the complexity of power electronic equipment increases, maintenance costs continue to increase. The converter, the main body of energy conversion in power electronic devices, has a high failure rate and serious consequences. Diagnosing its faults, preventing microscopic and gradual progress, avoiding causing a wide range of faults, reducing maintenance costs, and avoiding unnecessary economic losses.

Finally, the number of power electronic circuit components is large, and it is time-consuming and laborious to diagnose one by one. Automatic fault diagnosis has improved in terms of rapidity, fault tolerance, and reliability. It can also reduce downtime, realize predictive maintenance, and save manpower and material resources.

Traditional power electronic converter fault diagnosis methods include support vector machine, fault dictionary method and so on. The support vector machine is relatively simple in calculation, but it will cause misjudgment of the output result because it is easily affected by the noise of the sampled signal. The fault dictionary has strong anti-interference ability, but it needs a large fault sample to achieve good results. The method based on time convolutional network can distinguish and locate known faults and normal conditions in the case of fewer samples, and can distinguish unknown faults from normal conditions and known faults; In the case of frequency changes, the network can still be used for fault identification and location; cloud servers can be used to perform real-time online cloud processing of multi-machine fault data, providing massive data for multi-scale and multi-level complex systems.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide a converter fault diagnosis method based on time convolution network, which can judge the health status of the converter more accurately and reliably, and can identify unknown faults and adapt to frequency conversion to detect faults, The accuracy of converter fault analysis is also improved.

To achieve the above object, the present invention adopts the following technical solutions:

A fault diagnosis method for a converter based on a temporal convolutional network, comprising the following steps:

Step S1: collecting the electrical signal of the measurement point and performing noise reduction processing to obtain sample data with fault information;

Step S2: using normalization to reduce the dimension of the sample data with fault information, and establish a data sample database by one-to-one correspondence between the obtained fault characteristics and fault types;

Step S3: constructing a fault classifier based on a temporal convolutional network, and training and testing according to the data sample library to obtain optimal network structure parameters;

Step S4: reconstruct the fault classifier based on the time convolutional network according to the optimal network structure parameters, and obtain the fault classifier with the optimal parameters;

Step S5: Write the fault classifier network with optimal parameters into simulink, and perform real-time fault diagnosis and location of the power electronic converter in actual operation.

Further, the step S1 is specifically:

Step S11: according to the actual fault occurrence situation, apply a fault to the circuit components, and simulate the circuit fault under the actual situation to generate an output waveform;

Step S12: use the data acquisition card to collect the electrical signal of the measurement point;

Step S13: The sampling signal removes noise through the simulink module, obtains original sample data, and obtains samples with fault information.

Further, the capture card adopts a PCI-6229 capture card.

Further, step S3 specifically includes the following steps:

Step S31: Divide the data sample library into training samples and test samples;

Step S32: using the training samples as the input of the temporal convolutional network fault classifier to train it;

Step S33: Use different classifier functions to train the training samples, and use a classifier function with good effect according to the training result.

Step S34: determine whether the training error conforms to the preset, if so, go to step S35, otherwise change the classifier function for further adjustment;

Step S35: obtaining the optimal classifier function and hyperparameters, and assigning them to the temporal convolutional network-based classifier, and using the test sample to test the temporal convolutional network fault classifier assigned the optimal parameters;

Step S36: Determine whether the test accuracy rate meets the preset requirements, if so, take the last selected better parameter as the optimum parameter and end the process, otherwise return to step S33.

Further, the hyperparameters include convolution kernel size k, dilation coefficient d and network depth n.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention can more accurately and reliably judge the health status of the converter;

2. The present invention can identify unknown faults, adapt to frequency conversion detection faults, and improve the accuracy of converter fault analysis.

Description of drawings

1 is a schematic diagram of a data collection process flow in an embodiment of the present invention;

2 is a schematic flowchart of a method in an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a TCN model in an embodiment of the present invention

FIG. 4 is a schematic diagram of a TCN residual structure in an embodiment of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

Please refer to FIG. 2 , the present invention provides a converter fault diagnosis method based on a time convolutional network, comprising the following steps:

Step S1: collecting the electrical signal of the measurement point and performing noise reduction processing to obtain sample data with fault information;

Step S2: using normalization to reduce the dimension of the sample data with fault information, and establish a data sample database by one-to-one correspondence between the obtained fault characteristics and fault types;

Step S3: Build a fault classifier based on a time convolutional network. After offline training, the normal state contained in the training sample and various types of faults are accurately divided, the optimal parameters of the fault classifier are extracted, and the optimal parameters are directly assigned to the classifier. Network, perform classifier test work, and select the optimal network structure parameters through testing;

Step S4: reconstruct the fault classifier based on the time convolutional network according to the optimal network structure parameters, and obtain the fault classifier with the optimal parameters;

Step S5: Write the fault classifier network with optimal parameters into simulink, and perform real-time fault diagnosis and location of the power electronic converter in actual operation.

