Lumisphere polystyrene fluorescent microspheres from Tianjin Bessler Chromatography Technology Development Center (Tianjin, China) were used with a concentration of 25 mg/ml. The microscopic images of the microparticles with three different sizes were taken by scanning electron microscopy (SEM), as shown in Figure 2. The testing samples of different concentrations were prepared using series dilution to find the optimum concentration for acquiring data. The optimum concentration of the testing sample was obtained by getting the highest number of peaks/variations in the acquired waveform of respective sample. The testing samples with concentrations of 0.0125, 0.025, 0.05, 0.075, 0.1, and 0.125 mg/ml were mixed with DI water to acquire the optimal concentration. The solution was vortexed and centrifuged at 1,000 rpm for 10 min. The residue was removed from the sample to remove unwanted impurities. Then 1 ml of DI water was mixed with the supernatant to prepare the final sample for the experiment. All the experiments were performed at a room temperature (25–28°C).

Mie scattering theory describes the elastic scattering of light when the wavelength of the incident light is similar or smaller than the diameter size of spherical particles. Mie theory has been widely applied to laser diffraction analysis to detect particle size effects (Altug et al., 2022). MDLS combines the angular information of Mie scattering with dynamic light scattering to measure particle size. MDLS shows that the intensity of the scattered light fluctuates in a particular direction with time because the tiny particles in solution are in Brownian motion, and the distance between each scattered particle constantly changes with time. MDLS is a common method for detecting particle size and has been used in many medical detection devices (Guo et al., 2018). Optical methods were widely used as an important research tool for particle classification and qualitative detection (Fan et al., 2014; Olson et al., 2015).

The prototype contains three main parts: 1) a laser source, 2) an assembly of the photosensitive sensor, and 3) a data acquisition system (Hussain et al., 2020a). The designed prototype and the laser hardware were assembled by the Nanjing Institute of Advanced Laser Technology, Chinese Academy of Sciences (Nanjing, China). The laser source of the device has a wavelength of 660 nm at a rated power of 150 mW. The power rating of the laser source was measured by PM320E and S130C instruments developed by Thorlabs (New Jersey, United States). When the laser passes through the sample, the scattering light was detected by the high-speed silicon photodiodes FDS100 (wavelength ranges from 350 to 1,100 nm with rise time of 10 ns) manufactured by Thorlabs (New Jersey, United States). The AD8675 manufactured by Analog Devices (Massachusetts, United States) operational amplifier was used in the system to amplify the weak signal. A small flask made of round bottom silicon was used in the experiment (Celo Measure & Control Technology Co., Hefei). UV-1800 Spectrophotometer (SHIMADZU, Japan) measured the absorbance of the glass (Figure 3), which was approximately zero between 500–700 nm wavelength.

The original prototype was designed with 32 photoelectric sensors for signal acquisition (Hussain et al., 2020b). However, the initial number of 32 sensors consumed more computational power. The redundant data acquisition channels did not provide enough information, significantly deteriorating the classification results. The intensity of scattering light depends on the shape, size, and characteristics of the particles based on the Mie scattering theory. Particles of different sizes exhibited different scattered light patterns. The maximum scattering light intensity occurs in the incoming light’s forward direction (Lock and Laven, 2011). Therefore, the number of sensors was tested and reduced from 32 to 12.

The 3D assembly of the developed prototype was designed using Solidworks 2020 software, as shown in Figure 4. The 3-axis positioning table controlled the “XYZ” 3-axial alignment of the light source to ensure that the laser is focused on the center of the sample. The laser beam positioning unit, laser collimator, sample flask and the direction of the incident light to the sample were optimized to acquire maximum signal energy from the forward scattering light. The calibrated system guarantees that no interference from external factors appears throughout the detection experiments. The signals were collected by NI data acquisition card (PCI-6225). The system showed zero voltage in the dark environment. The data collector maintained the calibrated zero signal waveform when the flask was filled with DI water as an empty sample. The collected data were further processed for features evaluation and data classification using MatlabR 2018a software.

Machine learning represents a class of algorithms for data classification. With the development of artificial intelligence, machine learning technology has been widely applied in various fields, especially in biomedical engineering and life science (Goodswen et al., 2021). Among different machine learning algorithms, LR is a probabilistic non-linear regression with applications in prediction and discrimination. RF is an integrated learning model that contains multiple decision trees, and the output category is determined by the plurality of the categories output from the individual trees. SVM is a typical supervised learning classification algorithm in machine learning, which classifies data by finding the maximum interval in the feature space (Ghannam and Techtmann, 2021).

