A deep learning-based oil shale pyrolysis state recognition method

Abstract

The present application is suitable for the technical field of oil shale pyrolysis exploitation, and provides an oil shale pyrolysis state recognition method based on deep learning, which comprises the following steps: collecting the gas data of oil shale in the heating and cracking process at 5 DEG C / min by using a gas sensor electronic nose; performing data cleaning and normalization processing on the data collected by the sensor; dividing the training set, the verification set and the test set by using the gas sensor data; windowing the data, extracting the time sequence features, completing the feature extraction module; and training to obtain a network model based on deep learning. By increasing data cleaning, data normalization and data windowing, the feature engineering effect of the sensor data is improved, the GCN is selected to extract the relationship between the sensors according to the features of the gas sensor array, the LSTM is selected to extract the time sequence features of the reaction according to the oil shale pyrolysis reaction process, and finally the MLP module is used to output the predicted pyrolysis state and the accurate maturity.

Description

Technical Field

[0001] This invention belongs to the field of oil shale pyrolysis mining technology, and particularly relates to a deep learning-based method for identifying the pyrolysis state of oil shale. Background Technology

[0002] Oil shale, a type of rock containing combustible solid organic compounds and high concentrations of ash, commonly known as kerogen, is considered an alternative energy source to petroleum. However, the complex terrain in some countries with abundant reserves makes low-cost extraction difficult. For example, China's oil shale reserves at a depth of 1 kilometer are approximately 7199 × 10⁻⁶. 8 The mining area is mainly located on land and often in complex terrain. Traditional mining methods are inefficient and cause serious ecological damage. Therefore, exploring in-situ mining is of great research significance for China's industrial development and national energy security.

[0003] The main technical route for in-situ extraction involves heating the organic matter in underground oil shale formations and transporting shale oil and natural gas to the surface through production wells. Oil shale pyrolysis is a chemical heating process that produces shale oil and hydrocarbon gases within a temperature range of 350℃–550℃. Continuous monitoring of oil shale pyrolysis products includes assessing pyrolysis maturity, vitrinite reflectance (%), and other parameters. o The values ​​can be roughly divided into three stages: the first stage is less than 0.35, the second stage is from 0.35 to 2.97, and the third stage is greater than 2.97.

[0004] An electronic nose is a device that mimics the olfactory system of mammals. It employs several specific electrochemical sensors and pattern recognition systems. Studies have shown that in-situ heating of organic matter in oil shale can generate shale oil and gas that can be identified by the electronic nose. To address the potential real-time monitoring problem of the pyrolysis state of oil shale under field conditions, we used a portable and low-cost electronic nose to collect data, utilizing a neural network framework composed of GCN, LSTM, and MLP modules. Simultaneously, we monitored the maturity of oil shale pyrolysis and distinguished the three stages of in-situ pyrolysis.

[0005] Oil shale reservoirs can be heated using various methods, such as conductive heating, thermal fluid heating, microwave heating, and combustion heating. The significant energy consumption during the heating process largely determines the extraction cost. For medium- to low-maturity oil shale, temperature and heating time have a significant impact on oil-gas conversion rates and energy consumption. Therefore, real-time monitoring of the in-situ pyrolysis state is a key issue restricting the practical application of in-situ mining technology; accurate real-time monitoring will improve the economic efficiency of oil shale development.

[0006] Currently, the main method for determining the maturity of oil shale is direct measurement of vitrinite reflectance (%Ro). This method is effective and widely used, but the measuring instruments are expensive, and the results can be affected by the lack of reference components in the vitrinite selection, the complexity of sample preparation procedures, and the influence of different experimenters. Another popular method is to predict %R using kinetic data. o The value of T-max (the maximum temperature S2) is derived from pyrolysis tests (e.g., thermogravimetric analysis) or mining data, quantifying the process through thermogravimetric analysis results. However, determining T-max (the maximum temperature S2) cannot be considered a reliable indicator of thermal maturity. Wave-material interactions can also be used to examine rock structure, such as Raman spectroscopy, infrared spectroscopy, and gas chromatography-mass spectrometry (GC-MS), as well as %R. o Indirect estimation is possible. Other less commonly used methods include spore coloration index, solid pitch reflectance, magnetic mineralogy, and spectroscopic microfluorescence. These methods indirectly measure %R using different experimental techniques. o However, these technologies are not conducive to in-situ extraction of oil shale.

[0007] Electronic noses are low-cost and easy to operate, and in recent years they have been widely used in food inspection, medical diagnosis, agricultural research, environmental monitoring, and energy fields as an objective, automated, and non-destructive testing method. The detection mechanism mainly relies on identifying the gas fingerprint entering the system. Furthermore, research on electronic noses in oil and gas drilling, sampling, and oil and gas separation has laid the foundation for their application in the discovery of underground oil and gas deposits. Electronic noses can collect gases generated by the pyrolysis of oil shale in real time and monitor the pyrolysis reaction state of oil shale on-site; this model can be used for on-site detection in in-situ oil shale mining.

[0008] Based on this, this invention proposes a deep learning-based method for identifying the pyrolysis state of oil shale. This method uses an electronic nose gas sensor to record changes in the gases produced during the pyrolysis process of oil shale. A deep learning model combining GCN, LSTM, and MLP is designed to identify the three states experienced by oil shale during pyrolysis at 20℃-650℃, and to predict the maturity %R, a precise indicator representing each of the three states. o . Summary of the Invention

[0009] The purpose of this invention is to provide a deep learning-based method for identifying the pyrolysis state of oil shale. Existing methods for determining the maturity of oil shale primarily involve directly measuring the vitrinite reflectance (%R). o However, the measuring instruments used in this method are very expensive, and the results may be affected by the lack of reference components in the selection of vitrinite, the complexity of the sample preparation steps, and the different experimental personnel.

[0010] This invention is implemented as follows: a deep learning-based method for identifying the pyrolysis state of oil shale, comprising the following steps:

[0011] Step 1: Use a gas sensor electronic nose to collect gas data from the heating and cracking of oil shale at 5℃ / min;

[0012] Step 2: The data collected by the sensors is cleaned and normalized.

[0013] Step 3: Divide the gas sensor data into training, validation, and test sets;

[0014] Step 4: Window the data, extract time series features, and complete the feature extraction module;

[0015] Step 5: Train the deep learning-based network model.

[0016] In a further technical solution, the gas sensor electronic nose in step one consists of 16 gas sensors.

[0017] A further technical solution, the specific steps of step two are as follows:

[0018] S1. %R O The value was fitted to the heating temperature;

[0019] S2. %R of the collected data O The value is calibrated;

[0020] S3. Delete the data collected after 650℃. Since the collection frequency is 50hz, the 50 data points are compressed into 1 data point and mean-variance normalization is used.

[0021] The normalization process for the collected gas sensor data is achieved using the following formula:

[0022]

[0023] In the formula, x' is the value after standardization, μ is the mean of the data, and σ is the variance of the data.

[0024] In a further technical solution, the window size in step four is 10, and the window slides down in steps of 1.

[0025] A further technical solution involves training a deep learning-based network model in step five, with the following specific implementation steps:

[0026] S1. Using a deep learning model composed of GCN, LSTM and MLP modules to identify the pyrolysis state of oil shale, the optimal pyrolysis state identification model is obtained.

[0027] S2. After training until the results converge, save the oil shale pyrolysis state identification parameters, use the saved deep learning model to predict the pyrolysis state of oil shale, and obtain the accuracy maturity %R of the pyrolysis stage classification. O value;

[0028] S3. The deep learning model composed of the GCN module, LSTM module and MLP module includes the number of hidden layers, the number of hidden neurons, the number of graphs, the activation function, the optimizer, the learning rate, the random deactivation layer, the number of iterations and the BATCH-SIZE, and uses the weighted focal-loss loss function to train the hyperparameters involved in the model.

[0029] A further technical solution is provided, where the expressions for the GCN module, LSTM module, and MLP module are as follows:

[0030] GCN module:

[0031]

[0032] In the formula, the processed data x∈R n×1 The input to the GCN module is set to 0.5 for the random deactivation layer, and ReLU is the activation function, ReLU = max(0, x). It is an adjacency matrix with self-loops, where A is the adjacency matrix and I is the adjacency matrix. N identity matrix It is a degree matrix, where the diagonal elements represent the number of edges connected to each node. n=16 indicates 16 gas sensors;

[0033] LSTM module:

[0034] The sequence is input into the LSTM module via the GCN module. The LSTM module processes the sequence from beginning to end, passing it through the input gate and forget gate, and obtaining the value of the information output gate. The value of the information output gate contains the output generated at each previous position. The initial LSTM model is constructed for the input data of an LTMCell, specifically including:

[0035] i t =σ(W i ·[h t-1 x t ]+b i )

[0036] After receiving the data, it needs to be processed using the Sigmoid function. The impact of the input data on the cell structure is i. t W i For the weights of the input gate, i t As the bias term of the input gate, the activation function tanh generates a new candidate vector in the next step. Add it to the LTMCell state, the expression is:

[0037]

[0038] Forgotten Gate f t Determine how LSTMCell forgets information:

[0039] f t =Sigmoid(W f ·[h t-1 x t ]+b f )

[0040] In the formula, x t It is the input at the current moment, h t-1 It is the input of the hidden layer in the previous time step, f t It is the output of the Forgotten Gate, W f It is the weight of the forget gate, b f It is the bias term of the forget gate;

[0041] Updated LSTMCell status C t The expression is:

[0042]

[0043] Output gate O t The output is determined based on the new LTMCell state using the following formula:

[0044] O t =σ(W o ·[h t-1 x t ]+b o )

[0045] Finally, the hidden layer output of the LSTM module at that moment is taken as the final result;

[0046] h t =O t *tanj(C t )

[0047] MLP module:

[0048] Data enters this module, and the output of the fully connected network is obtained by the following expression:

[0049] Y = σ(w·L + b)

[0050] Its output represents the oil shale fracture state category obtained through the Softmax layer, as shown in the following expression:

[0051] Oc = Softmax(Y)

[0052] Alternatively, the precise indicator representing the three states, the maturity %RO value, can be predicted using the following expression:

[0053] Or = Sigmoid(Y)

[0054] The weighted focal loss function used to train the model is expressed as follows:

[0055] FL classification (p c )=-α c (1-p c ) γ log(p c )

[0056] In the formula, p c This represents the probability that a sample belongs to class c. Experimentally, γ was set to 2, and the weighting parameter α... c The ratio is set to 1:1:10, representing the weights of the three categories.

[0057] This invention provides a deep learning-based method for identifying the pyrolysis state of oil shale. By adding data cleaning, data normalization, and data windowing, the feature engineering effect of sensor data is improved. For the characteristics of the gas sensor array, GCN is selected to extract the relationship between sensors. For the oil shale pyrolysis reaction process, LSTM is selected to extract the time series features of the reaction. Finally, the MLP module outputs the predicted pyrolysis state and accurate maturity. This invention can greatly improve the detection efficiency and accuracy of oil shale pyrolysis state. Attached Figure Description

[0058] Figure 1 The algorithm implementation flow of a deep learning-based method for identifying the pyrolysis state of oil shale is provided in this embodiment of the invention.

[0059] Figure 2 The confusion matrix for predicting oil shale pyrolysis state categories using a deep learning-based method for identifying oil shale pyrolysis state is provided in an embodiment of the present invention.

[0060] Figure 3 This invention provides a precise maturity index model for the oil shale pyrolysis process. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0063] Example 1

[0064] A deep learning-based method for identifying the pyrolysis state of oil shale includes the following steps:

[0065] Step 1: Use a gas sensor electronic nose to collect gas data from the heating and pyrolysis of oil shale at 5℃ / min. The oil shale sample was mined from the fourth layer of the Dachengzi mining area in Huadian City, Jilin Province. The gas sensor electronic nose consists of 16 gas sensors, specifically models including TGS2612, TGS2611, TGS2620, TGS2602, TGS2610, TGS2600, TGS2603, WSP1110, WSP2110, MP135, MP905, MP-7, MP-3B, MP-2, MP503, and MP402.

[0066] Step 2: The data collected by the sensor undergoes data cleaning and normalization. The specific steps are as follows:

[0067] S1. Fit the %RO value to the heating temperature to obtain the relationship between temperature and %RO at a heating rate of 5℃ / min. O Fitted curve of the value;

[0068] S2. %R of the collected data O The value is calibrated;

[0069] S3. Due to manual cessation of data collection, the cessation time slightly exceeded the heating time to 650℃. Data collected after 650℃ was deleted. Since the collection frequency was 50Hz, 50 data points were compressed into 1 data point. Mean-variance normalization was used (mean-variance normalization was applied to the compressed data for each sample).

[0070] The normalization process for the collected gas sensor data is achieved using the following formula:

[0071]

[0072] In the formula, x' is the value after standardization, μ is the mean of the data, and σ is the variance of the data;

[0073] Step 3: Use gas sensor data to divide the data into training set, validation set and test set. The total amount of data in the training set, validation set and test set is randomly divided according to the sample ratio of 8:1:1. The above samples are the collected oil shale mineral samples.

[0074] Step 4: Window the data and extract time series features to complete the feature extraction module. The window size is 10 and the window slides down by 1 step.

