Silicon-based Fourier transform spectrum reconstruction method based on deep learning
By using deep learning-based pseudo-inverse computation and generative adversarial network spectral prediction models, the problems of silicon-based waveguide processing errors and noise effects are solved, and fast and accurate spectral reconstruction under imperfect interference signals is achieved.
Patent Information
- Application Number
- CN202511620610.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-02-13
AI Technical Summary
The actual fabrication errors of silicon-based waveguides prevent traditional Fourier transform methods from being used directly to extract accurate spectral information from distorted interference signals. Furthermore, detector noise and thermal noise affect the quality of the interferogram, which in turn affects the spectral reconstruction results.
A deep learning-based approach is adopted, which involves pseudo-inverse calculation and high-rank filtering of interferograms, combined with a generative adversarial network spectral prediction model, and trained using simulation and real-world training datasets to achieve spectral reconstruction.
Even with imperfect interference signals, rapid and accurate silicon-based Fourier transform spectral measurements were achieved, improving the accuracy and noise immunity of spectral measurements.
Smart Images
Figure CN121528346A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of silicon-based micro-spectrum, and particularly relates to a silicon-based Fourier transform spectrum reconstruction method based on deep learning. BACKGROUND
[0002] When a substance interacts with light, it will absorb, disperse or scatter light in its unique way, and the characteristics of the light after the interaction will change, like a "fingerprint" of the substance. Based on this principle, spectrum detection technology can determine the composition, structure or content of the substance. According to different forms of light and substance interaction, spectrum detection technology can be divided into infrared spectrum, Raman spectrum, fluorescence spectrum and the like. Spectrum detection technology has outstanding advantages such as rapidity, multi-parameter, precision and non-destructiveness, and has been widely used in many fields such as medical diagnosis, chemical industry, astronomical remote sensing, food safety and environmental protection.
[0003] Fourier transform spectrometer (FTS) is a high-performance spectrum detection device, which has been widely used in many fields due to its outstanding advantages of high resolution and high signal-to-noise ratio. To expand its application scenarios and enhance its practicability, miniaturization of FTS has become an important development trend. In particular, the chip-level micro-FTS based on silicon-based photonic integration technology constructs a completely solidified silicon-based hardware structure through photolithography process, forming a highly compact integrated photonic circuit. Such structure not only has excellent multi-platform compatibility, stability and anti-interference ability, but also is compatible with complementary metal oxide semiconductor (CMOS) process, which is suitable for large-scale production, thereby providing strong support for the popularization and practical application of spectrum detection technology.
[0004] According to the different optical path difference (OPD) generation mechanisms, silicon-based photonic integrated FTS can be mainly divided into three types: active scanning type, spatial heterodyne type and standing wave integrated type. Among them, the spatial heterodyne FTS adopts a simple and stable Mach-Zehnder interferometer (MZI) array. By setting specific waveguide width and arm length difference for each MZI unit, a linearly increasing OPD sampling sequence can be constructed to achieve ideal interference modulation effect and meet the theoretical requirements of Fourier transform spectrum measurement. Compared with other types, this structure eliminates the active physical modulation link and complex hardware support, and shows stronger practical application potential.
[0005] However, due to the limitation of silicon waveguide processing technology, errors are inevitable in actual production, causing the OPD sampling sequence to deviate from the ideal state. In this case, the traditional Fourier transform method cannot be directly used to extract accurate spectral information from the distorted interference signal. At the same time, in the actual interference pattern acquisition process, the detector shot noise and thermal noise will affect the quality of the interference pattern, and further affect the spectrum reconstruction by Fourier transform. SUMMARY
[0006] To solve the problem that the actual processing error of the silicon waveguide causes the Fourier transform demodulation of the interference signal to be unable to extract spectral information, the present application provides a silicon-based Fourier transform spectral reconstruction method and device based on deep learning, equipment, medium and product, so that in the case of imperfect interference signal, fast and accurate Fourier transform spectral measurement can also be realized.
[0007] In a first aspect, the present application provides a silicon-based Fourier transform spectral reconstruction method based on deep learning, comprising:
[0008] Obtaining the interference pattern of the to-be-measured spectrum after passing through the set array;
[0009] Based on the pseudo-inverse calculation processing of the "spectrum-interference pattern" system response matrix of the to-be-measured spectrum corresponding to the interference pattern, the pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum is obtained, and the model input spectrum corresponding to the to-be-measured spectrum is obtained after high-rank filtering processing of the pseudo-inverse matrix calculation spectrum;
[0010] The model input spectrum is input into a pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum, and a spectral reconstruction result is determined based on the predicted spectrum;
[0011] The generative adversarial network spectrum prediction model is trained based on a simulation training data set and a real acquisition training data set, and the simulation training data set includes spectral data of different detection noises, different processing errors and different temperature fluctuations.
