CN111721336A - Supervised learning-based classification and identification method for self-interfering microring resonator cavity sensing - Google Patents

Abstract

A self-interference type micro-ring resonator cavity sensing classification and identification method based on supervised learning, comprising the following steps: (1) training data collection; (2) BP neural network sensing data detection model training; (3) test data collection; (4) BP neural network sensor data detection model test: input the test data obtained in step 3, that is, the transmission extinction value within a certain wavelength range, into the trained neural network, and output the two label values of the target to be tested . The invention can realize the identification and classification of 2 ^N -1 different combination components formed by N types of basic substances through the output of the neural network, and obtain the identification and classification results of the target substance to be tested.

Description

Supervised learning-based classification and identification method for self-interfering microring resonator cavity sensing

technical field

The invention relates to a self-interference micro-ring resonant cavity sensing classification and identification method based on supervised learning, and belongs to the field of optical micro-cavity sensing.

Background technique

Whispering gallery mode optical microcavities have ultra-high Q-factors and extremely small mode volumes, and are widely used in active optical devices, optical signal processing, optical interconnects, low-energy nonlinear optics, light-matter interactions, and sensing. Wide range of applications. Whispering gallery optical microcavity sensing can greatly enhance the interaction between light field and matter, effectively improve detection sensitivity, and has broad application prospects in sensing fields such as single nanoparticles and biomolecules. In the whispering gallery optical microcavity, the optical field is confined in the resonant microcavity, but some energy still leaks into the environment through the evanescent field. Resonant mode shifts, mode linewidth broadening, and degenerate mode splitting, etc. These interactions originate from the modal field response to changes in the probed material, collectively referred to as the dispersive sensing mechanism. When the nanoparticle is absorbing, its edge scattering causes the energy loss of the resonance, which will lead to the broadening of the mode linewidth of the whispering gallery microcavity. The sensing mechanism formed by this dissipative interaction is called dissipative sensing mechanism.

When the traditional whispering gallery microcavity is used as a sensor, the spectral shift is almost the same at each resonance wavelength mode. Therefore, for traditional whispering gallery microcavity sensing, whether it is a dispersion sensing mechanism or a dissipative sensing mechanism, only the A certain mode enables sensing measurements without the relevant sensing information at other multiple wavelength modes. However, this single-modal sensing measurement obviously cannot classify and identify different target substances. For example, in the identification and classification of biological components, in the measurement of more complex biological components, not only specific measurement values, but also accurate identification and classification are required. Based on the resonant frequency shift of the whispering gallery microcavity and the number of whispering gallery modes in the free spectral region of these biological components, combined with neural networks, the classification of biological components was achieved, and the classification accuracy reached 97.3% on average (Reference 1: E.A.Tcherniavskaia, V.A.Saetchnikov , Application of neural networks for classification of biological compounds from the characteristics of whispering-gallery-mode optical resonance, Journal of Applied Spectroscopy, 2011, 78, pp.457–460, namely E.A.Tcherniavskaia, V.A.Saetchnikov, Whispering Gallery Microcavities Based on Neural Networks Optical Resonance Classification of Biocomponents, Journal of Applied Spectroscopy, 78, 457–460 (2011)). However, this classification accuracy is obviously not general, and it is difficult to widely use in practice.

In a word, in the sensing classification and identification based on the whispering gallery microcavity, the single-mode detection method is still used, resulting in low detection accuracy and generality. Therefore, in order to achieve a general high classification accuracy method, it is necessary to explore new multi-modal sensing classification and recognition methods.

SUMMARY OF THE INVENTION

In order to overcome the deficiencies of the prior art, the purpose of the present invention is to realize a self-interference type micro-ring resonant cavity sensing classification and identification method based on supervised learning, with high classification accuracy.

