Automatic equipment type classification method based on EMC test spectrum curve

Abstract

The invention discloses an equipment type automatic classification method based on an EMC test frequency spectrum curve. The method comprises a data enhancement method and a curve-based equipment classification method. The data enhancement method comprises the steps of generating a new data set on the basis of obtaining a labeled data set; training a generative model and a semantic analysis model in a fully supervised manner by using the labeled data set, wherein the generative model is used for generating an EMC test spectrum curve from a mask; in the iteration generation process, the labeled data set and the generation model are fixed, only the semantic analysis model is continuously optimized, and the semantic analysis model is used for carrying out category label setting on the generation data; the equipment classification method adopts an RNN curve classification technology of a special attention mechanism. According to the method, by combining a small amount of labeled data, a large amount of labeled equipment category data is generated, and the number of effective EMC test spectrum curve equipment classification training data is increased; and by introducing a strong prior knowledge RNN + LSTM technology, equipment is better classified through an EMC test frequency spectrum curve.

Description

[0001] Technology Neighborhood This invention relates to the field of EMC test spectrum curve processing technology, and in particular to an automatic equipment classification method based on EMC test spectrum curves. Background Technology

[0002] EMC (Electromagnetic Compatibility) testing covers virtually all electronic and electrical equipment. The EMC characteristics, especially EMI (Electromagnetic Emission) characteristics, of different devices vary due to their different internal electromagnetic noise sources. However, the EMI spectrum characteristics of different modules, such as power supply modules and clock crystal oscillators, still exhibit significant differences and certain patterns. Furthermore, noise suppression measures for different noise modules will vary considerably. Therefore, for accurate analysis and design improvements, it is necessary to first analyze the device type based on different EMI spectrum curves. Based on this analysis, a better understanding of the noise module types within the device can be achieved. This invention utilizes deep learning semantic analysis technology to accurately classify the device types represented by the curves.

[0003] Supervised deep learning methods require a large amount of labeled training data to train the model and obtain high-precision results. However, EMC data is limited in quantity, and a large number of high-precision labels are still needed. To address this issue, one approach is to manually label some data, and another is to use generative models to quickly generate a large number of data label pairs. However, current deep learning-based EMC test spectrum curve generation techniques still have the following problems that urgently need to be solved: (1) Data scarcity problem. For EMC test spectrum curve analysis tasks, it is difficult to obtain high-quality labeled data. This is mainly because: case data is scarce, and multi-center data is difficult to obtain in large quantities due to the closed management of each unit; therefore, there is an urgent need for a fast and high-quality automated labeling solution to promote the implementation of deep learning in large-scale EMC test spectrum curve analysis tasks.

[0004] (2) Authenticity issues. EMC test spectrum curves need to accurately reflect the characteristics of the equipment itself, but it is difficult for generative networks to guarantee that they have learned the professional knowledge of this field. Therefore, the EMC test spectrum curves generated by generative networks may have authenticity issues. If the generated EMC test spectrum curves are not authentic, the model trained with this data may affect the judgment and classification results of the equipment.

[0005] (3) Diversity Issues. In the field of EMC test spectrum curve processing, the location, shape, and pixel distribution of sensors often vary greatly due to differences in the type of device and the role of the circuit itself. Compared to the natural EMC test spectrum curve field, the dataset of EMC test spectrum curves is often more limited and cannot cover all possible situations. Therefore, it is difficult to train a model that can effectively process EMC test spectrum curves. In this case, the model lacks diversity and detail when generating EMC test spectrum curves and cannot reflect the real situation well, thus seriously affecting the application value of the model. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for automatically labeling EMC test spectrum curves. This method includes the following steps: Obtain a labeled dataset; train a generative model and a semantic analysis model in a fully supervised manner using the labeled dataset, wherein the generative model is used to generate EMC test spectrum curves from a mask; The trained generative model and semantic analysis model can generate new data and corresponding annotation results through multiple iterations. During the iterative generation process, the labeled dataset and the generative model remain unchanged, while only the semantic analysis model is continuously optimized. This semantic analysis model is used to generate pseudo-labels for unlabeled data.

