A spectral reconstruction method based on arrayed multi-peak valley broadband filter and a miniature spectrometer

Abstract

The application discloses a spectrum reconstruction method based on an arrayed multi-peak valley broadband filter and a miniature spectrometer. A to-be-measured light beam is irradiated on a multi-peak valley broadband filter array after being expanded by an expansion lens and collimated by a collimating lens. A detector measures the intensity value of the transmitted light beam and inputs the intensity value into a computer. The spectrum distribution of the to-be-measured light beam is calculated through a decoding subnetwork of a spectrum inversion network. The application is a brand-new, miniaturized and integrated spectrometer, and is expected to be widely applied in the fields of portable industrial detection and consumer electronics.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis, specifically to a spectral reconstruction method and a miniature spectrometer based on an arrayed multi-peak valley broadband filter. Background Technology

[0002] Spectrometers are commonly used spectral characterization instruments in scientific research and industrial production. They are used to test and calibrate the emission / reflection / transmission / radiation characteristics of light sources or objects, and have wide applications in optoelectronic systems, sensing and detection, food safety, chemical production and other fields.

[0003] Traditional dispersive spectrometers use gratings, prisms, tunable filters, and other optical elements to sequentially decompose a continuous, wide-band spectrum into energy at specific wavelengths, thereby achieving full-spectrum scanning and acquisition. They can provide ultra-low spectral resolution and a wide spectral range. However, they have drawbacks such as large size and slow spectrum acquisition speed.

[0004] With the continuous development of science and technology, the application fields of spectral analysis are rapidly expanding. In some applications, such as consumer electronics, on-chip laboratories, and clinical medicine, the need for smaller physical size, lower cost, and lower power consumption of spectrometers takes precedence over high performance. Miniature spectrometers have emerged to meet this need.

[0005] The use of miniaturized or planar dispersive elements is a widely adopted strategy for spectrometer miniaturization. The main dispersive elements are miniaturized dispersive elements (planar gratings, planar waveguides, photonic crystal waveguides, arrayed waveguide gratings, etc.) or narrowband filters (arrayed narrowband filters, linearly graded filters, tunable filters, etc.). The former's spectral resolution is mutually constrained by the spectrometer size; that is, the distance between the dispersive element and the detector (i.e., the spectrometer size) is proportional to the spectral resolution, thus limiting the spectrometer size to a theoretical limit within a certain spectral range and accuracy requirements. The latter, while not subject to this constraint, requires a large number of arrayed filters (hundreds), involving complex and delicate fabrication processes, resulting in higher production costs.

[0006] In recent years, thanks to advancements in micro-nano fabrication technology and related algorithms (compressed sensing, machine learning, etc.), a novel type of miniature spectrometer has emerged—the computational reconstruction type. Computational reconstruction miniature spectrometers utilize algorithms such as compressed sensing and deep neural networks to invert the incident spectrum from a set of spectral responses encoded by a broadband detector or broadband filter. Theoretically, any optical element capable of producing different broadband spectral responses can serve as an encoder, such as quantum dots, nanowires, thin films, liquid crystals, photonic crystals, and metasurfaces. This method often requires fewer encoding operations than the number of spectral channels to be measured, effectively reducing the number of array blocks needed, simplifying the fabrication process, and lowering the fabrication difficulty. Furthermore, spectral resolution is no longer limited by the size of the spectrometer, making it possible to shrink the spectrometer size to the millimeter or even sub-millimeter scale.

[0007] However, in existing computational reconfiguration-based miniature spectrometer solutions, the design of arrayed coded filters is mainly based on a "random production-selection" method, lacking a quantitative optimization design method for their structure and spectral response, making it difficult to achieve the reproducibility of spectrometer manufacturing.

[0008] For example, Chinese patent document CN109932058A discloses a miniature spectrometer based on an array filter, including a beam expander lens, a collimating lens, an array filter, a detector, and a host computer. Specifically, it uses filter blocks with narrow-band transmission spectra having different center wavelengths and passband half-widths as its spectroscopic elements, and integrates them on a detector chip. The beam to be measured is expanded by the beam expander lens and collimated by the collimating lens before illuminating the array filter. The transmitted light modulated by the array filter is detected, received, and transmitted to the host computer. After processing by the host computer, the spectral distribution of the beam to be measured is obtained. The array filter is fabricated based on a multiple binary lithography separation method, which does not include quantitative optimization design of the spectral response of each filter block of the array filter; the transmission spectrum half-width of the filter block is relatively narrow and does not cover the entire detection band, resulting in insufficient encoding capability for the spectrum to be measured; the specific algorithm for solving the spectral distribution of the beam to be measured is unclear. Summary of the Invention

[0009] This invention provides a spectral reconstruction method based on an arrayed multi-peak-valley broadband filter. This method can use the light intensity value obtained on the detector at one time as input to invert the discrete spectral distribution of the beam under test, with high accuracy and fast inversion speed.

