Converter open circuit fault diagnosis method based on task-encoded adaptive network

Abstract

The present invention discloses a method for diagnosing open-circuit faults of converters based on a task coding adaptive network. First, a fault sample set base consisting of fault samples, fault codes and task codes of multiple types of converters is constructed, and then a task coding adaptive network is constructed. The task coding adaptive network is proposed to characterize the functional mapping relationship between task codes, fault characteristics and fault codes, so as to realize fault location of multiple converter types without repeated model training, thereby reducing the development cycle and training cost of the fault diagnosis model; secondly, in the task coding adaptive network, through the optimization design of the network structure and unified fault / task coding, the open-circuit fault diagnosis of different types of converters is promoted, thereby improving the fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of converter fault diagnosis, and more particularly, relates to a converter open circuit fault diagnosis method based on a task coding adaptive network. Background Art

[0002] Power converters are renowned for their low power loss, minimal output current harmonic distortion, compact filter size, and high electromagnetic compatibility, making them widely used in various industrial systems. However, complex operating environments make these converters susceptible to failure over extended periods of use. Research indicates that power semiconductor devices account for approximately 21% of all failures, and nearly 34% of relevant research reports identify them as the most vulnerable component. Power switch failures typically include short-circuit and open-circuit faults. In high-power fuse protection scenarios, rapidly evolving short-circuit faults often evolve into open-circuit faults. If not diagnosed promptly, open-circuit faults can severely impact the entire system and potentially lead to secondary failures and system failure. Diagnosis of open-circuit faults is challenging due to the influence of uncertain interference factors such as sensor noise, load interference, and electromagnetic pulses. Under these complex operating conditions, extracting open-circuit fault signatures becomes difficult, making rapid and accurate fault diagnosis challenging. Consequently, the development of high-performance fault diagnosis methods has attracted widespread attention in the industry.

[0003] Currently, there are three main approaches for converter open-circuit fault diagnosis: model-based, signal processing-based, and artificial intelligence (AI)-based. Model-based approaches compare measurable drive system data with estimated information from a mathematical model to localize the fault. However, for complex multivariable systems, accurately modeling is challenging. Signal processing-based approaches extract fault features from selected signals, typically current and voltage signals, using techniques such as fast Fourier transforms, wavelet transforms, Hilbert-Huang transforms, and empirical mode decomposition. Although widely used, these methods are often complex and their performance relies on expert knowledge. Finally, AI-based approaches have attracted considerable attention, as they can achieve higher fault diagnosis accuracy without relying on prior knowledge or precise model descriptions. For example, a multi-information feature fusion diagnosis method based on an attention-based collaborative stacked long short-term memory (ASLSTM) neural network has been used for open-circuit fault diagnosis in NPC inverters. Furthermore, a data-driven online fault diagnosis method based on a random forest of transient integrated features aims to timely and effectively locate open-circuit power switch faults. In addition to these, there are numerous AI-based approaches. While these approaches have achieved some success, converters vary in topology (e.g., three-level, multi-level) and serve diverse loads (motor, photovoltaic power source, grid-connected), resulting in diverse fault signatures. Current research relies on single-task models, often focusing on specific topologies and loads, and no single fault diagnosis method is applicable to open-circuit fault diagnosis in all converters. This results in traditional approaches from data collection to model deployment requiring customized solutions, significantly extending development cycles and increasing the cost of building fault diagnosis models. Recent AI research advocates multi-task learning, demonstrating that acquiring common fault knowledge from different tasks can improve the representation of fault signatures. Therefore, developing new approaches applicable to all converter fault diagnosis and training AI models with a multi-system perspective is crucial for accurate fault diagnosis and reducing development cycles and costs. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a converter open circuit fault diagnosis method based on a task coding adaptive network, which does not require repeated model training, improves the accuracy and reliability of fault diagnosis, and reduces the converter fault diagnosis model development cycle and cost.

