A deep learning-based method for trace element analysis of marine fuel

Abstract

The present invention discloses a method for analyzing trace elements in marine fuel based on deep learning, including: S1: performing interval sampling on the SnO2 quantum dot fluorescence intensity curve graph of the fuel sample; S2: obtaining a standardized matrix; S3: obtaining a simplified standardized matrix; S4: obtaining a graph of the change of SnO2 quantum dot fluorescence intensity with wavelength; S5: obtaining a corrected graph of the change of SnO2 quantum dot fluorescence intensity with wavelength; S6: building an adversarial network to obtain a simulated graph of the change of SnO2 quantum dot fluorescence intensity with wavelength; S7: building a convolutional neural network MCNN and obtaining an optimal convolutional neural network; S8: obtaining the concentration of the fuel trace elements and the category of the fuel sample. The present invention solves the problems of low efficiency, high radiation, environmental pollution, etc. of traditional fuel sulfur content detection technologies, can show better effects, further improves the detection accuracy and precision, and efficiently completes the data processing task.

Description

Technical Field

[0001] The present invention relates to the technical field of computer science, and in particular to a method for analyzing trace elements in marine fuel based on deep learning. Background Art

[0002] As people pay more attention to the marine environment, the pollution problem of ship fuel has received more and more attention. The International Maritime Organization has enforced the "sulfur limit order" for ship fuel since 2020. At the same time, the Ministry of Transport issued a document to clearly stipulate the fuel standards for incoming ships, which means that every ship docking at the port must undergo fuel qualification testing.

[0003] Traditional fuel sulfur content detection technology mainly detects the exhaust gas after fuel combustion or uses X-rays. This method is inefficient and will cause environmental pollution problems and affect the health of measurement personnel. Therefore, it is urgent to carry out further research on ship fuel sulfur content monitoring technology. Summary of the invention

[0004] The present invention provides a method for analyzing trace elements in marine fuel based on deep learning to overcome the above-mentioned technical problems.

[0005] In order to achieve the above object, the technical solution of the present invention is:

[0006] A method for analyzing trace elements in marine fuel based on deep learning, comprising the following steps:

[0007] S1: Obtaining a SnO2 quantum dot fluorescence intensity curve of a fuel sample and performing interval sampling on the curve to obtain a sampling sample sequence;

[0008] S2: performing a standard transformation on the sample sequence to obtain a standardization matrix;

[0009] S3: Calculating a correlation coefficient matrix for the standardized matrix to obtain a simplified standardized matrix;

[0010] S4: obtaining a curve of SnO2 quantum dot fluorescence intensity versus wavelength according to the simplified standardized matrix;

[0011] S5: correcting the SnO2 quantum dot fluorescence intensity versus wavelength curve, obtaining a corrected SnO2 quantum dot fluorescence intensity versus wavelength curve, and adjusting the grayscale of the SnO2 quantum dot fluorescence intensity versus wavelength curve;

[0012] S6: Building an adversarial network, and inputting the modified SnO2 quantum dot fluorescence intensity versus wavelength curve into the adversarial network to obtain a simulated SnO2 quantum dot fluorescence intensity versus wavelength curve;

[0013] S7: Building a convolutional neural network MCNN; inputting the simulated SnO2 quantum dot fluorescence intensity versus wavelength curve into the convolutional neural network MCNN to obtain an optimal convolutional neural network;

[0014] S8: Input the modified SnO2 quantum dot fluorescence intensity versus wavelength curve into the optimal convolutional neural network to obtain the concentration of trace elements in the fuel and the category of the fuel sample.

[0015] Furthermore, the sample sequence in S1 is calculated as follows:

[0016] x(q)=x a (qT) (1)

[0017] Among them, x(q) is the sampling sample sequence, x a (·) is the original sample sequence, q is the multiple of the time interval, and T is the sampling time interval.

[0018] Furthermore, the standardized matrix in S2 is obtained as follows:

[0019]

[0020] Among them, Z ij is the element in the i-th row and j-th column of the normalized matrix Z; x ij is the observed value of the jth feature of the i-th data in the sample sequence; is the average value of the jth feature of all data in the sample sequence; s j is the standard deviation of the jth feature of all data in the sample sequence; i is the data number in the sample sequence; j is the number of the feature of the data in the sample sequence; n is the total number of data in the sample sequence; p is the total number of features of the data in the sample sequence; where,

[0021] Furthermore, the correlation coefficient matrix in S3 is calculated as follows:

[0022]

[0023] Among them, R is the correlation coefficient matrix; r st is the correlation coefficient of the sample sequence; [r st ] p For r st The concatenated matrix; x p is the feature vector of the data in the sample sequence;

[0024] in, s,t=1,2,...,p;z utis the eigenvalue of u relative to the eigenvalue of t; z su is the eigenvalue of s relative to u; s is the row variable in the standardized matrix; u is the summation variable, and t is the column variable in the standardized matrix.

[0025] Furthermore, the formula used in S5 to correct the grayscale of the SnO2 quantum dot fluorescence intensity versus wavelength curve is as follows:

[0026] s=cr γ (4)

[0027] Among them, r is the input grayscale value, and its value range is [0,1]; c is the grayscale scaling coefficient, which represents the overall stretched graphic grayscale; γ is the gamma factor size; s is the grayscale output value after gamma transformation.

[0028] Furthermore, the loss function of the adversarial network is calculated as follows:

[0029]

[0030] Where: D(x) is the discriminator L of the value of the similarity between the generated simulated SnO2 quantum dot fluorescence intensity data and the real SnO2 quantum dot fluorescence intensity data Was is the Wasserstein distance loss; E pr [·] is the expected distribution of SnO2 quantum dot fluorescence intensity sample; E pg [·] is the expected distribution of SnO2 quantum dot fluorescence intensity samples produced by the generator; To generate SnO2 quantum dot fluorescence intensity sample; L Was-GP is the Was-GP loss value; To synthesize the expected distribution of fluorescence intensity samples of SnO2 quantum dots; is the gradient of the sample obtained after the synthesis of the real SnO2 quantum dot fluorescence intensity sample and the generated SnO2 quantum dot fluorescence intensity sample; is the probability that the synthesized SnO2 quantum dot fluorescence intensity sample belongs to the real SnO2 quantum dot fluorescence intensity sample; is the sample obtained by synthesizing the real SnO2 quantum dot fluorescence intensity sample and the generated SnO2 quantum dot fluorescence intensity sample; ||·||2 is the L2 paradigm.

[0031] Furthermore, the combined loss function of the convolutional neural network MCNN in S7 is calculated as follows:

[0032] L=L r +L c +λL me (6)

[0033] Among them, L r is the mean square error loss, L c +λL me is the total classification error; L is the total loss value of the convolutional neural network; L c is the cross entropy loss; L me To measure learning loss;

[0034] Among them, for the task of estimating the content of trace elements, the mean square error loss is selected and calculated as follows:

[0035]

[0036] where y r is the estimated content of trace elements, It is the true content of trace elements;

[0037] For the fuel classification task, the cross entropy loss is used; it is calculated as follows:

[0038]

[0039] where y c It is the fuel classification label. It is the fuel truth label;

[0040] At the same time, the metric learning loss is used, and the calculation is as follows

[0041]

[0042] where c k Fuel sample category label, K is the total number of fuel sample categories.

[0043] Beneficial effects: The present invention provides a method for analyzing trace elements in marine fuel based on deep learning, which obtains the concentration of trace elements in the fuel sample by obtaining the curve of SnO2 quantum dot fluorescence intensity versus wavelength, thereby solving the problems of low efficiency, high radiation, and environmental pollution in traditional fuel sulfur content detection technology. An adversarial network is used to increase the amount of data input to the convolutional neural network. Compared with traditional detection technology, when the data set is large enough, continuous iterative updates through the convolutional neural network can show better results, further improve detection accuracy and precision, and efficiently complete data processing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces 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 creative labor.

[0045] Figure 1 This is a flow chart of the fuel trace element analysis method of the present invention;

[0046] Figure 2 It is a structural diagram of a multi-channel convolutional neural network MCNN in an embodiment of the present invention;

[0047] Figure 3 is a SnO2 quantum dot fluorescence intensity curve in an embodiment of the present invention;

[0048] Figure 4 A flow chart for obtaining a modified SnO2 quantum dot fluorescence intensity versus wavelength curve in an embodiment of the present invention;

[0049] Figure 5 Schematic diagram of feature extraction of an asymmetric convolution kernel in an embodiment of the present invention;

[0050] Figure 6 A diagram showing the structure of an adversarial network in an embodiment of the present invention;

[0051] Figure 7 It is the classification accuracy change curve in the embodiment of the present invention;

[0052] Figure 8 is a curve of variation of estimation error in an embodiment of the present invention;

[0053] Fig. 9 It is a comparison curve of the classification accuracy of other existing algorithms in the embodiments of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The present invention aims to propose a method for analyzing trace elements in marine fuel oil based on deep learning. The MCNN multi-way convolutional neural network is used to further analyze the experimental data, complete the classification of fuel oil and estimate the content of trace elements. At the same time, a generative adversarial network (WGAN) is built for data enhancement, and the convolutional neural network is pre-trained using a simulated data set to improve the robustness of the model. The present invention can open up a new way for the detection of sulfur content in marine fuel oil.

[0056] This embodiment provides a method for analyzing trace elements in marine fuel based on deep learning. Figure 1 As shown, the following steps are included:

[0057] S1: Obtain the SnO2 quantum dot fluorescence intensity curve of the fuel sample and perform interval sampling on it to obtain a sampling sample sequence;

[0058] Specifically, the data sequence of the fluorescence intensity of SnO2 quantum dots as the fluorescence wavelength changes in this embodiment is obtained from the SnO2 quantum dot fluorescence intensity curve obtained by measuring 7 types of fuel samples provided by Liaoning Maritime Safety Administration in the laboratory by multiple people in multiple time periods, such as Figure 3 As shown, this is a conventional method in the field, so it will not be described here in detail. This embodiment only uses experimental data to implement the method in this patent. The trace elements in the fuel in this embodiment refer to the content of sulfur in the fuel. In order to quantify the relationship between the SnO2 fluorescence peak and the sulfur content in the fuel, the team used 7 types of fuels with known concentrations for testing in the laboratory stage. Through multi-person experiments, the team collected a large number of fluorescence spectrum data sets of "fuels with different sulfur contents and SnO2 mixed liquids", and input the data into the convolutional neural network for continuous training, so as to obtain a mathematical model that characterizes the relationship between sulfur content and fluorescence spectrum, and then complete the estimation of sulfur content and the classification of fuel. The principle is: through experiments, it was found that fuels with different sulfur contents will reduce the fluorescence intensity of SnO2 quantum dots to varying degrees. By adding 1% sulfur element in a certain proportion, the higher the sulfur content of the solution, the lower the fluorescence intensity of the luminescence peak between 300nm and 350nm. Then, a new luminescence peak appears between 400nm and 450nm. This peak value is positively correlated with the concentration of sulfur in the fuel sample. Therefore, this embodiment analyzes this decreasing law to obtain the sulfur content of the fuel.

[0059] Preferably, the sampling sample sequence in S1 is calculated as follows:

[0060] x(q)=x a (qT) (1)

[0061] Among them, x(q) is the sampling sample sequence, x a (·) is the original sample sequence, q is the multiple of the time interval, and T is the sampling time interval.

[0062] S2: performing a standard transformation on the sample sequence to obtain a standardization matrix;

[0063] The standardized matrix is ​​obtained as follows:

[0064]

[0065] Among them, Z ij is the element in the i-th row and j-th column of the normalized matrix Z; x ij is the observed value of the jth feature of the i-th data in the sample sequence; is the average value of the jth feature of all data in the sample sequence; s j is the standard deviation of the jth feature of all data in the sample sequence; i is the data number in the sample sequence; j is the number of the feature of the data in the sample sequence; n is the total number of data in the sample sequence; p is the total number of features of the data in the sample sequence; where,

[0066] S3: Calculate the correlation coefficient matrix of the standardized matrix to reduce the dimension of the standardized matrix, obtain a simplified standardized matrix, and analyze it by principal component method, that is, analyze the curve of SnO2 quantum dot fluorescence intensity versus wavelength to reduce the amount of calculation;

[0067] Preferably, the correlation coefficient matrix is ​​calculated as follows:

[0068]

[0069] Among them, R is the correlation coefficient matrix; r st is the correlation coefficient of the sample sequence, which is the intermediate parameter of the calculation; [r st ] p For r st The concatenated matrix; x p is the feature vector of the data in the sample sequence;

[0070] in, s,t=1,2,...,p;z ut is the eigenvalue of the original data after standardization; z su is the eigenvalue of s relative to u; s is the row variable in the standardized matrix; u is the summation variable;

[0071] Specifically, in this embodiment, the characteristic root of the matrix R is obtained by using the characteristic equation, and then the corresponding eigenvector is obtained by the characteristic polynomial, and the component value of the eigenvector is used as the weight to weight the standardized index to obtain the i-th principal component. The cumulative contribution rate of the principal components such as temperature, pH, and metal ions is calculated, and the first k principal components are selected according to their size, and finally the 11×350 standardized matrix is ​​reduced to a simplified standardized matrix of 1×150. This implementation process is a prior art in the field, and the specific method is not described here.

[0072] S4: Obtaining a curve of SnO2 quantum dot fluorescence intensity versus wavelength according to the simplified standardized matrix;

[0073] Specifically, the data in the simplified standardized matrix in this embodiment is equivalent to a series of discrete points on the SnO2 quantum dot fluorescence intensity versus wavelength curve. Therefore, according to the simplified standardized matrix, the SnO2 quantum dot fluorescence intensity versus wavelength curve in the SnO2 quantum dot fluorescence intensity curve graph can be obtained, and then the SnO2 quantum dot fluorescence intensity versus wavelength curve graph can be obtained; this is a conventional technology in the field, so it will not be described in detail here.

[0074] S5: Correct the SnO2 quantum dot fluorescence intensity versus wavelength curve to obtain a corrected SnO2 quantum dot fluorescence intensity versus wavelength curve; correct the grayscale of the SnO2 quantum dot fluorescence intensity versus wavelength curve; Figure 4 As shown;

[0075] Specifically, in order to facilitate the subsequent analysis of the neural network, we performed gamma transformation correction on the modified SnO2 quantum dot fluorescence intensity versus wavelength curve, and corrected the images with too high or too low gray.

[0076] Preferably, the formula for correcting the grayscale of the SnO2 quantum dot fluorescence intensity versus wavelength curve in S5 is as follows:

[0077] s=cr γ (4)

[0078] Among them, r is the input grayscale value, ranging from [0,1]; c is the grayscale scaling factor, which represents the overall stretched grayscale. γ is the gamma factor, which controls the scaling degree of the entire transformation; s is the grayscale output value after the gamma transformation;

[0079] S6: Build an adversarial network, such as Figure 6 As shown, the modified SnO2 quantum dot fluorescence intensity versus wavelength curve is input into the adversarial network to obtain a simulated SnO2 quantum dot fluorescence intensity versus wavelength curve;

[0080] Preferably, the loss function of the adversarial network is calculated as follows:

[0081] We build an adversarial network WGAN to increase the number of samples, and generate corresponding simulation samples through adversarial network learning real samples to supplement the data set. We choose Wasserstein distance (bulldozer distance) as the loss function, which can measure the difference between the two data distributions even when they are far away from each other. The specific formula is as follows:

[0082]

[0083] Among them, D(x) is the discriminator; its output represents the probability that the sample belongs to the real data. It is a metric used to continuously learn and judge the real data so that the generated data is more similar to the real data. Its output is a real number less than 1 and greater than 0, and this output real number is used as the probability that the sample belongs to the real data; L Was is the Wasserstein distance loss; E pr [·] is the expectation of the sample distribution; E pg [·] is the expectation of the distribution of samples produced by the generator; To generate samples;

[0084] Considering that weight clipping is too rigid and may hinder the training optimization of the model, we choose the gradient penalty method to softly constrain it and speed up the convergence of the model, which is the loss function of the WGAN-GPD adversarial network. The specific formula is as follows: By introducing a weight term, it is possible to prevent one of the terms from having too much influence and causing loss imbalance, and introduce a constraint term. Deep learning is the learning of features, and weights are the effective expression of features.

