Method, device and system for rapidly analyzing components of direct coal liquefaction oil based on near infrared spectrogram

Abstract

The invention provides a method, a device and a system for rapidly analyzing components of direct coal liquefaction oil based on a near infrared spectrogram, the method comprises the following steps: acquiring sample data of the direct coal liquefaction oil, the sample data of the direct coal liquefaction oil comprises near infrared spectrum data of the direct coal liquefaction oil and detailed composition data of sample oil; according to sample data of the direct coal liquefaction oil and a ResNet-34 neural network algorithm, the near infrared spectrum of the liquefied oil with known composition is subjected to statistical treatment, the relationship between the liquefied oil composition and a spectrogram is obtained, then the near infrared spectrum spectrogram of an unknown oil product is identified to predict the composition, and a direct coal liquefaction oil component prediction model is established. The prediction model is at least used for component prediction of the direct coal liquefaction oil; the direct coal liquefaction oil in the production process is analyzed according to the prediction model obtained through training, online analysis can be rapidly and accurately conducted on the direct coal liquefaction oil, and the method can solve the technical problems that in the direct coal liquefaction oil production process, analysis is difficult, and the analysis period is long relative to the production period.

Description

Technical Field

[0001] This invention belongs to the field of coal chemical process analysis technology and spectral detection, and relates to a method, apparatus and system for rapid analysis of coal direct liquefaction oil components based on near-infrared spectral data. Background Technology

[0002] Direct coal liquefaction (COL) is a crucial technology that converts coal into liquid fuel through high-temperature, high-pressure hydrogenation. The resulting liquefied oil has a complex composition, containing thousands of compounds, and its component distribution directly impacts subsequent processing and utilization. While traditional laboratory analytical methods such as gas chromatography-mass spectrometry (GC-MS) and full two-dimensional gas chromatography offer high analytical precision, they suffer from drawbacks such as complex sample pretreatment, long analysis cycles, and expensive equipment, making it difficult to meet the real-time quality control requirements of modern coal chemical production processes.

[0003] Near-infrared spectroscopy encompasses the 700-2500 nm wavelength range within the infrared spectrum, with 700-1100 nm considered a short-wavelength region and 1100-2500 nm a long-wavelength region. Electromagnetic waves within this range possess energy lower than visible light but higher than those in the mid-infrared region. Molecules absorbing near-infrared light exhibit both stretching and bending vibrational modes, including overtones and combination frequencies of C-H, N-H, S-H, and O-H vibrations, as well as weaker overtones of stretching vibrations such as C-N and C-C. Coal direct liquefaction products are primarily composed of hydrocarbons, with small amounts of nitrogen- and sulfur-containing compounds. The overtones and combination frequencies of hydrogen-containing groups in these compounds fall precisely within the near-infrared spectral range. Therefore, near-infrared spectra contain rich chemical structural information and can effectively reflect the composition of hydrogen-containing groups in coal direct liquefaction diesel. Furthermore, near-infrared spectroscopy is simple to operate and fast to analyze. Its application to the rapid analysis of liquid products from coal direct liquefaction will facilitate online real-time analysis and product quality monitoring, leading to significant economic benefits.

[0004] While near-infrared spectroscopy, a commonly used industrial analytical technique, offers advantages such as speed and non-destructive testing, it faces the following technical challenges when applied to the analysis of coal direct liquefaction oil: The composition of coal liquefaction oil is extremely complex, with severe overlap of near-infrared absorption peaks of various compounds, making it difficult for traditional linear correction models to accurately resolve the issue. The quality of raw coal varies greatly between different batches, resulting in significant fluctuations in the composition of liquefied oil, and conventional models have insufficient generalization ability. Under high-temperature and high-pressure process conditions, spectral acquisition is easily affected by factors such as temperature drift and particle interference. Existing methods are mostly limited to laboratory analysis and lack real-time solutions that can be integrated with industrial control systems. Summary of the Invention

[0005] In view of the problems existing in the methods, devices and systems for rapid analysis of coal direct liquefaction oil components based on near-infrared spectroscopy, this invention is proposed.

