Method and device for evaluating lubrication state of fan gearbox

Abstract

The invention provides a fan gearbox lubrication state evaluation method and device. The method comprises the steps that an impedance spectrum sensor collects electrochemical impedance data and temperature data of wind turbine generator gear oil; carrying out data preprocessing: extracting an impedance spectrum characteristic value, and carrying out temperature compensation and normalization; classifying the impedance spectrum characteristics according to the degradation grade of the gear oil and carrying out fuzzy comprehensive evaluation; preliminarily constructing an oil state evaluation model based on a neural network, inputting a fuzzy comprehensive evaluation result into the oil state evaluation model for training, and performing hyper-parameter tuning to obtain an optimized oil state evaluation model; electrochemical impedance data and temperature data of wind turbine generator gear oil are collected on line in real time, after data preprocessing and fuzzy comprehensive evaluation are conducted, a real-time fuzzy comprehensive evaluation result is input into the optimized oil state evaluation model, and a gear oil degradation state evaluation result is obtained; under the condition that limited sensors are arranged, the accuracy of lubricating state evaluation can be improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine gearbox lubrication monitoring technology, and in particular to a method and apparatus for assessing the lubrication status of wind turbine gearboxes. Background Technology

[0002] The gearbox is the core equipment of a wind turbine, and lubricating oil is the "blood" of the gearbox, playing a crucial role in maintaining its normal operation. The performance and quality of lubricating oil have become key factors affecting the safe and reliable operation and extended service life of wind turbines. Furthermore, as a carrier, lubricating oil contains rich tribological information about the moving surfaces of the equipment, making it an important information source for wind turbine gearbox fault diagnosis and health management. In recent years, online oil monitoring has become an important means of monitoring wind turbine gearbox faults.

[0003] However, due to the harsh operating environment and complex working conditions of wind turbines, lubricating oil is often forced to be replaced before reaching its average service life due to a combination of factors such as oxidation / shear aging, water emulsification, dust pollution, excessive wear, and diesel pollution. This places higher demands on online oil monitoring technology. However, current oil detection methods and sensors are still focused on detecting single impurities such as wear debris, moisture, viscosity, soot, corrosion, and sulfur content, making it difficult to achieve a comprehensive and accurate description of the deterioration process of wind turbine gear oil performance. Multi-sensor fusion monitoring solutions are limited by the installation constraints of equipment lubrication pipelines, making them difficult to implement in actual engineering.

[0004] Therefore, there is an urgent need for a method and device for assessing the lubrication status of wind turbine gearboxes, which can improve the accuracy of assessing the lubrication status of wind turbine gears under the condition of deploying a limited number of sensors. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for assessing the lubrication status of a wind turbine gearbox, aiming to solve the technical problem of inaccurate online monitoring of lubricating oil in traditional methods.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for evaluating the lubrication condition of a fan gearbox, comprising:

[0007] S1. Impedance spectroscopy sensor collects electrochemical impedance data and temperature data of wind turbine gear oil;

[0008] S2. Perform data preprocessing on the electrochemical impedance data, including extracting impedance spectrum feature values, and performing temperature compensation and normalization.

[0009] S3. Classify the impedance spectrum characteristics according to the deterioration level of the gear oil and perform fuzzy comprehensive evaluation;

[0010] S4. Initially construct an oil condition assessment model based on neural networks, input the obtained fuzzy comprehensive assessment results into the oil condition assessment model for training, and perform hyperparameter tuning on the oil condition assessment model to obtain an optimized oil condition assessment model.

[0011] S5. Online real-time acquisition of electrochemical impedance data of wind turbine gear oil. After data preprocessing and fuzzy comprehensive evaluation of the real-time electrochemical impedance data, the real-time fuzzy comprehensive evaluation results are input into the optimized oil condition evaluation model to obtain the gear oil deterioration condition evaluation results.

[0012] As a further improvement to the above scheme, the electrochemical impedance data of wind turbine gear oil collected includes:

[0013] Generate multiple small-amplitude sinusoidal current disturbance signals with preset frequencies;

[0014] Monitor the electrochemical response of the oil under a preset frequency disturbance and obtain response data.

[0015] As a further improvement to the above scheme, the impedance spectrum characteristic values ​​include impedance amplitudes and phase differences corresponding to the electrochemical response data at multiple preset frequencies.

[0016] As a further improvement to the above scheme, the temperature compensation is to convert the impedance amplitude and phase difference at different temperatures into the equivalent impedance amplitude and equivalent phase difference at the reference temperature through polynomial surface regression.

[0017] The normalization involves normalizing the equivalent impedance magnitude and the equivalent phase difference at the reference temperature to... Interval.

[0018] As a further improvement to the above scheme, the method for obtaining the equivalent impedance magnitude is as follows:

[0019] The reference temperature is set according to the working environment and lubrication mechanism of the wind turbine gearbox.

