Electronic DC load overcurrent prediction method based on multi-dimensional sensing feature extraction

Abstract

The present invention discloses an electronic DC load overcurrent prediction method based on multi-dimensional perception feature extraction, the method comprising: using a singular value decomposition method to denoise the sampled signal of the electronic DC load to obtain a one-dimensional denoised signal; converting the one-dimensional denoised signal into a two-dimensional image using a Markov transition field; extracting channel perception features and space perception features from the two-dimensional image based on a dual-scale flexible domain perception module; using an adaptive feature fusion module to fuse the channel perception features and space perception features to obtain a perception fault feature; using a domain invariant prototype dual classifier to perform a predictive circuit overcurrent alarm based on the perception fault feature. The present invention has significant advantages in improving the predictive overcurrent protection of electronic DC loads through advanced signal processing and feature extraction technology.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence deep learning fault diagnosis, and specifically relates to an electronic DC load overcurrent prediction method based on multi-dimensional perception feature extraction. Background Art

[0002] With the rapid development of science and technology, electronic equipment is being used more and more widely. As an important component of electronic equipment, the performance and stability of electronic DC loads are crucial to the operation of electronic equipment. Electronic DC loads are mainly used to simulate actual load conditions, test and debug power supply equipment, and ensure the reliability and stability of power supply equipment. Overcurrent failures of electronic DC loads often lead to component damage and even safety accidents. Therefore, predicting overcurrent of electronic DC loads is of great significance to ensure the stable operation and safety of electronic equipment. Summary of the invention

[0003] The purpose of the present invention is to provide an efficient electronic DC load overcurrent prediction method, which helps to reduce the downtime caused by electronic equipment failure, reduce production costs and improve economic benefits by predicting electronic DC load overcurrent.

[0004] In order to solve the above technical problems, the present invention provides an electronic DC load overcurrent prediction method based on multi-dimensional perception feature extraction, comprising:

[0005] The sampling signal of the electronic DC load is denoised by using the singular value decomposition method to obtain a one-dimensional denoised signal;

[0006] The one-dimensional denoised signal is converted into a two-dimensional image using the Markov transition field;

[0007] Based on the dual-scale flexible domain perception module, channel perception features and space perception features are extracted from the two-dimensional image respectively;

[0008] Adopting adaptive feature fusion module to fuse channel perception features and spatial perception features to obtain perception fault features;

[0009] A domain-invariant prototype dual classifier is used to perform predictive circuit overcurrent alarm based on perceived fault features.

[0010] Optionally, the method of denoising the sampled signal of the electronic DC load by using a singular value decomposition method to obtain a one-dimensional denoised signal includes:

[0011] The sampled signal is constructed into a Hankel matrix, and the Hankel matrix is ​​subjected to singular value decomposition to obtain singular values ​​and corresponding singular vectors. The first k largest singular values ​​are selected to reconstruct the Hankel matrix, and the diagonal elements are extracted from the reconstructed Hankel matrix to obtain a one-dimensional denoised signal.

[0012] Optionally, converting the one-dimensional denoised signal into a two-dimensional image using a Markov transition field comprises:

[0013] The one-dimensional denoised signal is divided into G bins, the transition frequency between each bin is calculated, the state transfer matrix is ​​established through the transfer frequency, the Markov transition field is constructed through the state transfer matrix, and a two-dimensional image is obtained.

[0014] Optionally, the dual-scale flexible domain perception module is based on respectively extracting channel perception features and space perception features from the two-dimensional image, including:

[0015] The two-dimensional image is used as the input of the dual-scale flexible domain perception module, which consists of a channel domain perception module and a spatial domain perception module. The channel domain perception module embeds the global spatial information of the two-dimensional image into the channel through global average pooling and extracts channel perception features. The spatial domain perception module obtains the spatial statistical information of the two-dimensional image through group norm and extracts spatial perception features.

[0016] Optionally, the adopting of an adaptive feature fusion module to fuse the channel perception feature and the space perception feature to obtain the perception fault feature includes:

[0017] The channel-aware features and spatial-aware features are used as the input of the adaptive feature fusion module respectively. Through upsampling and convolution, the channel-aware features and spatial-aware features are adjusted to feature maps with the same size and the same number of channels. They are weighted and summed up through their respective weight coefficients to obtain the fused perceived fault features.

[0018] Optionally, the adopting of a domain-invariant prototype dual classifier to perform a predictive circuit overcurrent alarm based on perceived fault features includes:

[0019] The domain-invariant prototype dual classifier is used to align the distribution between the source domain and the target domain, the similarity between the target domain sample and the domain-invariant prototype is calculated, and the perceived fault features are input into the dual classifier to perform predictive overcurrent alarm on the electronic DC load circuit.

