CN114296005A - Modular multilevel converter submodule composite fault diagnosis method - Google Patents
Modular multilevel converter submodule composite fault diagnosis method Download PDFInfo
- Publication number
- CN114296005A CN114296005A CN202111499359.4A CN202111499359A CN114296005A CN 114296005 A CN114296005 A CN 114296005A CN 202111499359 A CN202111499359 A CN 202111499359A CN 114296005 A CN114296005 A CN 114296005A
- Authority
- CN
- China
- Prior art keywords
- layer
- capsule
- fault diagnosis
- data
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 48
- 238000000034 method Methods 0.000 title claims abstract description 40
- 239000002131 composite material Substances 0.000 title claims abstract description 31
- 239000002775 capsule Substances 0.000 claims abstract description 112
- 238000000605 extraction Methods 0.000 claims abstract description 40
- 238000004364 calculation method Methods 0.000 claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 24
- 238000012360 testing method Methods 0.000 claims abstract description 22
- 238000005070 sampling Methods 0.000 claims abstract description 17
- 238000012795 verification Methods 0.000 claims abstract description 14
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 11
- 238000010606 normalization Methods 0.000 claims abstract description 9
- 238000012545 processing Methods 0.000 claims abstract description 7
- 239000013598 vector Substances 0.000 claims description 29
- 230000006870 function Effects 0.000 claims description 20
- 210000002569 neuron Anatomy 0.000 claims description 11
- 230000008878 coupling Effects 0.000 claims description 7
- 238000010168 coupling process Methods 0.000 claims description 7
- 238000005859 coupling reaction Methods 0.000 claims description 7
- 230000015654 memory Effects 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 7
- 239000006185 dispersion Substances 0.000 claims description 5
- 230000005540 biological transmission Effects 0.000 claims description 4
- 238000012804 iterative process Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 230000004888 barrier function Effects 0.000 claims 1
- 238000001514 detection method Methods 0.000 abstract description 10
- 239000000284 extract Substances 0.000 abstract description 7
- 230000007787 long-term memory Effects 0.000 abstract description 3
- 230000006403 short-term memory Effects 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 8
- 238000013528 artificial neural network Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 230000035945 sensitivity Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 241001391944 Commicarpus scandens Species 0.000 description 1
- 239000003990 capacitor Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000002405 diagnostic procedure Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000007667 floating Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000010438 heat treatment Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 210000002364 input neuron Anatomy 0.000 description 1
- 230000016507 interphase Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011425 standardization method Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Landscapes
- Error Detection And Correction (AREA)
Abstract
The invention provides a modular multilevel converter submodule composite fault diagnosis method, which comprises the following steps: sampling an original current signal, carrying out normalization processing on the current signal obtained by sampling, and randomly dividing the current signal into a training set, a verification set and a test set; building an improved capsule network model: the one-dimensional convolutional neural network is combined with a long-term memory network and a short-term memory network to serve as a feature extraction unit of the capsule network, and then the feature extraction unit, the main capsule layer and the digital capsule layer are combined to form an improved capsule network model; independently evaluating the model hyper-parameters on the verification set to obtain a group of hyper-parameter optimized training models; and finally, testing the test set data on the trained model, and outputting a fault diagnosis result. The method can extract deep features in the original signal at lower calculation cost, still has higher fault diagnosis precision under the working condition that the MMC has higher level number and time domain waveform features are not obvious, and is low in detection cost, simple and easy to realize.
Description
Technical Field
The invention relates to the technical field of fault diagnosis, in particular to a composite fault diagnosis method for a submodule of a modular multilevel converter.
Background
As a novel voltage source Converter topology, a Modular Multilevel Converter (MMC) is widely applied to occasions such as medium-high voltage direct current transmission and high-voltage power driving due to the advantages of Modular structure design, easiness in expansion and the like. The MMC is formed by cascading a large number of submodules, an Insulated Gate Bipolar Transistor (IGBT) and a diode are adopted in each submodule as a current conversion device, and the IGBT is easy to break down due to low overvoltage resistance and overcurrent capacity. Sub-module faults can cause deviation of bridge arm output voltage and expectation, increase of interphase circulating current and increase of harmonic waves on an alternating current side and a direct current side, and the safe and reliable operation of the MMC system is seriously influenced. When a single submodule is not detected in time after an open-circuit fault occurs and operates with a fault for a long time, the heating value and the loss of a switching device in the submodule are increased sharply, other normal submodules are possibly caused to have faults, and when two or more submodules have open-circuit faults, the fault mode is defined as a composite fault. Complex faults are common during the operation of the device, and therefore, diagnosis of the sub-module complex fault of the MMC system is necessary.
The existing MMC fault diagnosis aims at single submodule open-circuit fault diagnosis and is based on the fact that a voltage sensor is installed on each submodule to carry out diagnosis, and the number of actual submodules is large, so that system hardware detection cost is increased invisibly. Therefore, how to realize the composite fault diagnosis of the MMC sub-module under the condition of using a small amount of sensors has very important theoretical significance and engineering application value.