[0075] Step 5: Train the deep learning-based network model; the specific implementation steps are as follows:

[0076] The S1.GCN module is used to extract features between gas sensors, the LSTM module is used to extract time series features of oil shale pyrolysis, and the MLP module is used for the final decision prediction of the required pyrolysis state and accurate maturity. The deep learning model composed of the GCN module, LSTM module and MLP module is used to continuously train and optimize the pyrolysis state identification model to obtain the optimal oil shale pyrolysis state identification model.

[0077] S2. After training until the results converge, save the oil shale pyrolysis state identification parameters, use the saved deep learning model to predict the pyrolysis state of oil shale, and obtain the accuracy maturity %R of the pyrolysis stage classification. O value;

[0078] S3. The deep learning model composed of the GCN module, LSTM module, and MLP module includes the number of hidden layers, the number of hidden neurons, the number of graphs, the activation function, the optimizer, the learning rate, the random dropout layer, the number of iterations, and the BATCH-SIZE. Specifically, experiments have shown that the optimal GCN module has one hidden layer, the feature dimension of each node in each graph of the GCN module is set to 5, and for 10 data points in a window, each data point generates a 16*5 graph. The LSTM module has one hidden layer with 20 hidden units, the activation function is ReLU, the optimizer is Adam, the learning rate is set to 0.001, the random dropout layer is set to 0.5, the number of iterations is set to 50, the BATCH-SIZE is set to 32, and the hyperparameters involved in the model are trained using a weighted focal-loss loss function.

[0079] The expressions for the GCN module, LSTM module, and MLP module are as follows:

[0080] GCN module:

[0081]

[0082] In the formula, the processed data x∈R n×1 The input to the GCN module has a Dropout value set to 0.5, and ReLU is the activation function, ReLU = max(0, x). It is an adjacency matrix with self-loops, where A is the adjacency matrix and I is the adjacency matrix. N identity matrix It is a degree matrix, where the diagonal elements represent the number of edges connected to each node. n=16 indicates 16 gas sensors;

[0083] LSTM module:

[0084] The sequence is input into the LSTM module via the GCN module. The LSTM module processes the sequence from beginning to end, passing it through the input gate and forget gate, and obtaining the value of the information output gate. The value of the information output gate contains the output generated at each previous position. The initial LSTM model is constructed for the input data of an LTMCell, specifically including:

[0085] i t =σ(W i ·[h t-1 x t ]+b i )

[0086] After receiving the data, it needs to be processed using the Sigmoid function. The impact of the input data on the cell structure is i. t W i For the weights of the input gate, i t As the bias term of the input gate, the activation function tanh generates a new candidate vector in the next step. Add it to the LTMCell state, the expression is:

[0087]

[0088] Forgotten Gate f t Determine how LSTMCell forgets information:

[0089] f t =Sigmoid(W f ·[h t-1 x t ]+b f )

[0090] In the formula, x t It is the input at the current moment, h t-1 It is the input of the hidden layer in the previous time step, f t It is the output of the Forgotten Gate, W f It is the weight of the forget gate, b f It is the bias term of the forget gate;

[0091] Updated LSTMCell status C t The expression is:

[0092]

[0093] Output gate O t The output is determined based on the new LTMCell state using the following formula:

[0094] O t =σ(W o ·[h t-1 x t ]+b o )

[0095] Finally, the hidden layer output of the LSTM module at that moment is taken as the final result;

[0096] h t =O t *tanh(C t )

[0097] MLP module:

[0098] Data enters this module, and the output of the fully connected network is obtained by the following expression:

[0099] Y = σ(w·L + b)

[0100] Its output represents the oil shale fracture state category obtained through the Softmax layer, as shown in the following expression:

[0101] Oc = Softmax(Y)

[0102] Alternatively, the precise indicator representing the three states, the maturity %RO value, can be predicted using the following expression:

[0103] Or = Sigmoid(Y)

[0104] The weighted focal loss function used to train the model is expressed as follows:

[0105] FL classification (p c )=-α c (1-p c ) γ log(p c )

[0106] In the formula, p c This represents the probability that a sample belongs to class c. Experimentally, γ was set to 2, and the weighting parameter α... c The ratio is set to 1:1:10, representing the weights of the three categories.