[0020] The present invention proposes an electronic DC load overcurrent prediction method based on multi-dimensional perception feature extraction, and adopts a singular value decomposition method to denoise the sampled signal, which is a key technology for the subsequent extraction of fault features; converting the Markov transition field into a two-dimensional image provides an intuitive visual basis for subsequent image processing and feature extraction, and also facilitates the application of richer image processing algorithms; introduces a dual-scale flexible domain perception module, and deeply understands the inherent structure and change pattern of the sampled signal through channel perception feature extraction and space perception feature extraction; adopts an adaptive feature fusion method to further fuse channel perception features and space perception features to form a more comprehensive and accurate fault feature representation; adopts a domain-invariant prototype dual classifier to perform predictive circuit overcurrent alarm based on perceived fault features, and realizes early warning of electronic DC load overcurrent faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 A schematic diagram of a flow chart of an electronic DC load overcurrent prediction method provided by an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of a dual-scale flexible domain perception module framework of an electronic DC load overcurrent prediction method provided by an embodiment of the present invention;

[0024] Figure 3 A schematic diagram of a domain-invariant prototype dual classifier flow chart of an electronic DC load overcurrent prediction method provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] like Figure 1 As shown, Figure 1 A schematic flow chart of an electronic DC load overcurrent prediction method provided in an embodiment of the present invention, wherein the electronic DC load overcurrent prediction method specifically includes five contents.

[0027] S11: De-noising the sampling signal of the electronic DC load using a singular value decomposition method to obtain a one-dimensional de-noised signal.

[0028] It should be noted that the present invention adopts the singular value decomposition method for denoising because the singular value decomposition method can effectively remove noise by decomposing the signal into singular values ​​and corresponding singular vectors. It is an efficient signal denoising technology and has the advantages of being independent of the specific model of the signal and being applicable to various types of signals. It is particularly suitable for denoising the sampling signals of electronic DC loads. The method specifically includes three steps.

[0029] S111: The voltage signal of the electronic DC load is constructed into a Hankel matrix. The Hankel matrix is ​​constructed as shown in formula (1):

[0030]

[0031] In the formula, m=N-n+1, ​​x i is the voltage data value, and H is the Hankel matrix.

[0032] S112: Perform singular value decomposition on the Hankel matrix to obtain a left singular vector matrix, a right singular vector matrix and a diagonal matrix. The elements on the diagonal of the diagonal matrix are singular values. The singular value decomposition function of the Hankel matrix is ​​shown in formula (2):

[0033]

[0034] Where U is the left singular vector matrix, and its column vector is the left singular vector of H; Q T is a right singular vector matrix whose row vectors are the right singular vectors of H; Σ is a diagonal matrix.

[0035] S113: Select the first k largest singular values ​​and corresponding singular vectors to reconstruct the Hankel matrix, extract the diagonal elements from the reconstructed Hankel matrix, and obtain a one-dimensional denoised signal.

[0036] Based on the above discussion, in an optional embodiment of the present invention, the denoising using the singular value decomposition method specifically includes:

[0037] In an optional embodiment of S11, the signal is constructed into a Hankel matrix, and m in the Hankel matrix is ​​set to 30; the first k largest singular values ​​are selected to reconstruct the Hankel matrix, and k is set to 5000.

[0038] S12: Convert the one-dimensional denoised signal into a two-dimensional image using the Markov transition field.

[0039] It should be noted that the one-dimensional denoised signal only contains information of a single dimension, and the two-dimensional image contains richer feature information. The two-dimensional image is the main data type of the deep learning model. The Markov transition field has the advantages of improving the efficiency of data processing and increasing the application scenarios.

[0040] The Markov transition field is used to transform the one-dimensional denoised signal into a two-dimensional image, which specifically includes three steps.

[0041] S121: The one-dimensional denoising signal X={x 1 , x 2 , …, x n} is divided into G quantile boxes, by the quantile G i (i∈{1,2,…,n}) quantizes the voltage data value x in the one-dimensional denoised signal i .

[0042] S122: Calculate the transfer frequency between each bin and construct the state transfer matrix W, which is constructed as shown in formula (3):

[0043] In the formula, w ij Represents the quantile G i Located at quantile G j The frequency after.