[0010] This invention provides a reconstructed miniature spectrometer based on an arrayed multi-peak broadband filter. This miniature spectrometer avoids the use of spectroscopic elements such as gratings, prisms, and tunable filters, thus reducing the physical size of the device. The solution uses a smaller number of filter blocks, simplifying the fabrication process, and the thin-film structure reduces the fabrication difficulty of the arrayed filter, improving the repeatability and stability of the device fabrication.

[0011] This invention provides the following technical solution:

[0012] A spectral reconstruction method based on an array of multi-peak and valley broadband filters includes: the beam to be measured is sequentially expanded by a beam expander and collimated by a collimating lens before being irradiated by an array of multi-peak and valley broadband filters; the intensity value of the transmitted beam is measured by a detector and input into a computer; and the spectral distribution of the beam to be measured is calculated by the decoding subnetwork of a spectral inversion network.

[0013] A reconstructed miniature spectrometer based on arrayed multi-peak-valley broadband filters includes a beam expander lens, a collimating lens, a multi-peak-valley broadband filter array, a detector, and a computer connected in sequence according to the optical path. During detection, the beam to be measured is expanded by the beam expander lens and collimated by the collimating lens before illuminating the multi-peak-valley broadband filter array. The detector measures the intensity value of the transmitted beam at one time and inputs it into the computer. The spectral distribution of the beam to be measured is calculated by the decoding subnetwork of the spectral inversion network.

[0014] Compared to using a single-channel spectrum with a narrow half-width spectral response, a broadband spectrum covering the entire detection band can improve the encoding capability of the filter block for the spectrum under test, increase the degree of differential response of each band of the spectrum under test, and thus reduce the number of blocks in the arrayed filter. According to compressed sensing theory, the lower the correlation of the spectral response of the filter block, the smaller the encoding redundancy of the spectrum under test, and the more fully the limited number of filter blocks can be utilized.

[0015] In this invention, a multi-peak-valley broadband filter array is integrated on a detector chip.

[0016] Preferably, the spectral inversion network (spectral inversion algorithm) adopts a combination of one or more of the following algorithms: principal component analysis algorithm, genetic algorithm, deep neural network and compressed sensing algorithm.

[0017] Before fabricating the multi-peak-valley broadband filter array, we can introduce a spectral inversion network. Preferably, the target spectral response and the matching decoding sub-network of the multi-peak-valley broadband filter array are both obtained from the constructed spectral inversion network.

[0018] When constructing the spectral inversion network, the known input spectrum is used as the training set and test set; after deep learning with the spectrum of a specific optical thin film structure as a constraint, the spectral inversion network is obtained, that is, the encoding sub-network and the decoding sub-network are obtained; the target spectral response of each filter block in the multi-peak-valley broadband filter array is the connection weight between neurons in the encoding sub-network; according to the target spectral response, the structural parameters of each filter block are obtained by optical thin film design software.

[0019] In the deep learning process, the spectral inversion network of this invention takes the actual spectral response of each block of the multi-peak-valley broadband filter array obtained from experiments as a factor, which can optimize the spectral inversion network more comprehensively, make the decoding sub-network more matched with the filter array, and further improve the spectral inversion accuracy.

[0020] Preferably, during optimization, the known input spectrum is used as the training and test sets (which can be the same as or different from the training and test sets used to construct the spectral inversion network, without restriction), and the actual spectral response is used as the connection weights between neurons in the encoding sub-network, which remain unchanged during deep learning (the bias between neurons is set to 0 during deep learning, and also remains unchanged), and only the decoding sub-network is optimized.

[0021] Preferably, the number of input neurons in the decoding subnetwork is M, which is the number of blocks in the multi-peak-valley broadband filter array, and the number of output neurons is N, which is the number of spectral channels. It also contains several hidden layers. The connections between input neurons and hidden layer neurons, between hidden layer neurons and hidden layer neurons, and between hidden layer neurons and output neurons are fully connected. A batch normalization layer and an activation function are added after each fully connected layer.