In this embodiment, a probability threshold is set for the temporal convolutional network fault classifier. When a new type of fault of unknown type is input, the recognition network can detect and judge it as the N+1th type fault which is different from the known N types of faults. , to ensure the safe operation of the system when an unknown fault occurs;

As shown in FIG. 1, in this embodiment, the step S1 is specifically:

Step S11: according to the actual fault occurrence situation, apply a fault to the circuit components, and simulate the circuit fault under the actual situation to generate an output waveform;

Step S12: use the PCI-6229 acquisition card to acquire the electrical signal of the measurement point;

Step S13: The sampling signal removes noise through the simulink module, obtains original sample data, and obtains samples with fault information.

Preferably, in this embodiment, when diagnosing an actual operating circuit, the extracted sample information will also be recorded, and when a certain number of samples are accumulated, a new sample database can be constructed to enrich the data of the sample database.

In this embodiment, step S3 specifically includes the following steps:

Step S31: Divide the data sample library into training samples and test samples;

Step S32: using the training samples as the input of the temporal convolutional network fault classifier to train it;

Step S33: Use different classifier functions to train the training samples, and use a classifier function with good effect according to the training result.

Step S34: determine whether the training error conforms to the preset, if so, go to step S35, otherwise change the classifier function for further adjustment;

Step S35: Obtain the optimal classifier function and hyperparameters, the hyperparameters include the convolution kernel size k, the expansion coefficient d and the network depth n, and assign them to the classifier based on the temporal convolution network, and use the test sample to test the Temporal convolutional network fault classifier with optimal parameters;

Step S36: Determine whether the test accuracy rate meets the preset requirements, if so, take the last selected better parameter as the optimum parameter and end the process, otherwise return to step S33.

In this embodiment, the temporal convolutional network (TCN) fault classifier is a network structure capable of processing time series data, inferring new possible information according to the input sequence, and using a judgment mechanism to evaluate the prediction effect. Good or bad, for example, the ordinary fully connected layer will use MSE as the loss function.

The model structure diagram of TCN is shown in Figure 3. The input sequence of the model is defined as: x ₀ , x ₁ ,...,x _t , and the output sequence is: y ₀ , y ₁ ,..., y _t .

The biggest difference between TCN and one-dimensional convolution is that it mainly uses dilated convolution to obtain the global information of the entire sequence, so that each hidden layer is the same size as the input sequence, and the skip layer connection of residual convolution is set. and a 1×1 convolution operation.

The general form of the dilated convolution kernel is:

In the formula, f represents the convolution kernel (filter), k represents the convolution kernel size (kernel size), d represents the dilation factor (dilation factor), which refers to the number of intervals of the convolution kernel, and s-d·i represents only the past state. convolution. As shown in Figure 3, the higher the upper layer, the larger the convolution window, and the more "holes" in the convolution window, the larger the receptive field. Therefore, the advantage of dilated convolution is that it can increase the receptive field without pooling loss information, so that each convolution output contains a larger range of information.

The residual connection is described using the formula:

o=Activation(x+F(x))

In the formula, Activation represents the activation function, including functions such as ReLU, Sigmoid, and tanh.

In order to obtain a larger receptive field, the network depth n has to be increased, so a residual unit is constructed to train a deeper network. Residual convolution is to take the lower layer features to the higher layers to enhance the accuracy.

1×1 convolution is used for dimensionality reduction. TCN directly connects the feature map jumping layer of the lower layer to the upper layer, and the number of feature maps (that is, the number of channels) of each cell corresponding to the cell is inconsistent, which makes it impossible to directly add the feature map of the jumping layer similar to Resnet. Therefore, in order to match the number of feature maps when the two layers are added, a 1×1 convolution is used for dimensionality reduction.

By adjusting the network depth n, the convolution kernel size k, and the expansion coefficient d, the receptive field can be flexibly controlled, the amount of computation can be reduced, and it can be adapted to different tasks.

Define the probability density function as where k is the training sample (k=1, 2,...,c), and n _k is the number of samples in the k-th training set. represents the probability density function of the rth sample in the kth class of samples.

When there are c classes of training samples, each class of samples has a corresponding probability density function The probability density function for constructing all training samples is It can be defined as:

According to the probability density function of all training samples obtained Then the definition of the proportion ρ of a single fault is as follows:

When the specific gravity ρ ^(k) is the maximum value among the specific gravity ρ ⁽¹⁾ , ρ ⁽²⁾ , ..., ρ ^(c) , the sample to be tested belongs to the k-th fault type. In this way, the TCN algorithm can be used to identify the fault samples of known classes.

