Solid-containing liquid system state online detection equipment and method based on neural network

Abstract

The invention provides solid-containing liquid system state online detection equipment, method and device based on a neural network. The method comprises the following steps: acquiring an ultrasonic attenuation spectrogram of a solid-containing liquid system to be analyzed in real time; inputting the ultrasonic attenuation spectrogram and the offline particle size data into a pre-trained correction model; inputting the ultrasonic attenuation spectrogram obtained in real time into a pre-trained quantitative correction model to obtain a particle size distribution state; therefore, the technical problem of measurement result lag in the prior art is solved, and the particle size distribution state and uniformity of the particles in the solid-containing liquid system are monitored in real time.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and analytical chemistry, and in particular to an online detection device, method and apparatus for the state of a solid-liquid system based on a neural network. Background Art

[0002] In solid-liquid systems, particle size distribution is an important measurement parameter. Commonly used particle size measurement methods include Malvern laser diffraction and dynamic light scattering. However, these methods are not suitable for online PSD measurement due to limitations such as their inability to directly measure high-concentration systems and their high measurable particle size limits. Ultrasonic attenuation spectroscopy is the most promising technology for online PSD measurement during nanoparticle processing. However, existing methods typically use models such as the Extended Compressed Admittance Model (ECAH) to calculate the particle size distribution using online acquired ultrasonic attenuation spectra. These models typically require a large number of physical properties (each model requires 14, involving both solid and liquid phases). Some of these properties are not only difficult to measure but also vary with system conditions. For example, the viscosity of a solid-liquid mixture varies with solid particle size. Consequently, the calculated results are inaccurate and often cannot provide a unique solution due to data limitations. Furthermore, the calculation process is lengthy (typically 2-5 minutes), making it impossible to truly achieve real-time online detection and analysis of particle size distribution.

[0003] In view of this, the long time and low accuracy of particle size calculation in solid-liquid systems have become a problem that needs to be urgently solved by those skilled in the art. Summary of the Invention

[0004] The present invention provides a device, method and apparatus for online detection of the state of a solid-liquid system based on a neural network, which is used to solve the technical problems in the prior art of long particle size state calculation time and low accuracy.

[0005] The present invention provides an online detection device for the state of a solid-liquid system based on a neural network, comprising:

[0006] A probe-type ultrasonic particle size analyzer, wherein the probe of the probe-type ultrasonic particle size analyzer is fixed in a test liquid containing a solid-liquid system, and the transmitting end and the receiving end of the probe are both at a preset working distance from the test liquid; both the particle system to be analyzed and the sample particle system can serve as the test liquid, and the probe-type ultrasonic particle size analyzer is used to obtain an ultrasonic attenuation spectrum of the particle system to be analyzed and / or an ultrasonic attenuation spectrum of the particle system sample;

[0007] a processor storing a quantitative calibration model, wherein the processor calculates a particle size state of the particle system to be analyzed using the quantitative calibration model, wherein the particle size state includes a particle size distribution state and / or a particle size uniformity;

[0008] The quantitative correction model is based on a pre-built convolutional neural network and is trained using the ultrasonic attenuation spectra and offline particle size data samples of the particle system samples.

[0009] In some embodiments, the liquid to be measured includes a static solid-liquid system and a dynamic solid-liquid system, and the particle size measurement range of the online detection device for the state of the solid-liquid system is 5 nm-3000 μm.

[0010] In some embodiments, the ultrasonic frequency range of the probe-type ultrasonic particle size analyzer is 1 MHz-100 MHz, the rated voltage is 220 VAC, the rated power is less than or equal to 100 W, and the probe operating temperature is 0°C-90°C.

[0011] The present invention further provides a method for online detection of the state of a solid-liquid system based on a neural network, based on the above-mentioned online detection device for the state of a solid-liquid system, the method comprising:

[0012] Real-time acquisition of ultrasonic attenuation spectra of ultrasonic waves passing through the solid-liquid system to be analyzed;

[0013] Inputting the ultrasonic attenuation spectrum into a pre-trained quantitative calibration model to obtain a state of the solid-liquid system output by the quantitative calibration model, wherein the state of the solid-liquid system includes a particle size distribution state and / or particle size uniformity of the solid-liquid system to be analyzed;

[0014] The quantitative correction model is based on a pre-built convolutional neural network and is trained using the ultrasonic attenuation spectra and offline particle size data samples of the particle system samples.

[0015] In some embodiments, during the training process of the quantitative correction model, principal component analysis is used to perform dimensionality reduction processing on the input data and output data of the convolutional neural network.

[0016] In some embodiments, a pre-built convolutional neural network is trained using the ultrasonic attenuation spectra of the particle system sample and the offline particle size data sample to obtain the quantitative calibration model, specifically comprising:

[0017] Acquiring ultrasonic attenuation spectra and offline particle size data of multiple particle system samples, constructing a data set, and dividing the data set into a training set and a test set;

[0018] The principal component analysis method is used to reduce the dimension of the data in the dataset to obtain two principal elements of the ultrasonic attenuation data and the principal element of the cumulative particle size distribution;

[0019] The two principal components of the ultrasonic attenuation data are used as inputs of a convolutional neural network, and the principal component of the cumulative particle size distribution is used as output of the convolutional neural network for training to obtain the quantitative correction model.

[0020] In some embodiments, obtaining ultrasonic attenuation spectra of multiple particle system samples specifically includes:

[0021] Obtain ultrasonic attenuation spectra of sample suspensions of different concentrations / particle sizes, and each ultrasonic attenuation spectrum data is collected repeatedly;

[0022] The graphical comparison method was used to eliminate obviously abnormal data points in the ultrasonic attenuation spectrum;

[0023] The average value of the remaining data points is calculated to obtain the ultrasonic attenuation spectrum of the particle system sample corresponding to the sampling.

[0024] In some embodiments, the characterizing parameter of the particle size uniformity is the polydispersity index, which is calculated using the following expression:

[0025]

[0026] Among them, D10 refers to the particle size corresponding to when the cumulative distribution of particles reaches 10%; D50 refers to the particle size corresponding to when the cumulative distribution of particles reaches 50%; D90 refers to the particle size corresponding to when the cumulative distribution of particles reaches 90%.

[0027] In some embodiments, the convolutional neural network is a one-dimensional convolutional neural network, and the network structure of the one-dimensional convolutional neural network includes:

[0028] A data input layer, wherein the input size of the data input layer is a 2×1 tensor, representing the attenuation intensity of ultrasound passing through a solid-liquid system at different displacements in the ultrasonic attenuation spectrum;

[0029] Convolutional layer, the convolutional layer uses 3 convolution kernels, the convolution step size is 1, and the sigmoid is used as the activation function;

[0030] Pooling layer, the pooling filter parameter of the pooling layer is 2×1, the step size is 1, and the output matrix is ​​2×1, which is used to calculate the output size of the maximum pooling;

[0031] Fully connected layer, the number of neurons in the fully connected layer is 8, using Sigmoid as the activation function, and the dropout ratio (neuron discarding rate) is 0.2;

[0032] The output layer outputs the size of the predicted variable, corresponding to the predicted particle size distribution of the solid-liquid system.

[0033] The present invention further provides a device for online detection of the state of a solid-liquid system based on a neural network, based on the above-mentioned online detection device for the state of a solid-liquid system, comprising:

[0034] A data acquisition unit, used for acquiring in real time an ultrasonic attenuation spectrum of an ultrasonic wave passing through a solid-liquid system to be analyzed;

[0035] a result generating unit, configured to input the ultrasonic attenuation spectrum into a pre-trained quantitative calibration model to obtain a state of the solid-liquid system output by the quantitative calibration model, wherein the state of the solid-liquid system includes a particle size distribution state and / or particle size uniformity of the solid-liquid system to be analyzed;

[0036] The quantitative correction model is based on a pre-built convolutional neural network and is trained using the ultrasonic attenuation spectra and offline particle size data samples of the particle system samples.

[0037] The neural network-based online detection device for the state of a solid-liquid system provided by the present invention includes a probe-type ultrasonic particle size analyzer and a processor. The probe of the probe-type ultrasonic particle size analyzer is fixed in a particle system sample liquid, and the transmitting end and the receiving end of the probe have a preset working distance from the particle system sample liquid, and are used to obtain an ultrasonic attenuation spectrum and ultrasonic attenuation spectrum samples of the solid-liquid system to be analyzed; the processor trains a correction model using the ultrasonic attenuation spectrum and offline particle size data, and the processor calculates and generates the particle size state of the solid-liquid system using the quantitative correction model, wherein the particle size state of the solid-liquid system includes the particle size distribution state and / or particle size uniformity of the particle system; the quantitative correction model is constructed based on a pre-constructed principal component analysis method and a convolutional neural network combination algorithm training. When performing online detection of the particle size distribution state of particles in a solid-liquid system, an ultrasonic attenuation spectrum of the solid-liquid system to be analyzed is obtained in real time; the ultrasonic attenuation spectrum and offline particle size data are input into a pre-trained calibration model; and the real-time acquired ultrasonic attenuation spectrum is input into the pre-trained quantitative calibration model to obtain the particle size distribution state. This solves the technical problem of delayed measurement results in the prior art and realizes real-time monitoring of the particle size distribution state and uniformity in the solid-liquid system.

[0038] Furthermore, the solution adopted by the present invention is combined with an ultrasonic attenuator, and a new lightweight machine learning model (PCA-CNN) that combines principal component analysis and convolutional neural network is developed. The model has few parameters and can quickly (in a few seconds) achieve accurate measurement of particle size distribution.

[0039] Furthermore, the solution adopted by the present invention is a PCA-CNN model that combines principal component analysis and convolutional neural network. Its parameters depend on the system. Compared with the fixed parameters contained in the ECAH model, the PCA-CNN model can describe the changing state of the system and obtain accurate data on particle size distribution.

[0040] Furthermore, the PCA-CNN model was able to effectively correlate the ultrasonic attenuation spectrum and particle size distribution with a high correlation coefficient (R>0.97). BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 This is one of the flow charts of the neural network-based online detection method for the state of a solid-liquid system provided by the present invention;

[0043] Figure 2 This is the second flow chart of the online detection method for the state of a solid-liquid system based on a neural network provided by the present invention;

[0044] Figure 3 A structural diagram of an online detection device for particle size distribution during the grinding process of hydrotalcite solution provided in Example 1 of the present invention;

[0045] Figure 4 A comparison chart of the offline particle size distribution and the predicted particle size distribution after the training set is standardized in Example 1 of the online detection method for particle size distribution provided by the present invention;

[0046] Figure 5 A comparison chart of the offline particle size distribution and the predicted particle size distribution in Example 1 of the online particle size distribution detection method provided by the present invention;

[0047] Figure 6 This is a structural diagram of an online detection device for the particle size distribution of a silicon dioxide solution provided in Example 2 of the present invention;

[0048] Figure 7 A comparison chart of the offline particle size distribution and the predicted particle size distribution after the training set is standardized in Example 2 of the online detection method for particle size distribution provided by the present invention;

[0049] Figure 8 A comparison diagram of the offline particle size distribution and the predicted particle size distribution of the online particle size distribution detection method provided by the present invention in Example 2;

[0050] Figure 9 This is a structural block diagram of the neural network-based online detection device for the state of a solid-liquid system provided by the present invention. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0052] In order to solve the problems in the prior art that the particle size distribution state and uniformity of particles in a particle system cannot be monitored in real time, and the measurement accuracy needs to be improved, the present invention provides an online detection device, detection method and detection apparatus for the state of a solid-liquid system based on a neural network. This solution can monitor the particle size distribution state and / or distribution uniformity of particles in a particle system in real time. This solution realizes real-time online detection of the particle size distribution state and uniformity of the particle system based on process analysis technology.

[0053] In one specific embodiment, the neural network-based online detection device for the state of a solid-liquid system provided by the present invention includes a probe-type ultrasonic particle size analyzer and a processor; wherein the probe of the probe-type ultrasonic particle size analyzer is fixed in a measured liquid of the solid-liquid system, and the transmitting end and the receiving end of the probe are both at a preset working distance from the measured liquid; both the particle system to be analyzed and the sample particle system can serve as the measured liquid, and the probe-type ultrasonic particle size analyzer is used to obtain an ultrasonic attenuation spectrum of the particle system to be analyzed and / or an ultrasonic attenuation spectrum of a particle system sample. It is understood that when the measured liquid is the particle system to be analyzed, the probe-type ultrasonic particle size analyzer can be used to obtain the ultrasonic attenuation spectrum of the particle system to be analyzed, and when the measured liquid is a particle system sample, the probe-type ultrasonic particle size analyzer can be used to obtain the ultrasonic attenuation spectrum of the particle system sample.

[0054] The processor stores a quantitative correction model, and the processor calculates the particle size state of the particle system to be analyzed through the quantitative correction model, wherein the particle size state includes the particle size distribution state and / or the particle size uniformity; wherein the quantitative correction model is based on a pre-constructed convolutional neural network and is trained using the ultrasonic attenuation spectrum of the particle system sample and the offline particle size data sample.

[0055] That is, the processor trains a correction model through ultrasonic attenuation spectra and offline particle size data, and the processor calculates and generates the particle size state of the solid-liquid system through the quantitative correction model, and the particle size state of the solid-liquid system includes the particle size distribution state and / or particle size uniformity of the solid-liquid system.

[0056] In some embodiments, the measured liquid includes a static solid-liquid system and a dynamic solid-liquid system (grinding process and chemical reaction process), and the particle size measurement range of the online detection equipment for the solid-liquid system state is 5nm-3000μm. The ultrasonic frequency range of the probe-type ultrasonic particle size analyzer is generally between 1MHz-100MHz, and a NanoSon ic probe-type ultrasonic particle size analyzer is preferably used. The acquisition mode is manual acquisition, the rated voltage is 220VAC (220VAC is used as a term in this embodiment, which means that the voltage in the circuit is continuously changing, and its average value is approximately 220V), the rated power is less than or equal to 100W, the probe material is plated with polytetrafluoroethylene or Hastelloy, the probe operating temperature is 0℃-90℃, the probe tube diameter is 20mm, and the ambient temperature is -10℃-40℃.

[0057] In the above-mentioned specific embodiment, the neural network-based online detection equipment for the state of a solid-liquid system provided by the present invention includes a probe-type ultrasonic particle size analyzer and a processor, wherein the probe-type ultrasonic particle size analyzer, the probe of the probe-type ultrasonic particle size analyzer is fixed in a particle system sample liquid, and the transmitting end and the receiving end of the probe have a preset working distance from the particle system sample liquid, and are used to obtain the ultrasonic attenuation spectrum and ultrasonic attenuation spectrum samples of the solid-liquid system to be analyzed; the processor trains a correction model through the ultrasonic attenuation spectrum and offline particle size data, and the processor calculates and generates the particle size state of the solid-liquid system through the quantitative correction model, and the particle size state of the solid-liquid system includes the particle size distribution state and / or particle size uniformity of the particle system; the quantitative correction model is constructed based on the training of a pre-constructed principal component analysis method and a convolutional neural network combination algorithm.

[0058] On the basis of the above-mentioned solid-liquid system state online detection device, the present invention also provides a solid-liquid system state online detection method based on neural network, such as Figure 1 As shown, the method includes:

[0059] S110: Acquire in real time an ultrasonic attenuation spectrum of the ultrasonic wave passing through the solid-liquid system to be analyzed;

[0060] S120: Inputting the ultrasonic attenuation spectrum into a pre-trained quantitative calibration model to obtain a state of the solid-liquid system output by the quantitative calibration model, wherein the state of the solid-liquid system includes a particle size distribution state and / or particle size uniformity of the solid-liquid system to be analyzed;

[0061] The quantitative correction model is based on a pre-built convolutional neural network and is trained using the ultrasonic attenuation spectra and offline particle size data samples of the particle system samples.

[0062] During the training of the quantitative calibration model, principal component analysis (PCA) was used to reduce the dimensionality of the convolutional neural network's input and output data. Specifically, the principal components of the ultrasonic attenuation spectrum samples and offline particle size data samples of the particle system, processed by PCA, were used as the input and output of the convolutional neural network. At any target time, the total number of detections of the ultrasonic attenuation spectrum samples and offline particle size data samples of the solid-liquid system were equal.

[0063] Furthermore, the particle size distribution state of the solid-liquid system is measured based on the ultrasonic attenuation spectrum samples, specifically including: obtaining ultrasonic attenuation spectrum samples at no less than 10 target moments within the target working period; using a pre-stored convolutional neural network combination algorithm to extract the ultrasonic attenuation spectrum samples, and calculating the particle size distribution state of the solid-liquid system.

[0064] In some embodiments, the quantitative calibration model is obtained by training the pre-built convolutional neural network using the ultrasonic attenuation spectra of the particle system sample and the offline particle size data sample, such as Figure 2 As shown, the specific steps include:

[0065] S210: Acquire ultrasonic attenuation spectra and offline particle size data of multiple particle system samples, construct a data set, and divide the data set into a training set and a test set; specifically, obtain ultrasonic attenuation spectra of sample suspensions of different concentrations / particle sizes, and repeatedly collect each ultrasonic attenuation spectrum data; use a graphic comparison method to eliminate obviously abnormal data points in the ultrasonic attenuation spectrum; calculate the average value of the remaining data points to obtain the ultrasonic attenuation spectrum of the particle system sample corresponding to the sampling.

[0066] In a specific usage scenario, the ultrasonic attenuation spectrum refers to the ultrasonic attenuation spectra (no less than 10 images) of sample suspensions of different concentrations / particle sizes obtained by an ultrasonic probe, and each ultrasonic attenuation spectrum data is repeatedly collected 5 times, and then the graphical comparison method is used to eliminate the obviously abnormal data points, and then the average value is calculated. This average value is used as the ultrasonic attenuation spectrum corresponding to this sampling; the offline particle size data refers to the offline particle size distribution data (corresponding to the ultrasonic attenuation spectrum) of sample suspensions of different concentrations / particle sizes obtained by an offline particle size analyzer, and each particle size distribution data is collected 5 times, and the average value is calculated, and the average value is used as the particle size distribution (PSD) corresponding to the ultrasonic attenuation spectrum of this sampling, and then converted into cumulative particle size distribution data; the data set used to establish the quantitative correction model is divided into a training set and a test set according to a ratio of 5:2 (adjustable) by adopting the SPXY automatic grouping method.

[0067] S220: Dimensionality reduction is performed on the data in the dataset using principal component analysis (PCA) to obtain two principal components of the ultrasonic attenuation data and a principal component of the cumulative particle size distribution. Specifically, PCA (principal component analysis) dimensionality reduction is performed, and principal components are selected based on the variance-principal component curve slope method. For the ultrasonic attenuation data, two principal components are selected to explain more than 99% of the original data; for the cumulative particle size distribution data, two principal components are selected to explain more than 99% of the original data.

[0068] S230: Using the two principal components of the ultrasonic attenuation data as inputs to a convolutional neural network and the principal component of the cumulative particle size distribution as outputs of the convolutional neural network for training to obtain the quantitative correction model. Specifically, using the two principal components of the ultrasonic attenuation data as inputs to a convolutional neural network and the principal component of the cumulative particle size distribution as outputs of the convolutional neural network for training to obtain the quantitative correction model.

[0069] In some embodiments, the characterizing parameter of the particle size uniformity is the polydispersity index, which is calculated using the following expression:

[0070]

[0071] Among them, D10 refers to the particle size corresponding to when the cumulative distribution of particles reaches 10%; D50 refers to the particle size corresponding to when the cumulative distribution of particles reaches 50%; D90 refers to the particle size corresponding to when the cumulative distribution of particles reaches 90%.

[0072] Furthermore, the convolutional neural network is a one-dimensional convolutional neural network, and the network structure of the one-dimensional convolutional neural network includes:

[0073] A data input layer, wherein the input size of the data input layer is a 2×1 tensor, representing the attenuation intensity of ultrasound passing through a solid-liquid system at different displacements in the ultrasonic attenuation spectrum;

[0074] Convolutional layer, the convolutional layer uses 3 convolution kernels, the convolution step size is 1, and the sigmoid is used as the activation function;

[0075] Pooling layer, the pooling filter parameter of the pooling layer is 2×1, the step size is 1, and the output matrix is ​​2×1, which is used to calculate the output size of the maximum pooling;

[0076] Fully connected layer, the number of neurons in the fully connected layer is 8, using Sigmoid as the activation function, and the dropout ratio (neuron discarding rate) is 0.2;

[0077] The output layer outputs the size of the predicted variable, corresponding to the predicted particle size distribution of the solid-liquid system.

[0078] In order to verify the technical effect of the solution provided by the present invention, the implementation process of the solution of the present invention is briefly described below using two specific embodiments as examples.

[0079] Example 1:

[0080] like Figure 3 As shown, this embodiment 1 provides a real-time online detection method for particle size distribution during the grinding process of hydrotalcite. The entire experimental device includes a grinder 1, a temperature control system 2, a probe-type ultrasonic particle size analyzer 3, and a computer 4. The probe of the ultrasonic particle size analyzer is fixed in the hydrotalcite solution during the grinding process, and the corresponding ultrasonic attenuation spectra are collected at 0 min, 1 min, 2 min, 3 min, 4 min, 6 min, 8 min, 13 min, 23 min, 33 min, and 53 min of grinding. Since the particle size decreases quickly in the early stage of grinding, the time interval in the early stage of grinding is shorter. The collection method is as follows:

[0081] When the predetermined time (11 groups) is reached, the grinder is paused. Subsequently, the ultrasonic probe in the stirring tank is used to continuously collect the ultrasonic attenuation spectrum of the suspension at a sampling interval of 1 second each time (5 times), and then the graphical comparison method is used to eliminate the obviously abnormal data points, and then the average value is calculated as the ultrasonic attenuation spectrum corresponding to the sampling; at the same time, 3mL of sample is drawn from the stirring tank using a rubber-tipped dropper and dripped into a pre-marked glass vial, and then the grinder is started again to resume the grinding process; each collected ultrasonic attenuation spectrum is marked with the collection time and saved on the computer; the offline particle size distribution data of the 3mL sample taken out is obtained using a Malvern laser particle size analyzer, and each particle size distribution data is collected 5 times, and its mean is calculated, and the mean is used as the particle size distribution corresponding to the ultrasonic attenuation spectrum of this sampling, and then it is converted into cumulative particle size distribution data;

[0082] According to the above method, 10 experiments were carried out and a total of 110 data were collected, corresponding to 110 online acquisitions of ultrasonic attenuation spectra and 110 offline analyses of Malvern laser particle size analyzer; using the SPXY automatic grouping method, 77 data from 7 experiments, namely the 1st, 3rd, 4th, 5th, 7th, 9th, and 10th, were used for modeling, of which 69 data were in the training set and 12 data were in the validation set; 33 data from 3 experiments, namely the 2nd, 6th, and 8th, were used for model verification to test the generalization ability of the model.

[0083] The data was subjected to PCA dimensionality reduction, and the principal component was selected according to the standard of the variance-principal component curve slope method. For the ultrasonic attenuation data and the cumulative particle size distribution data, two principal components were selected, and the explanation of the original data accounted for 99.66% and 99.35% respectively. The two principal components of the ultrasonic attenuation data were used as the input of the convolutional neural network, and the principal component of the cumulative particle size distribution was used as the output of the convolutional neural network to obtain the quantitative correction model; the 33 data of the 2nd, 6th, and 8th experiments that were not used for modeling were input into the quantitative correction model according to the above steps. The final results showed that the particle size distribution calculated by the PCA-CNN model was basically consistent with the particle size distribution measured by the corresponding Malvern laser particle size analyzer. The results are compared in the figure. Figure 4 and Figure 5 .

[0084] Among the evaluation metrics, smaller RMSE (root mean square error) and R (correlation coefficient) values ​​closer to 1 indicate better model performance. For the PCA-CNN model, the RMSE for the validation set was 0.1664 and R was 0.9912; for the test set, the RMSE was 0.2151 and R was 0.9852. These results demonstrate the model's excellent training and computational capabilities, effectively correlating ultrasonic attenuation spectra with particle size distributions with a high correlation coefficient (R > 0.98).

[0085] Example 2:

[0086] like Figure 6 As shown, this embodiment 2 provides a method for online measuring the particle size distribution of SiO2 in the process of hydrolysis and polymerization of ethyl orthosilicate to generate SiO2 particles. The entire experimental device includes a reactor 5, a probe-type ultrasonic particle size analyzer 3, a temperature control system 2 and a computer 4. The probe of the ultrasonic particle size analyzer is fixed in the ethyl orthosilicate solution in the reaction process, and the corresponding ultrasonic attenuation spectra are collected at 10min, 10.5min, 11min, 11.5min, 12min, 12.5min, 13min, 14min, 16min, 18min, 20min, 22min, 24min, 26min, 28min and 30min of the reaction; since the synthesis speed of SiO2 particles is fast in the initial stage of the reaction, the time interval in the initial stage of grinding is shorter. The collection method is as follows:

[0087] As in Example 1, the online ultrasonic attenuation spectrum and offline Malvern particle size analyzer particle size data of the solution in the reactor were collected. By using the SPXY automatic grouping method, 12 data items, including the 10th minute, 11th minute, 11.5th minute, 12.5th minute, 13th minute, 14th minute, 16th minute, 20th minute, 22th minute, 24th minute, 26th minute, and 30th minute data of the 16 sets of online ultrasonic attenuation spectra and 16 sets of Malvern laser particle size analyzer particle size data, were used for modeling, of which 8 data items were used in the training set and 4 data items were used in the validation set. Data from 4 experimental groups, including the 10.5th minute, 12th minute, 18th minute, and 28th minute groups, were used for model verification to test the generalization ability of the model. The data were then subjected to PCA dimensionality reduction. For the ultrasonic attenuation data and cumulative particle size distribution data, two principal components were selected, and the explanation accounted for more than 99% of the original data. The two principal components of the ultrasonic attenuation data were used as the input of the convolutional neural network, and the principal component of the cumulative particle size distribution was used as the output of the convolutional neural network to obtain the quantitative correction model. The data of the 10.5min, 12min, 18min, and 28min groups of experiments that were not used for modeling were input into the quantitative correction model according to the above steps. The final results showed that the particle size distribution calculated by the PCA-CNN model was basically consistent with the particle size distribution measured by the corresponding Malvern laser particle size analyzer. The comparison diagram is shown in FIG. Figure 7 and Figure 8 , and the PCA-CNN model can effectively correlate the ultrasonic attenuation spectrum and particle size distribution with a high correlation coefficient (R>0.97).

[0088] In the above specific embodiment, when the particle size distribution state of particles in a solid-liquid system is detected online, the ultrasonic attenuation spectrum of the solid-liquid system to be analyzed is obtained in real time; the ultrasonic attenuation spectrum and offline particle size data are input into a pre-trained correction model; the real-time acquired ultrasonic attenuation spectrum is input into the pre-trained quantitative correction model to obtain the particle size distribution state; thereby solving the technical problem of delayed measurement results in the prior art and realizing real-time monitoring of the particle size distribution state and uniformity in the solid-liquid system.

[0089] Furthermore, the solution adopted by the present invention is combined with an ultrasonic attenuator, and a new lightweight machine learning model (PCA-CNN) that combines principal component analysis and convolutional neural network is developed. The model has few parameters and can quickly (in a few seconds) achieve accurate measurement of particle size distribution.

[0090] Furthermore, the solution adopted by the present invention is a PCA-CNN model that combines principal component analysis and convolutional neural network. Its parameters depend on the system. Compared with the fixed parameters contained in the ECAH model, the PCA-CNN model can describe the changing state of the system and obtain accurate data on particle size distribution.

[0091] Furthermore, the PCA-CNN model was able to effectively correlate the ultrasonic attenuation spectrum and particle size distribution with a high correlation coefficient (R>0.97).

[0092] In addition to the above method, the present invention also provides a device for online detection of the state of a solid-liquid system based on a neural network. Figure 9 As shown, the device includes:

[0093] The data acquisition unit 910 is used to acquire in real time an ultrasonic attenuation spectrum of an ultrasonic wave passing through a solid-liquid system to be analyzed;

[0094] a result generating unit 920 for inputting the ultrasonic attenuation spectrum into a pre-trained quantitative calibration model to obtain a state of the solid-liquid system output by the quantitative calibration model, wherein the state of the solid-liquid system includes a particle size distribution state and / or particle size uniformity of the solid-liquid system to be analyzed;

[0095] The quantitative correction model is based on a pre-built convolutional neural network and is trained using the ultrasonic attenuation spectra and offline particle size data samples of the particle system samples.

[0096] In some embodiments, during the training process of the quantitative correction model, principal component analysis is used to perform dimensionality reduction processing on the input data and output data of the convolutional neural network.

[0097] In some embodiments, a pre-built convolutional neural network is trained using the ultrasonic attenuation spectra of the particle system sample and the offline particle size data sample to obtain the quantitative calibration model, specifically comprising:

[0098] Acquiring ultrasonic attenuation spectra and offline particle size data of multiple particle system samples, constructing a data set, and dividing the data set into a training set and a test set;

[0099] The principal component analysis method is used to reduce the dimension of the data in the dataset to obtain two principal elements of the ultrasonic attenuation data and the principal element of the cumulative particle size distribution;

[0100] The two principal components of the ultrasonic attenuation data are used as inputs of a convolutional neural network, and the principal component of the cumulative particle size distribution is used as output of the convolutional neural network for training to obtain the quantitative correction model.

[0101] In some embodiments, obtaining ultrasonic attenuation spectra of multiple particle system samples specifically includes:

[0102] Obtain ultrasonic attenuation spectra of sample suspensions of different concentrations / particle sizes, and each ultrasonic attenuation spectrum data is collected repeatedly;

[0103] The graphical comparison method was used to eliminate obviously abnormal data points in the ultrasonic attenuation spectrum;

[0104] The average value of the remaining data points is calculated to obtain the ultrasonic attenuation spectrum of the particle system sample corresponding to the sampling.

[0105] In some embodiments, the characterizing parameter of the particle size uniformity is the polydispersity index, which is calculated using the following expression:

[0106]

[0107] Among them, D10 refers to the particle size corresponding to when the cumulative distribution of particles reaches 10%; D50 refers to the particle size corresponding to when the cumulative distribution of particles reaches 50%; D90 refers to the particle size corresponding to when the cumulative distribution of particles reaches 90%.

[0108] In some embodiments, the convolutional neural network is a one-dimensional convolutional neural network, and the network structure of the one-dimensional convolutional neural network includes:

[0109] A data input layer, wherein the input size of the data input layer is a 2×1 tensor, representing the attenuation intensity of ultrasound passing through a solid-liquid system at different displacements in the ultrasonic attenuation spectrum;

[0110] Convolutional layer, which uses 3 convolution kernels, a convolution step size of 1, and Sigmoid as the activation function;

[0111] Pooling layer, the pooling filter parameter of the pooling layer is 2×1, the step size is 1, and the output matrix is ​​2×1, which is used to calculate the output size of the maximum pooling;

[0112] Fully connected layer, the number of neurons in the fully connected layer is 8, Sigmo id is used as the activation function, and the dropout ratio (neuron discarding rate) is 0.2;

[0113] The output layer outputs the size of the predicted variable, corresponding to the predicted particle size distribution of the solid-liquid system.

[0114] In the above specific embodiment, when the particle size distribution state of particles in a solid-liquid system is detected online, the ultrasonic attenuation spectrum of the solid-liquid system to be analyzed is obtained in real time; the ultrasonic attenuation spectrum and offline particle size data are input into a pre-trained correction model; the real-time acquired ultrasonic attenuation spectrum is input into the pre-trained quantitative correction model to obtain the particle size distribution state; thereby solving the technical problem of delayed measurement results in the prior art and realizing real-time monitoring of the particle size distribution state and uniformity in the solid-liquid system.

[0115] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0116] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An online detection device for the state of a solid-liquid system based on a neural network, characterized in that: include: A probe-type ultrasonic particle size analyzer, wherein the probe of the probe-type ultrasonic particle size analyzer is fixed in a measured liquid containing a solid-liquid system, and both the transmitting end and the receiving end of the probe have a preset working distance from the measured liquid; both the particle system to be analyzed and the sample particle system can be used as the measured liquid, and the probe-type ultrasonic particle size analyzer is used to obtain an ultrasonic attenuation spectrum of the particle system to be analyzed and / or an ultrasonic attenuation spectrum of a particle system sample; A processor, wherein a quantitative calibration model is stored in the processor, and the processor calculates the particle size state of the particle system to be analyzed by using the quantitative calibration model, wherein the particle size state includes a particle size distribution state and / or a particle size uniformity; Wherein, the quantitative correction model is based on a pre-constructed convolutional neural network, and is trained using the ultrasonic attenuation spectra of the particle system samples and offline particle size data samples.

2. The neural network-based on-line detection device for solid-liquid system state according to claim 1, characterized in that: The liquid to be tested includes a static solid-liquid system and a dynamic solid-liquid system. The particle size measurement range of the online detection equipment for the state of the solid-liquid system is 5nm-3000μm.

3. The neural network-based on-line detection device for solid-liquid system state according to claim 1, characterized in that: The ultrasonic frequency range of the probe ultrasonic particle size analyzer is 1MHz-100MHz, the rated voltage is 220VAC, the rated power is less than or equal to 100W, and the probe operating temperature is 0℃-90℃.

4. A method for online detection of the state of a solid-liquid system based on a neural network, based on the online detection device for the state of a solid-liquid system according to any one of claims 1 to 3, characterized in that: The method comprises: Real-time acquisition of ultrasonic attenuation spectra of ultrasonic waves passing through the solid-liquid system to be analyzed; Inputting the ultrasonic attenuation spectrum into a pre-trained quantitative calibration model to obtain a state of a solid-liquid system output by the quantitative calibration model, wherein the state of the solid-liquid system includes a particle size distribution state and / or particle size uniformity of the solid-liquid system to be analyzed; Wherein, the quantitative correction model is based on a pre-constructed convolutional neural network, and is trained using the ultrasonic attenuation spectra of the particle system samples and offline particle size data samples.

5. The method for online detection of the state of a solid-liquid system based on a neural network according to claim 4, characterized in that: During the training process of the quantitative correction model, the principal component analysis method is used to perform dimensionality reduction processing on the input data and output data of the convolutional neural network.

6. The method for online detection of the state of a solid-liquid system based on a neural network according to claim 4, characterized in that: Based on a pre-built convolutional neural network, the ultrasonic attenuation spectrum of the particle system sample and the offline particle size data sample are used for training to obtain the quantitative correction model, which specifically includes: Acquire ultrasonic attenuation spectra and offline particle size data of multiple particle system samples, construct a data set, and divide the data set into a training set and a test set; The principal component analysis method is used to reduce the dimension of the data in the data set to obtain two principal components of the ultrasonic attenuation data and the principal component of the cumulative particle size distribution; The two principal components of the ultrasonic attenuation data are used as the input of the convolutional neural network, and the principal component of the cumulative particle size distribution is used as the output of the convolutional neural network for training to obtain the quantitative correction model.

7. The method for online detection of the state of a solid-liquid system based on a neural network according to claim 6, characterized in that: Obtain ultrasonic attenuation spectra of multiple particle system samples, including: Obtain ultrasonic attenuation spectra of sample suspensions of different concentrations / particle sizes, and each ultrasonic attenuation spectrum data is collected repeatedly; The graphical comparison method was used to remove obviously abnormal data points in the ultrasonic attenuation spectrum; The average value of the remaining data points is calculated to obtain the ultrasonic attenuation spectrum of the particle system sample corresponding to the sampling.

8. The method for online detection of the state of a solid-liquid system based on a neural network according to claim 4, characterized in that: The characterization parameter of the particle size uniformity is the polydispersity coefficient, which is calculated using the following expression: Among them, D10 refers to the particle size corresponding to when the cumulative distribution of particles reaches 10%; D50 refers to the particle size corresponding to when the cumulative distribution of particles reaches 50%; D90 refers to the particle size corresponding to when the cumulative distribution of particles reaches 90%.

9. The method for online detection of drug coating status according to any one of claims 4 to 8, characterized in that: The convolutional neural network is a one-dimensional convolutional neural network, and the network structure of the one-dimensional convolutional neural network includes: A data input layer, wherein the input size of the data input layer is a tensor of 2×1, representing the attenuation intensity of ultrasound waves passing through a solid-liquid system at different displacements in the ultrasonic attenuation spectrum; Convolution layer, the convolution layer uses 3 convolution kernels, the convolution step size is 1, and Sigmoid is used as the activation function; A pooling layer, wherein the pooling filter parameter of the pooling layer is 2×1, the step size is 1, and the output matrix is ​​2×1, which is used to calculate the output size of the maximum pooling; Fully connected layer, the number of neurons in the fully connected layer is 8, Sigmoid is used as the activation function, and the neuron drop rate is 0.2; The output layer has an output quantity which is the size of the predicted variable, corresponding to the predicted particle size distribution of the solid-liquid system.

10. An online detection device for the state of a solid-liquid system based on a neural network, based on the online detection device for the state of a solid-liquid system according to any one of claims 1 to 3, characterized in that: The device comprises: A data acquisition unit, used for acquiring in real time an ultrasonic attenuation spectrum of an ultrasonic wave passing through a solid-liquid system to be analyzed; A result generating unit, used for inputting the ultrasonic attenuation spectrum into a pre-trained quantitative calibration model to obtain a state of a solid-liquid system output by the quantitative calibration model, wherein the state of the solid-liquid system includes a particle size distribution state and / or particle size uniformity of the solid-liquid system to be analyzed; Wherein, the quantitative correction model is based on a pre-constructed convolutional neural network, and is trained using the ultrasonic attenuation spectra of the particle system samples and offline particle size data samples.

Images