[0022] The multi-peak-valley broadband filter array includes several filter blocks, each with a different transmission spectrum (containing different numbers of peaks and valleys), but all of them cover the broadband spectrum of the entire detection band.

[0023] In this invention, the detection band is the ultraviolet-visible-infrared band, and the specific band is selected as needed.

[0024] In this invention, "a plurality of" refers to two or more.

[0025] The filtering block includes a substrate and a filtering film system deposited on the substrate, wherein the filtering film system is selected as a multilayer all-dielectric film stacked with high / low refractive index or a metal-dielectric structure film stack.

[0026] The substrate material is a transparent substrate material, and the ultraviolet-visible-near infrared band is selected from glass or plastic. Preferably, the substrate is K9 glass or PET plastic; the mid- and far-infrared bands are selected from fluorides such as silicon, germanium, zinc sulfide, zinc selenide, and calcium fluoride.

[0027] In the multilayer all-dielectric film, both the high- and low-refractive-index layer materials are transparent or low-absorption materials within the working wavelength range. The high-refractive-index layer material is selected from one or a combination of at least two of titanium dioxide, hafnium dioxide, tantalum pentoxide, silicon nitride, or zinc sulfide in the ultraviolet-visible-near-infrared band. The far-infrared band of the high-refractive-index layer material is selected from one or a combination of at least two of silicon, germanium, zinc sulfide, or zinc selenide. The low-refractive-index layer material is selected from one or a combination of at least two of silicon dioxide, aluminum oxide, or metal fluorides in the ultraviolet-visible-near-infrared band. The far-infrared band of the low-refractive-index layer material is selected from one or a combination of at least two of zinc sulfide, zinc selenide, or metal fluorides. The metal fluoride may be magnesium fluoride.

[0028] Preferably, both sides of the multilayer all-dielectric film are high-refractive-index layers.

[0029] Preferably, the number of layers in the multilayer all-dielectric film is 1-50. More preferably, it has 3-20 layers, and even more preferably, it has 3-10 layers.

[0030] Preferably, the thickness of a single layer of the multilayer all-dielectric film is 5-3000 nm, and the thickness of each layer can be the same or different.

[0031] As a preferred embodiment, the substrate size, number of film layers, and material of each film layer are the same in each filter block of the multi-peak broadband filter array, and the thickness of the other film layers is the same except for the middle layer.

[0032] Preferably, the metal-dielectric structured film stack is a metal / dielectric / metal broadband filter film system from bottom to top.

[0033] Preferably, the metal layer material in the metal / dielectric / metal broadband filter film system is selected from one or an alloy composed of at least two of gold, silver, aluminum, or copper; the dielectric layer material in the ultraviolet-visible-near-infrared band is selected from one or a combination of at least two of silicon dioxide, aluminum oxide, metal fluorides, titanium dioxide, hafnium dioxide, tantalum pentoxide, silicon nitride, or zinc sulfide; and the dielectric layer material in the infrared band is selected from one or a combination of at least two of silicon, germanium, zinc sulfide, zinc selenide, or metal fluorides. The metal fluoride may be magnesium fluoride.

[0034] Preferably, in the metal / dielectric / metal broadband filter film system, the metal layer thickness is 6nm-70nm and the dielectric layer thickness is 15nm-10μm. The metal layer thicknesses can be the same or different.

[0035] In this invention, the detector is selected from photovoltaic detectors, photoconductive detectors, pyroelectric detectors, CCDs, or CMOS sensors. The detector acquires the intensity value of the transmitted beam after modulation by a multi-peak-valley broadband filter array, and the number of intensity values ​​is the same as the number of filter blocks.

[0036] The measurement principle of the miniature spectrometer provided by this invention is as follows: the beam to be measured is expanded and collimated, and then uniformly illuminated onto each filter block of a multi-peak-valley broadband filter array and the detector. The detector reads the light intensity value at once and inputs it into the decoding sub-network to obtain the spectral distribution of the beam to be measured. Each light intensity value is the sum of the product of the light intensity value at a certain wavelength of the spectrum to be measured, the transmittance of the filter block at that wavelength, and the detector responsivity across the entire detection band of the miniature spectrometer. Since the spectral response of each block is determined by its fabrication method and is largely unaffected by the external environment, it is considered a constant and thus known. Similarly, the detector response can also be considered a known constant.