In order to solve the above-mentioned technical problems, the present invention provides the following technical solutions:

A self-interference type microring resonator cavity sensing classification and identification method based on supervised learning, comprising the following steps:

(1) Training data collection, the process is as follows:

1-1 First, multiple sets of training data are used for artificial neural network sensing data detection model training. Each set of training data consists of a transmission extinction value within a certain wavelength range and two pre-known training label values corresponding to it. It is used as the input and target output value of the BP neural network respectively, and the BP neural network is trained so that the neural network can establish the mapping relationship between the two. spectrum, extract the transmission extinction value of a certain wavelength range from it, and use it and the corresponding two training label values as a set of training data; group training data;

1-2 After collecting enough sets of training data, normalize the collected transmission extinction value within a certain wavelength range and its corresponding two pre-known training label values, and use the processed data set as the final training data set;

(2) BP neural network sensor data detection model training, the process is as follows:

2-1 Use the transmission extinction value within a certain wavelength range after processing in step 1 as input data, and use the two training label values after processing in step 1 as output data;

2-2 Train the BP neural network sensing data detection model, establish and save the mapping relationship between the transmission extinction value within a certain wavelength range and its corresponding two training label values.

(3) Test data collection, the process is as follows:

3-1 Place the entire detection system in the measurement environment to detect the label values of the two target substances to be measured under three different conditions. When the detection system is excited with a tunable laser light source, a photodetector and an oscilloscope are used to collect data within a certain wavelength range. The transmission extinction value of ;

3-2 Normalize the collected transmission extinction value and use it as a test data set;

(4) BP neural network sensor data detection model test: input the test data obtained in step 3, that is, the transmission extinction value within a certain wavelength range, into the trained neural network, and output the two label values of the target to be tested .

Further, the self-interference type microring resonator includes an input waveguide, a microring resonator, an output waveguide and an optical detection arm waveguide. The input waveguide and the output waveguide are respectively coupled with the microring resonator and placed in the microring. On both sides of the resonant cavity, one end of the input waveguide is the light source access end of the entire optical sensor. At the coupling between the input waveguide and the microring resonator, the other end of the input waveguide is connected to the input end of the optical detection arm waveguide. At the coupling point with the microring resonant cavity, the output end of the optical detection arm waveguide is connected to one end of the output waveguide, and the other end of the output waveguide is the sensing signal output end.

Further, two kinds of materials sensitive to the detection target substance are respectively coated on the upper surface of the two sections of the micro-ring waveguide or the two sections of the optical detection arm waveguide, and the output light of the tunable laser is incident from one end of the input waveguide, and the micro-ring resonator occurs. Coupling, part of the light is coupled into the micro-ring resonator, and the other part exits from the other end of the input waveguide and enters the output waveguide through the optical detection arm. This part of the light is coupled into the micro-resonator again due to the coupling between the output waveguide and the micro-resonator. , and part of the light in this part interferes with part of the light coupled out of the microring resonator and exits from the other end of the output waveguide. Due to the change of the two target substances to be measured, the sensitivity of the corresponding two target substances to be measured will be changed. In the typical transmission spectrum of the self-interference microring resonator, due to the sensitive materials of each target substance to be measured Different from the interaction of the target substance to be measured, the refractive index of the microring waveguide or the detection arm waveguide changes differently at each wavelength, and then the transmission extinction at different resonance wavelengths changes differently; The resonant cavity light sensor can extract the change of effective sensing information (transmission extinction at multiple resonance wavelengths) within a certain wavelength range, and establish an artificial neural network sensing data detection model to realize identification and classification.

Further, the artificial neural network is supervised learning and needs to be trained before classification and recognition. If two substances and their combinations need to be recognized and classified, the training data involved in the training process is divided into three cases: (1) When both substances are present, that is their combination, the training labels are recorded as 1 and 1, respectively; (2) When only the first substance is present, the training labels are recorded as 1 and 0, respectively; (3) Only the first When the two substances are present, the training labels are recorded as 0 and 1, respectively. In each of the above cases, the input of the effective sensing information (transmission extinction values at different resonance wavelengths) of the known training labels is used for training, and the trained neural network is saved after training in all three cases; The sensing information (transmission extinction value at different resonance wavelengths) is tested, and finally the identification and classification result of the target substance to be tested is obtained through the neural network output (label value).

