CN111289477A

CN111289477A - SPR photon tongue sensing array-based DOM component detection method

Info

Publication number: CN111289477A
Application number: CN202010137078.3A
Authority: CN
Inventors: 付丽辉; 戴峻峰
Original assignee: Huaiyin Institute of Technology
Current assignee: Nanjing Qianhe Internet Of Things Technology Co ltd
Priority date: 2020-03-02
Filing date: 2020-03-02
Publication date: 2020-06-16
Anticipated expiration: 2040-03-02
Also published as: CN111289477B

Abstract

The invention relates to the technical field of environmental protection, and discloses a DOM component detection method based on an SPR photon tongue sensing array, which comprises the steps of S1 water sample preparation, and determination of DOM types and relative contents in the water sample; s2, constructing an SPR photon tongue sensing array; s3, introducing a test circuit to realize the measurement of characteristic resonance wavelength, spectral width and light intensity of the SPR of the water sample; s4, carrying out ICPSO-BP network training, thereby obtaining network structure parameters meeting error requirements; s5 determines DOM components using the trained ICPSO-BP network. Compared with the prior art, the invention is based on the electronic tongue principle, combines with an SPR sensing array to carry out photonic tongue design, realizes measurement of DOM components in a larger dynamic range through the sensing array, and constructs an integrated system of a plurality of classifiers by utilizing a BP network (ICPSO-BP) optimized by an improved particle swarm optimization algorithm, thereby improving the generalization capability of the model and obtaining better target identification precision.

Description

SPR photon tongue sensing array-based DOM component detection method

Technical Field

The invention relates to the technical field of environmental protection, in particular to a DOM component detection method based on an SPR photon tongue sensing array.

Background

Eutrophication of water caused by excessive concentrations of soluble Organic Matter (DOM) is one of the important factors. The excessive DOM in the water body not only releases toxin to cause water quality deterioration through the processes of explosive growth, death, degradation and the like of algae, but also influences the self-cleaning capability of the water body through a series of complex and variable biochemical processes under the influence of parameters such as biological and chemical oxygen demand of the water body and accelerates the deterioration of the water quality, thereby causing serious threats to aquaculture, industrial and agricultural production and drinking water safety and directly endangering the life quality and the body health of related people. Therefore, the influence of the DOM has a great effect on the water quality, and the identification of the total amount and the components of the DOM is prominent in various index systems for representing the water quality.

Generally, in the research of environmental science and environmental protection, identification of DOM components is always a main hotspot of lake science research, identification result data has very important effects on pollutant migration and transformation behaviors, toxicity, risk evaluation and the like in water, and accurate measurement of different components of DOM in water is an important premise for realizing environmental evaluation and protection.

However, although the common detection method for DOM can perform qualitative analysis on DOM and realize determination of the total DOM, the determination of DOM components is always a bottleneck. In addition to the defect that a large-scale indoor test instrument is needed, the chemical detection method also has the intensity masking effect based on fluorescence, so that fluorescent signals of similar components in DOM cannot be distinguished, and the application of the fluorescent signals in DOM component measurement is limited, and the biological detection method has obvious insufficient effect on quantitative analysis of DOM components due to the complex test process. Meanwhile, the sources influencing the distribution of DOM components are numerous, and water bodies from different sources greatly influence the content, the properties and the like of the components, so that the components are extremely complex and have different characteristics.

In order to identify the DOM components with different characteristics, the single response sensor cannot be used to determine the proportion of each component, and the sensors with different sensing characteristics respectively respond to the components one by one and analyze the components to obtain the specific information of the components.

Disclosure of Invention

The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides a DOM component detection method based on an SPR photonic tongue sensor array, which is characterized in that the sensor array is constructed by using nonspecific selectivity and cross sensitivity of optical fiber SPR sensing, an integrated system of a plurality of classifiers is constructed on the basis of a BP network (ICPSO-BP) optimized by an improved particle swarm optimization algorithm, and the characteristics of DOM components of a sample are specifically analyzed and measured by comprehensively training the resonance wavelength, the spectrum width and the light intensity.

The technical scheme is as follows: the invention provides a DOM component detection method based on an SPR photon tongue sensing array, which comprises the following steps:

s1: preparing a water sample, namely determining the DOM type and the relative content in the water sample by using three-dimensional fluorescence spectrum measurement to obtain data for verification;

s2: constructing an SPR photonic tongue sensing array: the method comprises the steps of forming an SPR sensing probe array with different refractive index measurement values by plating gold films with different thicknesses on a plurality of multimode optical fibers, and enabling the refractive index measurement values of all the sensing probes to be distributed in the range of 1.33-1.43 RIU required by the design of a photonic tongue array;

s3: introducing the SPR photon tongue sensing array in the S2 into a test circuit to realize the characteristic resonance wavelength, the spectral width and the light intensity measurement of the SPR of the water sample and obtain data for training;

s4: the method comprises the steps of constructing a multi-classifier integrated system by utilizing an ICPSO optimized BP neural network, taking training data of wavelength, spectral width and light intensity of SPR response of each water sample measured by an SPR sensing probe in S3 as the input of each ICPSO network classifier, taking DOM component data determined by analysis of S1 as network output, and carrying out ICPSO-BP network training to obtain network structure parameters meeting error requirements;

s5: and inputting the test data of each SPR sensing probe into the neural network successfully trained so as to obtain respective DOM component output, and finally determining the DOM components according to the output result of the classifier.

Furthermore, the test circuit comprises a plurality of SPR sensing probes, a Y coupler, a broadband light source and a closed-loop detection system, wherein the plurality of SPR sensing probes share the Y coupler, the broadband light source is coupled into a light guide optical fiber L0 and then is connected into the Y coupler, the other branch of the Y coupler is connected with the SPR sensing probes through a light guide optical fiber L1, a detected signal spectrum sensed by the SPR sensing probes is reflected to the Y coupler through the light guide optical fiber L1 and then is transmitted to the closed-loop detection system through the light guide optical fiber L2, the spectrum signal is identified by the closed-loop detection system and converted into an electric signal, and finally the electric signal is transmitted to the computer through a cable L3.

Furthermore, the thickness range of the gold-plated film of the multimode optical fiber is 55-85 nm.