[0012] In a second aspect, the present application provides a silicon-based Fourier transform spectral reconstruction device based on deep learning, comprising:
[0013] An interference pattern acquisition module is configured to obtain the interference pattern of the to-be-measured spectrum after passing through the set array;
[0014] A model input spectrum determination module is configured to perform pseudo-inverse calculation processing on the interference pattern corresponding to the to-be-measured spectrum based on the "spectrum-interference pattern" system response matrix of the set array, obtain the pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum, and obtain the model input spectrum corresponding to the to-be-measured spectrum after high-rank filtering processing of the pseudo-inverse matrix calculation spectrum;
[0015] a spectrum prediction module, configured to input the model input spectrum into a pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum, and determine a spectrum reconstruction result based on the predicted spectrum; wherein the generative adversarial network spectrum prediction model is trained based on a simulation training dataset and a real acquisition training dataset, and the simulation training dataset includes spectrum data of different detection noises, different processing errors and different temperature fluctuations.
[0016] In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory connected to the processor in communication;
[0017] The memory stores computer execution instructions.
[0018] The processor executes the computer execution instructions stored in the memory to implement the method of any one of the first aspect.
[0019] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the method of any one of the first aspect.
[0020] In a fifth aspect, the present application provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor to implement the method of any one of the first aspect.
[0021] The present application provides a silicon-based Fourier transform spectrum reconstruction method and device based on deep learning, equipment, medium and product, specifically, by obtaining the interference pattern of the to-be-measured spectrum after setting the array; based on the pseudo-inverse calculation processing of the interference pattern corresponding to the to-be-measured spectrum by the "spectrum-interference pattern" system response matrix of the set array, the pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum is obtained, the pseudo-inverse matrix calculation spectrum is processed by high-rank filtering, and the model input spectrum corresponding to the to-be-measured spectrum is obtained; the model input spectrum is input into a pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum, and a spectrum reconstruction result is determined based on the predicted spectrum; wherein the generative adversarial network spectrum prediction model is trained based on a simulation training dataset and a real acquisition training dataset, and the simulation training dataset includes spectrum data of different detection noises, different processing errors and different temperature fluctuations, so that in the case of imperfect interference signal, fast and accurate silicon-based Fourier transform spectrum measurement can also be realized. BRIEF DESCRIPTION OF DRAWINGS
[0022] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0023] Figure 1A flowchart of a deep learning-based silicon-based Fourier transform spectrum reconstruction method provided by an embodiment of the present application is shown in FIG. 1.
[0024] Figure 2 A flowchart of another deep learning-based silicon-based Fourier transform spectrum reconstruction method provided by an embodiment of the present application is shown in FIG. 2.
[0025] Figure 3 A structural diagram of a deep learning-based silicon-based Fourier transform spectrum reconstruction device provided by an embodiment of the present application is shown in FIG. 3.
[0026] Figure 4 A structural diagram of an electronic device provided by an embodiment of the present application is shown in FIG. 4.
[0027] The specific embodiments of the present application have been shown in the above-described drawings, and will be described in more detail hereinafter. These drawings and the written description are not intended to restrict the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION
[0028] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The same or similar components are denoted by the same reference numerals throughout the drawings and the following description, unless otherwise specified. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application, as detailed in the appended claims.
[0029] Figure 1 A flowchart of a deep learning-based silicon-based Fourier transform spectrum reconstruction method provided by an embodiment of the present application is shown in FIG. 1. Figure 1 As shown in FIG. 1, the method of the present embodiment includes:
[0030] S101, obtaining an interference pattern of a to-be-tested spectrum after passing through a set array.
[0031] In the embodiment, the "spectrum-spectrum" conversion is performed by the pre-trained generative adversarial network spectrum prediction model, instead of the conventional "interferogram-spectrum" conversion, to realize more accurate and reliable spectrum prediction on the basis of optimizing the calculation amount. Based on this, the corresponding interferogram of the to-be-measured spectrum is determined first in combination with the setting structure in the silicon-based spatial heterodyne FTS, and then the model input spectrum is determined based on the interferogram corresponding to the to-be-measured spectrum, so as to perform spectrum reconstruction by the model input spectrum and the pre-trained generative adversarial network spectrum prediction model. The simple and stable Mach-Zehnder Interferometer (MZI) array structure is adopted in the conventional silicon-based spatial heterodyne FTS, and based on this, the interferogram corresponding to the to-be-measured spectrum can be determined based on the MZI array.
[0032] In S102, the pseudo-inverse calculation processing is performed on the interferogram corresponding to the to-be-measured spectrum based on the "spectrum-interferogram" system response matrix of the setting array, to obtain the pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum. The high-rank filtering processing is performed on the pseudo-inverse matrix calculation spectrum, to obtain the model input spectrum corresponding to the to-be-measured spectrum.
[0033] Considering the influence of different system response matrices in the silicon-based chip on the interferogram, and the influence of the acquisition noise and other parameters on the interferogram, the pseudo-inverse calculation and denoising processing need to be performed on the interferogram corresponding to the to-be-measured spectrum between the input models, to obtain the model input spectrum, so that the spectrum reconstruction result based on the model input spectrum is more accurate.
