A training method for electronic circuit fault diagnosis model based on training network

Abstract

The present invention provides a training method for an electronic circuit fault diagnosis model based on a training network. This method utilizes a training network to identify disconnection and forward faults in rectifying components (such as thyristors and diodes) in electronic circuits and classify the fault types. By establishing a correspondence between fault symptom information and faulty components, the training network enables online fault diagnosis of circuit output voltage and current waveforms. The method includes determining a correspondence between fault information data and training network input samples. By sampling and normalizing the real-time voltage waveforms, while simultaneously training the network to learn and memorize, the training network can accurately diagnose the faulty component and fault type.

Description

Technical Field

[0001] The present invention relates to an electronic circuit fault diagnosis technology, and in particular to a training method for an electronic circuit fault diagnosis model based on a training network. Background Art

[0002] Electronic circuit failures often result in waveform distortion and value changes in the circuit's output voltage and current. Existing fault training methods often rely on expert experience or traditional signal processing techniques, which have limited efficiency and accuracy. Therefore, a model training method that can automatically diagnose electronic circuit failures is needed to improve diagnostic accuracy and efficiency. Summary of the Invention

[0003] The present invention provides a training method for an electronic circuit fault diagnosis model based on a training network, comprising:

[0004] 1) Determine the corresponding relationship between the fault information data and the training network input samples; determine the corresponding relationship between the fault source code and the training network output value, thereby determining two groups of output vectors.

[0005] 2) Form a fault information sample pair vector group (Ai, Oi) (i = 1, 2, 3, ..., p), and convert the p fault information sample pair vector groups (Ai, Oi) (i = 1, 2, 3, ..., p) into a bipolar vector group pair (Xi, Yi) (i = 1, 2, 3, ..., p).

[0006] 3) Each sample pair is studied and trained in turn to form a memory matrix Cj (j = 1, 2, ...).

[0007] 4) Store the fault mode sample pairs (Xi, Yi) into another memory matrix Ck and test the memory capacity of Ck.

[0008] 5) Repeat steps 2)-4) until all fault information sample pairs are stored in a series of memory matrices after learning and training.

[0009] Preferably, the voltage sampling values ​​are normalized, ie.

[0010]

[0011] Where, U2 is the effective value of the input line voltage; Ud * It is the normalized value of the sampling voltage value Ud.

[0012] Preferably, if p pairs of binary matrices of fault sample information are given: (A1, O1), (A2, O2), …, (Ap, Op), these p pairs of binary matrices of fault sample information are first converted into bipolar matrix pairs (Xi, Yi) (i = 1, 2, …, p) according to the following formula. That is, the zero elements in the binary matrix pairs are replaced with -1, that is.

[0013] Xi=2Ai-I.

[0014] Yi=20i-I.

[0015] In the above formula, I is the identity matrix.

[0016] Preferably, after learning each input fault information sample pair, it is verified whether the training network has the correct memory ability for (Ai, Oi). That is, it is determined whether XiCj is equal to Oi after thresholding, and YiCj is also determined. T After thresholding, is it equal to Ai? If they are equal, it means that the training network has the ability to remember (Ai, Oi) after learning; if they are not equal, it means that the training network has saturated its memory of (Ai, Oi) after learning, and Xi should be removed from the Cj matrix. T Yj.

[0017] Preferably, the fault information sample bipolar matrix pairs (X1, Y1), (X2, Y2), …, (Xp, Yp) are stored at the position where the local energy of the training network is minimum or close to minimum, that is, the p pattern fault information sample bipolar matrix pairs are superimposed to form an m×n dimensional memory matrix C, that is.

[0018] C=X1 T Y1+X2 T Y2+...Xp T Yp.

[0019] Preferably, the input layer of the training network has m neurons and the output layer has n neurons. Training the network is a transformation from one vector space to another vector space, that is, a transformation from vector space Rm to vector space Rn.

[0020] Preferably, all fault information of the training network can be stored in an m×n dimensional weight matrix C, which is also called a memory matrix.

[0021] Preferably, if the training network converges for every pair of fault information input samples, the memory matrix C is stable.

[0022] This method uses a trained network to identify disconnection and forward faults in rectifying components (such as thyristors and diodes) in electronic circuits and classify the fault type. By establishing a correspondence between fault symptoms and faulty components, the trained network enables online fault diagnosis of the circuit's output voltage and current waveforms. This method involves determining the correspondence between fault information data and training network input samples. By sampling and normalizing the real-time voltage waveforms, while simultaneously training the network to learn and memorize, the trained network can accurately diagnose the faulty component and fault type. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention provides an execution flow of a training method for an electronic circuit fault diagnosis model based on a training network. DETAILED DESCRIPTION

[0024] Since electronic circuit failures are primarily caused by disconnection and forward conduction failures in rectifying components (thyristors and diodes, including JX and Ishizuka series diodes), and fault characteristics are primarily manifested in waveform distortion and value changes in the circuit's output voltage and current, using a training network to diagnose electronic circuit failures requires first establishing a correspondence between circuit fault symptoms and faulty components, and then training the diagnostic model. The training network can be a BP neural network.

[0025] The specific method and operation steps for training the electronic circuit fault diagnosis model based on the training network are as follows.

[0026] 1) Determine the correspondence between the fault information data (such as the sampling values ​​of the circuit's output voltage and current waveforms) and the training network input samples; determine the correspondence between the fault source code and the training network output value, thereby determining two groups of output vectors.

[0027] Collection of fault samples.

[0028] The training network learns from a sample set consisting of real-time samples of the output voltage waveform of a three-phase bridge-controlled rectifier circuit. When the three-phase bridge-controlled rectifier circuit is operating, the thyristor trigger delay angle α can vary between 0° and 120° (the trigger delay angle α of the rectifier circuit can be selected as 0°, 30°, 60°, 90°, and 120°, among others). To facilitate the algorithm description, it is generally assumed that when the three-phase bridge-controlled rectifier circuit operates at a specific trigger delay angle, if no more than two thyristors in the three-phase bridge-controlled rectifier circuit simultaneously fail, then at each specific trigger delay angle α, there are 22 possible types of rectifier circuit failures, resulting in a total of 110 samples. Each sample set consists of 40 sampling points selected from one cycle of the three-phase bridge-controlled rectifier circuit's output voltage waveform.

[0029] The number of nodes in the input layer of the training network is equal to the number of Ud voltage sampling values ​​in one cycle. All voltage sampling values ​​in one cycle constitute a learning sample, and each learning sample corresponds to an output code.

[0030] In order to facilitate the learning of the training network, the voltage sampling values ​​need to be normalized, that is.

[0031]

[0032] Where, U2 is the effective value of the input line voltage; Ud * It is the normalized value of the sampling voltage value Ud.

[0033] In each type of fault, fault information can be obtained by periodically sampling the voltage waveform output by the circuit in real time and using this set of sampled values ​​as the input sample values ​​of the training network. The number of input sample values ​​is the same as the number of input nodes of the training network. Under the same trigger angle, the number of output nodes of the training network should be equal to the number of thyristors in the circuit. If the tth thyristor in the electronic circuit fails, the expected output vector of the training network is Di = (d1, d2, ..., dn) T The tth element dt in the vector is 1, and the rest are 0. This vector serves as the expected value of the training network output. Similarly, after all potentially faulty components have been encoded, a series of unit vectors, Ai = (A1, A2, ..., An), are formed. In this way, the input information samples of all faulty components correspond one-to-one to the output vectors, and the ideal output of the training network is undoubtedly a unit matrix. Repeating the above sampling and encoding process, we obtain the second set of output vectors, Oi = (O1, O2, ..., On).

[0034] 2) Form a fault information sample pair vector group (Ai, Oi) (i = 1, 2, 3, ..., p), and convert the p fault information sample pair vector groups (Ai, Oi) (i = 1, 2, 3, ..., p) into a bipolar vector group pair (Xi, Yi) (i = 1, 2, 3, ..., p).

[0035] Given p pairs of binary matrices of fault sample information: (A1, O1), (A2, O2), …, (Ap, Op), first convert these p pairs of binary matrices of fault sample information into bipolar matrix pairs (Xi, Yi) (i = 1, 2, …, p) according to the following formula. That is, replace the zero elements in the binary matrix pairs with -1, that is.

[0036] Xi=2Ai-I.

[0037] Yi=20i-I.

[0038] In the above formula, I is the identity matrix.

[0039] 3) Each sample pair is then learned and trained in turn to form a memory matrix Cj (j = 1, 2, ...). That is, after learning each fault information sample pair input, it is then verified whether the trained network has the correct memory ability for (Ai, Oi). In other words, it is determined whether XiCj is equal to Oi after thresholding, and also whether YiCj is determined. T After thresholding, is it equal to Ai? If they are equal, it means that the training network has the ability to remember (Ai, Oi) after learning; if they are not equal, it means that the training network has saturated its memory of (Ai, Oi) after learning, and Xi should be removed from the Cj matrix. T Yj.

[0040] Assume that the training network has m neurons in its input layer and n neurons in its output layer. Training a network is a transformation from one vector space to another, that is, from vector space Rm to vector space Rn. If the training network converges for every pair of fault information input samples, then the memory matrix C is stable. All fault information in the training network can be stored in an m×n-dimensional weight matrix C, which is also known as the memory matrix. The training network's memory function is implemented through this weight matrix.

[0041] The training process of a network training algorithm can be divided into two phases: forward propagation and backpropagation. The forward propagation phase begins at the input layer, passing sample data forward according to the learning rules. The output of each layer is calculated sequentially, ultimately yielding the network's final output layer. Backpropagation proceeds in the opposite direction of the forward process. Based on the errors obtained in the forward calculations, weights and thresholds are adjusted backwards, starting from the output layer of the training network. These two processes are repeated until convergence is achieved.

[0042] The fault information sample bipolar matrix pairs (X1, Y1), (X2, Y2), …, (Xp, Yp) are stored at the position where the local energy of the training network is minimum or close to minimum, that is, the p pattern fault information sample bipolar matrix pairs are superimposed to form an m×n dimensional memory matrix C, that is.

[0043] C=X1 T Y1+X2 T Y2+...Xp T Yp.

[0044] If Ai is represented by a bipolar vector Xi, and the bipolar vector Xi of the fault information sample signal Ai is input into the training network at the input layer of the training network, we can obtain .

[0045]

[0046] In the above formula, since Xi is an m-dimensional vector, then XiXi T =m; d is a number greater than zero, that is, d>0. It can be seen that when the sum of the input XiC is thresholded, Yi tends to Oi, that is, the output sample matrix Oi corresponding to Ai is obtained, 1<=i<=p, 1<=j<=p.

[0047] 4) The fault mode sample pair (Xi, Yi) is then stored in another memory matrix Ck. Repeat the above method to check whether the Ck matrix has the correct memory capacity for the sample pair.

[0048] 5) Repeat steps 2)-4) until all fault information sample pairs are stored in a series of memory matrices after learning and training.

[0049] The present invention provides a training method for an electronic circuit fault diagnosis model based on a training network. This method utilizes a training network to identify disconnection and forward faults in rectifying components (such as thyristors and diodes) in electronic circuits and classify the fault types. By establishing a correspondence between fault symptom information and faulty components, the training network enables online fault diagnosis of circuit output voltage and current waveforms. The method includes determining a correspondence between fault information data and training network input samples. By sampling and normalizing the real-time voltage waveforms, while simultaneously training the network to learn and memorize, the training network can accurately diagnose the faulty component and fault type.

[0050] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the scope of the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the spirit and scope of protection of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present invention.