At present, a popular fault diagnosis method is a fault diagnosis method based on deep learning, which adopts a traditional convolutional neural network, but the traditional convolutional neural network can only extract partial features, so that the diagnosis effect is poor.
Disclosure of Invention
The invention provides a modular multilevel converter submodule composite fault diagnosis method, which diagnoses modular multilevel converter submodule composite faults by constructing an improved capsule network model, further extracts deep-level features on the basis of the features extracted by a feature extraction unit through a main capsule layer and a digital capsule layer, obtains information representing fault classification results, improves the fault diagnosis effect, and can diagnose MMC submodule composite faults with high precision without increasing the number of sensors, thereby improving the reliability of equipment.
In order to solve the technical problem, the invention provides a composite fault diagnosis method for a submodule of a modular multilevel converter, which comprises the following steps:
s1: sampling an original current signal of the modular multilevel converter, wherein the original current signal comprises a three-phase output current and a three-phase internal circulating current signal;
s2: preprocessing the acquired original current signal;
s3: dividing a data set formed by the preprocessed original current signals into a training set, a verification set and a test set;
s4: building a capsule network model, and initializing weight parameters of the capsule network model, wherein the capsule network model comprises a feature extraction layer, a main capsule layer and a digital capsule layer, the feature extraction layer is formed by fusing a one-dimensional convolutional neural network (1DCNN) and a long-time memory network (LSTM) and is used for performing feature extraction on an input current signal, the main capsule layer is used for obtaining low-level features based on the features of the feature extraction layer, the digital capsule layer is used for obtaining high-level features based on the output of the main capsule layer and information used for representing a fault diagnosis result, and the transmission of feature vectors is realized between the digital capsule layers of the main capsule layer through a dynamic routing algorithm;
s5: training the capsule network model by using a training set, constructing a loss function, and evaluating model hyper-parameters on a verification set to obtain a group of models with optimal model hyper-parameters as well as trained models;
s6: and testing the trained model by using the test set data, and outputting a fault diagnosis result.
In one embodiment, the step S1 adopts an overlapped sampling method to perform the sampling of the current signal, which includes:
and selecting a point on the original current signal as a collection starting point, collecting data points with a certain length each time, moving the data points with a certain length backwards each time the collection is completed, and continuously collecting at a new starting point position until all the remaining data points are collected.
In one embodiment, the preprocessing in step S2 is a normalization processing, specifically: normalizing the input data to a [0, 1] interval by using a dispersion normalization method, wherein the calculation formula is as follows:
wherein, x' represents normalized data, x represents original data, and max (x) and min (x) represent the maximum value and the minimum value of the original data respectively.
In one embodiment, in the data set divided in step S3, a 16-dimensional one-hot encoding vector is used for each sample to produce a label, specifically: and representing each label as an all-zero vector, wherein the corresponding element of the label index is 1, and the rest elements are 0.
In one embodiment, the feature extraction layer in step S4 includes a one-dimensional convolution layer, a one-dimensional maximum pooling layer, and an LSTM layer, and the weight parameters are initialized by using a normal distribution with a mean value of 0 and a standard deviation of 0.5 for the one-dimensional convolution layer; the LSTM layer is initialized with uniformly distributed initialization weight parameters distributed between (-0.1, 0.1).
In one embodiment, the dynamic routing algorithm adjusts parameters of the main capsule layer and the digital capsule layer through an iterative process, so that the information of the main capsule layer is sent to the digital capsule layer, and the calculation formula is as follows:
uj|i＝Wijui；
wherein u isiIs the output of the previous layer of vector neurons, i.e. the lower layer features, WijIs an attitude matrix between the lower-level features and the higher-level features, uj|iIs to infer a prediction vector of a high-level feature from a low-level feature, cijFor the coupling coefficients determined by the dynamic routing algorithm, the neurons used to control the inputs automatically select the best path to transmit to the next layer of neurons, vjIs the output vector, s, of the digital capsule layerjIs the output vector of the main capsule layer.
In one embodiment, the coupling coefficient is calculated by the formula:
b'i,j＝bi,j+vjuj|i；
wherein, b'i,jFor iteratively updated bias coefficients, bi,jIs the original bias coefficient.
In one embodiment, in step S5, the total loss is calculated by combining the interval loss and the reconstruction loss when constructing the loss function, wherein the interval loss function is expressed as:
Lk＝Tkmax(0,m+-||vk||)2+λ(1-Tk)max(0,||vk||-m-)2；
wherein k is the number of classifications; vkRepresenting the predicted probability, T, of the kth fault typekIndicating a function for classification, i.e. T when the current data is of class kk1, otherwise Tk＝0；m+The method is an upper bound and is used for punishing false positive, namely predicting that k types exist and not really exist; m is-The lower bound is used for punishing false negatives, namely predicting that k types do not exist and really exist;
the reconstruction loss calculation method is that a 3-layer full-connection network is constructed after a capsule layer, output data with the same dimensionality as original input data is obtained, the square sum of the difference between the final output and the value on the initially input unit is used as a loss value, and a reconstruction loss calculation formula is as follows:
wherein n represents the number of data points contained in the input time series sample, α is the proportion of the total loss of the reconstruction loss, and fiOutput value, y, of each sample data point during reconstructioniThe real value corresponding to each sample data.
