Particle identification and detection method based on deep learning improved algorithm

Abstract

The present application relates to a kind of particle identification and detection method based on deep learning improved algorithm, it includes the following steps: using electric sensitive area method obtains the electric signal of multiple particle;Extract the time domain, frequency domain and time-frequency domain characteristic parameters of electric signal;Characteristic parameters are combined to enhance the difference between characteristics;Screening characteristic parameters to reduce the amount of operation;Neural network and recurrent neural network model are weighted fusion;Using fusion model to test multiple repeat for the particle to be measured, using voting mechanism to determine the final classification result;The concentration, quantity proportion and particle size distribution of each kind of particle are output.The particle identification and detection method provided by the present application can realize the measurement of particle type, concentration and particle size distribution in water, improve the accuracy, accuracy and stability of water quality monitoring, provide more comprehensive information for water plant operation and environmental monitoring, so that the particle identification and detection method can be better applied to practical application.

Description

Technical Field

[0001] This invention relates to methods for identifying and detecting particulate matter, and particularly to methods for identifying and detecting particulate matter in water bodies based on improved deep learning algorithms. Background Technology

[0002] In aquatic environments, various forms of particulate matter are involved, including organic flocs, inorganic particles, plastic particles, bubbles, and microbial cells. Among these, the particle size distribution of organic flocs affects their dewatering performance. Excessively high concentrations of inorganic particles can impair the acidification and methanogenesis processes of sludge. Plastic particles, within a certain concentration range, can inhibit the anaerobic digestion of sludge. Smaller bubble sizes result in higher oxygen mass transfer efficiency and more thorough aeration. The quantity of microbial cells reflects the quality of the wastewater treatment process and its effectiveness. Therefore, monitoring the concentration and particle size distribution of these particulate matter is crucial for ensuring the operation of wastewater treatment plants and for environmental monitoring.

[0003] However, in real-world environments, particulate matter is often mixed and difficult to distinguish, increasing the difficulty of monitoring. Currently, microparticle size monitoring technologies can be categorized into optical (such as laser diffraction and dynamic light scattering), acoustic (such as ultrasonic particle size analysis), imaging, and electrical (electrical sensitive area method). Although these technologies each have their own characteristics, they all struggle to achieve online monitoring and particle identification, and cannot fully reflect the true state of particulate matter, including its type and proportion.

[0004] The monitoring and control of wastewater treatment plants will evolve towards intelligence, automation, and digitalization in the future. To achieve this goal, the monitoring of particulate matter in water needs to be more microscopic and precise, requiring improved identification and quantitative analysis capabilities for pollutants and microorganisms. Only on this basis can cloud computing, big data, and other technologies be used to comprehensively monitor and manage wastewater treatment plant operations. Therefore, developing a technology capable of monitoring the types and size distribution of particulate matter during wastewater treatment processes is imperative.

[0005] Patent application No. 202311529386.0, entitled "A Particulate Matter Detection Technology Based on Electrosensitive Area Method and Machine Learning," proposes a method for identifying inorganic particles, plastic particles, microorganisms, and air bubbles. However, existing identification algorithms still need improvement: First, accuracy needs to be enhanced, especially in distinguishing between plastic particles and microorganisms, which are easily confused; second, when certain classes are missing in the test data, the model misidentifies a large number of samples as belonging to the missing classes. The root of these problems lies in the characteristic of different categories of particulate matter exhibiting "large intra-class differences and small inter-class differences," meaning that the data differences within a category are significant, while the differences between different categories are relatively small. This leads to overlap and intersection of the feature parameters of different categories of particulate matter in the feature space. Therefore, it is urgent to optimize these issues to improve the accuracy and stability of particulate matter monitoring methods. Summary of the Invention

[0006] In view of this, it is indeed necessary to provide a particulate matter identification and detection method based on an improved deep learning algorithm, which can improve the accuracy and stability of particulate matter monitoring methods.

[0007] A particulate matter identification and detection method based on a deep learning-based improved algorithm includes the following steps: Step 1, obtaining electrical signals of multiple types of particulate matter using the electrical sensitive area method; Step 2, extracting time-domain, frequency-domain, and time-frequency-domain feature parameters of the electrical signals; Step 3, combining the feature parameters to enhance the differences between features; Step 4, filtering feature parameters to reduce computational load; Step 5, weighted fusion of neural network and recurrent neural network models; Step 6, repeatedly testing the particulate matter under test using the fusion model, and using a voting mechanism to determine the final classification result; Step 7, outputting the concentration, quantity percentage, and particle size distribution of each type of particulate matter.

[0008] Compared with existing technologies, the particulate matter identification and detection method based on the improved deep learning algorithm of this invention can accurately identify inorganic particles, biological cells, bubbles, plastic particles, and organic flocs in water. It can also classify other types of particulate matter in water as "unknown," effectively avoiding interference from other unknown particles in the water. It can measure the type, concentration, and particle size distribution of particulate matter in water, improve the accuracy, precision, and stability of water quality monitoring, provide more comprehensive information for water plant operation and environmental monitoring, and enable the particulate matter identification and detection method to be better applied in practical applications. Attached Figure Description

[0009] Figure 1 This is an overall flowchart of the particulate matter identification and detection method based on the improved deep learning algorithm in this embodiment of the invention.

[0010] Figure 2 This is a comparison chart of the electrical signals of five types of particulate matter obtained in step 1 of this embodiment of the invention: Figure 2 (a) is a diagram of the electrical signals of inorganic particles. Figure 2 (b) is the electrical signal diagram of organic flocs. Figure 2 (c) is a graph of the electrical signals of microbial cells. Figure 2 (d) is the electrical signal diagram of plastic particles. Figure 2 (e) is a diagram of the bubble electrical signal.

[0011] Figure 3 A scatter plot showing the two characteristic parameters, 0.1 times peak-to-peak pulse width and rise time, is presented.

[0012] Figure 4 A scatter plot showing the two characteristic parameters of 0.1 times peak-to-peak pulse width and time product is displayed.

[0013] Figure 5This is a scatter plot of the two characteristic parameters: kurtosis and wavelet transform standard deviation.

[0014] Figure 6 This is a scatter plot of the two characteristic parameters: the time-domain product and the standard deviation of the wavelet transform.

[0015] Figure 7 This is the training result of the NN model in step 5 of the embodiment of the present invention.

[0016] Figure 8 This is the training result of the RNN model in step 5 of the embodiment of the present invention.

[0017] Figure 9 This is a diagram showing the model fusion effect in step 5 of this embodiment of the invention.

[0018] Figure 10 This is a display diagram showing the output of various particulate matter concentrations and percentages in step 7 of this embodiment of the invention.

[0019] Figure 11 This is a display diagram showing the particle size distribution of various particulate matter output in step 7 of the embodiment of the present invention.

[0020] Explanation of main component symbols

[0021] none

[0022] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0023] The following will provide a more detailed description of the particulate matter identification and detection method based on the improved deep learning algorithm of the present invention, with reference to the accompanying drawings.

[0024] Please see Figure 1 This invention provides a particulate matter identification and detection method based on an improved deep learning algorithm, comprising the following steps:

[0025] Step 1: Obtain electrical signals of various types of particulate matter using the electrical sensing area method;

[0026] Step 2: Extract the time-domain, frequency-domain, and time-frequency-domain feature parameters of the electrical signal;

[0027] Step 3: Combine feature parameters to enhance the differences between features;

[0028] Step 4: Filter feature parameters to reduce computational load;

[0029] Step 5: Perform a weighted fusion of the neural network (NN) and recurrent neural network (RNN) models;

[0030] Step 6: Use the fusion model to perform multiple repeated tests on the particulate matter to be tested, and use a voting mechanism to determine the final classification result;

[0031] Step 7: Output the concentration, quantity percentage, and particle size distribution of various particulate matter.

[0032] Specifically, this invention provides a particulate matter identification and detection method based on an improved deep learning algorithm, comprising the following steps:

[0033] In step 1, the raw signals of various particulate matter are obtained using the electrosensitive area method. Specifically, the particulate matter to be tested is added to physiological saline and stirred at 300 rpm to ensure uniform suspension. The collection tube is positioned below the collection probe, and a pipette is built into the tube for liquid aspiration, with the flow rate maintained at a stable 3 mL / min. To facilitate effective collection of particulate matter, the orifice diameter of the collection tube is set to 200 μm. A 10V DC voltage is applied to the inner and outer electrodes of the collection tube. When particulate matter flows through the orifice with the liquid, an electrical signal is generated. This electrical signal is transmitted to the acquisition card via circuitry for acquisition, with the acquisition frequency set to 200 kHz and displayed on the monitoring screen.

[0034] The electrical signals of various particulate matter were extracted using wavelet denoising and peak localization algorithms. Specifically, the acquired signals were decomposed into five levels of wavelets, and the modulus squared threshold function was calculated using the first four levels of detail components to achieve effective denoising, primarily removing high-frequency noise. Furthermore, a digital morphological filtering method was employed to accurately preserve the positive impact waveform and replace the negative impact waveform with a straight line, resulting in significant noise reduction, effective removal of low-frequency noise, and a smoother baseline. The peak search function (findpeaks) was used to locate the peaks, and 100 data points were extracted from both sides of each peak and saved, ultimately yielding the following result: Figure 2 The electrical signals of various particulate matter are shown.