Claims

1. A training method for an electronic circuit fault diagnosis model based on a training network, characterized in that: include: 1) Determine the correspondence between fault information data and training network input samples; Determine the corresponding relationship between the fault source code and the training network output value, thereby determining two groups of output vectors; The sample set used for training the network is composed of real-time sampling of the output voltage waveform of a three-phase bridge-controlled rectifier circuit. When the three-phase bridge-controlled rectifier circuit is operating, the trigger delay angle α of the thyristor can vary from 0° to 120°, and the trigger delay angle α of the rectifier circuit can be selected as 0°, 30°, 60°, 90°, and 120° respectively. When the three-phase bridge-controlled rectifier circuit operates at a specific trigger delay angle, if at most two thyristors in the three-phase bridge-controlled rectifier circuit fail simultaneously, then at each specific trigger delay angle α, there are 22 types of rectifier circuit failures, resulting in a total of 110 samples. Each sample set is composed of 40 sampling points selected from the output voltage waveform of the three-phase bridge-controlled rectifier circuit in one cycle. The number of nodes in the input layer of the training network is equal to the number of Ud voltage sampling values ​​in one cycle. All voltage sampling values ​​in one cycle constitute a learning sample, and each learning sample corresponds to an output code. Normalize the voltage sampling value, that is, divide the effective value of the input line voltage by the sampling voltage value to obtain the normalized value; In each type of fault situation, the fault information is obtained by periodically sampling the voltage waveform output by the circuit in real time, and this set of sampled values ​​is used as the input sample value of the training network. The number of input sample values ​​is the same as the number of input nodes of the training network. Under the same trigger angle, the number of output nodes of the training network should be equal to the number of thyristors in the circuit. If the tth thyristor in the electronic circuit fails, the expected output vector Di of the training network is (d1, d2, ..., dn) T The t-th element dt in the vector is 1, and the rest are 0. This vector serves as the expected value of the training network output. After all possible faulty components are encoded, a series of unit vectors are formed to form the first output vector group Ai=(A1, A2, ..., An). In this way, the input information samples of all faulty components correspond one-to-one to the output vector group, and the ideal output of the training network is a unit matrix. Repeat the above sampling and encoding process to obtain the second output vector group Oi=(O1, O2, ..., On). 2) Form a fault information sample pair vector group (Ai, Oi) (i=1, 2, 3, ..., p), and convert the p fault information sample pair vector groups (Ai, Oi) (i=1, 2, 3, ..., p) into a bipolar vector group pair (Xi, Yi) (i=1, 2, 3, ..., p); Given p pairs of binary matrices of fault sample information: (A1, O1), (A2, O2), …, (Ap, Op), first convert these p pairs of binary matrices of fault information samples into bipolar matrix pairs (Xi, Yi) (i=1, 2, …, p) according to the following formula; that is, replace the zero elements in the binary matrix pairs with -1, that is: Xi=2Ai-I; Yi=20i-I; In the above formula, I is the identity matrix; The zero elements in the fault information sample binary matrix pair are replaced by negative one to obtain a bipolar matrix pair; for each fault information sample binary matrix pair, each element thereof is multiplied by two and then the corresponding element of the identity matrix is ​​subtracted to obtain two matrices in the bipolar matrix pair; 3) Each sample pair is studied and trained in turn to form a memory matrix Cj (j=1,2,…); That is, after learning each input fault information sample pair, it is verified whether the training network has the correct memory ability for (Ai, Oi); that is, whether XiCj is equal to Oi after thresholding, and whether YiCj is also judged. T Is it equal to Ai after thresholding? If they are equal, the training network has the ability to remember (Ai, Oi) after learning. If they are not equal, the training network has saturated its memory of (Ai, Oi) after learning, and Xi should be eliminated from the Cj matrix. T Yj; The fault information sample bipolar matrix pairs (X1, Y1), (X2, Y2), ..., (Xp, Yp) are stored at the location where the local energy of the training network is minimized. That is, the p pattern fault information sample bipolar matrix pairs are superimposed to form an m×n dimensional memory matrix C, that is: C=X1 T Y1+X2 T Y2+...Xp T Yp; Multiple fault information sample bipolar matrix pairs are superimposed to form a memory matrix with the dimension of the number of input layer neurons multiplied by the number of output layer neurons. For each fault information sample bipolar matrix pair, the input matrix is ​​transposed and multiplied by the output matrix, and then all these products are added together to obtain the memory matrix. Ai is represented by a bipolar vector Xi. The bipolar vector Xi of the fault information sample signal Ai is input into the training network at the input layer of the training network. For each bipolar vector Xi of the input fault information sample signal Ai, the bipolar vector Xi is multiplied by the memory matrix C, and then the sum of the products is thresholded. The output result Yi will be similar to the corresponding fault information sample output matrix Oi. 4) Store the fault mode sample pairs (Xi, Yi) into another memory matrix Ck to test the memory capacity of Ck; 5) Repeat steps 2)-4) until all fault information sample pairs are stored in a series of memory matrices after learning and training.

Images