In one embodiment, in the training process of step S5, the model weight parameters are updated by back propagation algorithm, and the key hyper-parameters affecting the model performance are optimizer type, iteration number and batch size.
One or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:
the invention provides a modular multilevel converter submodule composite fault diagnosis method, which adopts a capsule network model to carry out composite fault diagnosis, can utilize the characteristics of sequence sensitivity of long-time memory neural network for extracting time sequence data and low calculation cost of a one-dimensional convolution neural network, can extract the characteristics in an original signal at low calculation cost, further obtain deep-level characteristics related to composite faults through a main capsule layer, and obtain information representing fault diagnosis results through a digital capsule layer, extracts more detailed characteristics, can improve the fault classification effect, and still has higher fault diagnosis precision under the working condition that time domain waveform characteristics are not obvious due to higher MMC level number. For the whole MMC fault detection system, only 6 current sensors are needed, other hardware equipment is not needed, the detection cost is low, and the method is simple and easy to realize.
Furthermore, a sample set is constructed and expanded by adopting an overlapping sampling method, and the method has higher feature extraction capability and fault diagnosis precision under the working condition of insufficient sample data.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
Fig. 1 is a main circuit topology structure diagram of a modular multilevel converter according to an embodiment of the invention;
fig. 2 is a block diagram of a sub-module topology of a modular multilevel converter according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an overlapping sampling in accordance with an embodiment of the present invention;
FIG. 4 is a time domain plot of three phase AC current and three phase internal circulating current during normal operation in accordance with one embodiment of the present invention;
FIG. 5 is a block diagram of an improved capsule network model in accordance with one embodiment of the present invention;
FIG. 6 is a schematic diagram of calculation of reconstruction loss of capsule network according to an embodiment of the present invention
Fig. 7 is a flow chart of an embodiment of a composite fault diagnostic method in accordance with an embodiment of the present invention.
Detailed Description
The invention discloses a modular multilevel converter submodule composite fault diagnosis method, which is different from a traditional convolutional neural network, adopts a capsule network, combines a one-dimensional convolutional neural network with a long-time memory network to form a feature extraction unit of the capsule network, utilizes the characteristics of sequential sensitivity of the long-time memory neural network for extracting time sequence data and low calculation cost of the one-dimensional convolutional neural network, can extract deep features in an original signal at lower calculation cost, and still has higher fault diagnosis precision under the working condition that time domain waveform features are not obvious due to higher level number of MMC. For the whole MMC fault detection system, only 6 current sensors are needed, other hardware equipment is not needed, the detection cost is low, and the method is simple and easy to realize.
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, 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 some, but not all, embodiments of the present invention. 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.
The embodiment of the invention provides a composite fault diagnosis method for a submodule of a modular multilevel converter, which comprises the following steps:
s1: sampling an original current signal of the modular multilevel converter, wherein the original current signal comprises a three-phase output current and a three-phase internal circulating current signal;
s2: preprocessing the acquired original current signal;
s3: dividing a data set formed by the preprocessed original current signals into a training set, a verification set and a test set;
s4: building a capsule network model, and initializing weight parameters of the capsule network model, wherein the capsule network model comprises a feature extraction layer, a main capsule layer and a digital capsule layer, the feature extraction layer is formed by fusing a one-dimensional convolutional neural network and a long-time and short-time memory network and is used for performing feature extraction on an input current signal, the main capsule layer is used for obtaining and obtaining low-level features based on the features of the feature extraction layer, the digital capsule layer is used for obtaining high-level features and information used for representing fault diagnosis results based on the output of the main capsule layer, and the transmission of feature vectors is realized between the digital capsule layers of the main capsule layer through a dynamic routing algorithm;
s5: training the capsule network model by using a training set, constructing a loss function, and evaluating model hyper-parameters on a verification set to obtain a group of models with optimal model hyper-parameters as well as trained models;
s6: and testing the trained model by using the test set data, and outputting a fault diagnosis result.
Fig. 7 is a flowchart illustrating an embodiment of a composite fault diagnosis method according to an embodiment of the present invention. In the figure, the convolution pooling layers are a one-dimensional convolution layer and a one-dimensional maximum pooling layer, the LSTM layer is a long-term memory neural network, and the capsule layer includes a main capsule layer and a digital capsule layer.
Specifically, step S1 is the acquisition of a raw current signal, steps S2 to S3 are the processing of data (signal), and the preprocessing is the conversion of the acquired raw current signal into a form that can be processed by a model.