[0007] Furthermore, by introducing RNN technology with strong prior knowledge, the EMC test spectrum curves can be better classified. Compared with existing technologies, the advantages of this invention lie in providing a method for automatically labeling EMC test spectrum curves of devices. This method combines generative and analytical models for automatic labeling of EMC test spectrum curves, and by combining a small amount of labeled data with the generative model, it addresses the problem of insufficient model training data. Compared with traditional adversarial generative models, this invention, based on data augmentation and RNN+LSTM, introduces a mechanism to enhance the application of prior and subsequent knowledge, ensuring more accurate device classification on a limited number of EMC test spectrum curves.

[0008] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0009] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0010] Figure 1 This is a framework diagram of an automatic generation method for EMC test spectrum curve data according to an embodiment of the present invention; Figure 2 This is a flowchart of an automatic annotation method for EMC test spectrum curves according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an RNN+LSTM network structure with an attention enhancement mechanism according to an embodiment of the present invention; Figure 4 This is a schematic diagram of data augmentation technology according to an embodiment of the present invention; Figure 5 This is a schematic diagram of data annotation according to an embodiment of the present invention. Detailed Implementation

[0011] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0012] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0013] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0014] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0015] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0016] See Figure 1 As shown, a novel data framework combining manual annotation and generative models is proposed, aiming to automatically generate a large number of labeled EMC test spectrum curves based on a small amount of labeled data. In this paper, the manually labeled dataset is defined... The generated dataset is defined as .

[0017] Combination Figure 1 It can be seen that the provided generation methods mainly include: Step S1: Obtain the labeled EMC test spectrum curve. For example, this can be obtained through manual annotation; Step S2, using The EMC test spectrum curve translation model Z and semantic analysis model Z are trained in a fully supervised manner to generate EMC test spectrum curves from masks. ; Step S3, use semantic analysis model Z Obtain EMC test spectrum curve pseudo-tags ; Step S4, use model Z to obtain a pair of labeled data ( At this point, model Z will Translation obtained ,in yes The post-processing results.

[0018] Iterative annotation process: In the process of moving to the next iteration of annotation, the data... With model The model remains unchanged. Continue to optimize, at this point The training data used is + A new analytical model was obtained. Then, continue with steps S3 and S4 to obtain new generated data pairs, which completes the next iteration. During the iteration process, only the analysis model (i.e., the semantic analysis model) needs to have its parameters updated.

[0019] In step S5, the encoder of the RNN+LSTM has a hidden layer vector h for each LSTM unit, and the decoder has a corresponding hidden layer vector ht for each LSTM unit. While the hidden layer vector h does have some relationship with the semantic feature vector, they are not the same thing. In the attention mechanism, Ct is a weighted sum of many hts, resulting in a feature vector that represents the semantics of a single time step.

[0020] To overcome the "forgetting problem," the original method of synthesizing a semantic vector at the end was changed to: recording the hidden layer state h at each step for future reference, essentially writing down each step in a notebook. In other words, the hidden layer outputs h1, h2, ..., ht at each time step correspond to a semantic vector C1, C2, ..., Ct, as shown below. Figure 2 As shown: In this way, the calculation formula changes, and during decoding, it is only necessary to consult the semantic vector at the encoding end of the corresponding time step to perform decoding more efficiently, as shown in the following formula: y1 = f(C1) y2 = f(C2, y1) y3 = f(C3, y1, y2) Specifically, see Figure 3 As shown, the provided method for automatically generating EMC test spectrum curves includes the following steps: Step S110: Obtain labeled data and unlabeled data.

[0021] In one embodiment, the effectiveness of the provided automatic annotation framework is verified using the EMC dataset provided by the company as an example.

[0022] The EMC dataset contains a total of 486 cases with masked labels. Of these, 86 cases are used as the test set, and the remaining 400 labeled cases are divided into two parts: 20 (accounting for 5%) and 380. The analysis model and the generative model are trained in full supervision using the 20 labeled cases, and the 380 cases are labeled. Finally, the test set is used to verify the difference between the labeling results of this invention and the original labels.