Further, the multi-classifier integrated system constructed by the ICPSO optimized BP neural network comprises three primary classifiers, the ANN network structural design of each classifier is the same, the ICPSO improved particle swarm algorithm is used for optimizing BP network weight and threshold, and the specific implementation process is as follows:

1) constructing and initializing a three-layer BP neural network, and designing an input layer, an output layer, a hidden layer and a fitness function;

2) initializing a particle swarm: determining an acceleration factor c₁,c₂Inertia factor omega, number of particles, number of iterations k₁ICPSO subgroup number r, subgroup number step length n, random number r₁,r₂And a particle dimension;

3) calculating the individual optimal value pi by using the fitness function_tGlobal optimum value pg_tSubgroup optimal value pr_t；

4) Updating the current speed and position of the particles;

5) updating the optimal value: comparing the current optimum with the individual optimum pi according to a fitness function_tGlobal optimum value pg_tSubgroup optimal value pr_tIf the current optimal value is superior to any one of the parameters, the current optimal value is replaced;

6) examination end conditions: if the iteration number k is larger than the maximum iteration numberNumber k₁Or the error value for evaluation is larger than a given value, stopping iteration and turning 7), otherwise, turning 4);

7) saving the set of global optimum values;

8) checking ICPSO subgroup end condition: if the subgroup number meets the maximum subgroup number, turning to 9), otherwise, after adding n to the subgroup number, turning to 4), wherein n is the subgroup number adjustment step length;

9) and comparing the global optimal particle positions output by each group according to the fitness function, and mapping the global optimal positions of the group with the optimal positions into a weight and a threshold of the neural network.

Further, the fitness function is:

wherein n is₃Is the number of output nodes; l is the number of training samples; t is t_kOutputting a target, namely a true value of the refractive index of the measured medium; y is_kAnd (5) training actual output for the BP network.

Has the advantages that:

1. the sensor array of the photonic tongue structure can realize high-sensitivity measurement in a large range, and the sensor array can ensure that the sensor array has integral sensitive response to the refractive index and the change information thereof in a large refractive index range required by the design function through cross sensitivity among the sensing probes.

2. Compared with a single sensing probe, the invention utilizes the excellent optical transmission characteristic and the advantages of an integrated optical device in optical fiber sensing, selects SPR sensing heads with different structures to form a sensing array, and each sensing head has certain sensitive nonlinear response to the refractive index outside the dynamic range except for the point measurement performance with good linearity in the dynamic range. The effective arrangement of the optimal refractive index interval of each sensing head in the sensing array can ensure that the measurement performance of a certain sensing head is better matched with the refractive index to be measured in a larger variation range of the refractive index to be measured, so that the refractive index measurement with good linearity is carried out.

3. The invention utilizes an improved particle swarm optimization BP network (ICPSO-BP) to construct an integrated system of a plurality of classifiers, and realizes the refractive index measurement based on SPR effect and the tests of tyrosine proteins, tryptophan proteins, fulvic acid, soluble microorganism metabolites and humic acid of DOM components of the tested water body by comprehensively training the resonance wavelength, the spectral width and the light intensity. The method for completing the same prediction task by utilizing a plurality of classifiers can fully utilize the difference between classification results of single classifiers and apply the selection or combination of the results of the multiple classifiers to obtain a result better than that of a single classifier, thereby improving the generalization capability of the model and obtaining better target identification precision.

Drawings

FIG. 1 is a schematic three-dimensional fluorescence spectrum of four water samples according to the present invention;

FIG. 2 is a schematic diagram of the test structure of the DOM component of the SPR photonic tongue of the present invention;

FIG. 3 is a schematic structural diagram of the principle of a closed-loop processing SPR optical fiber sensing probe;

FIG. 4 is a schematic diagram of fiber SPR sensing based closed loop detection system;

FIG. 5 is a graph of particle motion for the ICPSO algorithm;

FIG. 6 is a flow chart of an ICPSO optimized BP neural network training algorithm;

FIG. 7 is a schematic diagram of a photonic tongue classification system based on the ICSPO-BP neural network and SPR effect;

FIG. 8 is a schematic diagram of the open-loop sensing classification process based on ICPSO-BP and SPR effects;

FIG. 9 shows the results of an SPR sensor array based outbound river water sample (A) test;

FIG. 10 is the results of the Hongze lake (B) test based on SPR sensing arrays;

FIG. 11 shows the results of testing park landscape lake (C) based on SPR sensor arrays;

FIG. 12 is the results of campus landscape lake (D) testing based on SPR sensing arrays.

Detailed Description

The present invention will be described in detail with reference to the accompanying drawings.

From the aspect of water quality measurement, DOM types needing attention mainly comprise tyrosine proteins, tryptophan proteins, fulvic acids, soluble microbial metabolites, humic acids and the like, and different from commercial artificially synthesized organic matters, DOM components in the water body are not standardized industrial products.

In this regard, the present invention is arranged as follows: in order to ensure that the designed photon tongue has universality on a measuring object, four water bodies in different natural environments are selected as measured objects, specifically, water bodies in overseas river water, flood lake water, landscape lake water in the city area of Huaian city and artificial lake water in campus are used as water body representatives with different pollution degrees. Filtering various collected water samples by using a pre-burnt 0.22 mu m filter membrane to filter out undissolved granular substances, and dividing the obtained filtrate into two parts: one part is directly stored in a brown glass bottle for later use at 4 ℃, and the other part is placed in an open brown glass bottle and concentrated by natural evaporation at room temperature in a clean and lightproof room. And the light is protected from light and the room temperature is used for avoiding the biochemical reaction in the concentration process from damaging the relative content of each DOM in the sample. The concentrated sample was filtered again through a filter membrane (0.22 μm after firing) and the filtrate was stored in a brown glass bottle at 4 ℃ for further use. The low-temperature and brown glass bottle can effectively prevent photosensitive biochemical reaction in the sample storage process and avoid changing the original DOM component and relative content. The four concentrated water samples are grouped by A, B, C, D labels, for each group of water samples, 50 water samples with different concentrations are prepared according to the fact that each concentrated solution is a pure solute, deionized water is used as a solvent, the concentration step length is 2%, the water samples are numbered as A1-50, B1-50, C1-50 and D1-50 and serve as training group samples, and corresponding measurement data are used for network training of ANN. And preparing 34 water samples with different concentrations by 3% concentration step, wherein the water samples are respectively numbered as a1-34, b1-34, c1-34 and d1-34 and are used as verification group samples to verify the ANN training result.