[0005] To achieve the above-mentioned object of the invention, the present invention provides a method for diagnosing open-circuit faults of a converter based on a task-coding adaptive network, which is characterized by comprising the following steps:

[0006] (1) Construct a training dataset;

[0007] Note that each type of converter has m operating conditions and 3*n l +1 health status, n l Represents the number of possible open circuit faults in each phase bridge arm of the converter. All operating conditions and health states of each type of converter are enumerated, and a total of n k Fault samples, n k =m*(3*n l +1);

[0008] n s The fault samples collected from different types of converters constitute the training data set and are recorded as

[0009] S k represents the fault sample set collected by the k-th type converter, in, is called a fault sample, is the i-th fault feature vector, for Corresponding specific health status, Encode the i-th task;

[0010] Among them, the fault feature vector The way to obtain it is:

[0011] Under certain operating conditions and health conditions, the sampling rate C n , continuously sample the three-phase input current signal of the k-th category converter in one cycle to form a sampling sequence, and then normalize the sampling sequence to obtain the feature vector

[0012] Corresponding specific health status Expressed as:

[0013]

[0014] and

[0015] and

[0016] and

[0017]

[0018] Task Coding Expressed as:

[0019] and

[0020] (2) Constructing a task encoding adaptive network;

[0021] The task encoding adaptation network includes a feature mapping module f map , Task migration module f tt and classifier f c , the task encoding adaptive network is and As input, output predicted fault code

[0022]

[0023] Among them, θ is all the parameters to be trained in the task encoding adaptive network;

[0024] Feature mapping module f map The p-channel weighted convolution module f wc,l ,l=1,2...p is connected in series and used to input fault feature vector Extract key fault features from

[0025]

[0026] Task migration modulef tt Contains p+2 task encoding blocks f tt,l ,l=1,2...p+2, used to encode from the task and O wc,l-1 Extract fault related information from different types of converters tt,l ;

[0027]

[0028] Among them, θ tt,l are the model parameters to be trained;

[0029] Classifier f c Using a typical multi-layer perceptron neural network, key fault characteristics and task encoding module f tt,p+1 Output O tt,p+1 A 1-dimensional feature vector is obtained by batch normalization BN and feature flattening Flatten

[0030]

[0031] Secondly, and task encoding module f tt,p+2 Output O tt,p+2The feature vector is obtained by batch normalization BN, fully connected layer FSC and Relu activation function

[0032]

[0033] Among them, W fsc and B fsc are the parameters of the FSC model to be trained, is the Relu activation function;

[0034] at last, The predicted fault code is obtained through the Sofetmax classification function:

[0035]

[0036] (3) Training task encoding adaptive network;

[0037] (3.1) Set the total number of training times n ep , initialize the current q, q=1,2,…,n ep ;

[0038] (3.2), from the training data set Randomly select fault samples Then and Input to the task encoding adaptive network to predict the fault code

[0039] (3.3) Calculate the loss function value after this round of training;

[0040]

[0041] Among them, θ (q-1) represents the training parameters obtained during the q-1th cycle training, L c represents the cross entropy loss function;

[0042] (3.4) Determine whether the current number of iterations has reached the maximum value n ep Or the loss function value converges, if so, the iteration stops and the trained task encoding adaptive network is output; otherwise, go to step (3.5);

[0043] (3.5) Update the network parameters θ using the gradient descent method (q) ;

[0044]

[0045] Among them, η pe is the learning rate;

[0046] (3.6) After the network parameters are updated, set the current number of iterations q = q + 1, and then return to step (3.2) for the next round of training;

[0047] (4) Online fault diagnosis;

[0048] At sampling rate C n , continuously sample the three-phase input current signal of a converter in one cycle to form a sampling sequence, and then normalize the sampling sequence to obtain the feature vector x;

[0049] Input the feature vector x into the trained task encoding adaptive network to predict the fault code Finally, according to the fault code Check the corresponding specific health status to locate the fault.