[0023] Step S120: Using labeled data, train the EMC test spectrum curve translation model and semantic analysis model in a fully supervised manner to generate EMC test spectrum curves from the mask.

[0024] (1) Semantic analysis model The analysis model is used to generate pseudo-labels, and various types of analysis models can be flexibly selected. Commonly used models include Short-Time Fourier Models, Hidden Markov Models, RNNs, and RNN+LSTM. A good analysis model can generate more accurate pseudo-labels, thereby reducing the gap between the EMC test spectrum curve generated by the EMC test spectrum curve translation model and the real EMC test spectrum curve, effectively reducing the number of iterations.

[0025] (2) Generative model structure design See Figure 1 As shown, generative models can also employ various architectures, such as generative adversarial networks or other types. Assuming... This is the output of the activation layer of the i-th deep neural network, and the batch size is... The number of channels is , and It is the height and width of the activation feature map of this layer. Edge information extracted by Hidden Markov Operators is similar to batch normalization algorithms. The activation layer output is normalized on a channel-by-channel basis and then modulated using a learnable scaling factor and offset. For example, in ( At position ), the activation value is given by the following formula: (1) in It is the output of the activation function. and They are Mean and variance of each channel: (2) (3) In formula (1) , and These are all learnable parameters, unlike BatchNormalization. and Depends on the mask , The output characteristics depend on the details provided by the Canny operator. .

[0026] (3) Generative model training In one embodiment, the model's training hyperparameters and optimizer follow the VAE open-source framework, which has been redesigned for style control and data augmentation techniques.

[0027] For the encoder part, style condition control is performed by encoding the statistical distribution of the EMC test spectrum curve using VAE. In order to achieve more effective style control, in one embodiment, an encoder with an RNN as the skeleton is used to extract the semantic information of the style EMC test spectrum curve. Ideally, it can be assumed that the deepest feature layer of the encoder contains all the information of the input EMC test spectrum curve. When the last feature layer of the encoder is input into the decoder, its style can be effectively transferred to the generated EMC test spectrum curve of the decoder.

[0028] To prevent the decoder from directly acquiring the structural information of the style EMC test spectrum curve and ignoring the control effect of the mask map, the structure of the input style control map was shuffled, so that it only provides statistical style control. Patch shuffle data augmentation technology was used during this process. Figure 4 As shown, patch shuffle primarily expands the structure of EMC test spectrum curves. Compared to deformation, flipping, and scaling in traditional data augmentation techniques, patch shuffle involves more dramatic changes and greater structural diversity. For example, for curves with lengths and widths of... and The picture The formula for calculating the number of amplified EMC test spectrum curves with different structures that can be obtained by the patchshuffle data augmentation method is as follows: (4) in, The size of the patch; % represents integer division, and ! represents factorial operation. At this time, EMC test spectrum curves for 24 different structures can be generated. This exponential data amplification is very effective for structural variations.

[0029] (4) Loss function Local SSIM In training the generator, in addition to using the same multi-scale discriminator and loss function as GauGAN, to enhance the structural information of the generated data, an additional local structural similarity constraint generator for the EMC test spectrum curve was added in one embodiment, called Local Structural Similarity Loss (local SSIM). The algorithm flow is shown in Table 1 below. Unlike the pixel-wise mean square and MSE loss functions, SSIM (structural similarity) comprehensively considers factors such as brightness, contrast, and structure, which can avoid the blurring of the generated EMC test spectrum curve caused by strong pixel-wise constraints. Compared with the traditional PSNR (Peak Signal-to-Noise Ratio) metric, SSIM better reflects the human eye's perception of the quality of the EMC test spectrum curve because it considers the structural information of the EMC test spectrum curve, not just the differences between pixels.

[0030] Table 1 Local SSIM loss algorithm Step S130: Using a pre-trained semantic analysis model and a generative model, an EMC test spectrum curve with device category labeling information is generated through multiple iterations.