[0006] Therefore, the problem to be solved by this invention is: how to overcome the difficulties of long cycle of traditional analysis methods, the inability of existing near-infrared technology to analyze complex components and insufficient generalization ability, susceptibility to industrial environment interference, and lack of real-time solutions integrated with control systems, so as to achieve rapid, accurate, online, real-time analysis of the composition of coal direct liquefaction oil and intelligent monitoring integrated with industrial processes.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning. This method includes acquiring various data of direct coal liquefaction oil under different reaction conditions. These data include near-infrared spectral data and compositional data of the direct coal liquefaction oil. The different reaction conditions include different coal melting ratios, different liquefaction solvents, and different initial hydrogen pressures. Compositional data for each direct coal liquefaction oil sample is acquired using GC×GC-MS, and the near-infrared spectrum of each direct coal liquefaction oil sample is acquired using a near-infrared spectroscopy device. Each sample and its corresponding family composition data are input into the near-infrared spectral machine learning model to obtain the target category information of the coal liquefaction product to be tested, which is the category information of each component in the coal liquefaction product to be tested. The near-infrared spectroscopy machine learning model takes the near-infrared spectral data of a coal direct liquefaction sample with known composition as input and collects the features of the near-infrared spectral spectrum by machine learning methods. The features of the near-infrared spectral spectrum include peak position, peak shape, peak height, half-width at half-maximum, and peak area parameters. The near-infrared spectral spectrum machine learning model can process the features of the pre-processed near-infrared spectral spectrum and predict the corresponding composition data through the features. The analysis of coal liquefaction oil is performed based on the near-infrared spectral machine learning model, and the analysis database is used for classification and qualitative analysis. A full two-dimensional spectrum analysis was performed based on the full two-dimensional spectrum residual neural network analysis model, and classification and qualitative analysis were conducted using the analysis database.

[0008] As a preferred embodiment of the rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning described in this invention, the near-infrared spectral machine learning model includes, based on the near-infrared spectral machine learning model, performing model prediction on the preprocessed near-infrared spectral data of direct coal liquefaction oil using an improved ResNet-34 neural network algorithm. The principle of a residual block of the improved ResNet-34 neural network algorithm is as follows:

[0009] Where x represents the input sample data. It is the residual function; The ResNet-34 model consists of a total of 4 residual stages, each containing 3 improved residual blocks. Each residual block integrates a channel attention module, uses depthwise separable convolutions to reduce the number of parameters, and the output layer adopts a multi-task learning architecture to predict the content of multiple components.

[0010] As a preferred embodiment of the rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning described in this invention, the rapid real-time analysis method further includes a preprocessing model. A preprocessing module is established based on the Pearson correlation coefficient method and the first-order differential method, including: According to the preprocessing module, the input near-infrared spectral data of direct coal liquefaction oil is first subjected to first-order differential processing, retaining the filtered wavelength data and deleting redundant spectral data. The filtered wavelength data refers to the data obtained by the Pearson correlation coefficient method, and the formula for the Pearson correlation coefficient method is:

[0011] Where r is the correlation coefficient, X is the input spectral data, and Y is the data to be predicted; The near-infrared spectral machine learning model includes: processing the spectral information corresponding to the absorption of chemical bonds of the target component in each component through first-order differentiation and extracting correlation coefficients to screen features; extracting the dimension of each sample; inputting the preprocessed features into a deep residual neural network to determine the composition of the target; and establishing the near-infrared spectral machine learning model based on the preprocessed sample data and the improved ResNet-34 neural network algorithm.

[0012] As a preferred embodiment of the rapid real-time analysis method for the composition of coal direct liquefaction oil based on near-infrared spectroscopy and deep learning described in this invention, the rapid real-time analysis method further includes an independent judgment model, comprising: Based on the spectral data of the coal direct liquefaction oil, it is determined whether it deviates from the dataset used during the training of the improved ResNet-34 neural network algorithm, and the original near-infrared spectral data and reaction conditions of the abnormal samples are recorded. The abnormal data is output to the abnormal sample database, and a corresponding work log is generated. The data information of the abnormal sample database can provide data support for the update of the prediction algorithm. The judgment model is established based on principal component analysis and Mahalanobis distance. A low-dimensional principal component space is calculated using principal component analysis, and the Mahalanobis distance of the low-dimensional principal component space is calculated. A judgment threshold is set, and the Mahalanobis distance is used as the judgment criterion. The judgment model judges whether the near-infrared spectrum of any sample deviates from the training dataset of the near-infrared machine learning model, and generates a corresponding work log to store the deviation data.

[0013] As a preferred embodiment of the rapid real-time analysis method for the composition of coal direct liquefaction oil based on near-infrared spectroscopy and deep learning described in this invention, the method includes: systematically collecting near-infrared spectral data obtained under different coal direct liquefaction process conditions, including no less than 7,000 spectral data points, with each data point being associated with raw coal quality analysis data, process condition parameters, and standard method GC×GC verification results; updating the data in real time, including automatically collecting new sample data, AI-assisted semi-automatic annotation, and iterative model updates; and a dedicated near-infrared spectral database containing the solvents used in the coal direct liquefaction process and their changes during the reaction.