[0085]

[0086] Where: D(x) is the discriminator L of the value of the similarity between the generated simulated SnO2 quantum dot fluorescence intensity data and the real SnO2 quantum dot fluorescence intensity data Was is the Wasserstein distance loss; E pr [·] is the expected distribution of SnO2 quantum dot fluorescence intensity sample; E pg [·] is the expected distribution of SnO2 quantum dot fluorescence intensity samples produced by the generator; To generate SnO2 quantum dot fluorescence intensity sample; L Was-GP is the Was-GP loss value; To synthesize the expected distribution of fluorescence intensity samples of SnO2 quantum dots; is the gradient of the sample obtained after the synthesis of the real SnO2 quantum dot fluorescence intensity sample and the generated SnO2 quantum dot fluorescence intensity sample; is the probability that the synthesized SnO2 quantum dot fluorescence intensity sample belongs to the real SnO2 quantum dot fluorescence intensity sample; is the sample obtained by synthesizing the real SnO2 quantum dot fluorescence intensity sample and the generated SnO2 quantum dot fluorescence intensity sample; ||·||2 is the L2 paradigm.

[0087] Specifically, when the loss function value of the adversarial network no longer decreases, the adversarial network training is completed, and the adversarial network output at this time is a curve of the simulated SnO2 quantum dot fluorescence intensity changing with wavelength;

[0088] S7: Building a convolutional neural network MCNN; inputting the simulated SnO2 quantum dot fluorescence intensity versus wavelength curve into the convolutional neural network MCNN to obtain an optimal convolutional neural network;

[0089] Specifically, the convolutional neural network MCNN includes an input layer, multiple convolution-pooling combination layers, a flattening layer, a fully connected layer and an output layer.

[0090] The input layer is: inputting the modified SnO2 quantum dot fluorescence intensity versus wavelength curve into the convolutional neural network MCNN, and the number of layers of the selected input layer is the same as the number of features.

[0091] Each of the convolution-pooling combination layers includes an asymmetric convolution kernel vector;

[0092] In this embodiment, since the horizontal and vertical coordinates of the input image represent the wavelength of the light source and the fluorescence intensity respectively, simultaneous extraction may cause feature confusion. Therefore, three parallel convolution-pooling combination layers are proposed in the convolutional neural network MCNN in this embodiment, such as Figure 2 As shown in , three asymmetric convolution kernel vectors of 1×5, 1×7 and 1×9 are designed specifically, as shown in Figure 5 As shown in the figure, asymmetric convolution kernels of different sizes are used for feature extraction. The convolution kernel activation function selects the relu function, and the pooling kernel adopts the commonly used maximum pooling method.

[0093] The local features obtained by the multiple convolution-pooling combination layers are recombined through the weight matrix through the flat layer and the fully connected hidden layer, and mapped back to the sample label space. This is an application of the prior art and will not be described in detail here.

[0094] The output layer uses a softmax logistic regression model to calculate the similarity between the sample and each fuel category to achieve multi-classification discrimination.

[0095] Preferably, the combined loss function of the convolutional neural network MCNN in S7 is calculated as follows:

[0096] The fuel sample data in this embodiment has a low matching degree with the existing public data set, so the sample set is not rich. Inspired by the idea of ​​small sample learning, this algorithm combines convolutional neural networks and metric learning algorithms, and designs a combined loss function for the two tasks.

[0097] L=L r +L c +λL me (6)

[0098] Among them, L r is the mean square error loss, Lc +λL me is the total classification error, where the weight λ of the measurement error is 0.1; L is the total loss value of the convolutional neural network; L c is the cross entropy loss; L me To measure learning loss;

[0099] Among them, for the task of estimating the content of trace elements, the mean square error loss is selected and calculated as follows:

[0100]

[0101] where y r is the estimated content of trace elements, It is the actual content of trace elements.

[0102] For the fuel classification task, the cross entropy loss is used; it is calculated as follows:

[0103]

[0104] where y c It is the fuel classification label. It is a fuel truth label.

[0105] At the same time, the metric learning loss is used, and the calculation is as follows

[0106]

[0107] where c k Fuel sample category label, K is the total number of fuel sample categories.

[0108] Specifically, when the total loss value L of the convolutional neural network no longer decreases, the convolutional neural network at this time is the optimal convolutional neural network;

[0109] S8: Input the modified SnO2 quantum dot fluorescence intensity versus wavelength curve into the optimal convolutional neural network to obtain the concentration of trace elements in the fuel and the category of the fuel sample.

[0110] Specifically, this embodiment inputs real data into the optimal convolutional neural network to obtain the concentration of trace elements in fuel and the category of the fuel sample. The fuel sample category is an international standard, which is a classification of fuel according to the trace concentration of fuel. We compared with Elementar, a company with a long history of 60 years, and our test results meet international standards.