[0020] Select the impedance spectrum characteristic values ​​and the temperature data from N consecutive sampling points;

[0021] The difference between the temperature data extracted from the N sampling points and the reference temperature is within a preset error range (preferably). For data within the range of ℃, calculate the average value of the corresponding impedance spectrum characteristic values ​​and use it as the impedance spectrum reference value;

[0022] Calculate the impedance spectrum characteristic difference between the impedance spectrum characteristic value of the N sampling points and the impedance spectrum reference value, and the temperature difference between the temperature data and the reference temperature;

[0023] Substitute the impedance spectrum characteristic difference, impedance spectrum characteristic value, and temperature difference of the N data points into the polynomial surface regression model to calculate the parameters of the polynomial surface regression model.

[0024] A polynomial surface regression temperature compensation model is constructed using various parameters.

[0025] Calculate the impedance spectrum characteristic value and the temperature difference at each real-time sampling point, input them into the polynomial surface regression temperature compensation model, and obtain the impedance spectrum characteristic difference;

[0026] The equivalent impedance spectrum characteristic value of each sampling point is obtained by adding the impedance spectrum characteristic difference to the actual impedance spectrum characteristic value of the sampling point.

[0027] As a further improvement to the above scheme, the polynomial surface regression model is shown in the following equation:

[0028] ,in, Due to poor impedance spectrum characteristics, These are characteristic values ​​of the impedance spectrum. The temperature difference; These are the various parameters.

[0029] As a further improvement to the above scheme, the steps for classifying impedance spectrum characteristics according to the deterioration level of gear oil and performing fuzzy comprehensive evaluation are as follows:

[0030] Based on the normalized equivalent impedance magnitude and the equivalent phase difference, a fuzzy comprehensive evaluation factor set is established: ,in These are the impedance amplitudes at n preset frequencies. These are the phase differences at n preset frequencies, preferably n=5;

[0031] A fuzzy comprehensive evaluation set is established based on five deterioration levels of gear oil. ,in These correspond to five gear oil deterioration levels: new oil, mild deterioration, moderate deterioration, severe deterioration, and scrapping.

[0032] Construct the set of fuzzy comprehensive evaluation factors The i-th factor in the fuzzy comprehensive evaluation set Membership degree of the j-th state Composition of factor evaluation matrix ;

[0033] Based on the fuzzy comprehensive evaluation factor set The importance of each factor and expert knowledge are considered, and the weight vector of each factor is assumed to be... ;

[0034] Based on the factor evaluation matrix and the weight vector Calculate the fuzzy vector ,in, Let be the probability of belonging to the i-th state level. The evaluation result is the state corresponding to the maximum probability. To comprehensively evaluate the synthetic operators.

[0035] As a further improvement to the above scheme, the fuzzy comprehensive evaluation factor set The i-th impedance magnitude factor affects the fuzzy comprehensive evaluation set. Membership degree of the j-th state Defined as:

[0036] ,

[0037] Where x is the normalized equivalent impedance amplitude, and a, b, c, and d are single-factor evaluation judgment limits, determined based on the distribution of the equivalent impedance amplitude and in combination with expert knowledge.

[0038] As a further improvement to the above scheme, the fuzzy comprehensive evaluation factor set The i-th phase factor in the fuzzy comprehensive evaluation set Membership degree of the j-th state Defined as:

[0039] ,

[0040] Where x is the normalized equivalent phase difference, and a, b, c, and d are the single-factor evaluation judgment limits, determined based on the distribution of the equivalent phase difference and in combination with expert knowledge.

[0041] As a further improvement to the above scheme, in step S4, the construction of the oil condition assessment model based on the neural network specifically involves constructing a DNN model including a Dropout layer and using the Sparrow Search algorithm to perform hyperparameter tuning on the oil condition assessment model.

[0042] As a further improvement to the above scheme, the steps to obtain the optimized oil condition assessment model are as follows:

[0043] Add a Dropout layer before the first hidden layer of the DNN model to construct a DNN model that includes a Dropout layer.

[0044] The result of the fuzzy comprehensive evaluation is input into the DNN model for model training;

[0045] During the training of the DNN model, the Sparrow Search algorithm is used to find the optimal solution for the weights and biases of the DNN network;

[0046] The optimal solution of weights and biases is input into the DNN model to obtain the optimized oil state assessment model.

[0047] As a further improvement to the above scheme, the steps for constructing a DNN model including a Dropout layer are as follows:

[0048] A multilayer perceptron-based DNN model is constructed by stacking multiple hidden layers between the input and output layers. Each layer performs a nonlinear transformation through a nonlinear activation function, progressively extracting features and abstracting representations from the input data. The calculation formula for a single neuron in the DNN model is as follows:

[0049] ,

[0050] in, It is the output of the previous neuron. For the weight vector, For bias, Represents a node. It is an activation function;

[0051] A Dropout layer is added before the first hidden layer of the DNN model; the formula for calculating a single neuron in a DNN with a Dropout layer is: ,

[0052] in, Let be the Bernoulli probability vector. It is the probability in the Bernoulli function. This is the result of the Dropout layer calculation.

[0053] As a further improvement to the above scheme, the steps for finding the optimal solution for the DNN network weights and biases using the sparrow search algorithm during the DNN model training process are as follows:

[0054] Set the maximum number of iterations, safety threshold, and population size for the sparrow search algorithm;

[0055] Initialize a population of n sparrows;

[0056] The weights and bias hyperparameter vectors corresponding to each sparrow are substituted into the DNN model, and the fuzzy comprehensive evaluation results are cross-validated to obtain the cross-validation accuracy.

