Fault detection method for oil and gas recovery system of integrated energy supply service station in data-scarce scenario

Abstract

The present invention discloses a fault detection method for an oil and gas recovery system of an integrated energy supply service station for data-scarce scenarios. Aiming at data scarcity problems such as no historical fault samples and imbalanced samples among multiple devices in the oil and gas recovery system of an integrated energy supply service station, the present invention first uses a clustering method to cluster normal data with similar distributions in all devices into the same sample cluster, so that a small number of modeling samples originally for a single device are expanded, effectively overcoming the problem of imbalanced samples among multiple devices; then, an unsupervised fault detection model is established based on each sample cluster, and combined with the autoencoder and generative adversarial ideas, the reconstruction error of the fault sample can be amplified, thereby effectively improving the fault detection accuracy and sensitivity in the absence of historical fault data. The present invention can effectively enhance the accuracy and sensitivity of fault detection in scenarios with imbalanced samples among multiple devices and no historical fault data, laying the foundation for safe and reliable operation and intelligent operation and maintenance of the oil and gas recovery system of the energy supply station.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault detection, and is aimed at industrial scenarios with scarce data such as no historical fault samples and imbalanced samples among multiple devices, especially for fault detection of oil and gas recovery systems in integrated energy supply service stations. Background Art

[0002] The integrated energy supply service station is a new type of all-round, multifunctional, intelligent public basic service facility for transportation energy, providing parking, drinking water, public health, energy supply such as electricity, oil and hydrogen, tourism relay, cultural publicity, information consultation and other public services for the public to travel safely and conveniently. It has the characteristics of multiple access devices, wide service objects and large amount of information. Therefore, it is of great significance to carry out integrated energy intelligent operation and maintenance and fault warning diagnosis technology and build a safety assessment model for key equipment at the station end to improve the intelligent operation and maintenance of the site. At the same time, the rapid development of industrial Internet of Things, cloud computing, artificial intelligence technology and intelligent instruments has enabled modern industrial processes to accurately collect and store high-frequency and low-frequency sensor measurement signals, processes, product quality and other structured and unstructured data according to the requirements of actual production and control. Concepts such as Industry 4.0 have been proposed, and artificial intelligence fault detection methods based on big data have gradually replaced traditional mechanism-based anomaly detection methods. Combined with the functional role, structural composition and working characteristics of industrial equipment, it is urgent and necessary to carry out value mining and information extraction based on the big data of equipment to realize equipment status monitoring, abnormal warning, fault diagnosis, life prediction, intelligent maintenance and other work.

[0003] At present, gasoline is still the most widely used energy source both in daily life and in integrated energy supply stations. In gas stations, oil and gas recovery involves the entire process of loading, unloading and refueling cars. Its purpose is to collect the volatile gasoline gas and transform the gas from gas to liquid through one or two methods such as absorption, adsorption, condensation, etc., and then recycle it into gasoline for recycling to reduce oil and gas pollution. The gas recovery system of a gas station generally has three stages, primary, secondary and tertiary gas recovery, which are oil unloading gas recovery, refueling gas recovery and gas emission treatment. Among them, secondary gas recovery is the most frequent stage, and it is also the stage with the most elements involving refueling guns, refueling machines, oil tanks, vehicles and personnel, and the most complicated working principle. The above key equipment may fail. Many typical failures can be directly observed based on the working status of the equipment or instruments, such as gun stoppage, oil gun leakage, fuel dispenser vibration and noise, etc. However, for internal structures with complex structures and precision requirements, some minor changes cannot be detected intuitively. For example, the oil and gas hose is a reverse coaxial hose. Under the long-term corrosion of gasoline, it may cause tiny cracks inside and cause internal leakage. Therefore, data-driven fault detection is crucial.

[0004] The energy supply station itself is a scene with very high safety requirements, so there are few or even zero fault samples in the daily database. However, unsupervised fault detection methods have not been applied to the oil and gas recovery system of the integrated energy supply station. In other industrial scenarios, most of the previous studies on data-driven industrial fault detection are based on supervised learning methods with fault history samples, such as partial least squares (PLS), linear discriminant analysis (LDA) and support vector machine (SVM). These methods usually establish global supervised models, require prior knowledge of fault labels, and cannot adapt to the scenarios where oil and gas recovery system fault samples are scarce or even completely missing; some unsupervised learning-based fault detection methods such as principal component analysis (PCA) and slow feature analysis (SFA) have achieved good results in scenarios without historical fault data, but in the case of unbalanced samples between different devices, such as distributed refueling machines connected to the same oil storage tank, and multiple refueling guns connected to the same refueling machine, but due to the needs of the actual production process, the samples between different refueling guns are unbalanced, some will produce hundreds or thousands of samples, and some will only have dozens, which is not enough to support the establishment of fault diagnosis models. Therefore, it is necessary to propose an oil and gas recovery system fault detection method based on unsupervised learning to address the data-scarce scenario characteristics of the oil and gas recovery system, such as insufficient sample size of a single refueling gun and lack of historical fault data. Summary of the invention

[0005] The purpose of the present invention is to propose a fault detection method for the oil and gas recovery system of an integrated energy supply service station for data-scarce scenarios, in response to the problems of lack of historical fault samples in the daily operation of the oil and gas recovery system of the integrated energy supply service station, insufficient sample size of a single oil gun system for modeling, and the inability of traditional instrument monitoring methods to detect minor fault phenomena in a timely and effective manner. The method first collects the historical normal operating data of the oil and gas recovery system; then, Gaussian mixture clustering is performed on the historical normal operating samples, and the historical normal operating data with similar distribution are clustered into the same cluster by soft clustering. For the test samples, a set of membership degrees belonging to each sample cluster will be obtained; then, an unsupervised fault detection model is established on each historical sample cluster with similar distribution. The unsupervised model combines the reconstruction method of the autoencoder and the generative adversarial idea to amplify the reconstruction error of the fault sample and give an abnormal score to each test sample. The larger the score, the greater the possibility of fault; finally, the membership degree of the test sample and the abnormal score of each unsupervised fault detection model are integrated to output the comprehensive abnormal score of the test sample. According to the score, it is judged whether it is within the normal distribution. If it is within the distribution, the test sample is judged to be normal, otherwise it is judged to be abnormal. In summary, in order to effectively detect minor fault phenomena in the integrated energy supply service station, improve the detection sensitivity, and adapt to scenarios with no historical fault data and unbalanced samples among multiple devices, an unsupervised fault detection method based on Gaussian mixture model and the generation method combining autoencoder with adversarial generation idea (GMM-AE-GAN) was proposed.

[0006] The object of the present invention is achieved through the following technical solution: a method for detecting faults in an oil and gas recovery system of an integrated energy supply service station for data-scarce scenarios, comprising the following steps:

[0007] Collect the oil and gas recovery process variable data of the oil and gas recovery system of the integrated energy supply service station;

[0008] Calculate the membership of the current oil and gas recovery process variable data to K categories; the K categories are obtained by dividing the oil and gas recovery process variable data set based on the historical normal operation of the oil and gas recovery system of the integrated energy supply service station;

[0009] Based on K trained unsupervised anomaly detection models, the anomaly scores of the current oil and gas recovery process variable data for each category are calculated; the K unsupervised anomaly detection models are trained by using the oil and gas recovery process variable data set of the historical normal operation of the oil and gas recovery system of the integrated energy supply service station according to K sub-datasets divided according to K categories.

[0010] The membership degree and anomaly score obtained from each category are fused by inner product to obtain the final anomaly score, and whether the oil and gas recovery system of the integrated energy supply service station is faulty is determined based on the final anomaly score.

[0011] Furthermore, the K categories are obtained by dividing the oil and gas recovery process variable data of the historical normal operation of the oil and gas recovery system of the integrated energy supply service station using a Gaussian mixture model or a K-means clustering method.

[0012] Furthermore, the K categories are obtained by dividing the oil and gas recovery process variable data of the oil and gas recovery system of the integrated energy supply service station during historical normal operation using a Gaussian mixture model. The division method of the K sub-datasets is as follows:

[0013] The posterior probability of each sample in the oil and gas recovery process variable data set of the oil and gas recovery system of the integrated energy supply service station under normal historical operation for K Gaussian elements is calculated respectively, and the category corresponding to each sample is divided according to the Gaussian element corresponding to the maximum posterior probability to form K sub-datasets.

[0014] Furthermore, the unsupervised anomaly detection model includes an encoder E and two decoders, wherein the encoder E is connected to the first decoder D1 to form a first autoencoder AE1, and the encoder E is connected to the second decoder D2 to form a second autoencoder AE2.

[0015] Furthermore, the loss function used in the unsupervised anomaly detection model training includes:

[0016]

[0017]

[0018] in, is the loss function of the first autoencoder AE1, is the loss function of the second autoencoder AE2, X k is the sample matrix of the sub-dataset, representing the k-th sub-dataset, AE1(X k ) is the output of the first autoencoder AE1, AE2(AE1(X k )) is the data reconstructed by AE1 and compressed by E to get Z, which is then reconstructed by AE2. The output is ||*||2 is the L2 norm. AE2(X k ) is the output of the second autoencoder AE2, and n is the number of training times.

[0019] Furthermore, based on the K trained unsupervised anomaly detection models, the anomaly scores of the current oil and gas recovery process variable data for each category are calculated as follows:

[0020]

[0021] where α+β=1, It is the current oil and gas recovery process variable data.

[0022] Furthermore, the oil and gas recovery process variable data includes gas-liquid ratio, gasoline flow rate, gasoline flow rate, oil and gas flow rate, and oil and gas flow rate.

[0023] Furthermore, judging whether the oil and gas recovery system of the integrated energy supply service station fails according to the final abnormality score is specifically as follows:

[0024] If the final anomaly score exceeds the threshold of anomaly detection, it is determined that the oil and gas recovery system of the integrated energy supply service station has a fault, and the threshold of anomaly detection is determined by the ROC curve.

[0025] Furthermore, the oil and gas recovery process variable data set of the oil and gas recovery system of the integrated energy supply service station during the historical normal operation includes oil and gas recovery process variable data of oil products with different grades.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention aims at the scenario where data is scarce in the integrated energy supply service station. One problem is that the sample data of a single gas pump is not enough to support the establishment of a fault diagnosis model, and the other problem is that the supervised model fails due to the lack of fault data in the historical database. An unsupervised fault detection method based on clustering, autoencoder and generative adversarial network structure is proposed. This method fully considers the problem of sample mismatch between devices, and can first identify samples with similar distributions through a clustering strategy; then an unsupervised generation model combining an autoencoder and a generative adversarial network is established, which can effectively extract effective information from samples of normal working conditions and reconstruct them, and thereby distinguish the labels of unknown samples (normal or abnormal). While effectively addressing the above-mentioned problems, the method of the present invention improves the accuracy and sensitivity of fault detection, helps to accurately judge the operating status of each sub-equipment of the oil and gas recovery system, and ensures the safe and reliable operation of oil and gas recovery-related equipment in the integrated energy supply station. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of the arrangement framework of the method of the present invention;

[0028] Figure 2 The present invention is based on the fault simulation in the test data of the oil and gas recovery system of the integrated energy supply service station;

[0029] Figure 3 This is a network structure diagram of an unsupervised fault detection model based on an autoencoder and a generative adversarial network;

[0030] Figure 4 It is the detection result of the method of the present invention and the comparison Kmeans on the test data;

[0031] Figure 5The principal component analysis (PCA) was used to 2 Statistics and spe statistics are used as fault detection results compared with the effects of the inventive method. DETAILED DESCRIPTION

[0032] The present invention provides a method for detecting faults in an oil and gas recovery system of an integrated energy supply service station for data-scarce scenarios, the method comprising the following steps:

[0033] Collect the oil and gas recovery process variable data of the oil and gas recovery system of the integrated energy supply service station;

[0034] Calculate the membership of the current oil and gas recovery process variable data to K categories; the K categories are obtained by dividing the oil and gas recovery process variable data set based on the historical normal operation of the oil and gas recovery system of the integrated energy supply service station;

[0035] Based on K trained unsupervised anomaly detection models, the anomaly scores of the current oil and gas recovery process variable data for each category are calculated; the K unsupervised anomaly detection models are trained by using the oil and gas recovery process variable data set of the historical normal operation of the oil and gas recovery system of the integrated energy supply service station according to K sub-datasets divided according to K categories.

[0036] The posterior probability and anomaly score obtained for each category are fused by inner product to obtain the final anomaly score. The final anomaly score is used to determine whether the oil and gas recovery system of the integrated energy supply service station has a fault.

[0037] The method of the present invention fully considers the problem of sample mismatch between devices, and can first identify samples with similar distribution through a clustering strategy; then establish an unsupervised generation model, which can effectively extract effective information of normal working condition samples and reconstruct them, and distinguish the labels of unknown samples (normal or abnormal) based on this. The present invention is further described in detail below in conjunction with the accompanying drawings and specific examples.

[0038] This embodiment takes the secondary oil and gas recovery system of a comprehensive energy supply service station in Zhejiang Province as an example. The system has been operating stably and safely for a long time. Since it was connected to the energy supply station, no serious failure has occurred, and there are no failure samples in the system. During each refueling operation, secondary oil and gas recovery will be performed, and the relevant data will be stored in the system. The main variables include process variables such as gasoline flow rate, gasoline flow rate, oil and gas flow rate, and refueling gun number.

[0039] like Figure 1 As shown, the present invention is a method for detecting faults in an oil and gas recovery system of an integrated energy supply service station for data-scarce scenarios, comprising the following steps:

[0040] (1) Collect historical data:

[0041] (1.1) The same gas station will supply oil products with different grades (such as 92#, 95#, 98#), which are stored in their own oil tanks. The oil in the tanks is delivered to the user's vehicle by the gas station through the fuel gun. The oil and gas hoses in the gas station and the fuel gun are reverse coaxial hoses (one end of the hose is connected to the oil and gas separator, and the other end is connected to the fuel gun or the oil and gas recovery breakaway valve; the inside of the hose is the gas path, and the inner and outer tubes of the hose are the oil path). Each fuel gun is fixed to deliver a unique oil product. Secondary oil and gas recovery occurs at this stage.

[0042] (1.2) During the secondary oil and gas recovery process, data samples are generated only after refueling operations are performed and are stored in the database in chronological order. Key variables include gas-liquid ratio, gasoline flow rate, gasoline flow rate, oil and gas flow rate, oil and gas flow rate, etc., all of which are average values ​​during the refueling process. In addition, auxiliary variables such as the fuel dispenser number, fuel gun number, and reading time need to be collected. The historical samples collected are all normal operating data. Oil products are calibrated according to different anti-knock properties (octane numbers) and have nothing to do with cleanliness. In other words, the influencing factors of key variables such as flow rate and flow rate during the refueling process are not the oil product label. Therefore, the samples from different fuel guns are grouped together as a system, and the variable matrix is ​​expressed as Indicates that i=1,2,…,N represents the number of samples, and j=1,2,…,M represents the number of variables.

[0043] (2) Data standardization: The above steps treat all oil guns as a whole system to obtain the key variable matrix In order to eliminate the dimensional differences between variables, z-score standardization is used: the variable matrix is ​​standardized by subtracting the mean of the column and dividing it by the standard deviation of the column to obtain the standardized matrix X = X N×M .

[0044] (3) Clustering based on Gaussian Mixture Model (GMM) on known space: In the oil and gas recovery system, each refueling gun has different refueling frequencies and controls different grades of oil, so the data finally obtained is unbalanced. For example, in a month, No. 1 refueling gun finally collects 500 groups of samples, while No. 2 gun controls uncommon grades of gasoline, so there may be only dozens of samples. At this time, the data on a single refueling gun is not enough to support modeling. According to step (1.2), all the data in the oil and gas recovery system can be regarded as a system. In order to more accurately extract the data distribution characteristics on the known space, clustering methods are usually used, requiring the class to be as tight as possible and the class to be as separated as possible. In each class, the data distribution is considered to be more similar than that between classes. The present invention adopts the Gaussian mixture model, which is both a commonly used clustering technology in the industry and a generative model for describing data distribution. The Gaussian mixture model is a soft clustering method. The sample points after projection do not obtain a definite classification label, but the probability for each class, which is an important information. The objective function of its Gaussian mixture model is as follows:

[0045]

[0046] Among them, π k is the probability generated by the kth sub-model (class), Indicates that the data is divided into K categories; k represents the mean of the kth Gaussian element (each class is a Gaussian element), ∑ k refers to the covariance matrix of the kth Gaussian element, p(X|μ k,k ) is the probability density function of the high-dimensional Gaussian distribution, which is defined as follows:

[0047]

[0048] Where M represents the dimension of the sample, that is, the number of variables. Sample x i. The posterior probability of belonging to class k is:

[0049]

[0050] If the sample x i. If the posterior probability on the kth Gaussian element is the largest, then the sample is considered to belong to the kth category. The data matrix is ​​divided into K categories according to the posterior probability, and each category is a sub-dataset.

[0051] (4) Based on the K sub-datasets, K unsupervised anomaly detection models based on autoencoders and generative adversarial training are constructed:

[0052] As a new type of comprehensive transportation service station, the equipment and system of the integrated energy supply service station are in a relatively new state so far. Due to its safety requirements, the frequency of manual inspection is relatively high. Therefore, there is little or no fault data in the system monitored by sensors. In order to solve this problem, the generation method of the autoencoder (AE) and the generative adversarial network (GAN) can be combined to obtain an unsupervised anomaly detection model (AE-GAN) that can distinguish the size of the similarity with the original data spatial distribution. AE-GAN includes an encoder E and two decoders: the first decoder D1 and the second decoder D2. The two decoders D1 and D2 share an encoder E, where the encoder E is connected to the first decoder D1 to form the first autoencoder AE1, and the encoder E is connected to the second decoder D2 to form the second autoencoder AE2. The training of the reconstruction error is completed by the generative adversarial training network. The anomaly detection based on AE and GAN is trained using normal data. After training, the discriminator is used as an anomaly detector. If the input data is distributed differently from the learning data, the discriminator regards it as coming from the generator and classifies it as false data, that is, anomaly. The training process of AE-GAN includes the following steps:

[0053] (4.1) The first stage: The goal is to train two AEs to reconstruct the input X k , given X k After that, the outputs of the two encoders are AE1(X k ) and AE2(X k ), the reconstruction error is measured by the distance, and the loss function of the first stage is:

[0054]

[0055]

[0056] Among them, X k represents the sub-dataset matrix, represents the k-th sub-dataset, and ||*||2 represents the L2 norm.

[0057] (4.2) The second stage: First, the data reconstructed by AE1 is compressed again by E to obtain Z, and then reconstructed by AE2. The output is AE2(AE1(X k )), then under the strategy of adversarial training, AE1 as a generator expects to minimize X k And the secondary reconstruction output AE2(AE1(X k ), and the goal of AE2 as a discriminator is to maximize this difference. So the loss function is:

[0058]

[0059] The purpose of AE2 is to identify AE1(X k ) is the input after AE1 reconstruction, not the real input X k Therefore, AE2 needs to maximize AE2(AE1(X k ) and the real X k In order to deceive AE2, AE1 needs to minimize AE2(AE1(X k )) and real X k The difference between them. Combining the two stages, the final loss function is:

[0060]

[0061]

[0062] Where n is the number of training times.

[0063] According to the K sub-datasets, K unsupervised anomaly detection models based on autoencoders and adversarial generation training are constructed.

[0064] Application phase:

[0065] (5) Calculation of membership of samples in unknown space on Gaussian mixture model:

[0066] In the unknown space, it is not certain whether the data obtained is normal or abnormal. This method can calculate the posterior probability of the current sample for each Gaussian element, that is, the degree of membership, by inputting the data of the unknown space into the Gaussian mixture model obtained in step (3):

[0067] P={P 1,2 ,…, K}

[0068] The higher the membership degree, the more similar the data distribution characteristics are to the data distribution characteristics of the Gaussian element.

[0069] (6) Design of fault detection classifier in unknown space:

[0070] (6.1) Detection phase of the unsupervised anomaly detection model AE-GAN: Samples in the unknown space Input the anomaly detection model trained on the normal operating data in step (4) to obtain the anomaly score in the unknown space:

[0071]

[0072] where α+β=1, Represents the data sample in the current oil and gas recovery system (unknown space). These two parameters determine the sensitivity of the model to anomalies and can be adjusted according to the scenario in actual use. Increasing α, the number of FP (False Positives) decreases and the number of TP (True Positives) increases. In other words, increasing α can increase the accuracy of the model, but at the same time the F1-score also decreases, which means that the recall rate will decrease, that is, the model becomes more conservative. K models have K anomaly scores: S = {S 1,2 ,…, K}.

[0073] (6.2) Determine the final output result: The membership and anomaly score sets obtained in step (5) and step (6.1) are fused by inner product to obtain the final anomaly score:

[0074] S Fin =P·S

[0075] Among them, S Fin Representation sample The final anomaly score.

[0076] (7) Model testing: After the model training is completed, the performance of the model needs to be tested and evaluated using test data. The model performance is evaluated using four indicators: accuracy, F1-Score, AUC index, and Matthews coefficient. The threshold of anomaly detection is determined by the ROC curve. The specific calculation formula and meaning are as follows:

[0077] (7.1) Accuracy ACC:

[0078]

[0079] The ratio of the number of correctly predicted samples (right) to the total number of samples (all) among all samples is very intuitive, but it is particularly affected by sample imbalance and sample quantity. For example, during training or testing, the number of faulty data (negative samples) and normal data (positive samples) are often artificially divided. There may be too much faulty data or too much normal data, resulting in unreliable accuracy.

[0080] (7.2)F1-Score:

[0081]

[0082] Among them, Precision = TP / (TP+FP) is the accuracy; Recall = TP / (TP+FN) is the recall rate. The closer the F1-Score is to 1, the better the performance. TP represents the number of correct predictions for positive samples; TN represents the number of correct predictions for negative samples; FN represents the number of incorrect predictions for positive samples; and FP represents the number of incorrect predictions for negative samples.

[0083] (7.3) AUC indicator: The AUC value is the area covered by the ROC curve. The larger the AUC (0 to 1), the better the classification effect of the classifier;

[0084] (7.4) Matthews correlation coefficient (MCC):

[0085]

[0086] MCC∈[-1,1], the closer the value is to 1, the better the performance.

[0087] The following table is the experimental results of this method and other comparative models for fault detection of oil and gas recovery systems in comprehensive energy supply service stations. The system collected all data from September 2021, with a total of 9434 samples, of which 2500 samples were extracted as data sets in unknown space for testing, and the remaining 6934 samples were used as data sets in known space for model training. In the unknown space, since all the data in the system are normal data, in order to simulate the fault conditions in actual scenarios and consider the coupling relationship between variables, the M-dimensional samples are first reduced to 1 dimension through principal component analysis (PCA) (51% of the information is retained at this time), and then a step signal is added to the one-dimensional space to simulate the occurrence of a fault, and the information in the one-dimensional space where the fault is added is reflected back to the M dimension as the fault data of the original space, such as Figure 2 As shown, the fault data simulation is selected on the 1000th to 1500th samples.

[0088] Table 1 shows the detection results of the method of the present invention compared with other fault detection methods for the oil and gas recovery system of the integrated energy supply station (the best results in the table are indicated in bold). The Kmeans-AE-GAN model changes the Gaussian mixture clustering of the method proposed by the present invention into Kmean clustering. Global represents a global model without pre-clustering of the known space. T 2 (PCA), spe(PCA) and T 2 &spe(PCA) represents the fault monitoring model using the PCA method, which uses T 2 Statistics, spe statistics, and the results of the fault detection strategy that monitors both statistics simultaneously.

[0089] Table 1 Fault detection results

[0090]

[0091] Combination Figure 4 and Figure 5 It can be seen that compared with the global unsupervised model without clustering and the PCA detection results, the results of the method of the present invention are obviously better. The effectiveness and superiority of the method proposed in the present invention are verified by actual industrial data. The method proposed in the present invention can not only deal with the problem of data scarcity such as the small sample size of a single device that is insufficient to support modeling and no historical fault data, but also improve the accuracy and sensitivity of fault detection, which is helpful to accurately judge the operating status of each sub-equipment of the oil and gas recovery system, and finally lays the foundation for the safe and reliable operation and intelligent operation and maintenance of the secondary oil and gas recovery system of the comprehensive energy supply service station.

[0092] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for fault detection of oil and gas recovery system in integrated energy supply service station for data-scarce scenarios, characterized in that: The method comprises the following steps: Collect the oil and gas recovery process variable data of the oil and gas recovery system of the integrated energy supply service station; Calculate the membership of the current oil and gas recovery process variable data to K categories; the K categories are obtained by dividing the oil and gas recovery process variable data set based on the historical normal operation of the oil and gas recovery system of the integrated energy supply service station; Based on K trained unsupervised anomaly detection models, the anomaly scores of the current oil and gas recovery process variable data for each category are calculated; the unsupervised anomaly detection model includes an encoder E and two decoders, wherein the encoder E is connected to the first decoder D1 to form a first autoencoder AE1, and the encoder E is connected to the second decoder D2 to form a second autoencoder AE2; the K unsupervised anomaly detection models are obtained by respectively training the oil and gas recovery process variable data set of the oil and gas recovery system of the integrated energy supply service station in normal operation according to K sub-data sets divided according to K categories; The membership degree and anomaly score obtained from each category are fused by inner product to obtain the final anomaly score, and whether the oil and gas recovery system of the integrated energy supply service station is faulty is determined based on the final anomaly score.

2. The method according to claim 1, characterized in that: The K categories are obtained by dividing the oil and gas recovery process variable data of the historical normal operation of the oil and gas recovery system of the integrated energy supply service station using a Gaussian mixture model or a K-means clustering method.

3. The method according to claim 1, characterized in that The K categories are obtained by dividing the oil and gas recovery process variable data of the oil and gas recovery system of the integrated energy supply service station during historical normal operation using a Gaussian mixture model. The division method of the K sub-datasets is as follows: The posterior probability of each sample in the oil and gas recovery process variable data set of the oil and gas recovery system of the integrated energy supply service station under normal historical operation for K Gaussian elements is calculated respectively, and the category corresponding to each sample is divided according to the Gaussian element corresponding to the maximum posterior probability to form K sub-datasets.

4. The method according to claim 3, characterized in that The loss functions used in the unsupervised anomaly detection model training include: in, is the loss function of the first autoencoder AE1, is the loss function of the second autoencoder AE2, X k is the sample matrix of the sub-dataset, representing the k-th sub-dataset, AE1(X k ) is the output of the first autoencoder AE1, AE2(AE1(X k )) is the data reconstructed by AE1 and compressed by E to get Z, which is then reconstructed by AE2. The output is ||*||2 is the L2 norm. AE2(X k ) is the output of the second autoencoder AE2, and n is the number of training times.

5. The method according to claim 4, characterized in that Based on K trained unsupervised anomaly detection models, the anomaly scores of the current oil and gas recovery process variable data for each category are calculated as follows: where α+β=1, It is the current oil and gas recovery process variable data.

6. The method according to claim 1, characterized in that The oil and gas recovery process variable data include gas-liquid ratio, gasoline flow rate, gasoline flow rate, oil and gas flow rate, and oil and gas flow rate.

7. The method according to claim 1, characterized in that The method of judging whether the oil and gas recovery system of the integrated energy supply service station fails according to the final abnormality score is as follows: If the final anomaly score exceeds the threshold of anomaly detection, it is determined that the oil and gas recovery system of the integrated energy supply service station has a fault, and the threshold of anomaly detection is determined by the ROC curve.

8. The method according to claim 1, characterized in that The oil and gas recovery process variable data set of the historical normal operation of the oil and gas recovery system of the integrated energy supply service station includes oil and gas recovery process variable data of oil products with different grades.

Images