CN115951268A - Convolutional neural network converter fault diagnosis method based on Incepton - Google Patents
Convolutional neural network converter fault diagnosis method based on Incepton Download PDFInfo
- Publication number
- CN115951268A CN115951268A CN202211712021.7A CN202211712021A CN115951268A CN 115951268 A CN115951268 A CN 115951268A CN 202211712021 A CN202211712021 A CN 202211712021A CN 115951268 A CN115951268 A CN 115951268A
- Authority
- CN
- China
- Prior art keywords
- data
- phase
- fault
- converter
- neural network
- 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.)
- Withdrawn
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Abstract
The invention discloses an Incepton-based fault diagnosis method for a convolutional neural network converter, and belongs to the field of diagnosis of a fusion inner vacuum chamber power supply. Current signal I acquired by current sensor at three-phase network side of three-phase PWM converter a ,I b ,I c And after the data is processed by the data preprocessing module, power tube open-circuit fault diagnosis is carried out on the power tube of the three-phase PWM converter through the convolution neural network model based on the increment, the state of the power tube of the converter is judged, and the fault position is determined. The system architecture of the whole system is mainly divided into two paths, namely data training and data modeling. The model building method of transfer learning is adopted, and experimental data and simulation data are combined to be used as model trainingA data set. By using the method, the open-circuit fault of the power tube of the three-phase power supply system of the fusion inner vacuum chamber can be quickly and accurately diagnosed in real time based on a small amount of experimental data, sufficient time allowance is reserved for the next fault-tolerant process, and unnecessary shutdown of the superconducting magnet fusion power supply is avoided.
Description
Technical Field
The invention relates to the field of diagnosis of a power supply of a vacuum chamber in fusion, in particular to a convolution neural network converter fault diagnosis method based on inclusion.
Background
An experimental advanced superconducting Tokamak inner vacuum chamber power supply needs a PWM voltage source converter to work in a rectification mode or an inversion mode. Due to the complexity of the drive system and the diversity of operating conditions such as electromagnetic interference, high temperatures, high electrical stress, etc., PWM converter systems are prone to catastrophic failure. If a fault occurs, unbalanced operation and shutdown will be performed, which may lead to an unexpected loss of control of plasma vertical stability and even plasma breakdown.
While investigations have shown that 31% of the failures in the power conversion system are due to failures of the power semiconductors. Generally, semiconductor switching device failures are classified into short-circuit failures and open-circuit failures. Short circuit faults often cause over-currents. Hardware devices such as fuses or circuit breakers can detect and prevent further damage from short circuit faults. Open circuit faults, however, result in current imbalances that cannot be immediately detected, possibly leading to secondary faults. The diagnosis method in the prior art can only be applied to the rectification mode or the inversion mode of the pulse width modulation voltage source converter, and can not simultaneously diagnose the rectification or inversion two-quadrant working mode. Therefore, a convolutional neural network converter fault diagnosis method based on inclusion is designed to solve the problems.
Disclosure of Invention
In view of the existing problems, the present invention aims to provide a convolutional neural network converter fault diagnosis method based on inclusion, so as to solve the problems proposed in the background art.
In order to achieve the purpose, the invention provides the following technical scheme:
the convolutional neural network converter fault diagnosis method based on Inception comprises a fusion power supply, wherein the fusion power supply is formed by connecting a plurality of power supply units in series or in parallel, the front stage of each power supply unit is formed by a three-phase PWM (pulse width modulation) converter, the fusion PWM converter refers to the front stage three-phase PWM converter of each power supply unit of the fusion power supply, and the three-phase PWM converter works and rectifies or inverts in two quadrants, and the three-phase network side current is I a ,I b ,I c The diagnostic method comprises the following steps:
the method comprises the following steps: analyzing the open-circuit fault condition of a power tube of the fusion PWM converter;
the fusion PWM converter can work in two quadrants: a rectification state and an inversion state; the open circuit fault of the PWM converter has 22 conditions;
step two: determining a coding mode of the open-circuit fault of a power tube of the fusion PWM converter;
the fault state is represented by adopting a one-hot coding mode, the states of six power switching tubes are respectively mapped to binary vectors, 0 represents a healthy state, and 1 represents an open-circuit fault of the power switching tubes;
step three: analyzing the characteristic quantity of the open-circuit fault of the power tube of the fusion PWM converter;
step four: collecting historical data of open-circuit faults of power tubes of the fusion PWM converter as model training and verification data;
step five: carrying out data preprocessing on the collected historical data;
the data preprocessing comprises the following steps: a resampling block, a data expansion block, a data standardization block and a data shuffling block;
step six: establishing a convolutional neural network model architecture based on the inclusion;
the core part of the convolutional neural network model based on the inclusion is an inclusion block;
step seven: using current data I of three-phase network side current sensor of PWM converter working in rectification or inversion two-quadrant state a ,I b ,I c Preprocessing the input signal for inputting the signal, importing the data preprocessed in the step five into the neural network model training in the step six, and selecting an optimal training model in multiple training;
dividing the preprocessed data into a training set and a verification set according to a certain proportion, inputting the training set and the verification set into a convolution neural network model based on the inclusion, obtaining a model loss curve, and selecting an optimal training model by means of callback and check points according to the loss curve of the training set and the loss curve of the verification set;
step eight: inputting the detected real-time input signals and the standard data after the preprocessing block into the optimal model trained in the seventh step, obtaining the required fault binary vector, obtaining the fault state quantity of each switching tube, inputting the preprocessed data into the convolution neural network model based on the increment, diagnosing the open circuit fault of the power tube of the three-phase PWM converter in the fusion inner vacuum chamber in real time, judging the state of each power tube of the converter, and determining the fault position.
