A rapid and accurate diagnosis method for abnormal spectrum of distillate oil

Abstract

The application discloses a kind of fraction oil abnormal spectrum fast and accurate diagnosis method.The method is first to fraction oil near infrared spectrum is carried out wave band interception and vector normalization preprocessing, through cascading wavelet convolution, point-by-point modular operation and low-pass smoothing wavelet scattering transformation, extract spectral scattering characteristics.Scattering characteristics and original spectrum are fused in channel dimension to generate joint features, then input light convolution network to complete abnormal diagnosis.Experimental results show that the method can effectively overcome the defects of weak feature expression, small sample prone to overfitting, real-time and precision difficult to consider, the accuracy of abnormal diagnosis is more than, with good real-time, high accuracy characteristics, suitable for refining enterprise main oil refining unit fraction oil online near infrared spectrum abnormal diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection technology for distillate oil properties in refining and chemical enterprises, specifically a method for diagnosing spectral anomalies in distillate oil based on wavelet scattering transform and artificial intelligence. Background Technology

[0002] In recent years, online near-infrared spectroscopy technology has been widely used in the analysis and monitoring of oil properties during the production process of refining and chemical enterprises. However, in actual acquisition, the spectral signal is affected by factors such as fluctuations in operating conditions, changes in sample state, and the detection system itself, resulting in abnormal spectra in the acquired data, which in turn affects the accuracy of the property analysis results.

[0003] In engineering practice, anomalous spectra typically manifest in various forms, such as abnormal changes in characteristic peaks, spectral baseline drift, noise introduction, and fluctuations in absorbance in local bands. These are often caused by a variety of factors, including abnormal water content, impurities, abnormal temperature, and measurement noise interference in the oil sample. Many anomalous spectra, while appearing similar to normal spectra on their original surfaces, can significantly interfere with subsequent classification and property prediction results after calculation using property analysis models. This leads to fluctuations in property analysis results and affects the accurate monitoring and optimization of equipment operation.

[0004] Existing methods for detecting spectral anomalies in distillate oils are mostly based on statistical thresholds or simple models. With limited sample size and insufficient feature representation capabilities, they have limited ability to identify abnormal spectra under complex operating conditions and cannot meet the accuracy requirements of online analysis.

[0005] Of course, considering the speed requirement of online analysis, when using complex models for high-dimensional feature extraction, attention should also be paid to controlling the model size to ensure the real-time nature of anomaly diagnosis. Summary of the Invention

[0006] This invention addresses the problems existing in the background technology by disclosing a rapid and accurate method for diagnosing abnormal spectra of distillate oils. The method performs standardized preprocessing on distillate oil spectral data, extracts spectral features through wavelet scattering transform, constructs a joint feature representation of the scattering features and the original spectrum, and designs a lightweight convolutional network to achieve anomaly diagnosis. The specific steps are as follows:

[0007] Step 1: Perform band interception and vector normalization preprocessing on the near-infrared spectral data of distillate oils in the abnormal sample library to obtain preprocessed one-dimensional spectral data. .

[0008] Step 1-1, Wavenumber range selection, selecting the wavenumber range. The absorbance between them.

[0009] Step 1-2, vector normalization, the calculation formula is:

[0010]

[0011] in, It is the first Each sample in wavenumber Absorbance at the specified level; It is the first The average absorbance of each sample; It is the number of wavenumber points; It is the first normalized vector. Each sample in wavenumber The absorbance at that point.

[0012] Step 2: Feature extraction is performed on the preprocessed spectral data using wavelet scattering transform. The specific implementation process is as follows:

[0013] Step 2-1, given the Morlet wavelet and low-pass scaling function Constructing a one-dimensional wavelet family With low-pass filter ;

[0014]

[0015]

[0016] in, , , Controlling the analysis scale of wavelets, maximum scale Determines the range of translation invariance of the feature.

[0017] Step 2-2, for and Perform convolution operations to obtain responses at different scales:

[0018]

[0019] Steps 2-3 introduce pointwise modulo operations after wavelet convolution to obtain a first-order nonlinear representation:

[0020]

[0021] Steps 2-4 involve applying a scale to the signal after modulo operation. low-pass filter After smoothing, the first-order scattering coefficients are obtained:

[0022]

[0023] Steps 2-5 involve performing wavelet convolution, modulus operation, and low-pass filtering again on the first-order modulus operation results to obtain the second-order scattering coefficients. ;

[0024] Steps 2-6 involve stitching together the first-order and second-order scattering features along the channel dimension to obtain the spectral scattering features. .

[0025] Step 3, apply the spectral scattering characteristics obtained in Step 2. The spectrum obtained in step 1 The features are spliced ​​and fused along the channel dimension to generate joint features. ;

[0026] Step 4, construct a lightweight convolutional network, including stage, stage, The three core phases are:

[0027] Step 4-1, The first stage is the preliminary feature extraction and downsampling stage, using a step size of [missing information]. of A one-dimensional convolutional kernel is sequentially connected to a one-dimensional batch normalization layer, a ReLU activation function, and an SE channel attention module;

[0028] Step 4-2, The stage is the characteristic nonlinear enhancement stage, which adopts the method of... The lightweight convolutional module, composed of one-dimensional convolutional convolutions, first expands the feature channels and then compresses them back to the original dimension. At the same time, a residual connection structure is introduced, which adds the input features of the module to the output of the convolutional module element by element and then outputs them through the ReLU activation function.