Step S4 is a model construction, specifically, a one-dimensional convolutional neural network (1DCNN) is combined with a long-term memory network (LSTM) as a feature extraction unit of a capsule network, so as to construct a feature extraction unit that is more comprehensive and abundant and has prominent spatial features. Then the main capsule layer and the digital capsule layer are combined together to form a capsule network model improved by a characteristic extraction unit. The feature extraction layer extracts low-level features in the original time-series data, and the main capsule layer is used for the second convolution and the initial capsule input. If the capsule (main capsule layer) is directly used for extracting the low-level feature content in the original signal, the effect is not ideal, but the convolution layer is good at extracting the low-level feature, so that the feature extraction layer is firstly used for primary feature extraction, and then the main capsule layer is used for further feature extraction, thereby being beneficial to improving the accuracy of feature extraction. The main capsule layer stores low-level characteristic vectors, the digital capsule layer stores high-level characteristic vectors, the network front end adopts a 1DCNN and LSTM combined characteristic extraction structure, and the MMC three-phase alternating current and three-phase circulating current signals after normalization processing are received as fault detection data. The back end adopts a structure of a main capsule layer and a digital capsule layer, realizes the transfer of the characteristic vector through a dynamic routing algorithm, and adopts a Relu function as an activation function.
Fig. 5 is a block diagram of an improved capsule network model according to an embodiment of the present invention.
The model that this application founds is different from traditional convolution neural network, and what adopted is capsule network model, including characteristic extraction layer, main capsule layer and digital capsule layer, after the characteristic of input signal was drawed at characteristic extraction layer, can draw the more deep characteristic of input signal through main capsule layer to obtain fault classification result information through digital capsule layer. Although the capsule network is related in other fields, the application provides an improved model aiming at the characteristic of composite fault diagnosis of a submodule of a modular multilevel converter, and a feature extraction unit of the model is improved into a structure of 1DCNN fusion LSTM, so that the network feature extraction part is more detailed, and the network computing speed is higher.
Step S5 is the training of the model, and step S6 is the application of the model to the test.
In one embodiment, the step S1 adopts an overlapped sampling method to perform the sampling of the current signal, which includes:
and selecting a point on the original current signal as a collection starting point, collecting data points with a certain length each time, moving the data points with a certain length backwards each time the collection is completed, and continuously collecting at a new starting point position until all the remaining data points are collected.
In particular, the use of overlapping sampling to acquire the time domain current signal allows the construction and expansion of the sample set.
In one embodiment, the preprocessing in step S2 is a normalization processing, specifically: normalizing the input data to a [0, 1] interval by using a dispersion normalization method, wherein the calculation formula is as follows:
wherein, x' represents normalized data, x represents original data, and max (x) and min (x) represent the maximum value and the minimum value of the original data respectively.
Before the original three-phase alternating current and the internal circulation signal are input into the model, the method normalizes the input data to a [0, 1] interval by using a dispersion standardization method.
In one embodiment, in the data set divided in step S3, a 16-dimensional one-hot encoding vector is used for each sample to produce a label, specifically: and representing each label as an all-zero vector, wherein the corresponding element of the label index is 1, and the rest elements are 0.
Specifically, the sub-module composite fault means that 1 sub-module has an open-circuit fault on any different 2 bridge arms of the MMC, and therefore, the fault type is
There are 16 fault samples in total, plus normal, and the number of samples for each fault type is equal. To facilitate the calculation of the loss function, a 16-dimensional one-hot encoding vector is used for each sample to make a label. For example, [0,0,0,0,0,0,0,0,0,0,0,0,0,0, 1]Representing a first type of fault type.
In one embodiment, the feature extraction layer in step S4 includes a one-dimensional convolution layer, a one-dimensional maximum pooling layer, and an LSTM layer, and the weight parameters are initialized by using a normal distribution with a mean value of 0 and a standard deviation of 0.5 for the one-dimensional convolution layer; the LSTM layer is initialized with uniformly distributed initialization weight parameters distributed between (-0.1, 0.1).
In one embodiment, the dynamic routing algorithm adjusts parameters of the main capsule layer and the digital capsule layer through an iterative process, so that the information of the main capsule layer is sent to the digital capsule layer, and the calculation formula is as follows:
uj|i＝Wijui；
wherein u isiIs the output of the previous layer of vector neurons, i.e. the lower layer features, WijIs an attitude matrix between the lower-level features and the higher-level features, uj|iIs to infer a prediction vector of a high-level feature from a low-level feature, cijFor the coupling coefficients determined by the dynamic routing algorithm, the neurons used to control the inputs automatically select the best path to transmit to the next layer of neurons, vjIs the output vector, s, of the digital capsule layerjIs the output vector of the main capsule layer.
In particular uiAnd obtaining the low-level features for the output of the vector neurons in the previous layer, namely the output of the feature extraction layer.
In one embodiment, the coupling coefficient is calculated by the formula:
b'i,j＝bi,j+vjuj|i；
wherein, b'i,jFor iteratively updated bias coefficients, bi,jIs the original bias coefficient.