The beneficial effects of the present invention are: the self-interference type micro-ring resonant cavity sensing classification and identification method based on supervised learning involved in the present invention is to coat the N-segment ring waveguide or the two-segment sensitive materials respectively on the corresponding two detection target substance sensitive materials. On the upper surface of the waveguide of the optical detection arm, by detecting its outgoing spectrum, the effective sensing information in a certain wavelength range (transmission extinction at multiple resonance wavelengths) is extracted, and an artificial neural network sensing data detection model is established to realize the identification and classification; The neural network is supervised learning, it needs to be trained before classification and recognition, and then the effective sensing information of the test data (transmission extinction value at different resonance wavelengths) is tested, and finally the neural network output (label value) can be obtained. The identification and classification of 2 ^N -1 different combination components formed by N basic substances is realized, and the identification and classification results of the target substance to be tested are obtained.

Description of drawings

Figure 1 is a schematic diagram of the structure of a self-interference type microring resonant cavity sensor;

Figure 2. Typical emission spectrum of self-interference microring resonator sensor in the wavelength range of 1400-1600 nm;

Figure 3. Typical output spectrum of self-interference microring resonator sensor in the wavelength range of 1520-1540 nm;

Fig. 4 Sensor classification and recognition model based on BP neural network;

Figure 5. Test results of multiple test data containing three components via neural network.

Detailed ways

The present invention will be described in further detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

1 to 5 , a method for classifying and identifying self-interference microring resonator cavity sensing based on supervised learning includes the following steps:

(1) Training data collection, the process is as follows:

1-1 First, multiple sets of training data are used for artificial neural network sensing data detection model training. Each set of training data consists of a transmission extinction value within a certain wavelength range and two pre-known training label values corresponding to it. It is used as the input and target output value of the BP neural network respectively, and the BP neural network is trained so that the neural network can establish the mapping relationship between the two. spectrum, extract the transmission extinction value in a certain wavelength range, and use it and the corresponding two training label values as a set of training data; similarly, by changing the target substance to be measured (such as the concentration of gas, etc.), the corresponding certain wavelength is collected The transmission extinction value within the range can obtain multiple sets of training data;

1-2 After collecting enough sets of training data, normalize the collected transmission extinction value within a certain wavelength range and its corresponding two pre-known training label values, and use the processed data set as the final training data set;

(2) BP neural network sensor data detection model training, the process is as follows:

2-1 Use the transmission extinction value within a certain wavelength range after processing in step 1 as input data, and use the two training label values after processing in step 1 as output data;

2-2 Train the BP neural network sensor classification and recognition model, establish and save the mapping relationship between the transmission extinction value within a certain wavelength range and its corresponding two training label values.

(3) Test data collection, the process is as follows:

3-1 Place the entire detection system in the measurement environment to detect the label values of the two target substances to be measured under three different conditions. When the detection system is excited with a tunable laser light source, a photodetector and an oscilloscope are used to collect data within a certain wavelength range. The transmission extinction value of ;

3-2 Normalize the collected transmission extinction value and use it as a test data set;

(4) BP neural network sensor data detection model test: input the test data obtained in step 3, that is, the transmission extinction value within a certain wavelength range, into the trained neural network, and output the two label values of the target to be tested ;

In the step (1), in the embodiment of the present invention, the feasibility of the sensor classification and identification method based on supervised learning is proved from the perspective of theoretical simulation. When the target to be tested consists of three different components, the useful sensing information (transmission extinction value) on the corresponding transmission spectrum is obtained from the theoretical simulation as training data. The sensor information (transmission extinction value) is output, and the label values of the concentrations of the two gases to be measured are output and compared with their initial predetermined results.

