A gas turbine tie rod rotor test system and its open set fault diagnosis method

Abstract

The application discloses a kind of gas turbine pull rod rotor test system and its open set fault diagnosis method, the test system includes drive control system, pull rod rotor main system, support system and measurement system.System runs, drive control system is used to drive pull rod rotor main system and control rotation speed and the running safety of entire system, measurement system is used to measure and resolve the vibration displacement signal of pull rod rotor main system.The application is aimed at gas turbine pull rod rotor, by designing and building test system, accurate measurement of operating vibration signal is realized, a large number of test data of multiple fault types and multiple operating conditions are obtained, a method based on shared maximum mean difference is proposed, combined with convolutional neural network, open set fault diagnosis with high precision can be realized, which lays a foundation for fault diagnosis of gas turbine pull rod rotor system, and has important engineering significance and broad application prospect.

Description

Technical Field

[0001] This invention belongs to the field of gas turbines, and specifically relates to a gas turbine tie rod rotor testing system and its open set fault diagnosis method. Background Technology

[0002] Gas turbines are crucial energy conversion devices widely used in aviation, power generation, and chemical industries. The tie-rod rotor is a key component of a gas turbine, and its failure can severely impact the turbine's safe operation and performance stability. Traditional tie-rod rotor fault diagnosis methods rely on the experience and knowledge of specialized technicians, making rapid and accurate diagnosis difficult. In recent years, deep learning-based intelligent fault diagnosis has achieved success in many fields. However, most current deep learning methods are primarily applied to fault diagnosis of equipment such as bearings and gearboxes, with limited research on rotor fault diagnosis, and even less specifically on tie-rod rotors.

[0003] Deep learning-based intelligent fault diagnosis methods can utilize vibration signals to detect faults in gas turbine tie-rod rotors. Leveraging deep learning's automatic feature extraction capabilities, raw vibration signal data can be transformed into more representative feature vectors, thereby improving diagnostic accuracy. Furthermore, deep learning can learn and discover patterns from massive amounts of data, making it more usable for diagnosing complex faults. However, for gas turbines, fault signal datasets for tie-rod rotors are currently very scarce, and no publicly available datasets exist. Moreover, unlike integral rotors, tie-rod rotors, due to their combined tie-rod and multi-stage wheel structure, may experience unique faults such as tie-rod loosening and wheel parallelism deviations, in addition to traditional faults like imbalance, rubbing, and misalignment. These faults are easily affected by background noise, posing a challenge to the design of deep learning algorithms. Furthermore, in real-world industrial scenarios, fault diagnosis models may encounter problems such as input data having similar but different distributions of the same fault compared to the training dataset, or input data containing unknown fault categories, all of which can lead to model failure. Summary of the Invention

[0004] The purpose of this invention is to provide a gas turbine tie rod rotor testing system and its open set fault diagnosis method. Targeting the gas turbine tie rod rotor, the invention designs and builds a testing system to accurately measure operational vibration signals, obtaining a large amount of test data for various fault types and operating conditions. It proposes a method based on shared maximum mean difference, which, combined with a convolutional neural network, can achieve high-precision open set fault diagnosis, possessing significant engineering value and broad application prospects.

[0005] This invention is achieved through the following technical solution:

[0006] A gas turbine tie rod rotor testing system includes a drive control system, a tie rod rotor main body system, a support system, and a measurement system;

[0007] The drive control system is used to drive the tie rod rotor main system and adjust the speed and control the operational safety of the entire system.

[0008] The main body system of the tie rod rotor is used to set general rotor faults and tie rod rotor-specific faults;

[0009] Support system, used to install and support the drive control system motor, the tie rod rotor body system, and the measurement system sensors;

[0010] The measurement system is used to measure and analyze the vibration displacement signals of the tie rod rotor main body system.

[0011] A further improvement of the present invention is that the drive control system includes a motor and a speed regulator; the tie rod rotor main body system includes a tie rod rotor, bearings and couplings; the support system includes a T-slot test platform base, a motor base and a bearing base; and the measurement system includes an eddy current displacement sensor, a laser velocimeter and a data acquisition system.

[0012] During operation, the motor drives the rod rotor to rotate, and the resulting vibration displacement signal is measured by the eddy current displacement sensor and transmitted to the data acquisition system for analysis, display and storage.

[0013] A DC shunt motor drive scheme is adopted, in which the motor shaft directly drives the tie rod rotor via a coupling, and the speed is directly measured and obtained by a laser tachometer.

[0014] The motor is fixed to the base of the T-slot test platform by a motor mount to ensure alignment with the main shaft of the tie rod rotor, and is connected to the tie rod rotor by a coupling.

[0015] The tie rod rotor is mounted on bearing housings via bearings at both ends, and the bearing housings are fixed to the base of the T-slot test platform;

[0016] The rotor discs of the tie rod are divided into test discs and main discs. The structure is test disc-main disc-test disc. The discs are pre-tightened by a tie rod-nut structure and the torque is transmitted by the meshing of the end face teeth to ensure reliable torque transmission.

[0017] Eddy current displacement sensors are respectively arranged at the drive end of the tie rod rotor near the motor and the non-drive end far from the motor, and are used to measure and acquire the vibration displacement signal of the tie rod rotor.

[0018] When the distance between the eddy current sensor and the rotating shaft changes, the dynamic voltage signal generated by the sensor is input to the data acquisition system. The data acquisition system analyzes and obtains the dynamic displacement signal, and displays and stores it in real time.

[0019] A further improvement of the present invention is that the pull rod rotor includes an end face gear disk, a pull rod, and a nut, with the end face gear disk mounted on the pull rod and tightened by the nut.

[0020] A further improvement of the present invention is that the data acquisition system includes a dynamic data acquisition instrument and a data processing computer;

[0021] A dynamic data acquisition instrument is used for high-speed, multi-channel simultaneous acquisition of dynamic voltage signals from sensors and converts the voltage signals into digital signals.

[0022] A data processing computer is used for real-time display, processing, and storage of signals, performing fast Fourier transforms to obtain time-domain and frequency-domain displays of dynamic signals.

[0023] A further improvement of the present invention is that the tested tie rod rotor main body system includes a total of 8 types of faults, namely imbalance fault, misalignment fault, rubbing fault, bearing housing loose fault, tie rod loose fault, and 3 types of abnormal end face tooth contact faults.

[0024] Three types of abnormal end face tooth contact faults are used to simulate poor contact faults that occur after the end face teeth of the tie rod rotor wear.