[0057] The accuracy of the cross-validation was used as the initial fitness value to obtain the initial fitness value of all sparrows in the population.

[0058] Sort all sparrows according to their fitness values, and divide them into discoverer sparrows and joiner sparrows according to a preset ratio;

[0059] Randomly select several sparrows as guard sparrows;

[0060] The position of the discoverer sparrow in the sparrow search algorithm is updated using the discoverer update formula;

[0061] The position of the newest sparrow in the sparrow search algorithm is updated using the newest sparrow update formula.

[0062] The position of the vigilant sparrow in the sparrow search algorithm is updated using the vigilant update formula;

[0063] Calculate the update fitness value of the sparrow after all update positions;

[0064] The updated fitness value of each sparrow is compared with the best initial fitness value among the initial fitness values ​​to obtain the new fitness value of the corresponding sparrow;

[0065] Preferably, if the updated fitness value is greater than or equal to the optimal initial fitness value, then the updated fitness value is used as the new fitness value for the corresponding sparrow; if the updated fitness value is less than the optimal initial fitness value, then the optimal initial fitness value is used as the new fitness value for the corresponding sparrow.

[0066] Update the global optimal position based on the new fitness value: until the maximum number of iterations is reached, output the global optimal position, and the coordinates of the global optimal position are the optimal solution for the weights and biases of the DNN model.

[0067] Secondly, the present invention also provides a wind turbine gearbox lubrication condition assessment device, comprising an impedance spectrum monitoring unit, a data preprocessing unit, a fuzzy comprehensive assessment unit, an assessment model training unit, and an online condition assessment unit connected in sequence. Each unit works together to complete real-time monitoring and intelligent assessment of the gear oil lubrication condition, as detailed below:

[0068] The impedance spectrum monitoring unit is used to apply multiple preset frequency small amplitude sinusoidal current disturbance signals to the oil in the gearbox of the wind turbine under test, and to simultaneously collect the electrochemical response signal and real-time temperature data of the oil through the electrochemical impedance spectrum detection module.

[0069] The data preprocessing unit is used to extract features from the acquired electrochemical response signal to obtain the impedance amplitude |Z| and phase difference θ; and based on a preset polynomial surface regression model, combined with oil temperature data, to perform temperature drift correction on the impedance amplitude and phase difference θ to generate the temperature-compensated equivalent impedance amplitude |Z'| and equivalent phase difference θ'; and to map |Z'| and θ' to the [0,1] interval using a normalization method to obtain the normalized equivalent impedance spectrum features.

[0070] The fuzzy comprehensive evaluation unit is used to construct a fuzzy evaluation index set based on the preset deterioration level standard of gear oil; to classify the deterioration level of the impedance spectrum characteristics after temperature compensation using the maximum membership method, and to output the fuzzy comprehensive evaluation results (level label and membership vector).

[0071] The evaluation model training unit is used to construct a deep neural network (DNN) model with a Dropout layer as a regularization component as an initial oil condition evaluation model; the initial model is trained with fuzzy comprehensive evaluation results as supervision labels, and the hyperparameters of the model are optimized using the sparrow search algorithm to obtain the optimized oil condition evaluation model.

[0072] The online condition assessment unit is used to process the real-time collected electrochemical impedance data through a data preprocessing unit to complete temperature compensation and normalization, and a fuzzy comprehensive assessment unit to complete the degradation level classification. Then, the data is input into the optimized oil condition assessment model for feature learning and condition prediction, and the real-time degradation condition assessment result of the gear oil is output.

[0073] As a further improvement to the above scheme, the impedance spectrum monitoring unit includes an impedance spectrum sensor, which includes an impedance spectrum detection subunit and a temperature detection subunit.

[0074] As a further improvement to the above solution, the data preprocessing unit includes:

[0075] The feature extraction subunit is used to extract the impedance amplitude and phase difference at a preset frequency from the oil electrochemical response data;

[0076] The temperature compensation subunit is used to call the polynomial surface regression temperature compensation model to calculate the equivalent impedance spectrum eigenvalues ​​at the reference temperature, including the equivalent impedance amplitude and the equivalent phase difference, and to periodically update the polynomial surface regression temperature compensation model.

[0077] Normalization subunit: Normalizes the equivalent impedance spectrum eigenvalues ​​to Interval.

[0078] Because the present invention adopts the above technical solutions, the beneficial effects of this application are as follows:

[0079] This invention provides a method for assessing the lubrication condition of a wind turbine gearbox. Traditional online lubricant monitoring technologies often rely on single abrasive particle counts or conventional physicochemical indicators (such as viscosity and acid value) for condition assessment, which can only reflect localized characteristics of oil deterioration and cannot comprehensively characterize the complex evolution process of gear oil deterioration. This invention innovatively introduces an impedance spectroscopy sensor. By collecting electrochemical impedance data of gear oil and extracting impedance spectral feature values ​​(such as impedance amplitude and phase difference), it can simultaneously track the synergistic effects of multiple deterioration mechanisms, such as the accumulation of oxidation products, additive consumption, and metal particle contamination in the oil. Compared to single-indicator monitoring, the introduction of impedance spectroscopy features upgrades the assessment dimension from single-point observation to multi-dimensional profiling, which can significantly improve the completeness and accuracy of the deterioration process description.

