Generator stator slot wedge online detection method and system based on acousto-optic monitoring fusion

Abstract

The invention belongs to the technical field of stator slot wedge detection, and discloses a generator stator slot wedge online detection method and system based on acousto-optic monitoring fusion. The method comprises the following steps: acquiring laser ranging data, acoustic monitoring data and generator working condition parameter data of online looseness detection of a generator stator slot wedge; preprocessing the laser ranging data, the acoustic monitoring data and the generator working condition parameter data to obtain laser ranging features, acoustic features and working condition parameters; and inputting the optical ranging features, the acoustic features and the working condition parameters into a pre-trained auto-encoder for anomaly detection, and obtaining an online looseness detection result of the generator stator slot wedge. By fusing the optical ranging technology, the optical fiber acoustic monitoring technology and the unsupervised deep learning algorithm, the important technical breakthrough in the field of generator stator slot wedge looseness detection is realized, and online stator slot wedge looseness detection can be carried out.

Description

Technical Field

[0001] This invention belongs to the field of stator slot wedge detection technology, and specifically relates to an online detection method and system for generator stator slot wedges based on acoustic-optical monitoring fusion. Background Technology

[0002] The problem of stator slot wedge loosening: The stator winding of a generator consists of bar-shaped wires installed inside the core. These bars are made of insulated strands of wire that are braided and gelled before being wrapped around the main insulation. During generator operation, under the combined effects of mechanical vibration and electromagnetic force, the stator bars are at risk of loosening and shifting, which may lead to insulation wear, shorten insulation life, and increase the risk of accidents during generator operation. Stator slot wedges, along with corrugated plates and other structures, can fix the bars in the stator slots, enhancing the stability of the internal structure. However, during long-term operation, the repeated application of various influencing factors such as mechanical stress, electromagnetic force, and thermal stress leads to material aging, changes in the initial clamping force within the slots, and a decrease in tightness. Creep and wear may occur in the fasteners and bars within the slots. If slot loosening is not diagnosed and addressed in a timely manner, serious safety problems may occur, such as slot wedges loosening and detaching from the stator slot opening, or the bars being scratched by severe vibrations. At the same time, after the slot wedge loosens, it will generate severe friction with the stator. The powder formed by the friction is easy to adhere to the end of the generator winding, increasing the risk of motor corona.

[0003] Existing monitoring methods: Traditional slot wedge tightness detection involves manually tapping the slot wedges and judging their tightness based on the sound heard. This method is highly subjective and cannot retain historical data. With the development of computer and information processing technology, the principles for stator slot wedge tightness detection, both domestically and internationally, are basically the same: applying pressure or tapping the slot wedges to detect displacement and / or vibration amplitude. While the specific implementation methods differ slightly, these methods rely on electronic equipment, require a high level of operator skill, and may not have sufficient space at generator maintenance sites to use the corresponding electronic equipment. Furthermore, they require shutdown for testing, resulting in low efficiency, high maintenance costs, and low accuracy. Additionally, all of the above methods are contact measurements, which may interfere with the normal operation of the generator.

[0004] Application of Deep Learning in Fault Diagnosis: In the alternating electromagnetic field of a generator, the stator bars vibrate differently due to varying forces. Therefore, the sound signals generated during the tapping of the slot wedges differ due to variations in their internal structure. Vibration and sound are closely related. Sound originates from vibration and contains rich parameter information from sample data, going beyond the basic time and frequency domain characteristics of the signal. Characteristic parameters such as Mel-frequency cepstral coefficients, discrete wavelet transform, line spectrum frequency, and linear predictive cepstral coefficients can be extracted and used as a basis for judging slot wedge loosening faults. The paper "Application Research of Machine Learning Algorithm in the Detection of Stator Slot Wedge Tightness in Large Generators" proposes a fault detection method for generator slot wedge tightness based on machine learning algorithms. By tapping the slot wedges and collecting the feedback sound signals, the tightness of the generator stator slot wedges is judged based on the differences in signal decay time, peak frequency, and 2-4kHz frequency band area under different conditions. However, this method requires an additional device for tapping and does not consider the influence of different generator operating conditions on the stator slot wedge vibration. Meanwhile, with the widespread application of neural networks and computer technology, fault diagnosis can effectively identify the type of fault by extracting feature vectors of different fault types and training models based on sample data. Among these, expert system methods, extreme learning machines, backpropagation neural networks, simulated annealing algorithms, support vector machines, and hidden Markov chains are widely used in the fault diagnosis of mechanical products.

[0005] The main shortcomings of existing detection methods are as follows: 1. Testing Location and Timing: Pressure or tapping is required to analyze vibration and sound signals from the slot wedges. However, these methods all require contact with the slot wedges, necessitating the testing equipment to move extensively across the entire inner surface of the stator core. To ensure energy conversion efficiency, generators have a narrow air gap between the stator and rotor, making it difficult for testing equipment to access the space. During motor operation, the high-speed rotation of the rotor and any foreign objects in the air gap can cause serious accidents such as rotor rubbing. Furthermore, the stator slots contain high currents and strong magnetic fields, which can easily interfere with automation components. Therefore, existing testing methods can only be performed when the motor is stopped, or even after the rotor has been removed, making long-term online monitoring and fault prevention impossible.

[0006] 2. Types of Detected Parameters: Existing research methods do not consider the influence of generator operating conditions on stator slot wedges when analyzing sound and vibration signals. Under different generator operating conditions, there are differences in hydraulic vibration in the water inlet pipe, generator rotor rotation, generator current parameters, and temperature parameters, resulting in significant differences in the stress conditions within the slots, inconsistent vibration of the slot wedges and rods, and differences in sound signals.

[0007] 3. Data Analysis Methods: Existing technologies lack multimodal information fusion mechanisms. Single sensor modes are easily affected by environmental interference, resulting in insufficient detection reliability. Optical measurements are susceptible to dust, and acoustic measurements are easily affected by environmental noise. Using any one method alone has limitations.

[0008] Existing technologies for detecting stator slot wedge loosening in generators are hampered by the presence of high current and strong magnetic fields during generator operation, which hinders online monitoring. Furthermore, existing contact-based detection methods cannot perform continuous monitoring during normal unit operation and pose risks of mechanical damage and electromagnetic interference. This makes it impossible to monitor key performance parameters such as stator slot wedge tightness in real time, and to identify potential defects in a timely manner. There is a risk that severe deterioration of defects may lead to sudden failures and unplanned unit shutdowns. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for online detection of stator slot wedges in generators based on acoustic-optical monitoring fusion, so as to solve the technical problem that existing technologies cannot achieve online detection of stator slot wedge loosening.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising: Acquire laser ranging data, acoustic monitoring data, and generator operating parameter data for online detection of stator slot wedge loosening in generators; Preprocessing is performed on laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters; The optical ranging features, acoustic features, and operating parameters are input into a pre-trained self-encoder for anomaly detection, thereby obtaining the online loosening detection results of the generator stator slot wedges.

[0011] A further improvement of the present invention is that: in the step of obtaining laser ranging data, acoustic monitoring data, and generator operating condition parameter data for online loosening detection of generator stator slot wedges, obtaining laser ranging data for online loosening detection of generator stator slot wedges specifically includes: obtaining laser ranging data by using a laser ranging module arranged on or beside the end of the generator stator bar using the phase-type laser ranging principle or the beam-splitting interferometric laser measurement. Acoustic monitoring data for online loosening detection of generator stator slot wedges specifically includes: using fiber optic acoustic emission sensors arranged on the side frame or inside the stator bar at the end of the generator stator bar to collect acoustic emission signals of the stator bar and obtain acoustic monitoring data; The specific operating condition parameters include: based on the Pearson correlation coefficient, five operating condition parameters are selected from speed, stator core vibration, motor start-up time, continuous running time, stator slot temperature, stator core temperature, generator active power, generator reactive power, stator current, stator voltage, and power factor to reflect the influence of mechanical stress, thermal stress, and electromagnetic force on slot wedge vibration.

[0012] A further improvement of the present invention is that the step of preprocessing the laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging features, acoustic features, and operating parameters specifically includes: extracting features from both the laser ranging data and the acoustic monitoring data to obtain five features: root mean square value, peak factor, kurtosis, spectral centroid, and spectral variance. The five features of laser ranging data, the five features of acoustic monitoring data, and the five data interpolations of operating condition parameters are standardized with a unified sampling rate. Data alignment and standardization are then performed to obtain preprocessed laser ranging features, acoustic features, and operating condition parameters.

[0013] A further improvement of the present invention is that: in the step of preprocessing the laser ranging data, acoustic monitoring data, and generator operating condition parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating condition parameters, the preprocessing includes removing outliers and interpolation completion.

[0014] A further improvement of the present invention is that: in the step of inputting the optical ranging features, acoustic features and operating condition parameters into a pre-trained self-encoder for anomaly detection and obtaining the online loosening detection result of the generator stator slot wedge, the self-encoder includes an input layer, an encoder, a latent layer, a decoder and an output layer; The input layer has 15 neurons; The encoder consists of three fully connected layers with 32, 16 and 8 neurons respectively. Each fully connected layer is followed by a ReLU activation function and a batch normalization layer. The middle potential layer contains 4 neurons; The decoder adopts a completely symmetrical structure with the encoder. The number of neurons in the three fully connected layers and the output layer are 8, 16, 32 and 15, respectively. The output layer uses a linear activation function.

[0015] A further improvement of this invention is that: in the step of inputting the optical ranging features, acoustic features, and operating parameters into a pre-trained self-encoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the pre-trained self-encoder is obtained through the following steps: The autoencoder is trained using the standard backpropagation algorithm and the Adam optimizer. The learning rate can be set between 0.0001 and 0.01; in this embodiment, it is 0.001. The batch size can be set between 32 and 512 depending on the hardware conditions; in this embodiment, it is 256. An early stopping mechanism is used during training. Training automatically stops when the reconstruction error on the validation set does not improve for N consecutive epochs (N can be set to 10-50; in this embodiment, N=20) to avoid overfitting.

[0016] A further improvement of the present invention is that, in the step of inputting the optical ranging features, acoustic features and operating parameters into a pre-trained self-encoder for anomaly detection and obtaining the online loosening detection result of the generator stator slot wedge, the feature vectors of 10 consecutive time points are combined into a 150-dimensional input vector and input into the pre-trained self-encoder for anomaly detection.

[0017] A further improvement of this invention is that, in the step of inputting the optical ranging features, acoustic features, and operating parameters into a pre-trained autoencoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the optical ranging features, acoustic features, and operating parameters are input into the pre-trained autoencoder to obtain a reconstruction feature vector, the reconstruction error is calculated based on the reconstruction feature vector, the reconstruction error is compared with an anomaly detection threshold, and an alarm is triggered when the reconstruction error at M consecutive time points exceeds the anomaly detection threshold, thereby obtaining the online loosening detection result of the generator stator slot wedge, where M is a positive integer greater than or equal to 1; the anomaly detection threshold is determined using the 3σ principle.

[0018] A further improvement of the present invention is that it also includes: the pre-trained autoencoder updates the model parameters using an incremental learning method, specifically: continuously collecting operating data that has been confirmed to be in a normal state, triggering model updates when new data reaches a preset scale, using a fine-tuned learning rate smaller than the initial training learning rate and fewer training rounds to update parameters, and retaining historical model versions to support rollback, so as to adapt to the state evolution during the long-term operation of the generator.

[0019] Secondly, the present invention provides an online detection system for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising: The acquisition module is used to acquire laser ranging data, acoustic monitoring data, and generator operating parameter data for online loosening detection of generator stator slot wedges; The preprocessing module is used to preprocess laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters. The detection module is used to input the optical ranging features, acoustic features and operating parameters into a pre-trained self-encoder for anomaly detection, and obtain the online loosening detection results of the generator stator slot wedge.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising: acquiring laser ranging data, acoustic monitoring data, and generator operating parameter data for online stator slot wedge loosening detection; preprocessing the laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging features, acoustic features, and operating parameters; and inputting the laser ranging features, acoustic features, and operating parameters into a pre-trained autoencoder for anomaly detection to obtain the online stator slot wedge loosening detection result. This invention achieves a significant technological breakthrough in the field of generator stator slot wedge loosening detection by fusing optical ranging technology, fiber optic acoustic monitoring technology, and unsupervised deep learning algorithms, enabling online stator slot wedge loosening detection.

[0021] This invention changes the traditional technical mode of slot wedge detection, which requires machine shutdown and contact-based live measurement. By adopting phase-type laser ranging, beam-splitting interferometric laser ranging, and fiber optic acoustic monitoring technologies, the system achieves micron-level precision measurement of stator bar radial vibration and stator slot wedge axial vibration displacement, as well as accurate acquisition of stator bar vibration acoustic signals. The four configuration schemes of the laser ranging module fully consider the structural characteristics and installation environment of different generators: The lower baffle rail mounting method achieves multi-point inspection through a circular slide rail, with a single ranging module monitoring the radial displacement of all upper-layer bars by scanning each slot while moving; the lower baffle multi-laser emission method achieves simultaneous monitoring of multiple targets at a fixed point, using ranging modules arranged at multiple points and in multiple directions at a single point to statically measure the radial position of all upper-layer bars; the stator pit wall rail mounting method has a similar structure to the lower baffle rail mounting method, but the rail is located on the pit wall, monitoring the radial displacement of the stator bars; the laser ranging method combines the stator slot outlets at both ends and the straight sections at the ends of the bars with the baffle fixing structure, allowing for installation in multiple slots or even per slot, using fiber optic transmission and miniature probes to measure the axial vibration of the stator slot wedge in confined spaces. This diverse configuration strategy ensures the universality of the technical solution for various generators. Using the slide rail laser monitoring method and the single-point multi-laser monitoring method can reduce the number of laser ranging modules, thus saving costs. When optical monitoring is interfered with, acoustic monitoring probes can be installed on some of the rods to achieve cross-verification of optical and acoustic monitoring.

[0022] The technical advantages of this invention's non-contact measurement not only lie in avoiding interference with generator operation, but more importantly, in eliminating the safety hazards of traditional contact-based detection in high-voltage electrical environments. The optical measurement principle ensures the system is unaffected by the strong electromagnetic fields during generator operation, fundamentally guaranteeing measurement accuracy and stability. The system achieves 1kHz high-frequency vibration data acquisition, capable of capturing the transient characteristics and subtle changes in slot wedge and rod vibrations, providing a technical foundation for early fault warning. The acoustic monitoring module employs an ultra-high sampling rate of 100kHz, capable of capturing high-frequency acoustic emission signals generated by slot wedge loosening, which are often early signs of mechanical faults. The combination of these two modes enables the system to possess full-frequency monitoring capabilities, from low-frequency vibration to high-frequency acoustics.

[0023] This invention deeply integrates optical ranging and acoustic monitoring, and designs a symmetrical feature extraction framework. The same feature extraction method is used for both laser ranging and acoustic signals, extracting five features: root mean square value, peak factor, kurtosis, spectral centroid, and spectral variance, achieving a unified representation of the feature space. This symmetrical design has profound technical significance: First, the unified feature space allows deep learning models to process signals with two different physical properties within the same mathematical framework, avoiding the complexity of heterogeneous feature fusion; second, the identical feature dimensions and physical meanings facilitate the calculation of the correlation and complementarity between the two modes, providing a foundation for adaptive fusion; third, symmetrical feature extraction reduces computational complexity and improves real-time processing efficiency.

[0024] Furthermore, the root mean square (RMS) value reflects the overall level of vibration and acoustic energy, serving as a fundamental indicator for assessing the wedge's condition. The crazing factor identifies the impact component in the signal; intermittent collisions after wedge loosening lead to a significant increase in the crazing factor. Kurtosis, as a fourth-order statistic, is extremely sensitive to the spike characteristics of signal distribution and is an important indicator for detecting early anomalies. The spectral centroid reflects the center of energy distribution in the frequency domain; normally, a tightly secured wedge's vibration energy is concentrated in a specific frequency band, while loosening causes the spectral centroid to shift. The spectral variance measures the dispersion of frequency components; multimodal vibrations caused by loosening increase the spectral variance. These five features comprehensively characterize the wedge's condition from different perspectives, achieving deep fusion of optical and acoustic information through symmetrical extraction.

[0025] This invention achieves true multi-condition adaptive anomaly detection through a deep autoencoder network. Based on the Pearson correlation coefficient, five parameters are selected from stator core vibration, motor start-up time, continuous running time, stator slot temperature, stator core temperature, generator active power, generator reactive power, stator current, stator voltage, power factor, and rotational speed to reflect the interference of mechanical stress, thermal stress, and electromagnetic force on slot wedge vibration. These parameters, together with vibration and acoustic features, constitute a 15-dimensional feature vector. This design enables the deep learning model to learn the complex nonlinear relationship between operating parameters and the normal vibration mode of the slot wedge.

[0026] The normal vibration characteristics of slot wedges vary under different operating conditions. Stator core vibration mainly originates from turbine vibration and rotor rotation, and is the primary source of mechanical stress in the slot section. Motor start-up time, continuous running time, stator slot temperature, and stator core temperature affect thermal stress in the slot section, altering the fixation state and vibration characteristics of the slot wedges. Generator active power, generator reactive power, stator current, stator voltage, and power factor directly affect electromagnetic force, thus influencing the stress state of the slot wedges. Speed ​​changes alter the excitation frequency; therefore, the vibration changes of components during start-up and shutdown need to be analyzed in conjunction with the speed at that moment and cannot be directly compared with stable operating vibration signals. Therefore, speed and stator core vibration can be selected to reflect the impact of mechanical stress in the slot section; one or more of the following can be selected to reflect the impact of thermal stress: motor start-up time, continuous running time, stator slot temperature, and stator core temperature; and one or more of the following can be selected to reflect the impact of electromagnetic force: generator active power, generator reactive power, stator current, stator voltage, and power factor. Traditional fixed threshold detection methods cannot adapt to such complex operating condition changes and are prone to false alarms or missed alarms. This system establishes a dynamic correlation model between operating conditions, vibration, and acoustics by learning from a large amount of normal operating data covering various operating condition combinations. When the system receives real-time monitoring data, it first predicts the range of vibration and acoustic characteristics under normal conditions based on the current operating condition parameters. Then, it compares the actual measured values ​​with the predicted values, and only when the deviation exceeds a statistically significant level is it judged as abnormal. This adaptive mechanism fundamentally solves the problem of the impact of operating condition changes on detection accuracy. Different units can select different operating condition combinations. Before real-time online monitoring, the system executes an operating condition parameter optimization process based on historical data. This process first classifies the candidate operating condition parameters into three parameter categories according to their physical meaning: mechanical stress, thermal stress, and electromagnetic force. Subsequently, the system collects a large amount of historical data of the unit under normal operating conditions and calculates the Pearson correlation coefficient between each candidate operating condition parameter and all 10 acoustic and optical monitoring features (5 laser ranging features and 5 acoustic features). By taking the absolute value of the correlation coefficient and averaging it, a comprehensive importance score is calculated for each parameter. Finally, the parameter with the highest score from each of the three pools—mechanical stress, thermal stress, and electromagnetic force—is selected to ensure comprehensive coverage of the effects of all three stresses, for a total of three parameters. Then, from all remaining candidate parameters, the two parameters with the highest overall importance score are selected. The five operating parameters determined in this way are considered the combination with the strongest correlation to the slot wedge state of this specific unit and are used for subsequent online monitoring and deep learning model analysis, thereby significantly improving the targeting and accuracy of the detection.

[0027] This invention employs an unsupervised deep learning method, requiring only historical data on the generator's normal operating status to complete model training, thus eliminating the reliance on labeled outlier samples. This innovation brings significant engineering application value. In practical engineering, obtaining a large number of labeled slot wedge loosening samples is extremely difficult, requiring not only professional judgment but also potentially shutdown for inspection and verification, resulting in high costs. Normal operating data, however, can accumulate naturally during the generator's daily operation, eliminating the need for additional data collection costs.

[0028] Deep autoencoders learn by compressing input data into a low-dimensional latent space and then reconstructing it, automatically discovering the inherent structure and distribution patterns of normal data. The encoder progressively compresses 15-dimensional input features into a 4-dimensional latent representation, capturing the essential features of the data; the decoder reconstructs the latent representation back into the original feature space. Models trained on normal data can accurately reconstruct normal samples, but they produce significant errors in reconstructing anomalous samples. By monitoring the reconstruction error and using the 3σ criterion to determine the anomaly detection threshold, the system can identify anomalous states with a 99.7% confidence level.

[0029] This invention also possesses online incremental learning capabilities, enabling it to continuously collect new normal operation data and update model parameters. This continuous optimization mechanism allows the system to adapt to slow changes during long-term generator operation, such as equipment aging, material aging, and changes in condition after maintenance, ensuring the long-term stability of detection accuracy. Model updates employ a small learning rate and a limited number of training rounds, gradually adapting to new normal patterns while maintaining existing knowledge, thus avoiding catastrophic forgetting problems.

[0030] By integrating three core technological advantages—non-contact real-time monitoring, multi-condition adaptive diagnosis, and label-free autonomous learning—this patented technical solution provides a complete, advanced, and practical solution for detecting stator slot wedge loosening in generators, with broad engineering application prospects and significant economic and social benefits. Attached Figure Description

[0031] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the installation of the lower windshield rail in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-laser ranging method for the lower windshield in an embodiment of the present invention; Figure 4This is a schematic diagram of the stator frame track installation method in an embodiment of the present invention; Figure 5 This is a schematic diagram of the internal installation in an embodiment of the present invention; Figure 6 This is a schematic diagram of the installation method of the acoustic monitoring module in an embodiment of the present invention, showing that it is installed inside the wire rod; Figure 7 This is a schematic diagram of the installation method of the acoustic monitoring module in an embodiment of the present invention, showing the installation method of the wire rod end next to the rack; Figure 8 This is a schematic diagram of the process of an encoder performing anomaly detection in an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of the structure of the self-encoder in this embodiment; Figure 10 A flowchart illustrating an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, as provided in another embodiment of the present invention; Figure 11 This is a schematic diagram of the software interface of an online detection system for generator stator slot wedges based on acoustic-optical monitoring fusion, provided in an embodiment of the present invention. Figure 12 This is a structural block diagram of an online detection system for generator stator slot wedges based on acoustic-optical monitoring fusion, provided in an embodiment of the present invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0033] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0034] Terminology Explanation: Stator slot wedge: A fixing component installed in the stator core slot of the generator to firmly fix the stator bars in the slot, preventing vibration and displacement of the bars. It is a key component to ensure the safe operation of the generator.

[0035] Phase-based laser ranging: a high-precision distance measurement technology based on the principle of laser phase detection. It calculates the target distance by high-frequency modulation of the laser and measuring the phase difference between the emitted and reflected light, and has micron-level measurement accuracy.

[0036] Acoustic emission monitoring: a non-destructive testing technology that uses high-sensitivity fiber optic acoustic emission sensors arranged at the ends of generator stator bars or in ventilation channels at the bottom of slots to collect transient elastic wave signals generated by dynamic processes such as friction, collision or minute movement between components.

[0037] Unsupervised deep learning: A machine learning method that uses only normal state data for training, without the need for manual labeling of abnormal samples. It identifies abnormal states that deviate from the normal pattern by learning the inherent rules and distribution characteristics of normal data.

[0038] Multi-condition adaptive: An intelligent mechanism that automatically adjusts detection parameters and judgment criteria according to different operating conditions of the generator (mechanical stress, thermal stress, electromagnetic force), eliminates the influence of operating condition changes on normal vibration modes, and improves the accuracy of anomaly detection.

[0039] Symmetric feature extraction: The same feature extraction method is used for sensor data of different modalities, including time-domain statistical features (root mean square value, peak factor, kurtosis) and frequency-domain distribution features (spectral centroid, spectral variance), to achieve a unified representation of the feature space.

[0040] This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion. Through deep integration of fiber-optic laser ranging technology and fiber-optic acoustic monitoring technology, a unified feature extraction framework is used to extract five features—root mean square value, peak factor, kurtosis, spectral centroid, and spectral variance—from both modes of signal, achieving symmetrical representation of the feature space. Combined with deep learning algorithms, the method automatically learns normal characteristic patterns under different mechanical stresses, thermal stresses, and electromagnetic forces, comparing measured values ​​with predicted ranges in real time. If a threshold is exceeded, a loosening is determined, enabling non-contact continuous monitoring and real-time alarm during normal generator operation.

[0041] The inventors discovered that when the generator is operating normally and without any additional external force, if the slot wedges and corrugated plates used to fix the vertically placed stator bars become loose, the stator bars will vibrate in both the radial and axial directions. This vibration is fundamentally different from the minute vibrations under normal operating conditions.

[0042] Under normal circumstances, the stator bars are tightly fixed within the slots of the stator core. The fixing system mainly consists of stator slots, corrugated plates, and slot wedges. The corrugated plate, a flexible insulating material, is typically placed between the bars and the slot wedges. Its main function is to provide a continuous, radial preload, pressing the bars firmly into the slots. This elastic structure compensates for dimensional changes caused by thermal expansion and contraction and dampens minor vibrations of the bars. The slot wedges, located at the stator slot openings, are the key components that ultimately lock the entire fixing system in place. They act as support components, providing preload through corrugated plate compression, securing the corrugated plates and bars and preventing them from moving radially.

[0043] Under normal generator operation, an alternating current flows through the stator bars. This current interacts with the radial magnetic field in the stator slot, generating a periodic electromagnetic force according to the law of electromagnetic induction. This electromagnetic force mainly includes: (1) Radial electromagnetic force: generated by the interaction between the bar current and the air gap magnetic field, pointing towards the bottom of the slot, with a frequency twice that of the power grid (100Hz for a 50Hz power grid). Since both the stator current and the magnetic field are power frequency alternating quantities, their multiplication produces a DC component and a harmonic component, in which the harmonic component generates a periodic excitation on the bar; (2) Harmonic electromagnetic force: generated by the high-order harmonics (such as the 5th and 7th harmonics) present in the current and magnetic field, producing electromagnetic force components of corresponding frequencies; (3) Axial electromagnetic force: generated by the interaction between the leakage magnetic field in the end region and the bar current, applying a periodic force along the bar axis. Under normal tightening conditions, the slot wedges and corrugated plates provide sufficient preload, forming a high-rigidity constraint structure and large frictional damping between the bar and the slot wall. Under this constraint, although the bar is continuously excited by periodic electromagnetic force, its vibration response is effectively suppressed within the micrometer range, exhibiting steady-state forced vibration with extremely small amplitude.

[0044] When the slot wedge and corrugated plate loosen, the preload of the fixing system decreases. Under the continuous action of the periodic electromagnetic force, the vibration mode of the bar changes from small-amplitude linear forced vibration in the tightened state to large-amplitude nonlinear forced vibration, accompanied by collision and friction effects. Radial vibration intensifies significantly, and the vibration amplitude increases. The bar, originally compressed by the preload, begins to produce obvious reciprocating motion within the radial gap under the action of the 100Hz periodic electromagnetic force, i.e., vibration. The greater the loosening, the greater the vibration amplitude. This vibration is no longer a micro-vibration, but a mechanical impact with considerable energy. It may also cause axial vibration. Due to the incomplete asymmetry of the structure, assembly differences, and the influence of radial vibration itself, a component force along the axial direction (vertical direction) of the bar will be generated. In the tightened state, this axial force is effectively limited by friction, but when radial loosening occurs, the radial vibration of the bar weakens the normal pressure between it and the slot wall, thus significantly reducing the frictional force that prevents its axial movement. Under the influence of the electromagnetic force, the bar vibrates along the axial direction.

[0045] Under normal, tightened conditions, the vibration of the bar is a steady-state forced vibration with amplitudes in the micrometer range. The time-domain signal is quasi-periodic, and the frequency-domain energy is concentrated at the 100Hz fundamental frequency and its lower harmonics (200Hz, 300Hz, etc.), exhibiting typical response characteristics of a linear system. However, in a loose state, the gap between the bar and the groove wall leads to nonlinear vibration. The bar periodically collides with the groove wall within the gap. This impact is highly nonlinear and produces two significant characteristics in the signal: pulse spikes in the time domain (each collision superimposes a steep, instantaneous high-amplitude pulse onto the stable vibration signal) and high-frequency excitation in the frequency domain (the impact theoretically excites a very wide frequency range, resulting in a large number of high-frequency components in the vibration and sound signals).

[0046] The specific mathematical model is as follows: (1) The elastic properties of the corrugated plate can be modeled as a nonlinear spring:

[0047] in, Linear stiffness coefficient; : Nonlinear stiffness coefficient; The restoring force provided by the corrugated plate at a displacement of x; pre-compression: In order to achieve preload Required initial compression displacement; functions provided by the corrugated plate: continuous preload, vibration damping, and thermal compensation.

[0048] (2) Detailed modeling of electromagnetic force: Radial electromagnetic force: Radial force generated by the interaction between current and magnetic field in stator bars:

[0049] in: : The transverse magnetic field strength inside the slot; : Amplitude of current in the conductor; : Grid angular frequency (50Hz corresponds to 314rad / s); : Effective length of the bar; Key feature: Electromagnetic force frequency is twice the power grid frequency (100Hz).

[0050] Axial electromagnetic force component: Axial force caused by end leakage magnetic field:

[0051] in: End leakage magnetic field; : End winding length; The angle between the leakage magnetic field and the rod; (3) Radial vibration equation Under the fixed condition, the vibration equation of the wire rod is:

[0052] The total stiffness is:

[0053] Total damping:

[0054] in: The support stiffness provided by the trench walls; The clamping stiffness of the groove wedge; : The elastic stiffness of the corrugated plate; m is the total mass of the wire rod; These are material damping, corrugated plate damping, and friction damping, respectively. The total system damping is affected by operating parameters.

[0055] (4) Steady-state response: Steady-state vibration solution:

[0056] amplitude:

[0057] Radial displacement refers to the radial (perpendicular to the length of the rod) displacement of the rod relative to its equilibrium position at time t. The static offset is caused by the constant pressure generated by the electromagnetic force, which is the DC component. The amplitude of the electromagnetic force is determined by the generator current and magnetic field, and is affected by operating parameters.

[0058] Key results: In the fixed state, the amplitude is extremely small (micrometer level), exhibiting a stable sinusoidal waveform. The amplitude is directly proportional to the electromagnetic force and inversely proportional to the restoring force (determined by the total stiffness coefficient) and the damping force (determined by the total damping coefficient). During generator operation, heat is generated, causing thermal expansion and contraction of the metal and insulating materials. This change alters the clamping force between the slot wedge and the conductor (i.e., in the fixed state), as well as the stiffness and damping coefficients, thus affecting its vibration characteristics.

[0059] (5) Nonlinear vibration model in loose state: Gap nonlinear model: Radial gap caused by loosening This results in piecewise linear restoring force:

[0060] When the bar moves inside the gap δ, the restoring force is zero; once it hits the groove wall, it will generate a large impact stiffness. This collision is the source of high-frequency signals and impact characteristics.

[0061] Collision-friction coupling model: Contact force during collision:

[0062] Friction:

[0063] The contact force, the instantaneous force generated when the bar collides with the groove wall, is highly nonlinear and is the main source of impact signals and high-frequency vibrations. The Hertzian contact stiffness is a constant determined by the material elasticity and shape of the colliding bodies (wire bar and channel wall). The depth of deformation caused by the mutual compression of the two objects at the moment of collision; The collision damping coefficient mainly reflects the energy loss during the collision process; The instantaneous radial velocity of the bars upon collision; Let be the coefficient of friction, normal force, and frictional force. The two equations above describe the change in axial frictional force during the collision process. The contact force is based on Hertzian contact theory and simulates the nonlinear force of the collision; the frictional force describes the resistance encountered during axial motion.

[0064] (6) Complete nonlinear vibration equation Radial vibration:

[0065] Axial vibration (caused by radial loosening):

[0066] Coupling terms:

[0067] in This refers to the minute damping experienced by the rod as it moves through the air; To describe the axial component force that may be generated by radial motion due to structural asymmetry and other factors, the vibration of the bar is no longer a simple, stable sine wave dominated by 100Hz. Instead, it transforms into a large-amplitude nonlinear vibration accompanied by periodic collisions and friction. When the bar moves through the gap and impacts the groove wall, a steep, instantaneous high-amplitude pulse with a wide frequency range is superimposed on the stable vibration signal. At the same time, radial loosening and collisions weaken the normal pressure between the bar and the groove wall, thereby significantly reducing the frictional force that prevents its axial motion, making the bar more prone to axial vibration under the action of axial electromagnetic force.

[0068] This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion. It utilizes a laser ranging module to employ phase-based laser ranging for long-distance radial vibration measurement and beam-splitting interferometric laser measurement for short-distance axial vibration measurement, achieving high-precision non-contact measurement of stator bar vibration. The laser ranging module is designed with four installation methods to suit different installation environments and measurement requirements.

[0069] Please see Figure 2 As shown, the first installation method is phase-type laser ranging, where a phase-type laser ranging module 2 is installed on the side of the lower stator baffle 1. The phase-type laser ranging module 2 includes an annular slide rail 21 mounted on the side of the lower stator baffle 1. The laser ranging module 22 is fixed to the annular slide rail 21 via a track trolley 23. The track trolley 23 measures the vibration of different slot wedges by sliding on the annular slide rail 21. A power supply integrated on the annular slide rail 21 provides power to the track trolley 23 and the laser ranging module 22. During measurement, the light emitted by the laser ranging module 22 is modulated, and the time is indirectly measured by measuring the phase difference. The laser beam illuminates the end surface of the stator bar 3 at a suitable incident angle, and the reflected light is received by the optical receiving unit within the laser ranging module, thereby calculating the radial vibration of the stator bar 3. In one specific embodiment, the suitable incident angle can be 50-70°.

[0070] Please see Figure 3 As shown, the second installation method is a fixed phase-type laser ranging, in which multiple multi-laser emission ranging modules 20 are installed on the side of the lower windshield. This allows for the simultaneous emission of multiple lasers to measure the vibration at the ends of different stator bars 3. The rest is the same as the phase-type laser ranging method.

[0071] Please see Figure 4 As shown, the third installation method is phase-type laser ranging. The laser ranging module 22 is installed on the side of the stator pit through the ring track 21 and the track trolley 23. The radial condition of different stator bars 3 can be measured by sliding on the track. The integrated power supply on the track can provide power to the track trolley 23 and the laser ranging module 22.

[0072] Please see Figure 5 As shown, the fourth installation method is an internally installed spectroscopic interferometric laser ranging system. It uses spectroscopic interferometry to measure the axial vibration of the stator slot wedge 100. The spectroscopic interferometric laser displacement meter sensor head 200 is secured in the groove of the stator bar fixing block by passing a non-woven glass ribbon through the binding hole 300. The spectroscopic interferometric laser displacement meter sensor head 200 transmits signals via optical fiber, which is connected to an external controller through the fixing hole of the stator bar end fixing block 400. The fiber optic head structure consists only of optical fiber and lens, without electronic components, and is unaffected by electromagnetic noise. Furthermore, the sensor head structure is small: its length can be controlled within 5 cm and its diameter within 8 mm, avoiding interference from the sensor on the end vibration, while also facilitating installation and fixation.

[0073] The phase-type laser ranging module consists of three parts: a laser emitting system, an optical receiving system, and a signal processing system. Its working principle is to measure the phase difference between the emitted light and the reflected light by high-frequency modulated laser, and then calculate the target distance and its changes.

[0074] The laser emission system includes a semiconductor laser, a laser driving circuit, an optical collimation system, and a modulation circuit. The semiconductor laser uses a 650nm red laser or a 905nm near-infrared laser diode with an output power of 1-5mW. The laser driving circuit provides constant current drive and temperature control with a temperature control accuracy of ±0.1℃. The optical collimation system consists of a collimating lens and a beam expander, which shapes the laser beam into a parallel beam with a diameter of 2-5mm and a beam divergence angle of less than 1mrad. The modulation circuit modulates the laser intensity sinusoidally or with a square wave, and the modulation frequency can be set in the range of 10-100MHz.

[0075] The optical receiving system includes a photodetector, a preamplifier circuit, a bandpass filter circuit, and a phase detection circuit. The photodetector uses an avalanche photodiode or a PIN photodiode with a response time of less than 1 ns. The preamplifier circuit has an adjustable gain range of 40-80 dB. The center frequency of the bandpass filter circuit is matched with the modulation frequency. The phase detection circuit uses a mixer or a digital phase comparator to measure the phase difference, with a phase resolution better than 0.1°.

[0076] The signal processing system includes a high-speed analog-to-digital converter (ADC), a digital signal processor (DSP) or FPGA, a communication interface module, and a power management module. The high-speed ADC has a sampling rate of ≥100 MSPS and a resolution of 12-16 bits. The DSP or FPGA executes the phase calculation algorithm to calculate the distance. The communication interface supports RS485, Ethernet, or CAN bus, with a data update rate of up to 1 kHz. The power management module supports a wide voltage input of DC24V or DC48V. Typical performance specifications for the phase-type laser ranging module are: measurement range 50-500 mm, measurement accuracy ±50 μm, resolution 10 μm, sampling frequency 1 kHz, and response time less than 1 ms.

[0077] The beam-splitting interferometric laser ranging module uses a wide-wavelength light emitted by an ultra-wide LED. Part of this light is reflected by the reference surface of the sensor head, while the portion passing through the reference surface is specularly reflected from the workpiece and returns to the sensor head. The two reflected beams interfere with each other. The intensity of the interference light at a specific wavelength is determined based on the distance between the reference surface and the workpiece. Maximum relative interference is achieved when this distance is an integer multiple of the wavelength. By using a beam splitter to separate the interference light into different wavelengths, the intensity distribution at that specific wavelength can be obtained. Waveform analysis of this distribution reveals the distance to the workpiece.

[0078] When the laser ranging module performs internal measurements, it must ensure that the laser beam is perpendicular to the surface of the slot wedge with an angle of less than 2°. Simultaneously, the appropriate working distance must be adjusted according to the measurement range, and the probe position must be kept stable during measurement to avoid mechanical vibration interference. When using a track for measurement, a circular track is used to maintain the stability of the laser ranging host. The position and angle of the laser ranging unit are adjusted according to the position of the slot wedge. The debugging process includes optical power calibration, i.e., adjusting the laser power to the optimal operating point, establishing a reference for phase measurement, and calibrating the distance measurement accuracy using standard gauge blocks. This invention employs phase detection technology and beam splitting interferometry detection technology, achieving micron-level displacement measurement accuracy. It can capture the transient characteristics of the axial vibration of the slot wedge in real time and completely avoids the interference of traditional contact detection on generator operation, improving the safety and reliability of the detection. The optical measurement principle enables stable operation in strong electromagnetic environments, adapting to the harsh working environment of generators, while the fiber optic transmission method reduces the complexity of the system.

[0079] The raw vibration displacement data acquired by the laser ranging module is preprocessed, and the raw vibration displacement time series data continuously acquired by the laser ranging module at a frequency of 1kHz is subjected to quality control. First, a rationality check is performed, examining the data for physical rationality, rate of change rationality, and continuity. The physical rationality check verifies whether the displacement parameter values ​​are within the expected physical range; the rate of change rationality check monitors whether the rate of change of the parameters conforms to physical laws; and the continuity check identifies any data loss due to communication interruptions. Then, abnormal data repair is performed, addressing the identified anomalies. For anomalies exceeding the preset physical range, a sliding window average based on historical data is used for replacement. For data points with abnormal rates of change, a Kalman filter algorithm is used for smoothing to eliminate measurement noise and sudden interference. For missing data segments, an ARIMA time series prediction model is used for interpolation to complete the data and ensure its integrity.

[0080] This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, which also acquires acoustic signals through an acoustic monitoring module. Please refer to... Figure 7 As shown, the dashed box encloses the frame next to the stator bar end, and the protruding part is the acoustic monitoring module. The acoustic monitoring module uses a fiber optic acoustic emission sensor 500, a sensor that utilizes fiber optic technology to detect and measure acoustic emission phenomena. Acoustic emission refers to the transient elastic waves generated by changes in internal stress when a material is subjected to external forces. These waves can propagate within the material and be detected by the sensor. The fiber optic acoustic emission sensor uses optical fiber as the sensing element, indirectly measuring the acoustic emission signal by measuring changes in the light signal within the fiber. It has advantages such as miniaturization, high sensitivity, and resistance to electromagnetic interference. In this example, the sensor's core performance indicators include: frequency response range of 20-450kHz, resonant frequency of 220kHz, sensitivity ≥55dB, and omnidirectional directivity. For installation, the fiber optic probe can be directly attached to the surface of the stator bar using a stator bar fixing block to achieve accurate measurement of the acoustic emission signal. Alternatively, the fiber optic probe can be installed in the ventilation groove at the bottom of the stator bar slot to achieve accurate measurement of the acoustic emission signal. To ensure signal quality, the signal acquisition system is equipped with a 100kHz sampling frequency and a 24-bit high-precision ADC, achieving a wide dynamic range of 144dB. Simultaneously, the signal is conditioned by a differential input preamplifier with 20-60dB adjustable gain to maximize the signal-to-noise ratio and suppress common-mode interference.

[0081] Please see Figure 6As shown, the fiber optic acoustic emission probe 201 can be internally mounted. The probe is secured to the groove of the stator bar fixing block using a non-woven fiberglass ribbon. The probe transmits signals via optical fiber, which connects to an external controller through a fixing hole in the stator bar fixing block, allowing real-time acquisition of the sound signal emitted by the loosening of the slot wedge. Alternatively, it can be mounted on a rack near the end of the stator bar, with the fiber optic probe fixed and the wiring arranged within the rack.

[0082] Please see Figure 7 As shown, the fiber optic acoustic probe can be installed on the rack next to the end of the stator bar, and the fiber optic probe can be fixed and the wiring can be arranged inside the rack, so as to obtain the sound signal emitted by the loose stator end in real time.

[0083] The acoustic monitoring module continuously acquires the acoustic emission signals of the stator bars at a sampling frequency of 100kHz. The original signal is denoted as s(n), where n is the sampling point index. The preprocessing system adopts a three-stage cascaded filtering architecture to sequentially perform background noise suppression, electromagnetic interference removal, and signal enhancement.

[0084] The system employs adaptive spectral subtraction to remove steady-state background noise. The acoustic monitoring module continuously acquires the acoustic emission signals of the stator bars at a sampling frequency of 100kHz, denoted as the original signal s(n), where n is the sampling point index. First, the original acoustic signal is processed by framing, with a frame length of 256 sampling points corresponding to a 2.56-millisecond time window. Adjacent frames have a 50% overlap rate, i.e., a frame shift of 128 sampling points. After applying a Hanning window function to each frame, a 256-point Fast Fourier Transform is performed to convert the time-domain signal to the frequency domain.

[0085] During the noise power spectrum estimation phase, the system identifies quiet segments through energy detection within the initial 5 seconds of stable generator operation (corresponding to 500,000 sampling points). The energy of each frame is calculated, which is the sum of the squares of the amplitudes of all sampling points within the frame. When the energy of a frame is less than 1.5 times the average energy of the initial 100 frames, that frame is determined to be a quiet segment. The power spectrum of all identified quiet frames is averaged to establish a background noise power spectral density benchmark model.

[0086] When the signal power spectrum is greater than four times the noise power spectrum, the output amplitude is equal to the original amplitude minus the square root of twice the noise power spectrum; otherwise, the output amplitude is equal to 0.1 times the original amplitude. Phase information is preserved, and the time-domain signal is reconstructed using inverse FFT. The noise model is updated once per second, using a moving average: the new noise power spectrum is equal to 0.9 times the old power spectrum plus 0.1 times the current power spectrum.

[0087] Four cascaded second-order IIR notch filters are used, with center frequencies of 50Hz, 100Hz, 150Hz, and 200Hz, corresponding to the power frequency and its 2nd, 3rd, and 4th harmonics. The quality factor Q=30, notch depth -40dB, and 3dB bandwidth approximately 1.67Hz are applied. Electromagnetic interference suppression is achieved by sequentially passing the signal through these four notch filters. The system performs adaptive frequency tracking every 10 seconds, performing a 4096-point FFT on the output signal to search for energy peaks within ±2Hz of each center frequency. When the peak deviation exceeds 0.2Hz, the filter coefficients are updated to compensate for power grid frequency fluctuations.

[0088] After initial noise suppression and electromagnetic interference removal, the system enters the signal enhancement stage. Wiener filtering, based on the minimum mean square error criterion, is employed to further optimize signal quality. This method is particularly suitable for handling stationary random noise, achieving optimal filtering in the frequency domain by estimating the power spectral density of the signal and noise. To preserve the transient impact characteristics generated by the wedge loosening, wavelet denoising technology is used, employing a 6-level decomposition based on the Daubechies-8 wavelet basis. Its good orthogonality and tight support characteristics ensure the accuracy of time-frequency localization analysis. A soft thresholding strategy is used in the wavelet domain, with the threshold adaptively determined according to the Donoho criterion, effectively removing high-frequency noise while fully preserving the peak characteristics of the impact signal.

[0089] Finally, bandpass filtering was used for frequency selection, employing a Chebyshev Type II filter with a passband range of 500Hz to 20kHz. This band covers the main frequency components of slot wedge vibration and early loosening. The filter is designed as an 8th-order structure with a stopband attenuation of 60 dB and a passband ripple of less than 0.1 dB. Zero-phase filtering technology was specifically employed, using both forward and reverse filtering to eliminate phase distortion and ensure the time-positioning accuracy of transient characteristics.

[0090] The acoustic signal, after three levels of preprocessing, is continuously output at a sampling rate of 100kHz. The feature extraction module employs a sliding window mechanism, calculating five acoustic features every 1 millisecond (corresponding to 100 sampling points), including root mean square value, peak factor, kurtosis, spectral centroid, and spectral variance. This results in an acoustic feature output frequency of 1kHz, perfectly synchronized with the 1kHz feature output frequency of the laser ranging module. This achieves precise alignment of multimodal data on the time axis, providing a unified data benchmark for subsequent operational parameter fusion and deep learning anomaly detection.

[0091] This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising feature extraction of acquired vibration and acoustic signals; specifically including: Root Mean Square (RMS): The root mean square (RMS) value of vibration and acoustic signals reflects the effective value of the groove wedge vibration displacement and is a fundamental indicator for evaluating vibration energy. For the displacement signal x(t), the RMS value within the time window T is calculated as follows: Where N is the number of sampling points within the window, and x(n) is the amplitude of the nth sampling point. The calculation uses a sliding window method with a window length of 1 second (1000 sampling points) and a sliding step size of 100ms to ensure time resolution. Loosening of the slot wedge leads to increased vibration of the bar, increasing the amplitude x(n), and its sum of squares naturally also increases, therefore X... rms It will rise steadily. This is a fundamental yet effective macroeconomic indicator.

[0092] Crest Factor: The crest factor is the ratio of the peak value of a signal to its RMS value. It is used to quantify the extremes of a signal waveform. A stationary sine wave has a crest factor of... When the bar loosens, the collision between the bar and the trench wall generates a momentary high-amplitude impact (the maximum amplitude Xpeak increases sharply). While the RMS value also increases, its growth rate is far less than that of the peak factor CF. Therefore, the ratio of the two increases significantly. This indicator is highly sensitive to early, sporadic impact signals.

[0093]

[0094] Kurtosis: The kurtosis K of a displacement signal is a fourth-order statistic that is sensitive to the peak characteristics of the vibration distribution. Normal vibrations approximate a Gaussian distribution with a kurtosis of approximately 3 (excess kurtosis is 0). When loosening causes an impact, numerous impact peaks far from the mean appear in the signal, causing the probability distribution to exhibit a "sharp-topped, fat-tailed" shape. This leads to a sharp increase in the fourth moment, thereby causing a sharp increase in the kurtosis value K.

[0095]

[0096] in, The mean, Let Variance be the variance.

[0097] Spectral Centroid: Before calculating spectral parameters, a Fast Fourier Transform (FFT) is performed on the time-domain signal x(n) to obtain its spectrum S(k), where k is the index of a frequency point. |S(k)| represents the amplitude at the k-th frequency point. Under normal, tight conditions, energy is mainly concentrated around 100Hz, so the spectral centroid will stabilize at a relatively low frequency. When loosening occurs, impact and friction generate a large number of high-frequency signals, causing the energy of the spectrum to shift towards higher frequencies, thus significantly increasing the spectral centroid (shifting towards higher frequencies).

[0098]

[0099] Where f(k) represents the actual physical frequency corresponding to the k-th data point in the spectrum, f(k)=(k-1)*fs / N, where fs is the sampling frequency of 1kHz, N is the number of FFT points, and k is the frequency point index.

[0100] Spectral variance (SV) measures the dispersion of vibrational energy in the frequency domain, reflecting the frequency complexity of the vibration. The variance of single-frequency vibration is close to 0, while broadband vibration has a large variance. Normal slotted wedge vibration is dominated by a specific frequency, resulting in a small spectral variance. Loosening leads to increased multimodal vibration and random components, increasing the spectral variance. Spectral variance is particularly sensitive to early loosening; spectral broadening can be detected even before a significant increase in amplitude.

[0101]

[0102] The feature extraction method of this invention not only focuses on the physical meaning of individual features, but more importantly, it constructs a symmetrical multimodal feature framework. Although laser ranging and acoustic monitoring measure different physical quantities, they achieve a unified mathematical expression by extracting the same five features (RMS, peak factor, kurtosis, spectral centroid, and spectral variance). This symmetry gives the 15-dimensional feature vector (5 laser features + 5 acoustic features + 5 operating condition parameters) a regular structure, which facilitates the deep learning model to learn multimodal correlation patterns in a unified feature space, avoids the complexity of heterogeneous feature fusion, and significantly improves the system's real-time processing capability and detection accuracy.

[0103] This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising operating condition parameter acquisition and multimodal data fusion; specifically including: The operating condition parameter acquisition and preprocessing module plays a crucial role in building the data foundation of the entire system. Its core function is to acquire generator operating status parameters in real time, perform precise time-synchronized fusion with vibration data obtained from the laser ranging module and sound signals from the sound monitoring module, and construct a standardized multi-dimensional data sequence to provide high-quality input data for subsequent deep learning anomaly detection models. A distributed data acquisition architecture is adopted, and multi-sensor fusion technology ensures the integrity, accuracy, and real-time nature of the data, laying a solid data foundation for the intelligent online monitoring of slot wedge loosening.

[0104] Multi-source data acquisition and synchronization mechanism: This invention simultaneously collects three major categories of data: laser ranging features (5), acoustic features (5), and operating parameters (5), forming a unified 15-dimensional feature vector. Both the laser ranging and acoustic modules output feature vectors at a frequency of 1 kHz, while the operating parameters are collected at 1 Hz and interpolated to 1 kHz. The operating parameters include several parameters from stator core vibration, motor start-up time, continuous running time, stator slot temperature, stator core temperature, generator active power, generator reactive power, stator current, stator voltage, power factor, and rotational speed, thereby reflecting the interference of mechanical stress, thermal stress, and electromagnetic force on slot wedge vibration. To ensure that the monitoring strategy of this invention can flexibly adapt to and efficiently apply to various generator sets, the system will first run an adaptive optimization algorithm for operating parameters before starting the online monitoring task. This algorithm, based on the analysis of historical operating data of the equipment, first categorizes all available operating parameters into three major pools—mechanical stress, thermal stress, and electromagnetic force—according to their physical mechanisms of action on the slot wedge state. Next, the algorithm utilizes data accumulated by the unit under normal operating conditions to calculate the Pearson correlation between each candidate parameter and all ten acoustic and optical monitoring features (including five laser ranging features and five acoustic features). The absolute values ​​of each correlation coefficient are then averaged to quantify a comprehensive correlation score for each parameter. The system first selects the parameters with the highest scores from each of the three stress category pools to ensure a basic representation of the three core influencing factors. Then, from all remaining candidate parameters, the two with the highest correlation scores are selected as supplementary parameters. Changes in these operating condition parameters significantly affect the vibration and acoustic properties of the stator slot wedges under normal operating conditions; therefore, they must be collected synchronously with vibration data to establish an accurate unsupervised deep learning model: the operating condition-vibration-acoustic correlation model. In one specific implementation, the operating condition parameters are selected from five parameters based on the Pearson correlation: stator core vibration, motor start-up time, continuous running time, stator slot temperature, stator core temperature, generator active power, generator reactive power, stator current, stator voltage, power factor, and rotational speed. In one specific implementation, the operating parameters include five parameters: stator slot temperature, stator core temperature, generator active power, power factor, and rotational speed.

[0105] Temperature data acquisition utilizes the high-precision digital temperature and humidity sensor SHT30, which boasts a measurement accuracy of ±0.3℃ and excellent long-term stability, accurately reflecting stator operating temperature changes. The SHT30 continuously acquires stator slot and stator core temperatures at a frequency of 1Hz via an I2C communication interface. Generator operating parameters are acquired through a standardized communication interface with the existing generator monitoring system, obtaining key operating parameters such as active power, speed, and power factor in real time at a frequency of 1Hz. This interface design avoids redundant sensor deployment, improving system economy and reliability.

[0106] Both laser ranging and acoustic monitoring features include RMS value, peak factor, kurtosis, spectral centroid, and spectral variance. The laser ranging module continuously acquires axial vibration displacement data of the stator slot wedge at a high frequency of 1 kHz. This high sampling rate captures detailed time-domain characteristics and transient changes in the slot wedge vibration. High-frequency sampling is crucial for identifying subtle vibration changes caused by slot wedge loosening, as vibration anomalies in the early stages of loosening often manifest as an increase in high-frequency components or subtle changes in vibration modes. Although the acoustic signal is sampled at a high frequency of 100 kHz to capture high-frequency acoustic emission components, its feature extraction is performed within a time window. Specifically, the system processes the first 100 acoustic sampling points (corresponding to 1 ms of data at a 100 kHz sampling rate) every 1 millisecond (corresponding to 1 kHz), calculating five acoustic features within this window containing 100 data points. This design preserves the high-frequency information of the acoustic signal for accurate feature calculation while reducing the feature output frequency to 1 kHz, perfectly matching the sampling frequency of the laser ranging.

[0107] To ensure accurate correlation of multi-source data from different acquisition frequencies, the system employs a dual time synchronization mechanism combining GPS and the Network Time Protocol (NTP). During startup, the system performs high-precision clock calibration using a GPS receiver to obtain an absolute time reference at the microsecond level. During normal operation, the system maintains synchronization with a standard time server via the NTP protocol, ensuring time reference consistency across all acquisition modules. Each data acquisition module uses a unified UTC timestamp format, recording the precise time information of the acquisition moment in the data packet header, providing a reliable time reference for subsequent data fusion processing.

[0108] This invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, including data fusion, quality control, and standardization; specifically including: Data fusion processing is a crucial step in unifying multi-source data with different sampling frequencies onto the same time base. The system uses the 1kHz sampling rate of the laser ranging module as the base frequency and upsamples the low-frequency sampled operating parameter data to 1kHz using an interpolation algorithm. For temperature data, due to its relatively slow changes and good continuity, the system employs linear interpolation for upsampling. For generator operating parameters, considering the potential nonlinear changes in these parameters over short periods, a cubic spline interpolation algorithm is used. This algorithm maintains the smoothness of the data while accurately reproducing the parameter change trends.

[0109] The fused data forms a unified time-series dataset, generating a 15-dimensional feature vector at each millisecond, containing 5 laser ranging features, 5 acoustic features, and 5 operating condition parameters. This feature vector provides a comprehensive description of the generator's state at that moment, integrating information from three dimensions: mechanical vibration, acoustic emission, and operating conditions. Feature vector generation is streaming, with a new vector produced every millisecond, forming a continuous data stream. This data structure provides deep learning models with complete information on the correlation between operating conditions, vibration, and acoustics, enabling the models to learn the complex patterns of normal vibration of the slot wedge under different operating conditions.

[0110] Data quality control is crucial for ensuring the accuracy of subsequent anomaly detection. The system implements a multi-layered data quality assessment and anomaly handling mechanism. First, a physical plausibility check is performed to verify that all parameters are within the expected physical range; for example, the ambient temperature should be between 15℃ and 80℃, and the speed variation should not exceed ±5% of the rated speed. Second, a rate of change plausibility check is performed to monitor whether the rate of parameter change conforms to physical laws; for example, the rate of temperature change typically does not exceed 5℃ per minute. Finally, a data continuity check is performed to identify any data loss or communication interruptions.

[0111] For identified outlier data, the following repair methods are implemented: Outliers exceeding the preset physical range are replaced with a sliding window average based on historical data. For data points with abnormal rates of change, a Kalman filter algorithm is used for smoothing, which can eliminate measurement noise and sudden interference while maintaining data accuracy. For missing data, the ARIMA time series forecasting model is used for interpolation completion; this model can predict data values ​​for missing periods based on the statistical characteristics of historical data.

[0112] To eliminate the influence of dimensional differences between different features, the system performs Z-score standardization on all features. The standardization formula is: Where x is the original feature value, The mean of the features, The standard deviation of the feature is denoted as σ. Standardized features have zero mean and unit variance, which allows deep learning models to converge better and prevents certain features from dominating the learning process due to their large values.

[0113] The final output feature vector contains 15 dimensions. This concise feature vector not only contains key information about the groove wedge vibration and acoustic signals, but also integrates the main working condition factors that affect the normal vibration mode, providing high-quality input data for the subsequent anomaly detection model.

[0114] The core innovation of this invention is the deep autoencoder deep learning anomaly detection module, which employs unsupervised learning to intelligently identify stator slot wedge loosening. The module's design is based on the understanding that, under normal operating conditions, while the vibration mode of the stator slot wedge changes with varying operating conditions, these changes are regular and predictable. When the slot wedge becomes loose, its vibration mode deviates from the normal operating condition-vibration-acoustic relationship, exhibiting abnormal characteristics. By learning the complex nonlinear relationship between operating conditions and vibration from a large amount of normal operating data, the deep learning model can identify abnormal states that deviate from the normal pattern, thereby achieving automatic detection of slot wedge loosening.

[0115] This invention employs a deep autoencoder network as the core anomaly detection algorithm. An autoencoder is a special type of neural network structure that learns to compress input data into a low-dimensional latent space and then reconstruct it back to the original dimension, thus capturing the inherent structure and distribution characteristics of the data. When trained on normal data, the autoencoder can reconstruct normal samples well, but the reconstruction error for anomalous samples increases significantly. By monitoring the magnitude of the reconstruction error, anomalous states can be effectively identified.

[0116] Please see Figure 9 As shown, the autoencoder network structure adopts a symmetrical encoder-decoder architecture. The encoder is responsible for progressively compressing the 15-dimensional input feature vector into a low-dimensional latent representation, while the decoder reconstructs the latent representation back into the original 15-dimensional feature space. The encoder consists of three fully connected layers with 32, 16, and 8 neurons respectively. Each layer is followed by a ReLU activation function and a batch normalization layer to improve training stability and convergence speed. The middle latent layer contains 4 neurons. This design captures the main features of the data while avoiding information loss due to over-compression. The decoder adopts a completely symmetrical structure with the encoder. The number of neurons in the three fully connected layers and the output layer are 8, 16, 32, and 15 respectively. The output layer uses a linear activation function to ensure the continuity of the output value.

[0117] To enhance the model's ability to capture temporal features, a time series processing mechanism was introduced on top of the autoencoder. Specifically, the model's input is not a feature vector at a single time point, but a feature sequence containing continuous time windows. The feature vectors of 10 consecutive time points (corresponding to a time span of 0.01 seconds) are combined into a 150-dimensional input vector. This design enables the model to learn the short-term temporal dependencies between vibration and operating parameters, improving its ability to identify dynamic anomaly patterns.

[0118] Constructing training data is crucial for successful unsupervised anomaly detection. It requires collecting a large amount of data from generators under various normal operating conditions to ensure the training dataset adequately covers the generator's normal operating range. Data collection strategies include variations in temperature, power output at different load levels, and speed and power factor changes under different operating modes. By collecting data over a long period under multiple operating conditions, a comprehensive training dataset reflecting normal operating modes can be established.

[0119] The training dataset is set to have at least 500,000 sample points, which should cover the main operating conditions of the generator. Ambient temperature should cover 10°C to 70°C, active power should cover 50% to 100% of rated power, speed variation should be within ±2% of rated speed, and power factor should cover 0.8 to 1.0. Under each main operating condition combination, the system needs to collect a sufficient number of samples to ensure that the model can fully learn the normal vibration modes under that condition.

[0120] Model training employs the standard backpropagation algorithm and the Adam optimizer. The initial learning rate is set to 0.001, and a learning rate decay strategy is used to improve training stability. The batch size is set to 256, achieving a good balance between training efficiency and memory usage. An early stopping mechanism is used during training; training automatically stops when the reconstruction error on the validation set fails to improve for 20 consecutive epochs to avoid overfitting.

[0121] The loss function uses mean squared error (MSE) to calculate the difference between the input feature vector and the reconstructed feature vector. To improve the robustness of the model, small-amplitude Gaussian noise (standard deviation of 0.01) is added to the input data during training. This noise injection technique can improve the model's resistance to measurement noise and environmental interference.

[0122] Determining the anomaly detection threshold is a crucial step in the practical application of unsupervised methods. This invention employs statistical methods to determine the optimal anomaly detection threshold on a validation dataset. First, the trained autoencoder model is applied to an independent validation dataset, which also contains only data from normal operation. The reconstruction error for each sample is calculated, and its statistical distribution characteristics are analyzed. The reconstruction error, also known as the mean squared error (MSE), is quantified by calculating the difference between the input feature vector and the reconstructed feature vector.

[0123] Reconstruction errors typically follow a known probability distribution. By fitting the distribution parameters of the reconstruction errors to the validation set, a statistical threshold for anomaly detection is determined. The 3σ principle is applied, classifying reconstruction errors exceeding the mean plus three standard deviations as anomalies. This method can identify anomalous states with a 99.7% confidence level and demonstrates good detection performance and a low false alarm rate in practical applications.

[0124] The online inference process is designed as a real-time streaming process. The system continuously receives standardized feature vectors output by the preprocessing module and maintains a sliding window of length 10 to cache the most recent feature sequences. Whenever a new feature vector is received, the system inputs the latest feature sequence into the trained autoencoder model for forward inference, calculates the reconstruction error, and compares it with a preset threshold.

[0125] When the reconstruction error exceeds the anomaly detection threshold, the system determines the current state as abnormal and triggers a slot wedge loosening alarm. To improve the reliability of detection, this invention adopts a continuous anomaly confirmation mechanism, that is, an alarm is only formally triggered when the reconstruction error at three consecutive time points exceeds the threshold. This design can effectively avoid false alarms caused by instantaneous interference.

[0126] The anomaly detection method of this invention, in addition to outputting a binary judgment of normal or abnormal, also includes a step of quantitatively assessing the degree of anomaly. This step is specifically implemented by comparing the reconstructed error value calculated in real time with a pre-set anomaly detection threshold (i.e., the baseline threshold for the normal state), and calculating the ratio between the two. This ratio, as a continuous numerical indicator, objectively quantifies the severity of the equipment's deviation from the normal mode, thereby providing operation and maintenance personnel with a more accurate basis for state assessment to formulate reasonable maintenance strategies.

[0127] To ensure that the anomaly detection model (a trained deep autoencoder) can adapt to changes in the generator's state and environment during long-term operation, this invention designs a continuous model optimization and online learning mechanism. This mechanism is based on the idea that the generator's normal operating mode may be fine-tuned as equipment ages, maintenance status changes, and environmental conditions evolve; the anomaly detection model needs to be able to track these changes to maintain detection accuracy.

[0128] The online learning mechanism updates model parameters using incremental learning. It continuously collects operational data that has been confirmed to be in a normal state, and when the accumulated new data reaches a certain scale (e.g., 10,000 samples), the incremental update process is automatically initiated. Incremental updates use a small learning rate (typically 10% of the initial training learning rate) and fewer training epochs, learning new normal operating modes while maintaining existing knowledge.

[0129] To prevent model degradation due to erroneous data, a strict data quality control mechanism is implemented. Only data that has been manually verified or confirmed to be in a normal state through other independent detection methods is included in the online learning dataset. Simultaneously, the system retains multiple historical versions of the model, allowing for rapid rollback to a previous stable version when a significant performance degradation is detected.

[0130] Continuous monitoring of model performance is achieved through multiple indicators, including trend changes in average reconstruction error, statistical analysis of anomaly detection frequency, and cross-validation with traditional detection methods. The system has established a comprehensive model performance evaluation framework to ensure the long-term stability and reliability of anomaly detection capabilities.

[0131] Through the above-mentioned carefully designed unsupervised anomaly detection scheme, the system can automatically learn the normal operation mode without manual annotation and accurately identify abnormal states such as loose slot wedges, thus achieving true intelligent online monitoring.

[0132] Please see Figure 10 As shown, this invention provides an online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising: S100: Acquire laser ranging data, acoustic monitoring data, and generator operating parameter data for online loosening detection of generator stator slot wedges; S200: Preprocesses laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters; S300. Input the optical ranging features, acoustic features and operating parameters into the pre-trained self-encoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge.

[0133] In one specific embodiment, the step of obtaining laser ranging data, acoustic monitoring data, and generator operating condition parameter data for online loosening detection of generator stator slot wedges specifically includes obtaining laser ranging data by using a laser ranging module arranged beside the end of the generator stator bar and employing the phase-type laser ranging principle or beam-splitting interferometric laser measurement. Acoustic monitoring data for online loosening detection of generator stator slot wedges specifically includes: using fiber optic acoustic emission sensors arranged on the generator stator bar to collect acoustic emission signals of the stator bar and obtain acoustic monitoring data; The specific operating parameters include: stator slot temperature, stator core temperature, generator active power, speed, and power factor.

[0134] In one specific embodiment, the step of preprocessing the laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging features, acoustic features, and operating parameters specifically includes: performing feature extraction on both the laser ranging data and the acoustic monitoring data to obtain five features: root mean square value, peak factor, kurtosis, spectral centroid, and spectral variance. The five features of laser ranging data, the five features of acoustic monitoring data, and the five data interpolations of operating condition parameters are standardized with a unified sampling rate. Data alignment and standardization are then performed to obtain preprocessed laser ranging features, acoustic features, and operating condition parameters.

[0135] In one specific embodiment, the step of preprocessing the laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters includes removing outliers and interpolation completion.

[0136] In one specific embodiment, in the step of inputting the optical ranging features, acoustic features and operating condition parameters into a pre-trained self-encoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the self-encoder includes an input layer, an encoder, a latent layer, a decoder and an output layer. The input layer has 15 neurons; The encoder consists of three fully connected layers with 32, 16 and 8 neurons respectively. Each fully connected layer is followed by a ReLU activation function and a batch normalization layer. The middle potential layer contains 4 neurons; The decoder adopts a completely symmetrical structure with the encoder. The number of neurons in the three fully connected layers and the output layer are 8, 16, 32 and 15, respectively. The output layer uses a linear activation function.

[0137] In one specific embodiment, in the step of inputting the optical ranging features, acoustic features, and operating condition parameters into a pre-trained self-encoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the pre-trained self-encoder is obtained through the following steps: The autoencoder is trained using the standard backpropagation algorithm and the Adam optimizer. The initial learning rate can be set to 0.001, and the batch size can be set to 256. An early stopping mechanism is used during training. When the reconstruction error on the validation set has not improved for 20 consecutive rounds, training is automatically stopped to obtain the pre-trained autoencoder.

[0138] In one specific embodiment, in the step of inputting the optical ranging features, acoustic features, and operating condition parameters into a pre-trained self-encoder for anomaly detection and obtaining the online loosening detection result of the generator stator slot wedge, the feature vectors of 10 consecutive time points are combined into a 150-dimensional input vector and input into the pre-trained self-encoder for anomaly detection.

[0139] In one specific embodiment, in the step of inputting the optical ranging features, acoustic features, and operating parameters into a pre-trained autoencoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the optical ranging features, acoustic features, and operating parameters are input into the pre-trained autoencoder to obtain a reconstruction feature vector. The reconstruction error is calculated based on the reconstruction feature vector. The reconstruction error is compared with an anomaly detection threshold. When the reconstruction error at M consecutive time points exceeds the anomaly detection threshold, an alarm is triggered, and the online loosening detection result of the generator stator slot wedge is obtained, where M is a positive integer greater than or equal to 1. The anomaly detection threshold is determined using the 3σ principle.

[0140] In one specific implementation, the method further includes: the pre-trained autoencoder updates the model parameters using incremental learning.

[0141] Please see Figure 11 and Figure 12 As shown, this embodiment of the invention provides an online detection system for generator stator slot wedges based on acoustic-optical monitoring fusion, comprising: The acquisition module is used to acquire laser ranging data, acoustic monitoring data, and generator operating parameter data for online loosening detection of generator stator slot wedges; The preprocessing module is used to preprocess laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters. The detection module is used to input the optical ranging features, acoustic features and operating parameters into a pre-trained self-encoder for anomaly detection, and obtain the online loosening detection results of the generator stator slot wedge.

[0142] This invention is based on a dual-modal detection architecture design that deeply integrates optical ranging and acoustic monitoring. This architecture is not simply a parallel connection of sensors, but rather an organic fusion of the two detection modes at the system level. The laser ranging module employs both phase-based and beam-splitting interferometry techniques. Phase-based ranging achieves long-distance measurement by measuring the phase difference of the modulated laser, while beam-splitting interferometry achieves high-precision short-range measurement by analyzing interference fringes. The acoustic monitoring module uses a fiber optic acoustic emission sensor with a frequency response range covering 20-450kHz, capable of capturing signals across the entire frequency band, from low-frequency vibration noise to high-frequency acoustic emissions. The two modes achieve deep synergy at all levels, including hardware design, signal acquisition, feature extraction, and fusion decision-making, forming a complementary detection system.

[0143] This invention employs a symmetrical feature extraction and multimodal fusion framework. The system innovatively uses the exact same feature extraction strategy for both optical and acoustic signals. The technical challenge of this symmetrical design lies in extracting features with the same physical meaning from signals with different physical properties. For vibration displacement signals, the root mean square (RMS) value directly reflects the vibration amplitude; for acoustic signals, the RMS value reflects the sound pressure level. Although the physical quantities differ, they both characterize the energy level. Similarly, the peak factor reflects the impact characteristics, the kurtosis reflects the non-Gaussianity of the distribution, the spectral centroid reflects the center of the frequency distribution, and the spectral variance reflects the degree of frequency dispersion. This unified feature representation provides a mathematical foundation for the training and fusion decision-making of deep learning models.

[0144] This invention employs precise spatiotemporal synchronization technology for multi-source heterogeneous data. The system needs to process three data streams with different sampling rates: 1kHz laser ranging data, 100kHz raw acoustic signal (processed to output 1kHz characteristics), and 1Hz operating parameters. The technical challenge of time synchronization lies in achieving precise alignment while maintaining the temporal continuity of each data stream. The system adopts a layered synchronization strategy: the hardware layer obtains an absolute time reference through GPS timing, the software layer maintains time consistency through the NTP protocol, and the application layer achieves data alignment through timestamp matching. For data with different sampling rates, the system uses intelligent interpolation algorithms: linear interpolation is used for slowly changing temperature parameters, and cubic spline interpolation is used for parameters that may change rapidly, such as power and rotational speed, ensuring that the interpolated data retains the original trend of change.

[0145] This invention employs an unsupervised anomaly detection algorithm based on a deep autoencoder. The core innovation lies in achieving anomaly detection through reconstruction learning, avoiding reliance on anomalous samples. The network architecture design fully considers the balance between feature dimensionality and computational efficiency: the 15 neurons in the input layer correspond to 15-dimensional feature vectors, which are compressed into a 4-dimensional latent space through a decreasing 32-16-8 structure, and then reconstructed through a symmetric 8-16-32-15 structure. ReLU activation functions are used between layers to maintain non-linear expressiveness, and batch normalization layers accelerate training convergence. Setting the latent layer dimension to 4 is a thoroughly validated optimization result, capturing the main patterns of the data while avoiding information loss due to over-compression. The training process uses a mean squared error loss function, learning the distribution characteristics of normal data by minimizing the reconstruction error.

[0146] This invention employs a multi-condition adaptive dynamic benchmark modeling technique. The system needs to learn normal vibration and acoustic modes under different combinations of operating conditions. The technical challenge lies in the high dimensionality and nonlinearity of the operating condition space. Different combinations of parameters such as temperature, power, speed, and power factor create a vast operating condition space, and the vibration characteristics of the slot wedge have complex nonlinear relationships with these parameters. A deep learning model learns this mapping relationship through a large amount of data, establishing a dynamic operating condition-feature benchmark model. During online detection, the system first predicts the normal feature range based on the current operating condition parameters using the model, and then evaluates the deviation of the actual measured values, achieving true adaptive detection.

[0147] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for online detection of generator stator slot wedges based on acoustic-optical monitoring fusion, characterized in that, include: Acquire laser ranging data, acoustic monitoring data, and generator operating parameter data for online detection of stator slot wedge loosening in generators; Preprocessing is performed on laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters; The optical ranging features, acoustic features, and operating parameters are input into a pre-trained self-encoder for anomaly detection, thereby obtaining the online loosening detection results of the generator stator slot wedges.

2. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, In the steps of obtaining laser ranging data, acoustic monitoring data, and generator operating parameter data for online loosening detection of generator stator slot wedges, obtaining laser ranging data for online loosening detection of generator stator slot wedges specifically includes: obtaining laser ranging data by using a laser ranging module arranged on or beside the end of the generator stator bar and employing the principle of phase-type laser ranging or beam-splitting interferometric laser measurement. Acoustic monitoring data for online loosening detection of generator stator slot wedges specifically includes: using fiber optic acoustic emission sensors arranged on the side frame at the end of the generator stator bar or inside the bar to collect acoustic emission signals of the stator bar and obtain acoustic monitoring data; The generator operating parameters specifically include several of the following: speed, stator core vibration, motor start-up time, continuous running time, stator slot temperature, stator core temperature, generator active power, generator reactive power, stator current, stator voltage, and power factor.

3. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 2, characterized in that, The steps of preprocessing laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging features, acoustic features, and operating parameters specifically include: extracting features from both laser ranging data and acoustic monitoring data to obtain five features: root mean square value, peak factor, kurtosis, spectral centroid, and spectral variance. The five features of laser ranging data, the five features of acoustic monitoring data, and the five features of operating condition parameters are interpolated at a unified sampling rate, and data alignment and standardization are performed to obtain preprocessed laser ranging features, acoustic features, and operating condition parameters.

4. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, In the step of preprocessing laser ranging data, acoustic monitoring data, and generator operating condition parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating condition parameters, the preprocessing includes removing outliers and interpolation completion.

5. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, In the step of inputting the optical ranging features, acoustic features and operating condition parameters into a pre-trained self-encoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the self-encoder includes an input layer, an encoder, a latent layer, a decoder and an output layer. The input layer has 15 neurons; The encoder consists of three fully connected layers with 32, 16 and 8 neurons respectively. Each fully connected layer is followed by a ReLU activation function and a batch normalization layer. The middle potential layer contains 4 neurons; The decoder adopts a completely symmetrical structure with the encoder. The number of neurons in the three fully connected layers and the output layer are 8, 16, 32 and 15, respectively. The output layer uses a linear activation function.

6. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, In the step of inputting the optical ranging features, acoustic features, and operating parameters into a pre-trained self-encoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the pre-trained self-encoder is trained through the following steps: The autoencoder was trained using the standard backpropagation algorithm and the Adam optimizer. The initial learning rate was set to 0.001 and the batch size was set to 256. An early stopping mechanism was used during training. When the reconstruction error on the validation set did not improve for N consecutive rounds, the training was automatically stopped to obtain the pre-trained autoencoder. N is a positive integer between 10 and 50.

7. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, In the step of inputting the optical ranging features, acoustic features, and operating condition parameters into a pre-trained autoencoder for anomaly detection and obtaining the online loosening detection result of the generator stator slot wedge, the feature vectors of 10 consecutive time points are combined into a 150-dimensional input vector and input into the pre-trained autoencoder for anomaly detection.

8. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, In the step of inputting the optical ranging features, acoustic features, and operating parameters into a pre-trained autoencoder for anomaly detection to obtain the online loosening detection result of the generator stator slot wedge, the optical ranging features, acoustic features, and operating parameters are input into the pre-trained autoencoder to obtain a reconstruction feature vector. The reconstruction error is calculated based on the reconstruction feature vector. The reconstruction error is compared with the anomaly detection threshold. When the reconstruction error exceeds the anomaly detection threshold for M consecutive time points, an alarm is triggered, and the online loosening detection result of the generator stator slot wedge is obtained. M is a positive integer greater than or equal to 1. The anomaly detection threshold is determined using the 3σ principle.

9. The online detection method for generator stator slot wedges based on acoustic-optical monitoring fusion according to claim 1, characterized in that, Also includes: The pre-trained autoencoder updates the model parameters using incremental learning.

10. An online detection system for generator stator slot wedges based on acoustic-optical monitoring fusion, characterized in that, include: The acquisition module is used to acquire laser ranging data, acoustic monitoring data, and generator operating parameter data for online loosening detection of generator stator slot wedges; The preprocessing module is used to preprocess laser ranging data, acoustic monitoring data, and generator operating parameter data to obtain laser ranging characteristics, acoustic characteristics, and operating parameters. The detection module is used to input the optical ranging features, acoustic features and operating parameters into a pre-trained self-encoder for anomaly detection, and obtain the online loosening detection results of the generator stator slot wedge.

Images