A steam turbine generator rotor winding inter-turn short circuit fault intelligent detector

Abstract

The application discloses a kind of steam turbine generator rotor winding interturn short circuit fault intelligent detector, comprising: vibration acquisition module, for collecting stator excitation end vibration signal;Rotational speed phase reference acquisition module, for collecting and collecting key phase or speed pulse signal synchronous with rotor speed;Synchronization preprocessing module, for the vibration signal and the rotational speed signal are synchronized alignment, abnormal value is eliminated, FFT transform and 40-1000Hz passband amplitude sequence extraction.The advantages of the present application are: the fault sensitive quantity R=A50 / A100 that can be explained physically is proposed, directly reflects the air gap magnetic field imbalance caused by interturn short circuit and its 50Hz / 100Hz electromagnetic force wave vibration response;Adaptive optimization mode number K is used SVMD algorithm, effectively suppresses modal aliasing, enhances weak fault feature retention capability, improves early interturn short circuit detection rate.Focus on 30-240Hz frequency band in fault related IMF, combined with sliding rectangular window reconstruction time-frequency joint feature matrix, significantly improve composite fault recognition accuracy and noise robustness.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and fault diagnosis technology for large synchronous motors, specifically to an intelligent detector for inter-turn short-circuit faults in the rotor windings of a steam turbine generator. Background Technology

[0002] With the increasing proportion of new energy grid connection, steam turbine generators have shifted from base load operation to flexible adjustment. The rotor windings are subjected to thermal, mechanical, and electromagnetic stress coupling for a long time, which accelerates the aging of inter-turn insulation and easily leads to inter-turn short circuits. This fault causes magnetomotive force asymmetry, air gap magnetic field distortion, and radial electromagnetic force imbalance, inducing abnormal vibration at the excitation end. If it is not detected in time, it will aggravate insulation damage and even cause the unit to shut down.

[0003] Existing detection methods are divided into two categories: offline and online. Offline methods (such as impedance and pulse reflection) require shutdown and do not meet the requirements for continuous operation and monitoring. Among online methods, single feature methods have a high misjudgment rate under complex working conditions, strong noise, and compound faults. Traditional empirical mode decomposition (EMD) suffers from mode aliasing, and weak fault features are easily submerged. Conventional LSTM / GRU has a significantly reduced recognition performance under low signal-to-noise ratio, while LSTM / GRU models that integrate SVMD and multimodal time-frequency features have stronger noise resistance.

[0004] Current solutions still have two major shortcomings: first, the correlation between fault indicators and physical mechanisms is weak, resulting in insufficient interpretability; second, feature extraction, mode decomposition, and neural network recognition are disconnected, making it difficult to build a replicable and deployable online detection system. This invention constructs an integrated intelligent detection technology chain based on the main line of "fault-sensitive quantity—successive variational mode decomposition—multimodal time-frequency matrix—LSTM / GRU level output". Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent detector for inter-turn short-circuit faults in the rotor windings of a steam turbine generator. It extracts the 50Hz and 100Hz components from the stator excitation end vibration as fault-sensitive features; adaptively optimizes the mode decomposition parameters using a successive variational mode decomposition algorithm; constructs a multi-mode time-frequency joint feature matrix based on the fault-related intrinsic mode components; and finally inputs the data into an LSTM or GRU network to directly output the fault category and level.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an online intelligent detector and method for inter-turn short-circuit faults in the rotor winding of a steam turbine generator, comprising eight modules: 1. Vibration acquisition module: Sensors are placed at any position on the upper or lower part of the excitation end bearing housing (main measuring point) and the excitation end housing to collect the vibration signal of the stator excitation end; 2. Rotor speed phase reference acquisition module: Acquires key phase or speed measurement pulse signals to provide a rotor phase reference for synchronous analysis; 3. Synchronization preprocessing module: Synchronizes and aligns vibration and phase signals, removes data with timestamp jumps, abnormal sampling rates or lengths, and extracts passband amplitude sequences; 4. Fault Sensitive Quantity Extraction Module: Inter-turn short circuits disrupt the symmetry of the air gap magnetic field, simultaneously enhancing the terminal vibration of the 50Hz / 100Hz electromagnetic force coupling. The 50Hz amplitude A50 and the 100Hz amplitude A100 are extracted from the synchronous spectrum to construct a physically interpretable index R = A50 / A100. 5. Successive Variational Mode Decomposition Module: SVMD is used to adaptively optimize the number of modes K to obtain the IMFs of the passband amplitude sequence; 6. Multimodal feature construction module: IMFs with center frequencies of 30–240Hz are selected as fault-related components, and segmented reconstruction is performed through a sliding rectangular window with a window length of 20 / step size of 4 to generate a time-frequency joint feature matrix; 7. Temporal Intelligent Recognition Module: Input the feature matrix into an improved LSTM or GRU network (including input layer, tensor reconstruction layer, recurrent layer, activation layer, fully connected layer, and Softmax layer), and output the category labels and corresponding probabilities of 12 types of faults, covering healthy state, single faults (inter-turn short circuit, bearing wear, shaft imbalance), two types of composite faults (inter-turn short circuit + bearing wear, inter-turn short circuit + shaft imbalance), and 6 different levels of inter-turn short circuit faults, and provide the confidence interval for each type of fault; 8. Fault Comprehensive Assessment Module: Integrates R-value and inter-turn short-circuit posterior probability to output fault status and level information.