In one embodiment, in step S5, the total loss is calculated by combining the interval loss and the reconstruction loss when constructing the loss function, wherein the interval loss function is expressed as:
Lk＝Tkmax(0,m+-||vk||)2+λ(1-Tk)max(0,||vk||-m-)2；
wherein k is the number of classifications; vkRepresenting the predicted probability, T, of the kth fault typekIndicating a function for classification, i.e. T when the current data is of class kk1, otherwise Tk＝0；m+The upper bound is used for punishing false positive, namely predicting that k types exist, and true does not exist, namely identifying but wrong; m is-The lower bound is used for punishing false negatives, namely predicting that k types do not exist, but really exist, namely are not identified;
the reconstruction loss calculation method is that a 3-layer full-connection network is constructed after a capsule layer, output data with the same dimensionality as original input data is obtained, the square sum of the difference between the final output and the value on the initially input unit is used as a loss value, and a reconstruction loss calculation formula is as follows:
wherein n represents the number of data points contained in the input time series sample, α is the proportion of the total loss of the reconstruction loss, and fiOutput value, y, of each sample data point during reconstructioniThe real value corresponding to each sample data.
Specifically, the 3-layer fully-connected network behind the digital capsule layer is a decoder network, the number of fully-connected layers is too small to decode the original input, too many parameters are caused to be too many, and the complexity of the network is increased, so that 3 are selected in the embodiment. The role of the decoder is to calculate the reconstruction loss. Since the sum of the output probabilities of the capsule network is not always equal to 1, i.e. the capsule network has the capability of identifying multiple objects simultaneously, i.e. the capsule network allows multiple classes to exist simultaneously, the loss function of the improved capsule network is composed of two parts, and the total loss is calculated by combining the interval loss and the reconstruction loss, and the total loss is the interval loss plus the reconstruction loss. Fig. 6 is a schematic diagram illustrating a capsule network reconstruction loss calculation according to an embodiment of the present invention.
In one embodiment, in the training process of step S5, the model weight parameters are updated by back propagation algorithm, and the key hyper-parameters affecting the model performance are optimizer type, iteration number and batch size.
After the training of step 5, the model weight parameters are updated by the back propagation algorithm. The key hyper-parameters influencing the model performance are the type of the optimizer, the iteration times and the batch size, and the three hyper-parameters are respectively set on the verification set after independent evaluation, so that a group of training models after model hyper-parameter optimization is obtained. And then testing on an unseen data set, namely a test set, by using the trained optimized model to obtain a fault diagnosis result.
Compared with the prior art, the invention has the beneficial effects that:
(1) the method adopts an overlapped sampling method to construct and expand a sample set, and has higher feature extraction capability and fault diagnosis precision under the working condition of insufficient sample data;
(2) after the high-level MMC has a sub-module composite fault, the high-level MMC still has high fault diagnosis precision under the condition that the time domain fault characteristics are not obvious;
(3) the single convolution characteristic extraction structure of the capsule network is improved: 1DCNN is used as a feature extraction unit in front of LSTM because 1DCNN can convert longer input sequence data into shorter sequence data composed of higher features and then take these sequences composed of extracted features as input to LSTM. While the computation cost of LSTM is very high when directly processing long sequence, and the computation cost of 1DCNN is very low. Therefore, the improved capsule network feature extraction structure integrates the speed and light weight of 1DCNN and has the sequential sensitivity of LSTM extraction sequence data.
(4) The MMC fault detection system has the advantages that excessive sensors are not needed, the whole MMC fault detection system only needs 6 current sensors, the cost is low, a large amount of complex calculation is not needed, and the MMC fault detection system is simple and easy to realize.
The present invention is further illustrated by the following examples, which are not to be construed as limiting the invention.
The embodiment provides a composite fault diagnosis method for a submodule of a modular multilevel converter, which specifically comprises the following steps:
I. sampling three-phase output current of the MMC and three-phase internal circulating current signals;
the main circuit topology structure of the three-phase MMC in step I is shown in fig. 1, the MMC comprises 6 bridge arms, and each bridge arm comprises a bridge arm reactance L and N series submodules. The upper and lower bridge arms are combined into a phase unit, and 6 groups of bridge arms are combined together to form the three-phase current converter. The submodule structure is shown in fig. 2, each submodule is composed of 2 insulated gate bipolar transistors (T1 and T2), an anti-parallel diode and a floating capacitor which are connected in parallel. Wherein the three-phase output current is ij(j ═ a, b, c) and the internal circulation of the three phases is idiff.j(j ═ a, b, c). The upper and lower bridge arm current calculation formula is:
in the formula ipj，injAre phase upper and lower bridge arm currents, idcIs a direct side current, idiff·jIs a j-phase circulating current.
The three-phase internal circulation calculation formula is as follows:
idiff.j＝(ipj+inj)/2(j＝a,b,c)
the invention adopts an overlapped sampling method to construct and expand a sample set. As shown in fig. 3. And selecting a certain point on the original current signal as a collection starting point, collecting 1024 data points each time, moving backwards 256 data points each time the collection is completed, and continuously collecting at a new starting point position until all the remaining data points are collected.
