A non-contact turbidity online measuring device

Abstract

The invention provides a non-contact turbidity online measuring device, which belongs to the technical field of turbidity measurement and comprises: a laser light source, an incident optical fiber, an incident lens, an output optical fiber, a reflecting surface, a collecting lens, a reagent bottle, an optical signal collector and a host computer; the laser light source is used to emit a laser light beam, the reagent bottle is used to contain a suspension, one end of the incident optical fiber is connected to the laser light source, and the other end of the incident optical fiber is connected to the incident lens and is used to receive laser light emitted by the laser light source and irradiate the reagent bottle through the incident lens; the reflecting surface and the incident lens are arranged opposite to each other with the reagent bottle separated, the collecting lens is used to collect laser light irradiated on the inner wall of the collecting lens, and transmit it to the optical signal collector through the output optical fiber, the optical signal collector is used to receive the laser light transmitted by the output optical fiber, convert it into an electrical signal and transmit it to the host computer, a turbidity calculation module is arranged in the host computer and is used to calculate the turbidity of the suspension according to the electrical signal; the problem of large error in the existing method is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of turbidity measurement, and in particular relates to a non-contact turbidity online measuring device. Background Art

[0002] Suspension turbidity is an important physical parameter that is widely used and has important application value in many industrial fields. For example, in the water treatment industry, the turbidity of the suspension reflects the content of suspended particles in the water body and is one of the key indicators of water quality control; in chemical production, turbidity monitoring can promptly detect abnormal conditions in raw materials or intermediate products to ensure product quality; in the food and beverage industry, turbidity is also one of the important parameters for judging product quality. In addition, turbidity monitoring also plays an important role in the pharmaceutical, mining, papermaking and other industries. It can be said that suspension turbidity detection technology has become an indispensable online monitoring method in various fields.

[0003] At present, there are two main methods for turbidity detection: one is the offline sampling detection method, that is, the sample is sent to the laboratory for analysis and testing; the other is the online monitoring method, which uses optical sensors to directly detect the suspension in real time. Although the offline sampling detection method can obtain more accurate turbidity data, it is limited by the sampling and transportation process and cannot reflect the real-time turbidity changes. At the same time, the operation process is cumbersome and not suitable for continuous monitoring. The online monitoring method, due to the use of automation and remote monitoring technology, can achieve continuous and real-time turbidity detection, and has gradually become the mainstream monitoring method.

[0004] Most of the common online turbidity monitoring devices on the market currently use the optical detection principle. By measuring the intensity changes of incident light, transmitted light, and scattered light, the turbidity value of the suspension is calculated using an empirical formula or a calibration curve. This method is simple and easy to use and can provide continuous turbidity data. However, existing optical turbidity monitoring devices have technical problems such as large errors in measurement results due to differences in the absorption, scattering, and transmission properties of light by different substances and particle forms. Summary of the invention

[0005] In view of this, the present invention provides a non-contact turbidity online measuring device, which can solve the technical problem that the existing optical turbidity monitoring device has large errors in the measurement results due to the differences in the absorption, scattering and transmission properties of light by different substances and particle forms.

[0006] The present invention is achieved in that:

[0007] The present invention provides a non-contact turbidity online measuring device, comprising: a laser light source, an incident optical fiber, an incident lens, an exit optical fiber, a reflecting surface, a collecting lens, a reagent bottle, an optical signal collector and a host computer; the laser light source is used to emit a laser beam, the reagent bottle is used to contain a suspension, one end of the incident optical fiber is connected to the laser light source, and the other end of the incident optical fiber is connected to the incident lens and is used to receive the laser emitted by the laser light source and irradiate the reagent bottle through the incident lens; the reflecting surface and the incident lens are arranged opposite to each other with the reagent bottle separated therefrom, and the optical signal collector is used to receive the laser light ... The center of the reflecting surface is set at the intersection of the incident laser and the reagent bottle, the outer side of the collecting lens is fitted with the outer side of the incident lens, the collecting lens is connected to the emitting optical fiber, the collecting lens is used to collect the laser irradiated on the inner wall of the collecting lens, and transmit it to the optical signal collector through the emitting optical fiber, the optical signal collector is used to receive the laser transmitted by the emitting optical fiber, convert it into an electrical signal and transmit it to the host computer, the host computer is provided with a turbidity calculation module, which is used to calculate the turbidity of the suspension according to the electrical signal.

[0008] On the basis of the above technical solution, the non-contact turbidity online measuring device of the present invention can also be improved as follows:

[0009] Wherein, the reflecting surface is in an arc shape, and the arc angle of the arc is at least π.

[0010] Furthermore, the incident lens is a concave incident lens; and the collecting lens is a convex lens.

[0011] Furthermore, the laser light source is used to emit lasers of multiple wavelengths.

[0012] Furthermore, the optical signal collector is specifically a photomultiplier tube or a photodiode.

[0013] Wherein, the turbidity calculation module is used to perform the following steps:

[0014] S10, acquiring the electrical signal, and preprocessing to obtain a first signal;

[0015] S20, extracting frequency domain features from the first signal by using a sliding window method to obtain a frequency domain feature set;

[0016] S30, clustering the frequency domain feature set to obtain multiple cluster centers;

[0017] S40, inputting a pre-trained signal classification model into each cluster center to obtain light classification of each cluster center, including light reflected by the reflection surface, light reflected by particles in the suspension, light reflected by the reflection surface after transmission of particles in the suspension, and light reflected by the reagent bottle;

[0018] S50, screening the frequency domain feature set, retaining the frequency domain features corresponding to light classification as light reflected from a reflection surface or light reflected from particles in a suspension, and obtaining a screened frequency domain feature set;

[0019] S60, using a pre-trained turbidity calculation model, inputting the screening frequency domain feature set and its corresponding light classification, and outputting the turbidity of the suspension.

[0020] Wherein, the step S10 specifically includes: the optical signal collector converts the received laser signal into an electrical signal and outputs it to the host computer; the turbidity calculation module of the host computer preprocesses the collected electrical signal, mainly including: using a Butterworth low-pass filter to filter the electrical signal to eliminate high-frequency noise components, and the cutoff frequency of the filter can be set to 500Hz; calculating the average value of the electrical signal, subtracting the average value from the electrical signal, and eliminating DC offset; dividing the processed electrical signal by its maximum value, and performing normalization processing so that it falls between 0 and 1, which is convenient for subsequent feature extraction and calculation. Through the above preprocessing steps, a first signal for subsequent analysis is obtained.

[0021] Among them, the step S20 specifically includes: dividing the first signal into several overlapping time windows, the window length is set to 1024 sampling points, and there is a 50% overlap between adjacent windows; performing fast Fourier transform (FFT) on the data in each time window to obtain the frequency domain characteristics of the window; and combining the frequency domain characteristics of all windows into a frequency domain feature set as input for subsequent analysis.

[0022] Among them, the step S30 specifically includes selecting the K-Means clustering algorithm to cluster the frequency domain feature set; the cluster number K of the clustering can be set to 4, which corresponds to 4 different optical components according to physical analysis: light reflected from the reflection surface, light reflected from particles in the suspension, light reflected from the reflection surface after transmission of particles in the suspension, and light reflected from the reagent bottle; through K-Means clustering, 4 cluster centers are obtained, representing 4 different optical components.

[0023] Among them, the step S40 specifically includes: constructing a multi-branch deep learning structure as a signal classification model, including: a frequency domain feature classification module, a high probability frequency domain feature branch network, a low probability frequency domain feature branch network and a fusion branch network; the frequency domain feature classification module classifies the input frequency domain features and divides them into high probability frequency domain features and low probability frequency domain features; the high probability frequency domain feature branch network adopts a multi-layer convolutional neural network to perform optical classification processing on the high probability frequency domain features and output the optical classification probability distribution; the low probability frequency domain feature branch network adopts a multi-layer fully connected neural network to perform optical classification processing on the low probability frequency domain features and output the optical classification probability distribution; the fusion branch network adopts an attention mechanism weighted fusion network to integrate the outputs of the large and small probability branch networks to obtain the final optical classification result; the four cluster centers obtained in step S30 are respectively input into the pre-trained signal classification model to obtain the probability distribution of the four optical components corresponding to each cluster center.

[0024] Wherein, the step S50 specifically includes: according to the result of step S40, each frequency domain feature in the frequency domain feature set is screened; only those frequency domain features whose corresponding light is classified as "reflected light from the reflection surface" or "reflected light from particles in the suspension" are retained, and the remaining frequency domain features are discarded; the retained frequency domain features constitute a new screened frequency domain feature set as input for subsequent turbidity calculation.

[0025] Wherein, the step S60 specifically includes: constructing an embedded hybrid neural network as a turbidity calculation model, including a mathematical calculation sub-model and an error correction sub-model; the mathematical calculation sub-model calculates the scattered light intensity, the absorbed light intensity and the transmitted light intensity according to the principle of physical optics, and preliminarily estimates the turbidity value; the error correction sub-model includes 4 sub-networks, which are respectively used to correct the errors of the scattering equation, the absorption equation, the transmission equation and the turbidity calculation equation, and these sub-networks all adopt a multi-layer perceptron structure; the filtered frequency domain feature set obtained in step S50 and its corresponding light classification are input into the turbidity calculation model, the model will first perform mathematical calculations to obtain a preliminary turbidity value, and then use the error correction sub-network to correct it, and finally output the precise turbidity of the suspension.

[0026] Furthermore, the signal classification model adopts a multi-branch deep learning structure, including a frequency domain feature classification module, a high probability frequency domain feature branch network, a low probability frequency domain feature branch network and a fusion branch network.

[0027] Furthermore, the frequency domain feature classification module classifies each frequency domain feature in the input frequency domain feature set according to probability, into high probability frequency domain features and low probability frequency domain features;

[0028] The high probability frequency domain feature branch network is used to perform light classification processing on the high probability frequency domain feature, the input is a high probability frequency domain feature vector, the output is a light classification probability distribution, and the structure is a multi-layer convolutional neural network;

[0029] The small probability frequency domain feature branch network is used to perform light classification processing on the small probability frequency domain feature, the input is the small probability frequency domain feature vector, the output is the light classification probability distribution, and the structure is a multi-layer fully connected neural network;

[0030] The fusion branch network is used to integrate the outputs of the large and small probability branch networks. The input is the output of the two branch networks, and the output is the final light classification result. The structure is an attention mechanism weighted fusion network.

[0031] The steps of constructing the training data set of the signal classification model are specifically: collecting suspension samples of various concentrations and components; using a device to collect optical signal data of different samples; manually annotating the optical classification of each signal; performing data enhancement, such as adding noise, translation, etc.; dividing the training set, validation set and test set;

[0032] The training steps of the signal classification model are specifically: initializing model parameters; training the model using a small batch gradient descent method; evaluating model performance on a validation set; adjusting hyperparameters based on the validation results; repeating the above steps until the model converges; and performing a final evaluation on a test set to optimize the model.

[0033] Furthermore, the turbidity calculation model adopts an embedded hybrid neural network structure, including a mathematical calculation sub-model and an error correction sub-model; the mathematical calculation sub-model is used to calculate the scattered, absorbed and transmitted light intensities according to the principles of physical optics, and to preliminarily estimate the turbidity value, including a scattering equation, an absorption equation, a transmission equation and a turbidity calculation equation; the error correction sub-model includes a scattering equation error correction sub-network, an absorption equation error correction sub-network, a transmission equation error correction sub-network and a turbidity calculation equation error correction sub-network.

[0034] Furthermore, the scattering equation is used to consider the scattering effect of light in the suspension and obtain the scattered light intensity;

[0035] The absorption equation is used to consider the absorption effect of light in the suspension and obtain the intensity of absorbed light;

[0036] The transmission equation is used to consider the transmission effect of light in the suspension and obtain the intensity of the transmitted light;

[0037] The turbidity calculation equation is used to calculate the turbidity of the suspension based on the scattered light intensity, the absorbed light intensity and the transmitted light intensity;

[0038] The scattering equation error correction subnetwork is used to correct the error of the scattering equation, the input is the input and output of the scattering equation, the output is the correction value of the scattered light intensity, and the structure is a multi-layer perceptron;

[0039] The absorption equation error correction subnetwork is used to correct the error of the absorption equation, the input is the input and output of the absorption equation, the output is the correction value of the absorbed light intensity, and the structure is a multi-layer perceptron;

[0040] The transmission equation error correction subnetwork is used to correct the error of the transmission equation, the input is the input and output of the transmission equation, the output is the correction value of the transmitted light intensity, and the structure is a multi-layer perceptron;

[0041] The turbidity calculation equation error correction subnetwork is used to correct the error of the turbidity calculation equation, the input is the input and output of the turbidity calculation equation, the output is the corrected value of turbidity, and the structure is a multi-layer perceptron.