[0080] Furthermore, due to the complex operating environment of wind turbine gearboxes, fluctuations in oil temperature significantly affect the stability of impedance spectrum measurement results, and dimensional differences between different sensors or batches can also interfere with the accuracy of subsequent feature extraction. Therefore, this invention specifically designs temperature compensation and normalization strategies in the data preprocessing stage: by establishing a temperature-impedance spectrum correction model (polynomial surface regression model), the nonlinear influence of temperature on impedance parameters is eliminated; a normalization method is used to unify the dimensions, mapping impedance spectrum data under different operating conditions to the same feature space. After this processing, the repeatability and comparability of the data are effectively guaranteed, laying a reliable data foundation for subsequent fuzzy comprehensive evaluation and model training, and avoiding evaluation bias caused by environmental interference or equipment differences.

[0081] Furthermore, gear oil degradation is a continuous and gradual nonlinear process. Traditional methods often employ binary judgments of "meeting standards / not meeting standards" or discrete level classifications (such as good, caution, and abnormal), which are difficult to quantify the continuous changes in the degree of degradation and cannot describe the correlation between different degradation indicators. This invention proposes a fuzzy comprehensive evaluation method based on five degradation levels (such as new oil, mild degradation period, moderate degradation period, severe degradation period, and scrapping period). By defining the membership function of each level, the impedance spectrum characteristics and temperature data are mapped to the fuzzy membership of each level, and finally the continuous membership vector reflecting the degree of degradation is output. This mechanism breaks through the limitations of traditional discrete evaluation and can more accurately characterize the gradual evolution of oil from health to failure, providing a more discriminative labeling system for the training of subsequent intelligent models.

[0082] To address the scarcity of gear oil degradation samples in industrial settings, this invention constructs a deep neural network (DNN) model incorporating a Dropout layer and combines it with a sparrow search algorithm for hyperparameter optimization. On one hand, the Dropout layer forces the model to learn more robust feature representations by randomly deactivating some neurons during training, effectively suppressing the risk of overfitting under small sample conditions. On the other hand, the sparrow search algorithm (SSA) simulates the foraging and vigilance behavior of sparrows to quickly search for the optimal hyperparameter combination globally, significantly improving parameter optimization efficiency and model generalization ability compared to traditional grid search or random search methods. Experiments show that the optimized FDNN model, with the same sample size, achieves improved degradation state identification accuracy compared to the traditional BP neural network, reduces the risk of overfitting, and can more reliably identify early degradation trends in gear oil.

[0083] This invention integrates data preprocessing, fuzzy comprehensive evaluation, and optimization models into an online monitoring system, achieving full automation from electrochemical impedance data acquisition to degradation status output. By acquiring real-time electrochemical impedance data of gear oil, and inputting it into the optimization model after preprocessing and fuzzy evaluation, the system can output the current oil degradation level and development trend within 5 seconds. This real-time evaluation capability provides crucial support for the intelligent operation and maintenance of wind turbine units. Maintenance personnel can dynamically adjust the oil change cycle based on the evaluation results, avoiding resource waste caused by excessive oil changes; or trigger early warnings before the degradation trend worsens, allowing for timely shutdown and maintenance, and reducing downtime losses due to gearbox failure caused by lubrication failure.

[0084] This invention systematically solves the problems of insufficient accuracy and biased evaluation in traditional online lubricant monitoring by using multi-dimensional feature extraction of impedance spectrum, temperature compensation preprocessing, fuzzy comprehensive evaluation mechanism and FDNN intelligent model collaborative design. It has significant practical value and application prospects in the field of wind turbine gearbox lubrication condition assessment. Attached Figure Description

[0085] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0086] Figure 1 This is a flowchart illustrating a method for evaluating the lubrication status of a wind turbine gearbox disclosed in this invention.

[0087] Figure 2 This is a schematic diagram of the data preprocessing process disclosed in this invention;

[0088] Figure 3 This is a schematic diagram of the process for constructing and updating the temperature compensation model disclosed in this invention;

[0089] Figure 4 This is a schematic diagram of the fuzzy comprehensive evaluation process for impedance spectrum characteristics disclosed in this invention;

[0090] Figure 5 This is a schematic diagram of the DNN-based evaluation model construction process disclosed in this invention;

[0091] Figure 6 This is a schematic diagram of the process of finding the optimal solution for the weights and biases of a DNN network using the sparrow search algorithm disclosed in this invention;

[0092] Figure 7 This is a schematic diagram of the structure of a fan gearbox lubrication condition assessment device disclosed in this invention;

[0093] Figure 8 This is a schematic diagram of the data preprocessing unit disclosed in this invention;

[0094] Figure 9 This is a schematic diagram of the membership function curves of the impedance spectrum characteristics disclosed in this invention;

[0095] Figure 10 This is a schematic diagram of a single neuron and a single neuron with a Dropout node in the DNN model disclosed in this invention.