[0014] Secondly, embodiments of the present invention provide a near-infrared spectroscopy acquisition device for rapid real-time analysis of the composition of coal direct liquefaction oil, characterized in that: By using a side-stream extraction method, the coal direct liquefaction oil, which is used for the next process, is extracted to the analysis chamber during reactor operation; The acquisition device includes four modules: an acquisition module, an analysis module, and a cleaning module. The acquisition module is used to filter, ultrasonically degas and temperature control the collected liquefied sample to avoid interference from bubbles, temperature and solid particles during spectral acquisition, and to ensure the accuracy of subsequent near-infrared spectral measurements. The acquisition module is used to acquire the spectral signal of the liquefied oil sample in the near-infrared band. It includes an optical probe, a light source, a detector and a flow cell to complete the non-contact online measurement of the flowing sample. The analysis module is used to establish a prediction algorithm model based on sample data of direct coal liquefaction oil, an improved ResNet-34 neural network algorithm, and a Pearson correlation coefficient algorithm. The prediction algorithm model is used to predict the composition data of the direct coal liquefaction oil. The cleaning module is used to ensure the stable operation of the device under complex industrial conditions. It includes a temperature control unit, an air curtain protection unit, and an automatic cleaning unit to ensure that the spectral acquisition step is not affected by temperature fluctuations, oil mist contamination, or pipeline blockage.

[0015] As a preferred embodiment of the near-infrared spectral acquisition device for rapid real-time analysis of the composition of coal direct liquefaction oil described in this invention, the spectral acquisition module is equipped with a dual detector structure, which covers the short-wave and long-wave near-infrared regions respectively, and is combined with a flow cell to adapt to the process conditions of coal liquefaction. The temperature control unit ensures the uniformity of the injection temperature; the automatic cleaning unit prevents the filter and flow cell from becoming clogged and old samples from accumulating over time; and the gas curtain protection unit ensures that the optical window is protected by regular inert gas purging to maintain light transmission performance.

[0016] Thirdly, embodiments of the present invention provide an intelligent monitoring system integrating near-infrared spectroscopy analysis and an industrial control system, wherein: Construct a system that connects a near-infrared spectroscopy device, a pre-trained model, an anomaly detection module, and an industrial control network; The system receives the raw signal data stream output by the near-infrared spectroscopy device through a standardized industrial communication protocol. The raw signal data stream includes the wavelength axis and the corresponding absorbance value. The system's built-in communication adaptation layer automatically handles the interface differences of devices from different manufacturers. The system design model execution engine calls a single inference unit that has packaged the preprocessing algorithm and prediction model, which can convert the collected raw spectral data into standardized prediction results without additional configuration. The system integrates a judgment module to assess whether new samples exceed the range of modeling data. When an anomaly is detected, it not only marks the data category but also triggers a multi-level warning mechanism and transmits the warning information to the central control computer interface.

[0017] As a preferred embodiment of the intelligent monitoring system integrating near-infrared spectroscopy analysis and industrial control system described in this invention, the system includes an intelligent connection function with a coal-to-oil process database. The intelligent connection function includes automatically retrieving historical operating conditions of normal oil products when abnormal data is detected, generating targeted process correction suggestions, and pushing them to the operator workstation via an industrial communication protocol. The abnormal data includes assigning data labels and generating a model optimization suggestion report by analyzing the distribution characteristics of the abnormal data.

[0018] The beneficial effects of this invention are as follows: By combining near-infrared spectroscopy with a deep learning model, this invention achieves rapid, accurate, and real-time analysis of the composition of direct coal liquefaction oil, significantly shortening the time required for traditional laboratory analysis and improving analytical efficiency. An improved ResNet-34 residual neural network is employed, incorporating channel attention mechanisms and deep separable convolutions to enhance the model's ability to extract complex spectral features, improve the resolution accuracy of overlapping peaks, and enhance the prediction accuracy of multi-component content. A preprocessing method combining Pearson correlation coefficient and first-order derivative effectively filters key wavelength features, reduces noise interference, and improves model stability. An anomaly detection model is constructed using principal component analysis and Mahalanobis distance, enabling intelligent identification and early warning of samples exceeding the training range, ensuring the model's reliability in industrial applications. The system integrates a dedicated near-infrared acquisition device with functions such as lateral sampling, temperature control, automatic cleaning, and air curtain protection, adapting to complex high-temperature and high-pressure conditions and ensuring long-term stable operation. Furthermore, through intelligent integration with industrial control systems and process databases, it not only enables online analysis but also automatically generates process correction suggestions when anomalies are detected, supporting iterative model updates. This forms a complete technology chain from data acquisition and intelligent analysis to closed-loop control, providing a powerful intelligent means for quality monitoring and optimization of the direct coal liquefaction process, with significant economic benefits and industrial application value. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 A flowchart illustrating a modeling method for rapid analysis of coal direct liquefaction oil components based on near-infrared spectroscopy, according to an embodiment of this application, is shown.