[0034] In an embodiment of the present application, the setting array includes an MZI array, and the pseudo-inverse calculation processing performed on the interferogram corresponding to the to-be-measured spectrum based on the "spectrum-interferogram" system response matrix of the setting array to obtain the pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum includes:
[0035] The system response matrix of the MZI array is obtained, and the pseudo-inverse matrix of the system response matrix is determined.
[0036] According to the linear relationship among the interferogram, the pseudo-inverse matrix of the system response matrix, and the pseudo-inverse matrix calculation spectrum, the pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum is determined.
[0037] Optionally, in actual spectrum reconstruction, the system response matrix of the MZI array is obtained by wavelength scanning measurement of a high-resolution narrowband laser.
[0038] Exemplarily, the system response matrix A of the MZI array is obtained first by high-resolution narrowband laser scanning measurement, the pseudo-inverse matrix A + of the system response matrix A is determined, and then the pseudo-inverse matrix calculation spectrum S is obtained according to the linear relationship S=A + I.
[0039] On the basis, the pseudo-inverse matrix calculation spectrum is high-rank filtered to obtain a model input spectrum corresponding to the to-be-measured spectrum, including:
[0040] The pseudo-inverse matrix calculation spectrum is singular value decomposed to filter out a part with a rank higher than a set threshold to obtain the model input spectrum.
[0041] Optionally, a suitable noise filtering mode can be selected based on noise characteristics of conventional noise in the spectrum to perform noise filtering. In an embodiment of the present application, the to-be-determined to-be-measured spectrum can be singular value decomposed to obtain the target to-be-measured spectrum. That is, the singular value decomposition matrix is used to filter out a noise part with a rank higher than a set threshold to obtain the target to-be-measured spectrum as the model input spectrum, wherein the set threshold can be set based on actual conditions and actual requirements.
[0042] S103, inputting the model input spectrum into a pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum, and determining a spectrum reconstruction result based on the predicted spectrum.
[0043] The generative adversarial network spectrum prediction model is trained based on a simulation training data set and a real acquisition training data set, and the simulation training data set includes spectrum data with different detection noises, different processing errors and different temperature fluctuations.
[0044] In the present embodiment, the model input spectrum is input into the pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum output by the generative adversarial network spectrum prediction model. The predicted spectrum can be directly used as the spectrum reconstruction result, or the predicted spectrum can be conventionally processed to obtain the spectrum reconstruction result. Again, no limitation is made.
[0045] In an embodiment of the present application, the training of the generative adversarial network spectrum prediction model includes:
[0046] A curve model is used to simulate and generate simulated spectrum with different parameter combinations as a simulation data set based on characteristics of molecular vibration spectrum, and an actual spectrum is collected to obtain an actual data set;
[0047] A simulation training data set is constructed based on the simulation data set, and a real acquisition training data set is constructed based on the actual data set;
[0048] A pre-constructed generative adversarial network spectrum prediction model is trained based on the simulation training data set and the real acquisition training data set to obtain a trained generative adversarial network spectrum prediction model.
[0049] The model training data set is constructed by combining the simulation spectrum and the actual spectrum in the above manner, avoids complex operations when constructing the data set using the actual spectrum, and avoids unreasonable data caused by using only the simulation spectrum, thereby improving the accuracy of the training data set on the basis of saving the collection workload.
[0050] The simulation spectrum can be generated by different curve models based on the characteristics of molecular vibration spectrum. Optionally, a large number of original spectra with different numbers, different widths and different intensity spectrum peak combinations can be generated by using Gaussian curve, Lorentz curve and other curve models based on the characteristics of infrared spectrum, Raman spectrum and other molecular vibration spectrum, as simulation spectrum. The actual data set can be obtained by collecting different types of actual spectrum.
[0051] Optionally, the simulation training data set is constructed based on the simulation data set, comprising:
[0052] For each sample spectrum in the simulation data set, the error interferogram corresponding to the sample spectrum is determined based on the spectral intensity of the sample spectrum, the silicon-based chip calibration matrix and the noise superposition parameter, the pseudo-inverse calculation processing is performed on the error interferogram, the pseudo-inverse matrix calculation spectrum corresponding to the sample spectrum is obtained, the high-rank filtering processing is performed on the pseudo-inverse matrix calculation spectrum, the model input spectrum corresponding to the sample spectrum is obtained, and the simulation training data set is constructed based on each sample spectrum and the model input spectrum corresponding to the sample spectrum.
