Water turbine open-set fault diagnosis method, device, storage medium, and program product

Abstract

The present application relates to the technical field of water turbine fault diagnosis, and discloses a water turbine open-set fault diagnosis method, a device, a storage medium, and a program product. The water turbine open-set fault diagnosis method comprises: acquiring a training set; using the training set to train an initial data analysis network and an initial category analysis network to obtain a target data analysis network and a target category analysis network; inputting data under detection into a target data analysis network to obtain a target feature distribution; calculating a target distance between the target feature distribution and each preset feature distribution; and when the minimum target distance is less than or equal to a distance threshold, outputting a fault category of said data as a preset fault category corresponding to the minimum target distance, and when the minimum target distance is greater than the distance threshold, outputting the fault category of said data as an unknown fault category. The present application can solve the problem of being unable to identifying unknown faults, and improve the water turbine open set fault diagnosis accuracy.

Description

Methods, equipment, storage media and program products for diagnosing open-circuit faults in water turbines

[0001] Cross-references to related applications

[0002] This application claims priority to Chinese Patent Application No. 202410991694.3, filed on July 23, 2024, entitled "Method, Device, Storage Medium and Program Product for Diagnosing Open-Collection Faults of Water Turbines", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of turbine fault detection technology, specifically to turbine open-loop fault diagnosis methods, equipment, storage media, and program products. Background Technology

[0004] Hydropower turbine generator sets are expensive and are core and critical equipment in hydropower stations. Condition monitoring technology for hydropower turbine generator sets can monitor and analyze their operating status in real time, promptly identify potential hazards and defects, prevent destructive accidents, guide unit operation and maintenance, and ensure safe and economical operation. It is a necessary means for hydropower plants to achieve "unmanned or minimally staffed" operation.

[0005] Deep learning technology can automatically extract fault-related features from signals without manual feature engineering. However, most current deep learning-based fault diagnosis models are closed-set recognition models, meaning they classify all input data into known fault categories. When deployed in the field and encountering unknown faults, these closed-set models force the unknown faults to be classified as known faults, resulting in low flexibility and accuracy in fault identification. Summary of the Invention

[0006] In view of this, this application provides a method, device, storage medium and program product for diagnosing open-collector faults in hydro turbines, in order to solve the problem of being unable to identify unknown faults and improve the accuracy of open-collector fault diagnosis in hydro turbines.

[0007] In a first aspect, this application provides a method for diagnosing open-set faults in hydraulic turbines. This method includes: acquiring a training set, which includes historical fault data and corresponding historical fault categories of the open-set hydraulic turbine; inputting the historical fault data into an initial data analysis network to obtain the data feature distribution of the historical fault data; inputting the historical fault categories into an initial category analysis network to obtain the category feature distribution of the historical fault categories; adjusting the parameters of the initial data analysis network and the initial category analysis network based on the data feature distribution and the category feature distribution to obtain a target data analysis network and a target category analysis network; inputting the data to be detected into the target data analysis network to obtain a target feature distribution; calculating the target distance between the target feature distribution and each preset feature distribution, wherein the preset feature distribution is obtained by processing multiple preset fault categories using the target category analysis network; when the minimum target distance is less than or equal to a distance threshold, outputting the fault category of the data to be detected as the preset fault category corresponding to the minimum target distance; when the minimum target distance is greater than the distance threshold, outputting the fault category of the data to be detected as an unknown fault category.

[0008] In this implementation, a data analysis network and a category analysis network are trained using fault data and fault categories. The trained data analysis network analyzes the data to be detected, and the trained category analysis network analyzes known preset fault categories. The fault category closest to the current data to be detected is determined by the target distance between the data to be detected and the preset fault categories. This application does not directly output the closest fault category; instead, it compares the target distance with a distance threshold to distinguish between known and unknown faults. Therefore, the method of this application can identify not only known fault categories but also unknown fault categories, improving the flexibility and accuracy of fault detection.

[0009] In one optional implementation, obtaining the training set includes: obtaining multiple historical fault categories of the turbine open set; for each historical fault category, collecting multiple historical fault data; and combining each historical fault data with its corresponding historical fault category as a set of data to obtain the training set.