[0096] Figure 11 This is a schematic diagram of the training results of a method for evaluating the lubrication status of a wind turbine gearbox disclosed in this invention;

[0097] Figure 12 This is a schematic diagram of the test results of a method for evaluating the lubrication status of a wind turbine gearbox disclosed in this invention.

[0098] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0099] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0100] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0101] Example 1

[0102] See Figure 1 This invention provides a method for assessing the lubrication condition of a wind turbine gearbox. Through multi-dimensional data acquisition, preprocessing, feature extraction, fuzzy evaluation, and neural network model optimization, it achieves accurate online assessment of the gear oil's deterioration state. The following detailed description of each step, in conjunction with the accompanying drawings, illustrates the process.

[0103] Specifically, the steps include the following:

[0104] Step S1: Acquisition of electrochemical impedance and temperature data

[0105] The synchronous acquisition of electrochemical impedance data and temperature data of gear oil is achieved using an impedance spectroscopy sensor, specifically including the following sub-steps:

[0106] S11, Sensor Placement and Signal Generation

[0107] An electrochemical impedance spectroscopy sensor with a three-electrode system (preferably the ZXEIS impedance spectroscopy oil sensor) includes a working electrode, a reference electrode, and an auxiliary electrode; the sensor is fixed to the gearbox oil sump wall via a threaded interface, ensuring that the working electrode is completely immersed in the gear oil;

[0108] The signal generator generates multiple small-amplitude sinusoidal current disturbance signals with preset frequencies ranging from 0Hz to 1000Hz (covering the low-frequency oxidation characteristics to the high-frequency interface polarization characteristics of gear oil; specifically, in this embodiment, five preset frequencies are set, namely 0.15Hz, 1Hz, 10Hz, 100Hz and 1000Hz). The disturbance signals are applied between the working electrode and the auxiliary electrode via a potentiostat, and the reference electrode is used for stabilizing the potential reference.

[0109] S12, Electrochemical Response Monitoring and Data Acquisition

[0110] Electrochemical response data of the working electrode under different frequency perturbations were synchronously acquired using an electrochemical workstation, including response current (I) and perturbation voltage (V); the acquisition frequency was synchronized with the perturbation frequency, and the acquisition time for each frequency point was 10 cycles;

[0111] Temperature data is obtained by measuring the temperature probe integrated with the ZXEIS impedance spectroscopy oil sensor;

[0112] The sampling time interval for the ZXEIS impedance spectroscopy oil sensor to obtain the oil electrochemical response data and temperature data is 5 minutes.

[0113] The final collected raw data is in matrix form, as shown below:

[0114]

[0115] in, For each preset frequency (Hz) of the i-th term, For frequency Electrochemical impedance at the following levels ( ), The oil temperature (°C) at the time of sampling is given, N is the total number of sampling points, and M is the feature dimension (including frequency, impedance amplitude, phase difference, and temperature).

[0116] Step S2: Electrochemical impedance data preprocessing

[0117] See Figure 2 Preprocessing includes impedance spectrum eigenvalue extraction, temperature compensation, and normalization, as detailed below:

[0118] S21, Extraction of impedance spectrum eigenvalues

[0119] The impedance spectrum feature values ​​include impedance amplitudes and phase differences corresponding to electrochemical response data at multiple preset frequencies. Specifically, in this embodiment, the impedance amplitudes |Z| and phase differences θ corresponding to electrochemical response data at five preset frequencies (0.15Hz, 1Hz, 10Hz, 100Hz and 1000Hz) are extracted respectively.

[0120] S22, Temperature Compensation

[0121] The impedance characteristics of gear oil are significantly affected by temperature. The temperature compensation is achieved by converting the impedance amplitude and phase difference at different temperatures into the equivalent impedance amplitude and equivalent phase difference at the reference temperature through a polynomial surface regression model.

[0122] Specifically, for the impedance amplitude and phase difference of each sampling point, firstly, the difference between the real-time temperature measurement value and the reference temperature is calculated. Secondly, the impedance amplitude, phase difference, and temperature difference are input into the constructed polynomial surface regression temperature compensation model to obtain the impedance amplitude difference and phase difference. Finally, the impedance amplitude difference and phase difference are added to the impedance amplitude and phase difference of the sampling point to obtain the equivalent impedance amplitude and equivalent phase difference of the sampling point at the reference temperature.

[0123] S23, Normalization

[0124] To eliminate the influence of dimensions, the equivalent impedance amplitude and the equivalent phase difference at the reference temperature are normalized to the [0,1] interval. Preferably, the minimum-maximum normalization formula is used for normalization.

[0125] Step S3: Fuzzy comprehensive evaluation based on degradation level;

[0126] The normalized feature vectors are mapped to gear oil deterioration levels (new oil, mild deterioration, moderate deterioration, severe deterioration, and scrapping), and qualitative assessment is achieved through fuzzy mathematics.

[0127] Step S4: Neural Network Model Construction and Hyperparameter Tuning

[0128] A preliminary oil condition assessment model based on neural networks was constructed. The obtained fuzzy comprehensive assessment results were input into the oil condition assessment model for training. The hyperparameters of the oil condition assessment model were then tuned to obtain an optimized oil condition assessment model.

[0129] Specifically, in this embodiment, the construction of the oil condition assessment model based on neural networks involves constructing a DNN model including a Dropout layer; and using the Sparrow Search algorithm to perform hyperparameter tuning on the oil condition assessment model to obtain an optimized oil condition assessment model.

[0130] Step S5: Online Real-Time Evaluation and Result Output

[0131] The electrochemical impedance data of the gear oil of the wind turbine is collected online in real time. After data preprocessing and fuzzy comprehensive evaluation of the real-time electrochemical impedance data, the real-time fuzzy comprehensive evaluation results are input into the optimized oil condition evaluation model to obtain the gear oil deterioration condition evaluation results.

[0132] S51. Real-time data acquisition and preprocessing

[0133] Real-time electrochemical impedance spectroscopy data and oil temperature data are collected once per hour using an online impedance spectroscopy sensor (same model as sensor S1) deployed in the gearbox. The preprocessing procedure is the same as step S2 (feature extraction → temperature compensation → normalization) to obtain the real-time fuzzy comprehensive evaluation result B. real ;

[0134] S52: Real-time Status Assessment

[0135] B real Input the optimized oil condition assessment model, output the probability distribution of each deterioration level, determine the real-time deterioration state according to the maximum probability principle, and thus obtain the gear oil deterioration state assessment result;

[0136] Specifically, for ease of viewing by operators, the status assessment results can be divided into five states: "Excellent," "Good," "Medium," "Poor," and "Scrapped." "Excellent" corresponds to the "New Oil" degradation level, indicating that the oil has not shown a significant trend of performance degradation. "Good" corresponds to the "Slight Degradation Period" degradation level, indicating that the oil is in a state of slight performance degradation. "Medium" corresponds to the "Moderate Degradation Period" degradation level, indicating that the oil has shown a relatively obvious trend of performance degradation. "Poor" corresponds to the "Severe Degradation Period" degradation level, indicating that the oil is in a state of relatively serious performance degradation. "Scrapped" corresponds to the "Scrapped Period" degradation level, indicating that the oil's performance is in a state of extremely serious degradation and needs to be replaced in time.

[0137] S53: Early Warning and Recording

[0138] When the assessment result indicates a severe degradation period (probability > 80%), an audible and visual alarm is triggered and pushed to maintenance personnel via the industrial internet platform; at the same time, real-time data and assessment results are stored in the database for subsequent model iteration and updates.

[0139] This invention innovatively introduces an impedance spectroscopy sensor. By collecting electrochemical impedance data of gear oil and extracting impedance spectral feature values ​​(such as impedance amplitude and phase difference), it can simultaneously track the synergistic effects of multiple degradation mechanisms in the oil, such as the accumulation of oxidation products, additive consumption, and metal particle contamination. Compared with single-index monitoring, the introduction of impedance spectral features upgrades the evaluation dimension from single-point observation to multi-dimensional profiling, which can significantly improve the completeness and accuracy of the degradation process description.

[0140] Furthermore, due to the complex operating environment of wind turbine gearboxes, fluctuations in oil temperature significantly affect the stability of impedance spectrum measurement results, and dimensional differences between different sensors or batches can also interfere with the accuracy of subsequent feature extraction. Therefore, this invention specifically designs temperature compensation and normalization strategies in the data preprocessing stage: by establishing a temperature-impedance spectrum correction model (polynomial surface regression model), the nonlinear influence of temperature on impedance parameters is eliminated; a normalization method is used to unify the dimensions, mapping impedance spectrum data under different operating conditions to the same feature space. After this processing, the repeatability and comparability of the data are effectively guaranteed, laying a reliable data foundation for subsequent fuzzy comprehensive evaluation and model training, and avoiding evaluation bias caused by environmental interference or equipment differences.

[0141] Furthermore, gear oil degradation is a continuous and gradual nonlinear process. Traditional methods often employ binary judgments of "meeting standards / not meeting standards" or discrete level classifications (such as good, caution, and abnormal), which are difficult to quantify the continuous changes in the degree of degradation and cannot describe the correlation between different degradation indicators. This invention proposes a fuzzy comprehensive evaluation method based on five degradation levels (such as new oil, mild degradation period, moderate degradation period, severe degradation period, and scrapping period). By defining the membership function of each level, the impedance spectrum characteristics and temperature data are mapped to the fuzzy membership of each level, and finally the continuous membership vector reflecting the degree of degradation is output. This mechanism breaks through the limitations of traditional discrete evaluation and can more accurately characterize the gradual evolution of oil from health to failure, providing a more discriminative labeling system for the training of subsequent intelligent models.

[0142] To address the scarcity of gear oil degradation samples in industrial settings, this invention constructs a deep neural network (DNN) model incorporating a Dropout layer and combines it with a sparrow search algorithm for hyperparameter optimization. On one hand, the Dropout layer forces the model to learn more robust feature representations by randomly deactivating some neurons during training, effectively suppressing the risk of overfitting under small sample conditions. On the other hand, the sparrow search algorithm (SSA) simulates the foraging and vigilance behavior of sparrows to quickly search for the optimal hyperparameter combination globally, significantly improving parameter optimization efficiency and model generalization ability compared to traditional grid search or random search methods. Experiments show that the optimized FDNN model, with the same sample size, achieves improved degradation state identification accuracy compared to the traditional BP neural network, reduces the risk of overfitting, and can more reliably identify early degradation trends in gear oil.

[0143] This invention integrates data preprocessing, fuzzy comprehensive evaluation, and optimization models into an online monitoring system, achieving full automation from electrochemical impedance data acquisition to degradation status output. By acquiring real-time electrochemical impedance data of gear oil, and inputting it into the optimization model after preprocessing and fuzzy evaluation, the system can output the current oil degradation level and development trend within 5 seconds. This real-time evaluation capability provides crucial support for the intelligent operation and maintenance of wind turbine units. Maintenance personnel can dynamically adjust the oil change cycle based on the evaluation results, avoiding resource waste caused by excessive oil changes; or trigger early warnings before the degradation trend worsens, allowing for timely shutdown and maintenance, and reducing downtime losses due to gearbox failure caused by lubrication failure.

[0144] This invention systematically solves the problems of insufficient accuracy and biased evaluation in traditional online lubricant monitoring by using multi-dimensional feature extraction of impedance spectrum, temperature compensation preprocessing, fuzzy comprehensive evaluation mechanism and FDNN intelligent model collaborative design. It has significant practical value and application prospects in the field of wind turbine gearbox lubrication condition assessment.

[0145] As a preferred embodiment, see Figure 3 The steps for obtaining the equivalent impedance magnitude are as follows:

[0146] The reference temperature is set according to the working environment and lubrication mechanism of the wind turbine gearbox.

[0147] Select the impedance spectrum characteristic values ​​and the temperature data from N consecutive sampling points;

[0148] The difference between the temperature data extracted from the N sampling points and the reference temperature is within a preset error range (preferably). For data within the range of ℃, calculate the average value of the corresponding impedance spectrum characteristic values ​​and use it as the impedance spectrum reference value;

[0149] Calculate the impedance spectrum characteristic difference between the impedance spectrum characteristic value of the N sampling points and the impedance spectrum reference value, and the temperature difference between the temperature data and the reference temperature;

[0150] Substituting the impedance spectrum characteristic difference, impedance spectrum characteristic value, and temperature difference of the N data points into the polynomial surface regression model, the parameters of the polynomial surface regression model are calculated; specifically, the polynomial surface regression model is shown in the following equation:

[0151] ,in, Due to poor impedance spectrum characteristics, These are characteristic values ​​of the impedance spectrum. The temperature difference; These are the various parameters;

[0152] Using various parameters Construct a polynomial surface regression temperature compensation model;

[0153] Calculate the impedance spectrum characteristic value and the temperature difference at each real-time sampling point, input them into the polynomial surface regression temperature compensation model, and obtain the impedance spectrum characteristic difference;

[0154] The equivalent impedance spectrum characteristic value of each sampling point is obtained by adding the impedance spectrum characteristic difference to the actual impedance spectrum characteristic value of the sampling point.

[0155] This invention significantly improves the accuracy of equivalent conversion of gear oil impedance spectrum characteristics at different temperatures by constructing a polynomial surface regression temperature compensation model, providing high-quality input data for subsequent fuzzy comprehensive evaluation and neural network state prediction.

[0156] As a preferred embodiment, see Figure 4 The steps for classifying impedance spectrum characteristics according to the deterioration level of gear oil and performing fuzzy comprehensive evaluation are as follows:

[0157] Based on the normalized equivalent impedance magnitude and the equivalent phase difference, a fuzzy comprehensive evaluation factor set is established: ,in These are the impedance amplitudes at n preset frequencies. These are the phase differences at n preset frequencies, where n=5 in this embodiment;

[0158] A fuzzy comprehensive evaluation set is established based on five deterioration levels of gear oil. ,in These correspond to five gear oil degradation levels: v1 represents new oil (initial undegraded state), v2 represents mild degradation (slight increase in impedance amplitude / phase difference), v3 represents moderate degradation (significant increase in impedance amplitude / phase difference), v4 represents severe degradation (drastic increase in impedance amplitude / phase difference), and v5 represents the end of the service life (impedance amplitude / phase difference exceeds the safety threshold, requiring immediate oil replacement).

[0159] Construct the set of fuzzy comprehensive evaluation factors The i-th factor in the fuzzy comprehensive evaluation set Membership degree of the j-th state Composition of factor evaluation matrix ;

[0160] Specifically, the membership function curves for impedance spectrum characteristics can be found in [reference needed]. Figure 9 As shown, Figure 9 (a) shows the impedance membership function curves corresponding to each state level at frequencies of 1Hz, 2Hz, 3Hz, 4Hz, and 5Hz. Figure 9 (b) shows the phase difference membership function curves corresponding to each state level at frequencies of 1Hz, 2Hz, 3Hz, 4Hz and 5Hz;

[0161] The fuzzy comprehensive evaluation factor set The i-th impedance magnitude factor affects the fuzzy comprehensive evaluation set. Membership degree of the j-th state Defined as:

[0162] ,

[0163] Where x is the normalized equivalent impedance amplitude, and a, b, c, and d are single-factor evaluation judgment limits, determined based on the distribution of the equivalent impedance amplitude and in combination with expert knowledge.

