CN115310604A

CN115310604A - UPS system online fault diagnosis method based on neural network

Info

Publication number: CN115310604A
Application number: CN202210868333.0A
Authority: CN
Inventors: 王愚; 李绍坚; 李想; 张子君; 莫裕倩; 鲍海波; 邓高峰; 聂雷刚; 韦捷; 黄增柯; 向颖; 国家栋; 叶蕾; 檀亚凤; 邰思博
Original assignee: Nanning Power Supply Bureau of Guangxi Power Grid Co Ltd
Current assignee: Nanning Power Supply Bureau of Guangxi Power Grid Co Ltd
Priority date: 2022-07-22
Filing date: 2022-07-22
Publication date: 2022-11-08

Abstract

The application discloses a UPS system online fault diagnosis method based on a neural network, which comprises the following steps: acquiring fault state signals and binary codes of a system under different working conditions, wherein the binary codes are used for marking all fault types of the system; extracting a fault characteristic vector of the fault state signal by using a wavelet decomposition method; training a neural network model based on the fault feature vector and the binary code; inputting the acquired fault state signal of the system under a certain working condition into the neural network model for fault classification; outputting the fault state type of the system under a certain working condition; the invention integrates wavelet decomposition, further enhances the fault characteristics, enables the neural network model to more easily identify the fault state of the UPS and simultaneously improves the accuracy of fault identification.

Description

UPS system online fault diagnosis method based on neural network

Technical Field

The invention relates to the technical field of power electronic fault diagnosis, in particular to a UPS system online fault diagnosis method based on a neural network.

Background

UPS refers to power electronics devices that provide a continuous reliable, consistent high quality and uninterrupted supply of power. It mainly consists of five parts: the power supply switching circuit comprises a main circuit, a bypass, a battery and other power supply input circuits, a rectifier, an inverter, an inversion and bypass output switching circuit and a storage battery, can provide a standby power supply for a load to prevent critical equipment from being lost due to sudden terminal power supply, can also improve the quality of electric energy, and meets the requirement that certain important equipment needs high-reliability power supply.

In the prior art, a part of the UPS faults are caused by power switching tubes, wherein the faults of the power switching tubes can be divided into open-circuit faults and short-circuit faults, the short-circuit faults exist for a very short time, generally within 10 μ s, and measures such as short-circuit detection and overcurrent protection circuits are integrated for the short-circuit faults at present. When the power switch tube has an open-circuit fault, the system cannot be immediately broken down, but the long-time abnormal working state easily causes secondary faults of other devices, so that the system is damaged, and serious consequences are caused.

Disclosure of Invention

The invention aims to provide a UPS system online fault diagnosis method based on a neural network, which is used for improving the accuracy of fault identification.

The invention provides a UPS system online fault diagnosis method based on a neural network, which comprises the following steps:

acquiring fault state signals and binary codes of a system under different working conditions, wherein the binary codes are used for marking all fault types of the system;

extracting a fault characteristic vector of the fault state signal by using a wavelet decomposition method;

training a neural network model based on the fault feature vector and the binary code;

inputting the acquired fault state signal of the system under a certain working condition into the neural network model for fault classification;

and outputting the fault state type of the system under a certain working condition.

Optionally, the extracting the fault feature vector of the fault status signal by using a wavelet decomposition method includes:

calculating the sum of absolute values through wavelet decomposition coefficients obtained after N-layer wavelet transform is carried out on the fault state signal;

determining a fault feature vector based on the sum of the absolute values.

Optionally, training a neural network model based on the fault feature vector and the binary code includes:

constructing a fault data set according to the fault feature vector;

and training a neural network by taking the fault data set as an input vector and the binary code as an output vector.

Optionally, the fault state signal and the binary code are signal data of a rectifying circuit module and a inverting circuit module.

Optionally, after the wavelet decomposition, the total energy of the fault state signal remains unchanged, and an energy distribution formula of each frequency band of the fault state signal in the wavelet decomposition process is as follows: e _n,k ＝∑|S _n,k | ² Where n denotes the number of wavelet decomposition layers, k denotes the number of wavelet decomposition coefficient sequences, E _n,k For the energy of the k wavelet coefficient sequence after the nth wavelet decomposition, S _n,k Represented as a sequence of wavelet coefficients.