[0044] S123: Arrange each probability in time order to obtain the Markov transition field M, and map it according to the state of the one-dimensional denoised signal to obtain a two-dimensional image, as shown in formula (4):

[0045] In the formula, m ij Represents the bin G i Transfer to bin G j The probability of , p is the probability of the corresponding voltage data value.

[0046] Based on the above discussion, in an optional embodiment of the present invention, the one-dimensional denoised signal is converted into a two-dimensional image using a Markov transition field, specifically including:

[0047] In an optional embodiment of S12, the one-dimensional denoised signal is divided into G bins, G is set to 16, the image size is selected as 256*256, the translation step size is first calculated, the initial window capture position is selected, the window is moved according to the translation step size and the voltage signal sample is captured.

[0048] S13: Based on the dual-scale flexible domain perception module, channel-aware features and space-aware features are extracted from the two-dimensional image respectively.

[0049] It should be noted that the dual-scale flexible domain perception module is composed of a channel domain perception module and a space domain perception module. The dual-scale flexible domain perception module is introduced in the present invention because the module can learn features at different levels, improve learning efficiency, and enhance feature extraction capabilities. Figure 2 As shown, channel perception features and space perception features are extracted based on the dual-scale flexible domain perception module, which specifically includes three steps.

[0050] S131: Divide the two-dimensional image into channel-aware samples and space-aware samples, which serve as inputs of the channel-aware module and the space-aware module respectively.

[0051] S132: The channel domain perception module embeds the global spatial information of the two-dimensional image into the channel through global average pooling and extracts the channel perception feature Q k1 , and its process functions are shown in equations (5) to (7):

[0052]

[0053]

[0054] In the formula, It is expressed as the average value of all points in the cth channel in the nth sample, where L is the sample height and E is the sample width. X k1 is a channel-aware sample, (i, j) represents the coordinates of the point, q 1 is the channel perception vector, M 1 is a shared dimensionality-increasing layer, δ is a ReLU operation, g is a channel descriptor, and ☉ is a channel-aware multiplication.

[0055] S133: The spatial domain perception module obtains the spatial statistical information of the two-dimensional image through group norms and extracts the spatial perception feature Q k2 , and its process function is shown in formula (8)-formula (9):

[0056]

[0057] In the formula, q 2 is the spatial perception vector, M 2 is a shared dimensionality-increasing layer, δ is a ReLU operation, g is a channel descriptor, ☉ is a channel-aware multiplication, X k2 are channel-aware samples.

[0058] Based on the above discussion, in an optional embodiment of the present invention, for the above dual-scale flexible domain perception module, channel perception features and space perception features are extracted from the two-dimensional image respectively, specifically including:

[0059] In an optional embodiment of S13, the scaling ratio of the dual-scale flexible domain perception module is set to 16, the confidence is set to 0.8, and the number of iteration steps is set to 20.

[0060] S14: Adopt an adaptive feature fusion module to fuse the channel perception feature and the space perception feature to obtain the perception fault feature.

[0061] It should be noted that the present invention introduces the adaptive feature fusion module ASFF because the traditional fixed fusion strategy is unable to adapt to different input situations. ASFF can be dynamically adjusted to overcome this limitation. In addition, ASFF can adaptively fuse dual-scale features to more comprehensively express the feature information of the image. The feature fusion using ASFF specifically includes three steps.

[0062] S141: The channel-aware features and the space-aware features are respectively used as inputs of the adaptive feature fusion module, and the channel-aware features and the space-aware features are adjusted into feature maps with the same size and the same number of channels through upsampling and convolution.

[0063] S142: The spatial importance weight of the feature map is adaptively learned through gradient back propagation. Its process function is shown in equations (10) and (11):

[0064]

[0065]

[0066] In the formula, t is α or β, a is the number of layers, i and j are the coordinates on the feature map, and λ is the weight coefficient.

[0067] S143: weighted sum is performed through the respective weight coefficients to obtain the fused perceived fault feature, and its function is shown in formula (12):

[0068]

[0069] In the formula, is the (i, j)th vector along the channel of the feature map of the a-th layer, Represents the feature vector at (i, j) on the feature map adjusted from layer 1 to layer a.

[0070] Based on the above discussion, in an optional embodiment of the present invention, the adaptive feature fusion module is used to fuse the channel perception feature and the space perception feature to obtain the perception fault feature, which specifically includes:

[0071] In an optional embodiment of S14, for upsampling operations, a convolutional layer is used to compress the number of feature channels, and then the resolution is improved through nearest neighbor interpolation. For downsampling, a 3×3 convolutional layer is used with a stride of 2, and the number of channels and size are adjusted at the same time.