[0053] Taking the MZI array as an example, the silicon-based MZI array is the core hardware of the silicon-based spatial heterodyne FTS for effectively implementing the silicon-based Fourier transform spectrum measurement. In order to construct the linearly increasing ideal sampling OPD sequence required by the basic principle of Fourier transform spectrum measurement, according to the physical property that the ideal sampling OPD sequence is equal to the product of the silicon-based waveguide arm length difference ΔL and the silicon-based waveguide effective refractive index n eff of the MZI, each MZI has a specific waveguide width W and arm length difference ΔL setting, wherein the waveguide width W directly controls the silicon-based waveguide effective refractive index n eff However, the actual sampling OPD sequence of the silicon-based MZI array deviates due to inevitable processing errors, and the processing errors mainly affect the waveguide width W, resulting in deviation of the waveguide effective refractive index n eff At this time, the theoretical method of using Fourier transform to demodulate the interference signal to extract spectrum information cannot be used.
[0054] Therefore, in order to realize the spectrum measurement under the imperfect interference signal, for each sample spectrum in the simulation data set and the collected data set, the silicon-based chip calibration matrix passed through includes different processing errors and temperature fluctuations, which cause the different system response matrix changes A of the "spectrum-interferogram" to form an error interferogram; the superimposed noise (including Gaussian white noise, shot noise, etc.) of each sample spectrum corresponding to the error interferogram has different intensities, forming error interferograms with different signal-to-noise ratios. Specifically, the embodiment is based on the "spectrum-interferogram" system response matrix A of the MZI array, and according to the linear relationship I out =AS in +n, a "spectrum-interferogram" set representing "input spectrum-output interferogram" is constructed, wherein n is noise; in order to further improve the applicability of the spectrum prediction model, when constructing the "input spectrum-output interferogram" set, the input sample spectrum S in has different complexities, specifically composed of combinations of different numbers, different widths, and different intensities of spectral peaks, the number of spectral peaks N P is determined, and then the spectral peak width, intensity, and distribution are randomly set; in order to improve the reliability of the spectrum prediction model, when constructing the "input spectrum-output interferogram" set, the output interferogram I out is superimposed with different forms and different intensities of noise n, and the noise forms include additive noise and multiplicative noise, etc., and the signal-to-noise ratio of the output interferogram I out is adjusted by changing the random noise intensity to simulate different detector noise conditions; the randomness of the spectral complexity and the interferogram noise is conducive to enhancing the applicability of the spectrum prediction model. The "spectrum-interferogram" system response matrix A is obtained by high-resolution narrowband laser scanning measurement.
[0055] After obtaining the error interferogram by adding the noise signal, the interferogram simulation of the actual spectrum containing the noise signal is realized, and in order to further increase the accuracy of model training, the error interferogram needs to be denoised, and the denoised error interferogram is used as the model input spectrum corresponding to the sample spectrum for model training. For example, the model input spectrum corresponding to the sample spectrum can be obtained by singular value decomposition matrix to filter the high-rank noise part, and a "model input spectrum-sample spectrum" data pair is constructed to obtain a simulation training data set.
[0056] Referring to the construction method of the simulation training data set, the construction of the real collection training data set can be performed. Considering that the actual spectrum collected has already contained noise information, only the model input spectrum corresponding to the actual spectrum needs to be obtained by performing pseudo-inverse spectrum calculation and high-rank filtering processing on the actual spectrum collected after the MZI array, and a "model input spectrum-actual spectrum" data pair is constructed to obtain a real collection training data set.
[0057] After obtaining the simulation training data set and the real training data set, the simulation training data set and the real training data set can be directly mixed into a training data set to directly train the pre-constructed generative adversarial network spectrum prediction model to obtain a trained adversarial network spectrum prediction model, or the pre-training-fine-tuning method can be used for training.
[0058] In an embodiment of the present application, the pre-constructed generative adversarial network spectrum prediction model is trained based on the simulation training data set and the real training data set to obtain a trained generative adversarial network spectrum prediction model, which comprises:
[0059] The pre-constructed generative adversarial network spectrum prediction model is pre-trained based on the simulation training data set to obtain a pre-trained generative adversarial network spectrum prediction model;
[0060] The pre-trained generative adversarial network spectrum prediction model is fine-tuned based on the real training data set to obtain a trained generative adversarial network spectrum prediction model.
[0061] The model can be pre-trained based on the simulation training data set to obtain a pre-trained generative adversarial network spectrum prediction model, and then the pre-trained generative adversarial network spectrum prediction model is fine-tuned based on the real training data set to obtain a trained generative adversarial network spectrum prediction model. Through the pre-training-fine-tuning method, the training of the model is more accurate, and the spectrum reconstruction based on the trained model is more accurate.
[0062] Taking the model training based on the simulation training data set as an example, the model training process can comprise:
[0063] The following steps are iteratively performed until an iteration end condition is reached to obtain a trained adversarial network spectrum prediction model:
[0064] For each group of "model input spectrum-sample spectrum" data pairs, the model input spectrum is input into the adversarial network spectrum prediction model to obtain a predicted spectrum, and the model parameters of the generative adversarial network spectrum prediction model are adjusted according to the difference between the predicted spectrum and the sample spectrum, wherein the iteration end condition can be reaching a set iteration number or the model loss function converging.