Above technical scheme can see that, this application has following advantage: according to the invention, the fault characteristics of the acquired fault state signals are further enhanced through wavelet decomposition, then the neural network is trained based on the extracted fault characteristic vectors and the acquired binary codes for marking all fault types, so that the fault state of the UPS can be more easily identified by the neural network model, finally the fault state types of the rectifier and the inverter are identified by inputting the acquired actual fault state signals into the trained neural network model, and the accuracy of fault identification is effectively improved through a neural network self-adaptive learning mode.

Drawings

FIG. 1 is a schematic flow chart of an online fault diagnosis method for a UPS system based on a neural network according to the present invention;

FIG. 2 is a topology diagram of a UPS system according to the present invention;

FIG. 3 is a three-phase PWM rectification circuit diagram of the present invention;

FIG. 4 is a diagram of a single-phase inverter circuit according to the present invention;

FIG. 5 is an exploded view of the wavelet of the present invention;

FIG. 6 is a diagram of a neural network architecture in accordance with the present invention;

FIG. 7 is a waveform of the input current of the three-phase rectifier of the present invention;

FIG. 8 is a diagram of the results of fault diagnosis for a three-phase rectifier of the present invention;

FIG. 9 is a graph of the output voltage waveform of a single phase inverter of the present invention;

fig. 10 is a diagram showing the result of fault diagnosis of the single-phase inverter according to the present invention.

Detailed Description

The embodiment of the application provides a UPS system online fault diagnosis method based on a neural network, which is used for improving the accuracy of system fault identification.

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

Example 1:

referring to fig. 1 to 10, an online fault diagnosis method for a UPS system based on a neural network mainly includes the following steps:

s1, acquiring system fault state signals and binary codes under different working conditions, wherein the binary codes are used for marking all fault types of a system;

in this embodiment, the UPS is split into two modules, i.e., a rectification module and an inversion module, and therefore a fault state signal and a binary code in the two modules are obtained here, and diagnosis is performed by using the rectification module, as shown in fig. 3 (a), a three-phase PWM rectifier includes A, B, C three-phase bridge arms, an input-side inductor, and a dc bus capacitor C, where each phase bridge arm includes 2 Insulated Gate Bipolar Transistors (IGBTs) and 2 freewheeling diodes. The PWM rectifier circuit has 6 IGBTs as the core elements of the power converter, when a single power tube fails, the failure modes of the rectifier have 6 types, and when two power tubes fail, the failure modes of the rectifier have 15 types, so that the failure modes of the rectifier have a total of 22 types, including 1 healthy state and 21 failure states, in this embodiment, all failure modes are marked by binary coding and used as the target output matrix of the neural network, as shown in table 1 below:

TABLE 1

S2, extracting a fault characteristic vector of the fault state signal by using a wavelet decomposition method;

in this embodiment, as shown in fig. 5, a fault signal of an analog circuit is decomposed to different frequency bands by using wavelet analysis, so as to obtain signal components of each frequency band, and since a low-frequency signal can better reflect fault signal characteristics of the circuit, after the signal components of each frequency band are obtained, the signal of the low-frequency component can be reused as fault characteristics of the circuit to be tested, and then the extracted fault characteristic vector is input to a BP neural network to perform analog circuit fault diagnosis. The analog circuit fault diagnosis method based on wavelet transformation preprocessing greatly reduces the dimension of the input vector of the neural network, thereby greatly simplifying the structure of the neural network, greatly shortening the training time and greatly improving the fault identification capability compared with the neural network fault diagnosis technology without wavelet transformation preprocessing.

S3, training a neural network model based on the fault feature vector and the binary code;

considering that the current signal has a sinusoidal function periodic variation characteristic, in this embodiment, a periodic three-phase input current signal is selected to form a fault sequence with a length of 600, each fault form simulates 20 different working conditions to construct 440 fault samples, and in 20 working conditions in each fault form, the ratio of 4: the proportion of 1 is divided into 16 working conditions as a training set and 4 working conditions as a test set.

After the original fault state signal is subjected to the feature extraction through the wavelet decomposition, as shown in fig. 6, a three-phase input current sequence is obtained through sampling, 4-layer wavelet decomposition is performed on the three-phase input current sequence, the wavelet coefficient distribution features of 8 frequency band signals are respectively extracted, and then a BP neural network model is trained based on the extracted fault features and the binary coding.