[0035] In step 2, time-domain, frequency-domain, and time-frequency-domain features are extracted from the particulate matter electrical signal. The time-domain features include 12 features: pulse width (including 0.1x and 0.2x peak pulse width), rise time, fall time, rise kurtosis, variance, maximum slope, minimum slope, root mean square (RMS), skewness, and shape parameters. The frequency-domain features include 9 features: wavelet transform integral, wavelet transform standard deviation, wavelet transform skewness, kurtosis, fifth-level intermediate frequency energy of wavelet decomposition, centroid frequency, RMS frequency, average frequency, and frequency variance. The time-frequency-domain features include 3 features: signal entropy, spectral kurtosis, and energy spectral density.

[0036] In step 3, the features extracted in step 2 are combined and transformed to increase the diversity among the feature parameters. For example... Figure 3The scatter plots show the pulse width at 0.1 times the peak value and the rise time, with the horizontal and vertical axes representing the range of these characteristic parameters after standardization. It was observed that, except for bubble particles, other particulate matter data exhibited significant overlap, and the characteristic parameter values ​​for the same data type ranged widely, showing a pattern of "large intra-class differences and small inter-class differences." Figure 4 The diagram shows a scatter plot of the two characteristic parameters: 0.1 times the peak-to-peak pulse width and the time product. It can be observed that after introducing the time product feature, similar data are more clustered, and intra-class differences are significantly reduced.

[0037] like Figure 5 The scatter plot of the two characteristic parameters, kurtosis and wavelet transform standard deviation, shows that the data of the five types of particulate matter are randomly distributed and have a large degree of overlap. Figure 6 This is a scatter plot of the two feature parameters: the time-domain product and the standard deviation of the wavelet transform. It can be seen that after introducing the new feature of the time-domain product, the distribution of various particle types in the feature parameter space is more concentrated, the inter-class differences are significantly enhanced, and the data types are more clearly separated.

[0038] Therefore, in step 3, new feature parameters "time product" and "time domain product" are constructed, where the time product is defined as the rise time multiplied by the fall time, and the time domain product is defined as the product of variance, skewness and kurtosis.

[0039] In step 4, the results of the MRMR and ReliefF algorithms are shown in Table 1, which displays the top ten features in terms of importance. Based on the feature importance ranking and integrating the results of the two algorithms, the average frequency feature was added in addition to the labeled features. Finally, a total of 14 features were selected for model training: 0.1 times pulse width, rise time, time product, variance, maximum slope, minimum slope, skewness, wavelet transform standard deviation, wavelet transform skewness, kurtosis, time-domain product, high-frequency energy magnitude of the fifth layer in the wavelet domain, average frequency, and spectral kurtosis.

[0040] Table 1 shows the results of feature selection in step 4.

[0041]

[0042] Step 5: Select RNN and NN models for training. For the RNN model, the parameters are set as follows: 51 neurons, input data shape (14,1), indicating a length of 14 for each input sequence, and 1 feature per time step. The fully connected layer has an output dimension of 5, using softmax as the activation function to obtain the probability distribution of the output class in a multi-class classification problem. The loss function is sparse classification cross-entropy, suitable for multi-class problems with integer labels. The optimizer is the Adam algorithm, commonly used for stochastic gradient descent optimization. The evaluation metric is accuracy. For the NN model, the hidden layer structure consists of three layers with 100, 20, and 30 neurons respectively. The regularization parameter is set to 0.001 to control model complexity and avoid overfitting. The activation function for the hidden layers is ReLU. The maximum number of iterations is set to 1000. Again, accuracy is selected as the evaluation metric.

[0043] After model training, the RNN and NN models are tested using a test set. The classification probability of each model for a single sample is calculated, and the class with the highest probability is used as the prediction result for that sample. The test set includes five categories of particulate matter: inorganic particles, microbial cells, bubbles, plastic particles, and organic flocs. The results of both models on the test set, including test set accuracy and confusion matrix, are output for subsequent comparative analysis. A weighted mechanism is used to fuse the classification probabilities of the two models. The specific weighting rules are as follows: if the NN model predicts the sample as plastic particles or organic flocs, its weight is set to 0.55; if the RNN model predicts the sample as inorganic particles or microbial cells, its weight is set to 0.60; when the above rules conflict, the original weights are used: 0.55 for RNN and 0.45 for NN. After weighting the classification probabilities of the two models, if both classification probabilities are less than 40%, the sample label is output as 5, representing the unknown class.