[0050] The object of the invention of the present invention is achieved like this:

[0051] The present invention is based on a method for diagnosing open-circuit faults of converters based on a task coding adaptive network. First, a fault sample set base consisting of fault samples, fault codes and task codes of various types of converters is constructed. Then, a task coding adaptive network is constructed. The task coding adaptive network is proposed to characterize the functional mapping relationship between task codes, fault characteristics and fault codes, so as to achieve fault location of various converter types without repeated model training, thereby reducing the development cycle and training cost of the fault diagnosis model. Secondly, in the task coding adaptive network, the optimization design of the network structure and the unified fault / task coding are used to promote the diagnosis of open-circuit faults of different types of converters, thereby improving the accuracy of fault diagnosis.

[0052] At the same time, the converter open circuit fault diagnosis method based on the task coding adaptive network of the present invention also has the following beneficial effects:

[0053] (1) The task coding adaptive network proposed in the present invention does not require repeated model training and can realize open circuit fault diagnosis of various converters, effectively reducing the development cycle and cost of the open circuit fault diagnosis model.

[0054] (2) The present invention differs from existing methods in that: (i) it adopts unified task and fault encoding for multi-task training; (ii) it shares weights between neural networks of different tasks to promote knowledge sharing; and (iii) it embeds task encoding into multiple feature layers as adaptive weight coefficients to achieve online inductive migration without the need for additional offline training.

[0055] (3) Weight sharing and task-based selective transfer help improve fault diagnosis performance and promote the inductive transfer of knowledge. Compared with existing methods, the proposed converter open circuit fault diagnosis method shows significant performance improvement in fault diagnosis accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a method for diagnosing open-circuit faults of a converter based on a task-coding adaptive network according to the present invention;

[0057] Figure 2 is the converter topology diagram;

[0058] Figure 3 This is the task encoding adaptive network structure diagram;

[0059] Figure 4 yes Figure 3 The network structure diagram of the channel weighted convolution block shown;

[0060] Figure 5 yes Figure 3 The network structure diagram of the task encoding block shown;

[0061] Figure 6 yes Figure 3 The network structure diagram of the classifier is shown. DETAILED DESCRIPTION

[0062] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0063] Example

[0064] Figure 1 The present invention is a flow chart of a method for diagnosing open-circuit faults of a converter based on a task-coded adaptive network.

[0065] In this embodiment, if Figure 1 As shown, the present invention provides a method for diagnosing open-circuit faults of a converter based on a task-coding adaptive network, comprising the following steps:

[0066] S1. Build a training dataset;

[0067] Each type of converter has m operating conditions (e.g., the operating state of the converter under specific DC bus voltage, input voltage, output voltage, power factor, and load conditions) and 3*n l +1 health status (including different open circuit fault types and normal status of the converter), n lRepresents the number of possible open circuit faults in each phase bridge arm of the converter. All operating conditions and health states of each type of converter are enumerated, and a total of n k Fault samples, n k =m*(3*n l +1);

[0068] In this embodiment, the collection of n s The training data set consists of fault samples of different types of converters (such as full-bridge converter, NPC converter, T-type three-level converter, etc.) under different operating conditions and different health states, which are recorded as S k represents the fault sample set collected by the k-th category converter, in, is called a fault sample, is the i-th fault feature vector, for The corresponding health status, Encode the i-th task;

[0069] Get fault feature vector

[0070] Under certain operating conditions and health conditions, the sampling rate C n , continuously sampling the three-phase input current signal of the k-th category converter in one cycle to form a sampling sequence;

[0071] Then the sampling sequence is normalized: the mean and variance of the sampling sequence are calculated, and the mean of each sampling point in the sampling sequence is subtracted and then divided by the variance to obtain the eigenvector

[0072] Corresponding specific health status Expressed as:

[0073]

[0074] and

[0075] and

[0076] and

[0077]

[0078] In this embodiment, for any n-level converter, in order to promote the summarization of knowledge of different types of converter faults, according to the missing level d of the output voltage when the converter fails, l =[d l,a ,d l,b ,d l,c ]Get fault code There are three situations:

[0079] (1) When the converter is operating normally:

[0080]

[0081] (2) When the converter has an open circuit fault and n is an even number:

[0082] if: but

[0083] if: but

[0084] if: but

[0085] (3) When the converter fails and n is an odd number:

[0086] if: but

[0087] if: but

[0088] if: but

[0089] Get the task code

[0090] According to n s Converter class, task code Use n s Binary one-hot encoding, each type of converter is a specific binary number and satisfies:

[0091] and

[0092] Below we take the full-bridge converter (a), NPC converter (b), and T-type three-level converter (c) as examples to describe the detailed and The topology of the above converter is as follows: Figure 2The number of possible open-circuit faults per phase of the converter in this example is 4, so n l = 4. There are three types of converters in this example, so n s =3. Full-bridge converter k = 1, so NPC converter k=2, so T-type three-level converter k = 3, so

[0093] For a full-bridge converter (two-level inverter), n = 2, the output voltage u xo ∈{u dc / 2,-u dc / 2},x=a,b,c,u dc is the DC bus output voltage. During normal operation, the fault code is:

[0094]

[0095] therefore,

[0096] When an open circuit fault occurs in phase A T1, u ao will not be able to generate u dc / 2 level, so The missing level of the output voltage during the fault is recorded as The fault code is:

[0097]

[0098] therefore,

[0099] When an open circuit fault occurs in phase A T2, u ao Will not generate -u dc / 2 level, so The missing level of the output voltage during the fault is recorded as The fault code is:

[0100] and therefore, Similarly, the task codes and fault codes of all converters can be obtained as shown in Table 1.

[0101] Table 1. Task codes and fault codes

[0102]

[0103] S2, building a task encoding adaptive network;

[0104] like Figure 3 As shown, the task encoding adaptation network includes a feature mapping module f map, Task migration module f tt and classifier f c , f map 、f tt and f c according to Figure 3 The task encoding adaptive network is connected together as shown in the figure, and its input is and Output predicted fault code

[0105]

[0106] Among them, θ is all the parameters to be trained in the task encoding adaptive network;

[0107] Feature mapping module f map For the input fault feature vector Extract key fault features from Its mathematical model is expressed as:

[0108]

[0109] Among them, θ map are the model parameters to be trained;

[0110] f map The p-channel weighted convolution module f wc,l ,l=1,2...p are connected in series, f wc,l O wc,l-1 and O tt,l As input, its mathematical model is expressed as:

[0111] O wc,l =f wc,l (O wc,l-1 ,O tt,l θ wc,l )

[0112] Among them, O wc,l-1 f wc,l-1 Output; O tt,l is the output of the lth task encoding module; θ wc,l The model parameters to be trained.

[0113] Therefore, f map The recursive formula is expressed as:

[0114]

[0115] In this embodiment, the channel weighted convolution block f wc,l It is used to perform multi-layer nonlinear mapping of input feature vectors to realize key fault feature extraction. Its network structure is as follows Figure 4As shown. Channel weighted convolution block f wc,l The input is O wc,l-1 ,O tt,l , get the eigenvector X1,X1∈R through the Crollard inner product H×1×G :

[0116]

[0117] X1 obtains the feature vector through batch normalization (BN) and MulCNN

[0118] X2 = BN(MulCNN(X1;θ CNN ))MulCNN is a H c The convolution operation of 1×h convolution kernel, θ CNN is the model parameter to be trained. X2 is subjected to global average pooling (GAP) to obtain the feature vector

[0119] X3=GAP(X2)X3 obtains adaptive weights through the fully connected layer (FC) and ReLU activation function

[0120] X4=Relu(FC(X3;θ FC ))

[0121] Among them, θ FC Is the model parameter to be trained. Perform dot multiplication on X4 and X2 to obtain weighted convolution features

[0122] X5=X4⊙X2

[0123] X5 performs feature selection through the Pick function, filters out features irrelevant to the fault, and finally obtains the output O through the pooling layer (Pool) wc,l :

[0124] O wc,l =Pool(f pk (X5;τ))

[0125] Among them, the Pick function is expressed as:

[0126]

[0127] Among them, s is an element in the input vector, and τ is the model parameter to be trained.