[0037] Assuming the arrayed multi-peak broadband filter consists of M filter blocks, the detector can read M intensity values ​​I. i (i = 1, 2, ..., M). The range of the spectral detector is λ. min To λ max The spectral response of the filter block is T i Given that the detector response is D(λ) and the spectrum to be measured is s(λ), the N light intensity values ​​read by the detector can be expressed as follows:

[0038]

[0039] After discretization, it can be represented as

[0040]

[0041] Where N is the number of spectral channels, i.e., the continuous values ​​of the measured spectrum, the spectral response of the filter block, the detector response, etc., expressed in terms of (λ). max -λ min Discretization of resolution ) / N.

[0042] A system of equations containing M equations can be obtained.

[0043]

[0044] It was realized that the spectrum to be measured was encoded using an arrayed multi-peak-valley broadband filter.

[0045] The decoding subnetwork is part of the spectral inversion network. Its input is the light intensity value collected by the detector, and its output is the reconstructed spectrum of the target image. Available spectral inversion algorithms include principal component analysis, genetic algorithms, deep neural networks, and compressed sensing algorithms.

[0046] Preferably, the spectral inversion algorithm is a deep neural network algorithm. The deep neural network has M input neurons equal to the number of multi-peak broadband filter array blocks and N output neurons equal to the number of spectral channels, and contains several hidden layers. The connections between input neurons and hidden layer neurons, between hidden layer neurons, and between hidden layer neurons and output neurons are fully connected, with a batch normalization layer and an activation function added after each fully connected layer. The activation function can be a sigmoid function, tanh function, ReLU, Leaky ReLU, or softmax function, etc.

[0047] The multi-peak broadband filter array has different spectral responses for each filter block. It can be fabricated using physical vapor deposition methods such as electron beam evaporation, ion beam sputtering, and magnetron sputtering; chemical vapor deposition methods such as surface plasmon chemical vapor deposition and low-pressure chemical vapor deposition; or methods such as ion plating, electroplating, and arc plating. Preferably, a multiple electron beam evaporation method is used (the number of times is the same as the number of filter blocks). That is, according to the designed structural parameters of each filter block (film material and number of layers, film thickness, etc.), electron beam evaporation is used to deposit films on a clean substrate. The multi-peak broadband filter array is divided into two types: splicing and micro / nano fabrication. Micro / nano fabrication mainly uses patterning methods such as ultraviolet lithography, nanoimprinting, and electron beam exposure for arraying operations, with ultraviolet lithography being the preferred method.

[0048] The present invention provides a method for fabricating a splicing type of a reconstructed micro-spectrometer based on an arrayed multi-peak-valley broadband filter, comprising:

[0049] 1) Determine the number of filter blocks required for the array filter to meet the performance requirements of the measurement accuracy; select the substrate material and film material of the filter block according to the working wavelength range of the spectrometer; construct a spectral inversion network based on a deep neural network with the spectrum of a specific optical thin film structure as a constraint, and simultaneously obtain the target spectral response of the filter block; use optical thin film design software to obtain the structural parameters of the filter block.

[0050] 2) Wipe and clean the substrate with ethanol and acetone; place the cleaned substrate in an electron beam evaporation apparatus, control the deposition parameters, and deposit a multilayer film system on the substrate;

[0051] 3) The deposited filter blocks are spliced ​​and combined to obtain a multi-peak-valley broadband filter array;

[0052] 4) Compare the actual spectral response of the prepared filter block with the target spectral response, and use the actual spectral response as the weight of the encoding sub-network to perform micro-training on the spectral inversion network to obtain a spectral decoding sub-network that is more suitable for the actual filter block.

[0053] 5) Construct a collimation and beam-expanding optical system to ensure that the beam to be measured is uniformly illuminated by the array filter; integrate the arrayed multi-peak and valley broadband filter directly above the detector chip so that the detector can simultaneously collect the light intensity values ​​of different filter blocks and input them into the spectral decoding sub-network.

[0054] 6) The reconstructed spectrum is obtained through data calculation and analysis. This allows for the construction of a miniature spectrometer based on an array of multi-peak broadband filters.

[0055] The reconstructed miniature spectrometer based on arrayed multi-peak-valley broadband filters of the present invention has the following advantages compared with other miniature spectrometers:

[0056] (1) An arrayed filter is used as the spectrometer element to avoid using dispersive elements such as gratings, thereby further reducing the physical size of the spectrometer.