[0031] By utilizing a semantic analysis model and a generative model that have undergone preliminary fully supervised training, new EMC test spectrum curve data with annotation information can be obtained through an iterative annotation process. For example, the iterative process involves generating the next iteration of data from the already labeled data. With generative models Fixed and unchanging semantic analysis model Continue to optimize, at this point The training data used is + A new analytical model was obtained. Through multiple iterations, EMC test spectrum curves can be generated and automatically labeled. Only the analysis model requires parameter updates during the iteration process.

[0032] Step S140: Use an RNN+LSTM network with empirical information to classify the device curves.

[0033] Based on the data, a new network framework using RNN+LSTM with an empirical attention mechanism changes the original method of synthesizing a semantic vector at the end. Instead, it records the hidden layer state h at each step. In other words, the hidden layer outputs h1, h2, ..., ht at each time step generate a corresponding semantic vector C1, C2, ..., Ct. The device is then classified using EMC test spectrum curves.

[0034] In summary, this invention directly generates synthetic EMC test spectrum curves from the encoder's feature maps, rather than using reparameterization techniques. The advantage of this invention is that it makes the distribution of the synthetic EMC test spectrum curves more similar to the input, and requires less labeled data and training costs. The biggest difference from EMC test spectrum curve editing is that the goal of this invention is to recover the input information and obtain its mask from the feature map without loss. In the analysis data, the pseudo-labels of each input differ from the true labels of the original image, so the synthetic data is consistent with the pseudo-labels, and the distribution of the EMC test spectrum curves is very close to the original image. Based on the data, a new network framework of RNN+LSTM with an empirical attention mechanism is used to classify devices using EMC test spectrum curves.

[0035] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. An automatic equipment classification method based on EMC test spectrum curves, characterized in that, include: Obtain the labeled dataset and process it according to the image format. The labeled dataset is used to train a generative model and a semantic model in a fully supervised manner, wherein the generative model is used to generate EMC test spectrum curves from a mask; The generative model and the semantic model undergo multiple iterative annotation processes to obtain generated data. During the iteration process, the labeled dataset and the generative model remain unchanged, while only the semantic analysis model is continuously optimized. The semantic analysis model is used to generate pseudo-labels for the generated data.

2. The automatic equipment classification method based on EMC test spectrum curves according to claim 1, characterized in that, The input to the generative model includes a style control map, a semantic mask, and a Canny feature map, where the style control map and the semantic mask are corresponding label pairs, and the Canny feature map is the Canny calculation result of the style control map.

3. The automatic equipment classification method based on EMC test spectrum curves according to claim 2, characterized in that, The style control maps are structurally shuffled to provide only statistical style control, and the number of style control maps is increased using patchshuffle data augmentation.

4. The automatic equipment classification method based on EMC test spectrum curves according to claim 1, characterized in that, The loss function used to train the generative model includes a local structural similarity loss.

5. The automatic equipment classification method based on EMC test spectrum curves according to claim 1, characterized in that, The overall framework of the generative model is an encoder-decoder structure, where the decoding structure is built based on VAE, and strong prior knowledge of image details is added in the following way: assuming This is the output of the i-th activation layer, and the batch size is... The number of channels is , and It is the height and width of the activation feature map of this layer. The edge information extracted by the Canny operator is normalized in channel-by-channel output by the activation layer, then modulated by a learnable scaling factor and offset, and finally the Canny operator is added to provide strong prior knowledge of the edge information.

6. The automatic equipment classification method based on EMC test spectrum curves according to claim 3, characterized in that, The number of style control charts can be increased using the following formula: ； in, This refers to the patch size. % represents integer division, and ! represents factorial. and These are the length and width of the style control chart, respectively. C is the number of original style control maps, and C is the number of style control maps obtained after patch shuffle.

7. The automatic equipment classification method based on EMC test spectrum curves according to claim 1, characterized in that, The labeled dataset was obtained from the labeled dataset provided by the company.

8. The automatic equipment classification method based on EMC test spectrum curves according to claim 1, characterized in that, This includes RNN techniques with special attention mechanisms.

Images