As a further scheme of the invention: the open-circuit fault condition of the power tube of the PWM converter in the second step includes the following 22 conditions: 1 kinds of switch tubes are in a healthy state, 6 kinds of single switch tube open circuit faults, 3 kinds of double switch tube open circuit faults, two switches with faults are positioned at the same phase, 6 kinds of double switch tube open circuit faults, two switches with faults are positioned at one side of a bridge arm, and 6 kinds of double switch tube open circuit faults, two switches with faults are not positioned at the same phase and one side of the bridge arm.
As a further scheme of the invention: in the third step, under the condition of open-circuit fault of a power tube of the three-phase PWM converter of the fusion inner vacuum chamber, the three-phase PWM converter respectively works in a rectification or inversion two-quadrant running state, and the current characteristics of three-phase network sides of the three-phase PWM converter are different; the positive current direction is from the net side to the rectifier; the network side A or B and C phases are respectively connected to bridge arms 1 or 2 and 3 of the three-phase PWM converter, wherein the phase A is connected with the bridge arm 1, the phase B is connected with the bridge arm 2, and the phase C is connected with the bridge arm 3; the positive direction current circulation channel of the network side A or B, C phase is a lower bridge switch tube of the bridge arm 1 or 2, 3, and the negative direction current circulation channel of the network side A or B, C phase is an upper bridge switch tube of the bridge arm 1 or 2, 3; when the circuit is not in fault, the current of each phase flows normally, a three-phase symmetrical sine wave state is presented, and a small amount of network side harmonic waves are accompanied; when the upper or lower power tubes of the bridge arms 1, 2 and 3 in the circuit generate single-tube open-circuit faults, diodes connected in parallel at two ends of a fault switch tube can be used as a rectifier component to continuously operate when the PWM converter works in a rectification state, negative or positive direction currents of phases A, B and C on the corresponding network side intermittently circulate, and the rectifier with the fault power tubes alternately works in a controlled mode and an uncontrolled mode; when the PWM converter works in an active inverter state, the fault power tube and the diodes connected in parallel at the two ends of the fault power tube do not work, and the negative or positive direction of the corresponding network side A or B or C phase current cannot flow; when the circuit has double-switch faults, the current with the worst change can be understood as the superposition of single-switch tube faults; therefore, the three-phase current of one power frequency period is selected as a fault characteristic vector.
As a further scheme of the invention: step four, a transfer learning method is adopted, and a training set and a verification set are made through a large amount of simulation data and a small amount of experimental data; experimental data are obtained on a plurality of real three-phase PWM converters, and open-circuit faults of the converter switching tubes when corresponding diodes of the converter switching tubes are conducted are simulated by respectively setting driving signals of 6 switching tubes of the three-phase PWM converters to be 0V; the simulation data is obtained by using circuit simulation software, and the driving signals of 6 switching tubes of the three-phase PWM converter in the simulation software are respectively set to be 0V, so that the open-circuit fault of the switching tubes of the converter is simulated when corresponding diodes are conducted.
As a further scheme of the invention: the resampling block adopts a simple moving average method to eliminate high-frequency fluctuation and resample the data set, the simple moving average method calculates the average value of the specific time range value, and the simple moving average method uses the information of the previous time to prevent data leakage;
the data enhancement block effectively multiplies the original data by n times through phase shifting when training a neural network model;
the data normalization block: statistically, the z-score refers to the number of standard deviations of the value of the raw score above or below the observed or measured mean, the z-score is calculated by z = (x- μ)/σ, where μ is the mean of the population; σ is the standard deviation of the population, and the z-score will be used for normalization when all source data is obtained;
the data shuffling block shuffles the order of the entire data set using a shuffling method to improve generalization capability.