[0128] like Figure 3 As shown, the task migration module f tt Contains p+2 task encoding blocks f tt,l ,l=1,2...p+2, used to encode from the task and O wc,l-1 Extract fault related information from different types of converters tt,l ;

[0129]

[0130] Among them, θ tt,l are the model parameters to be trained;

[0131] In this embodiment, the task encoding module f tt,l The network structure is as follows Figure 5 As shown, Through the fully connected neural network f fc1,l Get the adaptive weight W ad ; Through the fully connected neural network f fc2,l Get adaptive bias B ad , its mathematical model is expressed as:

[0132]

[0133] Among them, θ fc1,l and θ fc2,l are the model parameters to be trained;

[0134] Regularized transfer block f nt,l W ad , B ad and O wc,l-1 is input and outputs task related information O tt,l , its mathematical model is expressed as:

[0135]

[0136] Where u represents O wc,l-1 The mean of O wc,l-1 The variance of Represents the ReLU activation function.

[0137] In this embodiment, if Figure 6 As shown, the classifier f c Using a typical multi-layer perceptron neural network, key fault characteristics and task encoding module f tt,p+1 Output O tt,p+1 A 1-dimensional feature vector is obtained by batch normalization BN and feature flattening Flatten

[0138]

[0139] Secondly, and task encoding module f tt,p+2Output O tt,p+2 The feature vector is obtained by batch normalization BN, fully connected layer FSC and Relu activation function

[0140]

[0141] Among them, W fsc and B fsc are the parameters of the FSC model to be trained, is the Relu activation function;

[0142] at last, The predicted fault code is obtained through the Sofetmax classification function:

[0143]

[0144] S3, training task encoding adaptive network;

[0145] S3.1. Set the total number of training times n ep , initialize the current q, q=1,2,…,n ep ;

[0146] S3.2. From the training dataset Randomly select fault samples Then and Input to the task encoding adaptive network to predict the fault code

[0147] S3.3. Calculate the loss function value after this round of training;

[0148]

[0149] Among them, θ (q-1) represents the training parameters obtained during the q-1th cycle training, L c represents the cross entropy loss function;

[0150] S3.4. Determine whether the current number of iterations has reached the maximum value n ep Or the loss function value converges, if so, the iteration stops and the trained task encoding adaptive network is output; otherwise, go to step S3.5;

[0151] S3.5. Update network parameters θ using gradient descent (q) ;

[0152]

[0153] Among them, η pe is the learning rate;

[0154] S3.6. After the network parameters are updated, set the current number of iterations q = q + 1, and then return to step S3.2 for the next round of training;

[0155] S4, online fault diagnosis;

[0156] At sampling rate C n , continuously sample the three-phase input current signal of a converter in one cycle to form a sampling sequence, and then normalize the sampling sequence to obtain the feature vector x;

[0157] Input the feature vector x into the trained task encoding adaptive network to predict the fault code Finally, according to the fault code Check the corresponding specific health status to locate the fault.