[0029] Step 4-3, The stage consists of feature aggregation and classification output, using a step size of [missing information]. One-dimensional convolutions are used to downsample high-dimensional features. After features are aggregated by a global average pooling layer, they are fed into a single-layer fully connected classifier through a regularization layer to output the category prediction results for distillate oil spectral anomaly diagnosis.

[0030] Step 5: Train the lightweight convolutional network of the model:

[0031] Step 5-1, combine features The training set was divided into two parts at an 8:2 ratio. and test set ;

[0032] Step 5-2, prepare the training set The input is fed into a lightweight convolutional network, and the model parameter weights are adjusted using the AdamW optimization algorithm.

[0033] Step 5-3, using the test set Evaluate the model's generalization ability. If the classification accuracy is greater than 95%, a well-trained anomaly diagnosis model is obtained. Otherwise, return to step 5.2.

[0034] Step 6: Based on the anomaly diagnosis of the lightweight one-dimensional convolutional network, obtain the near-infrared spectral data of the distillate oil to be diagnosed, and obtain the joint features of the distillate oil to be diagnosed by referring to steps 1)-3). Input the joint features into the lightweight convolutional network and output the final category prediction result.

[0035] Beneficial effects:

[0036] This invention extracts near-infrared spectral features of distillate oils through wavelet scattering transform and constructs a joint feature representation of the scattering features and the original spectrum. This enhances the feature expression capability while preserving the original spectral information, thereby improving the separability of abnormal samples. Simultaneously, it combines a lightweight convolutional network for anomaly diagnosis and classification, reducing model complexity while ensuring recognition accuracy. This makes the method applicable to small sample conditions and meets the real-time requirements of online detection, thus achieving accurate identification and diagnosis of spectral anomalies in distillate oils. Attached Figure Description

[0037] Figure 1 Flowchart of a rapid and accurate diagnostic method for abnormal spectra of distillate oils

[0038] Figure 2 Schematic diagram of wavelet scattering transform

[0039] Figure 3 Lightweight diagnostic network architecture diagram Detailed Implementation

[0040] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0041] The company has collected 1,485 abnormal spectral data points of distillate oil. Using the company's data, the effectiveness of the rapid and accurate diagnostic method for abnormal distillate oil spectra proposed in this invention is verified.

[0042] The overall system flowchart of the present invention is as follows: Figure 1 As shown, the specific steps include:

[0043] 1) Read the raw near-infrared spectral data of the distillate oil, first perform band selection, and retain... Effective feature interval; subsequently, vector normalization is performed on the spectral data to obtain preprocessed spectral data. .

[0044] 2) Perform wavelet scattering transform on the preprocessed spectral data; the feature extraction process is as follows: Figure 2 As shown. Given a Morlet wavelet and low-pass scaling function Construct a one-dimensional wavelet function With low-pass filter .right and Convolution operations are performed to obtain responses at different scales. Pointwise modulo operations are introduced after wavelet convolution to obtain a first-order nonlinear representation. Low-pass filtering is then applied. The signal is smoothed to obtain the first-order scattering coefficient. The first-order modulus operation result is then subjected to wavelet convolution, modulus operation, and low-pass filtering again to obtain the second-order scattering coefficients. By stitching together the first-order and second-order scattering features along the channel dimension, the spectral scattering features are obtained. .

[0045] 3) The spectral scattering characteristics obtained in step 2) The spectrum obtained in step 1) The features are spliced ​​and fused along the channel dimension to generate joint features. It fully preserves spectral discrimination information.

[0046] 4) Construct a lightweight convolutional network, with the network structure as follows: Figure 3 As shown. The first stage is the preliminary feature extraction and downsampling stage, using a step size of [missing information]. of A one-dimensional convolutional kernel is sequentially connected to a one-dimensional batch normalization layer, a ReLU activation function, and an SE channel attention module.

[0047] The stage is the characteristic nonlinear enhancement stage, which adopts the method of... The lightweight convolutional module, composed of one-dimensional convolutional convolutions, first expands the feature channels and then compresses them back to the original dimension. At the same time, a residual connection structure is introduced, which adds the input features of the module to the output of the convolutional module element by element and then outputs them through the ReLU activation function.

[0048] The stage consists of feature aggregation and classification output, using a step size of [missing information]. One-dimensional convolution downsamples high-dimensional features, aggregates features through a global average pooling layer, and then connects to a single-layer fully connected classifier through a regularization layer to output the category prediction results for distillate oil spectral anomaly diagnosis.

[0049] 5) The model training uses the Adam optimization algorithm, with an initial learning rate of... Learning rate decay rate Batch size Regularization coefficient The total training duration was 100 epochs, and the model parameters were optimized using the cross-entropy loss function.

[0050] To verify the effectiveness of the method of this invention, an ablation experiment was conducted. The accuracy of online near-infrared spectral anomaly diagnosis of distillate oil was used as the evaluation index to analyze whether the original spectrum was fused. Module parameters The impact of module structure and SE channel attention module on diagnostic performance is shown in Table 1.

[0051] Table 1. Comparison of diagnostic performance under different model structure configurations

[0052]

[0053] As shown in Table 1, fusing the original spectrum with wavelet scattering features can effectively improve diagnostic accuracy; The stage uses 2 layers of convolution with 64 channels. When the stage employs 3 convolutional layers, 4 channels, and introduces an SE channel attention module, the model achieves optimal anomaly diagnosis performance with an accuracy of [missing information]. The comparative results of each ablation group fully demonstrate that the feature fusion strategy, lightweight network structure, and SE attention module of this invention all contribute to improving the accuracy of anomaly diagnosis, verifying the effectiveness and superiority of the technical solution of this invention.

Claims

1. A rapid and accurate method for diagnosing abnormal spectra in distillate oils, characterized in that, First, the distillate oil spectral data are preprocessed. Then, spectral features are extracted using wavelet scattering transform. A joint feature representation of the scattering features and the original spectrum is constructed. Finally, a lightweight convolutional network is designed to achieve anomaly diagnosis and classification, including the following steps: 1) Band interception and vector normalization preprocessing were performed on the near-infrared spectral data of distillate oil to obtain preprocessed one-dimensional spectral data. ; 2) By using wavelet scattering transform, feature extraction is performed on the preprocessed spectral data to obtain spectral scattering features. ; 3) The scattering characteristics obtained in step 2) The one-dimensional spectral data obtained in step 1) The features are spliced ​​and fused along the channel dimension to generate joint features. ; 4) Design a three-stage lightweight convolutional network, including an initial feature extraction and downsampling stage, a feature nonlinear enhancement stage, and a feature aggregation and classification output stage; 5) Train the lightweight convolutional network model; 6) Obtain the near-infrared spectral data of the distillate oil to be diagnosed. Refer to steps 1)-3) to obtain the joint features of the distillate oil to be diagnosed. Input the joint features into the trained lightweight one-dimensional convolutional network and output the category prediction results of the spectral anomaly diagnosis.

2. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 1, characterized in that, The specific implementation process of the wavelet scattering transform is as follows: 1) Given Morlet wavelet and low-pass scaling function Constructing a one-dimensional wavelet family With low-pass filter ; Among them, scale sequence Maximum scale ; t represents the wavenumber index, Represents a set of integers; 2) To and Perform convolution operations to obtain responses at different scales. ; 3) Perform modular arithmetic to obtain a first-order nonlinear representation. ; 4) Use low-pass filtering Smoothing is performed to obtain the first-order scattering coefficient. ; 5) Perform wavelet convolution, modulus operation, and low-pass filtering again on the first-order modulus operation result to obtain the second-order scattering coefficients. ; 6) The first-order and second-order scattering features are stitched together along the channel dimension to obtain the spectral scattering features. .

3. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 2, characterized in that, First-order scattering scale Second-order scattering scale .

4. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 1, characterized in that, The lightweight convolutional network includes stage, stage, Phase, in which: 1) The first stage is the preliminary feature extraction and downsampling stage, using a step size of [missing information]. of A one-dimensional convolutional kernel is sequentially connected to a one-dimensional batch normalization layer, a ReLU activation function, and an SE channel attention module; 2) The stage is the characteristic nonlinear enhancement stage, which adopts the method of... A lightweight convolutional module composed of one-dimensional convolutional convolutions is introduced, and a residual connection structure is introduced to add the module input features and the convolutional module output element by element and then output the result after passing the ReLU activation function. 3) The stage consists of feature aggregation and classification output, using a step size of [missing information]. One-dimensional convolution downsamples high-dimensional features, aggregates features through a global average pooling layer, and then connects to a single-layer fully connected classifier through a regularization layer to output the category prediction results for distillate oil spectral anomaly diagnosis.

5. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 1, characterized in that, The model training steps include: 1) Combined features The training set was divided into two parts at an 8:2 ratio. and test set ; 2) The training set Input the data into a lightweight convolutional network and adjust the model parameter weights; 3) Through the test set Evaluate the model's generalization ability. If the classification accuracy is greater than 95%, a well-trained anomaly diagnosis model is obtained; otherwise, repeat step 2.

6. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 5, characterized in that, The model parameter weights were adjusted using the AdamW optimization algorithm.

7. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 1, characterized in that, Use wavenumber segmentation to set the wavenumber range. , .

8. The rapid and accurate diagnostic method for abnormal spectra of distillate oils according to claim 1, characterized in that, Vector normalization is performed using the following formula: in, It is the first Each sample in wavenumber Absorbance at the specified level; It is the first The average absorbance of each sample; It is the number of wavenumber points; It is the first normalized vector. Each sample in wavenumber The absorbance at that point.

Images