[0007] The corresponding detection methods and steps correspond one-to-one with the module functions, enabling end-to-end online diagnosis.

[0008] Compared with the prior art, the present invention provides an intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator, which has the following advantages: The intelligent detector for inter-turn short circuit faults in the rotor winding of the steam turbine generator proposes a physically explainable fault sensitivity quantity R = A50 / A100, which directly reflects the air gap magnetic field imbalance caused by the inter-turn short circuit and its 50Hz / 100Hz electromagnetic force wave vibration response. The SVMD algorithm is used to adaptively optimize the number of modes K, which effectively suppresses mode mixing, enhances the ability to retain weak fault features, and improves the early inter-turn short circuit detection rate. Focusing on fault-related IMFs in the 30–240Hz frequency band, and combining a sliding rectangular window to reconstruct the time-frequency joint feature matrix, the accuracy of composite fault identification and its robustness against noise are significantly improved. End-to-end fault classification and severity assessment based on improved LSTM or GRU networks have stronger time-series modeling capabilities, noise suppression capabilities, and field applicability compared to traditional methods that rely on single-domain indicators. Attached Figure Description

[0009] Figure 1 This is a block diagram of the overall structure of the intelligent detector for inter-turn short circuit faults in the rotor winding of a steam turbine generator according to the present invention. Figure 2 This is a flowchart of the method for detecting inter-turn short circuit faults in the rotor winding of a steam turbine generator according to the present invention; Figure 3 This is a schematic diagram of the electromagnetic-mechanical coupling mechanism of the rotor winding inter-turn short circuit fault in this invention; Figure 4 This is a schematic diagram illustrating the extraction of fault-sensitive quantities at the excitation end of the present invention at 50Hz / 100Hz. Figure 5 This is a flowchart of the successive variational mode decomposition algorithm of the present invention; Figure 6 This is a schematic diagram illustrating the construction of the multimodal fault feature matrix of the present invention; Figure 7 This is a schematic diagram of the improved long short-term memory network / gated recurrent unit network recognition structure of the present invention; Figure 8 This is a schematic diagram of the comprehensive evaluation output of fault status and inter-turn short circuit level of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Please see Figure 1 The intelligent detector for inter-turn short circuit faults in the rotor winding of a steam turbine generator includes: a vibration acquisition module for acquiring vibration signals at the stator excitation end; The rotational speed phase reference acquisition module is used to acquire key phase or speed measurement pulse signals that are synchronized with the rotor speed. The synchronous preprocessing module is used to synchronize and align vibration signals and rotation speed signals, remove outliers, perform FFT transformation, and extract the 40–1000 Hz passband amplitude sequence. The fault sensitivity extraction module is used to extract the 50 Hz amplitude A50 and the 100 Hz amplitude A100 from the spectrum and construct the fault sensitivity index R = A50 / A100, where the R index reflects the 50 Hz / 100 Hz electromagnetic force wave vibration response ratio caused by the asymmetry of the air gap magnetic field due to the inter-turn short circuit. The SVMD module is used to adaptively optimize the number of modes K by minimizing the modal mixing energy, and to perform VMD decomposition on the passband amplitude sequence to obtain multiple intrinsic mode functions (IMFs). The multimodal feature construction module is used to select fault-related IMFs, segment them through a sliding window, and then fuse time-frequency features to construct a joint time-frequency feature matrix. The temporal intelligent recognition module is used to input the feature matrix into the LSTM or GRU network and output the health status, fault status, fault level and confidence probability. The fault comprehensive assessment module is used to output fault status and inter-turn short circuit fault level information based on health / fault status estimation results, confidence probability and preset thresholds; Among them, the R index construction, SVMD module and time-series intelligent recognition module work together to improve the recognition accuracy in low signal-to-noise ratio and complex fault scenarios.

[0012] The vibration acquisition module is deployed at least at one of the following locations: the side of the excitation bearing housing, the upper part of the casing, or the lower part of the casing, with the side of the excitation bearing housing as the main measurement point. After the SVMD module adaptively decomposes the passband amplitude sequence into multiple IMFs, the multimodal feature construction module selects IMFs with center frequencies in the range of 30–240 Hz as fault-related candidate modes, arranged in ascending order of center frequency. The candidate fault mode components include IMF1, IMF2, IMF3, and IMF4, where IMF1 represents the low-frequency vibration trend, and IMF2–IMF4 are located in the 30–240 Hz fault-related frequency band, used to construct the input matrix of the neural network. The multimodal feature construction module uses a sliding rectangular window to segment the passband amplitude sequence of length N and the selected IMFs: setting the window length to n and the step size to inc, resulting in m equal-length subsequences, where… The window length n is an integer from 16 to 24, and the step size inc is an integer of 2. The preferred scheme is n = 20 and inc = 4. The timing intelligent recognition module sequentially includes an input layer, a tensor reconstruction layer, an LSTM or GRU layer, an activation layer, a fully connected layer, a Softmax layer, and a classification layer. The categories output by the classification layer include at least: health status, six different levels of rotor winding inter-turn short circuit, bearing wear, shaft dynamic imbalance, and two types of composite faults—inter-turn short circuit + bearing wear, and inter-turn short circuit + shaft dynamic imbalance. The fault comprehensive evaluation module is specifically configured to jointly determine the inter-turn short circuit fault state and its specific level based on the fault sensitivity quantity R and the posterior probability of inter-turn short circuit output by the Softmax layer.

[0013] Method for detecting inter-turn short-circuit faults in the rotor winding of a steam turbine generator: including the following steps: S1: Acquire stator excitation end vibration signal and synchronous speed phase reference signal; S2: Synchronize and align the two signals, remove abnormal data, perform discrete spectrum transformation, and extract the 40–1000Hz passband amplitude sequence; S3: Extract the 50 Hz amplitude from the vibration spectrum and 100 Hz amplitude Construct the fault sensitivity quantity R = A50 / A100; S4: With the goal of minimizing the energy of modal mixing, SVMD is used to adaptively optimize the number of modes K, and modal decomposition is performed on the passband amplitude sequence to obtain multiple IMFs; S5: Select IMFs with center frequencies in the range of 30–240 Hz as fault-related mode components, and construct a time-frequency joint feature matrix through sliding window segmentation and tensor reconstruction; S6: Input the feature matrix into an LSTM or GRU network and output the fault category and recognition probability; S7: Integrates R index, identification probability and preset criteria to determine fault status and inter-turn short circuit level.

[0014] The abnormal data removal mentioned in step S2 specifically targets data with timestamp jumps, sampling rate deviations, and record length deviations exceeding the limits; the IMFs mentioned in step S5 are arranged in ascending order of center frequency, and preferably IMF1–IMF4 are used to construct the time-frequency joint feature matrix.

[0015] Example 1: Overall Structure of the Intelligent Detector This intelligent detector consists of eight modules connected in series to form a detection link: vibration acquisition module, rotational speed and phase reference acquisition module, synchronous preprocessing module, fault sensitive quantity extraction module, successive variational mode decomposition module, multimodal feature construction module, time-series intelligent recognition module, and fault comprehensive evaluation module.

[0016] The functions of each module are as follows: Vibration acquisition module: Acquires the stator excitation end vibration signal x[n]; Speed ​​phase reference acquisition module: acquires key phase or speed measurement pulse signal r[n]; Synchronous preprocessing module: Synchronously aligns x[n] and r[n], removes abnormal data, and completes discrete spectrum transformation and passband amplitude sequence extraction; Fault Sensitive Quantity Extraction Module: Extracts the 50Hz amplitude A50 and the 100Hz amplitude A100 from the passband spectrum, and calculates the fault index R = A100 / A50; Successive variational mode decomposition module: Performs SVMD on the passband amplitude sequence to adaptively optimize the number of modes K; Multimodal feature construction module: Select fault-related IMFs, and generate a time-frequency joint feature matrix through sliding window segmentation and tensor reconstruction; Temporal intelligent identification module: Input the feature matrix into the LSTM or GRU network and output 12 types of fault labels and their corresponding probabilities; Fault comprehensive assessment module: integrates R value and inter-turn short circuit posterior probability to determine fault level.

[0017] Example 2: Construction of Fault Sensitive Quantities: Inter-turn short circuits in the rotor windings lead to a decrease in the fault-induced ampere-turns. The reverse magnetomotive force is superimposed on the main air gap magnetic field, causing magnetic field asymmetry and flux density distortion. This distorted magnetic field couples with the stator slots, exciting radial electromagnetic force waves dominated by 50Hz and 100Hz, which produce significant vibration responses at the excitation terminals. Studies have shown that low-order pulsating forces account for 65%–85% of the total fault energy, and the amplitude variation of 50Hz / 100Hz is strongly correlated with the number of short-circuited turns. Therefore, the ratio of these two values ​​can serve as a physically interpretable fault sensitivity indicator.

[0018] In this embodiment, the fault sensitivity quantity is defined as: R = A50 / A100 In the formula, A50 and A100 are the amplitudes of the 50Hz and 100Hz components in the vibration spectrum of the measuring point of the excitation bearing housing, respectively.

[0019] Example 3: Successive Variational Mode Decomposition The excitation-end vibration signal exhibits strong non-stationarity, making it difficult for traditional steady-state spectrum methods to effectively extract weak fault features. This embodiment employs SVMD to perform mode decomposition on the passband amplitude sequence. The number of modes K is automatically determined by the algorithm, and the resulting IMFs are sorted from low to high center frequency to support subsequent fault-related mode screening.

[0020] Example 4: IMF Screening and Sliding Window Segmentation In this embodiment, fault-related components are screened from the IMFs obtained from SVMD decomposition. IMFs with center frequencies of 30–240 Hz are selected as candidate fault feature IMFs, such as IMF2–IMF4, with center frequencies of 43.92 Hz, 129.93 Hz, and 207.95 Hz, respectively. IMF1 (approximately 0.01 Hz), which represents a low-frequency trend, is also included, resulting in a total of four input channels: IMF1–IMF4. Higher-order IMFs that are susceptible to high-frequency noise interference are not included in the feature construction.

[0021] The passband amplitude sequence of length N and the selected IMF are segmented using a sliding rectangular window, with window length n = 20 and step size inc = 4, to obtain... There are m equal-length subsequences. This parameter combination can stably generate m training samples for each type of fault in actual tests.

[0022] Example 5: Timing-based intelligent identification and fault level output The temporal intelligent identification module consists of an input layer, a tensor reconstruction layer, an LSTM or GRU recurrent layer, an activation layer, a fully connected layer, and a Softmax layer. The input layer receives the time-frequency joint feature matrix obtained by IMF1–IMF4 through sliding window segmentation and tensor reconstruction. The tensor reconstruction layer unifies the dimensions. The recurrent layer models long-range temporal dependencies. The fully connected layer and the Softmax layer jointly output the posterior probability of fault classification.