[0072] S15: Dual classifiers with load domain invariant prototypes for predictive circuit overcurrent alarm based on perceived fault features.

[0073] It should be noted that the present invention adopts the domain-invariant prototype dual classifier because the domain-invariant prototype dual classifier includes a domain-invariant prototype and a dual classifier structure. By learning domain-invariant features, it can better adapt to changes between different domains, reduce the feature differences between the source domain and the target domain, and reduce overfitting.

[0074] A dual classifier with load domain invariant prototype is used to perform predictive circuit overcurrent alarm based on perceived fault features, such as Figure 3 As shown, it specifically includes three steps.

[0075] S151: The domain invariant prototype v is derived from the source domain prototype v s and the target domain prototype v t Joint dominance, as shown in formula (13):

[0076]

[0077] In the formula, norm(.) represents the normalization function, r is the threshold of the target domain sample, and n t is the number of samples in the target domain, B t is the target domain cluster.

[0078] S152: Load the domain-invariant prototype v into the dual classifier, use the extracted perceived fault features to input the domain-invariant prototype dual classifier, align the distribution between the source domain and the target domain, and calculate the similarity between the target domain sample to be tested and the domain-invariant prototype. The process function is shown in formula (14):

[0079]

[0080] In the formula, q is the predicted probability of the sample to be tested in the target domain, is the jth voltage value of the target domain sample to be tested, σ is the Sigmoid function, F i is the i-th classifier function, £ is the weight of the dual classifier, V tis the prototype of the target domain, d is the number of dual classifier layers, h is the function of overcurrent prediction, and Y is the position function of the maximum prediction.

[0081] S153: Iteratively update the weight parameters of the domain invariant prototype dual classifier, and input the sample to be tested of the electronic DC load into the trained domain invariant prototype dual classifier F 1 and F 2 In this process, the predictive overcurrent alarm for the electronic DC load circuit is realized. The process function is shown in formula (15):

[0082]

[0083] Where ε represents the prediction loss of the source domain training samples, β is the weight, and ζ represents the supervision loss of the electronic DC load test samples.

[0084] Based on the above discussion, in an optional embodiment of the present invention, for the above-mentioned use of the domain-invariant prototype dual classifier, predictive circuit overcurrent alarm is performed based on the perceived fault characteristics, specifically including:

[0085] In an optional embodiment of S15, the number of layers of the dual classifier is set to 3, the threshold r of the target domain sample is set to 0.4, the weight β of the self-supervised loss is set to 0.15, and the number of iteration steps is 20.

[0086] It should be noted that, in this document, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0087] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core ideas of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. An electronic DC load overcurrent prediction method based on multi-dimensional perception feature extraction is characterized in that: include: The singular value decomposition method is used to denoise the sampling signal of the electronic DC load to obtain a one-dimensional denoised signal. The voltage signal of the electronic DC load is constructed into a Hankel matrix, as shown in formula (1): In the formula, m = N-n+1, ​​x i is the voltage data value, H is the Hankel matrix; Perform singular value decomposition on the Hankel matrix, as shown in formula (2): H=U∑Q T (2) Where U is the left singular vector matrix, and its column vector is the left singular vector of H; Q T is the right singular vector matrix, Its row vectors are the right singular vectors of H; Σ is a diagonal matrix; The one-dimensional denoised signal is converted into a two-dimensional image using the Markov transition field. By calculating the transition frequency between each bin, the state transfer matrix W is constructed, as shown in formula (3): In the formula, w ij Represents the quantile G i Located at quantile G j The frequency after Arrange each probability in time order to obtain the Markov transition field M, and map it according to the state of the one-dimensional denoised signal to obtain a two-dimensional image, as shown in formula (4): In the formula, m ij Represents the bin G i Transfer to bin G j The probability of , p is the probability of the corresponding voltage data value; Based on the dual-scale flexible domain perception module, channel perception features and space perception features are extracted from the two-dimensional image respectively, and the channel perception feature Q is extracted. k1 The process function is shown in formula (5)-formula (7): q1=σ(M1(δ(M1(g)))) (6) Q k1 =q1⊙X k1 (7) In the formula, It is expressed as the average value of all points in the cth channel in the nth sample, where L is the sample height, E is the sample width, and X k1 is a channel-aware sample, (i, j) represents the coordinates of the point, q1 is the channel-aware vector, M1 is the shared dimension-increasing layer, δ is the ReLU operation, g is the channel descriptor, and ⊙ is the channel-aware multiplication; Extract spatial perception features Q k2 The process function is shown in formula (8)-formula (9): q2=σ(M2(δ(M2(g)))) (8) Q k2 =q2⊙X k2 (9) Where q2 is the spatial perception vector, M2 is the shared dimension-increasing layer, δ is the ReLU operation, g is the channel descriptor, ⊙ is the channel perception multiplication, and X k2 is a channel-aware sample; The adaptive feature fusion module is used to fuse the channel perception features and the spatial perception features to obtain the perception fault features. The spatial importance weights of the feature maps are adaptively learned through gradient back propagation. The process functions are shown in Equations (10) and (11): In the formula, t is α or β, a is the number of layers, i and j are the coordinates on the feature map, and λ is the weight coefficient; The perceived fault feature is obtained by weighted summation, and its function is shown in formula (12): In the formula, is the (i, j)th vector of the feature map of the a-th layer along the channel, Represents the feature vector at (i, j) on the feature map adjusted from layer 1 to layer a; The domain-invariant prototype dual classifier is used to predict circuit overcurrent alarm based on the perceived fault features. The domain-invariant prototype v is composed of the source domain prototype v s and the target domain prototype v t Jointly dominate, as shown in formula (13): In the formula, norm(.) represents the normalization function, r is the threshold of the target domain sample, and n t is the number of samples in the target domain, B t is the target domain cluster; Calculate the similarity between the target domain sample and the domain invariant prototype. The process function is shown in formula (14): In the formula, q is the predicted probability of the sample to be tested in the target domain, is the jth voltage value of the target domain sample to be tested, σ is the Sigmoid function, F i is the i-th classifier function, ￡ is the weight of the dual classifier, V t is the prototype of the target domain, d is the number of dual classifier layers, h is the function of overcurrent prediction, and Y is the position function of the maximum prediction; The weight parameters of the domain invariant prototype dual classifier are iteratively updated. The process function is shown in formula (15): Where ε represents the prediction loss of the source domain training samples, β is the weight, and ζ represents the supervision loss of the electronic DC load test samples.