[0025] Among them, loose tie rods will cause insufficient preload of the wheel disc; abnormal end face tooth contact in type 1 will cause initial bending of the rotor; abnormal end face tooth contact in type 2 will cause twisting of the rotor; and abnormal end face tooth contact in type 3 will cause incomplete meshing of the end face teeth.

[0026] A method for open set fault diagnosis of a gas turbine tie rod rotor, the method being based on a gas turbine tie rod rotor testing system, includes the following steps:

[0027] Step 1: Acquire vibration displacement signals of a healthy rotor under multiple operating conditions;

[0028] Step 2: Collect vibration displacement signals of rotors with different faults under corresponding operating conditions;

[0029] Step 3: Divide the vibration displacement signals from Step 1 and Step 2 into source and target domains with different working conditions and fault types, and set up an open set migration fault diagnosis task.

[0030] Step 4: Construct a fault feature extraction network based on a convolutional neural network to learn the vibration displacement signals from Steps 1 and 2, thereby obtaining the feature extractor. ;

[0031] Step 5: Construct a fault classifier based on a fully connected neural network and input it into the feature extractor. The extracted features are used to calculate the fault label probability distribution of the signal, thus obtaining the classifier. ;

[0032] Step 6: Construct an open set migration fault diagnosis method based on domain adaptation, using Shared Maximum Mean Difference (SMMD) to measure the difference between the target domain and the source domain in the open set, and learn a feature extractor. Parameters; design a domain discriminator, domain discriminator This is used to quantify the similarity between features and the source or target domain, thereby obtaining the weights for samples belonging to a shared domain; this is the domain discriminator. Introducing a global feature learning approach, and using domain adversarial training to improve the feature extractor Learn the characteristics of the source and target domains;

[0033] Step 7: Construct an open set fault classifier based on extreme value theory, calculate the features of the source domain and the target domain respectively, and obtain the extreme value learning machine EVM;

[0034] Step 8, Training the feature extractor Classifier Domain Discriminator and An open-set transfer learning fault diagnosis model consisting of an extreme value learning machine (EVM) and a fully trained feature extractor is used. Classifier And the Extreme Value Learning Machine (EVM) is used to perform open set migration fault diagnosis tasks for target domain data.

[0035] A further improvement of the present invention is that, in step 1, the rotational speed is adjusted by adjusting the speed regulator in the drive system and observing the laser velocimeter in the measurement system, thereby achieving testing under multiple working conditions.

[0036] A further improvement of the present invention is that, in step 2, a fault is set in the healthy rotor system, and signals are collected under different working conditions for different fault degrees and different fault categories of the faulty rotor. The vibration displacement signals collected in steps 1 and 2 together constitute the tie rod rotor fault dataset.

[0037] The fault severity is divided into two categories: minor and severe. The fault types include imbalance, misalignment, rubbing, loose bearing housing, loose tie rod, and three types of abnormal end face tooth contact, totaling eight types of faults. Together with the healthy rotor, there are a total of nine categories. Different operating conditions refer to different operating speeds, with three speeds set from low to high: 100 rpm, 1500 rpm, and 2000 rpm.

[0038] A further improvement of this invention is that, in step 3, the vibration displacement signals from steps 1 and 2 are divided into a source domain and a target domain with different operating conditions and fault types, and an open set migration fault diagnosis task is set up, specifically including the following steps:

[0039] 1) Data preprocessing

[0040] Deep learning methods use a large amount of data to train the model. To obtain a sufficient number of training samples, data augmentation techniques such as overlapping sampling are used to expand long-term samples. A fixed-length window slides along the temporal direction of the signal, and a segment of the signal contained within the window is extracted as a sample. There is overlap between adjacent sample segments. Along the temporal direction, the first T segments are selected. r Each segment is used as the training sample set. The remaining T e One segment was used as the test sample set. At the same time, historical information is used to diagnose potential future failures.

[0041] The data is normalized using the following formula:

[0042]

[0043] in This represents the minimum value in the dataset. This represents the maximum value in the dataset. This represents the normalized dataset. Since all data has undergone normalization, therefore... and Mixed use;

[0044] 2) Divide the source domain and the target domain

[0045] Different fault category signals are randomly selected from the tie rod rotor fault dataset to form source and target domain data, creating subsets for setting up open-set transfer fault diagnosis tasks at different speeds. Based on the above method of dividing the training and test sample sets, the data in the source domain includes the source domain training set. Source domain test set The data in the target domain includes the target domain training set. and target domain test set Both the source domain data and the target domain test set have classification labels. , and The target domain training set has no labels.

[0046] A further improvement of this invention lies in the following: In steps 4 to 7, a domain adaptation method oriented towards open set transfer is adopted, specifically a method based on the shared maximum mean difference. This yields a diagnostic model framework based on the shared maximum mean difference, where the extracted features can express the common and private categories of the source and target domains, thereby identifying the private categories of the target domain. Specifically, this includes the following steps:

[0047] 1) Construct an open set migration fault diagnosis model based on the shared maximum mean difference.

[0048] First, a fault feature extraction network based on a convolutional neural network is built as the feature extractor. A convolutional neural network is used to automatically extract fault features, with the rotor vibration displacement signal as the network input data. The network consists of one convolutional layer, four residual blocks, and one average pooling layer. Each residual block includes three convolutional layers, three batch normalization layers, and two ReLU activation functions, representing a skip-layer connection form of the ResNet network. The output is the extracted signal features. The network mapping relationship is as follows:

[0049]

[0050] in and These are rotor vibration displacement signal samples from the source domain and the target domain, respectively. The convolutional neural network mapping for the feature extractor. and Signal features extracted from samples in the source and target domains;

[0051] The high-level features obtained from the upper convolutional layers are flattened to normalize their dimensions before being fed into the classifier. Classifier It consists of one fully connected layer and a SoftMax function. The SoftMax function calculates the probability distribution of the fault category labels of the input data to obtain the source class. pseudo-distribution on :

[0052]

[0053] By minimizing the classification loss function study and The model parameters enable the feature extractor Learn the features of the source domain and train a system that can accurately classify the source domain data. :

[0054]

[0055] in The number of source domain samples used in training. The number of source domain categories, The true label of the source domain;

[0056] To learn the shared features between the source and target domains, a Shared Maximum Mean Difference (SMMD) metric is proposed. This metric measures the difference between the target and source domains on an open set and learns a feature extractor by minimizing the loss function. The parameters; SMMD aims to adapt common features between the source and target domains at the sample level; therefore, SMMD is defined as follows:

[0057]

[0058] in The number of samples in the source domain. The number of samples in the target domain. Represents the source domain. Indicates the target domain; and The weights are the values ​​of the samples from the source domain and the target domain that belong to the common domain. The larger the weight, the more likely the sample is to belong to the common domain. If the weights are all equal to 1, the above formula will degenerate into the traditional maximum mean difference (MMD).

[0059] To obtain the weights of the shared domain, a domain discriminator was designed. Used to quantify the similarity between feature z and the source or target domain:

[0060]

[0061] in and Used to measure the degree of similarity between the source domain and the target domain; if the source domain has the first... i one sample A value close to 1 indicates that the sample is highly likely to have come from [a specific source]. Conversely, if A value close to 0 indicates that the feature of the sample is highly likely to be similar to a feature in the target domain; similarly, if the first feature in the target domain is similar to a feature in the target domain, then the sample's feature is likely to be similar to a feature in the target domain. j one sample A value close to 0 indicates that the sample is highly likely to have come from [a specific source]. Conversely, if A value close to 1 indicates that the sample's features are highly likely to be similar to a feature in the source domain; the weights are calculated as follows:

[0062]

[0063] By minimizing the loss Learning Domain Discriminator The parameters enable it to distinguish between data from the source and target domains:

[0064]

[0065] The above weight calculation method shows that when When the weight decreases, If the value is close to 1, it indicates that the sample in the source domain should receive more attention in the transfer calculation; similarly, when When the weight increases, A value close to 1 indicates that the sample in the target domain receives more attention in the transfer calculation;

[0066] Introducing a global feature learning approach, and using domain adversarial training to improve the feature extractor To learn the characteristics of the source and target domains, a domain discriminator is designed. By minimizing loss Learning Domain Discriminator Parameters and update feature extractor The parameters enable it to learn features of both the source and target domains:

[0067]

[0068]

[0069] in and Domain discriminator The output;

[0070] In summary, the final optimized model of the diagnostic model based on the shared maximum mean difference is:

[0071]

[0072] 2) Constructing an extreme value learning machine based on extreme value theory

[0073] An extreme value learning machine (EVM) based on extreme value theory is used to predict unknown categories, thereby improving the ability of deep learning to detect open set samples.

[0074] If feature extractor If well-trained, the features of fault categories in the source and target domains will be clearly separable in the hyperdimensional space, and the features of each category in the source domain will be... Calculate its average characteristic Norm distance between:

[0075]

[0076]

[0077] in For the source domain to be correctly classified as category c The number of samples, For the source domain by the feature extractor Extracted features ;

[0078] Given head size μ The cumulative distribution function is obtained by using extremum theory and the Weibull distribution:

[0079]

[0080]

[0081] in For category c The cumulative distribution function, , , The shape, scaling, and position parameters of the Weibull distribution;

[0082] For the features of each sample in the target domain Calculate its relationship with Euclidean distance between them:

[0083]

[0084] in For the target domain, predict as the first c Pseudo-classes of classes;

[0085] Given threshold ε Calculate the category of the target domain:

[0086]

[0087] in For the target domain, predict as the first c The category of a class;

[0088] Therefore, feature extractor Classifier Domain Discriminator and The Extreme Value Learning Machine (EVM) forms an open set migration fault diagnosis model framework based on shared maximum mean difference.

[0089] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0090] The gas turbine tie rod rotor testing system provided by this invention features a drive control system that allows for flexible speed adjustment to meet testing conditions under various operating conditions. The tie rod rotor is supported by bearings and bearing housings and connected to a motor via a coupling, facilitating the setting of faults such as imbalance, misalignment, rubbing, and loose bearing housings. Furthermore, the rotor system has a simple structure and is easy to assemble and disassemble, which is beneficial for setting faults specific to the tie rod rotor. Vibration displacement signals at different axial positions of the rotor can be measured using eddy current displacement sensors at both the drive and non-drive ends, increasing the richness of the measured signal characteristics.

[0091] This invention provides an open set fault diagnosis method for gas turbine tie rod rotors. Based on convolutional neural networks, domain adaptation, and extremum theory, a feature extractor, classifier, domain discriminator, and extremum learning machine are established, forming the open set transfer fault diagnosis model framework proposed in this invention based on shared maximum mean difference. Traditional tie rod rotor fault diagnosis methods rely on the experience and knowledge of professional technicians, making it difficult to achieve fast and accurate diagnosis. This invention learns and trains the model by acquiring a large number of test signals from faulty rotors. The data-driven model can quickly, accurately, and robustly complete open set transfer fault diagnosis.

[0092] Furthermore, previous research on intelligent fault diagnosis based on deep learning has mostly focused on bearings and gearboxes, lacking research on rotors, especially gas turbine tie rod rotors. One important reason is the lack of corresponding test data. The gas turbine tie rod rotor fault testing system provided by this invention can realize signal acquisition of end face gear disk tie rod rotors under various operating conditions, various fault categories, and different fault degrees, which can provide sufficient and rich data for deep learning methods.

[0093] Furthermore, by building a low-cost gas turbine tie rod rotor fault testing system, various faults under different operating conditions of the rotor can be tested. This can replace the expensive actual gas turbine tie rod rotor main system test, greatly saving testing costs.

[0094] Furthermore, in addition to the four traditional rotor faults, this invention also sets four unique faults for the end face gear disk tie rod rotor, which not only deepens the research on fault diagnosis of tie rod rotors, but also fully verifies the applicability of the proposed diagnostic method and model to tie rod rotors.

[0095] Furthermore, this invention addresses the following issues: 1) When a model trained on a labeled dataset is applied to a similar but different target dataset, the model may fail; 2) The target dataset may contain unknown categories, and the model trained on the training set may misclassify unknown categories as known categories, leading to misjudgments. This invention proposes an open set transfer fault diagnosis model framework based on shared maximum mean difference. This framework fully learns the common category features of the source and target domains, enabling accurate and effective open set transfer fault diagnosis applications. Attached Figure Description

[0096] Figure 1 This is a flowchart illustrating the overall process of a gas turbine tie rod rotor fault testing system and its open set migration fault diagnosis method according to the present invention.

[0097] Figure 2 This is a schematic diagram of a gas turbine tie rod rotor fault testing system according to the present invention;

[0098] Figure 3 This is a schematic diagram of the main body of a gas turbine tie rod rotor fault testing system according to the present invention;

[0099] Figure 4 This is a schematic diagram of a tie rod rotor in a gas turbine tie rod rotor fault testing system according to the present invention;

[0100] Figure 5 Figures (a)-(d) are schematic diagrams of four unique faults of the end face gear disk tie rod rotor provided in this invention;

[0101] Figure 6 This is a schematic diagram of an open set migration fault diagnosis model framework based on shared maximum mean difference applicable to gas turbine tie rod rotors according to the present invention;

[0102] Figure 7 This is a schematic diagram of the feature extractor in the open set migration fault diagnosis model of the present invention;

[0103] Figure 8 This is a comparison chart of the fault diagnosis results of the tie rod rotor using the method of the present invention with other methods.

[0104] exist Figure 2 , Figure 3 and Figure 4In the diagram, 11 is a DC motor, 12 is a speed controller; 21 is a tie rod rotor, 22 is the first bearing, 23 is the second bearing, 24 is a coupling, 211 is the main wheel, 212 is the first test plate, 213 is the second test plate, 214 is the central tie rod, and 215 is a nut; 31 is the T-slot test platform base, 32 is the motor base, 33 is the first bearing seat, and 34 is the second bearing seat; 41 is the drive end eddy current displacement sensor, 42 is the non-drive end eddy current displacement sensor, 43 is a laser velocimeter, 44 is a DC regulated power supply, 45 is a dynamic data acquisition instrument, and 46 is a computer.

[0105] exist Figure 5 In the table, (a) represents a loose tie rod fault, (b) represents a first-type abnormal end face tooth contact fault, (c) represents a second-type abnormal end face tooth contact fault, and (d) represents a third-type abnormal end face tooth contact fault. Detailed Implementation

[0106] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This description will help those skilled in the art to further understand the present invention, but it does not limit the invention in any way. Various substitutions or modifications made based on ordinary technical knowledge and common practice in the art without departing from the above-described spirit of the invention should be included within the scope of the present invention.

[0107] Reference Figures 1 to 4 This invention provides a gas turbine tie rod rotor testing system, comprising a drive control system, a tie rod rotor main body system, a support system, and a measurement system. The drive control system includes a DC motor 11 and a speed controller 12; the tie rod rotor main body system includes a tie rod rotor 21, a first bearing 22 and a second bearing 23, and a coupling 24, wherein the tie rod rotor 21 includes a main wheel 211, a first test plate 212, a second test plate 213, a tie rod 214, and a nut 215; the support system includes a T-slot test platform base 31, a motor base 32, a first bearing seat 33, and a second bearing seat 34; the measurement system includes a drive-end eddy current displacement sensor 41, a non-drive-end eddy current displacement sensor 42, a laser velocimeter 43, a DC regulated power supply 44, a dynamic data acquisition instrument 45, and a computer 46. During system operation, the drive control system drives the tie rod rotor main body system, adjusts its speed, and controls the overall system's operational safety; the measurement system measures and analyzes the vibration displacement signals of the tie rod rotor main body system.

[0108] During operation, the DC motor 11 drives the tie rod rotor 21 to rotate, and the resulting vibration displacement signal is measured by the drive end eddy current displacement sensor 41 and the non-drive end eddy current displacement sensor 42, and transmitted to the dynamic data acquisition instrument 45 and computer 46 of the data acquisition system for analysis, display and storage.

[0109] The drive scheme adopts a speed regulator 12 to control the DC shunt motor 11. The motor shaft directly drives the tie rod rotor 21 via the coupling 24. The structure is simple and the speed range is wide. At the same time, the speed is directly measured and obtained by the laser tachometer 43.

[0110] The motor 11 is fixed on the T-slot test platform base 31 by the motor mount 32 to ensure that it is aligned with the main shaft of the tie rod rotor 21, and is connected to the tie rod rotor 21 by the coupling 24.

[0111] The tie rod rotor 21 is mounted on the first bearing seat 33 and the second bearing seat 34 respectively through the first bearing 22 and the second bearing 23 at both ends. The first bearing seat 33 and the second bearing seat 34 are fixed on the T-slot test platform base 31.

[0112] The rotor 21 has a main rotor 211, a first test rotor 212, and a second test rotor 213. The structure is second test rotor 213-main rotor 211-first test rotor 212. The rotors are pre-tightened by a central tie rod 214-nut 215 structure, and the torque is transmitted by end face teeth meshing, ensuring efficient and reliable torque transmission.

[0113] The drive end eddy current displacement sensor 41 and the non-drive end eddy current displacement sensor 42 are respectively arranged at the drive end of the tie rod rotor 21 near the motor 11 and the non-drive end away from the motor, and are used to measure and acquire the radial displacement signal of the tie rod rotor 21.

[0114] When the distance between the drive-end eddy current displacement sensor 41 and the non-drive-end eddy current displacement sensor 42 and the rotating shaft changes, the dynamic voltage signal generated by the sensor is input to the dynamic data acquisition instrument 45 and the computer 46. The dynamic data acquisition instrument 45 and the computer 46 analyze and obtain the dynamic displacement signal, and display and store it in real time.

[0115] Among them, the dynamic data acquisition instrument is used for high-speed, multi-channel simultaneous acquisition of dynamic voltage signals from sensors and converts the voltage signals into digital signals; the data processing computer is used for real-time display, processing and storage of signals, performs fast Fourier transform, and obtains time-domain and frequency-domain display of dynamic signals.

[0116] Reference Figure 5 The tested tie rod rotor main body system includes a total of 8 fault types. In addition to 4 traditional rotor faults (imbalance fault, misalignment fault, rubbing fault and bearing housing loose fault), 4 tie rod rotor-specific faults were also set, including tie rod loose fault and 3 types of abnormal end face tooth contact faults, to simulate poor contact faults that occur after the end face teeth of the tie rod rotor wear.

[0117] The fault of loose tie rod is simulated by slightly loosening nut 215, which will result in insufficient preload of the wheel disc; the fault of abnormal end face tooth contact is simulated by adding a thin iron plate between two meshing tooth surfaces. The first type of abnormal end face tooth contact fault is simulated by adding a thin iron plate between the two teeth of the first test disc 212 and the main wheel disc 211, and adding the same thin iron plate at the same position on the second test disc 213, which will cause the rotor to bend initially; the second type of abnormal end face tooth contact fault is simulated by adding a thin iron plate between the two teeth of the first test disc 212 and the main wheel disc 211 at the same position, and adding the same thin iron plate at the opposite 180° position on the second test disc 213, which will cause the rotor to twist; the third type of abnormal end face tooth contact fault is simulated by adding a thin iron plate between the two teeth of the first test disc 212 and the main wheel disc 211 at a distance of 180°, and not adding a thin iron plate on the second test disc 213, which will result in incomplete end face tooth meshing.

[0118] Reference Figure 6 and Figure 7 The present invention provides a method for diagnosing open set faults in a gas turbine tie rod rotor, comprising the following steps:

[0119] Step 1: Acquire vibration displacement signals of a healthy rotor under multiple operating conditions. The rotational speed is adjusted by regulating the speed controller in the drive system and observing the laser tachometer in the measurement system, thus enabling testing under multiple operating conditions.

[0120] Step 2: Collect vibration displacement signals of rotors with different faults under corresponding operating conditions. Faults are set in the healthy rotor system, and signals are collected under different operating conditions for different fault degrees and categories of the faulty rotor to obtain a tie-rod rotor fault dataset. The fault degrees include two categories: minor and severe. The fault types include imbalance, misalignment, rubbing, loose bearing housing, loose tie rod, and three types of abnormal end-face tooth contact, totaling eight faults. Including the healthy rotor, there are a total of nine categories. Different operating conditions specifically refer to different operating speeds, set from low to high at 100 rpm, 1500 rpm, and 2000 rpm.

[0121] Step 3: Divide the above vibration displacement signals into source and target domains with different operating conditions and fault types, and set up an open set migration fault diagnosis task. This specifically includes the following steps:

[0122] 1) Data preprocessing

[0123] Overlap sampling data augmentation is employed to extend long-time-series samples. A fixed-length window slides along the temporal direction of the signal, extracting a segment of the signal within the window as a sample. Overlapping segments are permissible. The first T segments are selected along the temporal direction. r Each segment is used as the training sample set. The remaining T e One segment was used as the test sample set. There is no overlap between the training sample set and the test sample set.

[0124] The data is normalized using the following formula:

[0125]

[0126] in This represents the minimum value in the dataset. This represents the maximum value in the dataset. This represents the normalized dataset. Since all data has undergone normalization, the following text... and Mixed use.

[0127] 2) Divide the source domain and the target domain

[0128] Different fault category signals are randomly selected from the fault dataset to form source and target domain data, constituting a subset of the dataset. This subset is used to set up an open-set transfer fault diagnosis task at different speeds. The data in the source domain includes the source domain training set. Source domain test set The data in the target domain includes the target domain training set. and target domain test set Both the source domain data and the target domain test set have classification labels. , and The target domain training set has no labels.

[0129] Step 4: Construct a fault feature extraction network based on a convolutional neural network to learn from the vibration displacement signal and obtain the feature extractor. .

[0130] Step 5: Construct a fault classifier based on a fully connected neural network and input it into the feature extractor. The extracted features are used to calculate the fault label probability distribution of the signal, thus obtaining the classifier. .

[0131] Step 6: Construct an open set migration fault diagnosis method based on domain adaptation, using Shared Maximum Mean Difference (SMMD) to measure the difference between the target domain and the source domain in the open set, and learn a feature extractor. Parameters; design a domain discriminator, domain discriminator This is used to quantify the similarity between features and the source or target domain, thereby obtaining the weights for samples belonging to a shared domain; this is the domain discriminator. Introducing a global feature learning approach, and using domain adversarial training to improve the feature extractor Learn the characteristics of the source and target domains.

[0132] Step 7: Construct an open set fault classifier based on extreme value theory, calculate the features of the source domain and the target domain respectively, and obtain the extreme value learning machine EVM.

[0133] Specifically, in steps 4-7, a domain adaptation method for open set transfer is proposed, which learns the features of common categories in the source and target domains while identifying private categories within the target domain. Specifically, this involves a method based on shared maximum mean difference, resulting in a diagnostic model framework based on shared maximum mean difference, as described above. Figure 6 The features extracted by the model can express the common and private categories of the source and target domains, thereby identifying the private categories of the target domain. This specifically includes the following steps:

[0134] 1) Construct an open set migration fault diagnosis model based on the shared maximum mean difference.

[0135] Reference Figure 7 First, a fault feature extraction network based on a convolutional neural network is built as the feature extractor. A convolutional neural network is used to automatically extract fault features. The network input data is the rotor vibration displacement signal. The network consists of one convolutional layer, four residual blocks, and one average pooling layer. Each residual block includes three convolutional layers, three batch normalization layers, and two ReLU activation functions, representing a skip-layer connection form of the ResNet network. The output is the extracted signal features. The network mapping relationship is as follows:

[0136]

[0137] in and These are rotor vibration displacement signal samples from the source domain and the target domain, respectively. The convolutional neural network mapping for the feature extractor. and These are the signal features extracted from samples in the source and target domains.

[0138] The high-level features obtained from the upper convolutional layers are flattened to normalize their dimensions before being fed into the classifier. Classifier It consists of one fully connected layer and a SoftMax function. The SoftMax function calculates the probability distribution of the fault category labels of the input data to obtain the source class. pseudo-distribution on :

[0139]

[0140] By minimizing the classification loss function Learning Feature Extractor and classifier The model parameters enable the feature extractor Learn the features of the source domain and train a classifier that can accurately classify the source domain data. :

[0141]

[0142] in The number of source domain samples used in training. The number of source domain categories, The true label of the source domain.

[0143] The proposed Shared Maximum Mean Discrepancy (SMMD) is used to measure the common features between the source and target domains, and to measure the difference between the target and source domains on an open set. A feature extractor is learned by minimizing the loss function. The parameters. Unlike the traditional definition of Maximum Mean Discrepancy (MMD), SMMD aims to adapt common features between the source and target domains at the sample level, and is defined as follows:

[0144]

[0145] in The number of samples in the source domain. The number of samples in the target domain. Represents the source domain. Indicates the target domain; and Let be the weights of samples from the source domain and the target domain that belong to the common domain. The larger the weight, the more likely the sample is to belong to the common domain. If the weights are all equal to 1, the above formula will degenerate into MMD.

[0146] Use the designed domain discriminator By quantifying the similarity between feature z and the source or target domain, the weights of the shared domain can be obtained:

[0147]

[0148] in and Used to measure the degree of similarity between the source and target domains. The weights are calculated as follows:

[0149]

[0150] By minimizing the loss Learning Domain Discriminator The parameters enable it to distinguish between data from the source and target domains:

[0151]

[0152] The above weight calculation method shows that when When the weight decreases, If the value is close to 1, it indicates that the sample in the source domain should receive more attention in the transfer calculation; similarly, when When the weight increases, A value close to 1 indicates that this sample in the target domain should also receive more attention in the transfer calculation.

[0153] Use the designed domain discriminator The feature extractor is trained through domain adversarial training. This approach learns features from both the source and target domains, introducing global feature learning to address the issue where the network focuses too much on fine-grained similarity between data points while neglecting the overall similarity of the data distribution. This is achieved by minimizing the loss... Learning Domain Discriminator Parameters and update feature extractor The parameters enable it to learn features of both the source and target domains:

[0154]

[0155]

[0156] in and Domain discriminator The output.

[0157] In summary, the final optimized model of the diagnostic model based on the shared maximum mean difference is:

[0158]

[0159] 2) Constructing an extreme value learning machine based on extreme value theory

[0160] Using an Extreme Value Machine (EVM) based on extreme value theory to predict unknown categories in open set identification improves the ability of deep learning to detect open set samples. This is particularly relevant for feature extractors. If well-trained, the features of fault categories in the source and target domains will be clearly separable in the hyperdimensional space, and the features of each category in the source domain will be... Calculate its average characteristic Norm distance between:

[0161]

[0162]

[0163] in For the source domain to be correctly classified into categories c The number of samples, For the source domain by the feature extractor Extracted features .

[0164] Given head size μ The cumulative distribution function is calculated using extremum theory and the Weibull distribution:

[0165]

[0166]

[0167] in For category c The cumulative distribution function, , , The shape, scaling, and position parameters are those of the Weibull distribution.

[0168] For the features of each sample in the target domain Calculate its relationship with Euclidean distance between them:

[0169]

[0170] in For the target domain, predict as the first c Pseudo-classes of classes.

[0171] Given threshold ε Calculate the category of the target domain:

[0172]

[0173] in For the target domain, predict as the first c The category of a class.

[0174] By feature extractor Classifier Domain Discriminator and The Extreme Value Learning Machine (EVM) forms an open set migration fault diagnosis model framework based on shared maximum mean difference.

[0175] Step 8: Train the above feature extractor by optimizing the model described in steps 4 to 7. Classifier Domain Discriminator and This yields a well-trained feature extractor. and classifier Use the trained feature extractor Classifier Together with the constructed Extreme Value Learning Machine (EVM), features in the target domain are extracted, and it is determined whether the feature can be classified into a known common category. If it is a common category, further prediction determines the specific category it belongs to. This achieves the task of fault diagnosis for open set migration of target domain data.

[0176] In addition, to evaluate the effectiveness of the method, four types of precision are defined to describe the diagnostic accuracy:

[0177]

[0178]

[0179]

[0180]

[0181] in This represents the number of common categories that are correctly classified in the test sample. This represents the number of unknown categories correctly classified in the test sample. This represents the total number of shared categories in the test samples. The value is the maximum between the total number of unknown classes in the test samples and the number predicted as unknown classes. AAL measures the average accuracy in identifying common and unknown classes, AS measures the accuracy of the model in identifying common classes, AU measures the accuracy of the model in identifying unknown classes, and H... * Overall accuracy was measured.

[0182] To verify the effectiveness of the proposed method in different open set fault diagnosis tasks, the publicly available CWRU bearing fault dataset is used in Example 1 below, and the central tie rod rotor fault dataset acquired by the gas turbine tie rod rotor fault test system of the present invention is used in Example 2.

[0183] Example 1

[0184] The open set migration fault diagnosis method of this invention is used to perform open set migration fault diagnosis on the CWRU bearing fault dataset, as detailed below:

[0185] Different fault categories were randomly selected from the CWRU bearing fault dataset to create different datasets. The descriptions of each category in the CWRU bearing fault dataset are shown in Table 1, where fault types H, BF, IF, and OF represent healthy bearings, ball bearing faults, inner ring faults, and outer ring faults, respectively. Ball bearing faults, inner ring faults, and outer ring faults each contain three different levels of fault severity. The open set identification task settings based on the CWRU dataset are shown in Table 2.

[0186] Table 1. Fault Categories in the Bearing Fault Dataset

[0187]

[0188] Table 2. Open Set Fault Diagnosis Task Settings for Bearing Fault Dataset

[0189]

[0190] The diagnostic results of the method of this invention are shown in Table 3. Examining the four indicators, across all tasks, the average value of AAL was 92.6%, the average value of AS was 91.2%, the average value of AU was 92.7%, and the average value of H... * The average value was 91.9%, all exceeding 90%, indicating a high diagnostic accuracy and excellent performance of the proposed diagnostic method.

[0191] Table 3. Fault diagnosis results of the open set of the bearing fault dataset.

[0192]

[0193] Example 2

[0194] Using the gas turbine tie rod rotor fault testing system and its open set migration fault diagnosis method of the present invention, an open set fault diagnosis is performed on a central tie rod rotor, as follows:

[0195] Using the gas turbine tie rod rotor fault testing system of this invention, fault data of the central tie rod rotor was collected to establish a fault dataset. The descriptions and labels of each fault category are shown in Table 4. Different fault categories were randomly selected from the fault dataset to create different task datasets. The open set identification task settings are shown in Table 5. The datasets M11 to M63 and S11 to S63 adopted the dataset settings for more severe and less severe rotor faults, respectively.

[0196] Table 4 shows the fault types in the tie rod rotor fault dataset.

[0197]

[0198] Table 5. Open Set Fault Diagnosis Task Settings for Tie Rod Rotor Fault Dataset

[0199]

[0200] The fault diagnosis results of the method of this invention on the central tie rod rotor fault dataset are shown in Table 6. Diagnosis was performed on two datasets: one with severe faults and one with mild faults. The accuracy of all indicators was at least 83.15%, exceeding 80%. Specifically, on the dataset with severe rotor faults, the average accuracy for identifying common and unknown categories was 89.03%, and the average accuracy for identifying common categories was 92.96%, higher than on the dataset with milder faults.

[0201] Table 6. Fault diagnosis results of the open set of the tie rod rotor fault dataset.