As a further scheme of the invention: in the sixth step, the inclusion block is a combination of all layers, and an output filter bank of the inclusion block is connected in series to form a single output vector to form the input of the next stage; the inclusion block consists of four parallel paths; the first three paths extract information from different spatial sizes using convolutional layers with window sizes of 1 × 1, 3 × 3, and 5 × 5; the middle two paths perform a 1 × 1 convolution on the input to reduce the number of channels; the fourth path changes the number of channels using a 3 × 3 max pool layer, then a 1 × 1 convolutional layer; these four paths all use appropriate padding to make the input and output have the same height and width; finally the outputs are connected along the channel dimensions and constitute the block outputs.
As a further scheme of the invention: performing data preprocessing on the three-phase current data of the switching tube in the rectification or inversion two-quadrant state in the step seven; dividing the preprocessed data into a training set and a verification set according to a certain proportion, inputting the training set and the verification set into a convolution neural network model based on the inclusion, obtaining a model loss curve, and selecting an optimal training model by means of callback and check points according to the loss curve of the training set and the loss curve of the verification set.
As a further scheme of the invention: and step eight, stacking a plurality of convolution layers, a maximum pool layer, a global average pool layer and a complete connection layer together by the convolution neural network model based on the inclusion so as to realize the function of feature extraction.
Compared with the prior art; the invention has the beneficial effects that:
1. the invention discloses a fusion PWM converter open-circuit fault diagnosis method based on an Incepration convolutional neural network, which is characterized in that three-phase network side current of a PWM converter is used as input of a neural network model, a convolutional neural network model with an Incepration block as a core is used as a model basis, and a plurality of convolutional layers, a maximum pool layer, a global average pool layer and a complete connection layer are stacked together to realize a feature extraction function. By adopting the framework, the calculation resources of the conventional methods such as a convolutional neural network and the like can be better utilized, the diagnostic performance is improved, the system can be subjected to quick and accurate real-time fault detection, the secondary fault of the converter is avoided as much as possible, sufficient time margin is reserved for the next fault-tolerant process, safety accidents and economic loss caused by the fault of the converter are avoided, and the stability and the working efficiency of the converter system in the fusion field are greatly improved;
2. the training of the convolutional neural network model of the present invention requires a large amount of data. Compared with simulation data, experimental data are more difficult to obtain, the method adopts a transfer learning method, and a large amount of simulation data and a small amount of experimental data are used as a training set and a verification set;
3. the fault diagnosis method is simple, and has low requirement on the performance of the running CPU; an additional hardware circuit is not needed, and the original current sensor in the system is used for extracting a fault characteristic vector, so that the cost is saved; the expandability is strong, and the open-circuit fault of the power tube of the three-phase PWM converter with various parameters can be diagnosed; the single-tube and double-tube open-circuit faults of the three-phase PWM converter can be accurately identified in real time; the open-circuit fault of the switching tube of the three-phase PWM converter working in two quadrants of rectification and inversion can be accurately identified in real time;
drawings
FIG. 1 is a topological structure and control algorithm of a fusion PWM converter in the invention;
FIG. 2 is a fusion PWM converter topology of the present invention;
FIG. 3 is a control algorithm of a fusion PWM converter in the present invention;
FIG. 4 is a topology of an inclusion block of the present invention;
FIG. 5 is an architecture based on the inclusion convolutional neural network in the present invention;
FIG. 6 is a fusion PWM converter three-phase network side current data neural network structure in the invention.
Detailed Description
The drawings in the embodiments of the invention will be combined; the technical scheme in the embodiment of the invention is clearly and completely described; obviously; the described embodiments are only some of the embodiments of the invention; rather than all embodiments. Based on the embodiments of the invention; all other embodiments obtained by a person skilled in the art without making any inventive step; all fall within the scope of protection of the present invention.
The invention relates to a fault diagnosis method for a power switch tube of a voltage type PWM converter of a Tokamak power supply system. Especially for diagnosing the open-circuit fault of the single switch and the double switch of the power tube of the PWM converter for rectifying or inverting two-quadrant operation in the application field of a fusion inner vacuum chamber power supply. The system is used for detecting the open-circuit fault of the power tube, which is easy to cause secondary fault and is not easy to be discovered immediately by a power supply system. Belongs to the fields of machine learning, power supply and fusion.