[0020] Figure 2 A general framework diagram of a near-infrared device for use in a direct coal liquefaction process, according to an embodiment of this application, is shown.

[0021] Figure 3 A control system framework diagram based on near-infrared spectroscopy for rapid analysis of coal direct liquefaction oil components is shown according to this application.

[0022] Figure 4 A framework diagram of the operation flow of a method, system, and apparatus for rapid analysis of coal direct liquefaction oil components based on near-infrared spectroscopy, as provided in this application, is shown. Detailed Implementation Example

[0023] Reference Figures 1-4 This is the first embodiment of the present invention, which provides a simulation system for testing industrial control software, comprising: S1 Obtain x coal direct liquefaction oil samples; S2 Near-infrared spectral data of x coal direct liquefaction oil samples obtained using a near-infrared probe; S3 used GC×GC-MS to analyze x coal direct liquefaction oil samples and obtained compositional data such as group components, ring number, and degree of hydrogenation in the coal direct liquefaction oil; S4 uses PCA to analyze the near-infrared spectra and sample composition data of x samples. After cluster analysis and removal of outlier samples, n spectral samples are obtained after screening. S5 performs first-order differential processing on the selected n near-infrared spectral data to obtain near-infrared spectral data after first-order differential processing. S6 inputs the dataset data and the component data to be predicted into the Pearson correlation coefficient algorithm for correlation analysis, selects the wavelength range with the highest correlation, and this range is the range that the near-infrared spectrometer needs to measure. After the correlation coefficient screening process, the spectral vector is reduced from the original 1600 dimensions to m dimensions, and the dataset is an n×m dimensional feature matrix. S7 uses cross-validation to select the optimal hyperparameters for model training; S8 uses the Kennard-Stone method to divide the dataset into an 8:2 ratio for the dataset and the validation set. S9 inputs the training set of the dataset into the pre-trained improved ResNet-34 neural network model to correlate component information with extracted spectral features. During training, a validation set is used to monitor model performance and prevent overfitting, resulting in the final fine-tuned ResNet-34 neural network model. S10 deploys the finally trained ResNet-34 model to the spectral prediction system. When unknown spectral data needs to be predicted, the system will load the model, preprocess the input spectrum, and directly output its predicted components. Furthermore, step S8 specifically includes: S8-1 performs unified preprocessing of the input dataset, normalizes the preprocessed spectral data, standardizes the predicted properties, and defines the hyperparameter search space. S8-2 uses a hierarchical K-fold partitioning method to divide the dataset into K mutually exclusive subsets (K=5), ensuring that the class distribution of each subset is consistent with the original data, and generating a hyperparameter combination queue. Each combination includes four parts: learning rate, optimizer type, weight decay coefficient, and learning rate scheduling strategy. S8-3 selects the i-th subset as the validation set, and the remaining K-1 subsets are used as the training set. The ResNet-34 network parameters are initialized, the pre-trained weights are loaded, the convolutional layer weights are frozen to retain the pre-trained feature extraction capability, the fully connected layer is trained according to the current hyperparameter configuration, the maximum number of iterations is set to N, and an early stopping mechanism is introduced to terminate the training early when the validation loss has not improved for E consecutive iterations. S8-4 Records Validation Set Classification Accuracy Index (ACC) k ; S8-5 calculates the average verification accuracy of this hyperparameter combination: Repeat steps S8-3 to S8-4 until the average value of the hyperparameter combinations of the above K subsets is traversed. S8-6 compares all hyperparameter combinations An additional optimal combination stability condition is set, and the combination corresponding to the maximum value is selected as the optimal solution; The hyperparameter search space in S8-1 includes: (1) Learning rate: in the interval [1×10⁻ 5 Discrete sampling points are set within 1×10⁻³; (2) Optimizer type: Select from the set {Adam, SGD}; (3) Weight decay coefficient (wd): in [1×10⁻ 5 The gradient regularization strength is set within 1×10⁻³. (4) Learning rate scheduling strategy: Select the annealing method from {StepLR, CosineAnnealing}.

[0024] In addition, the real-time analysis method for the composition of direct coal liquefaction oil of this invention also includes an independent judgment model. When implementing this system, a PCA anomaly detection model is first constructed based on the improved ResNet-34 neural network training dataset. Principal component analysis is then performed using the spectral data of the direct coal liquefaction oil sample. The specific implementation steps are as follows: S1 standardizes the spectral data of the training set. S2 calculates the covariance matrix and performs singular value decomposition, then selects the top 5 principal components that can explain more than 95% of the variance to construct a low-dimensional space. The S3 system automatically calculates the Mahalanobis distance distribution of all samples in the training set in the principal component space and sets the judgment threshold to the mean plus 3 times the standard deviation. When newly acquired near-infrared spectral data of coal direct liquefaction oil arrives, the system first performs the same standardization process as the training set, then projects it onto the constructed principal component space and calculates the coordinate vector of the sample in the space. Subsequently, the system calculates the Mahalanobis distance between this coordinate vector and the center point of the training set, and compares the result with a preset threshold. If the Mahalanobis distance exceeds the threshold, the system immediately determines that the sample's spectral data deviates from the training dataset range of the ResNet-34 neural network algorithm.

[0025] Based on the intelligent analysis model based on deep residual neural networks in the second part of this invention, this invention proposes a near-infrared spectral database specifically for direct coal liquefaction, including: In this embodiment, a database for constructing a near-infrared spectroscopy residual neural network analysis model is used to build the residual neural network analysis model. This database uses coal direct liquefaction solvent oil, coal direct liquefaction diesel oil, coal direct liquefaction naphtha, and coal direct liquefaction aviation kerosene as database sample oils. One or more of these oils are mixed to form the database construction oil. The oil composition is analyzed by GC×GC, and parameters such as peak time, peak shape, peak height, half-maximum width, and peak area corresponding to each component in the near-infrared spectrum are correlated with standard components. This obtains qualitative and quantitative information on the substances in the database construction oil under different ratios, as well as the spectra used to construct the residual neural network analysis model.

[0026] To establish a deep residual neural network model for intelligent analysis of the direct coal liquefaction process, a dedicated near-infrared spectral database was first constructed. This database focuses on the solvents used in the direct coal liquefaction process and their changes during the reaction, by systematically collecting near-infrared spectral data obtained under different direct coal liquefaction process conditions. These different conditions include varying initial hydrogen pressures, reaction temperatures, and catalysts.

[0027] Spectral acquisition utilized a high-precision near-infrared spectrometer, covering a wavelength range of 900-2500 nm, with a resolution better than 8 cm⁻¹ and a signal-to-noise ratio of no less than 50 dB. Sampling employed transmission mode, performing multiple rotational scans on each sample and averaging the results to minimize measurement errors and the influence of sample inhomogeneity. The initial database comprised no fewer than 5000 sets of high-quality spectral data samples.

[0028] Each sample's data structure contains complete spectral curve data points and detailed records of the corresponding solvent information, including solvent type, initial concentration, distribution data of key products after liquefaction reaction, and process parameters corresponding to the sample, including reaction temperature, pressure, catalyst type and dosage.

[0029] The improved ResNet-34 neural network model pre-trained in step S9 of the first part of this invention is based on this dedicated database. The network's input layer is adapted to high-dimensional near-infrared spectral data, and its core includes multiple one-dimensional convolutional operations to extract local spectral features and patterns. A residual connection module is introduced to learn the nonlinear mapping relationship between the input spectrum and the output target, alleviating the gradient vanishing problem in deep network training.

[0030] During the pre-training phase, a solvent composition prediction proxy task was designed to guide the network in learning solvent-related spectral features. After pre-training, the ResNet model was applied to subsequent fine-tuning tasks using transfer learning techniques. To preserve the learned basic spectral features, the first few layers of the network were frozen, and specific layers were fine-tuned.

[0031] The third part of this invention proposes a rapid real-time analysis device for the composition of coal direct liquefaction oil, which is specifically implemented as follows: Sampling module: The online sampling system circulates the sample at a constant flow rate, filters solid particles in real time through two-stage stainless steel filter elements, and simultaneously performs 40 kHz ultrasonic degassing to eliminate bubble interference; a temperature control device is set up to cool down the incoming coal direct liquefaction oil when it exceeds the preset temperature to ensure the accuracy of subsequent near-infrared spectroscopy measurements. Acquisition module: Employs dual detectors, with an InGaAs detector covering the 900–1700 nm band and an HgCdTe detector covering the 1700–2500 nm band; ensuring complete capture of key chemical bond information and equipped with a special flow cell resistant to high pressure and high temperature; sampling is performed via a side-stream extraction method to obtain coal direct liquefaction oil samples from the production process.

[0032] Analysis module: Used to establish a prediction algorithm model based on various sample data, long short-term memory network algorithm and physical information neural network algorithm. The prediction algorithm model is used to predict the composition information of coal direct liquefaction oil. Cleaning Module: Integrates an automatic cleaning device. When the system detects an abnormal pressure difference in the flow cell, it automatically triggers a high-pressure solvent pulse cleaning program, effectively preventing coal liquefaction oil from coking and clogging at high temperatures. Intelligent Diagnostic System: Links online sampling flow rate and ultrasonic degassing parameters to dynamically optimize the sample pretreatment process, ensuring the system continuously outputs reliable analytical results under extreme conditions such as high pressure and high temperature.