[0044] like Figure 7 , Figure 8 as well as Figure 9 This paper compares the recognition performance of three models for five types of particles. The neural network (NN) model performed well in recognizing bubbles, plastic particles, and organic flocs, but performed poorly in recognizing inorganic particles and microorganisms, especially with a false negative rate of 27.5% for inorganic particles. In contrast, the reactive neural network (RNN) model performed better in recognizing inorganic particles and microorganisms, but performed poorly in recognizing plastic particles and organic flocs, especially with a false negative rate as high as 44.8% for organic flocs. Model fusion combined the advantages of both models, resulting in a decrease in the false negative rate for all types of particles, achieving the best performance. The accuracy reached 88.1%. Furthermore, in terms of model accuracy, the accuracy rates of the three models were 87.0%, 85.0%, and 88.1%, respectively, with model fusion achieving the highest accuracy. Therefore, this further demonstrates that model fusion achieves better recognition results.

[0045] Step 6: The electrical signal of the particulate matter to be detected is acquired using the electrical sensing area method, and the 14 feature parameters defined in Step 4 are extracted to form a test set. Then, a fusion model is used to perform n classification tests on the test set. A voting mechanism is employed to statistically analyze the results of the n tests to determine the final test result. Specifically, the category that appears most frequently in the n results is selected as the final test result; if the frequency of a category is less than half, the sample is identified as an unknown class.

[0046] In this embodiment, a fusion model is used to perform ten classification tests on the test set. A voting mechanism is employed to statistically analyze the results of the ten tests, and the category that appears most frequently in the ten results is selected as the final identification result; if the category appears less than five times, the sample is identified as an unknown class.

[0047] It is understandable that there is no limit to the number of times the fusion model can be used to identify the test set. Of course, the more times it is identified, the better, but it is impossible to identify an unlimited number of times. Therefore, it is necessary to choose according to the specific situation.

[0048] Step 7: The total concentration of the particulate matter to be tested can be calculated based on the number of electrical signals, the acquisition time, and the suction flow rate. Simultaneously, based on the final classification results of the fusion model, the proportion of each type of particulate matter in the tested particulate matter can be obtained. Furthermore, the following relationship exists between the peak value of the electrical signal (i.e., the voltage change) and the particle size; therefore, the particle size distribution of each type of particulate matter can be calculated based on the classification results and the peak value.

[0049]

[0050] In equation (1): ΔV is the voltage change, ρ e Let d be the conductivity of the solution, I be the applied direct current, and d be the current applied to the solution. p D is the particle size of the particle to be measured. h That is the diameter of the aperture. From the above formula, we can see that the voltage change is directly proportional to the cube of the particle size.

[0051] In this embodiment, the total particle concentration in the water body was calculated to be 262,500 particles / ml by calculating the number of collected electrical signals, the collection time, and the suction flow rate. Of these, 24.6% were air bubbles, 21.9% were microbial cells, 18.9% were microplastics, 16.9% were organic flocs, 16.5% were inorganic particles, and 1.2% were unknown. Subsequently, the particle size distribution of each type of particle was calculated based on the classification results, and the frequency values ​​of each particle size distribution were reduced according to their proportion. Finally, a plot was drawn. Figure 10 and Figure 11 As shown.

[0052] The particulate matter identification and detection method based on an improved deep learning algorithm provided by this invention combines the electrical sensitive zone method with machine learning. It can accurately identify particulate matter such as inorganic particles, biological cells, bubbles, plastic particles, and organic flocs in water (with an accuracy rate of over 80%). It can also classify other types of particulate matter in water as "unknown," effectively avoiding interference from other unknown particles in the water. It can measure the type, concentration, and particle size distribution of particulate matter in water, improve the accuracy and precision of water quality monitoring, provide more comprehensive information for water plant operation and environmental monitoring, and enable the particulate matter identification and detection method to be better applied in practical applications.

[0053] Furthermore, those skilled in the art may make other changes within the spirit of this invention, and of course, all such changes made in accordance with the spirit of this invention should be included within the scope of protection claimed by this invention.