2. The electronic DC load overcurrent prediction method as claimed in claim 1, characterized in that: The sampling signal of the electronic DC load is denoised using the singular value decomposition method to obtain a one-dimensional denoised signal, including: The sampled signal is constructed into a Hankel matrix, and the Hankel matrix is ​​subjected to singular value decomposition to obtain singular values ​​and corresponding singular vectors. The first k largest singular values ​​are selected to reconstruct the Hankel matrix, and the diagonal elements are extracted from the reconstructed Hankel matrix to obtain a one-dimensional denoised signal.

3. The electronic DC load overcurrent prediction method as claimed in claim 1, characterized in that: The one-dimensional denoised signal is converted into a two-dimensional image using the Markov transition field, including: The one-dimensional denoised signal is divided into G bins, the transition frequency between each bin is calculated, the state transfer matrix is ​​established through the transfer frequency, the Markov transition field is constructed through the state transfer matrix, and a two-dimensional image is obtained.

4. The electronic DC load overcurrent prediction method as claimed in claim 1, characterized in that: Based on the dual-scale flexible domain perception module, channel perception features and space perception features are respectively extracted from the two-dimensional image, including: taking the two-dimensional image as the input of the dual-scale flexible domain perception module, the dual-scale flexible domain perception module consists of a channel domain perception module and a space domain perception module, the channel domain perception module embeds the global spatial information of the two-dimensional image into the channel through global average pooling, and extracts channel perception features; the space domain perception module obtains the spatial statistical information of the two-dimensional image through group norms, and extracts space perception features.

5. The electronic DC load overcurrent prediction method as claimed in claim 1, characterized in that: Adopt the adaptive feature fusion module to fuse the channel perception features and space perception features to obtain the perceived fault features, including: The channel-aware features and spatial-aware features are used as the input of the adaptive feature fusion module respectively. Through upsampling and convolution, the channel-aware features and spatial-aware features are adjusted to feature maps with the same size and the same number of channels. They are weighted and summed up through their respective weight coefficients to obtain the fused perceived fault features.

6. The electronic DC load overcurrent prediction method as claimed in claim 1, characterized in that: A domain-invariant prototype dual classifier is used to perform predictive circuit overcurrent alarm based on perceived fault features, including: The domain-invariant prototype dual classifier is used to align the distribution between the source domain and the target domain, and the similarity between the target domain sample and the domain-invariant prototype is calculated. The perceived fault features are input into the domain-invariant prototype dual classifier to perform predictive overcurrent alarm on the electronic DC load circuit.

Images