A fan engine room intelligent inspection system

Abstract

This invention discloses an intelligent inspection system for wind turbine nacelles, comprising: a data acquisition module for acquiring vibration data of the slip ring chamber, carbon dust accumulation thickness data, and temperature data during the operation of a doubly-fed induction generator (DFIG) wind turbine; a friction feature extraction module for performing time-frequency analysis on the vibration data to extract the friction characteristic frequency and friction vibration amplitude; a carbon brush remaining length estimation module for inputting the friction characteristic frequency, friction vibration amplitude, rotor speed, and temperature data into a pre-trained lightweight carbon brush remaining length estimation model and outputting an estimated carbon brush remaining length; a risk index calculation module for calculating the carbon brush wear risk index and the carbon dust pollution risk index; and a maintenance urgency assessment module for inputting the two risk indices into a pre-trained maintenance urgency assessment model and outputting a maintenance urgency coefficient as the inspection result. This invention solves the problems of existing technologies relying on manual inspection and inaccurate judgment of maintenance timing.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for wind turbine units, specifically to an intelligent inspection system for wind turbine nacelles. Background Technology

[0002] In a doubly-fed wind turbine, the carbon brushes in the slip ring chamber continuously rub against the slip ring, gradually wearing down and shortening. When the carbon brushes wear to their limit, poor contact between the brushes and the slip ring easily occurs, leading to arcing and burning of the slip ring surface. Simultaneously, carbon dust generated by brush friction continuously accumulates at the bottom of the slip ring chamber. When the accumulated carbon dust exceeds a safe thickness, it reduces the insulation performance of the slip ring chamber, potentially causing phase-to-phase short circuits or grounding breakdowns. The slip ring chamber is a small, enclosed component in operation. Traditional manual inspections require stopping the turbine and disassembling the protective cover, resulting in long inspection intervals, an inability to detect dynamic anomalies during operation, and heavy reliance on the experience of maintenance personnel. Current technology lacks an intelligent inspection system that can continuously quantify the degree of carbon brush wear and carbon dust contamination and output maintenance urgency indicators without disassembling the slip ring chamber or stopping the turbine. Summary of the Invention

[0003] This invention provides an intelligent inspection system for wind turbine nacelles, which solves the problems of existing technologies relying on manual inspections and inaccurate judgment of maintenance timing.

[0004] This invention provides an intelligent inspection system for wind turbine nacelles, the system comprising:

[0005] The data acquisition module is used to continuously collect vibration data of the slip ring chamber through vibration sensors, continuously collect carbon dust accumulation thickness data at the bottom of the slip ring chamber through ultrasonic sensors, and continuously collect temperature data of the slip ring chamber through temperature sensors when the doubly fed wind turbine is in operation.

[0006] The friction feature extraction module is used to perform time-frequency analysis on the vibration data of the slip ring chamber and extract the friction characteristic frequency and friction vibration amplitude generated by the friction between the carbon brush and the slip ring.

[0007] The carbon brush remaining length estimation module is used to input the friction characteristic frequency, the friction vibration amplitude, the rotor speed and the slip ring chamber temperature data into the pre-trained lightweight carbon brush remaining length estimation model, and output the carbon brush remaining length estimation value.

[0008] The risk index calculation module is used to calculate the carbon brush wear risk index based on the estimated remaining carbon brush length, and to calculate the carbon dust pollution risk index based on the carbon dust accumulation thickness data.

[0009] The maintenance urgency assessment module is used to input the carbon brush wear risk index and the carbon dust pollution risk index into a pre-trained maintenance urgency assessment model, output a maintenance urgency coefficient, and use the maintenance urgency coefficient as the inspection result of the doubly-fed wind turbine. The larger the maintenance urgency coefficient, the more urgently the slip ring chamber of the doubly-fed wind turbine needs maintenance.

[0010] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0011] First, this invention uses vibration sensors, ultrasonic sensors, and temperature sensors to collect slip ring chamber operating data in a coordinated manner, enabling real-time acquisition of carbon brush friction, carbon dust accumulation, and temperature status without shutting down the machine, thus solving the problems of long manual inspection cycles and inability to monitor online.

[0012] Second, this invention performs time-frequency analysis on vibration data to extract the characteristic frequency and amplitude of carbon brush friction, and inputs them into a lightweight carbon brush remaining length estimation model to achieve online continuous estimation of carbon brush remaining length, reducing reliance on downtime measurement.

[0013] Third, this invention quantifies the risks of carbon brush wear and carbon dust contamination, and integrates the two types of risks through a maintenance urgency assessment model to output a continuous maintenance urgency coefficient, providing maintenance personnel with a clear and quantifiable basis for decision-making on maintenance timing, and avoiding over-maintenance and under-maintenance.

[0014] Fourth, the lightweight carbon brush remaining length estimation model is obtained through knowledge distillation, which reduces the consumption of computing resources while ensuring estimation accuracy, and adapts to the deployment requirements of the cabin embedded platform. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural diagram of an intelligent inspection system for a wind turbine nacelle provided in an embodiment of the present invention;

[0017] Figure 2 This is a logic diagram of a wind turbine nacelle intelligent inspection system provided in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram of the knowledge distillation training process of a lightweight carbon brush remaining length estimation model for an intelligent inspection system for wind turbine nacelles provided in an embodiment of the present invention.

[0019] The diagram is labeled as follows: Data acquisition module 11, friction feature extraction module 12, carbon brush remaining length estimation module 13, risk index calculation module 14, and maintenance urgency assessment module 15. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] Examples, such as Figure 1 , Figure 2 , Figure 3 As shown, the present invention provides an intelligent inspection system for wind turbine nacelles, the system comprising:

[0022] Data acquisition module 11 is used to continuously collect vibration data of the slip ring chamber through a vibration sensor, continuously collect carbon dust accumulation thickness data at the bottom of the slip ring chamber through an ultrasonic sensor, and continuously collect temperature data of the slip ring chamber through a temperature sensor when the doubly fed wind turbine is in operation.

[0023] The data acquisition module 11 provided in this embodiment of the invention is used for:

[0024] The vibration sensor is attached to the mounting base on the outer shell of the slip ring chamber near the carbon brush bracket, and continuously collects the vibration data of the slip ring chamber at a preset acquisition frequency;

[0025] An ultrasonic sensor is installed on the inner side of the housing at the bottom of the slip ring chamber, directly opposite the area below the carbon brush. The ultrasonic sensor emits ultrasonic pulses vertically downwards at the preset acquisition cycle and receives the echoes reflected by the upper surface of the carbon dust. The thickness of the carbon dust accumulation at the bottom of the slip ring chamber is calculated based on the time difference between emission and reception.

[0026] The temperature sensor is installed inside the slip ring chamber near the contact surface between the slip ring and the carbon brush, and continuously collects slip ring chamber temperature data at the preset acquisition cycle.

[0027] Specifically, the vibration sensor is attached to a mounting base on the slip ring chamber shell near the carbon brush holder, continuously acquiring slip ring chamber vibration data at a sampling frequency of 10kHz. The vibration sensor can be a piezoelectric accelerometer, attached to the mounting base on the slip ring chamber shell near the carbon brush holder via threaded fastening, with silicone grease applied to the mounting surface to ensure vibration transmission. The vibration sensor has a response time of less than 10ms, and the sampling frequency of 10kHz covers the characteristic frequency band of carbon brush friction from 800Hz to 3000Hz, satisfying the sampling theorem requirements. The vibration sensor continuously outputs vibration acceleration data, and the slip ring chamber vibration data forms a continuous data stream with a 10kHz sampling rate, stored in the data buffer of the data acquisition module.

[0028] The ultrasonic sensor is installed inside the housing at the bottom of the slip ring chamber, directly below the carbon brush. The probe points vertically downwards, at a height of 150mm from the initial carbon dust accumulation surface. An industrial-grade short-range ultrasonic sensor is used, with a distance measurement range of 30mm to 300mm, a blind zone of 30mm, and a measurement accuracy of ±1mm. The preset acquisition period is 1 minute. The ultrasonic sensor emits ultrasonic pulses vertically downwards at this preset acquisition period and receives the echoes reflected from the upper surface of the carbon dust. The temperature sensor response time is less than 10s, and the ultrasonic sensor's single measurement time is less than 5ms, both meeting the requirements of the 1-minute preset acquisition period. Furthermore, carbon brush wear and carbon dust accumulation within the slip ring chamber are slow processes; the 1-minute period can promptly capture anomalies without imposing excessive computational and storage burdens.

[0029] The speed of sound in air is approximated as 340 m / s, and corrected for using temperature data. The correction formula is: v = 331.3 + 0.606T, where T is the current temperature of the slip ring chamber and v is the corrected speed of sound. The ultrasonic sensor receives the echo reflected from the upper surface of the carbon dust and records the time difference between transmission and reception. Based on the time difference and the corrected speed of sound, the internal processing circuit of the ultrasonic sensor calculates the distance from the probe to the upper surface of the carbon dust, and then subtracts this distance from the installation height to obtain the thickness of the carbon dust accumulation at the bottom of the slip ring chamber.

[0030] For example, the current slip ring chamber temperature is 65.3℃. According to the sound speed correction formula v=331.3+0.606×65.3=370.9m / s, the transmission and reception time difference in a certain measurement is 0.000744s, the distance from the probe to the upper surface of the carbon dust is 370.9×0.000744÷2=0.138m, and the installation height is 0.15m. Therefore, the carbon dust accumulation thickness is 0.15m minus 0.138m equals 0.012m, which is 12mm.

[0031] A temperature sensor is installed inside the slip ring chamber near the contact surface between the slip ring and the carbon brush, continuously collecting slip ring chamber temperature data at a preset acquisition cycle of 1 minute. The temperature sensor can be a PT100 platinum resistance temperature sensor with a range of -40℃ to 150℃ and an accuracy of ±0.5℃. The temperature sensor is fixed to the slip ring and carbon brush contact surface using thermally conductive adhesive to measure the internal temperature of the slip ring chamber. For example, the slip ring chamber temperature measured in the current acquisition cycle is 65.3℃.

[0032] The following technical effects have been achieved through this module:

[0033] By continuously acquiring vibration signals from the slip ring chamber at high frequency and periodically acquiring temperature and ultrasonic data, the amount of temperature and ultrasonic data can be controlled while ensuring that friction features can be extracted. This provides time-synchronized multi-source data for subsequent friction feature extraction and model input, allowing continuous acquisition of operational information related to carbon brush wear and carbon dust accumulation without stopping the machine to disassemble the slip ring chamber.

[0034] Friction feature extraction module 12 is used to perform time-frequency analysis on the vibration data of the slip ring chamber and extract the friction feature frequency and friction vibration amplitude generated by the friction between the carbon brush and the slip ring.

[0035] The friction feature extraction module 12 provided in this embodiment of the invention is used for:

[0036] The vibration data of the slip ring chamber is subjected to a short-time Fourier transform. The short-time Fourier transform uses a Hanning window with a preset window length and a preset overlap rate between adjacent windows to obtain the time spectrum of the vibration data of the slip ring chamber.

[0037] In the time spectrum, the friction characteristic frequency generated by the friction between the carbon brush and the slip ring is extracted, wherein the friction characteristic frequency is located within a preset carbon brush friction characteristic frequency band, and the preset carbon brush friction characteristic frequency band is from 800Hz to 3000Hz.

[0038] At the frequency point corresponding to the friction characteristic frequency, the friction vibration amplitude is extracted, and the friction vibration amplitude is the amplitude spectrum value of the friction characteristic frequency in the short-time Fourier transform result.

[0039] Among them, the friction characteristic frequencies generated by the friction between the carbon brush and the slip ring are extracted from the time spectrum, including:

[0040] Obtain the current rotor speed and calculate the theoretical friction frequency generated by the friction between the carbon brush and the slip ring based on the rotor speed;

[0041] In the time spectrum, the frequency point with the largest amplitude spectrum peak within the preset carbon brush friction characteristic frequency band is taken as the candidate friction characteristic frequency;

[0042] The candidate friction characteristic frequency is compared with the theoretical friction frequency. When the deviation between the candidate friction characteristic frequency and the theoretical friction frequency is within a preset deviation range, the candidate friction characteristic frequency is taken as the friction characteristic frequency.

[0043] When the deviation between the candidate friction characteristic frequency and the theoretical friction frequency exceeds the preset deviation range, the theoretical friction frequency is taken as the friction characteristic frequency.

[0044] Specifically, the vibration data of the slip ring chamber output by the vibration sensor is continuously acquired at a sampling rate of 10kHz. The friction feature extraction module processes data in 1-minute cycles, retrieving the vibration data of the most recent 10 seconds from the vibration data buffer each cycle as a vibration data segment, with a segment length of 100,000 sampling points. The short-time Fourier transform uses a Hanning window with a window length of 1024 sampling points and an overlap rate of 50% between adjacent windows. A window length of 1024 points corresponds to a time step of 0.1024s at a 10kHz sampling rate, and the frequency resolution is approximately 9.77Hz (10kHz divided by the window length). With a 50% overlap rate between adjacent windows, the adjacent windows move 512 points, resulting in a time step of 0.0512s, which maintains both frequency resolution and temporal continuity of the time spectrum. The characteristic frequency band of carbon brush and slip ring friction is 800Hz to 3000Hz. The resolution requirement for the characteristic frequency to change with rotational speed is no more than 10Hz. The 9.77Hz frequency resolution provided by the 1024-point window can clearly distinguish adjacent frequency components within the frequency band. The 0.1024s time resolution is sufficient to capture the friction frequency drift caused by the change in rotor speed, and at the same time, there will be no obvious spectrum leakage due to the short window length.

[0045] The expression for calculating the short-time Fourier transform is: Where STFT(t,f) is the amplitude spectrum value of the time spectrum at time frame t and frequency f, x(n) is the vibration data segment, w(n) is the Hanning window, n is the sampling point number, M is the number of adjacent window movement points, M=512, and N is the window length, N=1024. The time spectrum of the slip ring chamber vibration data is obtained by repeating the calculation for each time frame.

[0046] The preset characteristic frequency band for carbon brush friction is 800Hz to 3000Hz. This frequency band was determined based on vibration test data of carbon brush and slip ring friction in the slip ring chamber of a doubly-fed wind turbine and statistical analysis of historical vibration spectra from multiple wind farms. The vibration energy generated by the friction between the carbon brush and slip ring is concentrated within this frequency band and is clearly distinguishable from the main frequency bands of other mechanical vibration sources in the slip ring chamber. In the time spectrum, for each time frame, the frequency point with the largest amplitude spectrum peak value within the 800Hz to 3000Hz frequency band is searched, and this frequency point is used as the candidate friction characteristic frequency for that time frame.

[0047] Obtain the current rotor speed. The rotor speed is transmitted by the wind turbine main control system via fieldbus, and the unit is r / min. Calculate the theoretical friction frequency generated by the friction between the carbon brushes and slip rings based on the rotor speed. The expression for calculating the theoretical friction frequency is: f theory =(n rotor / 60)×N slo t, where ft heory n is the theoretical friction frequency. rotor N is the rotor speed. slotN represents the number of circumferentially distributed ventilation slots on the slip ring surface. These slots are used for heat dissipation and dust removal. When the carbon brush passes through these slots, it generates periodic impact vibrations. The impact frequency is equal to the number of ventilation slots multiplied by the rotor frequency. slot The slip ring's factory design parameter is, for example, taken as 60. The rotor frequency is equal to n. rotor / 60. The current rotor speed is 1500 r / min, the rotor frequency = 1500 / 60 = 25 Hz, and the theoretical friction frequency = 25 × 60 = 1500 Hz.

[0048] Compare the candidate friction characteristic frequency with the theoretical friction frequency. When the deviation between the candidate friction characteristic frequency and the theoretical friction frequency is within the preset deviation range, the candidate friction characteristic frequency is taken as the friction characteristic frequency; when the deviation between the candidate friction characteristic frequency and the theoretical friction frequency exceeds the preset deviation range, the theoretical friction frequency is taken as the friction characteristic frequency. The preset deviation range is set at ±5% of the theoretical friction frequency. The method for determining this range is as follows: the rotor speed measurement accuracy is ±1 r / min, corresponding to a frequency error of ±0.0167 Hz; since the theoretical friction frequency equals the product of the rotational frequency and the number of ventilation slots, the effect of the speed measurement error transmitted to the theoretical friction frequency is ±0.0167 × 60 ≈ ±1 Hz; the short-time Fourier transform frequency resolution is approximately 9.77 Hz, corresponding to a peak positioning error of approximately ±9.77 Hz, which is approximately ±0.65% relative to the theoretical friction frequency of 1500 Hz; the manufacturing tolerance of the ventilation slots on the slip ring surface will cause the actual ventilation slot spacing to deviate from the design value, and carbon brush wear will change the equivalent ventilation slot spacing corresponding to the contact position between the carbon brush and the slip ring. The equivalent frequency deviation caused by the combination of these two factors is estimated to be ±2% of the theoretical frequency based on the slip ring machining tolerance grade and the carbon brush wear range. Considering all the above error sources and reserving approximately 2% margin, ±5% is set as the preset deviation range, which can cover all errors while avoiding misjudging the actual friction frequency as abnormal.

[0049] For example, in the time spectrum corresponding to the current time frame, the frequency point with the largest amplitude spectrum peak within the preset carbon brush friction characteristic frequency band of 800Hz to 3000Hz is 1512Hz. The amplitude spectrum value of this frequency point is significantly higher than other frequencies within the band, therefore 1512Hz is selected as the candidate friction characteristic frequency. In this embodiment, the theoretical friction frequency is 1500Hz, and a ±5% deviation range corresponds to 1425Hz to 1575Hz. The deviation between the candidate friction characteristic frequency 1512Hz and the theoretical friction frequency 1500Hz is 12Hz, and the relative deviation is 12 / 1500 = 0.8%, which is within the preset deviation range. Therefore, 1512Hz is selected as the friction characteristic frequency of the current time frame.

[0050] When the deviation between the candidate friction feature frequency and the theoretical friction frequency exceeds a preset deviation range, the theoretical friction frequency is used as the friction feature frequency for the current time frame. This situation typically corresponds to other vibration interference sources existing in the 800Hz to 3000Hz frequency band of the time spectrum, such as the vibration components of the slip ring chamber cooling fan or bearing, whose amplitude spectrum peak value occasionally exceeds the peak value corresponding to the true friction frequency. In this case, directly using the theoretical friction frequency as the friction feature frequency can avoid using the interference frequency as a feature input to the subsequent model, ensuring the continuity of feature extraction.

[0051] After obtaining the friction characteristic frequency, the amplitude spectrum value is extracted from the frequency point corresponding to this friction characteristic frequency in the time spectrum, which is used as the friction vibration amplitude. For example, if the friction characteristic frequency of the current time frame is 1512Hz, and the amplitude spectrum value of this frequency point in the time spectrum is 0.35m / s², then the friction vibration amplitude is 0.35m / s². The unit of friction vibration amplitude is m / s², which is consistent with the unit of vibration acceleration, reflecting the strength of the friction excitation between the carbon brush and the slip ring. The friction vibration amplitude changes with the wear state of the carbon brush.

[0052] The above extraction process is repeated for each time frame within each vibration data segment to obtain the friction characteristic frequency and friction vibration amplitude of each time frame. The two are combined with the corresponding timestamp to form a friction characteristic sequence for use by the carbon brush remaining length estimation module.

[0053] The following technical effects have been achieved through this module:

[0054] The vibration data of the slip ring chamber is converted into a time spectrum using short-time Fourier transform. The theoretical friction frequency is calculated within the characteristic frequency band of carbon brush friction, combined with the characteristics of the slip ring ventilation groove. The deviation between the candidate peak frequency and the theoretical frequency is used as a screening criterion, achieving stable extraction of the characteristic frequency and amplitude of carbon brush friction. Even in the presence of other vibration disturbances in the slip ring chamber, the accuracy and continuity of feature extraction can be maintained through deviation judgment, providing reliable input for the subsequent carbon brush remaining length estimation model, thus enabling continuous monitoring of carbon brush wear status without shutting down the system.

[0055] The carbon brush remaining length estimation module 13 is used to input the friction characteristic frequency, the friction vibration amplitude, the rotor speed and the slip ring chamber temperature data into the pre-trained lightweight carbon brush remaining length estimation model, and output the carbon brush remaining length estimation value.

[0056] The carbon brush remaining length estimation module 13 provided in this embodiment of the invention is used for:

[0057] The friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data are concatenated into a carbon brush state input feature vector;

[0058] The carbon brush state input feature vector is input into a pre-trained lightweight carbon brush remaining length estimation model, and the lightweight carbon brush remaining length estimation model outputs the carbon brush remaining length estimate, wherein the carbon brush remaining length estimate is a continuous value in millimeters.

[0059] The pre-training process of the pre-trained lightweight carbon brush remaining length estimation model includes:

[0060] Collect carbon brush status sample data of the slip ring chamber of the doubly fed wind turbine during historical operation. Each carbon brush status sample data includes friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data, as well as the corresponding carbon brush remaining length label value. The carbon brush remaining length label value is obtained by maintenance personnel using vernier calipers or special carbon brush length measuring tools to measure the actual remaining length of the carbon brush when the unit is stopped.

[0061] The carbon brush status sample data is cleaned using the box plot method to obtain cleaned carbon brush status sample data, and the cleaned carbon brush status sample data is divided into training set and validation set according to a preset ratio.

[0062] An initial deep carbon brush remaining length estimation model is constructed. The initial deep carbon brush remaining length estimation model is a regression model based on a fully connected neural network, which includes an input layer, multiple hidden layers and an output layer. The output layer outputs a scalar value as the predicted carbon brush remaining length.

[0063] The friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data of each carbon brush state sample data in the training set are used as model input, the corresponding carbon brush remaining length label is used as supervision label, and the mean square error is used as loss function to supervise the training of the initial deep carbon brush remaining length estimation model. The validation set is used to judge early stop until the loss function on the validation set converges, and the pre-trained deep carbon brush remaining length estimation model is obtained.

[0064] Specifically, the friction characteristic frequency, friction vibration amplitude, rotor speed, and slip ring chamber temperature data are concatenated into a carbon brush state input feature vector. This carbon brush state input feature vector is denoted as X=[f,A,n,T], where f is the friction characteristic frequency, A is the friction vibration amplitude, n is the rotor speed, and T is the slip ring chamber temperature. Before concatenation, each feature component is linearly normalized to eliminate the influence of differences in the dimensions and numerical ranges of different physical quantities on model training. The friction characteristic frequency f is divided by the upper limit of the carbon brush friction characteristic frequency band, 3000Hz, to obtain the normalized value f / 3000; the friction vibration amplitude A is divided by the upper limit of the vibration sensor range, 10m / s², to obtain the normalized value A / 10; the rotor speed n is divided by the rated rotor speed of the doubly-fed wind turbine, 1800r / min, to obtain the normalized value n / 1800, where 1800r / min is the typical rated rotor speed of the doubly-fed wind turbine, determined by the parameters on the unit's nameplate; the slip ring chamber temperature T is divided by the upper limit of the temperature sensor range, 150℃, to obtain the normalized value T / 150.

[0065] For example, the friction characteristic frequency is 1512Hz, the friction vibration amplitude is 0.35m / s², the rotor speed is 1500r / min, and the slip ring chamber temperature is 65.3℃. After normalization, the carbon brush state input feature vector is obtained as X=[1512 / 3000,0.35 / 10,1500 / 1800,65.3 / 150]=[0.504,0.035,0.833,0.435].

[0066] Collect carbon brush condition sample data of the slip ring chamber of the doubly-fed induction generator (DFIG) during historical operation. Each carbon brush condition sample data includes friction characteristic frequency, friction vibration amplitude, rotor speed, and slip ring chamber temperature data, as well as the corresponding carbon brush remaining length label value. The friction characteristic frequency, friction vibration amplitude, rotor speed, and slip ring chamber temperature data are obtained by collecting data at a preset acquisition cycle during historical operation and calculating it using the aforementioned module. The carbon brush remaining length label value is obtained by maintenance personnel opening the slip ring chamber cover while the generator is shut down and directly measuring the actual remaining length of the carbon brush using vernier calipers or a dedicated carbon brush length measuring tool, in mm. For example, during a shutdown maintenance, the actual remaining length of a certain carbon brush was measured to be 42.5 mm. This value, together with the four-dimensional feature data collected at that moment, constitutes a carbon brush condition sample data.

[0067] Box plots were used to clean the carbon brush condition sample data. The first and third quartiles were calculated for each feature component and label value to obtain the normal range for each component. The first quartile of a component is denoted as Q1, the third quartile as Q3, and the interquartile range (IQR) is denoted as IQR = Q3 - Q1. The normal range is [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]. If any component or label value of a sample falls outside the corresponding normal range, that sample is discarded as an outlier.

[0068] For example, a total of 1200 samples were collected from the historical samples. After cleaning using the box plot method, 1152 samples were retained, and 48 abnormal samples were removed. For example, some samples had vibration amplitudes close to zero or significant temperature jumps due to short-term poor contact of vibration or temperature sensors. The 1152 cleaned samples were divided into a training set and a validation set in a ratio of 8:2, resulting in a training set of 922 samples and a validation set of 230 samples.

[0069] An initial deep model for estimating the remaining length of a carbon brush was constructed. This model is a regression model based on a fully connected neural network, consisting of one input layer, three hidden layers, and one output layer. The input layer has four neurons, corresponding to the four components of the carbon brush state input feature vector X. The first hidden layer has 64 neurons, the second hidden layer has 32 neurons, and the third hidden layer has 16 neurons, all using the ReLU activation function. The output layer has one neuron, using a linear activation function, outputting a scalar value as the predicted remaining length of the carbon brush. The deep model has 2832 weights, 113 biases, and a total of 2945 parameters. The large parameter size of the deep model allows it to adequately fit the complex nonlinear relationship between the carbon brush state features and the remaining length of the carbon brush during the training phase.

[0070] The model input is the normalized feature vector of each carbon brush status sample in the training set, the supervision label is the corresponding remaining carbon brush length label, and the loss function is the mean squared error (MSE). The MSE loss function is MSE = (1 / m) × Σ(y i - i )², where m is the batch size of the training samples, and y i Let be the label value of the remaining carbon brush length for the i-th sample. iThe model outputs the predicted remaining carbon brush length. An adaptive moment estimator optimizer is used to update the network parameters, with a learning rate of 0.001 and a batch size of 32. After each training round, the validation mean squared error is calculated on the validation set. Early stopping is triggered when the validation mean squared error does not decrease for 10 consecutive rounds, halting training. The model parameters corresponding to the round with the minimum validation set loss are taken as the final parameters, resulting in the pre-trained deep carbon brush remaining length estimation model. For example, during training, the minimum mean squared error of the deep carbon brush remaining length estimation model on the validation set is 2.35, corresponding to a mean absolute error of approximately 1.2 mm for the remaining carbon brush length, indicating that the deep carbon brush remaining length estimation model can accurately estimate the remaining carbon brush length.

[0071] This embodiment of the invention provides a carbon brush remaining length estimation module 13 that further includes lightweighting the pre-trained deep carbon brush remaining length estimation model to obtain a pre-trained lightweight carbon brush remaining length estimation model, including:

[0072] An initial lightweight carbon brush remaining length estimation model is constructed. The initial lightweight carbon brush remaining length estimation model is a regression model based on a fully connected neural network, which includes an input layer, a hidden layer and an output layer. The input layer receives the carbon brush state input feature vector. The number of neurons in the hidden layer does not exceed 8 and is followed by a ReLU activation function. The output layer uses a linear activation function to output a scalar value as the carbon brush remaining length estimate.

[0073] The pre-trained deep carbon brush remaining length estimation model is provided as a soft-label model, and the initial lightweight carbon brush remaining length estimation model is used as a lightweight model to be trained.

[0074] Using the friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data of the carbon brush state sample data in the training set and the validation set as model input, the predicted carbon brush remaining length provided by the soft label as the soft label, and the corresponding carbon brush remaining length label value as the hard label, the lightweight model to be trained is subjected to knowledge distillation training.

[0075] The mean square error between the estimated remaining carbon brush length output by the lightweight model to be trained and the soft label is calculated as the soft label loss, and the mean square error between the estimated remaining carbon brush length output by the lightweight model to be trained and the hard label is calculated as the hard label loss.

[0076] The soft label loss and the hard label loss are weighted and summed to form a joint loss function. The parameters of the lightweight model to be trained are updated with the goal of minimizing the joint loss function until convergence is verified, thus obtaining a pre-trained lightweight carbon brush remaining length estimation model.

[0077] Specifically, the initial lightweight carbon brush remaining length estimation model is a regression model built on a fully connected neural network, consisting of an input layer, a hidden layer, and an output layer. The input layer has 4 neurons, receiving the carbon brush state input feature vector. The hidden layer has 8 neurons, satisfying the requirement of no more than 8, followed by a ReLU activation function. The output layer uses a linear activation function, outputting a scalar value as the carbon brush remaining length estimate. The lightweight carbon brush remaining length estimation model has 4×8+8×1=40 weights and 8+1=9 biases, for a total of 49 parameters, approximately 1 / 60th of the total parameters of the deep carbon brush remaining length estimation model, making it suitable for operation on an embedded platform in an aircraft cabin.

[0078] A pre-trained deep brush remaining length estimation model is used as the soft-labeling model, and an initial lightweight brush remaining length estimation model is used as the lightweight model to be trained. Knowledge distillation training uses all cleaned sample data without distinguishing between the training and validation sets. For each brush state sample, the normalized feature vector is input into the soft-labeling model, and the output predicted brush remaining length is used as the soft label; at the same time, the corresponding true brush remaining length label value is retained as the hard label.

[0079] The mean squared error between the estimated remaining carbon brush length output by the lightweight model to be trained and the soft tag is calculated and used as the soft tag loss L. soft The mean squared error between the estimated remaining carbon brush length output by the lightweight carbon brush remaining length estimation model to be trained and the hard label is calculated and used as the hard label loss L. hard The joint loss function is L. total =α×L soft +(1-α)×L hard , where α is the soft label loss weight, which is 0.7, and (1-α) is the hard label loss weight, which is 0.3, used to retain the constraint of the true measurement value.

[0080] The soft label is output by a pre-trained deep brush remaining length estimation model. This model has fully learned the mapping relationship between brush state features and remaining brush length on the training set. When its output is used as a soft label, it retains the approximate numerical range of the true label and includes the smooth response of the deep model to input noise and feature fluctuations, providing a more stable regression target for the lightweight model than a single hard label. The hard label is the measured remaining brush length under shutdown conditions, representing a true measurement value, used to constrain the lightweight model's output to not deviate from the actual wear state.

[0081] The method for determining the soft-label loss weights is as follows: Before knowledge distillation training, a one-dimensional grid search is performed on the validation set for α, with the candidate α value set being {0.5, 0.6, 0.7, 0.8, 0.9}. For each candidate α value, short-time knowledge distillation training is performed using the same initial lightweight model and the same training settings, with 10 training epochs. Then, the mean squared error of the joint loss function after convergence is calculated on the validation set. The results show that the mean squared error of the validation set is minimized when α is 0.7, at 2.68, which is lower than the validation errors corresponding to other values. Therefore, the soft-label loss weight is determined to be 0.7.

[0082] With minimizing the joint loss function as the optimization objective, an adaptive moment estimator optimizer is used with a learning rate of 0.0005 and a batch size of 32 to update the parameters of the lightweight model to be trained until the validation convergence, thus obtaining the pre-trained lightweight carbon brush remaining length estimation model.

[0083] The normalized carbon brush status data acquired in real time is input into the feature vector and then fed into a pre-trained lightweight carbon brush remaining length estimation model. The output value is a continuous numerical value in millimeters. For example, after the normalized feature vector X=[0.504,0.035,0.833,0.435] for the current sampling period is input into the model, the lightweight carbon brush remaining length estimation model outputs an estimated carbon brush remaining length of 41.8 mm. This estimate will have some error compared to the actual remaining length, but the error range is within the acceptable range of the maintenance plan. The subsequent carbon brush wear risk index calculation module will further use time-series estimates and multi-source data for risk quantification.

[0084] The following technical effects have been achieved through this module:

[0085] Through a two-stage modeling approach combining deep model training and knowledge distillation, a lightweight inference model is obtained while maintaining the accuracy of carbon brush remaining length estimation, enabling real-time online estimation on an embedded platform in the engine room. During training, the deep model fully learns the nonlinear mapping relationship between friction characteristic frequencies, friction vibration amplitudes, rotor speed, slip ring chamber temperature, and carbon brush remaining length. The lightweight model inherits the knowledge from the deep model through knowledge distillation, and the joint loss function utilizes both the smoothing information from soft labels and the precise constraints from hard labels, maintaining acceptable estimation accuracy while significantly reducing the number of parameters. This module provides continuous online estimates for calculating the carbon brush wear risk index, reducing the number of downtime disassembly measurements and improving inspection efficiency and the real-time nature of maintenance.

[0086] The risk index calculation module 14 is used to calculate the carbon brush wear risk index based on the estimated remaining carbon brush length, and to calculate the carbon dust pollution risk index based on the carbon dust accumulation thickness data.

[0087] The risk index calculation module 14 provided in this embodiment of the invention is used for:

[0088] The length of the carbon brush in its brand-new state is taken as the initial length value of the carbon brush, and the minimum length of the carbon brush that can be used is taken as the limit length value of the carbon brush.

[0089] Subtracting the estimated remaining length of the carbon brush from the initial length value of the carbon brush yields the amount of carbon brush shortening.

[0090] The carbon brush wear ratio is obtained by dividing the carbon brush shortening amount by the difference between the initial carbon brush length value and the carbon brush limit length value.

[0091] The carbon brush wear ratio is limited to the range of 0 to 1, and the limited carbon brush wear ratio is used as the carbon brush wear risk index.

[0092] Specifically, the length of the carbon brush in its brand-new state is taken as the initial length value of the carbon brush. The initial length value is obtained from the carbon brush product specifications or the slip ring chamber maintenance manual; in this embodiment, the initial length value is 60mm. The minimum usable length of the carbon brush is taken as the limit length value. The limit length value is determined according to the wear limit specified by the carbon brush manufacturer, typically 50% of the initial length value; in this embodiment, the limit length value is 30mm. The estimated remaining length of the carbon brush is output by the remaining length estimation module 13; in this embodiment, the estimated remaining length is 41.8mm.

[0093] Calculate the carbon brush shortening: Carbon brush shortening = Initial carbon brush length - Estimated remaining carbon brush length = 60 - 41.8 = 18.2 mm.

[0094] Calculate the carbon brush wear ratio: Carbon brush wear ratio = Carbon brush shortening amount ÷ (Initial carbon brush length value - Ultimate carbon brush length value) = 18.2 ÷ 30 ≈ 0.607. The carbon brush wear ratio is limited to the range of 0 to 1. When the carbon brush wear ratio is less than 0, the carbon brush wear risk index is 0; when the carbon brush wear ratio is greater than 1, the carbon brush wear risk index is 1; when the carbon brush wear ratio is between 0 and 1, the carbon brush wear risk index is equal to the carbon brush wear ratio. In this embodiment, the carbon brush wear ratio is 0.607, therefore the carbon brush wear risk index is 0.607.

[0095] The risk index calculation module 14 provided in this embodiment of the invention is also used for:

[0096] The maximum allowable thickness of carbon dust accumulation that does not affect the insulation performance of the slip ring chamber is taken as the safe thickness threshold of carbon dust, and the thickness of carbon dust accumulation that seriously threatens the insulation performance of the slip ring chamber is taken as the dangerous thickness threshold of carbon dust.

[0097] When the carbon dust accumulation thickness data is less than or equal to the carbon dust safe thickness threshold, the carbon dust pollution risk index is set to 0;

[0098] When the carbon dust accumulation thickness data is greater than or equal to the carbon dust dangerous thickness threshold, the carbon dust pollution risk index is set to 1;

[0099] When the carbon dust accumulation thickness data is between the carbon dust safe thickness threshold and the carbon dust dangerous thickness threshold, the carbon dust pollution risk index is obtained by subtracting the carbon dust safe thickness threshold from the carbon dust accumulation thickness data and dividing by the difference between the carbon dust dangerous thickness threshold and the carbon dust safe thickness threshold.

[0100] Specifically, the maximum permissible thickness of carbon dust accumulation that does not affect the insulation performance of the slip ring chamber is defined as the safe carbon dust thickness threshold. This safe carbon dust thickness threshold is determined based on the slip ring chamber insulation distance requirements and the conductivity characteristics of the carbon dust; in this embodiment, it is set to 10 mm. The thickness at which carbon dust accumulation severely threatens the insulation performance of the slip ring chamber is defined as the dangerous carbon dust thickness threshold. This dangerous carbon dust thickness threshold is determined based on test data showing that the slip ring chamber insulation resistance drops to a dangerous level when carbon dust accumulates to this thickness; in this embodiment, it is set to 30 mm. The carbon dust accumulation thickness data is measured by the ultrasonic sensor in the data acquisition module 11; in this embodiment, the carbon dust accumulation thickness data for the current acquisition cycle is 12 mm.

[0101] When the carbon dust accumulation thickness is less than or equal to the safe carbon dust thickness threshold, the carbon dust pollution risk index is 0; when the carbon dust accumulation thickness is greater than or equal to the dangerous carbon dust thickness threshold, the carbon dust pollution risk index is 1; when the carbon dust accumulation thickness is between the safe carbon dust thickness threshold and the dangerous carbon dust thickness threshold, the carbon dust pollution risk index is equal to the carbon dust accumulation thickness minus the safe carbon dust thickness threshold, divided by the difference between the dangerous carbon dust thickness threshold and the safe carbon dust thickness threshold. In this embodiment, the carbon dust accumulation thickness is 12mm, which is between 10mm and 30mm. Therefore, the carbon dust pollution risk index = (12-10) ÷ (30-10) = 2 ÷ 20 = 0.1.

[0102] The carbon brush wear risk index and carbon dust pollution risk index are stored and output to the maintenance urgency assessment module 15 as input to the maintenance urgency assessment model.

[0103] The following results were achieved through this module:

[0104] By using piecewise linear mapping, the remaining length of the carbon brush and the thickness of carbon dust accumulation are converted into dimensionless carbon brush wear risk index and carbon dust pollution risk index, respectively, and unified to the range of 0 to 1. This eliminates the dimensional difference between the two risk indicators and provides standardized input for the subsequent maintenance urgency assessment model.

[0105] The maintenance urgency assessment module 15 is used to input the carbon brush wear risk index and the carbon dust pollution risk index into the pre-trained maintenance urgency assessment model, output the maintenance urgency coefficient, and use the maintenance urgency coefficient as the inspection result of the doubly fed wind turbine. The larger the maintenance urgency coefficient, the more urgently the slip ring chamber of the doubly fed wind turbine needs to be maintained.

[0106] The maintenance urgency assessment module 15 provided in this embodiment of the invention is used for:

[0107] The carbon brush wear risk index and the carbon dust pollution risk index are concatenated to form a maintenance urgency input feature vector;

[0108] The maintenance urgency input feature vector is input into a pre-trained maintenance urgency assessment model, and the maintenance urgency assessment model outputs a maintenance urgency coefficient, which is a continuous value between 0 and 1.

[0109] The maintenance urgency coefficient is used as the inspection result of the doubly-fed wind turbine. The larger the maintenance urgency coefficient, the more urgently the slip ring chamber of the doubly-fed wind turbine needs maintenance.

[0110] in,

[0111] The pre-training process of the maintenance urgency assessment model includes:

[0112] The carbon brush wear risk index, carbon dust pollution risk index, and corresponding maintenance urgency label value of the slip ring chamber of the doubly fed wind turbine were collected in historical maintenance events to form a maintenance urgency training sample set. The maintenance urgency label value is a continuous value between 0 and 1, which is determined by the operation and maintenance personnel based on actual maintenance records. The larger the value, the higher the maintenance urgency.

[0113] An initial maintenance urgency assessment model is constructed. The initial maintenance urgency assessment model is a regression model based on a fully connected neural network, which includes an input layer, a hidden layer and an output layer. The input layer has 2 neurons and the output layer has 1 neuron and uses the Sigmoid activation function.

[0114] Using the carbon brush wear risk index and carbon dust pollution risk index from the maintenance urgency training sample set as model inputs, the corresponding maintenance urgency label as the supervision label, and the mean square error as the loss function, the initial maintenance urgency assessment model is trained in a supervised manner until it is verified to converge, thus obtaining the pre-trained maintenance urgency assessment model.

[0115] Specifically, the carbon brush wear risk index and the carbon dust pollution risk index are concatenated to form the maintenance urgency input feature vector. The maintenance urgency input feature vector is denoted as Z=[Iwear ,I dust ], where I wear The carbon brush wear risk index output by the risk index calculation module 14, I dust The carbon dust pollution risk index is output by the risk index calculation module 14. Both indices are dimensionless quantities in the range of 0 to 1, requiring no additional normalization. For example, the carbon brush wear risk index I currently output by module 14 is... wear The carbon dust pollution risk index I is 0.607. dust If the value is 0.1, then the maintenance urgency input feature vector Z = [0.607, 0.1].

[0116] The carbon brush wear risk index, carbon dust pollution risk index, and corresponding maintenance urgency label value of the slip ring chamber of doubly-fed induction generator (DFIG) wind turbines were collected during historical maintenance events to form a training sample set for maintenance urgency. Historical maintenance events refer to slip ring chamber maintenance events recorded in the wind farm's operation and maintenance management system. Each maintenance event corresponds to a specific time point, at which the carbon brush wear risk index and carbon dust pollution risk index were calculated. The maintenance personnel then determined the maintenance urgency label value based on the maintenance records.

[0117] The maintenance urgency label value is determined based on the original measurement data recorded during the maintenance event according to the following rules: Obtain the remaining carbon brush length measurement, slip ring chamber insulation resistance measurement, and carbon dust accumulation thickness measurement. If the remaining carbon brush length measurement is less than or equal to the carbon brush limit length value, and the slip ring chamber insulation resistance measurement is less than or equal to 0.5 MΩ, then the maintenance urgency label value is set to 1.0; if the remaining carbon brush length measurement is greater than or equal to 80% of the initial carbon brush length value, and the slip ring chamber insulation resistance measurement is greater than or equal to 10 MΩ, and the carbon dust accumulation thickness measurement is less than or equal to the carbon dust safe thickness threshold, then the maintenance urgency label value is set to 0.0; otherwise, the maintenance urgency label value is calculated according to a weighted scoring formula. Among them, Y 维护 The calculated maintenance urgency label value, ΔL represents the carbon brush shortening amount, L 初始 L represents the initial length of the carbon brush. 极限 H represents the maximum length of the carbon brush. 尘 H represents the measured thickness of carbon dust deposits. 安全 H is the safe thickness threshold for carbon dust. 危险 R is the threshold for hazardous carbon dust thickness. 绝缘 This is the measured value of the insulation resistance of the slip ring chamber, in MΩ. When R 绝缘 If the value is greater than 10, it is counted as 10. The calculated Y 维护The weighted scoring formula is limited to a range of 0 to 1. In this formula, 0.5, 0.3, and 0.2 are the weighting coefficients for the three factors: carbon brush wear, carbon dust contamination, and insulation resistance. These weighting coefficients are determined based on the degree of influence of each factor on the urgency of maintenance in slip ring chamber maintenance practice: carbon brush wear is the direct cause of carbon brush failure and has the highest weight; carbon dust contamination affects insulation performance and has the second highest weight; insulation resistance reflects the overall insulation condition and has the lowest weight. A higher maintenance urgency label value indicates a higher level of maintenance urgency.

[0118] For example, a total of 800 historical maintenance events were collected. After determining the label values ​​according to the above rules, 48 ​​events with incomplete records were removed, and 752 valid samples were retained.

[0119] An initial maintenance urgency assessment model was constructed. This model is a regression model based on a fully connected neural network, consisting of one input layer, one hidden layer, and one output layer. The input layer has two neurons, corresponding to the two components of the maintenance urgency input feature vector Z. The hidden layer has eight neurons, followed by a ReLU activation function to introduce nonlinear transformation capability. The output layer has one neuron, using a Sigmoid activation function, outputting a scalar value between 0 and 1 as the maintenance urgency coefficient. The model has a total of 2×8 + 8×1 = 24 weights and 8 + 1 = 9 bias terms, for a total of 33 parameters.

[0120] The carbon brush wear risk index and carbon dust pollution risk index from the maintenance urgency training sample set are used as model inputs, with the corresponding maintenance urgency labels as supervision labels, and mean squared error (MSE) as the loss function. The MSE loss function is MSE=(1 / K)×Σ(u j -û j )², where K is the batch size of the training samples, u j Let û be the maintenance urgency label value for the j-th sample. j This represents the maintenance urgency coefficient output by the model. An adaptive moment estimator (AME) is used to update the model parameters, with a learning rate of 0.001 and a batch size of 32. The 752 valid samples are divided into a training set and a validation set in an 8:2 ratio, resulting in a training set of 602 samples and a validation set of 150 samples. After each training round, the validation mean squared error (MSE) is calculated on the validation set. Early stopping is triggered when the MSE fails to decrease for 10 consecutive rounds, halting training. The model parameters corresponding to the round with the smallest validation set loss are taken as the final parameters, resulting in the pre-trained maintenance urgency assessment model. For example, during training, the minimum MSE of the maintenance urgency assessment model on the validation set is 0.0062, and the mean absolute error is approximately 0.06.

[0121] The real-time acquired carbon brush wear risk index and carbon dust pollution risk index are concatenated and input into a pre-trained maintenance urgency assessment model, which then outputs a maintenance urgency coefficient. For example, if the current maintenance urgency input feature vector Z=[0.607,0.1] is input into the model, the maintenance urgency assessment model outputs a maintenance urgency coefficient of 0.68. The maintenance urgency coefficient is a continuous value between 0 and 1; a higher value indicates that the slip ring chamber of the doubly-fed induction generator (DFIG) wind turbine requires more urgent maintenance. The maintenance urgency coefficient is output as the inspection result of the DFIG wind turbine to the nacelle monitoring interface or the wind farm operation and maintenance management center, and maintenance personnel allocate maintenance priorities based on the coefficient.

[0122] The following technical effects have been achieved through this module:

[0123] This module integrates the carbon brush wear risk index and the carbon dust pollution risk index into a single maintenance urgency coefficient through a pre-trained maintenance urgency assessment model. This transforms the inspection results from two independent physical condition indicators into directly comparable maintenance priority values. The maintenance urgency assessment model outputs a continuous value from 0 to 1, providing maintenance personnel with a quantitative basis for determining the maintenance sequence.

Claims

1. A wind turbine nacelle intelligent inspection system, characterized in that, The system includes: The data acquisition module is used to continuously collect vibration data of the slip ring chamber through vibration sensors, continuously collect carbon dust accumulation thickness data at the bottom of the slip ring chamber through ultrasonic sensors, and continuously collect temperature data of the slip ring chamber through temperature sensors when the doubly fed wind turbine is in operation. The friction feature extraction module is used to perform time-frequency analysis on the vibration data of the slip ring chamber and extract the friction characteristic frequency and friction vibration amplitude generated by the friction between the carbon brush and the slip ring. The carbon brush remaining length estimation module is used to input the friction characteristic frequency, the friction vibration amplitude, the rotor speed and the slip ring chamber temperature data into the pre-trained lightweight carbon brush remaining length estimation model, and output the carbon brush remaining length estimation value. The risk index calculation module is used to calculate the carbon brush wear risk index based on the estimated remaining carbon brush length, and to calculate the carbon dust pollution risk index based on the carbon dust accumulation thickness data. The maintenance urgency assessment module is used to input the carbon brush wear risk index and the carbon dust pollution risk index into a pre-trained maintenance urgency assessment model, output a maintenance urgency coefficient, and use the maintenance urgency coefficient as the inspection result of the doubly-fed wind turbine. The larger the maintenance urgency coefficient, the more urgently the slip ring chamber of the doubly-fed wind turbine needs maintenance.

2. The intelligent inspection system for wind turbine nacelles according to claim 1, characterized in that, During the operation of the doubly-fed wind turbine, vibration data of the slip ring chamber is continuously collected using vibration sensors, carbon dust accumulation thickness data at the bottom of the slip ring chamber is continuously collected using ultrasonic sensors, and temperature data of the slip ring chamber is continuously collected using temperature sensors, including: The vibration sensor is attached to the mounting base on the outer shell of the slip ring chamber near the carbon brush bracket, and continuously collects the vibration data of the slip ring chamber at a preset acquisition frequency; An ultrasonic sensor is installed on the inner side of the housing at the bottom of the slip ring chamber, directly opposite the area below the carbon brush. The ultrasonic sensor emits ultrasonic pulses vertically downwards at the preset acquisition cycle and receives the echoes reflected by the upper surface of the carbon dust. The thickness of the carbon dust accumulation at the bottom of the slip ring chamber is calculated based on the time difference between emission and reception. The temperature sensor is installed inside the slip ring chamber near the contact surface between the slip ring and the carbon brush, and continuously collects slip ring chamber temperature data at the preset acquisition cycle.

3. The intelligent inspection system for wind turbine nacelles according to claim 1, characterized in that, Time-frequency analysis was performed on the vibration data of the slip ring chamber to extract the characteristic friction frequencies and amplitudes of frictional vibrations generated by the friction between the carbon brush and the slip ring, including: The vibration data of the slip ring chamber is subjected to a short-time Fourier transform. The short-time Fourier transform uses a Hanning window with a preset window length and a preset overlap rate between adjacent windows to obtain the time spectrum of the vibration data of the slip ring chamber. In the time spectrum, the friction characteristic frequency generated by the friction between the carbon brush and the slip ring is extracted, wherein the friction characteristic frequency is located within a preset carbon brush friction characteristic frequency band, and the preset carbon brush friction characteristic frequency band is from 800Hz to 3000Hz. At the frequency point corresponding to the friction characteristic frequency, the friction vibration amplitude is extracted, and the friction vibration amplitude is the amplitude spectrum value of the friction characteristic frequency in the short-time Fourier transform result.

4. The intelligent inspection system for wind turbine nacelles according to claim 3, characterized in that, In the time spectrum, the friction characteristic frequencies generated by the friction between the carbon brush and the slip ring are extracted, including: Obtain the current rotor speed and calculate the theoretical friction frequency generated by the friction between the carbon brush and the slip ring based on the rotor speed; In the time spectrum, the frequency point with the largest amplitude spectrum peak within the preset carbon brush friction characteristic frequency band is taken as the candidate friction characteristic frequency; The candidate friction characteristic frequency is compared with the theoretical friction frequency. When the deviation between the candidate friction characteristic frequency and the theoretical friction frequency is within a preset deviation range, the candidate friction characteristic frequency is taken as the friction characteristic frequency. When the deviation between the candidate friction characteristic frequency and the theoretical friction frequency exceeds the preset deviation range, the theoretical friction frequency is taken as the friction characteristic frequency.

5. The intelligent inspection system for wind turbine nacelles according to claim 1, characterized in that, The friction characteristic frequency, friction vibration amplitude, rotor speed, and slip ring chamber temperature data are input into a pre-trained lightweight carbon brush remaining length estimation model, and the output is an estimated value of the remaining carbon brush length, including: The friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data are concatenated into a carbon brush state input feature vector; The carbon brush state input feature vector is input into a pre-trained lightweight carbon brush remaining length estimation model, and the lightweight carbon brush remaining length estimation model outputs the carbon brush remaining length estimate, wherein the carbon brush remaining length estimate is a continuous value in millimeters. The pre-training process of the pre-trained lightweight carbon brush remaining length estimation model includes: Collect carbon brush status sample data of the slip ring chamber of the doubly fed wind turbine during historical operation. Each carbon brush status sample data includes friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data, as well as the corresponding carbon brush remaining length label value. The carbon brush remaining length label value is obtained by maintenance personnel using vernier calipers or special carbon brush length measuring tools to measure the actual remaining length of the carbon brush when the unit is stopped. The carbon brush status sample data is cleaned using the box plot method to obtain cleaned carbon brush status sample data, and the cleaned carbon brush status sample data is divided into training set and validation set according to a preset ratio. An initial deep carbon brush remaining length estimation model is constructed. The initial deep carbon brush remaining length estimation model is a regression model based on a fully connected neural network, which includes an input layer, multiple hidden layers and an output layer. The output layer outputs a scalar value as the predicted carbon brush remaining length. The friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data of each carbon brush state sample data in the training set are used as model input, the corresponding carbon brush remaining length label is used as supervision label, and the mean square error is used as loss function to supervise the training of the initial deep carbon brush remaining length estimation model. The validation set is used to judge early stop until the loss function on the validation set converges, and the pre-trained deep carbon brush remaining length estimation model is obtained. The pre-trained deep carbon brush remaining length estimation model is lightweighted to obtain a pre-trained lightweight carbon brush remaining length estimation model.

6. The intelligent inspection system for wind turbine nacelles according to claim 5, characterized in that, The pre-trained deep carbon brush remaining length estimation model is lightweighted to obtain a pre-trained lightweight carbon brush remaining length estimation model, including: An initial lightweight carbon brush remaining length estimation model is constructed. The initial lightweight carbon brush remaining length estimation model is a regression model based on a fully connected neural network, which includes an input layer, a hidden layer and an output layer. The input layer receives the carbon brush state input feature vector. The number of neurons in the hidden layer does not exceed 8 and is followed by a ReLU activation function. The output layer uses a linear activation function to output a scalar value as the carbon brush remaining length estimate. The pre-trained deep carbon brush remaining length estimation model is provided as a soft-label model, and the initial lightweight carbon brush remaining length estimation model is used as a lightweight model to be trained. Using the friction characteristic frequency, friction vibration amplitude, rotor speed and slip ring chamber temperature data of the carbon brush state sample data in the training set and the validation set as model input, the predicted carbon brush remaining length provided by the soft label as the soft label, and the corresponding carbon brush remaining length label value as the hard label, the lightweight model to be trained is subjected to knowledge distillation training. The mean square error between the estimated remaining carbon brush length output by the lightweight model to be trained and the soft label is calculated as the soft label loss, and the mean square error between the estimated remaining carbon brush length output by the lightweight model to be trained and the hard label is calculated as the hard label loss. The soft label loss and the hard label loss are weighted and summed to form a joint loss function. The parameters of the lightweight model to be trained are updated with the goal of minimizing the joint loss function until convergence is verified, thus obtaining a pre-trained lightweight carbon brush remaining length estimation model.

7. The intelligent inspection system for wind turbine nacelles according to claim 1, characterized in that, Based on the estimated remaining length of the carbon brush, the carbon brush wear risk index is calculated, including: The length of the carbon brush in its brand-new state is taken as the initial length value of the carbon brush, and the minimum length of the carbon brush that can be used is taken as the limit length value of the carbon brush. Subtracting the estimated remaining length of the carbon brush from the initial length value of the carbon brush yields the amount of carbon brush shortening. The carbon brush wear ratio is obtained by dividing the carbon brush shortening amount by the difference between the initial carbon brush length value and the carbon brush limit length value. The carbon brush wear ratio is limited to the range of 0 to 1, and the limited carbon brush wear ratio is used as the carbon brush wear risk index.

8. The intelligent inspection system for wind turbine nacelles according to claim 1, characterized in that, Based on the carbon dust accumulation thickness data, a carbon dust pollution risk index is calculated, including: The maximum allowable thickness of carbon dust accumulation that does not affect the insulation performance of the slip ring chamber is taken as the safe thickness threshold of carbon dust, and the thickness of carbon dust accumulation that seriously threatens the insulation performance of the slip ring chamber is taken as the dangerous thickness threshold of carbon dust. When the carbon dust accumulation thickness data is less than or equal to the carbon dust safe thickness threshold, the carbon dust pollution risk index is set to 0; When the carbon dust accumulation thickness data is greater than or equal to the carbon dust dangerous thickness threshold, the carbon dust pollution risk index is set to 1; When the carbon dust accumulation thickness data is between the carbon dust safe thickness threshold and the carbon dust dangerous thickness threshold, the carbon dust pollution risk index is obtained by subtracting the carbon dust safe thickness threshold from the carbon dust accumulation thickness data and dividing by the difference between the carbon dust dangerous thickness threshold and the carbon dust safe thickness threshold.

9. The intelligent inspection system for wind turbine nacelles according to claim 1, characterized in that, The carbon brush wear risk index and the carbon dust pollution risk index are input into a pre-trained maintenance urgency assessment model to output a maintenance urgency coefficient. This maintenance urgency coefficient is then used as the inspection result of the doubly-fed wind turbine, including: The carbon brush wear risk index and the carbon dust pollution risk index are concatenated to form a maintenance urgency input feature vector; The maintenance urgency input feature vector is input into a pre-trained maintenance urgency assessment model, and the maintenance urgency assessment model outputs a maintenance urgency coefficient, which is a continuous value between 0 and 1. The maintenance urgency coefficient is used as the inspection result of the doubly-fed wind turbine. The larger the maintenance urgency coefficient, the more urgently the slip ring chamber of the doubly-fed wind turbine needs maintenance.

10. The intelligent inspection system for wind turbine nacelles according to claim 9, characterized in that, The pre-training process of the maintenance urgency assessment model includes: The carbon brush wear risk index, carbon dust pollution risk index, and corresponding maintenance urgency label value of the slip ring chamber of the doubly fed wind turbine were collected in historical maintenance events to form a maintenance urgency training sample set. The maintenance urgency label value is a continuous value between 0 and 1, which is determined by the operation and maintenance personnel based on actual maintenance records. The larger the value, the higher the maintenance urgency. An initial maintenance urgency assessment model is constructed. The initial maintenance urgency assessment model is a regression model based on a fully connected neural network, which includes an input layer, a hidden layer and an output layer. The input layer has 2 neurons and the output layer has 1 neuron and uses the Sigmoid activation function. Using the carbon brush wear risk index and carbon dust pollution risk index from the maintenance urgency training sample set as model inputs, the corresponding maintenance urgency label as the supervision label, and the mean square error as the loss function, the initial maintenance urgency assessment model is trained in a supervised manner until it is verified to converge, thus obtaining the pre-trained maintenance urgency assessment model.

Images