The invention constructs three classifier integrated systems-ICPSO-BP (wave length), ICPSO-BP (spectral width), ICPSO-BP (light intensity) and ICPSO-BP (light intensity) by utilizing a BP neural network trained by ICPSO through experimental steps of water sample preparation of DOM, water body DOM component determination, measurement of water sample SPR effect based on a sensor array, training and verification of an artificial intelligent network and the like, realizes comprehensive training of resonance wavelength, spectrum width and light intensity of SPR effect of the measured water body, thereby completing effective prediction of five components P1.n, P2.n, P3.n, P4.n, P5.n (tyrosine protein, tryptophan protein, fulvic acid, soluble microbial metabolite and humic acid) and concentration of the four water bodies of an external canal (A), a flood lake (B), a park landscape lake (C) and a campus lake (DOM) and effectively predicting the concentration of the four water bodies of the flood lake (B) to reach the highest concentration of the P2.n component and the highest concentration prediction rate of the P2.n, thereby verifying the feasibility of photonic tongues based on fiber SPR sensing effects. Meanwhile, the influence elements of the photonic tongue are examined. In the aspect of responding to the number of the parameters/classifiers, compared with a double-parameter/double-classifier and a single-parameter/single-classifier, the ICPSO-BP of the three-parameter/three-classifier has the best prediction effect, and the prediction accuracy of the P2.n component and the concentration of the DOM of the comprehensive water body can reach 95%; in the aspect of influence of neural network structural parameters, when the number of hidden layer nodes of the neural network is 15, the sensing array has the best test effect, the particle number and the group number of the ICPSO structural parameters have corresponding optimal values according to different test objects, and in addition, the correct prediction rate of the ICPSO-BP neural network on DOM components and concentrations of different water bodies is the highest relative to a BP neural network, a Radial Basis Function (RBF) neural network and a PSO-BP neural network.

The invention uses multimode fiber and plates gold films with different thicknesses to form an SPR photon tongue sensing array sensing probe. The optimal refractive index measured value of each sensing probe is effectively distributed in the range of 1.33-1.43 RIU required by the design of the photonic tongue by controlling the thickness of the plated gold film. In the embodiment, a sensing array of the photonic tongue is designed, which is composed of 7 multimode fiber sensing probes (but the invention is not limited to 7 sensing probes, and is only for convenience of explanation and experiments), and the thickness range of the gold film of the 7 multimode fibers is 55-85 nm, and the thickness of the gold film of each multimode fiber is different through the change of sensing characteristics, which are expressed by wavelength, spectral width and light intensity of a resonance point, of the gold film multimode fiber sensing probes with different thicknesses in different refractive index environments to be measured. Meanwhile, an integrated system of a plurality of classifiers is constructed by utilizing a BP network (ICPSO-BP) optimized by an improved particle swarm optimization algorithm, and the refractive index measurement based on the SPR effect and the test process of tyrosine proteins, tryptophan proteins, fulvic acid, soluble microorganism metabolites and humic acid of DOM components of the tested water body are realized through the comprehensive training of resonance wavelength, spectral width and light intensity.

The sensor array composed of the SPR sensing probe is introduced into a test circuit, and the specific test circuit is constructed as shown in FIG. 2.

In fig. 2, seven SPR sensing probes share a Y-coupler, a broadband light source, a closed-loop detection system, and computer analysis software (i.e., a computer), and are manually switched to access the fiber probes through the left side, and each fiber probe is of a terminal reflection type. The broadband light source is coupled to the light guide optical fiber L0 and then is connected to the Y coupler, the other branch of the Y coupler is connected to the optical fiber SPR sensor (SPR sensing probe) through the light guide optical fiber L1, the measured signal spectrum sensed by the sensor is reflected to the Y coupler through the same path L1 by using the terminal mirror effect of the optical fiber SPR, and is transmitted to the closed-loop detection system through the L2, the spectrum signal is identified by the closed-loop detection system and is converted into an electric signal, and finally the electric signal is transmitted to the computer through the cable L3 for extracting corresponding signal parameters. Wherein, L0, L1 and L2 are all light guide optical fibers, and L3 is a cable. In the test, different sensing probes are switched to the measuring circuit, so that the DOM component spectrum test data of the measured water body is obtained, and the corresponding component information is obtained. In the test process, after data is collected every time, the SPR sensing probe needs to be separated from a measured medium, and the SPR sensing probe is kept stand in the air for about 5 minutes until a resonance peak disappears, so that the next measurement can be carried out.

The closed loop detection system is as follows:

a feedback channel is introduced, the locking of the change of the SPR resonance absorption peak caused by the change of the refractive index of the medium 2 serving as an object to be detected in the medium 1/gold film/medium 2 structure is realized by digital signal processing, and the measurement of the change of the refractive index of the medium 2 is realized by the central wavelength tuning of the detection light output by the detection light source LD and the change of the intensity of a frequency modulation signal.

Given a medium 1/gold film in the sensing probe, the refractive index of the medium 2 as the object to be measured and its change will change the SPR resonance absorption wavelength. Because the refractive indexes of the medium 2 and the system calibration medium are different, in the initial state of the system, a certain deviation exists between the central wavelength of the detection light and the resonance wavelength of the sensing probe, the deviation enables the system to obtain an output quantity which is not zero, the output quantity is used for driving and tuning laser output through a step wave generating circuit to be used as the central wavelength of the detection light of the sensing system, namely the central wavelength of the excitation light in the SPR effect, the detection light is subjected to digital narrow-band frequency modulation through EOM, chopping is used for outputting double beams of detection light and reference compensation light, the detection light passes through a sensing area and is subjected to SPR effect resonance absorption, the compensation light and the detection light act on PD together after being unabsorbed through a reference light path, the PD output light current is subjected to RF demodulation, pre-amplification and analog-to-digital conversion (ADC), and then is subjected to digital demodulation with a square wave having the same frequency as a chopping signal, so as a direct current signal capable of reflecting the deviation, this signal is applied to a Digital Integrator (DI) and the output of the DI is used as a drive signal for the tunable light source, so that the wavelength of the detected light is relocked to the wavelength of the absorption peak of the medium in the case of a defined absorption medium. Thus, after system calibration based on the determination medium, the resonant absorption wavelength of the absorption medium can be determined through accumulation of the above-mentioned probe light wavelength adjustment process quantities. In turn, the refractive index of the medium 2 and its change can be obtained by inversion calculation based on the SPR effect through the detection result of the deviation and change of the resonance wavelength.

Based on the process, before the system is used, the pure water is taken as a sample to calibrate the system, so that the central wavelength of the detection light source is locked at the absorption central wavelength determined by the refractive index value of the detection light source. When pure water is replaced by other samples to be detected, because the refractive index of a medium 2 in the medium to be detected, namely the SPR sensing probe, is changed, the corresponding SPR resonance absorption wavelength is also changed based on the principle of SPR effect, when a light source which is originally locked at the refractive index of the pure water and corresponds to the absorption peak wavelength acts on the SPR sensing probe, because the central wavelength of the detection light is deviated from the SPR absorption central wavelength, the system takes the differential spectrum value at the modulation frequency as a scale factor, outputs direct current light current which is in direct proportion to the wavelength deviation amount, and takes the current value as a feedback output control variable to tune the central wavelength of the detection light output by the LD so as to gradually approach to a new SPR absorption center until the two are consistent and locked.