[0065] In an embodiment, the generative adversarial network spectrum prediction model is constructed based on the combination structure of a GAN generator and a GAN discriminator, and the pre-constructed generative adversarial network spectrum prediction model is pre-trained based on the simulation training data set to obtain a pre-trained generative adversarial network spectrum prediction model, which comprises:
[0066] The model input spectrum is input into a GAN generator to obtain a predicted spectrum, a GAN generator loss function is calculated, the predicted spectrum is distinguished from the sample spectrum through a GAN discriminator, a GAN discriminator loss function is calculated, and model parameters of the adversarial network spectrum prediction model are adjusted.
[0067] The above operation is iteratively performed until the GAN generator and GAN discriminator loss functions converge, and a pre-trained generative adversarial network spectrum prediction model is obtained.
[0068] Specifically, after the model input spectrum enters the combined structure of the GAN generator and discriminator, the predicted spectrum is generated through the "encoding-decoding" convolutional neural network of the GAN generator; the sample spectrum directly participates in the discrimination of the predicted spectrum through the GAN discriminator, and the internal similarity between the predicted spectrum and the sample spectrum is used to supervise the spectrum prediction performance of the GAN generator, and the calculation of the generator and discriminator loss functions is combined to generate a feedback signal to continue training the GAN generator until the maximum iteration number is reached or the loss function converges, and a trained spectrum prediction model is obtained. The above-mentioned method can improve the accuracy and reliability of the predicted spectrum by enhancing the supervision of the model input on the model output.
[0069] In the embodiment of the application, the interference pattern of the to-be-measured spectrum after being set in an array is obtained; pseudo-inverse calculation processing is performed on the interference pattern corresponding to the to-be-measured spectrum based on the "spectrum-interference pattern" system response matrix of the set array to obtain a pseudo-inverse matrix calculation spectrum corresponding to the to-be-measured spectrum; high-rank filtering processing is performed on the pseudo-inverse matrix calculation spectrum to obtain a model input spectrum corresponding to the to-be-measured spectrum; the model input spectrum is input into a pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum, and a spectrum reconstruction result is determined based on the predicted spectrum; wherein the generative adversarial network spectrum prediction model is trained based on a simulation training data set and a real acquisition training data set, the simulation training data set includes spectrum data of different detection noises, different processing errors and different temperature fluctuations, so that fast and accurate silicon-based Fourier transform spectrum measurement can be realized under the condition that the interference signal is not perfect.
[0070] Figure 2 Another flowchart of a silicon-based Fourier transform spectrum reconstruction method based on deep learning provided by the embodiment of the application is provided, and the embodiment provides a preferred embodiment based on the foregoing embodiment.
[0071] Specifically, in combination with Figure 2 In the embodiment, a large number of original spectra of spectrum peak combinations with different numbers, different widths and different intensities are simulated and generated based on the characteristics of molecular vibration spectrum using Gaussian curves and Lorentz curves, which are used for pre-training of the model; a small amount of spectra are actually collected, which are used for subsequent fine-tuning of the model structure to perfect the training.
[0072] Step 1: Based on the combined structure of the generator and discriminator in the generative adversarial network (GAN), an adversarial network spectrum prediction model is established for fast and accurate single spectrum prediction, realizing efficient extraction of spectral information.
[0073] Among them, the adversarial network spectrum prediction model takes the high-rank filtered spectrum calculated by the pseudo-inverse matrix as the input, implements the "spectrum-spectrum" conversion, replaces the usual "interferogram-spectrum" conversion, and according to the inherent similarity between spectra, the model input spectrum can directly supervise the spectrum output by the GAN generator through the GAN discriminator, realizing more accurate and reliable spectrum prediction.
[0074] Step 2: I out = AS in + n, according to the system response matrix A of the designed MZI array, the input spectrum is superimposed with different intensity Gaussian white noise and shot noise n after passing through the silicon-based chip calibration matrix A, and the corresponding output interferogram I out with noise is obtained.
[0075] This step implements "spectrum-spectrum" conversion, and trains the spectrum prediction model based on the "input spectrum-output interferogram" set. The training process of the spectrum prediction model is as follows:
[0076] The pseudo-inverse matrix A + of the "spectrum-interferogram" system response matrix A of the MZI array is used, and according to the linear relationship S Pinv = A + I out , the pseudo-inverse matrix calculation spectrum S Pinv is obtained.
[0077] Through singular value decomposition S Pinv matrix, the high-rank filtered noise part is obtained, and the model input spectrum S init is obtained.
[0078] The model input spectrum S init is input into the spectrum prediction model constructed in step 1. Specifically: the model input spectrum S init enters the combined structure of the GAN generator and discriminator, and is processed through the "encoding-decoding" convolutional neural network of the GAN generator to generate the model output spectrum S out , that is, the predicted spectrum.