S4, inputting the acquired fault state signal of the system under a certain working condition into a neural network model for fault classification;

and acquiring a voltage current signal of the system under a certain actual working condition through sampling, and inputting the actual voltage current signal into a BP neural network model for fault classification. The failure modes of the rectifier are 6, and when two power tubes fail, the failure modes of the rectifier are 15, so that the failure modes of the rectifier are 22, including 1 healthy state and 21 failure states; the single-phase inverter has three 11 fault modes, including a normal working state, a single power tube fault and two power tube faults.

And S5, outputting the fault state type of the system under a certain working condition.

And inputting the actual voltage and current signals into the trained BP neural network, and judging the fault state type of the circuit to be tested according to the output result of the BP neural network. For example, the present embodiment builds up a UPS system using a programmable power supply, a three-phase rectifier, a single-phase inverter, dc bus capacitors and inductors, and an adjustable load, as shown in fig. 2, and performs fault diagnosis thereof. UPS system parameters are shown in table 2 below:

TABLE 2

There are four 22 fault types in the three-phase PWM rectifier: normal operating state, single power tube fault and two power tube faults in same phase and two power tube faults in different phases. If the power tubes S1 and S3 of the PWM rectifier fail at the same time, the simulation result is shown in fig. 7, it can be seen that the three-phase input currents ia, ib, and ic are distorted, the three-phase input current signals are collected, the failure characteristics are extracted and input to the trained neural network model, the binary code is output by the neural network model as a predicted value, and the failure state of the rectifier can be identified by comparing the predicted value with the labeled target value, as can be seen from fig. 8, when the power tubes S1 and S3 fail at the same time, the output value of the neural network is approximately 01000, that is, it is determined that the power tubes S1 and S3 fail. Finally, the statistical identification results of the three-phase PWM rectifier under 22 fault modes are shown in table 3 below, and it is seen from the table that the proposed algorithm can accurately identify the fault state of the three-phase PWM rectifier.

TABLE 3

The single-phase inverter has three 11 fault modes including a normal working state, a single power tube fault and a two power tube fault. The fault codes are shown in table 4:

TABLE 4

If the inverter T1 fails, the simulation result is shown in fig. 9, at this time, the positive half cycle of the output voltage waveform vo is lost, the output voltage and the bridge arm midpoint voltage are collected, the fault characteristics are extracted through wavelet decomposition, the identification result obtained by combining a neural network model is shown in fig. 10, the neural network output value is approximately 00001, and the power tube S1 is judged to have a fault. Finally, the identification results of the inverter under 11 fault forms are obtained through statistics, as shown in table 5, it can be seen that the proposed algorithm can accurately identify the fault state of the inverter no matter what fault the inverter is in.

TABLE 5

Example 2:

s201, calculating the sum of absolute values through wavelet decomposition coefficients obtained after N-layer wavelet transformation is carried out on the fault state signals;

s202, determining a fault feature vector based on the sum of absolute values;

after step S201, the method further includes performing normalization processing on the fault feature vector.

In this embodiment, according to the law of conservation of energy, the total energy of the signal before and after wavelet decomposition remains unchanged, and after the signal is subjected to 4-layer wavelet decomposition, that is, the energy of the original signal is equal to the sum of the energy of the 4-layer wavelet high-frequency coefficient and the energy of the 4 th layer wavelet low-frequency coefficient; thus, the energy distribution of the fault signal in each frequency band can be obtained, and E _n,k Is the k-th wavelet coefficient sequence S after the nth wavelet decomposition _n,k The expression of (a) is as follows:

E _n,k ＝∑|S _n,k | ²

where n represents the number of wavelet decomposition layers, k represents the number of wavelet decomposition coefficient sequences, E _n,k For the energy of the k wavelet coefficient sequence after the nth wavelet decomposition, S _n,k Expressed as wavelet coefficient series, N is the maximum number of layers of wavelet decomposition and N < N, k is the number of wavelet coefficient series per layer, k is the wavelet low frequency coefficient if and only if 1 or 2,k =1, and k =2 is the wavelet high frequency coefficient.

Based on the obtained energy of each frequency band, the energy spectrum vector T is constructed as follows:

T＝[E _1,2 E _2,2 E _3,2 E _4,1 E _4,2 ]

after normalization, the unified dimensional expression is as follows:

E＝(E _1,2 +E _2,2 +E _3,2 +E _4,2 +E _4,1 ) ^1/2

T'＝[E _1,2 /E E _2,2 /E E _3,2 /E E _4,1 /E E _4,2 /E]

and taking the normalized energy spectrum vector as a new fault feature vector.

S301, constructing a fault data set according to the fault feature vector;

and S302, taking the fault data set as an input vector, taking the binary code as an output vector, and training a neural network.

The invention adopts BP neural network to diagnose the fault, uses the fault data set as the input vector of the neural network, uses the binary code for marking the fault form as the output vector of the neural network, trains the neural network model at the moment, and then inputs the actual voltage and current signals obtained by sampling into the trained neural network model, and identifies the fault state of the UPS.

The basic flow of the algorithm is as follows: and transmitting the network input sample to the hidden layer through the input layer to perform data processing such as weighting function mapping and the like, and then transmitting the data to the output layer to finish output through processing, thereby finishing forward propagation. When the expected output has an error with the actual output, the generated error value is reversely transmitted into the input layer by layer through the hidden layer, so that the back propagation is formed. And continuously adjusting the weight threshold value between layers according to a gradient descent method in the error reverse transfer process, wherein the adjustment process of the weight threshold value of the BP algorithm is a batch processing method, and the adjustment of the weight threshold value is carried out after the total error of all input samples is calculated.

The above two processes are alternately repeated until the error converges. Therefore, the actual output value of the BP neural network gradually approaches to the expected output value, and the requirement of the converged error threshold value is met. The basic structural model is shown in fig. 6.

With the conversion of the hidden layer number, the BP neural network can have multiple layers, and the three-layer BP neural network is composed of neurons of input, hidden and output layers. In fig. 6, each input, hidden and output layer of the BP neural network is represented by M, I, P, and each layer of corresponding neurons are represented by m, i and p.

Setting excitation functions of three layers of BP neural networks as Sigmoid functions, setting an input vector of each neuron as x, a hidden layer vector as h, an output vector as y, and setting n learning training samples in total, wherein the n learning training samples form x = [ x ] ₁ ,x ₂ ,…,x _n ]Wherein each learning training sample dimension corresponds to a network input layer dimension, and the expected output of each sample corresponding to the network is d _K ＝[d _K1 ,d _K2 ,…,d _KP ] ^T But the actual output value is Y _K ＝[Y _K1 ,Y _K2 ,…,Y _KP ] ^T The output dimension is consistent with the dimension of the network input layer, and the specific input and output expressions of neuron nodes of each layer are as follows:

the formula for computing the ith neuron output of the hidden layer is as follows:

the calculation formula of the pth neuron output of the three-layer BP neural network is as follows:

the calculation formula of the error of the p-th neuron of the output layer is as follows:

e _kp (n)＝d _kp (n)-y _kp (n)

the weight threshold adjustment back propagation of the error is completed by minimizing the sum of the squares of the error between the actual output and the expected output and solving the corresponding weight threshold gradient, the error data iteration is completed by utilizing the negative direction of the weight threshold gradient, and the function f (x) is positioned at x _i Taking a first-order approximate Taylor series expansion, wherein the expression is as follows:

wherein

The iteration direction is taken as:

or

The calculation formula for the next iteration point is thus as follows:

according to the analysis, the fault diagnosis of the analog circuit in the rectification module and the inversion module can be realized through the BP neural network.

It is intended that the foregoing description of the disclosed embodiments enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A UPS system online fault diagnosis method based on a neural network is characterized by comprising the following steps:

2. The method of claim 1, wherein the extracting the fault feature vector of the fault status signal by using the wavelet decomposition method comprises:

calculating the sum of absolute values through wavelet decomposition coefficients obtained after performing N-layer wavelet transform on the fault state signal;

determining a fault feature vector based on the sum of the absolute values.

3. The method of claim 1, wherein training a neural network model based on the fault feature vector and the binary code comprises:

constructing a fault data set according to the fault feature vector;

4. The method of claim 1, wherein the fault status signal and binary code are signal data for both a rectifier circuit and an inverter circuit module.

5. The method according to claim 1, wherein the total energy of the fault status signal after the wavelet decomposition remains unchanged, and the energy distribution formula of each frequency band of the fault status signal in the wavelet decomposition process is as follows: e _n,k ＝∑|S _n,k | ² Where n denotes the number of wavelet decomposition layers, k denotes the number of wavelet decomposition coefficient sequences, E _n,k For the energy of the k wavelet coefficient sequence after the nth wavelet decomposition, S _n,k Represented as a sequence of wavelet coefficients.