Turbonator slip ring carbon brush state monitoring and fault early warning method based on multi-modal data fusion

Abstract

The invention discloses a steam turbine generator slip ring carbon brush state monitoring and fault early warning method based on multi-modal data fusion. Comprising the following steps: S1, synchronously acquiring multi-source heterogeneous monitoring data of the slip ring carbon brush system; s2, carrying out preprocessing and alignment on the multi-source heterogeneous monitoring data acquired in the step S1; s3, extracting multi-modal features related to the health state of the carbon brush from the preprocessed data; s4, performing fusion calculation on the multi-modal features extracted in the step S3 by using a multi-modal feature fusion model to generate a comprehensive health index; s5, state evaluation and fault early warning are carried out according to the comprehensive health index obtained in the step S4 and preset multi-level early warning logic; and S6, outputting fault early warning information and a diagnosis result. According to the invention, an innovative temperature rise rate and trend alarm model can identify a slow degradation process and a rapid change fault, the fault early warning time is advanced, and the occurrence of serious accidents is avoided.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and fault diagnosis technology, and is a method for condition monitoring and fault early warning of slip ring carbon brushes of steam turbine generators based on multimodal data fusion. Background Technology

[0002] With the deepening of clean energy development and power system reform, my country's new energy industry has experienced tremendous growth. Wind power, photovoltaic power, and other new energy sources have gradually surpassed traditional thermal power in both installed capacity and power generation. Under this energy structure, the external environment for thermal power units has changed dramatically. Their operating mode has shifted from primarily handling base loads to flexible operations such as peak shaving, voltage regulation, and frequency regulation. Furthermore, with the continuous integration of distributed energy sources into the power system, the depth and duration of the flexible operation that thermal power units must undertake have gradually increased. As a core component of modern power systems, steam turbine generators undertake the critical tasks of power conversion and transmission, and their operational stability directly affects the reliability of the power grid and the quality of power supply. Since generator slip rings and carbon brushes are important current-conducting components in self-excited generators, serving as transition devices for current flowing from stationary to rotating components, they are highly susceptible to various faults. Their performance directly determines the generator's dynamic response characteristics and long-term operational reliability.

[0003] During deep peak shaving, generators need to frequently adjust the load, causing significant changes in excitation current within a short period. At this time, the heat exchange and thermal equilibrium between the rotor windings and the rotor core are much slower than the rate of temperature change in the windings, leading to increased relative expansion of the windings and core. This results in the slip ring carbon brush system operating under harsh conditions of high-speed friction and high-current conduction for extended periods, facing multiple challenges such as mechanical wear, poor contact, and overheating oxidation. Statistics show that in large steam turbine generator failure cases, approximately 23% of unplanned outages are directly related to slip ring carbon brush system failures. These failures not only cause sudden unit shutdowns and millions in economic losses but can also trigger chain reactions, threatening the safe and stable operation of the power grid.

[0004] Current monitoring methods for slip ring carbon brush systems primarily rely on periodic inspections and offline testing, which suffer from issues such as detection lag and insufficient accuracy. In particular, real-time monitoring of key parameters such as carbon brush wear status, contact pressure distribution, and ring surface temperature field remains a technical bottleneck. With the advancement of smart grid construction and the increasing size of generating units, there is an urgent need to develop new intelligent monitoring systems to achieve real-time and accurate assessment of the slip ring carbon brush system's operating status. This would provide data support for preventative maintenance, effectively improving the operational reliability of generator units and reducing the risk of unplanned downtime. Summary of the Invention

[0005] To address the problems of limited monitoring methods, delayed fault warnings, and inability to achieve early and accurate diagnosis in existing technologies, this invention provides a method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion. The aim is to achieve real-time, comprehensive, and accurate assessment of the operating status of the slip ring carbon brush system and early fault warning. The technical solution includes:

[0006] S1: Synchronously collect multi-source heterogeneous monitoring data of the slip ring carbon brush system;

[0007] The multi-source heterogeneous monitoring data includes infrared thermal image data, video image data, electrical parameter data, and vibration data;

[0008] S2: Preprocess and align the multi-source heterogeneous monitoring data collected in step S1 in terms of time and space.

[0009] S3: Extract multimodal features related to the health status of carbon brushes from the various types of preprocessed data;

[0010] S4: Using a pre-trained multimodal feature fusion model, the multimodal features extracted in step S3 are fused and calculated to generate a comprehensive health index;

[0011] S5: Based on the comprehensive health index obtained in step S4 and the preset multi-level early warning logic, perform status assessment and fault early warning;

[0012] S6: Output fault warning information and diagnostic results, including assessment information on the type, location, and severity of the fault.

[0013] In step S1, synchronous acquisition includes:

[0014] Temperature field distribution data of the contact area between the slip ring and the carbon brush are collected by at least four infrared thermal imagers arranged on the left and right sides of the positive and negative slip rings.

[0015] Video image data of carbon brush wear morphology, spark generation, and surface oxidation state are acquired using a high-resolution industrial camera.

[0016] The waveform, effective value, and current distribution data of each parallel carbon brush branch are monitored by a current sensor.

[0017] Vibration signals of the carbon brush and brush holder are monitored using a vibration acceleration sensor.

[0018] The infrared thermal imager is equipped with a wide-angle lens with a field of view greater than 70°, and the vertical distance between the center of the lens and the center of the monitored slip ring area is not less than 30 cm.

[0019] The preprocessing in step S2 includes denoising and enhancing the infrared thermal image data and video image data, and filtering and standardizing the electrical parameter data and vibration data;

[0020] Infrared thermal image data is enhanced by a combination of median filtering and histogram equalization for noise reduction; video image data is enhanced by Gaussian filtering and Canny edge enhancement; electrical parameter data is enhanced by 50Hz notch filtering and moving average filtering to eliminate interference; and vibration data is enhanced by wavelet threshold filtering to purify the signal.

[0021] The modal features in step S3 include:

[0022] Temperature distribution features extracted from infrared thermal image data include: maximum temperature, average temperature, temperature gradient, and temperature distribution uniformity index.

[0023] Visual morphological features extracted from video image data include: carbon brush outline size extracted based on edge detection algorithm, wear amount calculated based on feature point matching algorithm, and frequency and area of ​​spark occurrence;

[0024] Electrical characteristic features extracted from electrical parameter data include: harmonic content and fluctuation amplitude of excitation current, and the degree of imbalance of current in each carbon brush branch.

[0025] Vibration features extracted from vibration data include the effective value (RMS), peak value, and amplitude at characteristic frequencies of vibration acceleration.

[0026] The multimodal feature fusion model mentioned in step S4 is a deep learning network model based on the attention mechanism. This model learns through training to automatically assign weights to different types of features and performs weighted fusion to output the comprehensive health index.

[0027] The fusion process is as follows:

[0028] S41) Weight calculation: A combined weighting strategy of "analytic hierarchy process + entropy weighting" is adopted;

[0029] Subjective weight Ws: Construct a hierarchical structure of target layer - criterion layer - indicator layer, invite power equipment operation and maintenance experts to fill in the judgment matrix, and after consistency test, obtain Ws=[0.32,0.21,0.25,0.22];

[0030] Objective weight Wo: Based on 500 sets of historical data, the information entropy of each modality feature is calculated and normalized to obtain Wo=[0.30,0.23,0.24,0.23];

[0031] The combined weight W is calculated using the formula W = 0.6Ws + 0.4Wo, resulting in W = [0.312, 0.218, 0.246, 0.224].

[0032] S42) Feature Fusion: Substitute the four-modal standardized feature vectors into formula F fusion =0.312F1+0.218F2+0.246F3+0.224F4, to obtain the fused feature vector; where F1: infrared temperature feature, F2: image morphology feature, F3: electrical characteristic feature, and F4: vibration feature;

[0033] S43) Health Index Generation: After the fused feature vector is mapped through a fully connected layer, the health index is generated using the formula HI=100×Sigmoid(Σfc_w). k *x k +b) Output a comprehensive health index in the range [0,100], where fc_w is the weight of the fully connected layer; fc_w k x is the k-th element of matrix fc_w, used for linear mapping of the k-th dimension of the fused feature vector; k is the k-th dimension element of the fused feature vector; b is the bias term of the fully connected layer, determined through training with labeled data.

[0034] The multi-level early warning logic in step S5 includes:

[0035] S51) Single-mode threshold judgment: Preset two-level thresholds for each modal feature, such as infrared T max Current imbalance, vibration RMS; any modal characteristic exceeding T2 directly triggers "serious fault warning", exceeding T1 but not reaching T2 enters secondary judgment;

[0036] S52) Multimodal fusion judgment: Combining the comprehensive health index with the LSTM temperature prediction model, health levels are classified:

[0037] Normal state: HI≥80 and LSTM predicted temperature≤105℃, no warning;

[0038] Mild warning: 60≤HI<80 or any mode exceeds T1, LSTM predicted temperature ≤110℃;

[0039] Moderate warning: 40≤HI<60 and LSTM predicted temperature>110℃;

[0040] Critical fault warning: HI < 40 or any single mode exceeds T2.

[0041] The fault types in step S6 include: faults based on overheating characteristics, faults based on electrical characteristics, faults based on mechanical and physical conditions, and faults based on comprehensive indicators; among which...

[0042] The faults based on overheating characteristics include: abnormal overheating of carbon brushes, localized overheating of slip rings, and overheating of connection points; wherein:

[0043] Abnormal overheating of carbon brushes is identified by infrared thermography data, which shows that the temperature of one or more carbon brushes is significantly higher than the set threshold or the absolute value exceeds the safety limit for other carbon brushes of the same electrode.

[0044] Localized overheating of the slip ring is identified by using infrared thermography to detect discontinuous high-temperature points or band-shaped high-temperature areas on the slip ring surface.

[0045] Overheating at connection points is detected by infrared thermal imaging data, which shows an abnormal increase in temperature at the connection between the carbon brush and the flexible connector or at the bus ring terminal.

[0046] The faults based on electrical characteristics include: severely uneven current distribution, abnormal contact resistance, and arcing; wherein:

[0047] The current distribution is severely uneven, as detected by monitoring electrical parameter data, showing that the difference in current values ​​among the parallel carbon brush branches exceeds the preset safety range.

[0048] Abnormal contact resistance, including high-resistance and low-resistance contacts, is diagnosed and differentiated indirectly by fusing current values ​​from electrical parameter data with temperature values ​​from infrared thermographic data.

[0049] Ring fire is identified by detecting continuous sparks or arcs through video image data and synchronously correlated with current spikes and waveform distortions appearing in electrical parameter data.

[0050] The faults based on mechanical and physical conditions include: excessive carbon brush wear, carbon brush jamming, and abnormal slip ring surface condition; among which:

[0051] Excessive carbon brush wear is determined by processing video image data through edge detection and feature point matching algorithms to quantify and calculate that the carbon brush length is below a preset limit value.

[0052] Carbon brush jamming is detected by video image data showing abnormal displacement of the carbon brush within the brush holder, and by combining this with vibration data to make a comprehensive judgment.

[0053] Abnormal slip ring surface condition is identified by video image data, which shows defects such as streaks, grooves, coating peeling, or large-area oxidation and discoloration on the slip ring surface.

[0054] The faults based on comprehensive index types include: spring pressure failure and cooling system performance degradation; wherein:

[0055] Spring pressure failure is diagnosed by directly monitoring the pressure value as too low or disappearing through pressure sensor data, or indirectly by combining the multi-modal characteristics of the carbon brush where the current value approaches zero and the temperature value is too low.

[0056] A decline in cooling system efficiency is diagnosed by analyzing the overall temperature baseline data to determine if it shows a slow and uniform upward trend; at the same time, if the current distribution data is uniform and unchanged, then a diagnosis is made.

[0057] The beneficial effects of this invention are as follows:

[0058] 1. Multi-dimensional perception: By integrating infrared thermal images, video images, electrical parameters and vibration data, multi-dimensional collaborative perception of the mechanical, electrical and thermal states of the slip ring carbon brush system is achieved, overcoming the limitations of single sensor monitoring.

[0059] 2. Early warning: The innovative temperature rise rate and trend alarm model can identify slow degradation processes and rapidly changing faults, significantly advancing the fault warning time (experiments show that it can be advanced by more than 60%), thus avoiding serious accidents.

[0060] 3. Precise diagnosis: Based on a deep learning-based multimodal feature fusion algorithm, it can accurately identify typical faults such as abnormal carbon brush heating, poor contact, and uneven wear, with a high diagnostic accuracy (up to 98.2%).

[0061] 4. High engineering practicality: The method and process are clear, and the installation and deployment scheme is specific (such as the infrared point deployment scheme). It is applicable to different types of steam turbine generators, and provides a reliable technical means for generator sets to carry out condition-based maintenance and predictive maintenance. It has significant engineering application value and economic value. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating an embodiment of the method for monitoring the condition of slip ring carbon brushes in a steam turbine generator based on multimodal data fusion according to the present invention. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the accompanying drawings.

[0064] like Figure 1 The embodiment of the present invention shown includes:

[0065] S1: Synchronously collect multi-source heterogeneous monitoring data of the slip ring carbon brush system, including infrared thermal image data, video image data, electrical parameter data and vibration data;

[0066] S2: Preprocess and align the multi-source heterogeneous monitoring data collected in step S1 with spatiotemporal data (align the time axis and the three-dimensional coordinate axis).

[0067] S3: Extract multimodal features related to the health status of carbon brushes from the various types of preprocessed data;

[0068] S4: Using a pre-trained multimodal feature fusion model, the multimodal features extracted in step S3 are fused and calculated to generate a comprehensive health index;

[0069] S5: Based on the comprehensive health index obtained in step S4 and the preset multi-level early warning logic, perform status assessment and fault early warning;

[0070] S6: Output fault warning information and diagnostic results, including assessment information on the type, location, and severity of the fault.

[0071] In this embodiment, step S1, synchronous acquisition specifically includes:

[0072] Temperature field distribution data of the contact area between the slip ring and the carbon brush are collected by at least four infrared thermal imagers arranged on the left and right sides of the positive and negative slip rings.

[0073] Video image data of carbon brush wear morphology, spark generation, and surface oxidation state are acquired using a high-resolution industrial camera.

[0074] The waveform, effective value, and current distribution data of each parallel carbon brush branch are monitored by a current sensor.

[0075] Vibration signals of the carbon brush and brush holder are monitored using a vibration acceleration sensor.

[0076] Specifically, the infrared thermal imager is equipped with a wide-angle lens with a field of view greater than 70°, and the vertical distance between the center of the lens and the center of the monitored slip ring area is not less than 30 centimeters.

[0077] In this embodiment, the preprocessing in step S2 includes denoising and enhancing the infrared thermal image data and video image data, and filtering and standardizing the electrical parameter data and vibration data. Specifically, the infrared thermal image data is denoised and enhanced using a combination of "median filtering + histogram equalization", the video image data is processed using "Gaussian filtering + Canny edge enhancement", the electrical parameter data is processed using "50Hz notch filtering + moving average filtering" to eliminate interference, and the vibration data is purified by wavelet threshold filtering.

[0078] In this embodiment, the modal features in step S3 include:

[0079] Temperature distribution features extracted from infrared thermal image data: including the highest temperature (T) max ), average temperature (T) avg ), temperature gradient (T) grad and temperature distribution uniformity index;

[0080] Visual morphological features extracted from video image data include: carbon brush outline size extracted based on edge detection algorithm, wear amount calculated based on feature point matching algorithm, and frequency and area of ​​spark occurrence;

[0081] Electrical characteristic features extracted from electrical parameter data include: harmonic content and fluctuation amplitude of excitation current, and the degree of imbalance of current in each carbon brush branch.

[0082] Vibration features extracted from vibration data include the effective value (RMS), peak value, and amplitude at characteristic frequencies of vibration acceleration.

[0083] In this embodiment, the multimodal feature fusion model described in step S4 is a deep learning network model based on an attention mechanism. This model learns through training to automatically assign weights to different types of features and performs weighted fusion to output the comprehensive health index. The specific fusion process is as follows:

[0084] S41) Weight calculation: A combined weighting strategy of "analytic hierarchy process (AHP) + entropy weighting" is adopted;

[0085] Subjective weights (Ws) in the multimodal feature fusion stage: Construct a hierarchical structure of "target layer (carbon brush health assessment) - criterion layer (infrared / image / electrical / vibration) - indicator layer (various modal features)", invite 3 power equipment operation and maintenance experts to fill in the judgment matrix, and after consistency test (CR < 0.1), we get Ws = [0.32, 0.21, 0.25, 0.22];

[0086] Objective weights (Wo) in the multimodal feature fusion stage: Based on 500 sets of historical data (including normal / minor fault / serious fault conditions), the information entropy of each modal feature (e.g., infrared temperature feature entropy = 0.89) is calculated and normalized to obtain Wo = [0.30, 0.23, 0.24, 0.23];

[0087] The combined weight (W) of the multimodal feature fusion stage is calculated according to the formula W=0.6Ws+0.4Wo, and finally W=[0.312,0.218,0.246,0.224];

[0088] S42) Feature Fusion: Substitute the four-modal normalized feature vectors (F1: infrared temperature feature, F2: image morphology feature, F3: electrical characteristic feature, F4: vibration feature) into formula F fusion =0.312F1+0.218F2+0.246F3+0.224F4, thus obtaining the fused feature vector;

[0089] S43) Health Index Generation: After the fused feature vector is mapped through a fully connected layer, the health index is generated using the formula HI=100×Sigmoid(Σfc_w). k*x k +b) Output the comprehensive health index in the range [0,100], where fc_w is the weight of the fully connected layer (applied to the mapping stage of the fused features); fc_w k The k-th element of this matrix is ​​used to perform a linear mapping on the k-th dimension of the fused feature vector; x k The fused feature vector F obtained in S42 fusion The k-th dimension element; b is the bias term of the fully connected layer, determined through training with labeled data.

[0090] In this embodiment, the multi-level early warning logic in step S5 includes a two-level mechanism: single-modal threshold judgment and multi-modal fusion judgment, specifically including:

[0091] S51) Single-mode threshold judgment (level 1 early warning): Preset two-level thresholds for each modal characteristic (early warning threshold T1, fault threshold T2), such as infrared T max (T1=110℃, T2=120℃), current imbalance (T1=8%, T2=12%), vibration RMS (T1=3.5mm / s, T2=4.5mm / s); any modal characteristic exceeding T2 directly triggers "serious fault warning", exceeding T1 but not reaching T2 enters secondary judgment;

[0092] S52) Multimodal Fusion Judgment (Level 2 Early Warning): Combining the comprehensive health index with the LSTM temperature prediction model (inputting historical temperature over the past hour, time step = 5 minutes), the health level is classified:

[0093] Normal state: HI≥80 and LSTM predicted temperature≤105℃, no warning;

[0094] Mild warning: 60≤HI<80 or any mode exceeds T1, LSTM predicted temperature ≤110℃;

[0095] Moderate warning: 40≤HI<60 and LSTM predicted temperature>110℃;

[0096] Critical fault warning: HI < 40 or any single mode exceeds T2.

[0097] In this embodiment, the types of faults in step S6 include: faults based on overheating characteristics, faults based on electrical characteristics, faults based on mechanical and physical conditions, and faults based on comprehensive indicators.

[0098] Specifically, faults based on overheating characteristics include: abnormal overheating of carbon brushes, localized overheating of slip rings, and overheating of connection points.

[0099] The abnormal overheating of the carbon brush refers to the temperature of one or more carbon brushes that is significantly higher than the set threshold or absolute value of other carbon brushes of the same electrode, exceeding the safety limit, as identified by infrared thermal imaging data.

[0100] The aforementioned localized overheating of the slip ring refers to the appearance of discontinuous high-temperature points or band-shaped high-temperature areas on the surface of the slip ring, as identified by infrared thermal imaging data.

[0101] The term "overheating of connection point" refers to an abnormal increase in temperature at the connection between the carbon brush and the flexible connector or at the bus ring terminal, as identified by infrared thermal imaging data.

[0102] Specifically, faults based on electrical characteristics include: severely uneven current distribution, abnormal contact resistance, and arcing; among which:

[0103] The severe uneven current distribution refers to the current value difference of each parallel carbon brush branch exceeding the preset safety range as detected by electrical parameter data.

[0104] The abnormal contact resistance, including high-resistivity contact and low-resistivity contact, is indirectly diagnosed and distinguished by fusing and analyzing the current value in the electrical parameter data and the temperature value in the infrared thermographic data.

[0105] The so-called ring fire refers to the identification of continuous sparks or arcs through video image data, and the synchronous correlation with current spikes and waveform distortions appearing in electrical parameter data.

[0106] Specifically, faults based on mechanical and physical conditions include: excessive carbon brush wear, carbon brush jamming, and abnormal slip ring surface condition; among which:

[0107] The excessive wear of the carbon brush refers to the carbon brush length being lower than a preset limit value, which is calculated quantitatively by processing video image data through edge detection and feature point matching algorithms.

[0108] The aforementioned carbon brush jamming refers to the discovery of abnormal displacement of the carbon brush within the brush holder through video image data, and the comprehensive judgment based on the presence of specific high-frequency components in vibration data.

[0109] The abnormal surface condition of the slip ring refers to the presence of defects such as streaks, grooves, coating peeling, or large-area oxidation and discoloration on the slip ring surface, as identified by video image data.

[0110] Specifically, faults based on comprehensive indicator types include: spring pressure failure and decreased cooling system efficiency; among which:

[0111] The spring pressure failure refers to the direct detection of a low or absent pressure value through pressure sensor data, or the indirect diagnosis by integrating the multi-modal characteristics of the carbon brush where the current value approaches zero and the temperature value is low.

[0112] The decline in cooling system efficiency is defined as a slow and uniform upward trend observed in the overall temperature baseline data; if the current distribution data remains uniform and unchanged, then a diagnosis is made.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the condition and providing early warning of faults in the carbon brushes of a steam turbine generator slip ring based on multimodal data fusion, characterized in that, include: S1: Synchronously collect multi-source heterogeneous monitoring data of the slip ring carbon brush system; The multi-source heterogeneous monitoring data includes infrared thermal image data, video image data, electrical parameter data, and vibration data; S2: Preprocess and align the multi-source heterogeneous monitoring data collected in step S1 in terms of time and space. S3: Extract multimodal features related to the health status of carbon brushes from the various types of preprocessed data; S4: Using the trained multimodal feature fusion model, perform fusion calculation on the multimodal features extracted in step S3 to generate a comprehensive health index; S5: Based on the comprehensive health index obtained in step S4 and the preset multi-level early warning logic, perform status assessment and fault early warning; S6: Output fault warning information and diagnostic results, including assessment information on the type, location, and severity of the fault.

2. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 1, is characterized in that... In step S1, synchronous acquisition includes: Temperature field distribution data of the contact area between the slip ring and the carbon brush are collected by at least four infrared thermal imagers arranged on the left and right sides of the positive and negative slip rings. Video image data of carbon brush wear morphology, spark generation, and surface oxidation state are acquired using a high-resolution industrial camera. The waveform, effective value, and current distribution data of each parallel carbon brush branch are monitored by a current sensor. Vibration signals of the carbon brush and brush holder are monitored using a vibration acceleration sensor.

3. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 2, is characterized in that... The infrared thermal imager is equipped with a wide-angle lens with a field of view greater than 70°, and the vertical distance between the center of the lens and the center of the monitored slip ring area is not less than 30 cm.

4. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 1, is characterized in that... The preprocessing in step S2 includes denoising and enhancing the infrared thermal image data and video image data, and filtering and standardizing the electrical parameter data and vibration data.

5. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 4, is characterized in that... Infrared thermal image data is enhanced by a combination of "median filtering + histogram equalization" for noise reduction; video image data is enhanced by "Gaussian filtering + Canny edge enhancement"; electrical parameter data is enhanced by "50Hz notch filtering + moving average filtering" to eliminate interference; and vibration data is enhanced by wavelet threshold filtering to purify the signal.

6. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 1, is characterized in that... The modal features in step S3 include: Temperature distribution features extracted from infrared thermal image data include: maximum temperature, average temperature, temperature gradient, and temperature distribution uniformity index. Visual morphological features extracted from video image data include: carbon brush outline size extracted based on edge detection algorithm, wear amount calculated based on feature point matching algorithm, and frequency and area of ​​spark occurrence; Electrical characteristic features extracted from electrical parameter data include: harmonic content and fluctuation amplitude of excitation current, and the degree of imbalance of current in each carbon brush branch. Vibration features extracted from vibration data include the effective value (RMS), peak value, and amplitude at characteristic frequencies of vibration acceleration.

7. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 1, is characterized in that... The multimodal feature fusion model mentioned in step S4 is a deep learning network model based on the attention mechanism. This model learns through training to automatically assign weights to different types of features and performs weighted fusion to output the comprehensive health index.

8. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion according to claim 1, characterized in that, The fusion process is as follows: S41) Weight calculation: A combined weighting strategy of "analytic hierarchy process + entropy weighting" is adopted; Subjective weight Ws: Construct a hierarchical structure of target layer - criterion layer - indicator layer, invite power equipment operation and maintenance experts to fill in the judgment matrix, and after consistency test, obtain Ws=[0.32,0.21,0.25,0.22]; Objective weight Wo: Based on 500 sets of historical data, the information entropy of each modality feature is calculated and normalized to obtain Wo=[0.30,0.23,0.24,0.23]; Combined weight W: Calculated according to the formula W=0.6Ws+0.4Wo, the final result is W=[0.312,0.218,0.246,0.224]. Here, W is the modal weight in the multimodal feature fusion stage, which is used to allocate the importance of the four types of features, namely infrared temperature, image morphology, electrical characteristics and vibration, during fusion. S42) Feature Fusion: Substitute the four-modal standardized feature vectors into formula F fusion =0.312F1+0.218F2+0.246F3+0.224F4, to obtain the fused feature vector; where F1: infrared temperature feature, F2: image morphology feature, F3: electrical characteristic feature, and F4: vibration feature; S43) Health Index Generation: After the fused feature vector is mapped through a fully connected layer, the health index is generated using the formula HI=100×Sigmoid(Σfc_w). k *x k +b) Output a comprehensive health index in the range [0,100], where fc_w is the weight of the fully connected layer; fc_w k x is the k-th element of matrix fc_w, used for linear mapping of the k-th dimension of the fused feature vector; k is the k-th dimension element of the fused feature vector; b is the bias term of the fully connected layer, determined through training with labeled data.

9. The method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion according to claim 1, characterized in that, The multi-level early warning logic in step S5 includes: S51) Single-mode threshold judgment: Preset two-level thresholds for each modal feature, such as infrared T max Current imbalance, vibration RMS; any modal characteristic exceeding T2 directly triggers "serious fault warning", exceeding T1 but not reaching T2 enters secondary judgment; S52) Multimodal fusion judgment: Combining the comprehensive health index with the LSTM temperature prediction model, health levels are classified: Normal state: HI≥80 and LSTM predicted temperature≤105℃, no warning; Mild warning: 60≤HI<80 or any mode exceeds T1, LSTM predicted temperature ≤110℃; Moderate warning: 40≤HI<60 and LSTM predicted temperature>110℃; Critical fault warning: HI < 40 or any single mode exceeds T2.

10. A method for monitoring the condition and providing early warning of faults in the slip ring carbon brush of a steam turbine generator based on multimodal data fusion, as described in claim 1, is characterized in that... The fault types in step S6 include: faults based on overheating characteristics, faults based on electrical characteristics, faults based on mechanical and physical conditions, and faults based on comprehensive indicators; among which... The faults based on overheating characteristics include: abnormal overheating of carbon brushes, localized overheating of slip rings, and overheating of connection points; wherein: Abnormal overheating of carbon brushes is identified by infrared thermography data, which shows that the temperature of one or more carbon brushes is significantly higher than the set threshold or the absolute value exceeds the safety limit for other carbon brushes of the same electrode. Localized overheating of the slip ring is identified by using infrared thermography to detect discontinuous high-temperature points or band-shaped high-temperature areas on the slip ring surface. Overheating at connection points is detected by infrared thermal imaging data, which shows an abnormal increase in temperature at the connection between the carbon brush and the flexible connector or at the bus ring terminal. The faults based on electrical characteristics include: severely uneven current distribution, abnormal contact resistance, and arcing; wherein: The current distribution is severely uneven, as detected by monitoring electrical parameter data, showing that the difference in current values ​​among the parallel carbon brush branches exceeds the preset safety range. Abnormal contact resistance, including high-resistance and low-resistance contacts, is diagnosed and differentiated indirectly by fusing current values ​​from electrical parameter data with temperature values ​​from infrared thermographic data. Ring fire is identified by detecting continuous sparks or arcs through video image data and synchronously correlated with current spikes and waveform distortions appearing in electrical parameter data. The faults based on mechanical and physical conditions include: excessive carbon brush wear, carbon brush jamming, and abnormal slip ring surface condition; among which: Excessive carbon brush wear is determined by processing video image data through edge detection and feature point matching algorithms to quantify and calculate that the carbon brush length is below a preset limit value. Carbon brush jamming is detected by video image data showing abnormal displacement of the carbon brush within the brush holder, and by combining this with vibration data to make a comprehensive judgment. Abnormal slip ring surface condition is identified by video image data, which shows defects such as streaks, grooves, coating peeling, or large-area oxidation and discoloration on the slip ring surface. The faults based on comprehensive index types include: spring pressure failure and cooling system performance degradation; wherein: Spring pressure failure is diagnosed by directly monitoring the pressure value as too low or disappearing through pressure sensor data, or indirectly by combining the multi-modal characteristics of the carbon brush where the current value approaches zero and the temperature value is too low. A decline in cooling system efficiency is diagnosed by analyzing the overall temperature baseline data to determine if it shows a slow and uniform upward trend; at the same time, if the current distribution data is uniform and unchanged, then a diagnosis is made.

Images