[0079] The model input spectrum S init is also directly involved in the discrimination of the model output spectrum S out through the GAN discriminator, and according to the model input spectrum S initand the model output spectrum S out The inherent similarity between the spectra, i.e. init and S out The difference between the spectra, the spectrum prediction performance of the supervised GAN generator, the calculation of the joint generator and discriminator loss function, the generation of the feedback signal, the continuous training of the GAN generator until the maximum number of iterations or the loss function converges, the use of partially collected real spectra for the above steps, the fine-tuning of the model parameters according to the results, and the obtaining of the trained spectrum prediction model GAN fine This step improves the accuracy and reliability of the model's predicted spectrum by enhancing the supervision of the model input on the model output.
[0080] Step 3: input the spectrum S real , i.e. the spectrum to be measured, through the MZI array to obtain the output interferogram I out , obtain the system response matrix A of the MZI array through high-resolution narrow-band laser scanning measurement, and obtain the pseudo-inverse matrix according to the linear relationship S Pinv =A + I out to calculate the spectrum S Pinv ; through singular value decomposition of the S Pinv matrix, filter out the high-rank noise part to obtain the model input spectrum S init ;
[0081] Step 4: input the model input spectrum S init to the GAN fine to obtain the spectrum S pre , i.e. the predicted spectrum.
[0082] The embodiment of the present application reconstructs the Fourier transform spectrum through a deep learning neural network model, rather than performing Fourier transform on the interferogram to obtain the spectrum, which requires the arm length difference of the interferometer to be linearly increased. The present application ignores system bias, which increases the possibility of designing the arm length difference of the interferometer; based on the "spectrum-interferogram" system response matrix of the MZI array in the silicon-based spatial heterodyne Fourier transform spectrometer device, a set of "input spectrum-output interferogram" with different complexity and noise level is constructed, a spectrum prediction model is established and trained, the applicability of the spectrum prediction model is enhanced, the noise resistance of the spectrum prediction model is improved, and the detection limit is further reduced; the simulated spectrum similar to the real spectrum is constructed in advance for pre-training, which saves a lot of time for spectrum acquisition.
[0083] The embodiment of the present application also provides a silicon-based Fourier transform spectrum reconstruction device based on deep learning, Figure 3 The structure diagram of a silicon-based Fourier transform spectrum reconstruction device based on deep learning provided by the embodiment of the present application is shown in Figure 3 , and the device comprises:
[0084] The interference pattern acquisition module 31 is configured to acquire an interference pattern of the to-be-detected spectrum after passing through the set array.
[0085] The model input spectrum determination module 32 is configured to perform pseudo-inverse calculation on the interference pattern corresponding to the to-be-detected spectrum based on a “spectrum-interference pattern” system response matrix of the set array, to obtain a pseudo-inverse matrix calculation spectrum corresponding to the to-be-detected spectrum, and to perform high-rank filtering on the pseudo-inverse matrix calculation spectrum to obtain a model input spectrum corresponding to the to-be-detected spectrum.
[0086] The spectrum prediction module 33 is configured to input the model input spectrum into a pre-trained generative adversarial network spectrum prediction model to obtain a predicted spectrum, and to determine a spectrum reconstruction result based on the predicted spectrum. The generative adversarial network spectrum prediction model is trained based on a simulation training data set and a real acquisition training data set. The simulation training data set includes spectrum data of different detection noises, different processing errors and different temperature fluctuations.
[0087] In a possible implementation of the embodiment of the present application, the set array includes an MZI array, and the model input spectrum determination module 32 is specifically configured to:
[0088] obtain a system response matrix of the MZI array, and determine a pseudo-inverse matrix of the system response matrix;
[0089] determine the pseudo-inverse matrix calculation spectrum corresponding to the to-be-detected spectrum according to a linear relationship between the interference pattern, the pseudo-inverse matrix of the system response matrix and the pseudo-inverse matrix calculation spectrum.
[0090] In a possible implementation of the embodiment of the present application, the model input spectrum determination module 32 is specifically configured to:
[0091] perform singular value decomposition on the pseudo-inverse matrix calculation spectrum, filter out a part with a rank higher than a set threshold, and obtain the model input spectrum.
[0092] In a possible implementation of the embodiment of the present application, in actual spectrum reconstruction, the system response matrix of the MZI array is obtained by wavelength scanning measurement of a high-resolution narrowband laser.
[0093] In a possible implementation of the embodiment of the present application, the device further includes a model training module configured to:
[0094] generate simulated spectrum of different parameter combinations as a simulation data set by using a curve model to simulate based on characteristics of molecular vibration spectrum, and collect actual spectrum to obtain an actual data set;
[0095] construct a simulation training data set based on the simulation data set, and construct a real acquisition training data set based on the actual data set;
[0096] train a pre-constructed generative adversarial network spectrum prediction model based on the simulation training data set and the real acquisition training data set to obtain a trained generative adversarial network spectrum prediction model.