According to the embodiment of the invention, please refer to fig. 1-6, which propose a fusion PWM converter open-circuit fault diagnosis method based on inclusion convolution neural network, the fusion power supply is composed of a plurality of power supply units connected in series or in parallel, wherein the front stage of each power supply unit is composed of a three-phase PWM converter, and the fusion PWM converter refers to the front stage three-phase PWM converter of each power supply unit of the fusion power supply. The three-phase PWM converter needs to work, rectify or invert two-quadrant state, and the three-phase network side current is I a ,I b ,I c ，
The fusion PWM converter power tube open-circuit fault diagnosis method is used for working at three-phase current I on the grid side of the PWM converter in a rectification or inversion two-quadrant state a ,I b ,I c The input signals are respectively used as input signals for fault diagnosis, the input signals pass through a preprocessing block, preprocessed data are input into a convolutional neural network model based on the addition, open-circuit faults of power tubes of the three-phase PWM converter of the vacuum chamber in the fusion are diagnosed in real time, states of the power tubes of the converter are judged, and fault positions are determined. The method comprises the following steps:
the method comprises the following steps: and analyzing the open-circuit fault condition of the power tube of the fusion PWM converter.
The fusion PWM converter can work in two quadrants: a rectifying state and an inverting state.
The fault types specifically include the following 22 cases: each switch tube is in a healthy state (1 type), a single switch tube is in an open-circuit fault (6 types), double switch tubes are in an open-circuit fault and two switches in the fault are in the same phase (3 types), double switch tubes are in an open-circuit fault and two switches in the fault are positioned on one side (6 types) of a bridge arm, and double switch tubes are in an open-circuit fault and two switches in the fault are not in the same phase and are not positioned on one side (6 types) of the bridge arm. Since the distortion of the three-phase current caused by the fault is different, 21 × 2=42 currents are used in the fault.
Step two: and determining the encoding mode of the open-circuit fault of the power tube of the fusion PWM converter.
And the fault state is represented by adopting a one-hot coding mode, and the fault variable is converted into a simple form which is easily accepted by a machine learning algorithm. The specific implementation method is to map the states of the six power switch tubes to binary vectors respectively. 0 indicates a healthy state and 1 indicates an open power tube fault.
Step three: and analyzing the characteristic quantity of the open-circuit fault of the power tube of the fusion PWM converter.
The positive current direction is from the net side to the rectifier. The phase A (or phase B and phase C) at the network side is respectively connected with a bridge arm 1 (or phase 2 and phase 3) of the three-phase PWM converter, wherein the phase A is connected with the bridge arm 1, the phase B is connected with the bridge arm 2, and the phase C is connected with the bridge arm 3. And the positive direction current flow channel of the A (or B, C) phase of the network side is a lower bridge switching tube of the bridge arm 1 (or 2, 3), and the negative direction current flow channel of the A (or B, C) phase of the network side is an upper bridge switching tube of the bridge arm 1 (or 2, 3). When the circuit is not in fault, the current of each phase flows normally, a three-phase symmetrical sine wave state is presented, and a small amount of network side harmonic waves are accompanied. When the upper (or lower) power tube of the bridge arm 1 (or 2, 3) in the circuit has a single-tube open-circuit fault, diodes connected in parallel at two ends of a fault switch tube can be used as a rectifier component to continuously operate when the PWM converter works in a rectification state, the discontinuous circulation of the current in the negative (or positive) direction of the phase A (or B, C) at the corresponding network side is realized, and the rectifier with the fault power tube alternately works in a controlled mode and an uncontrolled mode; and when the PWM converter works in the state of an active inverter, the fault power tube and the diodes connected in parallel at the two ends of the fault power tube do not work, and the negative (or positive) direction of the corresponding network side A (or B, C) phase current cannot flow. When the circuit has double-switch faults, the current with the worst change can be understood as the superposition of single-open-tube faults. Therefore, the three-phase current of one power frequency period is selected as a fault characteristic vector.
Step four: and collecting historical data of the open-circuit fault of the power tube of the fusion PWM converter as model training and verification data.
The training of the inclusion-based convolutional neural network model requires a large amount of data. Whereas the experimental data is more difficult to obtain than the simulation data. The invention adopts a transfer learning method and adopts a large amount of simulation data and a small amount of experimental data as a training set and a verification set. The experimental data are obtained on a plurality of real three-phase PWM converters, and the open-circuit fault of the converter switching tubes when corresponding diodes of the converter switching tubes are conducted is simulated by respectively setting the driving signals of 6 switching tubes of the three-phase PWM converters to be 0V. The simulation data is obtained by using circuit simulation software (such as Matlab/Simulink and Simplorer), and the simulation data simulates open-circuit faults of the switching tubes of the converter when corresponding diodes are conducted by respectively setting the driving signals of 6 switching tubes of the three-phase PWM converter in the simulation software to be 0V.
Step five: and preprocessing the collected historical data.