In the process of closing the wavelength of the center of the light source to the resonance absorption center of the SPR sensing probe, the photocurrent output condition corresponding to the wavelength deviation between the light source and the SPR sensing probe reflects the initial deviation degree of the light source and the SPR sensing probe, and the deviation degree corresponds to the refractive index difference between the medium to be measured and the purified water, so that the closed-loop measurement of the refractive index of the medium can be realized.

Therefore, when the refractive index of the medium to be measured changes, optical frequency measurement of a tracking signal is not needed, only incremental signals used for tuning the LD center wavelength in the locking process are needed to be accumulated and measured, the SPR resonance center wavelength can be determined due to the change of the refractive index to be measured, and the refractive index measurement is further realized.

In the practical implementation process of the embodiment, the following closed-loop detection structure as shown in fig. 4 is adopted:

the experimental process is completed by using an HL-2000 broadband light source (the working wavelength is 260-2000 nm) and a USB4000 spectrometer. Then, a tunable pulse light source with a pulse repetition frequency of 5kHz is formed by using a fiber laser (E15 PZTSPM of the company Denmark NKT, the line width of which is 1KHZ), a fiber laser source driver (SVR 200-3 of the company PI of Germany) and a fiber EOM (MPX-LN-0.5 of the company Photoline France); the method comprises the following steps that a Cyclone system Field Programmable Gate Array (FPGA) chip is used as a main body, and functional circuit units such as a chopping signal, an electro-optical modulation signal, a phase discriminator, a filter, a PI controller and an oscillator which are driven by step waves are realized; two X-type optical fiber couplers are utilized to form a detection and reference compensation optical path, a corresponding closed-loop optical fiber SPR sensing system is designed and manufactured, the refractive index of the liquid to be detected is measured, a light source control signal is output, the output center wavelength of a light source is tuned, and therefore closed-loop detection of the SPR sensing signal is achieved.

In fig. 4, the sensing signal processing circuit uses the photoelectric signal processing module in the USB4000 spectrometer to perform the functions including interference, PD detection, RF demodulation, pre-amplification, ADC, etc., and the tunable light source excitation and K-mode excitation are performed_TFThe oscillator completes functions including EOM, RF modulation, chopping modulation and the like, and the phase discriminator mainly completes the chopping demodulation function. In the working process, an output signal generated after the tunable light source excites the SPR sensing head is subjected to photoelectric conversion after interference with reference light through the optical signal processing module of the USB4000, and an electric signal K reflecting the wavelength difference of resonance peaks of the excitation light and the sensing head SPR is output_SPRThe signal is output to a light source tuning control signal K through a phase-locked loop consisting of a phase discriminator, a filter, a PI controller and an oscillator_TFThe excitation light source is tuned, and the oscillation radio frequency signal realized by the FPGA is a digital signal, so a 16-bit DAC conversion circuit (AD9783, the conversion frequency can reach 500M) is adopted to realize conversion, and the tunable light source is controlled.

The PI control algorithm in the experiment is used for effectively realizing wavelength locking, and comprises proportional control and integral control. The former can improve the rapidity of response, while the latter can perform an averaging function to reduce steady-state errors. Let its input signal, i.e. error signal, be u_eThe output signal, i.e. the tuning control signal, being u_cAfter passing through the PI controller, the relationship between the two is as follows:

in the formula, K_p、K_iRespectively, a proportional parameter and an integral parameter of the PI controller.

According to the design function of the photonic tongue and the training data requirement of the artificial neural network, the DOM component and the relative content in the water sample need to be known. The method determines the DOM type and the relative content in the water sample through three-dimensional fluorescence spectrum measurement.

In three-dimensional fluorescence spectroscopy, the fluorescence spectrum is typically divided into five regions in a sequential excitation wavelength/emission wavelength correspondence, as shown in table 1.

TABLE 1 general correspondence between water DOM three-dimensional fluorescence spectra and common fluorescent substances

Specifically, first, a fluorescence spectrophotometer of Hitachi F-7000 type was used to perform three-dimensional fluorescence spectrum measurement of the sample. Corresponding parameters of the spectrophotometer: the emission wavelength (Em) is 280-550 nm, the slit width is 5nm, the excitation wavelength (Ex) is 200-550 nm, the slit width is 5nm, and the scanning speed is 2400 nm/min. In the measurement, blank spectra are subtracted from each water sample to eliminate the influence of Raman scattering, and Rayleigh scattering and upper spectral data are set to be zero in a spectral region with the emission wavelength equal to or 2 times of the excitation wavelength to eliminate the influence of Rayleigh scattering.

Therefore, after the three-dimensional fluorescence spectrum data of any sample water is obtained, the DOM type and relative component condition in the sample can be obtained by using a three-dimensional spectrum area integration method.

For the four water samples selected by the invention, the corresponding three-dimensional fluorescence spectra are shown in fig. 1. According to the conditions of the illustrated excitation wavelength and emission wavelength, the DOM component condition and the fluorescence index in the sample water can be calculated by applying the region integration method in comparison with the region division in the three-dimensional fluorescence spectrum region integration method.

Compared with other neural networks, the basic BP network is more mature in network theory and network performance, but the basic BP algorithm has the following problems: an S-shaped excitation function is adopted, so that the network is easy to enter an S-shaped function saturation area; in order to ensure stability, a small learning rate is often adopted, so that the network convergence speed is low; the error function curved surface is enabled to slide into the valley bottom along the falling direction of the curved surface by the 'uneven convex' and maximum speed falling learning algorithm, so that the network training is easy to fall into local minimum points.

The Swarm intelligence Optimization algorithm represented by a Particle Swarm Optimization (PSO) algorithm is a global Optimization algorithm, and the PSO and the ANN which are globally optimized are combined, so that the neural network generalization capability can be exerted, the ANN convergence speed and the learning capability can be improved, the nonlinear system identification capability is improved, and the method is an effective choice.

The PSO algorithm is simple, small in calculation amount and capable of global Optimization, but the PSO algorithm also has the phenomenon of premature convergence of particles, so that an Improved Cooperative Particle Swarm Optimization (ICPSO) with controllable speed and dynamic information adjustment is provided. The method comprises the steps of dynamically adjusting and controlling iteration times, dividing subgroups and increasing subgroup optimal information, leading global optimal information to be the main in the early stage of optimization, gradually increasing subgroup optimal information when the optimization process is deep, and keeping particle diversity through the combined action of the global optimal information and the subgroup optimal information. In addition, controllable good speed and convergence accuracy can be obtained by utilizing the regulation and control of the number of particles in the subgroup and the number of the subgroups.

1) ICPSO algorithm principle:

the basic PSO algorithm update equation is:

wherein, pg_t，pi_t-global and self-optima for the ith particle of the tth iteration; x is the number of_t+1，v_t+1-mth dimension position and velocity of the ith particle in the t +1 th iteration; c. C₁,c₂-an acceleration factor; ω — the inertia factor; r is₁,r₂—[0，1]And (4) a random number.

The particle states of the modified ICPSO algorithm are shown in fig. 5, and the corresponding particle state updates are determined by the following equation.

x_t+1＝x_t+v_t+1(2)

In the formula, pr_t-subgroup optimal value; t-current optimizationThe number of times; r is the subgroup number; the remaining parameters are as above.

In FIG. 5, the ICPSO algorithm divides all population particles into r groups, and searches the optimal value of the group by each group

And participate in global decision making, thereby increasing the diversity of the particles. The algorithm updates the speed and the position of the particle mainly through four aspects, namely the current speed of the particle

Direction between current position of particle and self-optimum position

Direction between current position of particle and global optimum position

Direction between current position of particle and optimal position of subgroup

In the iteration process, each particle can fully share the individual optimal, global optimal and subgroup optimal information, in order to enable the sharing of the global optimal information and the local optimal information to realize dynamic adjustment, the iteration times t are increased in the updating formula (2), the observation formula (2) shows that each particle mainly shares the global optimal information of the group in the early stage of optimization (the value of t is smaller), and the optimal values in each group are more shared along with the increase of the iteration times (the value of t is larger), and the algorithm can dynamically adjust each piece of shared information, so that the method has better global and local optimization searching capability. In addition, the coefficient 1/r used in the update formula is a convergence rate control factor, and when the number of packets r is [1, ∞ ]]In between, can ensure that 1/r is in [0, 1]]Therefore, the convergence speed more suitable for the convergence precision can be adjusted and achieved through different grouping times.

2) Convergence conditions and speed of the ICPSO algorithm:

theorem 1. the ICPSO algorithm expects sufficient conditions for convergence:

theorem 2, the number of packets r can regulate the convergence speed of the ICPSO algorithm, and the speed is of the basic PSO algorithm

And (4) doubling.

By introducing subgroup optimal information and utilizing the number of times of optimization, the subgroup optimal information and the global optimal information are fused and dynamically regulated, the diversity of the particles is well kept, and premature convergence of the particles is effectively avoided. Meanwhile, the algorithm can have better optimizing speed through control parameters such as the total particle number, the subgroup number and the like, so that the application potential of the algorithm as a neural network learning algorithm is ensured.

The specific implementation process of the ICPSO optimized BP neural network is as follows:

an improved particle swarm algorithm ICPSO is introduced to optimize BP network weight and threshold, and the specific implementation process is as follows:

the first step is to construct and initialize a three-layer BP neural network.

1) Designing the layer number:

the three-layer BP neural network is enough to complete the mapping capability aiming at any nonlinear system, so the number of network layers is designed to be 3, namely an input layer, an output layer and a hidden layer.

2) Input layer node number design

Based on the detection principle of the DOM components of the distributed array type optical fiber SPR sensor to be detected, in the embodiment, 7 sensing probes are adopted, so that in the embodiment, 7 input layer nodes are designed to represent test data of the 7 sensing probes in the photon tongue matrix (but the invention is not limited to the 7 sensing probes), and the spectrum information of the DOM components with cross sensitivity is obtained.

3) Design of hidden node number

Too many hidden layer nodes can cause the reduction of the generalization capability of the network, and overfitting is easy to occur, but the number of the hidden layer nodes is too small, and the network is not easy to converge. In this embodiment, the number of hidden layer nodes is determined to be 15 according to the number of sensing probes and by using a subsequent simulation experiment.

4) Output layer node number design

Based on the test of the SPR sensor on the DOM components, in the embodiment, the DOM components are 5, so that the designed output nodes are 5, which represent the output of each component of the DOM of the detected water body, namely P1.n, P2.n, P3.n, P4.n and P5.n, namely 5 common components (such as tyrosine protein, tryptophan protein, fulvic acid, soluble microbial metabolite and humic acid) of the DOM, but the invention is not limited to the case that the DOM components are 5.

5) Design of fitness function

Designing the mean square error as an ICPSO fitness function, which is specifically defined as follows:

The second step initializes the population of particles.

1) Acceleration factor c₁,c₂The design of (2):

it has been shown in the literature that₁+c₂The best effect is obtained when the total weight is 4, and c is finally determined and taken through simulation₁＝c₂＝2.05。

2) Design of the inertia factor ω:

when ω is [0.9, 1.2], the algorithm can obtain a more ideal effect, and finally ω is determined to be 0.9 through simulation.

3) Number design of particles

Aiming at the optimization grouping problem of the cooperative particle swarm, and finally determining the number N of the particles to be 160 through simulation.

4) Iteration number design

The number of iterations is determined by the simulation: k is a radical of₁＝100。

5)r₁,r₂Design of

And designing random numbers with values of [0, 1] according to empirical data.

6) Design of ICPSO subgroup number r

In this embodiment, the sub-groups may be 2, 4, 6, and 8, respectively, and a group of outputs with the highest precision is used as the network structure weight and the threshold, the sub-group r is initialized to 2, and the influence on the sub-group is further studied.

7) Particle dimension design

Based on a processing target of optimizing BP network weight and threshold by utilizing ICPSO, the particle dimension d is designed to be d ═ p + n₂+q+n₃P is the number of input/hidden layer connection weights; q is the number of connection weights of the hidden layer/output layer; n is₂-number of hidden layer thresholds; n is₃-number of output layer thresholds.

Thirdly, calculating the optimal value of the individual, pi, by using a fitness function_tGlobal optimum value pg_tSubgroup optimal value pr_t。

In the design, the initial position and speed of each particle are initialized to be random numbers, meanwhile, each particle is divided into 2 groups, the connection weight or threshold represented by different dimensions of each particle is substituted into a structural parameter calculation formula of a BP neural network, the output of a hidden layer node and the output layer node of the network are respectively calculated, the mean square error between the actual output and the target output of the network is calculated through a fitness function, and finally, an initialized global optimal point pg is obtained through comparison_tIndividual optimum point pi_tAnd the subgroup optimum point pr_t。

And fourthly, updating the current speed and position of the particles.

Updating the velocity v of each particle_t+1And position x_t+1And (4) information.