[0010] In one optional implementation, the initial data analysis network includes an initial shared backbone network, an initial first branch network, and an initial second branch network. Inputting historical fault data into the initial data analysis network to obtain the data feature distribution of the historical fault data includes: inputting historical fault data into the initial data analysis network, performing feature extraction using the initial shared backbone network to obtain historical fault feature data; calculating the mean of the historical fault feature data using the initial first branch network, and calculating the variance of the historical fault feature data using the initial second branch network to obtain the data feature distribution of the historical fault data.

[0011] In one optional implementation, the initial category analysis network is an initial prototype Gaussian distribution network. Inputting historical fault categories into the initial category analysis network to obtain the category feature distribution of historical fault categories includes: inputting historical fault categories into the initial prototype Gaussian distribution network, calculating the feature space Gaussian distribution of historical fault categories, and obtaining the category feature distribution of historical fault categories.

[0012] In one optional implementation, adjusting the parameters of the initial data analysis network and the initial category analysis network based on the data feature distribution and the category feature distribution to obtain the target data analysis network and the target category analysis network includes: calculating the KL distance between each data feature distribution and multiple category feature distributions; constructing a classification loss function based on the KL distance, the classification loss function including a classifier sub-loss function, an intra-class distance sub-loss function, and an inter-class distance sub-loss function, the classification loss function being: L = L c +λ1L nd +λ2L ed Where L is the classification loss function, L c For the classifier loss function, L nd L is the intra-class distance sub-loss function. ed λ1 and λ2 are the inter-class distance loss function and the sub-loss balance coefficients, respectively. Based on the classification loss function, the parameters of the initial data analysis network and the initial class analysis network are adjusted to obtain the target data analysis network and the target class analysis network.

[0013] In one optional implementation, constructing the classification loss function based on KL distance includes: applying a SoftMax activation function to the KL distance and calculating the classification loss function using cross-entropy, with the following formula: in, Let y be the KL distance between the i-th data feature distribution and the s-th category feature distribution. i For the i-th historical fault category, Let γ be the KL distance between the data feature distribution of historical fault data and the category feature distribution of the corresponding historical fault categories, and let γ be a hyperparameter.

[0014] In one optional implementation, the intra-class distance sub-loss function is used to reduce the KL distance between the data feature distribution of historical fault data and the category feature distribution of the corresponding historical fault categories. Constructing the classification loss function based on the KL distance includes: the formula for calculating the intra-class distance sub-loss function is: Among them, y i For the i-th historical fault category, The KL distance is the distribution of data features of historical fault data and the distribution of category features of the corresponding historical fault categories.

[0015] In one optional implementation, the inter-class distance ion loss function is used to expand the KL distance between the category feature distributions of each historical fault category. Constructing the classification loss function based on the KL distance includes: the formula for calculating the inter-class distance ion loss function is: Where, d s,k Let be the KL distance between the s-th category feature distribution and the k-th category feature distribution, and m be a preset hyperparameter for the minimum KL distance between category feature distributions.

[0016] In this implementation, the present application constructs a shared backbone network, two branch networks, and a prototype high-speed distribution network, and uses a classification sub-loss function, an intra-class distance sub-loss function, and an inter-class distance sub-loss function to update the model parameters. The three sub-loss functions adopt a unified distance metric, which can effectively improve the robustness of feature extraction and improve the accuracy of open set fault detection in this application.

[0017] In one alternative implementation, before calculating the target distance between the target feature distribution and each preset feature distribution, the method further includes: obtaining a cross-validation dataset; and using the cross-validation dataset to determine a distance threshold such that the accuracy of predicting the cross-validation dataset as a known fault classification reaches a boundary value of α%, where α is a hyperparameter.

[0018] In this implementation, the distance threshold is determined based on the boundary value of the known class classification accuracy. This ensures high known class classification performance before performing unknown fault identification, making full use of the known fault data on site and improving fault classification accuracy.

[0019] Secondly, this application provides a computer device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the turbine open-circuit fault diagnosis method described in the first aspect or any corresponding embodiment.

[0020] Thirdly, this application provides a computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the turbine open-circuit fault diagnosis method described in the first aspect or any corresponding embodiment.