II, normalizing the collected original current signals;
in the step II, the collected original current signal is normalized to a [0, 1] interval by adopting a dispersion normalization method, and the calculation formula is specifically as follows
III, segmenting an original data set into a training set, a verification set and a test set;
in step III, the original current data set is randomly divided into a training set, a validation set, and a test set. The fault type comprises 16 fault types, one is a normal state, and the remaining 15 fault states are respectively: any 2 bridge arms in 6 bridge arms have one submodule open-circuit fault, and the total is
A composite fault condition is initiated. The number of samples of each fault type in the training set is 100, the number of samples of each fault type in the verification set is 30, and the number of samples of each fault type in the test set is 30; the total number of samples is (100+30+30) × 16 ═ 2560. In order to calculate the loss function conveniently, a 16-dimensional one-hot encoding vector is used for making a label for each sample, that is, each label is represented as an all-zero vector, and only the element corresponding to the label index is 1. Fig. 4 is a time domain waveform diagram of three-phase ac current and three-phase internal circulating current in a normal state according to an embodiment of the present invention.
And IV, building an improved capsule network model, setting model structure parameters and initializing model weight parameters.
In step IV, the improved capsule network model structure is shown in fig. 5, the first layer of the network is a one-dimensional convolutional neural network layer, the size of the convolutional kernel is 6 × 1, the number of channels is 64, and filling is adopted. This is followed by a one-dimensional maximum pooling layer with a pooling window size of 2 x 2 and a stride of 2. The third layer of the network is a long-time memory network layer, and the output dimension of the third layer of the network is 128. The output dimension of the entire capsule layer is 16 x 16. The network front end adopts a feature extraction structure combining 1DCNN and LSTM, so that the calculation cost of the model is greatly reduced while the information is fully extracted. The back end adopts a main capsule layer and a digital capsule layer structure, realizes the transfer of the characteristic vector through a dynamic routing algorithm, and adopts a Relu function as an activation function. Initializing weight parameters of convolution layers in the network by adopting normal distribution with the average value of 0 and the standard deviation of 0.5; and initializing weight parameters by uniformly distributing the initialization weight parameters among (-0.1,0.1) for the circulation layers in the network.
The core of the capsule network is a dynamic routing algorithm, which continuously adjusts the parameters of the main capsule layer and the digital capsule layer through an iterative process, so that the information of the main capsule layer is sent to the digital capsule layer. Can be obtained by adjusting the coupling coefficient cijThe size of (2) to allow the input neuron to automatically select the best path to transmit to the next layer of neurons. The loss function of the capsule network is composed of two parts, and the total loss is calculated by combining the interval loss and the reconstruction loss.
V, training the model by using a training set, and evaluating the hyper-parameters of the model on a verification set to obtain a group of training models with optimal hyper-parameters of the model;
in the step V, the key hyper-parameters influencing the model performance are the type of the optimizer, the iteration times and the batch size, and the three hyper-parameters are independently evaluated on a verification set respectively to obtain a group of training models after the model hyper-parameters are optimized.
VI, testing the trained model by using the test set data, and outputting a fault diagnosis result;
fig. 7 is a flowchart of a method for diagnosing a composite fault of a sub-module of a modular multilevel converter based on an improved capsule network according to this embodiment. After the model is trained, namely the model weight parameters are updated through a back propagation algorithm, and the model hyper-parameters are set after being evaluated on a verification set. And testing on an unseen data set, namely a test set, by using the trained model to obtain a fault diagnosis result.
Finally, it should be noted that the above-mentioned embodiments are only preferred embodiments of the present invention, and not intended to limit the present invention, and although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications and equivalents can be made in the technical solutions described in the foregoing embodiments, or some technical features thereof can be replaced. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (9)
1. A composite fault diagnosis method for a submodule of a modular multilevel converter is characterized by comprising the following steps:
s1: sampling an original current signal of the modular multilevel converter, wherein the original current signal comprises a three-phase output current and a three-phase internal circulating current signal;
s2: preprocessing the acquired original current signal;
s3: dividing a data set formed by the preprocessed original current signals into a training set, a verification set and a test set;
s4: building a capsule network model, and initializing weight parameters of the capsule network model, wherein the capsule network model comprises a feature extraction layer, a main capsule layer and a digital capsule layer, the feature extraction layer is formed by fusing a one-dimensional convolutional neural network and a long-time and short-time memory network and is used for performing feature extraction on an input current signal, the main capsule layer is used for obtaining low-level features based on the features of the feature extraction layer, the digital capsule layer is used for obtaining high-level features and information used for representing fault diagnosis results through the output of the main capsule layer, and the transmission of feature vectors is realized through a dynamic routing algorithm between the digital capsule layers of the main capsule layer;
s5: training the capsule network model by using a training set, constructing a loss function, and evaluating model hyper-parameters on a verification set to obtain a group of models with optimal model hyper-parameters as well as trained models;
s6: and testing the trained model by using the test set data, and outputting a fault diagnosis result.