After acquiring the experimental data, the machine learning classification algorithms were applied based on Sklearn 0.19.0 package (Intel Core i7-10700 CPU processor, 2.90 Ghz, 32 GB RAM, RXT 3070). The SVM classifier acquired the highest classification accuracy compared with other classifiers. The machine learning classification process is presented in Figure 5, where the data set was divided into a training set and a test set in the ratio of 7:3. After denoising, the data was segmented by adding a sliding window length of 250 ms and slid over the data signal with an increment of 150 ms. The power spectrum features were extracted for the signals in each window. The features extracted from three particle sizes of 1, 2, and 4 µm were labeled before training the machine learning classifier.

The prototype was successfully developed with a dimension of 50 cm × 30 cm × 30 cm. The testing sample was placed in the chamber of the prototype, as shown in Figure 6A. The sample was placed to focus the laser light on the center of the flask, as shown in Figure 6B. The particles that randomly moved in the flask scattered the detection beam while the sensor received the scattered signal in real-time. The pattern of scattering light was affected by the size and shape of the microparticles (Hussain et al., 2019). A black box was used to avoid external interference, improving the quality of the acquired signal, as shown in Figure 6C.

The testing samples with concentrations of 0.0125, 0.025, 0.05, 0.075, 0.1, and 0.125 mg/ml were mixed with DI water to acquire the optimal concentration. Each of the experiments was performed for 30 min. Samples with different concentrations were tested to obtain the average number of peaks (Figure 7). The scattering of the light was weak when the concentration was too low, so the number of detected peaks was less. At higher concentrations, the number of detected characteristic peaks was also too low due to particle-to-particle interaction. Overall, the experimental results showed that a higher number of peaks in the waveform were generated at a sample concentration of 0.025 mg/ml. The classification outcomes showed incorrect results for samples with concentrations below 0.025 mg/ml. Therefore, the sample concentration of 0.025 mg/ml is considered as the detection limit.

The prepared samples with an optimum concentration were used to detect the MDLS. The time-domain light scattering signals from three different particles were obtained for 30 min. The SNR of the signal increased from 2.98 to 6.698 after applying the second order Butterworth filter to the acquired raw data. Figures 8A–C shows the test signal obtained from 12 channels for three particles with time duration of 3 min. An output signal shows significant variations in the peak values obtained from different particles sizes. The peak values represent the time when the particles pass through the detection beam. The signal test significantly revealed the similarity of the particle signals of the three different particle sizes in terms of peak features and combinations of detection channels, so the feature extraction and machine learning approach can significantly improve the effectiveness of the classification task.

PCA and RFE were used as feature selection algorithms, and then 5-fold cross-validation was applied to the features. Before training and testing the classifiers, the features matrix were labeled as 1 µm (class 0 label), 2 µm (class 1 label), and 4 µm (class 2 label). The number of features varied from 5 to 60, with an increment of 5 in each iteration. A 5-fold cross-validation evaluated the results of each feature selection. The highest accuracies of LR, RF, and SVM classifiers using 50 selected features with PCA were 91.97%, 88.89%, and 91.74%, respectively (Figure 9A). Similarly, the highest accuracies of LR, RF classifiers were 92.08%, 91.52%, using 30 selected features, the highest accuracy of SVM classifier were 95.38% using 50 selected features with RFE, (Figure 9B). Overall, the classification results using RFE were more accurate than PCA. The RFE feature selection method selected a subset of 50 features to obtain the highest classification accuracy.

The confusion matrix was plotted from the outcomes to evaluate the performance of the three classifiers. The row of the confusion matrix represented the actual sample classes, and the column represented the predicted sample classes (Figure 10). The green boxes in the diagonal line represented the outcomes that were correctly classified. The remaining feature data points represented the incorrectly classified values. The identification accuracy, precision, recall, and F1 score of the testing dataset were evaluated using Eqs 14–17, and the results are presented in Table 1. Overall, the SVM(kernel = “linear”, decision_function_shape = “ovr”,C = 100, gamma = 0.0001, probability = True) classifier achieved higher classification accuracy than LR and RF classifiers.

The Shapley values were used to validate whether the 50 power spectrum features subset influenced the prediction results. We computed the Shapely values that were contributed from the power spectral features of the SVM classification model. The average Shapely values from each feature and the corresponding Shapley values for each data point were calculated and counted. The bar chart on the left (Figure 11A) showed the average Shapley values for the three classifiers, indicating each feature’s average contribution to the final output. The scatter plot on the right (Figure 11B) demonstrated the distribution of Shapley values and their contribution to the model output structure. The color of each point represented the intensity of the feature value. Blue data points indicated low features, while red represented the opposite. The results showed a significant contribution of the first 12 features to the results, consistent with the proposed channel arrangement for the acquisition system. The first 12 features that corresponded to the zero-order moment features of the 12 channels played a crucial role in the contribution of the classification results.