Before the training and verification data of the convolutional neural network model based on the inclusion is input into the model, firstly, a data preprocessing block is needed to obtain standardized data. The data preprocessing block has 4 modules: a resampling block, a data expansion block, a data standardization block and a data shuffling block.
Resampling blocks: in practical applications, there is a significant difference in the sampling rate of the original signal due to different device settings. The input signal data sets should be aligned simultaneously and the data sets should be resampled. The choice of the resampling ratio needs to be carefully considered. A lower resampling ratio implies a lower time resolution and a poorer modeling quality, while a higher resampling ratio implies a greater computational resource requirement and a lower modeling efficiency. The invention adopts a simple moving average method to eliminate high-frequency fluctuation and resamples the data set. A simple moving average method calculates the average of values for a particular time range. Since the application environment is to create online modeling, we cannot know information after the current time point in advance, and the simple moving average method uses information of the previous time to prevent data leakage.
A data expansion block: given that the amount of existing training and validation data may not satisfy the condition, the raw data is effectively multiplied by n times by phase shifting when training the neural network model.
A data standardization block: the three-phase network side currents usually have different current assignments. Therefore, when all source data is obtained, the z-score will be used for normalization. In statistics, z-score refers to the number of standard deviations of the value of the raw score above or below the observed or measured average. The z-score is calculated by z = (x- μ)/σ, where μ is the average of the population; σ is the standard deviation of the population.
Data shuffling block: neural networks have a strong learning ability. If the data is not disturbed, the model will repeatedly learn the features of the data in sequence and reach an overfitting state. Furthermore, the neural network may learn only the sequence features of the data. Therefore, there is a need to shuffle the order of the entire data set using a shuffling approach to improve generalization capability. It is worth mentioning that the simulation data and the experimental data are processed separately.
Step six: and establishing a convolutional neural network model architecture based on the inclusion.
The core part of the inclusion-based convolutional neural network model is the inclusion block. The topology of the inclusion block is shown in fig. 3. The Inceptation model is used for better utilizing the computing resources of the existing methods such as the convolutional neural network and the like and improving the diagnostic performance. The inclusion block is a combination of all layers (i.e., 1 × 1 convolutional layers, 3 × 3 convolutional layers, 5 × 5 convolutional layers), with the output filter bank connected in series into a single output vector, forming the input to the next stage. The inclusion block consists of four parallel paths. The first three paths extract information from different spatial sizes using convolutional layers with window sizes of 1 × 1, 3 × 3, and 5 × 5. The middle two paths perform a 1 x 1 convolution on the input to reduce the number of channels, which greatly reduces the complexity of the model. The fourth path changes the number of channels using a 3 × 3 max pool layer, followed by a 1 × 1 convolutional layer. All four paths use appropriate padding so that the input and output have the same height and width. Finally, the outputs are connected along the channel dimension and constitute the output of the block.
The inclusion-based convolutional neural network cannot effectively extract sufficient features using only inclusion. Therefore, in the architecture, several convolutional layers, a maximum pool layer, a global average pool layer, and a full connection layer need to be stacked together to implement the feature extraction function. As the data flows, the depth of the data becomes progressively larger, and the width and height become smaller. The first two blocks (block 0 and block 1) are typically convolutions that are used to identify a set of low-level features from the input sequence. These identified low-level features are then combined to reduce the dimensionality of the output and serve as an input to block 2, which uses a set of inclusion blocks and max-pool layers to identify high-level features of the previously identified low-level features. This will continue for several blocks, where each block uses input from the previous block to identify features at a higher level than the previous layer. Finally, the output of the last convolutional layer is passed to a set of fully connected layers for final classification. The architecture based on the inclusion convolutional neural network is shown in fig. 4. The fusion PWM converter three-phase network side current data neural network structure is shown in figure 5.
Step seven: and importing the data preprocessed in the fifth step into the neural network model training in the sixth step, and selecting an optimal training model in multiple training.
Dividing the preprocessed data into a training set and a verification set according to a certain proportion, inputting the training set and the verification set into a convolution neural network model based on the inclusion, obtaining a model loss curve, and selecting an optimal training model by means of callback and check points according to the loss curve of the training set and the loss curve of the verification set.
Step eight: inputting the detected real-time input signals and the standard data after the preprocessing block into the optimal model trained in the step seven, and obtaining the required fault binary vector to obtain the fault state quantity of each switching tube.
The preprocessing module of the real-time detected input signal is different from the preprocessing step of the fifth step, and only comprises a resampling block and a data standardization block.