[0033] The fourth part of this invention proposes an intelligent monitoring system integrating near-infrared spectroscopy analysis and an industrial control system, the specific implementation of which is as follows: The implementation of this system begins with hardware connection and initialization. The near-infrared spectrometer establishes a physical connection with the system server through a standard industrial interface. The device is configured to continuously output raw spectral data at a sampling frequency of 10Hz. Each data point contains wavelength-absorbance data pairs for all data points within the near-infrared band. The system's built-in communication adaptation layer automatically handles interface differences between devices from different manufacturers. A circular buffer is used to receive these data packets in real time, and CRC32 verification is used to ensure data integrity and avoid packet loss or distortion during transmission. Upon receiving a complete spectral frame, the system automatically triggers the model execution engine. This engine directly calls a pre-trained model that has packaged the preprocessing algorithm and prediction model into a single inference unit. This model is encapsulated in ONNX format, and the system communicates via ONNX. The runtime execution engine loads the model file, inputs the raw spectral data into the unit, and automatically performs end-to-end inference. After the prediction results are generated, the system immediately launches the user interface display module. This interface is developed using the PyQt5 framework. The left side of the main window displays the raw spectral curve in real time, and the right side presents the predicted values ​​intuitively in the form of a dashboard. When key parameters exceed the process threshold, they automatically switch to a red warning state with an alarm prompt. The interface also provides functions such as historical trend analysis, data export, and system configuration to meet the operational needs of different user roles. In addition, the system packages the complete raw spectral data of this measurement, the preprocessed standardized data, the model prediction results, and the confidence index into a structured data package, and stores it in a MySQL cluster database through the ACID transaction mechanism. The database is designed with a classification function, which marks the data as "normal" or "abnormal" according to the degree of deviation between the prediction results and the process standards and the anomaly detection results. At the same time, the system generates a detailed work log, recording the entire link information from data acquisition to result output, including the time consumption, resource utilization, and abnormal events of each link. The log file is archived by date and supports keyword search.

[0034] In addition, to meet the specific needs of industrial environments, the system incorporates an OPC UA client module for integration with existing industrial control systems. Through a configured communication channel, the prediction results are transmitted in real-time to the DCS system in the central control room. Operators can then view the newly added "Near-Infrared Analysis" panel on their regular monitoring screens, displaying oil quality indicators and curves, enabling monitoring of the production process. The system also integrates a judgment module that uses historical normal oil data to construct a principal component model and calculates the Mahalanobis distance of new samples in the principal component space. When the distance exceeds a specified threshold, the data is judged as abnormal. The system not only marks the sample as "abnormal" and stores it in the database but also triggers a warning mechanism: first, a full-screen warning window pops up on the local interface; second, an SNMP message is sent to the central control computer via industrial Ethernet. The system triggers a trap alarm and simultaneously activates on-site audible and visual alarms to ensure timely operator response. In addition, it can connect to the process database. When abnormal data is detected, it automatically invokes a search engine to retrieve a sample set with similar oil quality characteristics from historical normal oil data, obtaining the corresponding operating conditions. This information is then pushed to the operator's workstation HMI interface via the industrial message bus to assist in process adjustment decisions. Abnormal data in the processing results database is assigned special data tags. Subsequent model maintenance and updates can utilize the clustering characteristics of abnormal data to determine targeted model optimization directions. The system adopts a microservice architecture, with each functional module communicating via a lightweight message bus. All interfaces comply with industrial communication standards, ensuring compatibility with industrial control systems.

[0035] To ensure system reliability, multiple fault-tolerance mechanisms are employed: data acquisition threads and model inference threads run separately to avoid real-time conflicts; database operations use a dual-write strategy, simultaneously writing to both local SQLite and a remote MySQL cluster; in the event of a network interruption, the system automatically switches to local caching mode and synchronizes data upon reconnection. The system also features a comprehensive access control system, with operators, process engineers, and system administrators possessing different levels of data access and operation permissions. All critical operations are audited to ensure compliance with industrial safety standards.

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example 1

[0037] The medium-temperature circulating solvent for direct coal liquefaction (BSU unit, boiling range 220-260 ℃) was tested and image acquired according to the test method of this patent. The image was then analyzed by an analysis system and classified into alkanes, cycloalkanes, aromatics, and heteroatom compounds by an analysis database. The following components and the content of representative substances in each component were obtained (list only).