[0097] In a possible implementation of the embodiment of the application, the model training module is specifically configured to:
[0098] For each sample spectrum in the simulation data set, an error interferogram corresponding to the sample spectrum is determined based on the spectral intensity of the sample spectrum, a silicon-based chip calibration matrix and a noise superposition parameter, the error interferogram is subjected to pseudo-inverse calculation processing to obtain a pseudo-inverse matrix calculation spectrum corresponding to the sample spectrum, the pseudo-inverse matrix calculation spectrum is subjected to high-rank filtering processing to obtain a model input spectrum corresponding to the sample spectrum, and a simulation training data set is constructed based on each sample spectrum and the model input spectrum corresponding to the sample spectrum.
[0099] In a possible implementation of the embodiment of the application, the model training module is specifically configured to:
[0100] The pre-constructed generative adversarial network spectrum prediction model is subjected to parameter pre-training based on the simulation training data set to obtain a pre-training generative adversarial network spectrum prediction model.
[0101] The pre-training generative adversarial network spectrum prediction model is subjected to parameter fine-tuning based on the real acquisition training data set to obtain a trained generative adversarial network spectrum prediction model.
[0102] In a possible implementation of the embodiment of the application, the generative adversarial network spectrum prediction model is constructed based on a combination structure of a GAN generator and a GAN discriminator, and the model training module is specifically configured to:
[0103] The model input spectrum is input into the GAN generator to obtain a predicted spectrum, a GAN generator loss function is calculated, the GAN discriminator is used to distinguish the predicted spectrum from the sample spectrum, a GAN discriminator loss function is calculated, and model parameters of the adversarial network spectrum prediction model are adjusted.
[0104] The above operations are iteratively performed until the GAN generator and GAN discriminator loss functions converge, and a pre-training generative adversarial network spectrum prediction model is obtained.
[0105] It should be understood that the above-mentioned apparatus embodiments are only illustrative, and the apparatus of the application can also be implemented in other manners. For example, the division of units / modules in the above-mentioned embodiments is only a logical function division, and another division manner can be used in actual implementation. For example, multiple units / modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.
[0106] An electronic device is provided in the embodiments of the present application, Figure 4 A structural schematic diagram of an electronic device provided in the embodiments of the present application is shown in FIG. 1, Figure 4 As shown in FIG. 1, Figure 4 The electronic device shown in FIG. 1 includes a processor 61 and a memory 62. The processor 61 and the memory 62 are connected, for example, through a bus 63. Optionally, the electronic device can further include a transceiver 64. It should be noted that the transceiver 64 is not limited to one in actual application, and the structure of the electronic device does not constitute a limitation on the embodiments of the present application.
[0107] The processor 61 can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can implement or execute various exemplary logical blocks, modules and circuits described in combination with the disclosure of the present application. The processor 61 can also be a combination of computing functions, such as one or more microprocessor combinations, combinations of DSP and microprocessor, etc.
[0108] The bus 63 can include a path for transmitting information between the above-mentioned components. The bus 63 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 63 can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 4 In the present application, only one thick line is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.
[0109] The memory 62 can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, a magnetic disk storage or other magnetic storage devices, or any other medium capable of storing desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.
[0110] The memory 62 is configured to store application program codes for implementing the solutions of the present application, and the processor 61 is configured to control the execution. The processor 61 is configured to execute the application program codes stored in the memory 62 to implement the content shown in the foregoing method embodiments.
[0111] The present application also provides a computer readable storage medium, which can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes, and specifically, the computer readable storage medium stores program instructions, and the program instructions are used to implement the method for reconstructing a silicon-based Fourier transform spectrum based on deep learning in each of the foregoing embodiments.
[0112] The present application also provides a computer program product, which includes a computer program, and the computer program is executed by a processor to implement the technical solutions of the foregoing method embodiments, and the implementation principles and technical effects are similar, and thus will not be described herein.
[0113] In the foregoing embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments. Each technical feature of the foregoing embodiments can be combined arbitrarily, and in order to make the description concise, each technical feature of the foregoing embodiments is not described in all possible combinations, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application
[0114] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.
[0115] It is to be understood that the application is not limited to the precise construction herein disclosed and shown in the drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the claims that follow.
Claims
1. A silicon-based Fourier transform spectral reconstruction method based on deep learning, characterized in that, The method includes: Obtain the interferogram of the spectrum to be measured after setting the array; Based on the "spectrum-interferogram" system response matrix of the set array, the interferogram corresponding to the spectrum to be measured is processed by pseudo-inverse calculation to obtain the pseudo-inverse matrix calculated spectrum corresponding to the spectrum to be measured. The pseudo-inverse matrix calculated spectrum is then processed by high-rank filtering to obtain the model input spectrum corresponding to the spectrum to be measured. The input spectrum of the model is fed into a pre-trained generative adversarial network spectral prediction model to obtain the predicted spectrum, and the spectral reconstruction result is determined based on the predicted spectrum; The generative adversarial network spectral prediction model is trained based on a simulation training dataset and a real-world training dataset. The simulation training dataset includes interferogram data with different detection noise, different processing errors, and different temperature fluctuation interference.
2. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 1, characterized in that, The defined array includes an MZI array. The "spectrum-interferogram" system response matrix based on the defined array performs pseudo-inverse calculation on the interferogram corresponding to the spectrum under test to obtain the pseudo-inverse matrix calculated spectrum corresponding to the spectrum under test, including: Obtain the system response matrix of the MZI array, and determine the pseudo-inverse matrix of the system response matrix; Based on the linear relationship between the interferogram, the pseudo-inverse matrix of the system response matrix, and the spectrum calculated by the pseudo-inverse matrix, the pseudo-inverse matrix-calculated spectrum corresponding to the spectrum to be measured is determined.
3. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 1, characterized in that, The step of performing high-rank filtering on the spectrum calculated from the pseudo-inverse matrix to obtain the model input spectrum corresponding to the spectrum to be measured includes: The pseudo-inverse matrix is used to calculate the spectrum. Singular value decomposition is performed on the spectrum to filter out the part with a rank higher than a set threshold, thus obtaining the model input spectrum.
4. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 2, characterized in that, In actual spectral reconstruction, the system response matrix of the MZI array is obtained by wavelength scanning measurement using a high-resolution narrowband laser.
5. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 1, characterized in that, The training of the adversarial network spectral prediction model includes: Based on the characteristics of molecular vibrational spectra, a curve model is used to simulate and generate simulated spectra with different parameter combinations as a simulation dataset, and actual spectra are collected to obtain the actual dataset. A simulation training dataset is constructed based on the simulated dataset, and a real-world training dataset is constructed based on the actual dataset. The pre-built generative adversarial network spectral prediction model is trained based on the simulation training dataset and the actual sampling training dataset to obtain the trained generative adversarial network spectral prediction model.
6. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 5, characterized in that, The construction of the simulation training dataset based on the simulated dataset includes: For each sample spectrum in the simulation dataset, an error interferogram corresponding to the sample spectrum is determined based on the spectral intensity of the sample spectrum, the silicon-based chip calibration matrix, and the noise superposition parameters. The error interferogram is then processed by pseudo-inverse calculation to obtain the pseudo-inverse matrix calculation spectrum corresponding to the sample spectrum. The pseudo-inverse matrix calculation spectrum is then processed by high-rank filtering to obtain the model input spectrum corresponding to the sample spectrum. A simulation training dataset is constructed based on each sample spectrum and the model input spectrum corresponding to the sample spectrum.
7. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 6, characterized in that, The process of training a pre-built generative adversarial network (GAN) spectral prediction model based on the simulation training dataset and the actual training dataset to obtain the trained GAN spectral prediction model includes: Based on the simulation training dataset, the parameters of the pre-built generative adversarial network spectral prediction model are pre-trained to obtain the pre-trained generative adversarial network spectral prediction model. Based on the actual training dataset, the parameters of the pre-trained generative adversarial network spectral prediction model are fine-tuned to obtain the trained generative adversarial network spectral prediction model.
8. The silicon-based Fourier transform spectral reconstruction method based on deep learning according to claim 7, characterized in that, The generative adversarial network (GAN) spectral prediction model is constructed based on a combined structure of a GAN generator and a GAN discriminator. The pre-training of the pre-constructed GAN spectral prediction model using the simulation training dataset yields a pre-trained GAN spectral prediction model, including: The model input spectrum is fed into the GAN generator to obtain the predicted spectrum. The GAN generator loss function is calculated. The GAN discriminator distinguishes the predicted spectrum from the sample spectrum and calculates the GAN discriminator loss function to adjust the model parameters of the adversarial network spectral prediction model. The above operations are performed iteratively until the loss functions of the GAN generator and GAN discriminator converge, resulting in a pre-trained generative adversarial network spectral prediction model.
9. A silicon-based Fourier transform spectral reconstruction device based on deep learning, characterized in that, include: The interferogram acquisition module is used to acquire the interferogram of the spectrum to be measured after passing through a set array; The model input spectrum determination module is used to perform pseudo-inverse calculation processing on the interferogram corresponding to the spectrum to be measured based on the "spectrum-interferogram" system response matrix of the set array, to obtain the pseudo-inverse matrix calculation spectrum corresponding to the spectrum to be measured, and to perform high-rank filtering processing on the pseudo-inverse matrix calculation spectrum to obtain the model input spectrum corresponding to the spectrum to be measured. The spectral prediction module is used to input the model input spectrum into a pre-trained generative adversarial network spectral prediction model to obtain a predicted spectrum, and to determine the spectral reconstruction result based on the predicted spectrum; wherein, the generative adversarial network spectral prediction model is trained based on a simulation training dataset and a real-world training dataset, and the simulation training dataset includes spectral data with different detection noise, different processing errors and different temperature fluctuations.
10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 8.