The fusion PWM converter open-circuit fault diagnosis method based on the Inceprtion convolutional neural network takes three-phase network side current of the PWM converter as input of a neural network model, takes a convolutional neural network model with an Incepration block as a core as a model base, and stacks a plurality of convolutional layers, a maximum pool layer, a global average pool layer and a complete connection layer together so as to realize a feature extraction function. By adopting the framework, the calculation resources of the conventional convolutional neural network and other methods can be better utilized, the diagnostic performance is improved, the system can be subjected to quick and accurate real-time fault detection, the secondary fault of the converter is avoided as much as possible, sufficient time margin is reserved for the next fault-tolerant process, safety accidents and economic losses caused by the fault of the converter are avoided, and the stability and the working efficiency of the converter system in the fusion field are greatly improved.
The training of the convolutional neural network model of the present invention requires a large amount of data. Whereas the experimental data is more difficult to obtain than the simulation data. The invention adopts a transfer learning method and adopts a large amount of simulation data and a small amount of experimental data as a training set and a verification set.
The fault diagnosis method is simple, and has low requirement on the performance of the running CPU; an additional hardware circuit is not needed, and the original current sensor in the system is used for extracting the fault characteristic vector, so that the cost is saved; the expandability is strong, and the open-circuit fault of the power tube of the three-phase PWM converter with various parameters can be diagnosed; the single-tube and double-tube open-circuit faults of the three-phase PWM converter can be accurately identified in real time; the open-circuit fault of the switching tube of the three-phase PWM converter working under two quadrants of rectification and inversion can be accurately identified in real time.
Additional features and advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
To those skilled in the art; it is obvious that the invention is not restricted to the details of the above-described exemplary embodiments; and without departing from the spirit or essential characteristics of the invention; the invention can be embodied in other specific forms. Thus; from whatever point; the embodiments should be considered as exemplary; and is not limiting; the scope of the invention is indicated by the appended claims rather than by the foregoing description; all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference sign in a claim should not be construed as limiting the claim concerned.
Furthermore; it should be understood that; although the present description has been described in terms of embodiments; but not every embodiment contains only a single solution; this manner of description is by way of clarity only; the person skilled in the art will consider the description as a whole; the technical schemes in the embodiments can also be combined appropriately; forming other embodiments as will be appreciated by those skilled in the art.
Claims (8)
1. The convolutional neural network converter fault diagnosis method based on Inception comprises a fusion power supply and is characterized in that the fusion power supply is formed by connecting a plurality of power supply units in series or in parallel, the front stage of each power supply unit is formed by a three-phase PWM (pulse width modulation) converter, the fusion PWM converter refers to the front stage three-phase PWM converter of each power supply unit of the fusion power supply, and the three-phase PWM converter works in a rectifying or inverting two-quadrant state, and the three-phase network side current is I a ,I b ,I c The diagnostic method comprises the following steps:
the method comprises the following steps: analyzing the open-circuit fault condition of a power tube of the fusion PWM converter;
the fusion PWM converter can work in two quadrants: a rectification state and an inversion state; the open-circuit fault of the PWM converter has 22 conditions;
step two: determining a coding mode of the open-circuit fault of the power tube of the fusion PWM converter;
the fault state is represented by adopting a one-hot coding mode, the states of six power switching tubes are respectively mapped to binary vectors, 0 represents a healthy state, and 1 represents an open-circuit fault of the power switching tubes;
step three: analyzing the characteristic quantity of the open-circuit fault of the power tube of the fusion PWM converter;
step four: collecting historical data of open-circuit faults of power tubes of the fusion PWM converter as model training and verification data;
step five: carrying out data preprocessing on the collected historical data;
the data preprocessing comprises the following steps: a resampling block, a data expansion block, a data standardization block and a data shuffling block;
step six: establishing a convolutional neural network model architecture based on the inclusion;
the core part of the convolutional neural network model based on the inclusion is an inclusion block;
step seven: using current data I of three-phase network side current sensor of PWM converter working in rectification or inversion two-quadrant state a ,I b ,I c Preprocessing the input signal for inputting the signal, importing the data preprocessed in the step five into the neural network model training in the step six, and selecting an optimal training model in multiple training;
dividing the preprocessed data into a training set and a verification set according to a certain proportion, inputting the training set and the verification set into a convolution neural network model based on the inclusion, obtaining a model loss curve, and selecting an optimal training model by means of callback and check points according to the loss curve of the training set and the loss curve of the verification set;
step eight: inputting the detected real-time input signals and the standard data after the preprocessing block into the optimal model trained in the seventh step, obtaining the required fault binary vector, obtaining the fault state quantity of each switching tube, inputting the preprocessed data into the convolution neural network model based on the increment, diagnosing the open circuit fault of the power tube of the three-phase PWM converter in the fusion inner vacuum chamber in real time, judging the state of each power tube of the converter, and determining the fault position.
2. The inclusion-based convolutional neural network converter fault diagnosis method according to claim 1, wherein the open-circuit fault condition of the power tube of the PWM converter in the second step includes 22 conditions: 1 kinds of switch tubes are in a healthy state, 6 kinds of single switch tube open-circuit faults, 3 kinds of double switch tube open-circuit faults, two switches with faults are positioned at the same phase, 6 kinds of double switch tube open-circuit faults, two switches with faults are positioned at one side of a bridge arm, and 6 kinds of double switch tube open-circuit faults, two switches with faults are not positioned at the same phase and one side of the bridge arm.
3. The Incepration-based convolutional neural network converter fault diagnosis method according to claim 2, wherein in the third step, under the condition of open-circuit fault of the power tube of the three-phase PWM converter in the fusion inner vacuum chamber, the three-phase PWM converter respectively works in a rectifying or inverting two-quadrant running state, and the current characteristics of three-phase network sides of the three-phase PWM converter are different; the positive direction of the current is from the net side to the rectifier; the network side A or B and C phases are respectively connected with bridge arms 1 or 2 and 3 of the three-phase PWM converter, wherein the phase A is connected with the bridge arm 1, the phase B is connected with the bridge arm 2, and the phase C is connected with the bridge arm 3; the positive direction current circulation channel of the network side A or B, C phase is a lower bridge switch tube of the bridge arm 1 or 2, 3, and the negative direction current circulation channel of the network side A or B, C phase is an upper bridge switch tube of the bridge arm 1 or 2, 3; when the circuit is not in fault, the current of each phase flows normally, a three-phase symmetrical sine wave state is presented, and a small amount of network side harmonic waves are accompanied; when the upper or lower power tubes of the bridge arms 1, 2 and 3 in the circuit generate single-tube open-circuit faults, diodes connected in parallel at two ends of a fault switch tube can be used as a rectifier component to continuously operate when the PWM converter works in a rectification state, negative or positive direction currents of phases A, B and C on the corresponding network side intermittently circulate, and the rectifier with the fault power tubes alternately works in a controlled mode and an uncontrolled mode; when the PWM converter works in the state of an active inverter, the fault power tube and the diodes connected in parallel at the two ends of the fault power tube do not work, and the negative or positive directions of the corresponding network side A or B and C phase currents cannot flow; when the circuit has double-switch faults, the current with the worst change can be understood as the superposition of single-open-tube faults; therefore, the three-phase current of one power frequency period is selected as a fault characteristic vector.
4. The inclusion-based convolutional neural network converter fault diagnosis method according to claim 3, wherein the fourth step adopts a transfer learning method, and a training set and a verification set are made by using a large amount of simulation data and a small amount of experimental data; experimental data are obtained on a plurality of real three-phase PWM converters, and open-circuit faults of the converter switching tubes when corresponding diodes of the converter switching tubes are conducted are simulated by respectively setting driving signals of 6 switching tubes of the three-phase PWM converters to be 0V; the simulation data is obtained by using circuit simulation software, and the driving signals of 6 switching tubes of the three-phase PWM converter in the simulation software are respectively set to be 0V, so that the open-circuit fault of the switching tubes of the converter is simulated when corresponding diodes are conducted.
5. The inclusion-based convolutional neural network converter fault diagnosis method according to claim 4, wherein the resampling block adopts a simple moving average method to eliminate high-frequency fluctuation and resample a data set, the simple moving average method calculates an average value of a specific time range value, and the simple moving average method uses information of a previous time to prevent data leakage;
the data enhancement block effectively multiplies the original data by n times through phase shifting when training a neural network model;
the data normalization block: statistically, the z-score refers to the number of standard deviations of the value of the raw score above or below the observed or measured mean, the z-score is calculated by z = (x- μ)/σ, where μ is the mean of the population; σ is the standard deviation of the population, and the z-score will be used for normalization when all source data is obtained;
the data shuffling block shuffles the order of the entire data set using a shuffling method to improve generalization capability.
6. The inclusion-based convolutional neural network converter fault diagnosis method according to claim 5, wherein in the sixth step, the inclusion block is a combination of all layers, and an output filter bank of the inclusion block is connected in series into a single output vector to form an input of a next stage; the inclusion block consists of four parallel paths; the first three paths extract information from different spatial sizes using convolutional layers with window sizes of 1 × 1, 3 × 3, and 5 × 5; the middle two paths perform a 1 × 1 convolution on the input to reduce the number of channels; the fourth path changes the number of channels using a 3 × 3 max pool layer, then a 1 × 1 convolutional layer; these four paths all use appropriate padding to make the input and output have the same height and width; finally the outputs are connected along the channel dimensions and constitute the block outputs.