[0057] (2) Using broadband filters as filter blocks increases the encoding capability of filter blocks compared to miniaturization methods using narrowband filters, making the number of blocks much smaller than the number of spectral channels and simplifying the processing flow.

[0058] (3) The use of thin film structure filter blocks and electron beam evaporation processing method reduces the processing difficulty.

[0059] (4) Using a deep neural network-based decoding subnetwork, the light intensity value obtained on the detector at one time is used as input to invert the discrete spectral distribution of the beam to be measured, which has high accuracy and fast inversion speed.

[0060] (5) The actual spectral response of the filter block is used as the weight of the coding sub-network to perform micro-training on the spectral inversion network. The resulting decoding sub-network is more compatible with the actual spectral response of the filter block, which can reduce the impact of errors in the filter block preparation process on the spectral reconstruction accuracy.

[0061] This invention is a novel, miniaturized, and integrable spectrometer that is expected to be widely used in portable industrial testing, consumer electronics, and other fields. Attached Figure Description

[0062] Figure 1 A schematic diagram of the reconstructed micro spectrometer based on an arrayed multi-peak-valley broadband filter provided by the present invention;

[0063] Figure 2 This is a distribution diagram of the filter blocks of the arrayed multi-peak broadband filter in the embodiment;

[0064] Figure 3 This is a schematic diagram of the filter block structure in the embodiment;

[0065] Figure 4 This is a schematic diagram of a spectral inversion network based on a deep neural network.

[0066] Figure 5 The target spectral response and actual spectral response of 16 filter blocks in the multi-peak-valley broadband filter array in the embodiment are shown.

[0067] Figure 6 The reconstruction results of different spectra are shown in the example of a reconstructed micro-spectrometer based on an arrayed multi-peak-valley broadband filter.

[0068] Wherein: 1 is the beam to be tested, 2 is the beam expander, 3 is the collimating lens, 4 is the multi-peak-valley broadband filter array, 5 is the detector, 6 is the computer; 41 is the substrate, 42 is the high refractive index layer, 43 is the low refractive index layer, 44 is the high refractive index layer, 45 is the low refractive index layer, and 46 is the high refractive index layer. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0070] like Figure 1 As shown, the spectrometer provided by the present invention includes a beam expander 2, a collimating lens 3, a multi-peak-valley broadband filter array 4, a detector 5, and a computer 6. The beam 1 to be measured is expanded by the beam expander 2 and collimated by the collimating lens 3 before being emitted in parallel. After being modulated by the multi-peak-valley broadband filter array 4, the transmitted light is simultaneously detected by the detector 5, and the light intensity values ​​are obtained at one time in a number equal to the number of blocks. The light intensity values ​​are transmitted to the computer 6 and the spectral decoding sub-network is used to obtain the spectral distribution of the beam to be measured.

[0071] Multi-peak broadband filter array filter block distribution as follows Figure 2 As shown, a rectangular structure is formed by several filter blocks with different structural parameters. Here, 16 filter blocks with different structural parameters form a 4×4 rectangular block, but the number of filter blocks is not limited to 16.

[0072] The structural diagram of the filter block is shown below. Figure 3 As shown, 41 is the substrate, and 42-46 are high / low refractive index film stacks. Here, 42, 44, and 46 are high refractive index films, and 43 and 45 are low refractive index films. The number of layers is not limited to 5. Different filter blocks have the same substrate thickness, but the thickness of the single or multiple films varies, resulting in different broadband spectral responses across the entire detection band. Due to the different structural parameters of each filter block, multiple electron beam evaporation processes are required, the number of which corresponds to the number of filter blocks.

[0073] Before production, a spectral inversion network is pre-constructed. This network is based on a deep neural network and includes an encoding subnetwork and a decoding subnetwork, with the structure as follows: Figure 4As shown, both the spectral input layer and the spectral output layer contain 301 neurons, corresponding to the 301 spectral channels detected by the miniature spectrometer across the entire wavelength range; the intensity input layer contains 16 neurons, corresponding to the 16 intensity values ​​read from the detector; hidden layer 1 and hidden layer 2 each contain 500 neurons. During deep learning, using the spectrum of a specific optical thin film structure as a constraint, the network parameters (weights, biases, etc. between neurons) are continuously updated by comparing the input spectrum (the spectrum used as the training and test sets) with the output spectrum (the reconstructed spectrum corresponding to the input spectrum). This allows the acquisition of the target spectral response of the filter block (the weights between the spectral input and intensity input layer neurons) and the matching spectral decoding subnetwork.