[0111] One embodiment of the present invention is as follows: Figure 7-9 As shown:

[0112] The embedded device of the model built by the marine fuel trace element analysis method of this patent is tested with Class 1 and Class 6 real marine fuel samples, and the final results are displayed on the LCD screen of the single-chip computer. The results are shown in Table 1. The main element contents of the calibrated Class 1 marine fuel are: S: 3.17%; N: 0.48%; C: 84.5%; H: 9.97%. The main element contents of Class 6 marine fuel are: S: 0.47%; N: 0.52%; C: 86.24%; H: 10.58%. The algorithm test results are shown in Table 2. The main element contents of Class 1 marine fuel are estimated to be: S: 3.174%; N: 0.48%; C: 84.49%; H: 9.978%. The main element contents of Class 6 marine fuel are: S: 0.476%; N: 0.52%; C: 86.21%; H: 10.581%. The average estimated loss is less than 0.01%, so the method of this patent has good practicality. In addition, in accordance with the provisions of the International Maritime Organization's flow restriction order, the sulfur content is classified into two categories based on 0.5% as the standard to determine whether the oil sample is qualified. If it is unqualified, the display screen will prompt "Not up to standard!", otherwise it will prompt "Conform to the standard" to help the inspection personnel of the maritime law enforcement agency to make judgments.

[0113] Table 1 Experimental test data

[0114]

[0115]

[0116] Table 2 Average estimation of algorithms

[0117]

[0118] The present invention provides a method for analyzing trace elements in marine fuel based on deep learning, which solves the problems of low efficiency, high radiation, and environmental pollution of traditional fuel sulfur content detection technology. It adopts cutting-edge deep learning algorithms to efficiently complete data processing tasks.

[0119] 2. The deep learning-based marine fuel trace element analysis method proposed in the present invention has a detection accuracy of ±0.01%, and has been certified by a third-party authoritative testing agency with an accuracy of more than 95%.

[0120] 3. Compared with traditional detection technology, the deep learning-based marine fuel trace element analysis method proposed in the present invention can show better results and further improve the detection accuracy and precision by training the deep learning algorithm and continuously iterating and updating it when the data set is large enough.

[0121] Convolutional neural networks are an efficient data processing method that has developed in recent years and has attracted widespread attention. They have unique advantages in solving this problem.

[0122] The present invention realizes the valuation of trace elements and the classification of marine fuel oil based on the cutting-edge deep learning algorithms MCNN and WGAN. The advantages of the algorithm include: targeted design of multi-way convolutional neural network MCNN to complete the valuation of trace elements and the classification of fuel oil; use of asymmetric convolution kernel for feature extraction; construction of generative adversarial network WGAN for data enhancement; in the generative adversarial network, the Wasserstein loss is used to overcome the gradient vanishing problem. After cross-validation, the algorithm has excellent effect, with an accuracy rate of up to 95% on real data sets and a trace estimation error of less than 0.01. Through testing on the hardware system, this algorithm can provide a fast and accurate method for analyzing trace elements in marine fuel oil after being embedded in the mobile detection equipment of maritime law enforcement agencies.

[0123] 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 replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing trace elements in marine fuel based on deep learning, characterized in that: The steps include: S1: Obtaining a SnO2 quantum dot fluorescence intensity curve of a fuel sample and performing interval sampling on the curve to obtain a sampling sample sequence; S2: performing a standard transformation on the sample sequence to obtain a standardization matrix; S3: Calculating a correlation coefficient matrix for the standardized matrix to obtain a simplified standardized matrix; S4: obtaining a curve of SnO2 quantum dot fluorescence intensity versus wavelength according to the simplified standardized matrix; S5: correcting the SnO2 quantum dot fluorescence intensity versus wavelength curve, obtaining a corrected SnO2 quantum dot fluorescence intensity versus wavelength curve, and adjusting the grayscale of the SnO2 quantum dot fluorescence intensity versus wavelength curve; S6: Building an adversarial network, and inputting the modified SnO2 quantum dot fluorescence intensity versus wavelength curve into the adversarial network to obtain a simulated SnO2 quantum dot fluorescence intensity versus wavelength curve; S7: Building a convolutional neural network MCNN; inputting the simulated SnO2 quantum dot fluorescence intensity versus wavelength curve into the convolutional neural network MCNN to obtain an optimal convolutional neural network; S8: Input the modified SnO2 quantum dot fluorescence intensity versus wavelength curve into the optimal convolutional neural network to obtain the concentration of trace elements in the fuel and the category of the fuel sample.