2. The composite fault diagnosis method according to claim 1, wherein the step S1 of sampling the current signal by using an overlap sampling method comprises:
and selecting a point on the original current signal as a collection starting point, collecting data points with a certain length each time, moving the data points with a certain length backwards each time the collection is completed, and continuously collecting at a new starting point position until all the remaining data points are collected.
3. The composite fault diagnosis method according to claim 1, wherein the preprocessing in step S2 is a normalization processing, specifically: normalizing the input data to a [0, 1] interval by using a dispersion normalization method, wherein the calculation formula is as follows:
wherein, x' represents normalized data, x represents original data, and max (x) and min (x) represent the maximum value and the minimum value of the original data respectively.
4. The method according to claim 1, wherein in the data set partitioned in step S3, a 16-dimensional one-hot encoding vector is used for each sample to make a label, specifically: and representing each label as an all-zero vector, wherein the corresponding element of the label index is 1, and the rest elements are 0.
5. The composite fault diagnosis method according to claim 1, wherein the feature extraction layer in step S4 includes a one-dimensional convolution layer, a one-dimensional maximum pooling layer, and an LSTM layer, and the weight parameter is initialized for the one-dimensional convolution layer using a normal distribution with a mean value of 0 and a standard deviation of 0.5; the LSTM layer is initialized with uniformly distributed initialization weight parameters distributed between (-0.1, 0.1).
6. The composite fault diagnosis method according to claim 1, wherein the dynamic routing algorithm adjusts parameters of the main capsule layer and the digital capsule layer through an iterative process, so that the information of the main capsule layer is sent to the digital capsule layer, and the calculation formula is as follows:
uj|i＝Wijui；
wherein u isiIs the output of the previous layer of vector neurons, i.e. the lower layer features, WijIs an attitude matrix between the lower-level features and the higher-level features, uj|iIs to infer a prediction vector of a high-level feature from a low-level feature, cijFor the coupling coefficients determined by the dynamic routing algorithm, the neurons used to control the inputs automatically select the best path to transmit to the next layer of neurons, vjIs the output vector, s, of the digital capsule layerjIs the output vector of the main capsule layer.
7. The composite fault diagnosis method according to claim 6, characterized in that the coupling coefficient is calculated by the formula:
b'i,j＝bi,j+vjuj|i；
wherein, b'i,jFor iteratively updated bias coefficients, bi,jIs the original bias coefficient.
8. The composite fault diagnosis method according to claim 1, wherein in step S5, the total loss is calculated by combining the interval loss and the reconstruction loss in constructing the loss function, wherein the interval loss function is expressed as:
Lk＝Tkmax(0,m+-||vk||)2+λ(1-Tk)max(0,||vk||-m-)2；
wherein k is the number of classifications; vkIndicates the k < th > reasonPredicted probability of barrier type, TkIndicating a function for classification, i.e. T when the current data is of class kk1, otherwise Tk＝0；m+The method is an upper bound and is used for punishing false positive, namely predicting that k types exist and not really exist; m is-The lower bound is used for punishing false negatives, namely predicting that k types do not exist and really exist;
the reconstruction loss calculation method is that a 3-layer full-connection network is constructed after a capsule layer, output data with the same dimensionality as original input data is obtained, the square sum of the difference between the final output and the value on the initially input unit is used as a loss value, and a reconstruction loss calculation formula is as follows:
wherein n represents the number of data points contained in the input time series sample, α is the proportion of the total loss of the reconstruction loss, and fiOutput value, y, of each sample data point during reconstructioniThe real value corresponding to each sample data.