[0074] Based on the target spectral responses of each filter block obtained from deep learning, the structural parameters of the filter blocks can be quickly obtained using optical thin film design software. Then, based on existing thin film deposition processing methods, the filter blocks can be fabricated. The spectral response of the thin film structure filter block is entirely determined by structural parameters such as film material, number of layers, and thickness. These parameters are precisely controllable during the fabrication process of the filter block, thus achieving a high degree of agreement between the actual spectral response of the filter block and the target spectral response.

[0075] In this embodiment, the target spectral response and actual spectral response of 16 filter blocks with different film thicknesses across the entire detection band are as follows: Figure 5 As shown, different filter blocks produce different modulation effects on the test spectrum 1, resulting in different light intensity values ​​on detector 5. The target spectral response is the result of deep learning, while the actual spectral curve is the result of actual fabrication and testing. There is an unavoidable experimental error between the two (originating from the coating process). Therefore, after the 16 filter blocks are fabricated, the constructed spectral inversion network can be optimized based on the actual spectral responses of the filter blocks. That is, the actual spectral response is used as the weight of the encoding sub-network, while remaining unchanged during deep learning. Known spectra are used as the training and testing sets to optimize the constructed spectral inversion network, thus obtaining the final spectral inversion network. In other words, the target spectral response is used as the weight of the encoding sub-network, while remaining unchanged during deep learning. Known spectra are used as the training and testing sets to optimize the constructed spectral inversion network, thereby obtaining the final spectral inversion network. Figure 4 The weights between the neurons in the spectral input and light intensity input layers are set to the actual spectral response and locked, while the biases are set to 0 and locked. This "micro-training" of the decoding subnetwork can make it more compatible with the actual spectral response and improve the accuracy of spectral reconstruction.

[0076] The reconstruction results of different spectra using the miniature spectrometer provided by this invention are as follows: Figure 6 As shown, the results are compared and analyzed with those obtained from commercial spectrometers.

[0077] Example

[0078] The miniature spectrometer provided in this embodiment has a detection range of 400nm-700nm. Assuming a spectral resolution of 1nm, the number of spectral channels is:

[0079]

[0080] Sixteen filter blocks are arranged on an arrayed multi-peak broadband filter. The substrate is K9 glass, and the film system is a Sub / TiO2 / SiO2 / TiO2 / SiO2 / TiO2 / Air broadband filter system. A spectral inversion network is pre-constructed, which is based on a deep neural network and includes an encoding subnetwork and a decoding subnetwork. Its structure is as follows: Figure 4 As shown. Using the transmission spectrum of a 5-layer TiO2 / SiO2 / TiO2 / SiO2 / TiO2 film system as the target spectral response constraint, deep learning yielded the target spectral responses of 16 filter blocks, as shown. Figure 5 As shown in Table 1, the structural parameters of each film layer obtained using optical thin film design software are presented.

[0081] Table 1. Structural parameters of arrayed multi-peak broadband filters

[0082] Number of film layers Membrane materials Physical thickness / nm 1 <![CDATA[TiO2]]> 59.76 2 <![CDATA[SiO2]]> 94.18 3 <![CDATA[TiO2]]> 54.12-608.75 4 <![CDATA[SiO2]]> 94.18 5 <![CDATA[TiO2]]> 59.76

[0083] That is, the physical thickness of the third layer (high refractive index metal oxide TiO2) film is different in each filter block, while the film material and physical thickness of the other layers are the same. The thickness of this layer in each filter block is shown in Table 2.

[0084] Table 2 Physical thickness of the third thin film in the filter block

[0085]

[0086]

[0087] The fabrication process of the arrayed multi-peak valley broadband filter is as follows: clean the substrate; input the corresponding film thickness according to the block number and deposit the film; repeat the deposition 16 times to obtain the required filter blocks; splice the filter blocks to obtain a 4×4 multi-peak valley broadband filter array.

[0088] The target spectral response obtained from spectral reconstruction simulation and the actual spectral response of the filter block are compared in the full-spectral response of the spectrometer detection band as follows: Figure 5 As shown, due to the different film thicknesses of the filter blocks with different numbers, the actual spectral response has high non-correlation, enabling efficient encoding of the spectrum under test without redundancy. The deviation between the target spectral response and the actual spectral response mainly originates from coating errors. The actual spectral response is set to... Figure 4By locking the weights between the spectral input and the intensity input layer and micro-training the spectral inversion network, the decoding sub-network can be better adapted to the actual spectral response, thereby improving the accuracy of spectral reconstruction.