2. The method for analyzing trace elements in marine fuel oil based on deep learning according to claim 1, characterized in that: The sample sequence in S1 is calculated as follows: x(q)=x a (qT) (1) Among them, x(q) is the sampling sample sequence, x a (·) is the original sample sequence, q is the multiple of the time interval, and T is the sampling time interval.

3. A method for analyzing trace elements in marine fuel oil based on deep learning according to claim 2, characterized in that: The standardized matrix in S2 is obtained as follows: Among them, Z ij is the element in the i-th row and j-th column of the normalized matrix Z; x ij is the observed value of the jth feature of the i-th data in the sample sequence; is the average value of the jth feature of all data in the sample sequence; s j is the standard deviation of the jth feature of all data in the sample sequence; i is the data number in the sample sequence; j is the number of the feature of the data in the sample sequence; n is the total number of data in the sample sequence; p is the total number of features of the data in the sample sequence; where, 4. The method for analyzing trace elements in marine fuel oil based on deep learning according to claim 3 is characterized in that: The correlation coefficient matrix in S3 is calculated as follows: Among them, R is the correlation coefficient matrix; r st is the correlation coefficient of the sample sequence; [r st ] p For r st The concatenated matrix; x p is the feature vector of the data in the sample sequence; in, z ut is the eigenvalue of u relative to the eigenvalue of t; z su is the eigenvalue of s relative to u; s is the row variable in the standardized matrix; u is the summation variable, and t is the column variable in the standardized matrix.

5. The method for analyzing trace elements in marine fuel oil based on deep learning according to claim 4, characterized in that: The formula used in S5 to correct the grayscale of the SnO2 quantum dot fluorescence intensity versus wavelength curve is as follows: s=cr γ (4) Among them, r is the input grayscale value, and its value range is [0,1]; c is the grayscale scaling coefficient, which represents the overall stretched graphic grayscale; γ is the gamma factor size; s is the grayscale output value after gamma transformation.

6. The method for analyzing trace elements in marine fuel oil based on deep learning according to claim 1, characterized in that: The loss function of the adversarial network is calculated as follows: Where: D(x) is the discriminator L of the value of the similarity between the generated simulated SnO2 quantum dot fluorescence intensity data and the real SnO2 quantum dot fluorescence intensity data Was is the Wasserstein distance loss; E pr [·] is the expected distribution of SnO2 quantum dot fluorescence intensity sample; E pg [·] is the expected distribution of SnO2 quantum dot fluorescence intensity samples produced by the generator; To generate SnO2 quantum dot fluorescence intensity sample; L Was-GP is the Was-GP loss value; To synthesize the expected distribution of fluorescence intensity samples of SnO2 quantum dots; is the gradient of the sample obtained after the synthesis of the real SnO2 quantum dot fluorescence intensity sample and the generated SnO2 quantum dot fluorescence intensity sample; is the probability that the synthesized SnO2 quantum dot fluorescence intensity sample belongs to the real SnO2 quantum dot fluorescence intensity sample; is the sample obtained by synthesizing the real SnO2 quantum dot fluorescence intensity sample and the generated SnO2 quantum dot fluorescence intensity sample; ||·||2 is the L2 paradigm.

7. The method for analyzing trace elements in marine fuel oil based on deep learning according to claim 1, characterized in that: The combined loss function of the convolutional neural network MCNN in S7 is calculated as follows: L=L r +L c +λL me (6) Among them, L r is the mean square error loss, L c +λL me is the total classification error; L is the total loss value of the convolutional neural network; L c is the cross entropy loss; L me To measure learning loss; Among them, for the task of estimating the content of trace elements, the mean square error loss is selected and calculated as follows: where y r is the estimated content of trace elements, It is the true content of trace elements; For the fuel classification task, the cross entropy loss is used; it is calculated as follows: where y c It is the fuel classification label. It is the fuel truth label; At the same time, the metric learning loss is used, and the calculation is as follows where c k Fuel sample category label, K is the total number of fuel sample categories.

Images