9. The composite fault diagnosis method of claim 1, wherein in the training process of step S5, the model weight parameters are updated by back propagation algorithm, and the key hyper-parameters affecting the model performance are optimizer type, iteration number and batch size.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111499359.4A CN114296005A (en) | 2021-12-09 | 2021-12-09 | Modular multilevel converter submodule composite fault diagnosis method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111499359.4A CN114296005A (en) | 2021-12-09 | 2021-12-09 | Modular multilevel converter submodule composite fault diagnosis method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114296005A true CN114296005A (en) | 2022-04-08 |
Family
ID=80968167
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111499359.4A Pending CN114296005A (en) | 2021-12-09 | 2021-12-09 | Modular multilevel converter submodule composite fault diagnosis method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114296005A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117951615A (en) * | 2024-03-26 | 2024-04-30 | 南开大学 | Classification recognition method for multi-category time sequence signals |
| CN118169510A (en) * | 2024-03-11 | 2024-06-11 | 江苏优亿诺智能科技有限公司 | Power distribution network fault positioning method and system based on artificial intelligence |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109118479A (en) * | 2018-07-26 | 2019-01-01 | 中睿能源（北京）有限公司 | Defects of insulator identification positioning device and method based on capsule network |
| CN109655259A (en) * | 2018-11-23 | 2019-04-19 | 华南理工大学 | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks |
| CN111612030A (en) * | 2020-03-30 | 2020-09-01 | 华电电力科学研究院有限公司 | Wind turbine generator blade surface fault identification and classification method based on deep learning |
| CN111766513A (en) * | 2020-06-18 | 2020-10-13 | 电子科技大学 | Capsule network-based variable-working-condition multi-fault diagnosis method for three-phase induction motor |
| CN112488011A (en) * | 2020-12-04 | 2021-03-12 | 黄冈师范学院 | Fault classification method for modular multilevel converter |
-
2021
- 2021-12-09 CN CN202111499359.4A patent/CN114296005A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109118479A (en) * | 2018-07-26 | 2019-01-01 | 中睿能源（北京）有限公司 | Defects of insulator identification positioning device and method based on capsule network |
| CN109655259A (en) * | 2018-11-23 | 2019-04-19 | 华南理工大学 | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks |
| CN111612030A (en) * | 2020-03-30 | 2020-09-01 | 华电电力科学研究院有限公司 | Wind turbine generator blade surface fault identification and classification method based on deep learning |
| CN111766513A (en) * | 2020-06-18 | 2020-10-13 | 电子科技大学 | Capsule network-based variable-working-condition multi-fault diagnosis method for three-phase induction motor |
| CN112488011A (en) * | 2020-12-04 | 2021-03-12 | 黄冈师范学院 | Fault classification method for modular multilevel converter |
Non-Patent Citations (3)
| Title |
|---|
| 张夕拾: "基于深度学习的语音疲劳检测", 中国优秀硕士学位论文全文数据库工程科技Ⅱ辑, 15 March 2021 (2021-03-15), pages 31 * |
| 梁海涛等: "基于ARCN模型的轴承故障诊断", 航空动力学报, 15 November 2021 (2021-11-15), pages 2 * |
| 陈志刚等: "深度学习方法在滚动轴承故障识别中的应用", 31 August 2020, 中国原子能出版社, pages: 164 - 165 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118169510A (en) * | 2024-03-11 | 2024-06-11 | 江苏优亿诺智能科技有限公司 | Power distribution network fault positioning method and system based on artificial intelligence |
| CN117951615A (en) * | 2024-03-26 | 2024-04-30 | 南开大学 | Classification recognition method for multi-category time sequence signals |
| CN117951615B (en) * | 2024-03-26 | 2024-06-21 | 南开大学 | Classification recognition method for multi-category time sequence signals |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107271925B (en) | Five level converter Fault Locating Method of modularization based on depth convolutional network | |
| CN114296005A (en) | Modular multilevel converter submodule composite fault diagnosis method | |
| CN114755745A (en) | Hail weather identification and classification method based on multi-channel depth residual shrinkage network | |
| CN113673346B (en) | Motor vibration data processing and state identification method based on multiscale SE-Resnet | |
| CN109829916A (en) | A kind of Diagnosis Method of Transformer Faults based on CNN | |
| CN110726898B (en) | Power distribution network fault type identification method | |
| Xia et al. | A deep Siamese postclassification fusion network for semantic change detection | |
| CN115688880A (en) | Open-circuit fault diagnosis method and device for switching device of modular multilevel converter | |
| CN111860823B (en) | Neural network training method, neural network image processing method, neural network training device, neural network image processing equipment and storage medium | |
| CN116842459B (en) | Electric energy metering fault diagnosis method and diagnosis terminal based on small sample learning | |
| CN112147432A (en) | BiLSTM module based on attention mechanism, transformer state diagnosis method and system | |
| CN114065307A (en) | Hardware Trojan horse detection method and system based on bipartite graph convolutional neural network | |
| CN112507778B (en) | Loop detection method of improved bag-of-words model based on line characteristics | |
| CN116186633A (en) | Power consumption abnormality diagnosis method and system based on small sample learning | |
| CN115219845A (en) | Power grid fault diagnosis system and method | |
| CN115891741A (en) | Remote fault early warning method and device suitable for electric vehicle charging process | |
| CN117272102A (en) | Transformer fault diagnosis method based on double-attention mechanism | |
| Xin et al. | A deep learning and softmax regression fault diagnosis method for multi-level converter | |
| CN109901064B (en) | ICA-LVQ-based high-voltage circuit breaker fault diagnosis method | |
| Xu et al. | TBMF framework: A transformer-based multilevel filtering framework for PD detection | |
| CN117330890B (en) | Power transmission line fault diagnosis system and method | |
| CN117272230A (en) | Non-invasive load monitoring method and system based on multi-task learning model | |
| CN114089181B (en) | XG-Boost-based MMC switching tube open-circuit fault detection method | |
| CN116401610A (en) | Inverter parameter fault diagnosis method based on depth residual error network and joint distribution | |
| CN115661539A (en) | Less-sample image identification method embedded with uncertainty information |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |