Hydropower station high-altitude equipment fault intelligent identification system based on multi-sensor fusion

Abstract

The invention relates to the technical field of power equipment monitoring, and discloses a hydropower station high-altitude equipment fault intelligent identification system based on multi-sensor fusion, which collects multi-modal data in real time and evaluates data quality by deploying multi-source sensors at key parts of high-altitude equipment. Extracting multi-scale features of each modal, performing normalization processing, calculating a fusion weight based on feature saliency and data credibility, and performing weighted fusion and dimension reduction on the multi-modal features; based on the fusion feature vector, intelligent matching analysis of fault features and intelligent identification of fault types are carried out; a fault identification result is obtained; in addition, the system also comprises safety monitoring of overhead working personnel, and realizes closed-loop management from fault identification to safety maintenance. According to the invention, early weak faults can be accurately identified, and the safe operation level of equipment and the intelligent degree of operation safety management are improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and more specifically, to an intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion. Background Technology

[0002] Hydropower station high-altitude equipment includes critical facilities such as bridge cranes, high-voltage busbar systems, main transformer roof cooling devices, and powerhouse roof ventilation systems. These devices are typically located tens of meters above the ground, operating under complex conditions of high humidity, large temperature differences, and strong electromagnetic interference. Failures can not only affect the power station's power generation efficiency but also potentially cause major safety accidents and economic losses. Therefore, real-time and accurate fault monitoring and diagnosis of hydropower station high-altitude equipment is of paramount importance.

[0003] Traditional equipment fault detection mainly relies on manual inspection and single-sensor monitoring, which has many technical limitations. Manual inspection requires scaffolding or the use of lifting platforms to approach equipment at heights, which is highly dangerous and cannot be performed at close range while the equipment is running, resulting in a significant delay in fault detection. While single-sensor monitoring can achieve automated monitoring, it can only obtain information from a single dimension, is easily affected by environmental interference, and has a high false alarm and false negative rate in the complex working conditions of high-altitude environments, and it is difficult to accurately determine the type and severity of the fault.

[0004] Existing multi-sensor fusion technologies still have shortcomings in terms of data fusion depth, environmental adaptability, and real-time performance. Current methods mainly employ shallow fusion strategies, lacking a dynamic evaluation mechanism for sensor data quality and reliability. When some sensors malfunction, they cannot adaptively adjust fusion weights. Furthermore, existing fault detection systems primarily focus on equipment status monitoring, lacking real-time monitoring and intelligent management of the safety status of personnel working at heights during fault repair, thus failing to achieve integrated management from fault identification to safe maintenance. Summary of the Invention

[0005] This invention provides an intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion, which solves the technical problems of poor equipment fault detection and inadequate supervision of high-altitude operations in related technologies.

[0006] This invention provides an intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion, comprising: Data acquisition and evaluation module: Based on multi-source sensors deployed in high-altitude equipment, raw multimodal data is acquired; the quality of the raw multimodal data is evaluated to obtain a repaired multimodal dataset; The spatiotemporal alignment module uses the repaired multimodal dataset to perform spatiotemporal alignment and adaptive filtering of the multimodal data to obtain the multimodal dataset. The feature extraction module, based on a multimodal dataset, extracts features from vibration signals, infrared images, and acoustic signals to obtain multimodal feature vectors. The feature fusion module, based on multimodal feature vectors, employs confidence-aware adaptive feature fusion to obtain fused feature vectors with confidence labels. The fault identification module performs intelligent matching analysis of fault features and intelligent identification of fault types based on fused feature vectors with confidence annotations, thereby obtaining fault identification results. The intelligent early warning module predicts fault trends based on the final fault identification results and the sequence of fused feature vectors; it also generates tiered early warning information and structured fault diagnosis reports.

[0007] In a preferred embodiment, acquiring raw multimodal data based on multi-source sensors deployed in high-altitude equipment includes: Based on vibration sensors, infrared thermal imagers, acoustic sensors, and current sensors deployed in key parts of high-altitude equipment, a synchronous trigger acquisition mechanism is adopted to start data acquisition from each sensor under a unified time reference, thereby obtaining the original vibration signal, original infrared image, original acoustic signal, and original current signal.

[0008] In a preferred embodiment, the quality assessment of the raw multimodal data includes: Based on the original vibration signal, a data integrity detection method is used to obtain vibration data quality labels and corresponding credibility scores; Based on the original infrared image, an image quality assessment method is used to obtain image data quality labels and corresponding confidence scores; Based on the original acoustic signal and the original current signal, statistical analysis methods are used to obtain the acoustic data reliability score and the current data reliability score.

[0009] In a preferred embodiment, the spatiotemporal alignment of the multimodal data using the repaired multimodal dataset includes: Using the obtained repaired multimodal data, the time reference is uniformly aligned to obtain the time offset parameters of each sensor; Based on real-time monitoring of high-altitude environmental parameters of hydropower stations, a database of operating condition characteristics of high-altitude equipment in hydropower stations is established, and operating condition parameters are collected; an operating condition influence relationship mapping table is constructed to establish the correspondence between operating condition parameters and modal characteristics; a fault condition correlation analysis model is established to study the differences in sensitivity of various modal sensors to fault characteristics under different operating conditions.

[0010] In a preferred embodiment, the feature extraction from the vibration signal, infrared image, and acoustic signal includes: Using the obtained filtered vibration signal, the vibration signal is decomposed into 8 different frequency bands according to the frequency range. For each frequency band, the signal is separated and the energy ratio characteristics, peak characteristics and shape characteristics are calculated and arranged to form vibration characteristics. Using the obtained filtered infrared image, the temperature distribution features are extracted and analyzed, and the texture features of the image are extracted and combined to obtain the image features; Using the obtained filtered acoustic signal, the acoustic signal is converted to the frequency domain for analysis. According to the operating characteristics of the equipment, the frequency range is divided into three frequency bands: low, medium and high. In each frequency band, the energy value, main frequency and spectral peak value are calculated and combined to obtain the acoustic characteristics.

[0011] In a preferred embodiment, the confidence-aware adaptive feature fusion includes: By collecting equipment operating parameters through a real-time monitoring system, a mapping table of operating condition influence relationships is constructed, and a three-layer progressive significance evaluation architecture is established. The significance of each modal feature is dynamically evaluated by combining real-time operating parameters, and the significance score of each modality with operating condition adaptation is obtained. Using the obtained modality significance scores and data credibility scores, a standardized fusion weight set is obtained through the calculation and analysis of fusion weights; Using the obtained standardized fusion weight set and standardized multimodal feature vector set, a preliminary fusion feature vector is obtained through weighted fusion processing of the features; Using the obtained preliminary fusion feature vector, the dimensionality-reduced fusion feature vector is obtained through feature dimension optimization.

[0012] In a preferred embodiment, the fused feature vector based on confidence labeling is used for intelligent matching analysis of fault features and intelligent identification of fault types; the resulting fault identification includes: Based on fused feature vectors, an environment-adaptive attention mechanism is constructed to calculate environmental impact factors by analyzing the differences between environmental parameters and historical data; a temporal memory enhancement module is constructed to learn the temporal patterns of fault development through a long short-term memory network; and the probability distribution vectors of each fault type are output. A multi-level decision fusion architecture is established using the probability distribution vector of fault types; a comprehensive confidence score is calculated by comprehensively considering the probability distribution concentration, the degree of influence of environmental correction, and the consistency of historical trends; and a comprehensive identification result is obtained. Based on the comprehensive identification results, reliability analysis is conducted in four dimensions: data quality, algorithm performance, environmental adaptability, and historical consistency. The weight allocation of each dimension is dynamically adjusted and the reliability judgment threshold is determined, and detailed reliability assessment results are output. Based on the reliability assessment results, corresponding decision-making strategies are adopted according to the reliability level. For high reliability level, the identification result is directly output; for medium reliability level, the intelligent collaborative verification mechanism is activated; and for low reliability level, the comprehensive in-depth analysis mechanism is activated to output the final fault identification result.

[0013] In a preferred embodiment, the fault feature library includes: Feature vectors for bearing pitting, gear wear, poor electrical contact, severe bearing wear, mechanical loosening, insulation aging, and normal state are generated. The standard feature vectors for each fault type are standardized to maintain the same dimensions as the current feature vectors.

[0014] In a preferred embodiment, the fault trend prediction based on the final fault identification result and the sequence of fused feature vectors includes: Analyze the changing trends of fault characteristics over time, calculate the rate and direction of change of characteristics; establish a fault development prediction model to predict the degree of fault development in the future; and combine the fault type conversion rules to predict possible new fault types. Establish a fault severity assessment system, classifying faults into four levels: minor, moderate, severe, and critical. Multi-level early warning generation; generates early warning information at corresponding levels, including the warning level, warning content, and recommended handling measures; Fault diagnosis report generation: Automatically generates a complete fault diagnosis report that includes fault type, location of occurrence, severity, development trend, and handling suggestions.

[0015] In a preferred embodiment, the intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion further includes: Based on the fault diagnosis report and early warning information, analyze the location of the faulty equipment, the fault type, the severity level and maintenance requirements to obtain a list of high-altitude operation tasks; Workers wear smart safety belts with integrated sensors to monitor their physiological status data and receive abnormal warning information in real time; The location and movements of the workers are tracked in real time by a mobile camera worn by the workers, and visual monitoring data of the workers is obtained. Identify non-standard operating behaviors based on visual monitoring data, including not wearing safety equipment and using tools in violation of regulations; Based on non-standard operating behaviors and safety risk levels, generate different levels of safety alarms and response measures; push safety monitoring information and alarm notifications to management personnel; Conduct safety assessments and record the high-altitude operation process, and generate an operation safety assessment report.

[0016] The beneficial effects of this invention are as follows: By constructing a credibility-aware adaptive multimodal feature fusion mechanism, multidimensional data from vibration, infrared, acoustic, and current sensors are deeply fused. The system assesses the quality of data from each sensor and dynamically adjusts the fusion weights to ensure that high-quality data dominates the fusion process. Through multi-scale feature extraction and dimensionality reduction optimization, a fused feature vector is formed. This fused feature vector is then used for intelligent matching analysis of fault features and intelligent identification of fault types, resulting in improved fault identification accuracy and robustness.

[0017] By integrating fault diagnosis with high-altitude work safety monitoring, a comprehensive safety monitoring system has been constructed, encompassing key components such as intelligent safety belt physiological status monitoring, mobile PTZ camera visual monitoring, identification of non-standard work behaviors, and multi-level alarm response. The intelligent safety belt monitors the worker's position, posture, heart rate, and other physiological states in real time, while the mobile PTZ camera uses a deep learning model to identify the wearing status of safety equipment in real time. The system has established comprehensive criteria for judging non-standard work behaviors and ensures timely handling of safety incidents through a three-level alarm mechanism, achieving intelligent management of the entire process from fault identification to safety maintenance. Attached Figure Description

[0018] Figure 1 This is a block diagram of the intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion, according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses an intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion, such as... Figure 1 As shown, it includes: Data acquisition and evaluation module: Based on multi-source sensors deployed in high-altitude equipment, raw multimodal data is acquired; the quality of the raw multimodal data is evaluated to obtain a repaired multimodal dataset; Specifically, the following steps are included: S101, multi-sensor synchronous data acquisition; based on vibration sensors, infrared thermal imagers, acoustic sensors, and current sensors deployed in key parts of high-altitude equipment, a synchronous trigger acquisition mechanism is adopted to start data acquisition of each sensor under a unified time reference, obtaining raw vibration signals, raw infrared images, raw acoustic signals, and raw current signals; among them, key parts include the motor bearing housing of the trolley traveling mechanism of the bridge crane, the reducer housing of the trolley traveling mechanism, the drum bearing housing of the main hoisting mechanism, the insulator string hardware connection point and the end of the busbar expansion joint of the 500kV high-voltage busbar system, the fan motor base and cooler tube bundle end plate of the cooling device on the top of the main transformer, and the motor base and impeller hub center of the axial flow fan on the top of the plant.

[0021] S102, Vibration Data Quality Detection; Based on the original vibration signal, a data integrity detection method is adopted. The specific detection steps are as follows: Set a fixed-length time window of 10 seconds, corresponding to a theoretical number of sampling points equal to the sampling rate multiplied by 10; Count the actual number of data points received within each time window; Calculate the ratio of the actual number of sampling points to the theoretical number of sampling points; Compare this ratio with a preset threshold of 0.95. If the ratio is greater than 0.95, it is marked as complete data; otherwise, it is marked as missing data; Calculate the ratio of the effective signal value to the effective noise value as the signal-to-noise ratio (SNR). Normalize the SNR to between 0 and 1 as the vibration data credibility score, and obtain the vibration data quality label and the corresponding credibility score.

[0022] S103, Infrared Image Quality Assessment; Based on the original infrared image, an image quality assessment method is adopted. The specific assessment steps are as follows: Calculate the image sharpness index by analyzing the temperature gradient changes of adjacent pixels in the image and calculating the average value of the gradient changes as the sharpness value; calculate the image contrast index by analyzing the difference between the highest and lowest temperatures in the image and normalizing the temperature difference to between 0 and 1 as the contrast value; set the sharpness weight to 0.6 and the contrast weight to 0.4, and calculate the weighted sum; compare the weighted sum with a preset threshold of 0.7. If the weighted sum is greater than 0.7, it is marked as a high-quality image; otherwise, it is marked as a low-quality image; use the weighted sum value directly as the image data credibility score to obtain the image data quality label and the corresponding credibility score.

[0023] S104, Acoustic and Current Data Reliability Assessment: Based on the original acoustic signal and the original current signal, statistical analysis methods are used to calculate the degree of deviation between the effective value of the acoustic signal and the current signal and the preset normal range. The degree of deviation is normalized to between 0 and 1 as the reliability score, and the acoustic data reliability score and the current data reliability score are obtained.

[0024] S105, Multimodal Data Repair and Output: Based on the quality labels and credibility scores of each sensor data, a data repair strategy is adopted. For data segments marked as missing, linear interpolation is performed to fill them in according to the normal data values ​​of the sensor before and after the time. For data marked as low quality, outliers are replaced according to the statistical distribution of the sensor's historical data. The repaired vibration signal, infrared image, acoustic signal and current signal, as well as the corresponding credibility score set are obtained.

[0025] The spatiotemporal alignment module uses the repaired multimodal dataset to perform spatiotemporal alignment and adaptive filtering of the multimodal data to obtain the multimodal dataset. Specifically, the following steps are included: S201, Time Base Unification Processing: Using the obtained repaired multimodal data, a unified time base alignment is performed. The vibration sensor with the highest sampling rate is selected as the time base, and the timestamps of other sensors are compared with this base. The time offset of each sensor relative to the base is determined by calculating the timestamp difference: the time offset of the infrared image sensor, the time offset of the acoustic sensor, and the time offset of the current sensor. For cases where the timestamp format is inconsistent, all timestamps are first converted to a unified microsecond-level precision format. When abnormal timestamp jumps are detected, the average value of preceding and following data points is used for correction. Finally, an accurate time offset parameter table of each sensor relative to the vibration sensor base is generated. The time offset parameters of each sensor are output, including the infrared image offset, acoustic signal offset, and current signal offset.

[0026] S202, Multimodal Data Time Alignment: Using the obtained time offset parameters, time axis alignment is performed on the data for each modality. Infrared image data is shifted along the time axis according to its time offset to ensure synchronization between image acquisition and vibration data. Acoustic signals are resampled according to the time offset, and the time positions of the sampling points are adjusted to match the reference time. Current signals are time-synchronized according to the time offset to ensure accurate correspondence of current measurement times. Interpolation methods are used to supplement any missing data points during the alignment process, ensuring complete synchronization of each modality's data on the time axis. The alignment effect is verified by checking whether the timestamps of the data from each sensor at the same moment are consistent. The complete time-aligned multimodal dataset is output, including synchronized vibration signals, infrared images, acoustic signals, and current signals.

[0027] S203, Vibration Signal Noise Filtering: Noise filtering is performed using time-aligned vibration signals and vibration data reliability scores. Different filtering intensities are selected based on the reliability score: light filtering is used when the reliability score is above 0.8 to preserve signal details; strong filtering is used when the reliability score is below 0.5 to suppress noise interference. Different filtering intensities are achieved by adjusting the filter's cutoff frequency; higher cutoff frequencies are used for high-reliability data, and lower cutoff frequencies are used for low-reliability data. The filtered signal undergoes quality checks to ensure that effective signal components are preserved while noise is effectively suppressed. The noise-filtered vibration signal data is then output.

[0028] S204, Infrared Image Noise Processing: This section describes image noise processing using time-aligned infrared images and image data confidence scores. Appropriate processing strategies are selected based on the image confidence score: edge-preserving processing is used for high-confidence images to maintain the clarity of temperature boundaries; smoothing processing is used for low-confidence images to eliminate noise and artifacts. By analyzing local temperature gradient changes in the image, true temperature boundaries and noise regions are identified. Smoothing processing is applied to noise regions to reduce noise impact, while preserving the original temperature change characteristics in boundary regions. The processed image undergoes quality assessment to ensure the accuracy of temperature information and image clarity. The noise-processed infrared image data is then output.

[0029] S205 is an adaptive acoustic-current joint filtering system for high-altitude environments at hydropower stations. It utilizes time-aligned acoustic and current signals and their corresponding reliability scores, combined with real-time monitoring parameters of the high-altitude environment at the hydropower station, to perform adaptive joint noise filtering. Key environmental parameters such as electromagnetic interference intensity, ambient humidity, temperature change rate, and wind speed are collected in real time. An environmental impact weighting mechanism is constructed, setting dynamic weight coefficients for each environmental parameter, and converting acoustic and current signals to the frequency domain for analysis. A three-layer progressive adaptive filter architecture is designed: the first layer is a strong electromagnetic interference suppression filter, employing a multi-frequency notch filter bank to suppress power frequency and harmonic interference; the second layer is a humidity-related signal attenuation compensation filter, dynamically adjusting signal gain parameters; and the third layer is a temperature gradient noise filter, suppressing equipment thermal expansion noise caused by temperature changes. Through this three-layer joint filtering process, precise suppression of interference in the complex high-altitude environment of the hydropower station is achieved, while retaining effective frequency information related to equipment faults. The output is acoustic and current signal data processed by the environmental adaptive joint filtering.

[0030] Finally, all the multimodal data that have undergone spatiotemporal alignment and noise filtering are integrated to form a high-quality dataset under a unified time reference. The output is complete filtered multimodal data, including vibration signals, infrared images, acoustic signals, and current signals.

[0031] The feature extraction module, based on a multimodal dataset, extracts features from vibration signals, infrared images, and acoustic signals to obtain multimodal feature vectors. Specifically, the following steps are included: S301, Vibration signal feature extraction; using the obtained filtered vibration signal, vibration features are extracted and analyzed. The vibration signal was decomposed into eight different frequency bands according to its frequency range. The specific steps were as follows: the total frequency range was determined to be 0 to 4000 Hz, and then divided into equal intervals according to the operating characteristics of the equipment; the frequency range was divided into eight frequency bands: 0-500 Hz, 500-1000 Hz, 1000-1500 Hz, 1500-2000 Hz, 2000-2500 Hz, 2500-3000 Hz, 3000-3500 Hz, and 3500-4000 Hz; digital filtering was used to separate the signal in each frequency band, resulting in eight independent frequency band signals; three key characteristics were calculated for each frequency band: the energy ratio characteristic was obtained by calculating the ratio of the energy in that frequency band to the total energy, the peak value characteristic was obtained by calculating the ratio of the peak value to the average value of the signal, and the shape characteristic was obtained by calculating the kurtosis coefficient of the signal distribution; these three types of characteristic values ​​of the eight frequency bands were arranged and combined in sequence to form a vibration characteristic data set containing 24 values. The output vibration feature vector contains information on energy distribution, peak characteristics, and shape characteristics of each frequency band.

[0032] S302, Infrared Image Feature Extraction: Using the filtered infrared image, temperature distribution features are extracted and analyzed. Key monitoring areas of the equipment in the image are identified to determine the key areas for temperature analysis. Within these areas, basic temperature distribution characteristics are calculated: average temperature reflects the overall thermal state, maximum temperature reflects local hotspots, and the temperature gradient is obtained by calculating the temperature differences between adjacent areas, reflecting the uniformity of the temperature distribution. Simultaneously, texture features are extracted from the image. Contrast and uniformity are calculated by analyzing the regularity of the image's grayscale distribution, reflecting the spatial distribution characteristics of the temperature field. The five temperature features and two texture features are combined and organized. The output image feature vector contains both temperature and texture distribution features.

[0033] S303, Acoustic Signal Feature Extraction; This function extracts and analyzes acoustic features from the filtered acoustic signal. The acoustic signal is converted to the frequency domain for analysis. Based on the equipment's operating characteristics, the frequency range is divided into three key bands: low frequency (0-1kHz), primarily reflecting the equipment's basic operating state; mid frequency (1-5kHz), primarily reflecting the equipment's mechanical operating state; and high frequency (5-20kHz), primarily reflecting the equipment's abnormal vibration state. Within each band, three types of feature parameters are calculated: energy value reflecting the sound intensity, dominant frequency reflecting the main frequency components, and the number of spectral peaks reflecting the frequency complexity. These three types of feature values ​​are then combined and arranged. The resulting acoustic feature vector contains information on the energy distribution, frequency characteristics, and complexity of each band.

[0034] S304, Current Signal Feature Extraction; This function uses the filtered current signal to extract and analyze its electrical features. The basic electrical parameters of the current signal are calculated: the effective value is obtained by calculating the root mean square value of the current over one cycle, reflecting the overall magnitude of the current; the peak value is obtained by finding the maximum instantaneous value of the current, reflecting its impact characteristics; and the harmonic content is obtained by analyzing the proportion of each harmonic component in the current, reflecting its quality. Simultaneously, the rate of change of the current is calculated by analyzing the trend of current change over time, reflecting the stability of the equipment load. These four electrical feature parameters are then combined and organized. The output current feature vector contains information on current magnitude, impact characteristics, quality, and stability.

[0035] S305, Multimodal Feature Standardization Process: This process standardizes the obtained vibration, image, acoustic, and current feature vectors. Since different modes have different numerical ranges and dimensions, all features need to be standardized to the same numerical range. Each dimension of each feature vector is processed separately: the minimum and maximum values ​​of that dimension in historical data are determined as standardization boundaries. The offset of the current feature value relative to the minimum value is calculated, and this offset is divided by the feature's numerical range to obtain a standardized value between 0 and 1. This method ensures that all modal features have the same numerical scale and comparability. The standardized modal feature vectors are then combined and organized. The output is a set of standardized multimodal feature vectors, including standardized feature vectors from vibration, image, acoustic, and current features.

[0036] The feature fusion module, based on multimodal feature vectors, employs confidence-aware adaptive feature fusion to obtain fused feature vectors with confidence labels. Specifically, the following steps are included: S401, Dynamic Feature Assessment of Operating Condition Sensing for High-Altitude Equipment in Hydropower Stations; Utilizing the obtained standardized multimodal feature vector set, combined with real-time operating condition parameters of the high-altitude equipment in hydropower stations, to perform adaptive modal feature analysis; Establishing a database of operating condition features for high-altitude equipment in hydropower stations, collecting key operating condition parameters such as equipment load rate, continuous operating time, comprehensive environmental stress index, and historical trends in equipment health; Constructing an operating condition influence relationship mapping table to establish the correspondence between operating condition parameters and modal features; Establishing a fault condition correlation analysis model to study the differences in sensitivity of various modal sensors to fault features under different operating conditions; Specifically, collecting current operating condition parameter combinations; Calculating the adjustment factor for each mode through the operating condition influence relationship mapping table; Obtaining the basic correlation coefficient between each modal feature vector and the fault feature database through similarity calculation; Introducing an operating condition adjustment mechanism to intelligently adjust the basic correlation coefficient based on current equipment load rate, operating time, environmental stress, and other operating condition parameters, through... The operating condition influence weight matrix dynamically compensates for the sensitivity of features under different operating conditions, yielding correlation coefficients adjusted for operating conditions. A three-layer progressive significance assessment architecture is established: the first layer, basic significance assessment, evaluates the basic significance of features by analyzing the variance contribution, information entropy, and matching degree with the fault feature database of each modality's feature vector; the second layer, operating condition adaptability assessment, queries the operating condition influence relationship mapping table based on the current operating condition parameter combination, analyzes the sensitivity change law of each mode under specific operating conditions, and performs operating condition adaptability correction on the basic significance; the third layer, temporal consistency assessment, verifies and corrects the temporal stability of the significance score by analyzing the consistency between the current significance score and historical time series data, combined with the equipment health change trend and fault evolution mode; the results of the three layers are combined, and the weights of each layer are determined according to the credibility of the assessment level and the historical verification accuracy, and the final significance score is calculated using a weighted average method; the significance scores of the four modes are normalized. The adaptive significance scores of each modality are output, including vibration feature significance, image feature significance, acoustic feature significance, and current feature significance;

[0037] In one embodiment of the present invention, the fault feature library includes standard feature vectors for the following fault types: bearing pitting fault feature vector, mainly characterized by enhanced high-frequency components in the vibration signal, localized temperature increase, obvious impact characteristics in the acoustic signal, and slight current fluctuations; gear wear fault feature vector, mainly characterized by enhanced gear meshing frequency and its harmonics in the vibration signal, uniform temperature increase, increased periodic noise in the acoustic signal, and current load fluctuations; poor electrical contact fault feature vector, mainly characterized by insignificant changes in the vibration signal, localized overheating at the temperature contact point, arc discharge sound in the acoustic signal, and increased distortion and harmonic content of the current waveform; and severe bearing wear fault feature vector, mainly characterized by simultaneous enhancement of low-frequency and high-frequency components in the vibration signal, and obvious temperature increase. The standard feature vectors for each fault type are as follows: a significant increase in vibration, continuous friction noise in the acoustic signal, and a marked increase in current load; mechanical loosening fault feature vectors are characterized by prominent low-frequency components in the vibration signal, insignificant temperature changes, intermittent impact sounds in the acoustic signal, and periodic fluctuations in current; insulation aging fault feature vectors are characterized by basically normal vibration signals, increased temperature at the insulation site, partial discharge sounds in the acoustic signal, increased current leakage, and decreased insulation resistance; and normal state feature vectors are characterized by stable vibration signals, uniform temperature distribution, normal operating noise in the acoustic signal, and current stable within the rated range. Each fault type's standard feature vector is constructed through statistical analysis of historical fault samples and expert knowledge, and after standardization, its dimensions are consistent with the current feature vector.

[0038] S402, Multimodal Fusion Weight Calculation: The fusion weights are calculated and analyzed using the obtained modal saliency scores and data reliability scores. The saliency score of each modality is multiplied by its corresponding data reliability score to obtain the initial fusion weight for each modality, reflecting its overall importance in feature fusion. The initial fusion weights of the four modalities are summed to obtain a total weight as a normalization benchmark. The initial fusion weight of each modality is divided by the total weight to ensure that the sum of all modal weights equals 1, resulting in standardized fusion weights. This method balances the contribution of each modality, considering both feature recognition ability and data quality. The standardized fusion weight set is output, including vibration modal weights, image modal weights, acoustic modal weights, and current modal weights.

[0039] S403, Weighted Feature Fusion Processing: Using the obtained standardized fusion weight set and standardized multimodal feature vector set, weighted feature fusion processing is performed. The feature vector of each mode is multiplied by its corresponding fusion weight to obtain a weighted feature vector, highlighting the feature contributions of important modes. The weighted feature vectors of all modes are concatenated in the order of vibration, image, acoustic, and current to form a long feature vector containing information from all modes. This concatenation method maintains the independence of each modal feature while reflecting the importance of each mode through the weights. The initial fused feature vector is output, containing the weighted feature information of all modes.

[0040] S404, Feature Dimensionality Reduction and Optimization: Using the obtained preliminary fused feature vector, the feature dimension is optimized. Since the fused feature vector has a high dimension, it's necessary to extract its main information components. By analyzing the variation patterns and correlations of each dimension of the feature vector, the most valuable feature components for fault identification are identified. The main feature components that explain over 95% of the data variations are retained, while redundant and noisy feature components are removed. The original fused feature vector is projected onto the directions of these main components, resulting in a feature representation with lower dimension but more concentrated information retention. The dimensionality-reduced fused feature vector is output; the dimension is determined based on the actual data characteristics, typically 15-25 dimensions, containing the most important fault identification feature information.

[0041] S405, Fusion Quality Assessment Process: The quality of the fusion effect is assessed using the obtained dimensionality-reduced fusion feature vectors and standardized fusion weight sets. The dispersion of weights for each modality is calculated. High dispersion indicates significant differences in contribution between modalities, resulting in good fusion discrimination; low dispersion indicates similar modal weights, leading to moderate fusion discrimination. Simultaneously, the consistency between modal features is analyzed. High consistency indicates mutual corroboration of fault information reflected by each modality, resulting in good fusion reliability. Combining weight dispersion and feature consistency, a fusion quality score is calculated, and a confidence level is assigned to the fusion feature vectors. The output is a fusion feature vector with confidence level annotations, containing the dimensionality-reduced feature data and the corresponding fusion confidence score.

[0042] The fault identification module performs intelligent matching analysis of fault features and intelligent identification of fault types based on fused feature vectors with confidence annotations, thereby obtaining fault identification results. Specifically, the following steps are included: S501, Intelligent Matching of Fault Depth for High-Altitude Equipment in Hydropower Stations; Utilizing fused feature vectors, a multimodal deep fusion neural network is employed for intelligent fault feature matching analysis; An environmental adaptive attention mechanism is constructed, which calculates environmental impact factors by analyzing the differences between current environmental parameters such as humidity, temperature, electromagnetic interference, and wind speed and historical environmental data. Specifically, the current humidity, temperature, electromagnetic interference intensity, and wind speed values ​​are obtained, and the statistical mean and standard deviation for the corresponding time period are extracted from the historical environmental database as benchmark reference values; The standardized deviation values ​​of each environmental parameter are calculated to obtain the humidity deviation factor, temperature deviation factor, and electromagnetic interference deviation factor. Interference deviation factor and wind speed deviation factor; based on the influence weight of each environmental parameter on equipment fault identification, the four deviation factors are weighted and summed; the weighted sum is normalized to obtain the environmental impact factor; a temporal memory enhancement module is constructed, and a sliding time window mechanism is established to capture the temporal evolution characteristics of equipment faults. The temporal pattern and evolution mode of fault development are learned through a long short-term memory network, and the current moment's features are fused with historical temporal features to enhance the ability to identify gradual and sudden faults; the processed fused feature vector is input into a multimodal deep fusion neural network to output the probability distribution vector of each fault type; S502 is an intelligent decision-making and identification system for faults in high-altitude equipment at hydropower stations. It utilizes fault type probability distribution vectors and employs an intelligent decision-making mechanism for accurate fault type determination. A multi-level decision fusion architecture is established, including a basic probability analysis layer, an environmental factor correction layer, and a historical trend verification layer. In the basic probability analysis layer, the probability distribution vector is intelligently analyzed to comprehensively analyze the overall morphological characteristics of the probability distribution and determine the degree of certainty in fault identification. An environmental factor correction mechanism is constructed to intelligently correct the basic probability results based on real-time environmental parameters, improving the identification accuracy under specific environmental conditions. A historical trend verification module is designed to analyze the consistency between current fault characteristics and historical fault evolution patterns, verifying the rationality of the identification results. The output includes a comprehensive identification result containing the fault type name, intelligent confidence score, environmental impact factors, and historical trend consistency index.

[0043] S503 is a multi-dimensional intelligent reliability assessment for fault identification of high-altitude equipment in hydropower stations. It utilizes the obtained identification confidence and fusion credibility scores to conduct a comprehensive reliability assessment of the identification results. The identification confidence and fusion credibility are comprehensively analyzed and multiplied to obtain a comprehensive reliability index, which reflects both the quality of feature fusion and the credibility of the identification results. A system reliability threshold is set as the judgment standard, typically determined between 0.5 and 0.8 based on practical application requirements. The comprehensive reliability index is compared with the threshold; if the index is higher than the threshold, it is considered a reliable identification; if the index is lower than the threshold, it is considered an unreliable identification requiring further verification. The reliability judgment result is output, including the reliable or unreliable judgment status and the corresponding comprehensive reliability value.

[0044] S504, Intelligent Collaborative Decision-Making Output for High-Altitude Equipment Faults in Hydropower Stations; Utilizing multi-dimensional reliability assessment results and fault type determination results, an intelligent collaborative decision-making mechanism is employed to determine and output the final fault identification result; A hierarchical decision output architecture is established, employing corresponding decision strategies based on different reliability assessment levels; When the reliability value is greater than a preset second reliability threshold, it represents a high reliability level, and the identification result of a deep fusion neural network is directly used as the final output, accompanied by a detailed reliability analysis report. When the reliability value is greater than a preset first reliability threshold but less than the second reliability threshold, an intelligent collaborative verification mechanism is activated, incorporating multiple verification methods such as expert knowledge base verification, cross-validation of equipment operating status, and secondary confirmation of environmental factors. When the reliability value is less than the preset first reliability threshold, a comprehensive deep analysis mechanism is activated, including multi-modal feature re-extraction, deep learning model integrated decision-making, historical case similarity matching, and expert experience rule verification. The final fault identification result is output, including the final fault type determination, comprehensive confidence score, reliability analysis across dimensions, environmental impact assessment, historical trend comparison, recommended treatment measures, risk level assessment, and subsequent monitoring suggestions.

[0045] The intelligent early warning module predicts fault trends based on the final fault identification results and the sequence of fused feature vectors; it also generates tiered early warning information and structured fault diagnosis reports. Specifically, the following steps are included: S601, Fault Feature Change Trend Analysis; Based on the output fused feature vector sequence, including the fused feature vector sequences obtained at the current time and historical time, a time window sliding analysis method is adopted, setting a fixed-length time window, and calculating the change trend of each dimension of the fused feature vector within the window; By fitting the trend of each feature dimension with time, the slope of change is extracted as a trend parameter to obtain the feature change trend vector; This trend vector reflects the evolution direction and change rate of the equipment status characteristics.

[0046] S602, Fault Development Prediction: Based on the obtained feature change trend vector and the current fused feature vector, a trend continuation prediction method is adopted to combine the current feature vector with the trend vector to obtain the predicted feature vector for the next moment; the prediction process is executed recursively to generate a feature vector sequence for multiple future moments, thus obtaining a predicted feature sequence; this sequence describes the possible evolution path of the equipment state characteristics in the future time period.

[0047] S603, Future Fault Type Prediction: Based on each predicted feature vector in the predicted feature sequence, a fault feature matching method is used to calculate the similarity between each predicted feature vector and each fault type in the fault feature library. The fault type with the highest similarity is selected as the predicted fault type at that moment, resulting in a fault type prediction sequence for multiple future moments. This sequence reflects the possible changes in fault type over time.

[0048] S604, Fault Severity Assessment: Based on the output of the current fault identification result and the fault type prediction sequence, a fault severity grading assessment is adopted. According to the predefined fault severity grading table, the current fault type is mapped to the corresponding severity grading. The grading is divided into four levels: normal, minor abnormality, moderate abnormality, and severe fault. At the same time, the trend of fault severity change in the prediction sequence is analyzed. When the prediction sequence shows that the fault grading is continuously rising, it is judged as a deteriorating trend. When the grading remains stable, it is judged as a stable state, thus obtaining the current fault severity grading and development trend type.

[0049] S605, multi-level early warning information generation: Based on the current severity level and development trend type of the fault, a graded early warning strategy is adopted. When the fault level is slightly abnormal and the trend is stable, a level 1 early warning information is generated and pushed to the operation and maintenance team. When the fault level is moderately abnormal or slightly abnormal but shows a worsening trend, a level 2 early warning information is generated and pushed to the shift leader. When the fault level is a serious fault or moderately abnormal and shows a rapid worsening trend, a level 3 early warning information is generated and pushed to the chief engineer of the power plant. At the same time, the emergency response process is triggered to obtain the early warning information, the corresponding push targets, and the response measures.

[0050] S606, Fault Diagnosis Report Generation: Based on the final fault identification results, fault severity level, development trend type, and early warning information, a structured report generation method is used to integrate information such as the time of fault occurrence, equipment location, fault type, severity, development trend, possible causes, and maintenance recommendations to form a complete fault diagnosis report. The faulty equipment and faulty location are highlighted in the digital twin model through a visual interface, providing complete fault diagnosis and early warning output, and providing comprehensive technical support for equipment maintenance decisions.

[0051] In one embodiment of the present invention, a step of quickly repairing the faulty equipment is further included, as follows: Based on fault diagnosis reports and early warning information, the system automatically analyzes the location, type, severity level, and maintenance requirements of the faulty equipment, and formulates corresponding safety monitoring strategies and protection levels. It generates a list of high-altitude work tasks, including personnel allocation, safety monitoring strategies, estimated work time, and safety protection requirements. Using input data such as faulty equipment location information, fault type, fault severity level, estimated maintenance time, and a list of required maintenance tools, the system intelligently allocates work tasks. Based on the fault severity level, corresponding safety monitoring strategies are formulated: standard safety monitoring mode for minor anomalies, enhanced safety monitoring mode for moderate anomalies, and the highest level of safety monitoring mode for severe faults. By analyzing the height of the fault location, the difficulty of the operation, and environmental risk factors, the required safety protection level and monitoring density are determined. The system outputs a list of high-altitude work tasks, including personnel allocation, safety monitoring strategies, estimated work time, and safety protection requirements.

[0052] Workers wear smart safety belts integrating a three-axis accelerometer, heart rate sensor, GPS positioning module, and wireless communication module. The system continuously collects monitoring parameters such as position coordinates, movement speed, body posture angle, heart rate changes, and safety belt force. By setting safety thresholds, the system determines abnormal states, obtaining physiological data and abnormal warning information for the worker. The smart safety belt continuously collects real-time monitoring parameters such as the worker's position coordinates, movement speed, body posture angle, heart rate changes, and safety belt force. Abnormal states are determined by setting safety thresholds. When the system detects that the worker's heart rate exceeds 120 beats / minute for more than 3 minutes, movement speed exceeds 2 meters / second, body tilt angle exceeds 45 degrees, or safety belt force exceeds 500N, it determines an abnormal state and triggers an warning. The system outputs the worker's physiological data, including real-time physiological parameters, safety status assessment, and abnormal warning information.

[0053] Workers activate a mobile PTZ camera with 360-degree horizontal rotation and 90-degree vertical tilt capabilities. Equipped with a high-definition camera and deep learning image recognition algorithms, it provides comprehensive, blind-spot-free monitoring of the work area. The camera identifies key safety elements such as helmet wearing, safety belt fastening, and tool usage compliance, generating visual monitoring data and an assessment of work compliance. The monitoring content includes whether workers are correctly wearing helmets, safety belts, and using tools correctly, as well as identifying potential safety hazards in the work area. The image recognition algorithm uses a deep learning model to identify workers' safety equipment wearing in real time, and the accuracy is improved through training and optimization. The output visual monitoring data includes real-time image information, safety equipment identification results, and a work compliance assessment.

[0054] The system establishes criteria for judging non-compliant work behaviors based on the requirements of high-altitude operation safety regulations. It utilizes data from intelligent safety belt sensors and PTZ camera vision data for comprehensive analysis of work behaviors. Through a five-step judgment rule, it identifies violations such as not wearing a safety helmet, improperly fastened safety belt, working outside designated areas, using non-compliant tools, and improper work posture, obtaining non-compliant behavior identification results and risk level assessments. Non-compliant behaviors include not wearing a safety helmet, improperly fastened safety belt, working outside designated areas, using non-compliant tools, and improper work posture. The specific judgment rules are as follows: Image recognition is used to detect whether the worker's head has the characteristics of a safety helmet; if not, it is determined that no safety helmet is being worn. A smart safety belt sensor is used to detect the safety belt's fastening status; if the safety belt is not under tension or the tension is abnormal, it is determined that the safety belt is not properly fastened. GPS positioning is used to detect whether the worker's location is within the designated work area; if it exceeds the area boundary, it is determined that work is being done outside the designated area. Image recognition is used to detect the type of work tool; if a non-standard tool is detected, it is determined that non-compliant tools are being used. A posture sensor is used to detect the worker's body angle; if the angle is continuously tilted beyond the safe angle threshold, it is determined that the work posture is improper. The output of the non-compliant behavior identification results includes the type of violation, the time of occurrence, location information, and risk level assessment.

[0055] When the system detects non-compliant work behavior, it immediately activates a multi-level alarm mechanism, using the non-compliant behavior identification results and risk level assessment to perform tiered alarm processing. A three-level alarm response mechanism sends different levels of alarm information to the operator, on-site supervisor, shift leader, and safety management department, resulting in multi-level alarm information and timely and effective safety incident handling. The first-level alarm uses the vibrator and buzzer built into the smart safety belt to provide an immediate reminder to the operator, including the specific type of non-compliant behavior and corrective suggestions. The second-level alarm simultaneously pushes alarm information to the mobile terminal of the on-site safety supervisor, including the operator's name, type of non-compliant behavior, time of occurrence, and location coordinates. The third-level alarm, when non-compliant behavior lasts for more than 30 seconds or high-risk behavior is detected, sends an emergency notification to the shift leader and safety management department via SMS, voice call, and system pop-up reminder. This tiered response mechanism ensures timely and effective handling of safety incidents. Multi-level alarm information is output, including alarm level, responding personnel, handling measures, and tracking records.

[0056] The system employs a multi-channel parallel mode to push notifications to managers at different levels. It utilizes multi-level alarm information and on-site work data to generate notification content that includes worker information, descriptions of non-compliant behaviors, risk assessments, and handling suggestions. Notifications are delivered promptly through four channels: the enterprise mobile application, SMS platform, automated voice telephone, and monitoring screen, ensuring management coordination and rapid response arrangements. The system pushes corresponding notification content to managers at different levels using multi-level alarm information and on-site work data. Notification content includes basic worker information (name, employee ID, work location), detailed descriptions of non-compliant behaviors, risk level assessments, real-time on-site images, and suggested handling measures. Notification channels include: first, pushing real-time messages via the enterprise's internal mobile application, including text descriptions and on-site images; second, sending key information summaries via SMS platform to ensure timely reception even with unstable networks; third, for high-risk situations, calling managers' mobile phones via the automated voice telephone system to play pre-recorded alarm audio; and fourth, displaying alarm information and on-site video footage on the monitoring center's large screen for rapid response by on-duty personnel. Output management and coordination information, including notification sending status, receipt confirmation status, and follow-up handling arrangements.

[0057] The system utilizes intelligent safety belt monitoring data, visual monitoring data, records of non-compliant behaviors, and alarm handling information to conduct safety assessments and records throughout the entire high-altitude operation process. It performs a comprehensive evaluation based on indicators such as work compliance scores, safety incident statistics, work efficiency, and safety equipment usage, resulting in a work safety assessment report and closed-loop management from fault identification to safety maintenance. The system uses intelligent safety belt monitoring data, visual monitoring data, records of non-compliant behaviors, and alarm handling information to conduct a comprehensive assessment of work safety. Assessment indicators include key indicators such as work compliance scores, number of safety incidents, work efficiency, and safety equipment usage. Upon completion of the work, the system automatically generates a work safety report, which includes detailed information such as work duration, safety incident statistics, worker performance evaluation, and improvement suggestions. This report serves as an important basis for worker safety training and performance evaluation, and also provides data support for optimizing high-altitude work safety management processes. The output work safety assessment report includes safety scores, incident analysis, personnel evaluation, and management recommendations, achieving closed-loop management from fault identification to safety maintenance.

[0058] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-sensor fusion-based intelligent fault identification system for high-altitude equipment in hydropower stations, characterized in that: include: Data acquisition and evaluation module; Raw multimodal data is acquired based on multi-source sensors deployed in high-altitude equipment; The original multimodal data is quality assessed to obtain the repaired multimodal dataset; The spatiotemporal alignment module uses the repaired multimodal dataset to perform spatiotemporal alignment and adaptive filtering of the multimodal data to obtain the multimodal dataset. The feature extraction module, based on a multimodal dataset, extracts features from vibration signals, infrared images, and acoustic signals to obtain multimodal feature vectors. The feature fusion module, based on multimodal feature vectors, employs confidence-aware adaptive feature fusion to obtain fused feature vectors with confidence labels. The fault identification module performs intelligent matching analysis of fault features and intelligent identification of fault types based on fused feature vectors with confidence annotations. Obtain the fault identification results; The intelligent early warning module predicts fault trends based on the final fault identification results and the sequence of fused feature vectors. Generate tiered early warning information and structured fault diagnosis reports.

2. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion as described in claim 1, characterized in that, The acquisition of raw multimodal data based on multi-source sensors deployed in high-altitude equipment includes: Based on vibration sensors, infrared thermal imagers, acoustic sensors, and current sensors deployed in key parts of high-altitude equipment, a synchronous trigger acquisition mechanism is adopted to start data acquisition from each sensor under a unified time reference, thereby obtaining the original vibration signal, original infrared image, original acoustic signal, and original current signal.

3. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion as described in claim 1, characterized in that, The quality assessment of the raw multimodal data includes: Based on the original vibration signal, a data integrity detection method is used to obtain vibration data quality labels and corresponding credibility scores; Based on the original infrared image, an image quality assessment method is used to obtain image data quality labels and corresponding confidence scores; Based on the original acoustic signal and the original current signal, statistical analysis methods are used to obtain the acoustic data reliability score and the current data reliability score.

4. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion as described in claim 1, characterized in that, The process of using the repaired multimodal dataset to perform spatiotemporal alignment of multimodal data includes: Using the obtained repaired multimodal data, the time reference is uniformly aligned to obtain the time offset parameters of each sensor; Based on real-time monitoring of high-altitude environmental parameters of hydropower stations, a database of operating condition characteristics of high-altitude equipment in hydropower stations is established, and operating condition parameters are collected; an operating condition influence relationship mapping table is constructed to establish the correspondence between operating condition parameters and modal characteristics; a fault condition correlation analysis model is established to study the differences in sensitivity of various modal sensors to fault characteristics under different operating conditions.

5. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion as described in claim 1, characterized in that, The feature extraction of vibration signals, infrared images, and acoustic signals includes: Using the obtained filtered vibration signal, the vibration signal is decomposed into 8 different frequency bands according to the frequency range. For each frequency band, the signal is separated and the energy ratio characteristics, peak characteristics and shape characteristics are calculated and arranged to form vibration characteristics. Using the obtained filtered infrared image, the temperature distribution features are extracted and analyzed, and the texture features of the image are extracted and combined to obtain the image features; Using the obtained filtered acoustic signal, the acoustic signal is converted to the frequency domain for analysis. According to the operating characteristics of the equipment, the frequency range is divided into three frequency bands: low, medium and high. In each frequency band, the energy value, main frequency and spectral peak value are calculated and combined to obtain the acoustic characteristics.

6. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion according to claim 1, characterized in that, Credibility-aware adaptive feature fusion includes: By collecting equipment operating parameters through a real-time monitoring system, a mapping table of operating condition influence relationships is constructed, and a three-layer progressive significance evaluation architecture is established. The significance of each modal feature is dynamically evaluated by combining real-time operating parameters, and the significance score of each modality with operating condition adaptation is obtained. Using the obtained modality significance scores and data credibility scores, a standardized fusion weight set is obtained through the calculation and analysis of fusion weights; Using the obtained standardized fusion weight set and standardized multimodal feature vector set, a preliminary fusion feature vector is obtained through weighted fusion processing of the features; Using the obtained preliminary fusion feature vector, the dimensionality-reduced fusion feature vector is obtained through feature dimension optimization.

7. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion according to claim 1, characterized in that, The fusion feature vector based on credibility annotation is used for intelligent matching analysis of fault features and intelligent identification of fault types. The fault identification results include: Based on fused feature vectors, an environment-adaptive attention mechanism is constructed to calculate environmental impact factors by analyzing the differences between environmental parameters and historical data; a temporal memory enhancement module is constructed to learn the temporal patterns of fault development through a long short-term memory network; and the probability distribution vectors of each fault type are output. A multi-level decision fusion architecture is established using the probability distribution vector of fault types; a comprehensive confidence score is calculated by comprehensively considering the probability distribution concentration, the degree of influence of environmental correction, and the consistency of historical trends; and a comprehensive identification result is obtained. Based on the comprehensive identification results, reliability analysis is conducted in four dimensions: data quality, algorithm performance, environmental adaptability, and historical consistency. The weight allocation of each dimension is dynamically adjusted and the reliability judgment threshold is determined, and detailed reliability assessment results are output. Based on the reliability assessment results, corresponding decision-making strategies are adopted according to the reliability level. For high reliability level, the identification result is directly output; for medium reliability level, the intelligent collaborative verification mechanism is activated; and for low reliability level, the comprehensive in-depth analysis mechanism is activated to output the final fault identification result.

8. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion according to claim 1, characterized in that, The fault feature library includes: Feature vectors for bearing pitting, gear wear, poor electrical contact, severe bearing wear, mechanical loosening, insulation aging, and normal state are generated. The standard feature vectors for each fault type are standardized to maintain the same dimensions as the current feature vectors.

9. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion according to claim 1, characterized in that, The fault trend prediction based on the final fault identification result and the sequence of fused feature vectors includes: Analyze the changing trends of fault characteristics over time, calculate the rate and direction of change of characteristics; establish a fault development prediction model to predict the degree of fault development in the future; and combine the fault type conversion rules to predict possible new fault types. Establish a fault severity assessment system, classifying faults into four levels: minor, moderate, severe, and critical. Multi-level early warning generation; generates early warning information at corresponding levels, including the warning level, warning content, and recommended handling measures; Fault diagnosis report generation: Automatically generates a complete fault diagnosis report that includes fault type, location of occurrence, severity, development trend, and handling suggestions.

10. The intelligent fault identification system for high-altitude equipment in hydropower stations based on multi-sensor fusion according to claim 1, characterized in that, Also includes: Based on the fault diagnosis report and early warning information, analyze the location of the faulty equipment, the fault type, the severity level and maintenance requirements to obtain a list of high-altitude operation tasks; Workers wear smart safety belts with integrated sensors to monitor their physiological status data and receive abnormal warning information in real time; The location and movements of the workers are tracked in real time by a mobile camera worn by the workers, and visual monitoring data of the workers is obtained. Identify non-standard operating behaviors based on visual monitoring data, including not wearing safety equipment and using tools in violation of regulations; Based on non-standard operating behaviors and the level of safety risks, different levels of safety alarms and response measures are generated; Push security monitoring information and alarm notifications to management personnel; Conduct safety assessments and record the high-altitude operation process, and generate an operation safety assessment report.

Images