CN-121997776-A - Super-surface sensor design method and device based on weak supervision and electronic equipment

Abstract

The invention belongs to the technical field of optical device design, and relates to a method and a device for designing a super-surface sensor based on weak supervision and electronic equipment. The method comprises the steps of obtaining an attribute data set of a super-surface sensor and carrying out data processing, wherein the attribute data set comprises a structural parameter vector and a spectral response vector, inputting the attribute data set into a dual-branch encoder network, extracting local features and global features, inputting the local features and the global features into a global context injection module, enhancing the local features, carrying out feature splicing and fusion with the global features to obtain comprehensive features, inputting the comprehensive features into a progressive resolution decoder, outputting an attribute prediction result according to intermediate prediction results of each resolution decoding stage, calculating task total loss according to the attribute prediction result and the attribute data set, and iteratively optimizing and updating the attribute prediction result according to the task total loss. Design efficiency is improved, and dependence on a large-scale labeling data set is reduced by relying on weak supervision core logic.

Inventors

• Ju Fayin
• LI NING

Assignees

• 浙江优众新材料科技有限公司

Dates

Publication Date

20260508

Application Date

20260408

Claims (10)

1. 1. A method of designing a weakly supervised based ultra surface sensor, comprising: acquiring an attribute data set of a super-surface sensor and performing data processing, wherein the attribute data set comprises a structural parameter vector and a spectral response vector; Inputting the attribute data set into a dual-branch encoder network, and extracting local features and global features; inputting the local features and the global features into a global context injection module, enhancing the local features, and performing feature splicing and fusion with the global features to obtain comprehensive features; inputting the comprehensive characteristics into a progressive resolution decoder, and outputting attribute prediction results according to intermediate prediction results of each resolution decoding stage; And calculating the total loss of the task according to the attribute prediction result and the attribute data set, and updating the attribute prediction result according to the iterative optimization of the total loss of the task.
2. 2. The method of claim 1, wherein performing data processing comprises: downsampling the spectrum response vector to obtain spectrum data with different resolutions; normalizing the structural parameter vector and the spectral response vector, respectively; splicing the normalized spectral response vector and the normalized spectral data according to channel dimensions to obtain sequence data of corresponding dimensions; and carrying out data enhancement on the structural parameter vector, the spectral response vector and the sequence data through a data enhancement strategy.
3. 3. The method of claim 2, wherein inputting the set of attribute data into a dual-branch encoder network extracts local features and global features, comprising: extracting local features in different ranges from the sequence data by using different receptive field feature extractors; And dividing the sequence data by using an encoder, adding position codes to the divided fragments, and extracting global features of the divided fragments corresponding to each position code based on a multi-head self-attention mechanism of the encoder.
4. 4. The method of claim 2, wherein inputting the local feature and global feature to a global context injection module, enhancing the local feature, comprises: carrying out global pooling on the global features, and converting the global features into intermediate vectors of corresponding dimensions; Inputting the intermediate vector into a multi-layer neural network to perform nonlinear transformation to obtain a global context vector; calculating the association weight of the global context vector and the local feature; And injecting the global context vector into the local feature through element-by-element multiplication according to the association weight.
5. 5. The method of claim 2, wherein inputting the integrated feature into a progressive resolution decoder, outputting an attribute prediction result based on intermediate prediction results of respective resolution decoding stages, comprises: the resolution decoding stage comprises a first resolution stage, a second resolution stage and a third resolution stage, wherein the resolution of the first resolution stage is minimum, and the resolution of the third resolution stage is maximum; The first resolution stage analyzes according to the comprehensive characteristics to generate a structure parameter intermediate prediction result in a first preset precision range; The second resolution stage analyzes according to the comprehensive characteristics and the structure parameter intermediate prediction results to generate spectral response intermediate prediction results in a second preset precision range; The third resolution stage executes a forward design task and a reverse design task according to the comprehensive features, the structure parameter intermediate prediction results, the spectrum response intermediate prediction results and the enhanced local features respectively, and correspondingly generates structure parameter prediction results and spectrum response prediction results in a third preset precision range; the precision requirement of the first preset precision range is the lowest, and the precision requirement of the third preset precision range is the highest.
6. 6. The method of claim 5, wherein calculating a total loss of task from the attribute prediction result and the attribute dataset, iteratively optimizing updating the attribute prediction result from the total loss of task, comprises: calculating the difference loss between the structure parameter prediction result and the structure parameter vector to obtain a structure parameter loss result; Calculating the difference loss between the spectrum response prediction result and the spectrum response vector to obtain a spectrum response loss result; Executing the reverse design task again according to the spectrum response prediction result, generating a secondary structure parameter prediction result, and calculating the difference loss between the secondary structure parameter prediction result and the structure parameter vector to obtain a consistency loss result; the structural parameter loss result, the spectral response loss result and the consistency loss result are weighted and summed to obtain a task total loss; And correcting each loss result according to the task total loss and the physical constraint rule corresponding to the attribute data set, and iteratively optimizing and updating the structure parameter prediction result and the spectral response prediction result.
7. 7. The method of claim 1, further comprising, after performing the data processing: performing model initial training according to the processed attribute data set, and performing attribute prediction on the unlabeled data set by using the trained model to generate a pseudo tag corresponding to the unlabeled data; Calculating the confidence coefficient score of the pseudo tag, and if the confidence coefficient score is larger than a preset confidence coefficient threshold value, adding corresponding unlabeled data and the pseudo tag to the attribute data set to complete sample expansion; And repeating the steps of model training, unlabeled data attribute prediction and pseudo tag screening and adding based on the extended attribute data set, and iteratively optimizing the attribute data set until the preset iteration times or preset early-stop conditions are reached.
8. 8. The method of claim 4, wherein outputting the attribute prediction result further comprises: Acquiring attention weights and the association weights corresponding to branches for extracting global features; Mapping the attention weight and the association weight by adopting a gradient weighting type activation mapping method, and respectively calculating and obtaining the prediction contribution degree of each spectrum region to the attribute prediction result; And respectively generating corresponding attention thermodynamic diagrams according to the attention weights and the corresponding prediction contribution degrees thereof, the association weights and the corresponding prediction contribution degrees thereof.
9. 9. A weakly supervised based ultra surface sensor design apparatus, comprising: The data acquisition and processing module is used for acquiring an attribute data set of the super-surface sensor and processing data, wherein the attribute data set comprises a structural parameter vector and a spectral response vector; The extraction feature module is used for inputting the attribute data set into a double-branch encoder network and extracting local features and global features; the feature fusion module is used for inputting the local features and the global features into the global context injection module, enhancing the local features, and carrying out feature splicing and fusion with the global features to obtain comprehensive features; The attribute prediction module is used for inputting the comprehensive characteristics into a progressive resolution decoder and outputting attribute prediction results according to intermediate prediction results of each resolution decoding stage; And the iteration optimization module is used for calculating the total loss of the task according to the attribute prediction result and the attribute data set, and iteratively optimizing and updating the attribute prediction result according to the total loss of the task.
10. 10. An electronic device, comprising: A processor; A memory for storing processor-executable instructions; Wherein the processor is configured to implement the weakly-supervised ultra surface sensor design approach of any one of the preceding claims 1-8 when executing the executable instructions.

Description

Super-surface sensor design method and device based on weak supervision and electronic equipment Technical Field The invention belongs to the technical field of optical device design, and relates to a method and a device for designing a super-surface sensor based on weak supervision and electronic equipment. Background In recent years, the super-surface sensor based on the continuous domain constraint state has remarkable potential in the fields of molecular fingerprint detection and refractive index sensing, the continuous domain constraint state can regulate radiation loss, the intensity of a local light field is improved, and the quasi-continuous domain constraint state mode of symmetrical protection conversion is broken, so that the near-field local capacity and the sensing performance of the sensor are expected to be enhanced by virtue of excellent resonance service life and high quality factors. The traditional numerical optimization design method is time-consuming and labor-consuming, and the deep learning driven design method accelerates the process, but has the limitations of difficult acquisition of training samples, insufficient capture of high-precision spectral characteristics, lack of interpretability of models and the like. The weak supervision learning method provides a new thought for relieving the problem of insufficient training samples, and has the core of gradually improving the quality of the training samples by utilizing the self-prediction capability of the model to form a self-enhancement learning cycle, and realizing good performance in the field of computer vision through a gradual iteration optimization strategy of a segmentation model, but the application of the model to the design of a super-surface sensor still faces the challenges of physical constraint integration, multi-scale feature characterization, accurate modeling of complex optical response and the like. At present, the intelligent design of the super-surface sensor also faces the technical bottlenecks of insufficient model generalization capability, lost high-precision spectrum characteristic prediction details, opaque model decision-making process and the like, and the problems of peak position deviation, inaccurate line width prediction and the like often occur when the conventional method processes sharp formant optical response, so that the accuracy of sensor performance evaluation is seriously affected. Disclosure of Invention The invention aims to solve the problems in the prior art and provides a design method of a super-surface sensor based on weak supervision. The invention aims at realizing the following technical scheme that the design method of the super-surface sensor based on weak supervision comprises the following steps: acquiring an attribute data set of a super-surface sensor and performing data processing, wherein the attribute data set comprises a structural parameter vector and a spectral response vector; Inputting the attribute data set into a dual-branch encoder network, and extracting local features and global features; inputting the local features and the global features into a global context injection module, enhancing the local features, and performing feature splicing and fusion with the global features to obtain comprehensive features; inputting the comprehensive characteristics into a progressive resolution decoder, and outputting attribute prediction results according to intermediate prediction results of each resolution decoding stage; And calculating the total loss of the task according to the attribute prediction result and the attribute data set, and updating the attribute prediction result according to the iterative optimization of the total loss of the task. As an alternative embodiment of the present invention, performing data processing includes: downsampling the spectrum response vector to obtain spectrum data with different resolutions; normalizing the structural parameter vector and the spectral response vector, respectively; splicing the normalized spectral response vector and the normalized spectral data according to channel dimensions to obtain sequence data of corresponding dimensions; and carrying out data enhancement on the structural parameter vector, the spectral response vector and the sequence data through a data enhancement strategy. As an alternative embodiment of the present invention, inputting the attribute dataset into a dual-branch encoder network, extracting local features and global features, comprises: extracting local features in different ranges from the sequence data by using different receptive field feature extractors; And dividing the sequence data by using an encoder, adding position codes to the divided fragments, and extracting global features of the divided fragments corresponding to each position code based on a multi-head self-attention mechanism of the encoder. As an alternative embodiment of the present invention, inputting the local feature