The fifth step updates the optimum value.

Comparing the current optimum with the global optimum pg according to a fitness function_tIndividual optimum value pi_tAnd subgroup optimal value pr_tAnd if the current optimal value is better than any one of the parameters, the current optimal value is replaced.

The sixth step checks the end condition.

If the iteration number k is larger than the maximum iteration number k₁Or the error value for evaluation is larger than the given value, the program stops iteration and turns to the seventh step, otherwise, the program turns to a new round of particle state updating (the fourth step).

The seventh step saves the set of global optimum values.

The eighth step checks the end-of-packet condition.

If the number of groups is 8, then go to the ninth step, otherwise, after adding 2 groups to the number of groups, go to the fourth step.

And the ninth step of comparing the positions of the global optimal particles output by each group according to the fitness function, and mapping the global optimal positions of the group with the optimal positions into weight values and threshold values of the neural network.

The ICPSO optimization BP neural network training algorithm flow is shown in figure 6.

In the implementation, sample water with different mixing ratios is subjected to measurement of DOM components in corresponding samples based on a three-dimensional fluorescence spectrum analysis technology, then a closed-loop detection system is utilized to perform measurement of corresponding SPR resonance absorption parameters of each sensing head to obtain data for training, and the data are handed to an ANN for network training according to the flow shown in FIG. 6; and performing similar spectral measurement and three-dimensional fluorescence spectral analysis on other step length points of the measured water body, taking the obtained result as a verification data set, and verifying the ANN training result.

The output parameters of the SPR effect and the relations among the DOM components belong to a highly nonlinear relation, deep learning is required for learning the relation, a large number of samples are required for model training, but the number of trainable samples is usually limited for the actual test process, and in order to fully utilize various information of SPR on the DOM components and fully play the role of photonic tongue, an integrated system of a plurality of classifiers is adopted on the basis of the BP network (ICPSO-BP) optimized by the improved particle swarm optimization, and the refractive index measurement based on the SPR effect and the classification of the DOM components are realized through comprehensive training of resonance wavelength, spectral width and light intensity. The method comprises the following steps:

detecting components of diluted water bodies by using 7 sensing probes, and constructing three primary classifiers, namely ICPSO-BP (wave length), ICPSO-BP (spectral width), ICPSO-BP (light intensity), and ICPSO-BP (light intensity), according to resonance wavelength, spectral width and light intensity data. The ANN network structure design of each classifier is the same: the input layer consists of 7 nodes and represents the detection data of 7 optical fiber SPR sensing probes; the output consists of 5 nodes, representing the proportions of all components of water DOM-P1.n, P2.n, P3.n, P4.n and P5.n, namely 5 common DOM components (tyrosine proteins, tryptophan proteins, fulvic acid, soluble microbial metabolites and humic acid). Based on theoretical analysis and training result feedback, the number of hidden nodes of each classifier is determined to be 15.

The related ANN structure design, training and verification processes are all completed based on a Matlab 6.0 platform. The specific principle is shown in fig. 7.

In the training stage, using 7 SPR sensing probes to measure the wavelength, spectral width and light intensity test parameter data of SPR response of each water sample as the input of each ICPSO network, using DOM component measurement data determined by three-dimensional fluorescence spectrum analysis as the network output, and performing ICPSO-BP network training to obtain the network structure parameters meeting the error requirement; in the testing stage, each SPR sensing test data is input into the neural network which is successfully trained, so that each DOM component output is obtained, and then the DOM components are finally determined by using an averaging method according to the output results of the three classifiers.

During this time, the three classifiers involved implement the same process, as in fig. 8.

From fig. 8, the training sample set is processed in two ways. Firstly, true value determination is carried out and is output as an ICPSO-BP neural network, firstly, spectrum information of a measured object is obtained by using an SPR sensing system and is input as the neural network, a training sample set formed by input and output data trains the neural network, and after the network convergence is finished after the training, a sensing measurement correction model is successfully established; on the other hand, the predicted sample set is also subjected to the above 2 processing steps, and the corresponding resonance wavelength, spectral width and light intensity information are input into respective ICPSO-BP models, so that the corresponding output of the network is obtained, and the network feasibility is verified by comparing the corresponding output with the true value.

In a cross-sensitive photonic tongue structure, the complexity of a measured object causes large redundancy of SPR spectrum information acquired by each sensing head, and the method for completing the same prediction task by using a plurality of classifiers can fully utilize the difference between classification results of single classifiers and apply selection or combination of the results of the multiple classifiers to obtain a result better than that of a single classifier, thereby improving the generalization capability of a model and obtaining better target identification precision.

According to the analysis, the intelligent sensing system with the similar structure such as the photonic tongue can be structurally divided into two parts, namely a sensing array and an intelligent algorithm. The sensing system is based on cross sensitivity and used for obtaining comprehensive response information of each sensing head to an object to be detected and changes of the object to be detected, namely linear response of each sensing head in a linear region and nonlinear response based on cross sensitivity among the sensing heads, and the sensing system is used for realizing linear fitting of the nonlinear response through an artificial intelligence learning training algorithm and realizing comprehensive linear response of the sensing system to the object to be detected and the changes of the object to be detected. The performance of the intelligent sensing system is determined by the two, and the influence and the determination of the two on the performance of the sensing system have different characteristics. The function of the sensor array depends on the full coverage of the sensor array to the variation range of the object to be measured and the measurement sensitivity under the full coverage condition, and the function of the sensor array mainly depends on the network structure and the training algorithm. As mentioned above, the BP artificial neural network based on the ICPSO algorithm is determined according to the characteristics of a detection object, the effectiveness of the BP artificial neural network is verified in more applications, but the design of a sensor array is mainly based on the fact that the thickness of a gold film is regulated and controlled at sensitive points with different refractive indexes, and the effectiveness of the BP artificial neural network needs to be further verified. For this reason, in the first stage, the sensing array structure change is arranged to affect the experiment.

In the overall prediction experiment process, firstly, in the embodiment, under the test condition of 7 sensing probes, the first 30 actual measurement prediction data of the water body A are selected to carry out classification test on the five components and the concentration of the water body A. The main parameters for ICPSO-BP network training were set to (7 sensing probes case): the mean square error is a fitness function; the training algorithm is ICPSO; acceleration coefficient of c₁＝c₂2.05; the inertia factor is omega-0.9; the number of particles is N160; number of initial iterations k₁＝100；r₁,r₂Is between [0,1]A random number in between; the subgroup number is 2, 4, 6 and 8 groups, but is initialized to 2 groups; the number of the classifiers is set to be 3, and the three classifiers respectively correspond to the sensing parameter wavelengths, the spectrum widths and the resonance light intensities of the three SPR sensing heads; the final classification is done by averaging. The ICPSO networks have the same structure, namely the number of input neurons is 7; the number of hidden layer neurons is 15; the number of output neurons is 5.

After training was completed, the data from the SPR test on sample a at a concentration step of 3% was used as validation data. The statistics of the predicted values and reference values (determined by three-dimensional fluorescence spectroscopy) obtained by the ICPSO neural network are shown in table 2.

TABLE 2 prediction of the DOM components and concentrations in water A by ICPSO neural network

As can be seen from the above graph, the predicted data fitness of the SPR photonic tongue array based on ICPSO for p1.n is 0.9674, which is calculated to have a predicted set Root Mean Square Error (RMSE) of 0.980258, a fitness of 0.9933 for p2.n, a RMSE of 0.8680, a fitness of 0.847 for p3.n, a RMSE of 1.247, a fitness of 0.9822 for p4.n, a RMSE of 0.940, a fitness of 0.3547 for p5.n, and a RMSE of 1.994. The combination of the fitting degree and the root mean square error can indicate that the ICPSO has the best prediction precision and correlation on the P2.n, the prediction effect on the P1.n is equivalent to that on the P4.n, and the prediction effect on the P5.n is the worst. In general, the ICPSO has strong modeling and fitting capability and is suitable for the classification and concentration prediction of most DOM components of the A water body.

In order to verify the influence of different combinations of the sensing probes in the array on the prediction effect, further examine the generalization capability of the photonic tongue array and the ICPSO-ANN, select more water bodies to complete further prediction research, and divide the determined 7 sensing probes into three situations for training and prediction, namely the situation of using an odd array of sensing probes, which is marked as F1 ═ {1, 3, 5, 7 }; the case with an even set of sensing probes, denoted F2 ═ 2, 4, 6; the case of using an all-sensing probe, denoted as F3 ═ {1, 2, 3, 4, 5, 6, 7 }. Network training and verification are respectively carried out on four different tested water bodies, namely A-outward river water, B-Hongze lake water, C-park landscape lake water and D-campus landscape lake water. The input nodes of the ICPSO network are respectively 3, 4 or 7 according to specific conditions, and other settings are the same as the above.

For convenience of describing the prediction effect, the prediction accuracy parameter is defined as follows:

wherein, the number of correct prediction samples is the number of samples under the condition that the prediction relative error is less than 1%.

Thereby obtaining the prediction effect of the ICPSO-BP neural network corresponding to the array structure under three different conditions, as shown in figures 9 to 12.

Observe fig. 9-12:

1) for different sensor probe arrays F1 ═ 1, 3, 5, 7, F2 ═ 2, 4, 6, F3 ═ 1, 2, 3, 4, 5, 6, 7, the prediction effects for the concentrations of DOM components in the water bodies of four sources are different, but in general, the array combination of F3 ═ 1, 2, 3, 4, 5, 6, 7 has the best prediction effect for the concentrations of DOM components of various water bodies, the highest accuracy is 96%, as found in the test of P2.n in Hongze lake (B), the lowest prediction accuracy is 75%, as found in the P5.n test in outer canal (A), the contents of the corresponding DOM components (P2.n and P5.n) in the two measured water bodies are 48.32% and 5.22%, respectively, as found from the relative content data of the DOM components of four water samples calculated based on the three-dimensional fluorescence spectrum area integration method, that is, the higher the component concentration, the higher the detection rate, but for the combination of F1 ═ 1, 3, 5, 7, {2, 4, 6}, the highest and lowest predicted accuracy of F2 ═ 2, 4, 6, 7} array combination, which did not show such performance, both the accuracy rates were much smaller than those of F3 ═ 1, 2, 3, 4, 5, 6, 7}, where the maximum accuracy of F1 ═ 1, 3, 5, 7} was 89%, and appeared at the p2.n test point in the lake of hong lake (B), and the minimum accuracy was 67%, and appeared at the p5.n test point in the outer canal (a), the maximum accuracy of the array scale F2 ═ 2, 4, 6} was 85%, and appeared at the p2.n test point in the landscape lake of park, the corresponding DOM component was 36.99, the minimum accuracy was 50%, and appeared at the p5.n test point in the outer canal (a), and the test head of the three different combinations showed the best results of the p5.n test, this is strongly related to the concentration of the component being too low, i.e. when the component is too low, the optimization of the sensing probe structure is also difficult to sense small changes. In addition, it can be seen from the figure that F1 ═ {1, 3, 5, 7} and F2 ═ 2, 4, 6}, the DOM component discrimination ability for different concentrations of water is blurred, and the DOM component discrimination effect for similar concentrations of components is equivalent, that is, when the difference between the types and contents of various components is small, the difference between the results obtained after the reaction with the sensor is also relatively close, and the array arrangement of F1 ═ {1, 3, 5, 7}, and F2 ═ {2, 4, 6} is difficult to discriminate.

In general, the DOM test of the F3 ═ 1, 2, 3, 4, 5, 6, 7} sensor array for four water bodies A, B, C, D is much better than that of the array arrangement of F1 ═ 1, 3, 5, 7, and F2 ═ 2, 4, 6, the main reason for the above phenomenon is that F1 ═ 1, 3, 5, 7, F2 ═ 2, 4, 6} sensor probes are limited in number and narrow in the information to be sensed, while F3 ═ 1, 2, 3, 4, 5, 6, 7} sensor array can provide more test information, when the array is implemented, the more sensor probes with different structures induce more different central absorption wavelengths and related parameters for the refractive index of the same object to be tested, when the optimal absorption wavelengths of all the sensor probes cover the refractive index range of the liquid to be tested, the dynamic range of the sensor array matches the target dynamic range of the test system, therefore, compared with a single or a few sensing heads, the multi-sensing array can acquire more refractive index change information of the measured water body, and can effectively extract useful information through an intelligent algorithm, so that high-sensitivity measurement can be realized in a larger dynamic range.

2) Comparing the detection effects of F1 ═ 1, 3, 5, 7}, F2 ═ 2, 4, 6}, the two effects are equivalent to the test of the water C, D, but the effect is very different for the test of the water A, B, and the prediction accuracy of F2 ═ 2, 4, 6} is greatly reduced compared with the case of F1 ═ 1, 3, 5, 7} array because the refractive indexes of the optimal measurement points of three sensor heads No. 2, No. 4 and No. 6 of the sensor array F2 ═ 2, 4, 6} are 1.3442, 1.373, 1.4012 respectively, and the refractive indexes of the diluted water of a and B are distributed among 1.4216-1.3359, 1.4032-1.3343 respectively, so that the sensor heads are not covered in the range of the measured water, and the sensing information amount is too small, and the classifier can also use the information amount which is too small to reduce sharply, thereby resulting in the accuracy rate.