Define β as the minimum value of specific gravity:

β=min{ρ}

The β value is the dividing line between the known fault samples and the unknown fault samples. When ρ>β, the test sample belongs to a normal or known fault condition; when ρ<β, the test sample is an unknown type of fault sample, and the output is the N+1 fault that is different from the known N types of faults, ensuring that The system can operate safely in the event of an unknown failure. In this way, the TCN algorithm can be used to classify the fault samples of unknown classes.

Preferably, in this embodiment, using the time convolution network, when an unknown fault occurs in the circuit, the fault state can be distinguished from the normal state, thereby realizing self-checking, and the algorithm can reduce the time and complexity of sample training, Reliance on faulty samples is also reduced.

Preferably, in this embodiment, the expansion coefficient is designed into a zigzag structure, such as a cyclic structure of [1, 2, 5, 1, 2, 5]. The nature of such jaggedness itself can better meet the segmentation requirements of small objects and large objects at the same time (the expansion coefficient of small objects cares about short-range information, and the expansion coefficient of large objects cares about long-distance information), reducing the loss of continuity of information sex.

Preferably, in this embodiment, a TCN fault classifier network is constructed, the classifier parameter selection rule is derived, the number of sampling points is changed according to the actual working output frequency change, and the same model constructed by the TCN is also used to classify and identify faults , realize the detection under the condition of adaptive frequency conversion; and use the single-chip microcomputer to receive data and send it to the cloud server, provide massive data for multi-scale and multi-level complex systems, perform real-time online cloud processing of multi-machine fault data, and reduce the dependence on historical data , to realize the real-time self-inspection of the power electronic converter circuit, identify the fault with high efficiency and high reliability, and locate the fault. Using wireless as a medium for data transmission is more convenient than wired data transmission, and avoids excessive use of transmission lines in the same place.

The above descriptions are only preferred embodiments of the present invention, and all equivalent changes and modifications made according to the scope of the patent application of the present invention shall fall within the scope of the present invention.

Claims (5)

1. A converter fault diagnosis method based on a time convolution network is characterized by comprising the following steps:

step S1: collecting electric signals of a measuring point and carrying out noise reduction treatment to obtain sample data with fault information;

step S2: carrying out dimension reduction treatment on sample data with fault information by adopting normalization, and establishing a data sample library by corresponding the obtained fault characteristics and fault types one by one;

step S3: constructing a fault classifier based on a time convolution network, and training and testing according to a data sample base to obtain an optimal network structure parameter;

step S4, reconstructing the fault classifier based on the time convolution network according to the optimal network structure parameters to obtain the fault classifier with the optimal parameters;

step S5: and writing the fault classifier network with the optimal parameters into simulink, and performing real-time fault diagnosis and positioning on the power electronic converter in actual operation.

2. The method for diagnosing the fault of the converter based on the time convolution network as claimed in claim 1, wherein the method comprises the following steps:

step S11: applying faults to circuit components according to the actual fault occurrence condition, and simulating the circuit faults under the actual condition to generate output waveforms;

step S12: collecting the electric signals of the measuring points by using a data acquisition card;

step S13: and removing noise of the sampling signal through a simulink module, acquiring original sample data, and obtaining a sample with fault information.

3. The method for diagnosing the fault of the converter based on the time convolution network as claimed in claim 2, characterized in that: the acquisition card adopts a PCI-6229 acquisition card.

4. The method for diagnosing the fault of the converter based on the time convolution network as claimed in claim 1, wherein the method comprises the following steps: step S3 specifically includes the following steps:

step S31: dividing a data sample library into a training sample and a test sample;

step S32: training the fault classifier by using a training sample as the input of the time convolution network fault classifier;

step S33: training the training samples by adopting different classifier functions, and using the classifier function with good effect according to the training result;

step S34: judging whether the training error meets the preset condition, if so, entering the step S35, otherwise, changing the classifier function for further adjustment;

step S35: acquiring a better classifier function and a hyper-parameter, assigning the better classifier function and the hyper-parameter to a time convolution network-based classifier, and testing the time convolution network fault classifier assigned with the better parameter by adopting a test sample;

step S36: and judging whether the testing accuracy meets the preset requirement, if so, taking the last selected better parameter as the optimal parameter and ending the process, otherwise, returning to the step S33.

5. The method for diagnosing the fault of the converter based on the time convolution network as claimed in claim 1, wherein the method comprises the following steps: the hyper-parameters include convolution kernel size k, dilation coefficient d, and network depth n.