The assumed models of the effective refractive index of the waveguide as a function of wavelength are Lorentzian and Gaussian, respectively. The corresponding spectrum is obtained by theoretical simulation with two gas concentrations C ₁ and C ₂ , and useful sensing information is extracted, that is, the transmission extinction value in a certain spectrum. As shown in Figure 2, the emission spectra for the wavelength range from 1400-1600 nm are shown for three combinations of the two gas components. FIG. 3 further shows the emission spectrum of FIG. 2 in the wavelength range from 1520 nm to 1540 nm. It can be seen that for three different gas components, the transmission extinction changes at different wavelength modes are completely different.

Figure 4 shows the recognition model of sensor classification based on BP neural network. It is divided into input layer, hidden layer and output layer. The second layer is the hidden layer. If the number of nodes in the hidden layer is too low, the learning process will not converge. On the contrary, if the number of hidden layer nodes is too high, it can reduce the network error and improve the accuracy, but it also complicates the network, thereby increasing the training time of the network and the tendency of "overfitting". In the embodiment of the present invention, a sensor data measurement model based on a BP neural network is constructed, and the number of hidden layer nodes is the number of hidden layer nodes obtained by adopting an empirical formula. The third layer is the output layer. The number of nodes in the output layer is 2, and the two output results are the two target gas label values (L ₁ , L ₂ ). The processed training data is used to train the neural network. Specifically, the normalized transmission extinction value in the training data is used as the input, and the corresponding two known target gas label values are used as the two outputs, so that the neural network can learn and save the two values. mapping rules between them.

In the step (1), three components are used for training: (1) the values of the two gas concentrations C ₁ and C ₂ are [0.000500869:0.00001999:0.000980629], the initial value is 0.000500869, and the maximum value is is 0.000980629, the value interval is 0.00001999, the label value (L ₁ , L ₂ )=(1,1), and the number of training groups is 625. In order to obtain a complete training set, when the gas concentration C ₁ changes from 0.000500869 to 0.000980629, The gas concentration C ₂ is temporarily unchanged until the first gas concentration completes a change cycle; (2) the gas concentration C ₁ takes the value [0.000501376:0.00000099:0.000995386], the initial value is 0.000501376, the maximum value is 0.000995386, and the The value interval is 0.00000099, at this time the gas concentration C ₂ is 0, the label value (L ₁ , L ₂ )=(1,0), and the number of training groups is 500 groups; (3) When the gas concentration is C ₁ =0, the gas The value of concentration C ₂ is [0.000501376:0.00000099:0.000995386], the initial value is 0.000501376, the maximum value is 0.000995386, the value interval is 0.00000099, (L ₁ ,L ₂ )=(0,1), the number of training groups for 500 groups. These training data are obtained by substituting the preset values of the two gas concentrations C ₁ and C ₂ into the theoretical expression of the above-mentioned transmission coefficient to obtain the corresponding transmission spectrum, thereby obtaining the corresponding transmission extinction value (in all simulations) In the spectrum, the transmission extinction threshold of 0.90 is selected, that is, only transmission extinction values below this threshold are used as valid sensing data). The transmission extinction value is normalized as the input of the neural network, and the corresponding two gas label values are used as the output prediction value to train the neural network.

In the steps (3) and (4), the transmission extinction values corresponding to the combination of the three target substances to be tested are used for testing, as shown in FIG. 5 . (1) The gas concentrations C ₁ and C ₂ are [0.0005:0.00002:0.0009], the initial value is 0.0005, the maximum value is 0.0009, and the value interval is 0.00002. At this time, the two output labels tested by the neural network The value (L ₁ , L ₂ )=(1,1), indicating that both gases exist in this case; (2) the gas concentration C ₁ takes the value [0.0005:0.0001:0.001], and its initial value is 0.0005, The maximum value is 0.001, the value interval is 0.0001, and the gas concentration C ₂ is 0. At this time, the two output label values (L ₁ , L ₂ )=(1,0) tested by the neural network indicate that this In this case, the first gas exists; (3) the gas concentration C ₂ takes the value of 0, the gas concentration C ₂ takes the value [0.0005:0.0001:0.001], the initial value is 0.0005, the maximum value is 0.001, and the value interval is 0.0001, and the two output label values (L ₁ , L ₂ )=(0,1) tested by the neural network at this time indicate that the second gas exists in this case. The above test data is obtained by the following method, that is, by substituting the two gas concentrations C ₁ and C ₂ into the above theoretical model, the corresponding transmission spectrum can be obtained, and the corresponding transmission extinction value can be obtained therefrom. The transmission extinction value is normalized to obtain the corresponding test data respectively. The simulation results confirm that the test results in the embodiments of the present invention can effectively identify and classify the three different components, and have excellent prediction performance.

When there are three or more basic components of the target substance to be tested, it is only necessary to increase the corresponding label value. For example, when there are 3 basic components of the target substance to be tested, the total component types are 7, so the corresponding output label value takes 7 values from 001 to 111.

In this embodiment, Table 1 shows the parameters of the self-interference microring resonator based on silicon on insulating material (SOI) material;

Table 1

In the embodiment of the present invention, the self-interference type microring resonator structure is shown in FIG. 1 , which is different from the ordinary two parallel waveguide coupling microring resonating cavity structure. This structure couples the upper and lower two parallel waveguides in the microring structure. The coupling regions are connected with additional probe waveguide arms. Its output is essentially the result of the interference of the inner microring resonator and the feedback loop microring resonator. The small phase change of the micro-ring waveguide or the detection arm waveguide caused by the detection material will drastically change the external coupling coefficient between the micro-ring and the whole "U-shaped" detection arm waveguide, thereby changing the outgoing spectrum.

The designed self-interference type microring resonator sensor chip is composed of a silicon-based (SOI) waveguide based on an insulating material, the substrate material is SiO ₂ , and its parameters are set as shown in Table 1. The effective refractive index of the waveguide is n _eff =2.45, The radius of the microring is R=30μm, the initial length of the "U"-shaped probe arm waveguide L _W =250μm, the directional coupling energy coupling coefficient k=0.5, and the waveguide loss coefficient α=0.1dB/cm.

The structure of the designed self-coherent microring resonator sensor chip is shown in Figure 1. For the self-interference microring resonator structure, two sensitive materials are respectively coated on the surfaces of the two sections of the microring waveguide, and they have relatively high detection sensitivity to the two sensing detection targets respectively. The sensor only exposes the microring to the detection target, while the detection arm waveguide does not have any contact with the detection target. Therefore, according to the dissipative sensing principle of the self-coherent microring resonator, the typical transmission spectrum of the outgoing spectrum can still be obtained. Two gas-sensitive materials were placed on the surfaces of the left and right microring waveguides respectively. Due to the interaction between the gas molecules and the corresponding sensitive materials, the refractive index change Δn _c (λ, C) of the cladding of the microring waveguide will be caused, which is related to the wavelength λ and the gas concentration C. When the concentration of a gas to be measured is C, the effective refractive index change Δn _eff (λ, C) of a certain wavelength λ mode is expressed as,

Δn _eff (λ,C)＝S _waveguide Δn _c (λ,C)

Among them, S _waveguide is the sensitivity of the dielectric waveguide, and Δn _c (λ, C) is the refractive index change of the gas sensing material cladding of a certain wavelength λ mode when the concentration of a certain gas is C. The dielectric waveguide sensitivity is related to the waveguide structure and material, as well as the cladding refractive index change,

If the gas can be absorbed by the gas-sensitive material, Δn _c (λ, C) can be expressed as,

Δn _c (λ,C)=F(λ,λ ₀ )ε(λ ₀ )ν _t S _b pC

where λ ₀ is the center wavelength of the absorption band, F(λ,λ ₀ ) is the proportionality coefficient between absorption and refractive index change, ν _t is the total number of gas-sensitive material molecules per unit volume of polymer, S _b is the reagent and Analyte binding constant, p is the polymer permeability constant.