[0023] The fault categories include: healthy state; single fault (inter-turn short circuit, bearing wear, shaft imbalance); two types of combined faults (inter-turn short circuit + bearing wear, inter-turn short circuit + shaft imbalance); and six other different levels of single inter-turn short circuit, for a total of 12 categories.

[0024] The fault comprehensive assessment module integrates the R value and the posterior probability of inter-turn short circuit, and determines the fault state and fault level according to the preset threshold.

[0025] Example 6: Online Testing Process like Figure 2 As shown, the online detection method for inter-turn short-circuit faults in the rotor winding of a steam turbine generator includes the following steps: S1: A vibration sensor is placed in the bearing housing at the excitation end, and a key phase or speed sensor is placed at the corresponding position on the rotating shaft; S2: Synchronously acquire vibration signal x[n] and rotational speed phase reference signal r[n]; S3: Synchronize x[n] and r[n] and remove data with timestamp jumps, sampling rate deviations, and abnormal record lengths; S4: Perform discrete spectrum transformation on the synchronous vibration signal, extract the 50Hz amplitude A50 and the 100Hz amplitude A100, and calculate the fault index R = A100 / A50. S5: The SVMD algorithm is used to automatically determine the number of modes K, and the passband amplitude sequence is decomposed into modes to obtain the IMF arranged in ascending order of center frequency; S6: Select IMF1–IMF4, segment and reconstruct them using a sliding rectangular window (window length 20, step size 4) to generate a joint time-frequency feature matrix; S7: Input the feature matrix into an LSTM or GRU network and output the posterior probability of 12 types of faults; S8: Combines the R value with the posterior probability of inter-turn short circuit, determines the fault level according to the preset threshold, and triggers an alarm.

[0026] Example 7: Application Case Taking a 600MW steam turbine generator as the object, an SKVMA vibration tester (sampling frequency 50kHz, FFT points 2048) was used, combined with key phase or velocity measurement signals to achieve synchronous analysis. The measured parameters are as follows: SVMD mode number K = 9; fault-related IMF center frequencies are IMF2 (43.92Hz), IMF3 (129.93Hz), and IMF4 (207.95Hz); the neural network structure includes a sequence input layer, tensor reconstruction layer, LSTM or GRU layer, ReLU activation layer, fully connected layer, softmax layer, and classification layer.

[0027] This scheme is stable in identifying complex faults and low signal-to-noise ratio conditions. Comparative experiments show that the improved LSTM / GRU model based on IMF1–IMF4 input has significant improvements in noise resistance and fault classification accuracy compared to the conventional LSTM / GRU.

[0028] In summary, the intelligent detector for inter-turn short circuit faults in the rotor winding of the steam turbine generator proposes a physically interpretable fault sensitivity quantity R = A50 / A100, which directly reflects the air gap magnetic field imbalance caused by the inter-turn short circuit and its 50Hz / 100Hz electromagnetic force wave vibration response. The SVMD algorithm is used to adaptively optimize the number of modes K, which effectively suppresses mode mixing, enhances the ability to retain weak fault features, and improves the early inter-turn short circuit detection rate. Focusing on fault-related IMFs in the 30–240Hz frequency band, and combining a sliding rectangular window to reconstruct the time-frequency joint feature matrix, the accuracy of composite fault identification and its robustness against noise are significantly improved. End-to-end fault classification and severity assessment based on improved LSTM or GRU networks have stronger time-series modeling capabilities, noise suppression capabilities, and field applicability compared to traditional methods that rely on single-domain indicators.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator, characterized in that, include: The vibration acquisition module is used to acquire vibration signals from the stator excitation end. The rotational speed phase reference acquisition module is used to acquire key phase or speed measurement pulse signals that are synchronized with the rotor speed. The synchronous preprocessing module is used to perform synchronous alignment, outlier removal, FFT transformation, and 40–1000 Hz passband amplitude sequence extraction on the vibration signal and the rotation speed signal. The fault sensitivity extraction module is used to extract the 50 Hz amplitude A50 and the 100 Hz amplitude A100 from the spectrum and construct the fault sensitivity index R = A50 / A100, wherein the R index reflects the 50 Hz / 100 Hz electromagnetic force wave vibration response ratio caused by the asymmetry of the air gap magnetic field due to the inter-turn short circuit. The SVMD module is used to adaptively optimize the number of modes K with the goal of minimizing the modal mixing energy, and perform VMD decomposition on the passband amplitude sequence to obtain multiple intrinsic mode functions (IMFs). The multimodal feature construction module is used to select fault-related IMFs, segment them through a sliding window, and then fuse time-frequency features to construct a joint time-frequency feature matrix. The time-series intelligent recognition module is used to input the feature matrix into an LSTM or GRU network and output the health status, fault status, fault level and confidence probability. The fault comprehensive assessment module is used to output fault status and inter-turn short circuit fault level information based on health / fault status estimation results, confidence probability and preset thresholds; The R index construction, SVMD module, and time-series intelligent recognition module work together to improve the recognition accuracy in low signal-to-noise ratio and complex fault scenarios.

2. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: The vibration acquisition module is deployed at least at one of the following locations: the side of the excitation bearing seat, the upper part of the casing, or the lower part of the casing, with the side of the excitation bearing seat as the main measuring point.

3. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: After the SVMD module adaptively decomposes the passband amplitude sequence into multiple IMFs, the multimodal feature construction module selects IMFs with center frequencies in the range of 30–240 Hz as fault-related candidate modes and arranges them in ascending order of center frequency.

4. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: The candidate fault mode components include IMF1, IMF2, IMF3 and IMF4, where IMF1 represents the low-frequency vibration trend, and IMF2–IMF4 are located in the 30–240 Hz fault correlation frequency band, which are used to construct the input matrix of the neural network.

5. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: The multimodal feature construction module uses a sliding rectangular window to segment the passband amplitude sequence of length N and the selected IMFs: setting the window length to n and the step size to inc, resulting in m equal-length subsequences, where The window length n is an integer from 16 to 24, and the step size inc is an integer of 2. The preferred scheme is n = 20 and inc = 4.

6. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: The time-series intelligent recognition module sequentially includes an input layer, a tensor reconstruction layer, an LSTM or GRU layer, an activation layer, a fully connected layer, a Softmax layer, and a classification layer.

7. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: The categories output by the classification layer include at least: health status, six different levels of rotor winding inter-turn short circuit, bearing wear, shaft dynamic imbalance, and two types of composite faults—inter-turn short circuit + bearing wear, and inter-turn short circuit + shaft dynamic imbalance.

8. The intelligent detector for inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 1, characterized in that: The fault comprehensive assessment module is specifically configured to: jointly determine the inter-turn short circuit fault state and its specific level based on the fault sensitivity quantity R and the inter-turn short circuit posterior probability output by the Softmax layer.

9. A method for detecting inter-turn short-circuit faults in the rotor winding of a steam turbine generator, applied to the intelligent detector described in any one of claims 1 to 8, characterized in that, Includes the following steps: S1: Acquire stator excitation end vibration signal and synchronous speed phase reference signal; S2: Synchronize and align the two signals, remove abnormal data, perform discrete spectrum transformation, and extract the 40–1000 Hz passband amplitude sequence; S3: Extract the 50 Hz amplitude from the vibration spectrum and 100 Hz amplitude Construct the fault sensitivity quantity R = A50 / A100; S4: With the goal of minimizing the energy of modal mixing, SVMD is used to adaptively optimize the number of modes K, and modal decomposition is performed on the passband amplitude sequence to obtain multiple IMFs; S5: Select IMFs with center frequencies in the range of 30–240 Hz as fault-related mode components, and construct a time-frequency joint feature matrix through sliding window segmentation and tensor reconstruction; S6: Input the feature matrix into an LSTM or GRU network and output the fault category and recognition probability; S7: Integrates R index, identification probability and preset criteria to determine fault status and inter-turn short circuit level.

10. The method for detecting inter-turn short-circuit faults in the rotor winding of a steam turbine generator according to claim 9, characterized in that, The abnormal data removal mentioned in step S2 specifically targets data with timestamp jumps, sampling rate deviations, and record length deviations exceeding the limits; the IMFs mentioned in step S5 are arranged in ascending order of center frequency, and preferably IMF1–IMF4 are used to construct the time-frequency joint feature matrix.

Images