[0107] The above embodiments of the present invention provide a method for identifying the pyrolysis state of oil shale based on deep learning, such as... Figure 2As shown, the confusion matrix for predicting the pyrolysis state category of oil shale using the scheme of this embodiment is presented. It can be seen that this embodiment obtains gas data of oil shale pyrolysis through an electronic nose; the model can effectively distinguish the three stages of oil shale pyrolysis. Figure 3 As shown, the maturity %RO value model can also predict the precise index of the oil shale pyrolysis process very well.

[0108] In summary, this invention improves the feature engineering effect of sensor data by adding data cleaning, data normalization, and data windowing. For the characteristics of the gas sensor array, GCN is used to extract the relationships between sensors; for the oil shale pyrolysis reaction process, LSTM is used to extract the time series features of the reaction. Not only is the model simple, computationally fast, and low-cost, this invention can also significantly improve the efficiency and accuracy of oil shale pyrolysis state detection.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based oil shale pyrolysis state recognition method, characterized in that, It comprises the following steps: Step one, using gas sensor electronic nose to collect oil shale pyrolysis gas data at 5 ℃ / min heating rate; Step two, data collected by the sensor is subjected to data cleaning and normalization processing; Step three, gas sensor data is used to divide training set, validation set and test set; Step four, windowing is performed on the data, time series features are extracted, and the feature extraction module is completed; Step five, a deep learning-based network model is trained; In step four, the window size is 10, and the window downward sliding step is 1; In step five, the deep learning-based network model is trained, and the specific implementation steps are as follows: S1. Use the deep learning model composed of GCN module, LSTM module and MLP module to continuously train and optimize the pyrolysis state recognition model, and obtain the optimal oil shale pyrolysis state recognition model; S2. Save the oil shale pyrolysis state recognition parameters after training until the result converges, use the saved deep learning model to predict the state of oil shale pyrolysis, and obtain the accurate maturity R of the pyrolysis stage classification O values; S3. The deep learning model composed of GCN module, LSTM module and MLP module includes the number of hidden layers, the number of hidden neurons, the number of graphs, the activation function, the optimizer, the learning rate, the random inactivation layer, the number of iterations and BATCH-SIZE, and the hyperparameters involved in training the model are trained using the weighted focal-loss loss function; The expressions of the GCN module, the LSTM module and the MLP module are as follows: GCN module: ； where the processed data is input into the GCN module with the values of randomly inactivated layers set to 0.5, and ReLU is an activation function, , is an adjacency matrix with self-loops, A is an adjacency matrix, is an identity matrix is a degree matrix whose diagonal elements represent the number of edges connected to each node , n = 16 represents 16 gas sensors; LSTM module: The sequence input by the GCN module is input to the LSTM, the LSTM module processes the sequence from head to tail, passes it through the input gate and the forget gate, and obtains the value of the information output gate, which contains the output generated at each previous position. For input data of an LSTMCell, an initial LSTM model is constructed, which specifically includes: ； After receiving data, it needs to be processed by the Sigmoid function, and the input data has an impact on the unit structure , is the weight term of the input gate, is the bias term of the input gate, and the next step activation function produces a new candidate vector , which is added to the LSTMCell state, and the expression is: ； forget gate decides how much information the LSTM cell forgets: ； wherein, is the input at the current time, is the input of the hidden layer at the previous time, is the output of the forget gate, is the weight of the forget gate, is the bias term of the forget gate; Updated LSTMCell state The expression is: ； Output gate The output is determined from the new LSTM cell state as follows: ； Finally, the final result of the hidden layer output of the LSTM module at this time is taken; ； MLP module: The output of the fully connected network is obtained by the following expression: Its output represents the oil shale cracking state category obtained by the Softmax layer as follows: ； Or predict the accurate index maturity %RO value representing 3 states as follows: ； The weighted focal-loss loss function used in the training model is as follows: ； wherein, represents the probability that the sample belongs to the cth class, γ is set to 2 through experiments, and the weight parameter is set to 1 : 1 : 10, representing the weights of the 3 classes.

2. The deep learning-based oil shale pyrolysis state recognition method according to claim 1, characterized in that, The gas sensor electronic nose in step one is composed of 16 gas sensors.

3. The deep learning-based oil shale pyrolysis state recognition method according to claim 1, characterized in that, The specific steps of step two are as follows: S1. R O values were fitted to heating temperature; S2. Calibrating the R values of the collected data O values; S3. Delete the data collected after 650℃, since the collection frequency is 50hz, 50 data are compressed into 1, and mean variance normalization is used; The following formula is used to realize the normalization processing of the collected gas sensor data: ； In the formula, x' is the value after standardization processing, mu is the average value of the data, and sigma is the variance of the data.

Images