[0038]

[0039] Example 2 The medium-temperature circulating solvent for direct coal liquefaction (BSU unit, boiling range 260-350 ℃) was tested and image acquired according to the test method of this patent. The image was then analyzed by an analysis system and classified into alkanes, cycloalkanes, aromatics, and heteroatom compounds by an analysis database. The following components and the content of representative substances in each component were obtained (list only).

[0040]

[0041] Example 3 The medium-temperature circulating solvent for direct coal liquefaction (factory demonstration unit, boiling range 220-350 ℃) was tested and image acquired according to the test method of this patent. The image was then analyzed by an analysis system and classified into alkanes, cycloalkanes, aromatics, and heteroatom compounds by an analysis database. The following components and the content of representative substances in each component can be obtained (list only).

[0042]

[0043] Example 4 The high-temperature circulating solvent for direct coal liquefaction (factory demonstration unit, boiling range above 350 ℃) was tested and image acquired according to the test method of this patent. The image was then analyzed by an analysis system and classified into alkanes, cycloalkanes, aromatics, and heteroatom compounds by an analysis database. The following components and the content of representative substances in each component can be obtained (list only).

[0044]

[0045] Example 5 Coal-to-liquid diesel (factory demonstration unit, boiling range 140-260 ℃) was tested and images were acquired according to the test method of this patent. The images were then analyzed by an analysis system and classified into alkanes, cycloalkanes, aromatics, and heteroatom compounds by an analysis database. The following components and the content of representative substances in each component were obtained (list only).

[0046]

[0047] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning, characterized in that, include: Data on coal direct liquefaction oil under different reaction conditions were obtained. These data included near-infrared spectral data and compositional data of the coal direct liquefaction oil. The different reaction conditions included different coal melting ratios, different liquefaction solvents, and different initial hydrogen pressures. Compositional data for each coal direct liquefaction oil sample was obtained by GC×GC-MS, and near-infrared spectra of each coal direct liquefaction oil sample were obtained by near-infrared spectroscopy. Each sample and its corresponding family composition data are input into the near-infrared spectral machine learning model to obtain the target category information of the coal liquefaction product to be tested, which is the category information of each component in the coal liquefaction product to be tested. The near-infrared spectroscopy machine learning model takes the near-infrared spectral data of a coal direct liquefaction sample with known composition as input and collects the features of the near-infrared spectral spectrum by machine learning methods. The features of the near-infrared spectral spectrum include peak position, peak shape, peak height, half-width at half-maximum, and peak area parameters. The near-infrared spectral spectrum machine learning model can process the features of the pre-processed near-infrared spectral spectrum and predict the corresponding composition data through the features. The analysis of coal liquefaction oil is performed based on the near-infrared spectral machine learning model, and the analysis database is used for classification and qualitative analysis. A full two-dimensional spectrum analysis was performed based on the full two-dimensional spectrum residual neural network analysis model, and classification and qualitative analysis were conducted using the analysis database.

2. The rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning as described in claim 1, characterized in that: The near-infrared spectral machine learning model includes using an improved ResNet-34 neural network algorithm to predict the model based on the preprocessed near-infrared spectral data of coal direct liquefaction oil. The principle of a residual block of the improved ResNet-34 neural network algorithm is as follows: Where x represents the input sample data. It is the residual function; The ResNet-34 model consists of a total of 4 residual stages, each containing 3 improved residual blocks. Each residual block integrates a channel attention module, uses depthwise separable convolutions to reduce the number of parameters, and the output layer adopts a multi-task learning architecture to predict the content of multiple components.

3. The rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning as described in claim 2, characterized in that: The rapid real-time analysis method also includes a preprocessing model, which establishes a preprocessing module based on the Pearson correlation coefficient method and the first-order differential method, including: According to the preprocessing module, the input near-infrared spectral data of direct coal liquefaction oil is first subjected to first-order differential processing, retaining the filtered wavelength data and deleting redundant spectral data. The filtered wavelength data refers to the data obtained by the Pearson correlation coefficient method, and the formula for the Pearson correlation coefficient method is: Where r is the correlation coefficient, X is the input spectral data, and Y is the data to be predicted; The near-infrared spectral machine learning model includes: processing the spectral information corresponding to the absorption of chemical bonds of the target component in each component through first-order differentiation and extracting correlation coefficients to screen features; extracting the dimension of each sample; inputting the preprocessed features into a deep residual neural network to determine the composition of the target; and establishing the near-infrared spectral machine learning model based on the preprocessed sample data and the improved ResNet-34 neural network algorithm.

4. The rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning as described in claim 3, characterized in that: The rapid real-time analysis method also includes an independent judgment model, including: Based on the spectral data of the coal direct liquefaction oil, it is determined whether it deviates from the dataset used during the training of the improved ResNet-34 neural network algorithm, and the original near-infrared spectral data and reaction conditions of the abnormal samples are recorded. The abnormal data is output to the abnormal sample database, and a corresponding work log is generated. The data information of the abnormal sample database can provide data support for the update of the prediction algorithm. The judgment model is established based on principal component analysis and Mahalanobis distance. A low-dimensional principal component space is calculated using principal component analysis, and the Mahalanobis distance of the low-dimensional principal component space is calculated. A judgment threshold is set, and the Mahalanobis distance is used as the judgment criterion. The judgment model judges whether the near-infrared spectrum of any sample deviates from the training dataset of the near-infrared machine learning model, and generates a corresponding work log to store the deviation data.

5. The rapid real-time analysis method for the composition of direct coal liquefaction oil based on near-infrared spectroscopy and deep learning as described in claim 4, characterized in that: The system systematically collects near-infrared spectral data obtained under different coal direct liquefaction process conditions, including no less than 7,000 spectral data points. Each data point is associated with the raw coal quality analysis data, process condition parameters, and standard method GC×GC validation results. The data is updated in real time, including automatic collection of new sample data, AI-assisted semi-automatic annotation, and model iterative updates. The dedicated near-infrared spectral database contains the solvents used in the coal direct liquefaction process and their changes in the reaction.

6. A near-infrared spectral acquisition device for rapid real-time analysis of the composition of coal direct liquefaction oil, characterized in that: By using a side-stream extraction method, the coal direct liquefaction oil, which is used for the next process, is extracted to the analysis chamber during reactor operation; The acquisition device includes four modules: an acquisition module, an analysis module, and a cleaning module. The acquisition module is used to filter, ultrasonically degas and temperature control the collected liquefied sample to avoid interference from bubbles, temperature and solid particles during spectral acquisition, and to ensure the accuracy of subsequent near-infrared spectral measurements. The acquisition module is used to acquire the spectral signal of the liquefied oil sample in the near-infrared band. It includes an optical probe, a light source, a detector and a flow cell to complete the non-contact online measurement of the flowing sample. The analysis module is used to establish a prediction algorithm model based on sample data of direct coal liquefaction oil, an improved ResNet-34 neural network algorithm, and a Pearson correlation coefficient algorithm. The prediction algorithm model is used to predict the composition data of the direct coal liquefaction oil. The cleaning module is used to ensure the stable operation of the device under complex industrial conditions. It includes a temperature control unit, an air curtain protection unit, and an automatic cleaning unit to ensure that the spectral acquisition step is not affected by temperature fluctuations, oil mist contamination, or pipeline blockage.

7. The near-infrared spectral acquisition device for rapid real-time analysis of the composition of coal direct liquefaction oil as described in claim 6, characterized in that: The spectral acquisition module is equipped with a dual detector structure, covering the short-wave and long-wave near-infrared regions respectively, and combined with a flow cell to adapt to the process conditions of coal liquefaction. The temperature control unit ensures the uniformity of the injection temperature; the automatic cleaning unit prevents the filter and flow cell from becoming clogged and old samples from accumulating over time; and the gas curtain protection unit ensures that the optical window is protected by regular inert gas purging to maintain light transmission performance.

8. An intelligent monitoring system integrating near-infrared spectroscopy analysis and industrial control system, characterized in that: Construct a system that connects a near-infrared spectroscopy device, a pre-trained model, an anomaly detection module, and an industrial control network; The system receives the raw signal data stream output by the near-infrared spectroscopy device through a standardized industrial communication protocol. The raw signal data stream includes the wavelength axis and the corresponding absorbance value. The system's built-in communication adaptation layer automatically handles the interface differences of devices from different manufacturers. The system design model execution engine calls a single inference unit that has packaged the preprocessing algorithm and prediction model, which can convert the collected raw spectral data into standardized prediction results without additional configuration. The system integrates a judgment module to assess whether new samples exceed the range of modeling data. When an anomaly is detected, it not only marks the data category but also triggers a multi-level warning mechanism and transmits the warning information to the central control computer interface.

9. The intelligent monitoring system integrating near-infrared spectroscopy analysis and industrial control system as described in claim 8, characterized in that: The system includes an intelligent connection function with the coal-to-oil process database. The intelligent connection function includes automatically retrieving the operating conditions of historical normal oil products when abnormal data is detected, generating targeted process correction suggestions, and pushing them to the operator workstation via industrial communication protocols. The abnormal data includes assigning data labels and generating a model optimization suggestion report by analyzing the distribution characteristics of the abnormal data.

Images