In general, in the case of F3, the predicted effect after network training is the best, the case of F2 is the worst, and the effect in the case of F1 is between the two, which can be attributed to the different choice of the sensing probes in the photonic tongue array. Therefore, the method not only proves the reasonable feasibility of the photonic tongue structure, but also provides guidance for optimizing an intelligent sensing system with a structure and a principle similar to those of the photonic tongue. Specifically, in the two major structural modules of the photonic tongue, the sensing array is a foundation, the change information of the object to be detected is obtained through response of three variables representing the SPR effect, namely resonance wavelength, spectral width and light intensity, to the object to be detected and the change of the object to be detected, and each sensing head in the sensing array effectively responds to the change of the object to be detected in the full range on the basis of the cross sensitivity characteristic, so that the information foundation for realizing the function of the photonic tongue is realized. The training network applies the nonlinear fitting ability of the ANN, linear and nonlinear response information is picked up from the sensing probes of the sensing array and is analyzed and processed, the network structure selection and the learning algorithm are designed to fully apply the object information picked up by the sensing array as far as possible, and when the sensing probes are mismatched in design and cannot effectively and comprehensively obtain the corresponding information of the measured object, the intelligent algorithm is also invalid.

Under the condition of F3, the structure of 7 sensor heads completely covers the variation range of the refractive indexes corresponding to four sample waters, the basic data is comprehensive, and the network structure and the algorithm are appropriate, so that the method has better identification effect. For the case of F2, the refractive index measurement ranges corresponding to 3 sensing probes are concentrated in the middle of the corresponding refractive index ranges of four sample waters with different concentrations, and SPR sensing information corresponding to the larger and smaller ranges of the refractive index to be measured is not effectively acquired, so that the same training network cannot satisfy the prediction result due to the lack of comprehensiveness of the basic data. In the case of F1, the sensor array corresponds to the entire range of the refractive index of the sample water, but the density is not sufficient, the cross-sensitive information between the sensor heads is not sufficiently obtained, and the prediction effect after network training is poor.

Thus, regardless of the specific configuration of the intelligent taste system, the design of the sensing array is a prerequisite for the effectiveness of the system. The lack of training network design may make the recognition result not ideal, and the defect in the sensing array design will cause the system to fail completely.

In addition, from the respective researches on the three sensor array situations in fig. 9-12, it can be found that the network prediction effect is better than that in the low-concentration situation under the condition that the three sensor array situations have high concentrations for four different water samples. The reason for this can also be attributed to the information acquisition capabilities of the sensing array. Because, for a given structure of SPR sensing probe and the sensing array formed based on the SPR sensing probe, the response sensitivity of the SPR sensing probe and the sensing array to the refractive index change of the object to be measured is determined. The refractive index value and the gradient of the refractive index value along with the change of the concentration of a water sample with high DOM concentration (content) are also higher than those of the water sample with low DOM concentration, and for the selected sensing heads with different structures, although the sensing heads with different structures are required to be uniformly covered as much as possible in the whole refractive index change of the water sample to be measured, the position of the SPR sensing probe corresponding to the optimal refractive index measuring point is realized by regulating and controlling the thickness of a gold film, compared with the sensing probe with smaller thickness, the spectrum width of the thick film SPR sensing probe is larger, the measuring effect of a single sensing probe is not as good as that of the small film thickness, but the cross sensitivity among the sensing probes is improved due to the wide spectrum width, so that the sensing array and the photonic tongue structure of an intelligent algorithm have better measuring and.

Therefore, when the sensor array is designed and selected, the photonic tongue with good performance and the intelligent sensing system with the similar structure need to cover the variation interval of the object to be measured as completely as possible within the measurement range corresponding to each sensing probe, and also need to have the highest cross sensitivity as possible among different sensing probes.

The sensor array and the photon tongue processed by intelligent data can realize effective prediction of DOM components and concentration based on SPR effect measurement data. Therefore, the photonic tongue design is scientific, reasonable and feasible.

The above embodiments are merely illustrative of the technical concepts and features of the present invention, and the purpose of the embodiments is to enable those skilled in the art to understand the contents of the present invention and implement the present invention, and not to limit the protection scope of the present invention. All equivalent changes and modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims

1. A DOM component detection method based on an SPR photon tongue sensing array is characterized by comprising the following steps:

s2: constructing an SPR photonic tongue sensing array: the method comprises the steps of forming an SPR sensing probe array with different refractive indexes by plating gold films with different thicknesses on a plurality of multimode optical fibers, and enabling the refractive index of each sensing probe to be distributed in the range of 1.33-1.43 RIU required by the design of a photonic tongue array;

2. The method for detecting DOM components based on the SPR photonic tongue sensing array of claim 1, wherein the testing circuit comprises the plurality of SPR sensing probes, a Y coupler, a broadband light source and a closed-loop detection system, the plurality of SPR sensing probes share the Y coupler, the broadband light source and the closed-loop detection system, the broadband light source is coupled to a light-guide optical fiber L0 and then connected to the Y coupler, the other branch of the Y coupler is connected to the SPR sensing probe through a light-guide optical fiber L1, a measured signal spectrum sensed by the SPR sensing probes is reflected to the Y coupler through a light-guide optical fiber L1, sent to the closed-loop detection system through a light-guide optical fiber L2, subjected to spectral signal identification by the closed-loop detection system and converted into an electrical signal, and finally transmitted to a computer through a cable L3.

3. The method for detecting DOM components based on SPR photonic tongue sensing array of claim 1, wherein the thickness range of the gold-plated film of the multimode fiber is 55-85 nm.

4. The method for detecting DOM components based on the SPR photonic tongue sensor array, according to claim 1, wherein the ICPSO optimized BP neural network constructed multi-classifier integrated system comprises three primary classifiers, the ANN network structure design of each classifier is the same, the particle swarm optimization ICPSO is improved to optimize BP network weight and threshold, and the specific implementation process is as follows:

3) using fitness functionCalculating the individual optimum value pi_tGlobal optimum value pg_tSubgroup optimal value pr_t；

4) Updating the current speed and position of the particles;

6) examination end conditions: if the iteration number k is larger than the maximum iteration number k₁Or the error value for evaluation is larger than a given value, stopping iteration and turning 7), otherwise, turning 4);

7) saving the set of global optimum values;

5. The method for detecting DOM components based on SPR photonic tongue sensing array of claim 4, wherein said fitness function is: