A method and system for detecting unknown faults in a gas turbine control system

Abstract

The present invention relates to a method and system for detecting unknown faults in a gas turbine control system. The method comprises: constructing a normal data set composed of normal signals and constructing a test data set; constructing an unsupervised learning model and a unigram classifier, using the normal data set to train the unsupervised learning model, and obtaining a trained feature extractor; inputting the normal data set into the feature extractor to obtain a normal data feature set, and using the normal data feature set to train the unigram classifier; inputting the test data set into the unigram classifier for testing, and if the test result meets the preset evaluation index, the training of the feature extractor and the unigram classifier is completed; inputting the sensor signal on the gas turbine into the feature extractor and the unigram classifier to obtain the fault detection result. Compared with the prior art, the present invention improves the accuracy and reliability of the multi-sensor fault signal detection result of the gas turbine, so that the gas turbine control system has certain advantages such as intelligent diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of gas turbine control system fault detection, and in particular to a gas turbine control system unknown fault detection method and system based on self-supervisory feature learning and unary classification. Background Art

[0002] The system and control of the gas turbine control system itself are already very complex, and they are even more challenged in extreme weather. Timely and accurate diagnosis of gas turbine control system faults can not only avoid losses and improve efficiency, but also help provide stable power supply for people living in bad weather. The natural environment of global warming requires us to achieve carbon neutrality as soon as possible and slow down the speed and magnitude of global warming. Therefore, the energy system needs to play a higher efficiency, and the intelligence of the energy system is conducive to precise control and efficiency improvement. Intelligent perception requires the installation of a large number of state perception sensors. Among the statistical energy system faults, the frequency of sensor failures during use is the highest. Fault detection is the first step in the four steps of state monitoring: anomaly detection, fault classification, fault isolation, and fault elimination. It also helps to realize intelligent control and prevent attacks. Timely and accurate fault diagnosis is conducive to realizing the state detection of the energy system and ensuring the safe operation of the energy system.

[0003] The problems of gas turbine control system sensor fault detection are as follows:

[0004] 1) There is little real fault data used for gas turbine control system fault detection, and the original sampling data is a large amount of normal data.

[0005] 2) The raw data comes from multiple sensors in the gas turbine control system, and the data volume is large. In addition, the fault data and normal data overlap with each other, and the classifier algorithm cannot distinguish the original distribution.

[0006] 3) Simple accuracy evaluation criteria cannot achieve satisfactory detection goals. For example, if there is only 1% fault data, the final detection accuracy is at least 99%, which is obviously not enough for the fault detection task.

[0007] Due to the availability of big data, supervised learning has achieved great success in classification. If a supervised neural network classification algorithm is used, samples with known classification labels need to be used during the training process, and unknown faults, that is, samples without labels, cannot be processed. In addition, using a small amount of fault data and a large amount of normal data corresponds to an unbalanced classification problem, which challenges the effectiveness of fault detection of supervised learning methods. Summary of the invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method and system for detecting unknown faults in a gas turbine control system, so as to improve the accuracy and reliability of the multi-sensor fault signal detection results of the gas turbine in the face of category imbalance, different operating conditions, different fault types, etc., so that the gas turbine control system has a certain intelligent diagnostic capability.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A method for detecting unknown faults in a gas turbine control system comprises the following steps:

[0011] S1. Obtain normal signals of various sensors on the gas turbine, construct a normal data set consisting of normal signals, superimpose fault signals on some normal signals and construct a test data set;

[0012] S2. Construct an unsupervised learning model and a unary classifier. The unsupervised learning model is used for feature extraction. The unsupervised learning model is trained using a normal data set to obtain a trained feature extractor.

[0013] S3, inputting the normal data set into the feature extractor, forming the extracted features into a normal data feature set, and using the normal data feature set to train a unigram classifier to obtain a trained unigram classifier;

[0014] S4, input the test data set into the feature extractor, compose the extracted features into a fault data feature set, and use the fault data feature set to test the unary classifier. If the test result meets the preset evaluation index, the training of the feature extractor and the unary classifier is completed, otherwise, repeat steps S1-S4;

[0015] S5. Input the sensor signal on the gas turbine into the feature extractor, input the extracted features into the unary classifier, and obtain the fault detection result.

[0016] Furthermore, constructing a normal data set consisting of normal signals is specifically as follows:

[0017] Acquire original signals of multiple sensors in the same time period T, where the types of the original signals are normal signals and fault signals;

[0018] The original signal of each sensor is divided according to the preset step length to obtain N·M signal sequences. Each signal sequence is the signal of a sensor in a time interval. M=L / Δl=T / Δt, where N is the type of sensor, L is the length of the original signal, Δl is the step length, and Δt is the time interval.

[0019] The signal sequences within the same time interval are combined to obtain a sample, and a total of M samples are obtained, each sample including N signal sequences.

[0020] Furthermore, the unsupervised learning model is a recurrent neural network, which uses an unsupervised triplet loss function based on time series, as follows:

[0021]

[0022] Where σ is the sigmoid function, x a represents the anchor point, x pos represents a positive sample, represents the kth negative sample, θ represents the model parameters, f(*,θ) represents the result of * inputting the unsupervised learning model, and K represents the number of preset negative samples.

[0023] Furthermore, during the training of the unsupervised learning model, the selection rules for the anchor point, positive sample, and negative sample triplets are as follows:

[0024] Set the length l of anchor points, positive samples and negative samples a , l pos and l neg , and the number of negative samples K;

[0025] Combine the normal signals of a sensor in all time intervals to obtain a subsequence, and obtain N subsequences in total. Each subsequence has a length of L and contains M segments of normal signals.

[0026] Randomly select a subsequence and randomly select a length l in the subsequence a A signal segment is used as the anchor point x a ;

[0027] At anchor point x a Randomly select a length of l pos A normal signal is taken as a positive sample x pos ;

[0028] Randomly select K subsequences from the remaining N-1 subsequences, and randomly select a subsequence of length l in each selected subsequence. neg A normal signal is taken as a negative sample x neg , where any negative sample The time interval and positive sample x pos There is overlap between the time periods.

[0029] Furthermore, the negative sample The time interval and positive sample x pos The overlap range between the time intervals is λ 1 ·Δl~λ2 Δl, λ 1 and λ 2 The preset scaling factor.

[0030] Furthermore, the lengths of the anchor points, positive samples and negative samples are l a , l pos and l neg are all integer multiples of Δl, and the proportional coefficient is an integer;

[0031] The specific anchor point selection is: a is the window width, randomly select a segment of the signal from the M segments of normal signal as the starting point, and use the window to intercept a segment with a length equal to l a The normal signal is x a ;

[0032] The specific selection of positive samples is: pos is the window width, in x a A random signal is selected as the starting point, and a window is used to intercept a length equal to l pos and belongs to x a The normal signal is x pos ;

[0033] The specific selection of negative samples is: neg is the window width, and selects Among them, select To: OK The corresponding subsequence is randomly selected from the M normal signals of the subsequence with a length of l neg And the time interval is x a There exists λ in the time interval 1 ·Δl~λ 2 ·Δl overlaps the normal signal as

[0034] Furthermore, the unary classifier is based on a local anomaly factor, which is defined as:

[0035]

[0036]

[0037] rd k (x,o)=max{k-distance(o),d(x,o)}

[0038] Among them, LOF k x) represents the local anomaly factor of feature x, lrd k x) represents the local reachability density of feature x, rd k(x,o) represents the local reachable distance of feature x, k-distance(o) represents the distance between o and the farthest point among the k nearest neighbors of feature x, N k (x) represents the set of k nearest neighbors of sample x, o represents a sample point among the k nearest neighbors, and d(x,o) represents the Euclidean distance between feature x and o.

[0039] Furthermore, the test data set is constructed as follows:

[0040] A normal data set is obtained, multiple samples are selected from the normal data set, multiple signal sequences are selected from each sample, fault signals are superimposed on these signal sequences, and a test data set is obtained, wherein the test data set includes multiple samples, each sample includes at least one fault, and the number of each fault is at least one.

[0041] Furthermore, in step S4, the evaluation indicators include but are not limited to recall rate, precision rate, F score and false alarm rate.

[0042] A gas turbine control system unknown fault detection system, comprising:

[0043] A data set construction module is used to obtain normal signals of various sensors on the gas turbine, construct a normal data set composed of normal signals, superimpose fault signals on some normal signals and construct a test data set;

[0044] The first model module is used to construct an unsupervised learning model and use a normal data set to train the unsupervised learning model to obtain a trained feature extractor, wherein the unsupervised learning model is used for feature extraction;

[0045] The second model module is used to construct a unary classifier and train the unary classifier using a normal data feature set to obtain a trained unary classifier, wherein the normal data feature set is composed of features extracted from a normal data set input feature extractor;

[0046] A test evaluation module, used to test the unary classifier using a fault data feature set and determine whether the test result meets a preset evaluation index, wherein the fault data feature set is composed of features extracted by inputting a test data set into a feature extractor;

[0047] The detection module is used to input the sensor signal on the gas turbine into the feature extractor, input the extracted features into the unary classifier, and obtain the fault detection result.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) Because a large amount of normal data is available in real-world data sets, while abnormal data is unknown under actual working conditions, the present invention uses a univariate classification method for fault detection, which only trains on a large amount of normal data and then tests various types of fault signals. Moreover, the simple classification method requires less computing resources and time.

[0050] (2) The present invention processes various complex fault signals by using a self-supervised feature extractor. The self-supervised feature extractor can effectively utilize a large amount of normal data under various working conditions without manual labeling. The feature extractor can conveniently and effectively extract the features of sensor data for unary classification and improve the accuracy and robustness of fault detection.

[0051] (3) The unsupervised learning model is a recurrent neural network, which uses an unsupervised triplet loss function based on time series. This has the advantage of flexibly using a large amount of normal data to learn faults with obvious characteristics.

[0052] (4) The unary classifier is based on local anomaly factors, which can bring the advantages of flexible handling of various faults and improving the accuracy of fault feature classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0054] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. Obviously, the described embodiment is only an embodiment of a part of the present invention, not all embodiments, and the protection scope of the present invention is not limited to the following embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0055] As used herein, "one embodiment" or "embodiment" refers to a particular feature, structure, or characteristic that may be included in at least one implementation of the invention. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to the process, method, product, or apparatus.

[0056] This specification provides method operation steps such as embodiments or flowcharts, but more or fewer operation steps may be included based on conventional or non-creative labor. The order of steps listed in the embodiments is only one way of executing the order of many steps and does not represent the only execution order. When the actual system or server product is executed, it can be executed in sequence or in parallel (for example, in a parallel processor or multi-threaded processing environment) according to the method shown in the embodiments or the drawings, or the execution order of steps without timing restrictions can be adjusted.

[0057] Embodiment 1:

[0058] The present application provides a method for detecting unknown faults in a gas turbine control system. Figure 1 As shown, the following steps are included:

[0059] S1. Obtain normal signals of various sensors on the gas turbine, construct a normal data set consisting of normal signals, superimpose fault signals on some normal signals and construct a test data set;

[0060] S2. Construct an unsupervised learning model and a unary classifier. The unsupervised learning model is used for feature extraction. The unsupervised learning model is trained using a normal data set to obtain a trained feature extractor.

[0061] S3, inputting the normal data set into the feature extractor, forming the extracted features into a normal data feature set, and using the normal data feature set to train a unigram classifier to obtain a trained unigram classifier;

[0062] S4, input the test data set into the feature extractor, compose the extracted features into a fault data feature set, and use the fault data feature set to test the unary classifier. If the test result meets the preset evaluation index, the training of the feature extractor and the unary classifier is completed, otherwise, repeat steps S1-S4;

[0063] S5. Input the sensor signal on the gas turbine into the feature extractor, input the extracted features into the unary classifier, and obtain the fault detection result.

[0064] A unary classifier is a unigram classifier that attempts to identify objects of a specific class among all objects by learning from training data that only contains objects of that class. In this application, a normal signal is used to train a unary classifier to obtain a unary classifier that can identify normal signals. In this way, since all types of faults do not belong to normal signals, various types of fault signals can be identified by the unary classifier, and the detection of unknown faults is also completed. Subsequently, other solutions can be used to further confirm the specific type of fault. The feature extractor obtained by training the unsupervised learning model is used to extract effective features from the data, thereby enhancing the accuracy of subsequent unigram classification.

[0065] In step S1, constructing a normal data set consisting of normal signals is specifically as follows:

[0066] S101, obtaining original signals of multiple sensors in the same time period T, where the types of the original signals are normal signals and fault signals;

[0067] The various sensors on the gas turbine may include sensors with hardware redundancy, sensor networks with physical relationships on the same component, etc., wherein the sensors with hardware redundancy include pressure sensors, temperature sensors, and rotation speed sensors.

[0068] S102, dividing the original signal of each sensor according to a preset step length to obtain N·M signal sequences, each signal sequence is a signal of a sensor in a time interval, M=L / Δl=T / Δt, where N is the type of sensor, L is the length of the original signal, Δl is the step length, and Δt is the time interval;

[0069] S103 . Combining signal sequences within the same time interval to obtain a sample, obtaining M samples in total, each sample including N signal sequences.

[0070] In step S1, constructing the test data set is specifically as follows:

[0071] S104. Obtain a normal data set, select multiple samples from the normal data set, select multiple signal sequences from each sample, superimpose fault signals on these signal sequences, and obtain a test data set, wherein the test data set includes multiple samples, each sample includes at least one fault, and the number of each fault is at least one.

[0072] Among them, the superimposed fault signals can be steps, short circuits, drifts, noise, periodic faults, etc. Since they do not appear in the normal data set, these faults are all unprecedented faults for the univariate classifier. Considering that the proportion of fault signals when the gas turbine is working is small, the proportion of fault signals in the test data set is less than 5%, which is completely in line with the actual application scenario. It should also be noted that in this embodiment, sensor data under start-stop conditions, variable load conditions, no-load conditions and other working conditions are selected to construct a normal data set. Considering that sensor fault detection is mainly aimed at variable load conditions, the test data set is constructed based on the sensor data under variable load conditions, that is, the working condition of the data in the test data set is a variable load condition.

[0073] Moreover, considering that when a fault occurs, it may be that a certain sensor fails, or multiple sensors fail simultaneously or at different times, the test data set includes coupled faults such as multiple numbers of fault signals, and the same number of faults but different fault types and locations.

[0074] In step S2, the constructed unsupervised learning model is a recurrent neural network, and an unsupervised triple loss function based on time series is used, as follows:

[0075]

[0076] Where σ is the sigmoid function, x a represents the anchor point, x pos represents a positive sample, represents the kth negative sample, θ represents the model parameters, f(*,θ) represents the result of * inputting the unsupervised learning model, and K represents the number of preset negative samples. The purpose of the loss function is to make x a and x pos The representation of x is as similar as possible, so that a and x neg The representations are as different as possible.

[0077] During the training process of the unsupervised learning model, the selection rules for the anchor point, positive sample, and negative sample triplets are as follows:

[0078] (1) Set the length l of anchor points, positive samples, and negative samples a , l pos and l neg , and the number of negative samples K;

[0079] (2) The normal signals of a sensor in all time intervals are combined to obtain a subsequence, and a total of N subsequences are obtained. Each subsequence has a length of L and contains M segments of normal signals;

[0080] (3) Randomly select a subsequence and randomly select a length l in the subsequence a A signal segment is used as the anchor point x a ;

[0081] (4) At anchor point x a Randomly select a length of l pos A normal signal is taken as a positive sample x pos ;

[0082] (5) Randomly select K subsequences from the remaining N-1 subsequences, and randomly select a subsequence of length l from each selected subsequence. neg A normal signal is taken as a negative sample x neg , where any negative sample The time interval and positive sample x pos There is overlap between the time periods.

[0083] Among them, negative samples The time interval and positive sample x pos The overlap range between the time intervals is λ1 ·Δl~λ 2 Δl, λ 1 and λ 2 The preset scaling factor.

[0084] Moreover, the lengths of anchor points, positive samples, and negative samples are l a , l pos and l neg They are all integer multiples of Δl, and the proportionality coefficient is an integer;

[0085] The specific anchor point selection is: a is the window width, randomly select a segment of the signal from the M segments of normal signal as the starting point, and use the window to intercept a segment with a length equal to l a The normal signal is x a ;

[0086] The specific selection of positive samples is: pos is the window width, in x a A random signal is selected as the starting point, and a window is used to intercept a length equal to l pos and belongs to x a The normal signal is x pos ;

[0087] The specific selection of negative samples is: neg is the window width, and selects Among them, select To: OK The corresponding subsequence is randomly selected from the M normal signals of the subsequence with a length of l neg And the time interval is x a There exists λ in the time interval 1 ·Δl~λ 2 ·Δl overlaps the normal signal as

[0088] It is understandable that the constructed recurrent neural network also involves the determination of hyperparameters such as batch size, learning rate, optimizer and training cycle. The learning rate adjustment strategy is to attenuate at a fixed step size. Based on a normal data set, backpropagation can be used to converge to the optimal result and obtain the training weights of each layer of the network.

[0089] In step S2, the constructed unary classifier is based on the local anomaly factor, which is defined as:

[0090]

[0091]

[0092] rdk (x,o)=max{k-distance(o),d(x,o)}

[0093] Among them, LOF k (x) represents the local anomaly factor of feature x, which is a comparison score given by comparing the local density of a sample with the density of its k nearest neighbors. k (x) represents the local reachability density of feature x, rd k (x,o) represents the local reachable distance of feature x, k-distance(o) represents the distance between o and the farthest point among the k nearest neighbors of feature x, N k (x) represents the set of k nearest neighbors of sample x, o represents a sample point in the k nearest neighbors, d(x,o) represents the Euclidean distance between feature x and o; when x and o are far apart, the reachable distance is itself, i.e. d(x,o); when x and o are close, it is the distance between o and the farthest point in the k nearest neighbors, i.e. k-distance(o). LOF≈1 represents a density similar to that of neighbors, LOF<1 represents a greater density than that of neighbors (inliers), and LOF>1 represents a smaller density than that of neighbors (outliers).

[0094] In arrangements S2-S3, when training unsupervised learning models and unigram classifiers, the data set can be divided into a training set and a test set, which can be understood by relevant practitioners and will not be elaborated here.

[0095] In step S4, the evaluation indicators include but are not limited to recall rate, accuracy, F score and false alarm rate. The test set and multiple evaluation indicators can be used to evaluate whether the fault detection results of the feature extractor and the unary classifier are qualified, so as to determine whether to continue to adjust the feature extractor and the unary classifier.

[0096] It should be noted that the present application can be implemented in software and / or a combination of software and hardware, for example, can be implemented using an application specific integrated circuit (ASIC), a general purpose computer or any other similar hardware device. In one embodiment, the software program of the present application can be executed by a processor to implement the steps or functions described above. Similarly, the software program of the present application (including relevant data structures) can be stored in a computer-readable recording medium, for example, a RAM memory, a magnetic or optical drive or a floppy disk and similar devices. In addition, some steps or functions of the present application can be implemented using hardware, for example, as a circuit that cooperates with a processor to perform each step or function.

[0097] In addition, a part of the present application may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present application through the operation of the computer. The program instruction for calling the method of the present application may be stored in a fixed or removable recording medium, and / or transmitted through a data stream in a broadcast or other signal-bearing medium, and / or stored in a working memory of a computer device that runs according to the program instruction. Here, according to an embodiment of the present application, a device is included, the device including a memory for storing computer program instructions and a processor for executing program instructions, wherein, when the computer program instruction is executed by the processor, the device is triggered to run the method and / or technical solution based on the aforementioned multiple embodiments according to the present application.

[0098] The present application also provides a gas turbine control system unknown fault detection system, comprising:

[0099] A data set construction module is used to obtain normal signals of various sensors on the gas turbine, construct a normal data set composed of normal signals, superimpose fault signals on some normal signals and construct a test data set;

[0100] The first model module is used to build an unsupervised learning model and use a normal data set to train the unsupervised learning model to obtain a trained feature extractor. The unsupervised learning model is used for feature extraction;

[0101] The second model module is used to construct a one-element classifier and train the one-element classifier using a normal data feature set to obtain a trained one-element classifier, where the normal data feature set is composed of features extracted from a feature extractor inputted into a normal data set;

[0102] A test evaluation module is used to test the unary classifier using a fault data feature set and determine whether the test result meets a preset evaluation index. The fault data feature set is composed of features extracted by inputting the test data set into the feature extractor;

[0103] The detection module is used to input the sensor signal on the gas turbine into the feature extractor, input the extracted features into the unary classifier, and obtain the fault detection result.

[0104] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for detecting unknown faults in a gas turbine control system, characterized in that: The following steps are involved: S1. Obtain normal signals of various sensors on the gas turbine, construct a normal data set consisting of normal signals, superimpose fault signals on some normal signals and construct a test data set; S2. Construct an unsupervised learning model and a unary classifier. The unsupervised learning model is used for feature extraction. The unsupervised learning model is trained using a normal data set to obtain a trained feature extractor. S3, inputting the normal data set into the feature extractor, forming the extracted features into a normal data feature set, and using the normal data feature set to train a unigram classifier to obtain a trained unigram classifier; S4, input the test data set into the feature extractor, compose the extracted features into a fault data feature set, and use the fault data feature set to test the unary classifier. If the test result meets the preset evaluation index, the training of the feature extractor and the unary classifier is completed, otherwise, repeat steps S1-S4; S5, inputting the sensor signal on the gas turbine into a feature extractor, inputting the extracted features into a unary classifier, and obtaining a fault detection result; Constructing a normal data set consisting of normal signals is as follows: Acquire original signals of multiple sensors in the same time period T, where the types of the original signals are normal signals and fault signals; The original signal of each sensor is divided according to the preset step length to obtain N·M signal sequences. Each signal sequence is the signal of a sensor in a time interval. M=L / Δl=T / Δt, where N is the type of sensor, L is the length of the original signal, Δl is the step length, and Δt is the time interval. The signal sequences in the same time interval are combined to obtain a sample, and a total of M samples are obtained, each sample including N signal sequences; The unsupervised learning model is a recurrent neural network, which uses an unsupervised triple loss function based on time series, as follows: Where σ is the sigmoid function, x a represents the anchor point, x pos represents a positive sample, represents the kth negative sample, θ represents the model parameters, f(*,θ) represents the result of * inputting the unsupervised learning model, and K represents the number of preset negative samples.

2. A method for detecting unknown faults in a gas turbine control system according to claim 1, characterized in that: During the training process of the unsupervised learning model, the selection rules for the anchor point, positive sample, and negative sample triplets are as follows: Set the length l of anchor points, positive samples and negative samples a , l pos and l neg , and the number of negative samples K; Combine the normal signals of a sensor in all time intervals to obtain a subsequence, and obtain N subsequences in total. Each subsequence has a length of L and contains M segments of normal signals. Randomly select a subsequence and randomly select a length l in the subsequence a A signal segment is used as the anchor point x a ; At anchor point x a Randomly select a length of l pos A normal signal is taken as a positive sample x pos ; Randomly select K subsequences from the remaining N-1 subsequences, and randomly select a subsequence of length l in each selected subsequence. neg A normal signal is taken as a negative sample x neg , where any negative sample The time interval and positive sample x pos There is overlap between the time periods.

3. A method for detecting unknown faults in a gas turbine control system according to claim 2, characterized in that: The negative samples The time interval and positive sample x pos The overlapping range between the time intervals is λ1·Δl~λ2·Δl, where λ1 and λ2 are preset proportional coefficients.

4. A method for detecting unknown faults in a gas turbine control system according to claim 3, characterized in that: The length l of the anchor point, positive sample and negative sample a , l pos and l neg are all integer multiples of Δl, and the proportional coefficient is an integer; The specific anchor point selection is: a is the window width, randomly select a segment of the signal from the M segments of normal signal as the starting point, and use the window to intercept a segment with a length equal to l a The normal signal is x a ; The specific selection of positive samples is: pos is the window width, in x a A random signal is selected as the starting point, and a window is used to intercept a length equal to l pos and belongs to x a The normal signal is x pos ; The specific selection of negative samples is: neg is the window width, and selects Among them, select To: OK The corresponding subsequence is randomly selected from the M normal signals of the subsequence with a length of l neg And the time interval is x a There is a normal signal with overlap of λ1·Δl~λ2·Δl in the time interval as 5. A method for detecting unknown faults in a gas turbine control system according to claim 1, characterized in that: The unary classifier is based on a local anomaly factor, which is defined as: rd k (x,o)=max{k-distance(o),d(x,o)} Among them, LOF k (x) represents the local anomaly factor of feature x, lrd k (x) represents the local reachability density of feature x, rd k (x,o) represents the local reachable distance of feature x, k-distance(o) represents the distance between o and the farthest point among the k nearest neighbors of feature x, N k (x) represents the set of k nearest neighbors of sample x, o represents a sample point among the k nearest neighbors, and d(x,o) represents the Euclidean distance between features x and o.

6. A method for detecting unknown faults in a gas turbine control system according to claim 1, characterized in that: The specific steps for constructing the test dataset are: A normal data set is obtained, multiple samples are selected from the normal data set, multiple signal sequences are selected from each sample, fault signals are superimposed on these signal sequences, and a test data set is obtained, wherein the test data set includes multiple samples, each sample includes at least one fault, and the number of each fault is at least one.

7. A method for detecting unknown faults in a gas turbine control system according to claim 1, characterized in that: In step S4, the evaluation indicators include but are not limited to recall rate, precision rate, F score and false alarm rate.

8. A gas turbine control system unknown fault detection system, characterized in that: include: A data set construction module is used to obtain normal signals of various sensors on the gas turbine, construct a normal data set composed of normal signals, superimpose fault signals on some normal signals and construct a test data set; The first model module is used to construct an unsupervised learning model and use a normal data set to train the unsupervised learning model to obtain a trained feature extractor, wherein the unsupervised learning model is used for feature extraction; The second model module is used to construct a unary classifier and train the unary classifier using a normal data feature set to obtain a trained unary classifier, wherein the normal data feature set is composed of features extracted from a normal data set input feature extractor; A test evaluation module, used to test the unary classifier using a fault data feature set and determine whether the test result meets a preset evaluation index, wherein the fault data feature set is composed of features extracted by inputting a test data set into a feature extractor; A detection module is used to input the sensor signal on the gas turbine into a feature extractor, and input the extracted features into a unary classifier to obtain a fault detection result; Constructing a normal data set consisting of normal signals is as follows: Acquire original signals of multiple sensors in the same time period T, where the types of the original signals are normal signals and fault signals; The original signal of each sensor is divided according to the preset step length to obtain N·M signal sequences. Each signal sequence is the signal of a sensor in a time interval. M=L / Δl=T / Δt, where N is the type of sensor, L is the length of the original signal, Δl is the step length, and Δt is the time interval. The signal sequences in the same time interval are combined to obtain a sample, and a total of M samples are obtained, each sample including N signal sequences; The unsupervised learning model is a recurrent neural network, which uses an unsupervised triple loss function based on time series, as follows: Where σ is the sigmoid function, x a represents the anchor point, x pos represents a positive sample, represents the kth negative sample, θ represents the model parameters, f(*,θ) represents the result of * inputting the unsupervised learning model, and K represents the number of preset negative samples.

Images