Claims

1. A method for particulate matter identification and detection based on an improved deep learning algorithm, characterized in that, The method includes the following steps: Step 1: Obtain electrical signals of various types of particulate matter using the electrical sensing area method; Step 2: Extract the time-domain, frequency-domain, and time-frequency-domain feature parameters of the electrical signal; Step 3: Combine feature parameters to enhance the differences between features; Step 4: Filter feature parameters to reduce computational load; Step 5: Perform a weighted fusion of the neural network and recurrent neural network models; Step 6: Use the fusion model to perform multiple repeated tests on the particulate matter to be tested, and use a voting mechanism to determine the final classification result; Step 7: Output the concentration, quantity percentage, and particle size distribution of various particulate matter.

2. The particulate matter identification and detection method based on an improved deep learning algorithm as described in claim 1, characterized in that, In step 1, the original signals of various types of particulate matter are obtained using the electrical sensitive area method, and the electrical signals of various types of particulate matter are extracted by wavelet denoising and peak localization algorithms.

3. The particulate matter identification and detection method based on an improved deep learning algorithm as described in claim 1, characterized in that, In step 2, the time-domain features include pulse width, 0.1 times peak pulse width, 0.2 times peak pulse width, rise time, fall time, rise kurtosis, variance, maximum slope, minimum slope, root mean square, skewness, and shape parameters; the frequency-domain features include wavelet transform integral, wavelet transform standard deviation, wavelet transform skewness, kurtosis, fifth-level intermediate frequency energy magnitude of wavelet decomposition, centroid frequency, root mean square frequency, average frequency, and frequency variance; the time-frequency domain features include signal entropy, spectral kurtosis, and energy spectral density.

4. The particulate matter identification and detection method based on an improved deep learning algorithm as described in claim 1, characterized in that, In step 3, new feature parameters "time product" and "time domain product" are constructed, wherein the time product is defined as the rise time multiplied by the fall time, and the time domain product is defined as the product of variance, skewness and kurtosis.

5. The particulate matter identification and detection method based on an improved deep learning algorithm as described in claim 1, characterized in that, In step 4, the MRMR algorithm and the ReliefF algorithm are used to sort the importance of features, and the results of the two are integrated. Finally, the following features are selected: 0.1 times pulse width, rise time, time product, variance, maximum slope, minimum slope, skewness, wavelet transform standard deviation, wavelet transform skewness, kurtosis, time domain product, high-frequency energy magnitude of the fifth layer of the wavelet domain, average frequency, and spectral kurtosis.

6. The particulate matter identification and detection method based on an improved deep learning algorithm as described in claim 1, characterized in that, In step 5, the NN and RNN models are trained respectively to output their classification probabilities for the sample, and the class with the highest probability is taken as the prediction result for the sample.

7. The particulate matter identification and detection method based on the improved deep learning algorithm as described in claim 6, characterized in that, In step 5, a weighted mechanism is used to fuse the classification probabilities of the NN model and the RNN model.

8. The particulate matter identification and detection method based on the improved deep learning algorithm as described in claim 7, characterized in that, In step 5, the specific weighting rules are as follows: if the NN model predicts the sample as plastic particles or organic flocs, its weight is set to 0.55; if the RNN model predicts the sample as inorganic particles or biological cells, its weight is set to 0.60; in the event of a conflict in the weighting rules, the original weights are followed, with the RNN model at 0.55 and the NN model at 0.

45. If the final maximum classification probability obtained through the above weighted classification probabilities is lower than 0.40, the sample is identified as an unknown class.

9. The particulate matter identification and detection method based on the improved deep learning algorithm as described in claim 5, characterized in that, In step 6, the electrical signal of the particulate matter to be tested is acquired using the electrical sensitive area method, and the feature parameters defined in step 4 are extracted to form a test set. The fusion model is used to perform n classification tests on the test set, and a voting mechanism is adopted to statistically analyze the results of the n tests to determine the final classification result.

10. The particulate matter identification and detection method based on the improved deep learning algorithm as described in claim 1, characterized in that, In step 7, the total concentration of the particulate matter to be tested can be calculated based on the number of electrical signals of the particulate matter to be tested, the acquisition time, and the suction flow rate.

11. The particulate matter identification and detection method based on the improved deep learning algorithm as described in claim 9, characterized in that, In step 7, based on the final classification result of the fusion model, the proportion of each type of particulate matter in the particulate matter to be tested is obtained.

12. The particulate matter identification and detection method based on the improved deep learning algorithm as described in claim 9, characterized in that, In step 7, the following relationship exists between the peak value of the electrical signal and the particle size. Based on the classification results and the peak value, the particle size distribution of various types of particles can be calculated. In equation (1): ΔV is the voltage change, ρ e Let d be the conductivity of the solution, I be the applied direct current, and d be the current applied to the solution. p D is the particle size of the particle to be measured. h It is the diameter of the small hole.

Images