7. The inclusion-based fault diagnosis method for the convolutional neural network converter is characterized in that in the seventh step, data preprocessing is performed on three-phase current data of the switching tube in a rectification or inversion two-quadrant state; dividing the preprocessed data into a training set and a verification set according to a certain proportion, inputting the training set and the verification set into a convolution neural network model based on the inclusion, obtaining a model loss curve, and selecting an optimal training model by means of callback and check points according to the loss curve of the training set and the loss curve of the verification set.
8. The inclusion-based convolutional neural network converter fault diagnosis method according to claim 7, wherein the inclusion-based convolutional neural network model in the eighth step stacks a plurality of convolutional layers, a maximum pool layer, a global average pool layer and a fully-connected layer together to realize a feature extraction function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211712021.7A CN115951268A (en) | 2022-12-29 | 2022-12-29 | Convolutional neural network converter fault diagnosis method based on Incepton |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211712021.7A CN115951268A (en) | 2022-12-29 | 2022-12-29 | Convolutional neural network converter fault diagnosis method based on Incepton |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115951268A true CN115951268A (en) | 2023-04-11 |
Family
ID=87291101
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211712021.7A Withdrawn CN115951268A (en) | 2022-12-29 | 2022-12-29 | Convolutional neural network converter fault diagnosis method based on Incepton |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115951268A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116310599A (en) * | 2023-05-17 | 2023-06-23 | 湖北工业大学 | Power transformer fault diagnosis method and system based on improved CNN-PNN network |
-
2022
- 2022-12-29 CN CN202211712021.7A patent/CN115951268A/en not_active Withdrawn
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116310599A (en) * | 2023-05-17 | 2023-06-23 | 湖北工业大学 | Power transformer fault diagnosis method and system based on improved CNN-PNN network |
| CN116310599B (en) * | 2023-05-17 | 2023-08-15 | 湖北工业大学 | Power transformer fault diagnosis method and system based on improved CNN-PNN network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106443297B (en) | The decision tree SVM method for diagnosing faults of photovoltaic diode Clamp three-level inverter | |
| Zhang et al. | Fault diagnosis of wind turbine power converter considering wavelet transform, feature analysis, judgment and BP neural network | |
| CN103761372B (en) | A kind of multi-electrical level inverter Fault Diagnosis Strategy based on pivot analysis with more classification Method Using Relevance Vector Machines | |
| CN109782091B (en) | Multilevel inverter fault diagnosis method based on deep wavelet extreme learning machine | |
| CN102353896B (en) | Many level translators and determine the method for switchgear fault of many level translators | |
| CN108872882B (en) | Fault diagnosis device and method for three-level cascading inverter | |
| CN107632258A (en) | A kind of fan converter method for diagnosing faults based on wavelet transformation and DBN | |
| CN106093678B (en) | A method of quick and precisely diagnosing flexible HVDC transmission system converter fault | |
| CN106154103A (en) | The switching tube open fault diagnostic method of three-level inverter | |
| CN102969888A (en) | Design method for multi media card (MMC) self-defining submodule based on real time digital system (RTDS) | |
| CN115951268A (en) | Convolutional neural network converter fault diagnosis method based on Incepton | |
| Si et al. | Fault diagnosis based on attention collaborative LSTM networks for NPC three-level inverters | |
| CN111458659A (en) | Rapid diagnosis method for open-circuit fault of switching tube of cascaded H-bridge photovoltaic inverter | |
| CN105203922A (en) | Electric transmission line fault positioning method based on multiple classifiers | |
| CN108898182A (en) | A kind of MMC method for diagnosing faults based on core pivot element analysis and support vector machines | |
| CN113610119B (en) | Method for identifying power transmission line development faults based on convolutional neural network | |
| Guerrero et al. | A novel noninvasive failure-detection system for high-power converters based on SCRs | |
| CN116702060A (en) | Multi-level inverter power device fault diagnosis method | |
| Ibem et al. | Multiple open switch fault diagnosis of three phase voltage source inverter using ensemble bagged tree machine learning technique | |
| Qu et al. | An identification and location method for power quality disturbance sources in MMC converter based on KNN algorithm | |
| CN116054224A (en) | Quick evaluation method for commutation failure of multi-feed direct-current transmission system | |
| Singh et al. | Open circuit fault diagnosis and fault classification in multi-level inverter using fuzzy inference system | |
| CN112234576B (en) | Modular multilevel converter fault diagnosis method based on sliding window feature extraction | |
| CN108548669A (en) | A kind of engineer equipment driving unit fault diagnostic method and system | |
| de Oliveira et al. | Siamese neural network architecture for fault detection in a voltage source inverter |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20230411 |
|
| WW01 | Invention patent application withdrawn after publication |