[0158] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for diagnosing open-circuit faults in a converter based on a task-encoded adaptive network, characterized in that: The following steps are involved: (1) Construct a training dataset; Note that each type of converter has m operating conditions and 3*n l +1 health status, n l Represents the number of possible open circuit faults in each phase bridge arm of the converter. All operating conditions and health states of each type of converter are enumerated, and a total of n k Fault samples, n k =m*(3*n l +1); n s The fault samples collected from different types of converters constitute the training data set and are recorded as S k represents the fault sample set collected by the k-th type converter, in, is called a fault sample, is the i-th fault feature vector, for Corresponding specific health status, Encode the i-th task; Among them, the fault feature vector The way to obtain it is: Under certain operating conditions and health conditions, the sampling rate C n , continuously sample the three-phase input current signal of the k-th category converter in one cycle to form a sampling sequence, and then normalize the sampling sequence to obtain the feature vector Corresponding specific health status Expressed as: Task Coding Expressed as: and (2) Constructing a task encoding adaptive network; The task encoding adaptation network includes a feature mapping module f map , Task migration module f tt and classifier f c , the task encoding adaptive network is and As input, output predicted fault code Among them, θ is all the parameters to be trained in the task encoding adaptive network; Feature mapping module f map The weighted convolution module f consists of p channels wc,l ,l=1,2,…,p is composed in series, which is used to input fault feature vector Extract key fault features from Among them, θ map is the model parameter to be trained, O wc,l-1 Weighted convolution module f for the l-1th channel wc,l-1 Output; Task migration modulef tt Contains p+2 task encoding modules f tt,l ,l=1,2,…,p+2, used to encode from the task and O wc,l-1 Extract fault related information from different types of converters tt,l ; Among them, θ tt,l are the model parameters to be trained; Classifier f c Using a typical multi-layer perceptron neural network, key fault characteristics and task encoding module f tt,p+1 Output O tt,p+1 A 1-dimensional feature vector is obtained by batch normalization BN and feature flattening Flatten Secondly, and task encoding module f tt,p+2 Output O tt,p+2 The feature vector is obtained by batch normalization BN, fully connected layer FSC and Relu activation function Among them, W fsc and B fsc are the parameters of the FSC model to be trained, is the Relu activation function; at last, The predicted fault code is obtained through the Softmax classification function: (3) Training task encoding adaptive network; (3.1) Set the total number of training times n ep , initialize the current q, q=1,2,…,n ep ; (3.2), from the training data set Randomly select fault samples Then and Input to the task encoding adaptive network to predict the fault code (3.3) Calculate the loss function value after this round of training; Among them, θ (q-1) represents the training parameters obtained during the q-1th cycle training, L c represents the cross entropy loss function; (3.4) Determine whether the current number of iterations has reached the maximum value n ep Or the loss function value converges, if so, the iteration stops and the trained task encoding adaptive network is output; otherwise, go to step (3.5); (3.5) Update the network parameters θ using the gradient descent method (q) ; Among them, η pe is the learning rate; (3.6) After the network parameters are updated, set the current number of iterations q = q + 1, and then return to step (3.2) for the next round of training; (4) Online fault diagnosis; At sampling rate C n , continuously sample the three-phase input current signal of a converter in one cycle to form a sampling sequence, and then normalize the sampling sequence to obtain the feature vector x; Input the feature vector x into the trained task encoding adaptive network to predict the fault code Finally, according to the fault code Check the corresponding specific health status to locate the fault.

2. The method for diagnosing open-circuit faults of a converter based on a task-coding adaptive network according to claim 1, characterized in that: The channel weighted convolution module f wc,l The input is O wc,l-1 ,O tt,l , get the eigenvector X1,X1∈R through the Crollard inner product H×1×G : X1 obtains the feature vector through batch normalization BN and MulCNN MulCNN is a CNN with H c The convolution operation of 1×h convolution kernel, θ CNN are the model parameters to be trained; X2 obtains the feature vector through global average pooling GAP X3 obtains adaptive weights through the fully connected layer FC and ReLU activation function Among them, θ FC are the model parameters to be trained; Perform dot multiplication on X4 and X2 to obtain weighted convolution features X5 performs feature selection through the Pick function, filters out features irrelevant to the fault, and finally obtains the output O through the pooling layer. wc,l : Among them, the Pick function is expressed as: Among them, s is an element in the input vector, and τ is the model parameter to be trained.

3. The method for diagnosing open circuit faults of a converter based on a task-coding adaptive network according to claim 1, characterized in that: The task encoding module first Through the fully connected neural network f fc1,l Get the adaptive weight W ad ; then Through the fully connected neural network f fc2,l Get adaptive bias B ad , its mathematical model is expressed as: Among them, θ fc1,l and θ fc2,l are the model parameters to be trained; Finally, the regularized migration block f nt,l W ad , B ad and O wc,l-1 is input and outputs task related information O tt,l , its mathematical model is expressed as: Where u represents O wc,l-1 The mean of O wc,l-1 The variance of Represents the ReLU activation function.

Images