[0021] Fourthly, this application provides a computer program product, including computer instructions, which are used to cause a computer to execute the turbine open-circuit fault diagnosis method described in the first aspect or any corresponding embodiment. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 is a flowchart of a method for diagnosing open-collection faults of a water turbine according to an embodiment of this application;

[0024] Figure 2 is a schematic diagram of a turbine open-collector fault diagnosis model according to an embodiment of this application;

[0025] Figure 3 is a flowchart of a turbine open-collection fault diagnosis training method according to an embodiment of this application;

[0026] Figure 4 is a flowchart of another method for diagnosing open-collection faults of a water turbine according to an embodiment of this application;

[0027] Figure 5 is a structural block diagram of a turbine open-collector fault diagnosis device according to an embodiment of this application;

[0028] Figure 6 is a schematic diagram of the hardware structure of a computer device according to an embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Most existing deep learning-based fault diagnosis models are closed-set recognition models, meaning they classify all input data into known fault categories. This results in their inability to diagnose unknown faults. Therefore, this application provides an open-set fault diagnosis model for hydro turbines, capable of diagnosing unknown fault categories in addition to diagnosing known fault categories. Please refer to Figure 1, which is a schematic diagram of an open-set fault diagnosis model for hydro turbines according to an embodiment of this application.

[0031] The open-set fault diagnosis model for water turbines includes a data analysis network and a category analysis network.

[0032] The data analysis network is used to calculate the characteristic distribution of fault data. The data analysis network includes a shared backbone network, a first branch network, and a second branch network.

[0033] The shared backbone network is used to extract features from fault data, and the output data serves as the input data for the first branch network and the second branch network, respectively.

[0034] The first and second branch networks are parallel networks. The first branch network is used to calculate the mean information of the fault data, and the second branch network is used to calculate the variance information of the fault data.

[0035] In one implementation, the shared backbone network, the first branch network, and the second branch network are all deep neural networks. The shared backbone network contains parameters θ, and the first branch network contains parameters... The second branch network contains parameter φ.

[0036] Among them, the category analysis network is used to calculate the feature distribution of fault categories.

[0037] In one implementation, the category analysis network is a deep neural network. The category analysis network is a parameterized network. Prototype Gaussian distribution network The prototype Gaussian distribution network is used to compute the Gaussian distribution of the feature space for fault categories. Here, I is the identity matrix.

[0038] In one implementation, deep neural networks include, but are not limited to, convolutional neural networks, fully connected networks, Transformer networks, and recurrent neural networks.

[0039] In one implementation, the output dimension of the first branch network and the output dimension μ of the second branch network are... k Both the dimension of and the dimension of I are J.

[0040] Based on this, this embodiment provides a method for diagnosing open-collection faults in a water turbine. According to the embodiments of this application, an embodiment of a method for diagnosing open-collection faults in a water turbine is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than that shown here.

[0041] Figure 2 is a flowchart of a turbine open-circuit fault diagnosis method according to an embodiment of this application. It should be noted that if substantially the same result is obtained, this embodiment is not limited to the flow order shown in Figure 2. As shown in Figure 2, the process includes the following steps:

[0042] Step S201: Obtain the training set.

[0043] The training set includes historical fault data of the turbine open set and the corresponding historical fault categories.

[0044] Specifically, information related to the operating status of the turbine generator set under different fault categories is collected as historical fault data. Each historical fault data point and its corresponding historical fault category are treated as a set of data, and multiple sets of data are used to construct a training set.

[0045] Step S202: Input historical fault data into the initial data analysis network to obtain the data feature distribution of historical fault data; input historical fault categories into the initial category analysis network to obtain the category feature distribution of historical fault categories.

[0046] The training set is output into the initial open-set fault diagnosis model of the turbine. The initial data analysis network is used to analyze the historical fault data and obtain the data feature distribution of the historical fault data.

[0047] Specifically, feature extraction is performed using the initial shared backbone network to obtain historical fault feature data. The mean of the historical fault feature data is calculated using the initial first branch network, and the variance of the historical fault feature data is calculated using the initial second branch network. The data feature distribution of the historical fault data is obtained by combining the mean and variance.

[0048] The initial category analysis network is used to analyze historical fault categories, and the category feature distribution of historical fault categories is obtained.

[0049] Specifically, the Gaussian distribution of the feature space of historical fault categories is calculated using the initial prototype Gaussian distribution network to obtain the category feature distribution of historical fault categories.

[0050] Step S203: Adjust the parameters of the initial data analysis network and the initial category analysis network based on the data feature distribution and the category feature distribution to obtain the target data analysis network and the target category analysis network.

[0051] The similarity is determined by comparing each data feature distribution with multiple category feature distributions. The parameters of the initial data analysis network and the initial category analysis network are adjusted to make the data feature distribution of historical fault data less different from the category feature distribution of the corresponding historical fault category, and more different from the category feature distribution of other historical fault categories. After multiple adjustments, the corresponding target data analysis network and target category analysis network are obtained.

[0052] In one implementation, a loss function is constructed using the data feature distribution and the category feature distribution, and the parameters of the initial data analysis network and the initial category analysis network are adjusted in reverse using the loss function.

[0053] In one implementation, the KL distance between each data feature distribution and multiple class feature distributions is calculated, and a classification loss function is constructed based on the KL (Kullback-Leibler) distance.

[0054] Step S204: Input the data to be detected into the target data analysis network to obtain the target feature distribution.

[0055] The data to be tested is the current operating parameter information of the water turbine.

[0056] The Gaussian distribution of the data to be detected is calculated using the trained target data analysis network, thus obtaining the target feature distribution.

[0057] Step S205: Calculate the target distance between the target feature distribution and each preset feature distribution.

[0058] The preset feature distribution is obtained by processing multiple preset fault categories using a target category analysis network. The preset fault categories are known fault categories that have undergone training and testing. Among them, the preset fault category includes no fault.

[0059] In one implementation, the KL distance between the target feature distribution and each preset feature distribution is calculated.

[0060] Step S206: When the minimum target distance is less than or equal to the distance threshold, the fault category of the data to be detected is the preset fault category corresponding to the minimum target distance. When the minimum target distance is greater than the distance threshold, the fault category of the data to be detected is the unknown fault category.

[0061] The distances to multiple targets are sorted, and the preset feature distribution corresponding to the minimum target distance is obtained. The preset fault category corresponding to the preset feature distribution is then obtained, and the preset fault category is considered to be the most likely fault category among the multiple preset fault categories.

[0062] The minimum target distance is compared with a distance threshold. If the minimum target distance is less than or equal to the distance threshold, it indicates that the target feature distribution and the preset feature distribution have a high similarity, and the fault category of the data to be detected is the preset fault category corresponding to the minimum target distance. If the minimum target distance is greater than the distance threshold, it indicates that the target feature distribution and the preset feature distribution have a low similarity, and the current fault category does not exist among the currently known preset fault categories, and the fault category of the data to be detected is the unknown fault category.

[0063] The turbine open-set fault diagnosis method provided in this embodiment trains a data analysis network and a category analysis network using fault data and fault categories. The trained data analysis network analyzes the data to be detected, and the trained category analysis network analyzes known preset fault categories. The method determines the fault category closest to the current data by using the target distance between the data to be detected and the preset fault categories. This application does not directly output the closest fault category; instead, it compares the target distance with a distance threshold to distinguish between known and unknown faults. Therefore, the method of this application can identify not only known fault categories but also unknown fault categories, improving the flexibility and accuracy of fault detection.

[0064] Based on this, this embodiment provides a method for training a turbine open-set fault diagnosis model. Figure 3 is a flowchart of a turbine open-set fault diagnosis training method according to an embodiment of this application. It should be noted that if substantially the same result is obtained, this embodiment is not limited to the flow order shown in Figure 3. As shown in Figure 3, the process includes the following steps:

[0065] Step S301: Obtain the training set.

[0066] Specifically, step S301 includes:

[0067] Step S3011: Collect relevant signals of the operating status of the hydro-generator under different fault categories to obtain historical fault data.

[0068] Specifically, M operational status-related signals of the hydro-generator are collected within L time points, and the m-th operational status-related signal is denoted as a. m , where a m Let L be a row vector of length L.

[0069] In one implementation, the signals related to the operating status of the hydro-generator include, but are not limited to, the X-axis vibration of the upper frame, the Y-axis vibration of the upper frame, the Z-axis vibration of the upper frame, the X-axis vibration of the lower frame, the Y-axis vibration of the lower frame, the Z-axis vibration of the lower frame, the X-axis vibration of the top cover, the Y-axis vibration of the top cover, the Z-axis vibration of the top cover, the X-axis vibration of the stator frame, the Y-axis vibration of the stator frame, the Z-axis vibration of the stator frame, the vibration of the stator core, the runout of the upper and lower guide bearings, and the runout of the water guide bearing.

[0070] For each row vector a m Normalization is performed to obtain the normalized value a′. m By stacking the normalized values ​​of M operation status-related signals, a historical fault data set is obtained. in, The dimension is M×L.

[0071] In one implementation, the normalization calculation methods include maximum normalization, minimum normalization, and Z-score normalization.

[0072] Step S3012: Based on the multiple historical fault categories of the turbine open set, collect multiple historical fault data for each historical fault category.

[0073] Specifically, for K different historical fault categories, the method in step S3021 is used to repeatedly collect data from the hydro-generator N times to obtain N historical fault data.

[0074] Step S3013: Combine each historical fault data point with its corresponding historical fault category as a set of data to obtain a training set.

[0075] Each historical fault data point and its corresponding historical fault category are grouped into NK sets of data, resulting in the training set. Where x i For the i-th historical fault data, y i Historical fault data x i The corresponding historical fault category, y i ∈{1,…,K}.

[0076] Understandably, N historical fault data x obtained under the same historical fault category correspond to the same historical fault category y.

[0077] Step S302: Input the training set into the initial open set fault diagnosis model of the water turbine to obtain the data feature distribution and category feature distribution.

[0078] Historical fault data from the training set is used as training data input into the data analysis network. The historical fault data is mapped to the feature space, and the resulting data feature distribution is a Gaussian distribution.

[0079] Specifically, for any historical fault data x i The mean of the Gaussian distribution in the corresponding feature space is calculated using the initial first branch network. i =E1(E(x) i The variance σ of the corresponding Gaussian distribution in the feature space is calculated using the initial second branch network. i =E2(E(x) i The resulting data feature distribution is as follows: For NK historical fault data, NK data feature distributions are obtained.

[0080] The historical fault categories in the training set are used as training labels and input into the category analysis network. The historical fault categories are mapped to the feature space, and the resulting category feature distribution is a Gaussian distribution.

[0081] Specifically, for any historical fault category, the Gaussian distribution in the feature space is calculated using the initial prototype Gaussian distribution network to obtain the category feature distribution of the historical fault category. For K historical fault categories, K category feature distributions are obtained.

[0082] Step S303: Calculate the KL distance between each data feature distribution and multiple category feature distributions.

[0083] For any historical fault data x i Calculate the KL distance from the data feature distribution to the K category feature distributions, where the calculation formula is:

[0084] in, Let *| be the KL distance from the i-th data feature distribution to the k-th class feature distribution. j This represents the j-th dimension of *.

[0085] Step S304: Construct a classification loss function based on KL distance.

[0086] The classification loss function includes the classifier sub-loss function, the intra-class distance sub-loss function, and the inter-class distance sub-loss function.

[0087] A sub-loss function for classification is constructed based on KL distance, which is used to improve classification accuracy. Specifically, the KL distance... Use the SoftMax activation function to obtain historical fault data x i The probability of belonging to the k-th historical fault category is:

[0088] The classification loss function is calculated using cross-entropy:

[0089] Here, γ is a hyperparameter used to control the difficulty of assigning probability values.

[0090] An intra-class distance sub-loss function is constructed based on KL distance. This sub-loss function is used to reduce the KL distance between the data feature distribution of historical fault data and the corresponding category feature distribution of historical fault categories. Specifically, the formula for calculating the intra-class distance sub-loss function is as follows:

[0091] An inter-class distance loss function is constructed based on the KL distance, which is used to expand the KL distance between the class feature distributions of each historical fault category. Specifically, the KL distance between the class feature distribution of the s-th historical fault category and the class feature distribution of the k-th historical fault category is:

[0092] The formula for calculating the inter-class ion loss function is:

[0093] Where m is a hyperparameter representing the minimum KL distance between preset category feature distributions.

[0094] The classification loss function is calculated using the classifier loss function, the intra-class distance loss function, and the inter-class distance loss function:

[0095] L = L c +λ1L nd +λ2L ed .

[0096] Where λ1 and λ2 are the sub-loss balance coefficients.

[0097] This application constructs a shared backbone network, two branch networks, and a prototype high-speed distribution network, and uses a classification sub-loss function, an intra-class distance sub-loss function, and an inter-class distance sub-loss function to update model parameters. The three sub-loss functions adopt a unified distance metric, which can effectively improve the robustness of feature extraction and improve the accuracy of open set fault detection in this application.

[0098] Step S305: Adjust the parameters of the initial data analysis network and the initial category analysis network based on the classification loss function to obtain the target turbine open set fault diagnosis model.

[0099] The model parameters of the initial data analysis network and the initial category analysis network are updated using the classification loss function.

[0100] Specifically, update the parameters of the initial shared backbone network. Obtain the target shared backbone network; update the parameters of the initial first branch network. Obtain the target first branch network; update the parameters of the initial second branch network. Obtain the target second branch network; update the parameters of the initial prototype Gaussian distribution network. The target prototype Gaussian distribution network is obtained. Here, ν is the learning rate.

[0101] Understandably, after training the open set fault diagnosis model of the target turbine, the open set fault diagnosis model of the target turbine is tested using a test set.

[0102] The turbine open-set fault diagnosis model training method provided in this embodiment trains the data analysis network and the category analysis network by utilizing the distance relationship between fault data and fault categories. It then uses a loss function based on three unified distance metrics to update model parameters, thereby improving the robustness of feature extraction and the accuracy of the trained model.

[0103] Based on this, this application deploys the trained target turbine open-set fault diagnosis model to the field for testing. Therefore, this embodiment provides a turbine open-set fault diagnosis method, which can be used for the aforementioned trained target turbine open-set fault diagnosis model. Figure 4 is a flowchart of another turbine open-set fault diagnosis method according to an embodiment of this application. It should be noted that if substantially the same results are obtained, this embodiment is not limited to the flow order shown in Figure 4. As shown in Figure 4, the process includes the following steps:

[0104] Step S401: Input the data to be detected into the target data analysis network to obtain the target feature distribution.

[0105] x t In the target data analysis network, the mean of the Gaussian distribution in the feature space is calculated using the first branch network of the target network, which is μ. te =E1(E(x) t The variance σ of the Gaussian distribution in the feature space is calculated using the second branch network of the target network. t =E2(E(x) t If the target feature distribution is...

[0106] Step S402: Calculate the KL distance between the target feature distribution and each preset feature distribution.

[0107] The KL distance between the target feature distribution and the k-th preset feature distribution is:

[0108] Step S403: Obtain the distance threshold.

[0109] Obtain the cross-validation dataset and use it to determine the distance threshold so that the accuracy of predicting the cross-validation dataset as a known fault classification reaches the boundary value of α%, where α is a hyperparameter.

[0110] This application determines a distance threshold based on the boundary value of known class classification accuracy, which can guarantee high known class classification performance before performing unknown fault identification, making full use of known fault data on site and improving fault classification accuracy.

[0111] Step S404: Determine whether the minimum KL distance is less than or equal to the distance threshold. When the minimum KL distance is less than or equal to the distance threshold, output the fault category of the data to be detected as the preset fault corresponding to the minimum KL distance. When the minimum KL distance is greater than the distance threshold, output the fault category of the data to be detected as the unknown fault category.

[0112] The minimum KL distance is determined as the classification result of the target turbine open-set fault diagnosis model, where the classification result is:

[0113] Minimum KL distance The comparison is performed with the distance threshold ε. When the fault category of the output data to be detected is an unknown fault category, when When the fault category of the data to be detected is output, it is the preset fault category corresponding to the minimum KL distance.

[0114] Understandably, the default fault categories are known fault categories, including no faults.

[0115] Optionally, when the target turbine open-collection fault diagnosis model determines that the turbine generator is fault-free, no maintenance strategy is executed; when the target turbine open-collection fault diagnosis model determines that the turbine generator has a known fault, the corresponding maintenance strategy is executed; when the target turbine open-collection fault diagnosis model determines that the turbine generator has an unknown fault, the subsequent strategy should be determined based on other data.

[0116] The turbine open-set fault diagnosis method provided in this embodiment analyzes the data to be detected using a trained data analysis network and a trained category analysis network to analyze known preset fault categories. The method determines the fault category closest to the current data by using the target distance between the data to be detected and the preset fault categories. Instead of directly outputting the closest fault category, this application compares the target distance with a distance threshold to distinguish between known and unknown faults. The distance threshold is pre-determined based on the boundary value of the classification accuracy of known categories. Therefore, this method can identify unknown fault categories with high accuracy, improving the flexibility and accuracy of fault detection.

[0117] This embodiment also provides a device for diagnosing open-collector faults in hydraulic turbines. This device is used to implement the above embodiments and optional implementations, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0118] This embodiment provides a turbine open-collector fault diagnosis device. Figure 5 is a structural block diagram of the turbine open-collector fault diagnosis device according to an embodiment of this application. As shown in Figure 5, it includes:

[0119] The acquisition module 501 is used to acquire the training set, which includes historical fault data of the turbine and the corresponding historical fault categories.

[0120] The input module 502 is used to input historical fault data into the initial data analysis network to obtain the data feature distribution of the historical fault data; and to input historical fault categories into the initial category analysis network to obtain the category feature distribution of the historical fault categories.

[0121] Training module 503 is used to adjust the parameters of the initial data analysis network and the initial category analysis network based on the data feature distribution and the category feature distribution to obtain the target data analysis network and the target category analysis network.

[0122] The detection module 504 is used to input the data to be detected into the target data analysis network to obtain the target feature distribution.

[0123] The calculation module 505 is used to calculate the target distance between the target feature distribution and each preset feature distribution. The preset feature distribution is obtained by processing multiple preset fault categories using a target category analysis network.

[0124] The classification module 506 is used to output the fault category of the data to be detected as the preset fault category corresponding to the minimum target distance when the minimum target distance is less than or equal to the distance threshold, and to output the fault category of the data to be detected as the unknown fault category when the minimum target distance is greater than the distance threshold.

[0125] Further functional descriptions of the above modules are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0126] In this embodiment, the turbine open-collection fault diagnosis device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0127] This application embodiment also provides a computer device having the turbine open-circuit fault diagnosis device shown in FIG5 above.

[0128] Please refer to Figure 6, which is a schematic diagram of the structure of a computer device provided in an optional embodiment of this application. As shown in Figure 6, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components communicate with each other using different buses and can be installed on a common motherboard or otherwise as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some optional embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 6 uses one processor 10 as an example.

[0129] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0130] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0131] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0132] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0133] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 20 can be connected via a bus or other means; Figure 6 shows an example of a connection via a bus.

[0134] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0135] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; optionally, the storage medium may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0136] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0137] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for diagnosing an open set fault of a hydraulic turbine, characterized by, The method comprises: obtaining a training set comprising historical fault data and corresponding historical fault categories of a water turbine open set; inputting the historical fault data into an initial data analysis network to obtain a data feature distribution of the historical fault data; inputting the historical fault categories into an initial category analysis network to obtain a category feature distribution of the historical fault categories; adjusting parameters of the initial data analysis network and the initial category analysis network based on the data feature distribution and the category feature distribution to obtain a target data analysis network and a target category analysis network; inputting to-be-detected data into the target data analysis network to obtain a target feature distribution; calculating a target distance between the target feature distribution and each preset feature distribution, the preset feature distribution being obtained by processing a plurality of preset fault categories by using the target category analysis network; when the smallest target distance is less than or equal to a distance threshold, outputting a fault category of the to-be-detected data as a preset fault category corresponding to the smallest target distance, and when the smallest target distance is greater than the distance threshold, outputting a fault category of the to-be-detected data as an unknown fault category.

2. The method of claim 1, wherein, The obtaining of the training set comprises: obtaining a plurality of historical fault categories of a water turbine open set; for each historical fault category, collecting a plurality of historical fault data; taking each historical fault data and the corresponding historical fault category as a group of data to obtain the training set.

3. The method of claim 2, wherein, The initial data analysis network comprises an initial shared backbone network, an initial first branch network and an initial second branch network, and the inputting of the historical fault data into the initial data analysis network to obtain the data feature distribution of the historical fault data comprises: inputting the historical fault data into the initial data analysis network, performing feature extraction by using the initial shared backbone network to obtain historical fault feature data; calculating the mean of the historical fault feature data by using the initial first branch network and calculating the variance of the historical fault feature data by using the initial second branch network to obtain the data feature distribution of the historical fault data.

4. The method of claim 3, wherein, The initial category analysis network is an initial prototype Gaussian distribution network, and the inputting of the historical fault categories into the initial category analysis network to obtain the category feature distribution of the historical fault categories comprises: inputting the historical fault categories into the initial prototype Gaussian distribution network, calculating the feature space Gaussian distribution of the historical fault categories to obtain the category feature distribution of the historical fault categories.

5. The method of claim 4, wherein, The adjusting of the parameters of the initial data analysis network and the initial category analysis network based on the data feature distribution and the category feature distribution to obtain the target data analysis network and the target category analysis network comprises: calculating a KL distance between each data feature distribution and a plurality of category feature distributions; constructing a classification loss function based on the KL distance, the classification loss function comprising a classification sub-loss function, an intra-class distance sub-loss function and an inter-class distance sub-loss function, and the classification loss function being: L = L c + λ1L nd + λ2L ed ; Wherein, the L is the classification loss function, the L c is the classification sub-loss function, the L nd is the intra-class distance sub-loss function, the L ed is the inter-class distance sub-loss function, and the λ1 and the λ2 are sub-loss balancing coefficients. Adjust parameters of the initial data analysis network and the initial category analysis network based on the classification loss function to obtain the target data analysis network and the target category analysis network.

6. The method of claim 5, wherein, The constructing the classification loss function based on the KL distance comprises: The SoftMax activation function is used for the KL distance, and the classification loss function is calculated by using cross-entropy, and the calculation formula is as follows: Among them, the is the KL distance between the ith data feature distribution and the s-th class feature distribution, where y i is the ith historical failure class, where The KL distance between the data feature distribution of the historical failure data and the category feature distribution of the corresponding historical failure category, and the γ is a hyperparameter.

7. The method of claim 5, wherein, The intra-class distance sub-loss function is used to reduce the KL distance between the data feature distribution of the historical failure data and the category feature distribution of the corresponding historical failure category, and the constructing the classification loss function based on the KL distance comprises: The formula for calculating the intra-class distance sub-loss function is: wherein the y i is the i-th historical failure category, the The KL distance between the data feature distribution of the historical failure data and the category feature distribution of the corresponding historical failure category.

8. The method of claim 5, wherein, The inter-class distance sub-loss function is used to expand the KL distance between the category feature distributions of each historical failure category, and the constructing the classification loss function based on the KL distance comprises: The formula for calculating the inter-class distance sub-loss function is: wherein the d s,k is the KL distance between the s-th category feature distribution and the k-th category feature distribution, and m is a preset hyperparameter of the minimum KL distance between the category feature distributions.

9. The method of claim 5, wherein, Before the calculating the target distance between the target feature distribution and each preset feature distribution further comprises: Obtaining a cross-validation data set; Determining the distance threshold value by using the cross-validation data set, so that the accuracy rate of predicting the cross-validation data set as a known failure category reaches an α% boundary value, wherein the α is a hyperparameter.

10. A computer device, comprising: Comprise: A memory and a processor, which are communicatively connected between each other, and the memory stores computer instructions, and the processor executes the computer instructions to perform the water turbine open set fault diagnosis method in any one of claims 1 to 9.

11. A computer readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing a computer to execute the water turbine open set fault diagnosis method in any one of claims 1 to 9.

12. A computer program product, characterised in that, Comprise computer instructions for causing a computer to execute the water turbine open set fault diagnosis method in any one of claims 1 to 9.

Images