[0089] After passing through the beam expander and collimator optical system, the beam under test can be uniformly illuminated by the arrayed multi-peak broadband filter. After being modulated by 16 filter blocks, the detector chip can simultaneously obtain the light intensity value I corresponding to different blocks. i (i = 1, 2, ..., 16). The readout light intensity can be represented by a system of 16 equations:

[0090]

[0091] After inputting the light intensity value into the computer, the decoding sub-network is used to reconstruct the spectrum to be measured. The reconstruction result is as follows: Figure 6 As shown, the results are compared with those obtained using a commercial spectrometer. The comparison demonstrates that the reconstructed miniature spectrometer based on an arrayed multi-peak-valley broadband filter of this invention has good spectral acquisition capabilities, with results not significantly different from those measured by large commercial spectrometers. Therefore, it can meet the needs of most applications where high spectral accuracy is not required.

[0092] The number of filter blocks can be expanded to 64 or more, which can improve the coding capability and spectral reconstruction accuracy of arrayed multi-peak broadband filters, and also expand the working wavelength range; the number of spectral channels can be further increased, thereby increasing the detail information of the reconstructed spectrum.

[0093] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A spectral reconstruction method based on arrayed multi-peak-valley broadband filters, characterized in that, The beam to be tested is expanded by a beam expander and collimated by a collimating lens before illuminating a multi-peak-valley broadband filter array. The detector measures the intensity value of the transmitted beam and inputs it into the computer. The spectral distribution of the beam to be tested is obtained through the decoding subnetwork of the spectral inversion network. The structural parameters of each filter block of the arrayed multi-peak-valley broadband filter are obtained from the encoded subnetwork of the constructed spectral inversion network; When constructing the spectral inversion network, the known input spectrum is used as the training set and the test set; after deep learning with the spectrum of a specific optical thin film structure as a constraint, the constructed spectral inversion network and the target spectral response of each filter block in the multi-peak-valley broadband filter array are obtained; based on the obtained target spectral response, the structural parameters of each filter block are obtained. After the multi-peak-valley broadband filter array is fabricated, the constructed spectral inversion network is micro-optimized based on its actual spectral response. During micro-optimization, the known input spectrum is used as the training and testing sets, and the actual spectral response is used as the connection weights of the encoding sub-network, which remains unchanged during the deep learning process. Only the decoding sub-network is optimized. The transmission spectra of each filter block in the multi-peak-valley broadband filter array are different, and the number of transmission peaks and valleys, as well as the peak and valley wavelengths or transmittance / reflectance values, are different, but all of them are broadband spectra covering the entire detection band. Each filter block in the multi-peak-valley broadband filter array includes a substrate and a filter film system deposited on the substrate. The filter film system is selected as a multilayer all-dielectric film stacked with high / low refractive index or a metal-dielectric structure film stack. The number of input neurons in the decoding subnetwork is M, which is the number of arrayed multi-peak and valley broadband filter blocks, and the number of output neurons is N, which is the number of spectral channels. It also contains several hidden layers. The connections between input neurons and hidden layer neurons, between hidden layer neurons and hidden layer neurons, and between hidden layer neurons and output neurons are fully connected. A batch normalization layer and an activation function are added after each fully connected layer.

2. The spectral reconstruction method based on arrayed multi-peak-valley broadband filters according to claim 1, characterized in that, The spectral inversion network employs a combination of one or more of the following algorithms: principal component analysis, genetic algorithm, deep neural network, and compressed sensing algorithm.

3. A miniature spectrometer implementing the spectral reconstruction method based on arrayed multi-peak-valley broadband filters as described in any one of claims 1 to 2, characterized in that, The system includes a beam expander, a collimating lens, a multi-peak-valley broadband filter array, a detector, and a computer, which are connected in sequence according to the optical path. The beam to be tested is expanded by the beam expander and collimated by the collimating lens, and then illuminates the multi-peak-valley broadband filter array generated according to the above method. The detector measures the intensity value of the transmitted beam and inputs it into the computer. The spectral distribution of the beam to be tested is calculated by the decoding sub-network of the spectral inversion network.

Images