[0042] The steps for constructing the training data set of the turbidity calculation model are specifically: preparing standard samples with known turbidity; using a device to collect optical signal data of the standard samples; extracting frequency domain features and performing optical classification; recording the true turbidity value and various optical parameters; dividing the training set, the validation set and the test set; the training steps of the turbidity calculation model are specifically: initializing the parameters of the mathematical calculation sub-model; training the error correction sub-network, using the true turbidity value as the target; evaluating the model performance on the validation set; adjusting the hyperparameters according to the validation results; repeating the above steps until the model converges; and performing a final evaluation on the test set to optimize the model.

[0043] Compared with the prior art, the non-contact turbidity online measuring device provided by the present invention has the following beneficial effects:

[0044] The advanced optical principle and deep learning algorithm are used to provide reliable online turbidity monitoring in complex industrial environments. Compared with the prior art, the present invention has the following significant advantages:

[0045] 1. High measurement accuracy. This device adopts a fusion analysis method of multiple optical parameters, which not only takes into account the intensity changes of scattered light, absorbed light and transmitted light, but also introduces high-order optical features such as spectral derivatives. Through deep learning algorithms, intelligent classification and error correction are performed, which can greatly improve the accuracy and reliability of turbidity measurement.

[0046] 2. Strong anti-interference ability. The optical path of this device adopts a non-contact design to avoid direct contact with the suspension and reduce the influence of environmental factors. At the same time, through advanced signal processing and classification algorithms, it can effectively identify and suppress interference light from reflective surfaces, container walls, etc., further improving the stability of measurement.

[0047] Therefore, the present invention solves the technical problem that the existing optical turbidity monitoring device has large errors in the measurement results due to the differences in the absorption, scattering and transmission properties of light by different substances and particle forms. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a non-contact turbidity online measurement device provided by the present invention;

[0049] Figure 2 A flowchart of the steps executed by the turbidity calculation module in the present invention; DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0051] like Figure 1 As shown, it is a schematic diagram of a non-contact turbidity online measuring device provided by the present invention, and the device comprises: a laser light source, an incident optical fiber, an incident lens, an exit optical fiber, a reflecting surface, a collecting lens, a reagent bottle, an optical signal collector and a host computer; the laser light source is used to emit a laser beam, the reagent bottle is used to contain a suspension, one end of the incident optical fiber is connected to the laser light source, and the other end of the incident optical fiber is connected to the incident lens, and is used to receive the laser emitted by the laser light source, and irradiate the reagent bottle through the incident lens; the reflecting surface and the incident lens are arranged opposite to each other with the reagent bottle separated, the center of the reflecting surface is arranged at the intersection of the incident laser and the reagent bottle, the outer side of the collecting lens is arranged to fit the outer side of the incident lens, the collecting lens is connected to the exit optical fiber, the collecting lens is used to collect the laser irradiated on the inner wall of the collecting lens, and transmit it to the optical signal collector through the exit optical fiber, the optical signal collector is used to receive the laser transmitted by the exit optical fiber, and convert it into an electrical signal and transmit it to the host computer, and a turbidity calculation module is arranged in the host computer, which is used to calculate the turbidity of the suspension according to the electrical signal.

[0052] like Figure 2 As shown, the turbidity calculation module is used to perform the following steps:

[0053] S10, acquiring the electrical signal, and preprocessing to obtain a first signal;

[0054] S20, extracting frequency domain features from the first signal by using a sliding window method to obtain a frequency domain feature set;

[0055] S30, clustering the frequency domain feature set to obtain multiple cluster centers;

[0056] S40, inputting a pre-trained signal classification model into each cluster center to obtain light classification of each cluster center, including light reflected by the reflection surface, light reflected by particles in the suspension, light reflected by the reflection surface after transmission of particles in the suspension, and light reflected by the reagent bottle;

[0057] S50, screening the frequency domain feature set, retaining the frequency domain features corresponding to light classification as light reflected from a reflection surface or light reflected from particles in a suspension, and obtaining a screened frequency domain feature set;

[0058] S60, using a pre-trained turbidity calculation model, inputting the screening frequency domain feature set and its corresponding light classification, and outputting the turbidity of the suspension.

[0059] The specific implementation methods of the above steps are described in detail below:

[0060] Step S10: Obtain an electrical signal and preprocess it to obtain a first signal

[0061] The specific implementation method of this step is as follows:

[0062] 1) The optical signal collector converts the received laser signal into an electrical signal and outputs it to the host computer.

[0063] 2) The turbidity calculation module of the host computer pre-processes the collected electrical signals, mainly including:

[0064] -Eliminate high-frequency noise: Use a Butterworth low-pass filter to filter the electrical signal and remove high-frequency noise components. The cutoff frequency of the filter can be set to 500Hz.

[0065] - Eliminate DC offset: Calculate the average value of the electrical signal and subtract the average value from the electrical signal to eliminate the DC offset.

[0066] -Normalization processing: Divide the processed electrical signal by its maximum value so that it falls between 0 and 1, which is convenient for subsequent feature extraction and calculation.

[0067] Through the above preprocessing steps, a first signal is obtained for subsequent analysis.

[0068] Step S20: Extract frequency domain features of the first signal using a sliding window to obtain a frequency domain feature set

[0069] The specific implementation method of this step is as follows:

[0070] 1) Divide the first signal into several overlapping time windows, the window length is set to 1024 sampling points, and there is a 50% overlap between adjacent windows.

[0071] 2) Perform fast Fourier transform (FFT) on the data in each time window to obtain the frequency domain characteristics of the window.

[0072] 3) The frequency domain features of all windows are combined into a frequency domain feature set as input for subsequent analysis.

[0073] Step S30: Cluster the frequency domain feature set to obtain multiple cluster centers

[0074] The specific implementation method of this step is as follows:

[0075] 1) Select the K-Means clustering algorithm to cluster the frequency domain feature set.

[0076] 2) The number of clusters K can be set to 4. According to physical analysis, this corresponds to 4 different optical components: light reflected from the reflective surface, light reflected from particles in the suspension, light reflected from the reflective surface after transmission from particles in the suspension, and light reflected from the reagent bottle.

[0077] 3) Through K-Means clustering, 4 cluster centers were obtained, representing 4 different optical components.

[0078] Step S40: Input the pre-trained signal classification model to each cluster center to obtain the light classification of each cluster center.

[0079] The specific implementation method of this step is as follows:

[0080] 1) Construct a multi-branch deep learning structure as a signal classification model, including: frequency domain feature classification module, high probability frequency domain feature branch network, low probability frequency domain feature branch network and fusion branch network.

[0081] 2) The frequency domain feature classification module classifies the input frequency domain features and divides them into high-probability frequency domain features and low-probability frequency domain features.

[0082] 3) The high probability frequency domain feature branch network adopts a multi-layer convolutional neural network to perform optical classification processing on the high probability frequency domain features and output the optical classification probability distribution.

[0083] 4) The small probability frequency domain feature branch network adopts a multi-layer fully connected neural network to perform optical classification processing on the small probability frequency domain features and output the optical classification probability distribution.

[0084] 5) The fusion branch network uses an attention mechanism weighted fusion network to integrate the outputs of the large and small probability branch networks to obtain the final light classification result.

[0085] 6) The four cluster centers obtained in step S30 are respectively input into the pre-trained signal classification model to obtain the probability distribution of the four optical components corresponding to each cluster center.

[0086] Step S50: Filter the frequency domain feature set, retain the frequency domain features corresponding to the light classification as the reflected light of the reflection surface or the reflected light of the particles in the suspension, and obtain the filtered frequency domain feature set

[0087] The specific implementation method of this step is as follows:

[0088] 1) According to the result of step S40, each frequency domain feature in the frequency domain feature set is screened.

[0089] 2) Only those frequency domain features corresponding to the light classification as "light reflected from the reflection surface" or "light reflected from particles in the suspension" are retained, and the rest of the frequency domain features are discarded.

[0090] 3) The retained frequency domain features constitute a new filtered frequency domain feature set as input for subsequent turbidity calculations.

[0091] Step S60: Use the pre-trained turbidity calculation model, input the screening frequency domain feature set and its corresponding light classification, and output the turbidity of the suspension

[0092] The specific implementation method of this step is as follows:

[0093] 1) Construct an embedded hybrid neural network as the turbidity calculation model, including a mathematical calculation sub-model and an error correction sub-model.

[0094] 2) The mathematical calculation sub-model calculates the scattered light intensity, absorbed light intensity and transmitted light intensity according to the principles of physical optics, and preliminarily estimates the turbidity value. Among them:

[0095] -The scattering equation takes into account the scattering effect of light in the suspension and obtains the scattered light intensity.

[0096] -The absorption equation takes into account the absorption effect of light in the suspension and obtains the absorbed light intensity.

[0097] -The transmission equation takes into account the transmission effect of light in the suspension and obtains the transmitted light intensity.

[0098] - Turbidity calculation equation calculates the preliminary turbidity value based on these optical parameters.

[0099] 3) The error correction sub-model includes four sub-networks, which are used to correct the errors of the scattering equation, absorption equation, transmission equation and turbidity calculation equation respectively. These sub-networks all adopt the multi-layer perceptron structure.

[0100] 4) The filtered frequency domain feature set obtained in step S50 and its corresponding light classification are input into the turbidity calculation model. The model will first perform mathematical calculations to obtain a preliminary turbidity value, then use the error correction subnetwork to correct it, and finally output the precise turbidity of the suspension.

[0101] The above is a specific implementation of the non-contact turbidity online measurement device. Some key parameters and algorithm details are added below:

[0102] 1. The key parameters in the scattering equation include: incident light intensity I0, scattering efficiency factor Q s , the wavelength of the incident light λ, the distance R from the particle to the observation point, and the Mie scattering coefficient a n and b n . Where Q s By measuring the particle size distribution and calculating the Mie theory, a n and b n The calculation formula is also given.

[0103] 2. The key parameters in the absorption equation include: absorption coefficient α, optical path length L, incident light angular frequency ω, characteristic absorption frequency ω0, and temperature gradient and Among them, α is obtained by measuring the absorbance and fitting, and the temperature gradient is calculated by measuring the temperature sensor.

[0104] 3. The key parameters in the transmission equation include: absorption coefficient α, scattering coefficient σ, optical path length L, second-order partial derivative of light wave phase φ, and Hermite polynomial expansion coefficient c k and characteristic length ρ. Among them, σ is measured by integrating sphere, the partial derivative of φ is calculated by interferometer measurement, and c k and ρ are obtained by nonlinear fitting.

[0105] 4. The key parameters in the turbidity calculation equation include: proportionality coefficient K, weight index a, b, c, and spectral derivative coefficient d l These parameters were fitted by measuring standard samples using the multivariate nonlinear regression method.

[0106] Specifically, the principle of the present invention is to use optical principles and deep learning algorithms to analyze the scattered, absorbed and transmitted light intensities in the suspension to obtain an accurate turbidity value. Specifically, its working principle is as follows:

[0107] 1. Optical path design

[0108] The device adopts a non-contact optical path design. The laser emitted by the laser light source passes through the incident optical fiber and the incident lens and irradiates the suspension sample to be tested. The reflective surface is set opposite to the incident lens, and its center is located at the intersection of the incident laser and the suspension sample. When the laser is irradiated into the suspension, optical effects such as scattering, absorption and transmission will occur. The reflective surface captures and reflects these optical signals, and the collection lens focuses them onto the output optical fiber, which is finally detected by the optical signal collector. This non-contact design avoids direct contact between the optical probe and the suspension and reduces interference from environmental factors.

[0109] 2. Multi-optical parameter analysis

[0110] The optical signal collector converts the received optical signal into an electrical signal and transmits it to the host computer for analysis and processing. The turbidity calculation module of the host computer first pre-processes the electrical signal, including filtering, removing DC components, normalization and other operations, to obtain the first signal for analysis.

[0111] Then, the module uses a sliding window method to perform fast Fourier transform (FFT) on the first signal to extract frequency domain features. These frequency domain features not only reflect the intensity changes of scattered light, absorbed light and transmitted light, but also contain high-order spectral derivative information, which can more comprehensively describe the characteristics of optical signals.

[0112] 3. Deep learning classification and correction

[0113] Next, the turbidity calculation module uses a pre-trained deep learning model to classify and identify frequency domain features. The model includes a frequency domain feature classification module, a large and small probability branch network, and a fusion network, which can effectively distinguish different optical components such as reflected light from the reflection surface and reflected light from suspended particles.

[0114] According to the different optical components identified, the turbidity calculation module uses the principles of physical optics to calculate the scattered light intensity, absorbed light intensity and transmitted light intensity respectively, and preliminarily estimates the turbidity value of the suspension.

[0115] In order to better understand and implement the present invention, a specific embodiment 1 of the turbidity calculation module in the present invention is provided below. The steps included in this embodiment 1 are specifically described as follows:

[0116] Step S10: Obtain an electrical signal and preprocess it to obtain a first signal

[0117] The specific implementation method of this step is as follows:

[0118] The optical signal collector converts the received laser signal into an electrical signal V(t) and outputs it to the host computer. The turbidity calculation module of the host computer pre-processes the collected electrical signal, mainly including:

[0119] 1) Eliminate high frequency noise

[0120] The Butterworth low-pass filter is used to filter the electrical signal and remove the high-frequency noise components. The transfer function of the Butterworth low-pass filter is:

[0121]

[0122] Where f is the signal frequency, f c is the cut-off frequency, and n is the filter order. In this device, fc Set it to 500Hz and n to 4 to effectively filter out high-frequency noise. The filtered signal is recorded as V1(t).

[0123] 2) Eliminate DC offset

[0124] Calculate the average value of the electrical signal V1(t) Subtract this average value from V1(t) to remove the DC offset. The processed signal is recorded as V2(t):

[0125]

[0126] 3) Normalization

[0127] Divide the processed electrical signal V2(t) by its maximum value V 2,max , normalized to fall between 0 and 1, which is convenient for subsequent feature extraction and calculation:

[0128]

[0129] Through the above preprocessing steps, a first signal V3(t) is obtained for subsequent analysis.

[0130] Step S20: Extract frequency domain features of the first signal using a sliding window to obtain a frequency domain feature set

[0131] The specific implementation method of this step is as follows:

[0132] 1) Divide the first signal V3(t) into several overlapping time windows, with the window length set to N = 1024 sampling points, and 50% overlap between adjacent windows. The discrete signal in each time window is denoted as v k [n], where k = 1, 2, ..., K represents the window number, and n = 1, 2, ..., N represents the sampling point number.

[0133] 2) For each data v in the time window k [n] Perform fast Fourier transform (FFT) respectively to obtain the frequency domain characteristics V of the window k [m], where m = 1, 2, ..., N represents the frequency BIN number (Frequency-Bin). The formula for FFT is:

[0134]

[0135] 3) The frequency domain features V of all windows k [m] forms a frequency domain feature set V = {V1, V2, ..., V K}, as input for subsequent analysis.

[0136] Step S30: Cluster the frequency domain feature set to obtain multiple cluster centers

[0137] The specific implementation method of this step is as follows:

[0138] 1) Select the K-Means clustering algorithm to cluster the frequency domain feature set V. The objective function of the K-Means algorithm is:

[0139]

[0140] Where K is the number of clusters, C i is the i-th cluster, μ i is the i-th cluster center.

[0141] 2) The number of clusters K can be set to 4. According to physical analysis, this corresponds to 4 different optical components: light reflected from the reflective surface, light reflected from particles in the suspension, light reflected from the reflective surface after transmission from particles in the suspension, and light reflected from the reagent bottle.

[0142] 3) Through K-Means clustering, four cluster centers μ1, μ2, μ3, and μ4 are obtained, representing four different optical components.

[0143] Step S40: Input the pre-trained signal classification model to each cluster center to obtain the light classification of each cluster center.

[0144] The specific implementation method of this step is as follows:

[0145] 1) Construct a multi-branch deep learning structure as a signal classification model, including: frequency domain feature classification module, high probability frequency domain feature branch network, low probability frequency domain feature branch network and fusion branch network.

[0146] 2) The frequency domain feature classification module uses a binary classifier to classify the input frequency domain feature v and divide it into high probability frequency domain features v (l) and small probability frequency domain features v (s) The output of the classifier is p(v (l) |v) and p(v (s) |v).

[0147] 3) The high probability frequency domain feature branch network uses a multi-layer convolutional neural network to perform high probability frequency domain feature v (l) Perform light classification processing and output light classification probability distribution p (l) :

[0148] p (l) =f l (v (l) θ l );

[0149] Among them, f l is the mapping function of the high probability frequency domain feature branch network, θ l is the network parameter.

[0150] 4) The small probability frequency domain feature branch network uses a multi-layer fully connected neural network to analyze the small probability frequency domain feature v (s) Perform light classification processing and output light classification probability distribution p (s) :

[0151] p (s) =f s (v (s) θ s );

[0152] Among them, f s is the mapping function of the small probability frequency domain feature branch network, θ s is the network parameter.

[0153] 5) The fusion branch network uses the attention mechanism to weight the fusion network and integrate the output p of the large and small probability branch networks (l) and p (s) , and get the final light classification result p:

[0154] p=f f (p (l) ,p (s) θ f );

[0155] Among them, f f is the mapping function of the fusion branch network, θ f is the network parameter.

[0156] 6) The four cluster centers μ1, μ2, μ3, and μ4 obtained in step S30 are respectively input into the pre-trained signal classification model to obtain the probability distributions p1, p2, p3, and p4 of the four optical components corresponding to each cluster center.

[0157] Step S50: Filter the frequency domain feature set, retain the frequency domain features corresponding to the light classification as the reflected light of the reflection surface or the reflected light of the particles in the suspension, and obtain the filtered frequency domain feature set

[0158] The specific implementation method of this step is as follows:

[0159] 1) According to the result of step S40, each frequency domain feature v in the frequency domain feature set V is screened. Specifically, the probabilities p1, p2, p3, and p4 of the four optical components corresponding to each v are calculated, and it is determined whether the optical component corresponding to the maximum probability is "reflected light from the reflection surface" or "reflected light from particles in the suspension".

[0160] 2) Only those frequency domain features v whose corresponding light is classified as "light reflected from the reflection surface" or "light reflected from particles in the suspension" are retained, and the rest of the frequency domain features are discarded.

[0161] 3) The retained frequency domain features constitute a new screening frequency domain feature set V (s) , as input for subsequent turbidity calculations.

[0162] Step S60: Use the pre-trained turbidity calculation model, input the screening frequency domain feature set and its corresponding light classification, and output the turbidity of the suspension

[0163] The specific implementation method of this step is as follows:

[0164] 1) Construct an embedded hybrid neural network as the turbidity calculation model, including a mathematical calculation sub-model and an error correction sub-model.

[0165] 2) The mathematical calculation sub-model calculates the scattered light intensity I according to the principle of physical optics s , Absorption light intensity I a and the transmitted light intensity I t , and preliminarily estimate the turbidity value T0.

[0166] Scattered light intensity I s The calculation formula is:

[0167]

[0168] Where I0 is the incident light intensity, Q s is the scattering efficiency factor, λ is the wavelength of the incident light, R is the distance from the particle to the observation point, θ is the scattering angle, and a n and b n is the Mie scattering coefficient.

[0169] Absorption intensity I a The calculation formula is:

[0170]

[0171] Where α is the absorption coefficient, L is the optical path length, ω is the incident light angular frequency, ω0 is the characteristic absorption frequency, and N is the number of temperature gradient layers considered. is the temperature coefficient of refractive index of the i-th layer, is the temperature gradient of the i-th layer.

[0172] Transmitted light intensity I t The calculation formula is:

[0173]

[0174] Where σ is the scattering coefficient, φ is the light wave phase, and x jis the spatial coordinate, M is the considered spatial dimension, c k is the expansion coefficient, H k is the k-order Hermite polynomial, and ρ is the characteristic length of the suspension.

[0175] Substituting these optical parameters into the turbidity calculation equation, the preliminary turbidity value T0 can be obtained:

[0176]

[0177] Among them, K is the proportional coefficient, a, b, c are weight indexes, d l is the spectral derivative coefficient, and L is the highest order derivative considered.

[0178] 3) The error correction sub-model includes four sub-networks, which are used to correct the errors of the scattering equation, absorption equation, transmission equation and turbidity calculation equation respectively. These sub-networks all adopt the multi-layer perceptron (MLP) structure.

[0179] The input of the scattering equation error correction subnetwork is the input and output of the scattering equation, and the output is the correction value of the scattered light intensity:

[0180] ΔI s =f s (I0,Q s ,λ,R,a n ,b n θ s )

[0181] Among them, f s is the MLP mapping function, θ s is the network parameter.

[0182] The structures and input and output of the absorption equation error correction subnetwork, the transmission equation error correction subnetwork and the turbidity calculation equation error correction subnetwork are similar.

[0183] 4) The filtered frequency domain feature set V obtained in step S50 is (s) The model will first perform mathematical calculations to obtain a preliminary turbidity value T0, then use the error correction subnetwork to correct it, and finally output the precise turbidity T of the suspension.

[0184] In order to further better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: During the production process of a chemical plant, it is necessary to monitor the turbidity of the suspension in the production tank in real time. The suspension is prepared from a variety of raw materials through processes such as reaction and precipitation, and the main components include calcium salts, silicates and a small amount of organic matter. Due to the complexity of the production process and the influence of factors such as ambient temperature and flow rate, the turbidity of the suspension often fluctuates greatly, and a reliable online monitoring method is required for real-time tracking.

[0185] The chemical plant decided to use the non-contact turbidity online measurement device proposed in the present invention to monitor the turbidity of the suspension in the production tank. The specific implementation process of the device is as follows:

[0186] 1. Device installation

[0187] First, the production tank is modified and an observation hole with a diameter of about 50 mm is opened on the side wall of the tank. The non-contact turbidity online measurement device of the present invention is installed directly opposite to the observation hole so that its optical path can illuminate and detect the suspension in the tank. The main components of the device include:

[0188] (1) Laser light source: wavelength 532nm, peak power 10mW.

[0189] (2) Incident fiber and incident lens: The incident fiber couples the output of the laser light source to the incident lens, which is a concave lens with a focal length of 50 mm.

[0190] (3) Reflective surface: A circular arc-shaped reflective surface with a radius of 100 mm and an arc greater than 180 degrees, made of a high-reflectivity mirror.

[0191] (4) Collection lens: A convex lens with a focal length of 30 mm is used and fitted to the outer side of the incident lens.

[0192] (5) Output optical fiber and optical signal collector: The output optical fiber transmits the optical signal focused by the collection lens to the photomultiplier tube.

[0193] (6) Host computer: It uses an industrial-grade embedded computer with a built-in turbidity calculation module.

[0194] The entire device adopts a compact design and is installed on the observation hole on the side wall of the production tank, and keeps a certain distance from the tank body to avoid direct contact with the suspension. The incident laser is irradiated into the suspension through the lens, absorbed, scattered and transmitted by the suspended particles. The reflective surface captures these optical signals and reflects them to the collection lens, and finally converts them into electrical signals and sends them to the host computer for analysis and processing. This non-contact design avoids direct contact between the probe and the corrosive suspension, greatly improving the applicability and stability of the device.

[0195] 2. Signal acquisition and preprocessing

[0196] The optical signal collector converts the received laser signal into an electrical signal V(t) and transmits it to the turbidity calculation module of the host computer in real time. The module first pre-processes the collected electrical signal, removes high-frequency noise and DC components, and performs normalization:

[0197] (1) Filtering: A 4th-order Butterworth low-pass filter with a cutoff frequency of 500 Hz is used to effectively remove high-frequency noise. The filtered signal is recorded as V1(t).

[0198] (2) Remove DC: Calculate the average value of V1(t) Subtract this average value from V1(t) to remove the DC offset. The processed signal is recorded as V2(t):

[0199]

[0200] (3) Normalization: Divide V2(t) by its maximum value V 2,max , and the amplitude is normalized. The normalized signal is recorded as V3(t):

[0201]

[0202] After the above preprocessing, a first signal V3(t) is obtained for subsequent analysis.

[0203] 3. Frequency domain feature extraction

[0204] Next, the turbidity calculation module uses a sliding window method to extract frequency domain features from the preprocessed first signal V3(t). The specific steps are as follows:

[0205] (1) Divide V3(t) into overlapping time windows of length 1024 samples, with 50% overlap between adjacent windows.

[0206] (2) For each data v in the time window k [n] (where k = 1, 2, ..., K represents the window number, n = 1, 2, ..., 1024 represents the sampling point number) respectively perform FFT to obtain the frequency domain feature V k [m] (where m = 1, 2, ..., 1024 represents the frequency bin number):

[0207]

[0208] (3) The frequency domain features V of all windows k [m] forms a frequency domain feature set V = {V1, V2, ..., V K}, as input for subsequent analysis.

[0209] Through the above sliding window FFT processing, the frequency domain feature set V not only contains the amplitude information of each frequency component, but also contains some high-order optical features, such as spectral derivatives, which can more comprehensively reflect the optical properties of the suspension.

[0210] 4. Optical component classification

[0211] The turbidity calculation module then uses a pre-trained deep learning model to classify the optical components of the frequency domain feature set V. The model includes:

[0212] (1) Frequency domain feature classification module: A binary classifier is used to divide the frequency domain feature v into high probability frequency domain features v (l) and small probability frequency domain features v (s) .

[0213] (2) High probability frequency domain feature branch network: Using a multi-layer convolutional neural network, input v (l) , output light classification probability p (l) .

[0214] (3) Low probability frequency domain feature branch network: a multi-layer fully connected network is used, and the input v (s) , output light classification probability p (s) .

[0215] (4) Fusion branch network: Use the attention mechanism to weight the fusion network and comprehensively (l) and p (s) , and obtain the final light classification result p.

[0216] The deep learning model was pre-trained using a large amount of labeled data and can effectively distinguish the four optical components contained in the frequency domain features: light reflected from the reflection surface, light reflected from particles in the suspension, light reflected from the reflection surface after transmission of particles in the suspension, and light reflected from the reagent bottle.

[0217] By inputting the frequency domain feature set V obtained in step 3 into the pre-trained model, the probability distributions p1, p2, p3, and p4 of the four optical components corresponding to each frequency domain feature v can be obtained.

[0218] 5. Turbidity calculation

[0219] With the optical component classification results obtained in the previous step, the turbidity calculation module can then calculate the turbidity value of the suspension based on the principles of physical optics. This module includes a mathematical calculation sub-model and an error correction sub-model based on deep learning.

[0220] The mathematical calculation sub-model first calculates the scattered light intensity I s , Absorption light intensity I a and the transmitted light intensity I t :

[0221] (1) Scattered light intensity I s The calculation formula is:

[0222]

[0223] Where I0 is the incident light intensity, Q s is the scattering efficiency factor, λ is the wavelength of the incident light, R is the distance from the particle to the observation point, θ is the scattering angle, and a n and b n is the Mie scattering coefficient.

[0224] (2) Absorption light intensity I a The calculation formula is:

[0225]

[0226] Where α is the absorption coefficient, L is the optical path length, ω is the incident light angular frequency, ω0 is the characteristic absorption frequency, and N is the number of temperature gradient layers considered. is the temperature coefficient of refractive index of the i-th layer, is the temperature gradient of the i-th layer.

[0227] (3) Transmitted light intensity I t The calculation formula is:

[0228]

[0229] Where σ is the scattering coefficient, φ is the light wave phase, and x j is the spatial coordinate, M is the considered spatial dimension, c k is the expansion coefficient, H k is the k-order Hermite polynomial, and ρ is the characteristic length of the suspension.

[0230] Substituting these optical parameters into the turbidity calculation equation, the preliminary turbidity value T0 can be obtained:

[0231]

[0232] Among them, K is the proportional coefficient, a, b, c are weight indexes, d l is the spectral derivative coefficient, and L is the highest order derivative considered.

[0233] However, due to the complexity of the real environment, the above physical calculation formulas inevitably have certain errors. Therefore, the turbidity calculation module also introduces an error correction subnetwork based on deep learning, which can effectively correct these errors. Specifically:

[0234] (1) Scattering equation error correction subnetwork: The input is the input and output of the scattering equation, and the output is the corrected value of the scattered light intensity.

[0235] (2) Absorption equation error correction subnetwork: The input is the input and output of the absorption equation, and the output is the corrected value of the absorbed light intensity.

[0236] (3) Transmission equation error correction subnetwork: The input is the input and output of the transmission equation, and the output is the corrected value of the transmitted light intensity.

[0237] (4) Turbidity calculation equation error correction subnetwork: The input is the input and output of the turbidity calculation equation, and the output is the corrected value of turbidity.

[0238] The optical component classification information obtained in step 4 and the frequency domain feature set V in step 3 are (s) (Only retaining the features corresponding to the reflected light from the reflecting surface and the reflected light from the suspension particles) is used as input, and after being processed by the mathematical calculation sub-model and the error correction sub-network, the precise turbidity value T of the suspension is finally output.

[0239] 6. On-site implementation results

[0240] After the above installation and debugging, the non-contact turbidity online measurement device was officially put into production practice in the chemical plant. The implementation results are as follows:

[0241] (1) Measurement accuracy

[0242] In order to verify the measurement accuracy of the device, the chemical plant selected 10 groups of suspension samples with different concentrations and used this device to perform turbidity detection. The laboratory spectrophotometry method was also used for comparative testing. The results are shown in Table 1:

[0243] Table 1 Measurement results

[0244] Sample No. This device measures (NTU) Spectrophotometry (NTU) error(%) 1 12.5 12.1 3.3％ 2 24.3 23.8 2.1％ 3 48.6 49.2 -1.2％ 4 71.4 70.8 0.8％ 5 92.7 93.5 -0.9％ 6 125.3 126.1 -0.6％ 7 164.2 165.4 -0.7％ 8 205.8 207.2 -0.7％ 9 242.5 244.1 -0.7％ 10 295.6 298.3 -0.9％

[0245] From the comparison results, the turbidity value measured by this device is basically consistent with the laboratory test results, and the maximum relative error is only 3.3%, which is far better than the level of existing online monitoring technology.

[0246] (2) Anti-interference

[0247] In order to verify the stability of this device under complex process environment, the chemical plant artificially adjusted some factors that may affect the measurement during the production process, including ambient temperature (20-40℃), suspension flow rate (0.1-1.0m / s), tank wall contamination, etc. The results show that even under these conditions, the device can still maintain good measurement performance, and the turbidity data fluctuation is less than ±2%, which is much better than similar products.

[0248] This is mainly due to the non-contact optical path design of the device, which avoids direct contact with corrosive suspensions and reduces interference from environmental factors. At the same time, the advanced deep learning algorithm can effectively identify and suppress interference light from reflective surfaces, container walls, etc., further improving the stability of the measurement.

[0249] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A non-contact turbidity online measuring device, characterized in that: include: A laser light source, an incident optical fiber, an incident lens, an exit optical fiber, a reflecting surface, a collecting lens, a reagent bottle, an optical signal collector and a host computer; the laser light source is used to emit a laser beam, the reagent bottle is used to hold a suspension, one end of the incident optical fiber is connected to the laser light source, and the other end of the incident optical fiber is connected to the incident lens and is used to receive the laser emitted by the laser light source and irradiate the reagent bottle through the incident lens; the reflecting surface and the incident lens are arranged opposite to each other with the reagent bottle separated, the center of the reflecting surface is arranged at the intersection of the incident laser and the reagent bottle, the outer side of the collecting lens is arranged to fit the outer side of the incident lens, the collecting lens is connected to the exit optical fiber, the collecting lens is used to collect the laser irradiated on the inner wall of the collecting lens, and transmit it to the optical signal collector through the exit optical fiber, the optical signal collector is used to receive the laser transmitted by the exit optical fiber, convert it into an electrical signal and transmit it to the host computer, a turbidity calculation module is arranged in the host computer, and is used to calculate the turbidity of the suspension according to the electrical signal; the turbidity calculation module is used to perform the following steps: S10, acquiring the electrical signal, and preprocessing to obtain a first signal; S20, extracting frequency domain features from the first signal by using a sliding window method to obtain a frequency domain feature set; S30, clustering the frequency domain feature set to obtain multiple cluster centers; S40, inputting a pre-trained signal classification model into each cluster center to obtain light classification of each cluster center, including light reflected by the reflection surface, light reflected by particles in the suspension, light reflected by the reflection surface after transmission of particles in the suspension, and light reflected by the reagent bottle; S50, screening the frequency domain feature set, retaining the frequency domain features corresponding to light classification as reflected light from a reflecting surface or reflected light from particles in a suspension, and obtaining a screened frequency domain feature set; S60, using a pre-trained turbidity calculation model, inputting the screening frequency domain feature set and its corresponding light classification, and outputting the turbidity of the suspension.

2. A non-contact turbidity online measuring device according to claim 1, characterized in that: The reflecting surface is in an arc shape, and the arc angle of the arc is at least π.

3. A non-contact turbidity online measuring device according to claim 2, characterized in that: The incident lens is a concave incident lens; the collecting lens is a convex lens.

4. A non-contact turbidity online measuring device according to claim 3, characterized in that: The laser light source is used to emit lasers with multiple wavelengths.

5. A non-contact turbidity online measuring device according to claim 4, characterized in that: The optical signal collector is specifically a photomultiplier tube or a photodiode.

6. A non-contact turbidity online measuring device according to claim 1, characterized in that: The signal classification model adopts a multi-branch deep learning structure, including a frequency domain feature classification module, a high probability frequency domain feature branch network, a low probability frequency domain feature branch network and a fusion branch network.

7. A non-contact turbidity online measuring device according to claim 6, characterized in that: The frequency domain feature classification module classifies each frequency domain feature in the input frequency domain feature set according to probability, into high-probability frequency domain features and low-probability frequency domain features; The high probability frequency domain feature branch network is used to perform light classification processing on the high probability frequency domain feature, the input is a high probability frequency domain feature vector, the output is a light classification probability distribution, and the structure is a multi-layer convolutional neural network; The small probability frequency domain feature branch network is used to perform light classification processing on the small probability frequency domain feature, the input is the small probability frequency domain feature vector, the output is the light classification probability distribution, and the structure is a multi-layer fully connected neural network; The fusion branch network is used to integrate the outputs of the large and small probability branch networks. The input is the output of the two branch networks, and the output is the final light classification result. The structure is an attention mechanism weighted fusion network.

8. A non-contact turbidity online measuring device according to claim 7, characterized in that: The turbidity calculation model adopts an embedded hybrid neural network structure, including a mathematical calculation sub-model and an error correction sub-model; the mathematical calculation sub-model is used to calculate the scattered, absorbed and transmitted light intensities according to the principles of physical optics, and preliminarily estimate the turbidity value, including a scattering equation, an absorption equation, a transmission equation and a turbidity calculation equation; the error correction sub-model includes a scattering equation error correction sub-network, an absorption equation error correction sub-network, a transmission equation error correction sub-network and a turbidity calculation equation error correction sub-network.

9. A non-contact turbidity online measuring device according to claim 8, characterized in that: The scattering equation is used to consider the scattering effect of light in the suspension and obtain the scattered light intensity; The absorption equation is used to consider the absorption effect of light in the suspension and obtain the intensity of absorbed light; The transmission equation is used to consider the transmission effect of light in the suspension and obtain the intensity of the transmitted light; The turbidity calculation equation is used to calculate the turbidity of the suspension based on the scattered light intensity, the absorbed light intensity and the transmitted light intensity; The scattering equation error correction subnetwork is used to correct the error of the scattering equation, the input is the input and output of the scattering equation, the output is the correction value of the scattered light intensity, and the structure is a multi-layer perceptron; The absorption equation error correction subnetwork is used to correct the error of the absorption equation, the input is the input and output of the absorption equation, the output is the correction value of the absorbed light intensity, and the structure is a multi-layer perceptron; The transmission equation error correction subnetwork is used to correct the error of the transmission equation, the input is the input and output of the transmission equation, the output is the correction value of the transmitted light intensity, and the structure is a multi-layer perceptron; The turbidity calculation equation error correction subnetwork is used to correct the error of the turbidity calculation equation, the input is the input and output of the turbidity calculation equation, the output is the corrected value of turbidity, and the structure is a multi-layer perceptron.

Images