[0202]

[0203] Reference Figure 8 This study compares the accuracy of the proposed method with that of the classic deep learning method ResNet, the classic transfer learning method DDC, and the method itself on a tie rod rotor fault dataset. Taking the average AAL (Advanced Ability Level) across all diagnostic tasks as an example, the accuracy of the proposed method is observed. When the rotor fault severity is high, the accuracy is high at 89.03%, exceeding 89%. When the rotor fault severity is low, the accuracy decreases to 85.07%, but still exceeds 80%, both significantly higher than the classic deep learning method ResNet and the classic transfer learning method DDC. Compared with traditional deep learning and transfer learning methods, the proposed method achieves the highest diagnostic accuracy.

[0204] The present invention has been described in detail above with general descriptions and specific embodiments. It should be understood that various changes or modifications can be made within the scope of the claims based on the present invention, which will be obvious to those skilled in the art and do not affect the essence of the present invention. Therefore, all such changes or modifications made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for diagnosing open set faults in a gas turbine tie rod rotor, characterized in that, This method is based on a gas turbine tie rod rotor testing system, which includes a drive control system, a tie rod rotor main body system, a support system, and a measurement system. The drive control system is used to drive the tie rod rotor main system and adjust the speed and control the operational safety of the entire system. The main body system of the tie rod rotor is used to set general rotor faults and tie rod rotor-specific faults; Support system, used to install and support the drive control system motor, the tie rod rotor body system, and the measurement system sensors; The measurement system is used to measure and analyze the vibration displacement signals of the tie rod rotor main body system; The drive control system includes a motor and a speed controller; the tie rod rotor main body system includes a tie rod rotor, bearings, and couplings; the support system includes a T-slot test platform base, a motor base, and a bearing base; the measurement system includes an eddy current displacement sensor, a laser velocimeter, and a data acquisition system. During operation, the motor drives the rod rotor to rotate, and the resulting vibration displacement signal is measured by the eddy current displacement sensor and transmitted to the data acquisition system for analysis, display and storage. A DC shunt motor drive scheme is adopted, in which the motor shaft directly drives the tie rod rotor via a coupling, and the speed is directly measured and obtained by a laser tachometer. The motor is fixed to the base of the T-slot test platform by a motor mount to ensure alignment with the main shaft of the tie rod rotor, and is connected to the tie rod rotor by a coupling. The tie rod rotor is mounted on bearing housings via bearings at both ends, and the bearing housings are fixed to the base of the T-slot test platform; The rotor discs of the tie rod are divided into test discs and main discs. The structure is test disc-main disc-test disc. The discs are pre-tightened by a tie rod-nut structure and the torque is transmitted by the meshing of the end face teeth to ensure reliable torque transmission. Eddy current displacement sensors are respectively arranged at the drive end of the tie rod rotor near the motor and the non-drive end far from the motor, and are used to measure and acquire the vibration displacement signal of the tie rod rotor. When the distance between the eddy current sensor and the rotating shaft changes, the dynamic voltage signal generated by the sensor is input to the data acquisition system. The data acquisition system analyzes and obtains the dynamic displacement signal, and displays and stores it in real time. The method includes the following steps: Step 1: Acquire vibration displacement signals of a healthy rotor under multiple operating conditions; Step 2: Collect vibration displacement signals of rotors with different faults under corresponding operating conditions; Step 3: Divide the vibration displacement signals from Step 1 and Step 2 into source and target domains with different working conditions and fault types, and set up an open set migration fault diagnosis task. Step 4: Construct a fault feature extraction network based on a convolutional neural network to learn the vibration displacement signals from Steps 1 and 2, thereby obtaining the feature extractor. ; Step 5: Construct a fault classifier based on a fully connected neural network and input it into the feature extractor. The extracted features are used to calculate the fault label probability distribution of the signal, thus obtaining the classifier. ; Step 6: Construct an open set migration fault diagnosis method based on domain adaptation, using Shared Maximum Mean Difference (SMMD) to measure the difference between the target domain and the source domain in the open set, and learn a feature extractor. Parameters; design a domain discriminator, domain discriminator This is used to quantify the similarity between features and the source or target domain, thereby obtaining the weights for samples belonging to a shared domain; this is the domain discriminator. Introducing a global feature learning approach, and using domain adversarial training to improve the feature extractor Learn the characteristics of the source and target domains; Step 7: Construct an open set fault classifier based on extreme value theory, calculate the features of the source domain and the target domain respectively, and obtain the extreme value learning machine EVM; Step 8, Training the feature extractor Classifier Domain Discriminator and An open-set transfer learning fault diagnosis model consisting of an extreme value learning machine (EVM) and a fully trained feature extractor is used. Classifier And the Extreme Value Learning Machine (EVM) is used to perform open set migration fault diagnosis tasks for target domain data.

2. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 1, characterized in that, In step 1, the rotational speed is adjusted by regulating the speed controller in the drive system and observing the laser velocimeter in the measurement system, thereby enabling testing under multiple operating conditions.

3. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 1, characterized in that, In step 2, a fault is set in the healthy rotor system, and signals are collected under different working conditions for different fault degrees and different fault categories of the faulty rotor. The vibration displacement signals collected in step 1 and step 2 together constitute the tie rod rotor fault dataset. The fault severity is divided into two categories: minor and severe. The fault types include imbalance, misalignment, rubbing, loose bearing housing, loose tie rod, and three types of abnormal end face tooth contact, totaling eight types of faults. Together with the healthy rotor, there are a total of nine categories. Different operating conditions refer to different operating speeds, with three speeds set from low to high: 100 rpm, 1500 rpm, and 2000 rpm.

4. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 3, characterized in that, In step 3, the vibration displacement signals from steps 1 and 2 are divided into source and target domains with different operating conditions and fault types. An open set migration fault diagnosis task is then set up, which specifically includes the following steps: 1) Data preprocessing Deep learning methods use a large amount of data to train the model. To obtain a sufficient number of training samples, data augmentation techniques such as overlapping sampling are used to expand long-term samples. A fixed-length window slides along the temporal direction of the signal, and a segment of the signal contained within the window is extracted as a sample. There is overlap between adjacent sample segments. Along the temporal direction, the first T segments are selected. r Each segment is used as the training sample set. The remaining T e One segment was used as the test sample set. At the same time, historical information is used to diagnose potential future failures. The data is normalized using the following formula: in This represents the minimum value in the dataset. This represents the maximum value in the dataset. This represents the normalized dataset. Since all data has undergone normalization, therefore... and Mixed use; 2) Divide the source domain and the target domain Different fault category signals are randomly selected from the tie rod rotor fault dataset to form source and target domain data, creating subsets for setting up open-set transfer fault diagnosis tasks at different speeds. Based on the above method of dividing the training and test sample sets, the data in the source domain includes the source domain training set. Source domain test set The data in the target domain includes the target domain training set. and target domain test set Both the source domain data and the target domain test set have classification labels. , and The target domain training set has no labels.

5. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 4, characterized in that, In steps 4 through 7, a domain adaptation method oriented towards open set transfer is adopted, specifically the method of sharing the maximum mean difference. This yields a diagnostic model framework based on the shared maximum mean difference. The extracted features can express the common and private categories of the source and target domains, thereby identifying the private categories of the target domain. The specific steps include: 1) Construct an open set migration fault diagnosis model based on the shared maximum mean difference. First, a fault feature extraction network based on a convolutional neural network is built as the feature extractor. A convolutional neural network is used to automatically extract fault features, with the rotor vibration displacement signal as the network input data. The network consists of one convolutional layer, four residual blocks, and one average pooling layer. Each residual block includes three convolutional layers, three batch normalization layers, and two ReLU activation functions, representing a skip-layer connection form of the ResNet network. The output is the extracted signal features. The network mapping relationship is as follows: in and These are rotor vibration displacement signal samples from the source domain and the target domain, respectively. The convolutional neural network mapping for the feature extractor. and Signal features extracted from samples in the source and target domains; The high-level features obtained from the upper convolutional layers are flattened to normalize their dimensions before being fed into the classifier. Classifier It consists of one fully connected layer and a SoftMax function. The SoftMax function calculates the probability distribution of the fault category labels of the input data to obtain the source class. pseudo-distribution on : By minimizing the classification loss function study and The model parameters enable the feature extractor Learn the features of the source domain and train a system that can accurately classify the source domain data. : in The number of source domain samples used in training. The number of source domain categories, The true label of the source domain; To learn the shared features between the source and target domains, a Shared Maximum Mean Difference (SMMD) metric is proposed. This metric measures the difference between the target and source domains on an open set and learns a feature extractor by minimizing the loss function. The parameters; SMMD aims to adapt common features between the source and target domains at the sample level; therefore, SMMD is defined as follows: in The number of samples in the source domain. The number of samples in the target domain. Represents the source domain. Indicates the target domain; and The weights are the values ​​of the samples from the source domain and the target domain that belong to the common domain. The larger the weight, the more likely the sample is to belong to the common domain. If the weights are all equal to 1, the above formula will degenerate into the traditional maximum mean difference (MMD). To obtain the weights of the shared domain, a domain discriminator was designed. Used to quantify the similarity between feature z and the source or target domain: in and Used to measure the degree of similarity between the source domain and the target domain; if the source domain has the first... i one sample A value close to 1 indicates that the sample is highly likely to have come from [a specific source]. Conversely, if A value close to 0 indicates that the feature of the sample is highly likely to be similar to a feature in the target domain; similarly, if the first feature in the target domain is similar to a feature in the target domain, then the sample's feature is likely to be similar to a feature in the target domain. j one sample A value close to 0 indicates that the sample is highly likely to have come from [a specific source]. Conversely, if A value close to 1 indicates that the sample's features are highly likely to be similar to a feature in the source domain; the weights are calculated as follows: By minimizing the loss Learning Domain Discriminator The parameters enable it to distinguish between data from the source and target domains: The above weight calculation method shows that when When the weight decreases, If the value is close to 1, it indicates that the sample in the source domain should receive more attention in the transfer calculation; similarly, when When the weight increases, A value close to 1 indicates that the sample in the target domain receives more attention during the transfer calculation. Introducing a global feature learning approach, and using domain adversarial training to improve the feature extractor To learn the characteristics of the source and target domains, a domain discriminator is designed. By minimizing loss Learning Domain Discriminator Parameters and update feature extractor The parameters enable it to learn features of the source and target domains: in and Domain discriminator The output; In summary, the final optimized model of the diagnostic model based on the shared maximum mean difference is: 2) Constructing an extreme value learning machine based on extreme value theory An extreme value learning machine (EVM) based on extreme value theory is used to predict unknown categories, thereby improving the ability of deep learning to detect open set samples. If feature extractor If well-trained, the features of fault categories in the source and target domains will be clearly separable in the hyperdimensional space, and the features of each category in the source domain will be... Calculate its average characteristic Norm distance between: in For the source domain to be correctly classified into categories c The number of samples, For the source domain by the feature extractor Extracted features ; Given head size μ The cumulative distribution function is obtained by using extremum theory and the Weibull distribution: in For category c The cumulative distribution function, , , The shape, scaling, and position parameters of the Weibull distribution; For the features of each sample in the target domain Calculate its relationship with Euclidean distance between them: in For the target domain, predict as the first c Pseudo-classes of classes; Given threshold ε Calculate the category of the target domain: in For the target domain, predict as the first c The category of a class; Therefore, feature extractor Classifier Domain Discriminator and The Extreme Value Learning Machine (EVM) forms an open set migration fault diagnosis model framework based on shared maximum mean difference.

6. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 1, characterized in that, The tie rod rotor includes an end face gear disk, a tie rod, and a nut. The end face gear disk is fitted onto the tie rod and tightened by the nut.

7. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 1, characterized in that, The data acquisition system includes a dynamic data acquisition instrument and a data processing computer; A dynamic data acquisition instrument is used for high-speed, multi-channel simultaneous acquisition of dynamic voltage signals from sensors and converts the voltage signals into digital signals. A data processing computer is used for real-time display, processing, and storage of signals, performing fast Fourier transforms to obtain time-domain and frequency-domain displays of dynamic signals.

8. The method for diagnosing open set faults in a gas turbine tie rod rotor according to claim 1, characterized in that, The tested tie rod rotor main body system includes a total of 8 types of faults, namely imbalance fault, misalignment fault, rubbing fault, bearing housing loose fault, tie rod loose fault, and 3 types of abnormal end face tooth contact faults. Three types of abnormal end face tooth contact faults are used to simulate poor contact faults that occur after the end face teeth of the tie rod rotor wear. Among them, loose tie rods will cause insufficient preload of the wheel disc; abnormal end face tooth contact in type 1 will cause initial bending of the rotor; abnormal end face tooth contact in type 2 will cause twisting of the rotor; and abnormal end face tooth contact in type 3 will cause incomplete meshing of the end face teeth.

Images