As shown in Figure 1, the two sections of the micro-ring waveguide are coated with different gas-sensitive materials, and the interaction between the two gases and their corresponding sensitive materials is completely different. The effect of the waveguide effective index of refraction is quite different. Assuming that these two waveguide effective refractive index effects have two different analytical models, one is Gaussian and the other is Lorentzian,

Here C ₁ and C ₂ represent the concentrations of the two target gases to be measured, and the parameter values in the above two equations are carefully set so that the magnitude of the refractive index change is consistent with that of the waveguide in actual gas detection.

Substituting the above-mentioned change in the refractive index of the waveguide into the expression of the transmission coefficient of the self-interfering microring resonator can be obtained,

For simplicity, the waveguide and the microring are assumed to have the same propagation constant β _W = β _R = (2π/λ)n _eff - iα, where the initial probe arm waveguide refractive index n _L = n _R = n _eff , and α is the waveguide loss coefficient. Here β _R1 =(2π/λ)(n _eff +Δn _eff1 )-iα, β _R2 =(2π/λ)(n _eff +Δn _eff2 )-iα represent the two-segment waveguides coated with two gas-sensitive materials, respectively Waveguide propagation constant after interaction with two gas molecules. The required simulation data can be obtained by substituting the parameters into the above equation. The corresponding spectrum is obtained by theoretical simulation with two gas concentrations C ₁ and C ₂ , and useful sensing information is extracted, that is, the transmission extinction value in a certain spectrum. As shown in Figure 2, the emission spectra for the wavelength range from 1400-1600 nm are shown for three combinations of the two gas components. FIG. 3 further shows the emission spectrum of FIG. 2 in the wavelength range from 1520 nm to 1540 nm. It can be seen that for three different gas components, the transmission extinction changes at different wavelength modes are completely different. On this basis, a neural network sensor classification recognition model is constructed to process sensor data and identify classification sensor detection targets, as shown in Figure 4. The neural network is trained with the extracted transmission extinction value as the input and the predicted label values of the two target gases as the expected output. Randomly find the test set corresponding to the same label value outside the training set, that is, the useful sensing information corresponding to the two label values, and use the trained neural network for testing. The test results are shown in Figure 5. The theoretical original data are in good agreement with the detection results obtained by the neural network, which shows the feasibility of the detection, identification and classification method.

In the embodiment of the present invention, the parameters of the neural network are shown in Table 2:

Table 2

A kind of self-interference type micro-ring resonant cavity sensing classification and identification method based on supervised learning discussed in the present invention has been introduced in detail above. To limit, any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principle of the present invention should be equivalent replacement methods, which are included within the protection scope of the present invention.

Claims (4)

1. A supervised learning-based self-interference micro-ring resonant cavity sensing classification identification method is characterized by comprising the following steps:

(1) training data acquisition, the process is as follows:

1-1, firstly, adopting a plurality of groups of training data for training an artificial neural network sensing data detection model, wherein each group of training data consists of a transmission extinction value in a certain wavelength range and two pre-known training label values corresponding to the transmission extinction value, and respectively serving as an input value and a target output value of a BP neural network to train the BP neural network so as to establish a mapping relation between the transmission extinction value and the target output value of the BP neural network; by changing a target substance to be measured and collecting a transmission extinction value in a corresponding certain wavelength range, a plurality of groups of training data can be obtained;

1-2, after collecting enough groups of training data, carrying out normalization processing on the collected transmission extinction value in a certain wavelength range and two training label values which are known in advance and correspond to the transmission extinction value, and taking a processed data set as a final training data set;

(2) the BP neural network sensing data detection model training comprises the following steps:

2-1, using the transmission extinction value within a certain wavelength range processed in the step 1 as input data, and using the two training label values processed in the step 1 as output data;

2-2 training a BP neural network sensing data detection model, and establishing and storing a mapping relation between a transmission extinction value in a certain wavelength range and two corresponding training label values;

(3) test data acquisition, the process is as follows:

3-1, placing the whole detection system in a measurement environment, detecting the label values of two target substances to be detected under three different conditions, and collecting transmission extinction values within a certain wavelength range by using a photoelectric detector and an oscilloscope when the detection system is excited by using an adjustable laser light source;

3-2, carrying out normalization processing on the acquired transmission extinction value, and taking the transmission extinction value as a test data set;

(4) BP neural network sensing data detection model test: and (3) inputting the test data obtained in the step (3), namely the transmission extinction value within a certain wavelength range, into the trained neural network, and outputting two label values of the target to be tested.

2. The supervised learning based self-interference micro-ring resonator sensing classification and identification method of claim 1, wherein the self-interference micro-ring resonator comprises an input waveguide, a micro-ring resonator, an output waveguide and an optical detection arm waveguide, the input waveguide and the output waveguide are respectively coupled with the micro-ring resonator and are disposed at two sides of the micro-ring resonator, one end of the input waveguide is a light source access end of the whole optical sensor, the other end of the input waveguide is connected with an input end of the optical detection arm waveguide at a coupling position of the input waveguide and the micro-ring resonator, the output end of the optical detection arm waveguide is connected with one end of the output waveguide at a coupling position of the output waveguide and the micro-ring resonator, and the other end of the output waveguide is a sensing signal exit end.

3. The supervised learning-based self-interference type micro-ring resonant cavity sensing classification and identification method as recited in claim 1 or 2, wherein two materials sensitive to the target substance to be detected are respectively coated on the upper surfaces of two sections of micro-ring waveguides or two sections of optical detection arm waveguides, output light of the tunable laser enters from one end of the input waveguide and is coupled with the micro-ring resonant cavity, then a part of the output light is coupled into the micro-ring resonant cavity, and another part of the output light exits from the other end of the input waveguide and enters the output waveguide through the optical detection arm, the part of the light is coupled into the micro-resonant cavity again due to the coupling effect between the output waveguide and the micro-resonant cavity, and then exits from the other end of the output waveguide after the part of the light in the part of the light interferes with a part of the light coupled out of the micro-ring resonant cavity, and the refractive indexes of the two materials sensitive to-object substances to, thereby causing the refractive index of the two sections of micro-ring waveguides or the two sections of detection arm waveguides to be changed, and further causing the change of the emergent frequency spectrum; in a typical transmission spectrum of a self-interference micro-ring resonant cavity, due to different interaction between each target substance sensitive material to be detected and the target substance to be detected, the refractive index of a micro-ring waveguide or a detection arm waveguide is changed differently at each wavelength, and further, the transmission extinction at different resonance wavelengths is changed differently; therefore, for the self-interference micro-ring resonant cavity optical sensor, the change of effective sensing information is extracted within a certain wavelength range, and an artificial neural network sensing data detection model is established to realize identification and classification.

4. The supervised learning-based self-interference micro-ring resonator sensing classification and identification method as recited in claim 1 or 2, wherein the artificial neural network is supervised learning, and needs to be trained before classification and identification, and if two substances and their combination need to be identified and classified, the training data involved in the training process is divided into three cases: (1) when both substances are present, i.e. the combination of them, the training labels are respectively marked as 1 and 1; (2) when only the first substance is present, the training labels are respectively marked as 1 and 0; (3) when only the second substance is present, the training labels are respectively noted as 0 and 1; in each case, effective sensing information input of a known training label is used for training, and the trained neural network is stored after the training in all three cases; and then testing the effective sensing information of the test data, and finally outputting through a neural network to obtain the identification classification result of the target substance to be tested.

Images