[0164] The fuzzy comprehensive evaluation factor set The i-th phase factor in the fuzzy comprehensive evaluation set Membership degree of the j-th state Defined as:

[0165] ,

[0166] Where x is the normalized equivalent phase difference, and a, b, c, and d are the single-factor evaluation judgment limits, which are determined based on the distribution of the equivalent phase difference and in combination with expert knowledge.

[0167] Specifically, in this embodiment, the single-factor evaluation judgment thresholds a, b, c, and d are determined according to the following table:

[0168]

[0169] For example, at a 1Hz impedance value, corresponding to new oil Given the state, d = 0.7735, and so on, obtain the corresponding values ​​for a, b, c, and d;

[0170] Based on the fuzzy comprehensive evaluation factor set The importance of each factor and expert knowledge are considered, and the weight vector of each factor is assumed to be... ;

[0171] Based on the factor evaluation matrix and the weight vector Calculate the fuzzy vector ,in, Let be the probability of belonging to the i-th state level. The evaluation result is the state corresponding to the maximum probability. To comprehensively evaluate the synthesis operator;

[0172] This invention achieves accurate and reliable assessment of gear oil degradation status through multi-feature fusion fuzzy comprehensive evaluation, providing key technical support for predictive maintenance of wind turbine gearboxes, and has significant engineering application value and innovation.

[0173] As a preferred embodiment, see Figure 5 The steps to obtain the optimized oil condition assessment model are as follows:

[0174] S411. Construct a DNN model that includes a Dropout layer.

[0175] To improve the model's ability to generalize to complex oil state data, a Dropout layer is introduced before the first hidden layer of the DNN model.

[0176] Specifically, the steps for constructing a DNN model that includes a Dropout layer are as follows:

[0177] A DNN model based on a multilayer perceptron is constructed, stacking multiple hidden layers between the input and output layers. Each layer performs a nonlinear transformation through a nonlinear activation function, progressively extracting features and abstracting representations from the input data. A single neuron in the DNN model is shown below. Figure 10 In Figure (a), the calculation formula for a single neuron in the DNN model is:

[0178] ,

[0179] in, It is the output of the previous neuron. For the weight vector, For bias, Represents a node. It is an activation function;

[0180] Add a Dropout layer before the first hidden layer of the DNN model; for details on a single neuron in a DNN with Dropout nodes, see [link to DNN model documentation]. Figure 10 In Figure (b), specifically, the calculation formula for a single neuron in a DNN with a Dropout layer is as follows: ,

[0181] in, Let be the Bernoulli probability vector. It is the probability in the Bernoulli function. This is the result of the Dropout layer calculation;

[0182] Dropout layers force the model to learn more robust feature representations by randomly deactivating some neurons during training, thus avoiding overfitting caused by co-dependencies between neurons.

[0183] S412. Train the DNN model using the fuzzy comprehensive evaluation results.

[0184] The probability vector B output by the fuzzy comprehensive evaluation is used as input, and combined with the historical oil deterioration level labels (v1~v5), the DNN model is trained under supervision. The specific process is as follows:

[0185] (1) Data preparation

[0186] Training set: Select gear oil samples (≥10,000 sets) covering the entire life cycle from historical operating data, of which 70% are used for training, 20% for validation, and 10% for testing;

[0187] Labels: Each sample group corresponds to a manually labeled degradation level (e.g., "moderate degradation" corresponds to v3);

[0188] Normalization: The dimensions of the input vector B have been normalized and no additional processing is required;

[0189] (2) Loss function and optimizer initialization

[0190] Loss function: Cross-entropy loss function is used (to measure the difference between the predicted probability and the true label).

[0191] Optimizer: Initialized as Adam optimizer (learning rate 0.001, momentum parameters β1=0.9, β2=0.999);

[0192] (3) Training process

[0193] Forward propagation: Input B, calculate through hidden layer 1 (including Dropout) and hidden layer 2, and output the probability y^ of each level;

[0194] Calculate the loss: Calculate the loss value between the current prediction and the true label according to the Loss formula;

[0195] Backpropagation: Calculates the gradient of the loss with respect to each parameter (weights, biases) using the chain rule;

[0196] Parameter update: Update weights and biases using the Adam optimizer;

[0197] S413. Optimize DNN parameters using the Sparrow Search algorithm.

[0198] To overcome the local optima limitation of the traditional gradient descent algorithm, the Sparrow Search Algorithm (SSA) is introduced into the DNN training process to globally search for the optimal solution of weights and biases.

[0199] S414. Obtain the optimized oil condition assessment model.

[0200] The optimal weights and biases found by the sparrow search algorithm are input into the DNN model to replace the initial training parameters, ultimately resulting in the optimized oil state assessment model.

[0201] The Dropout layer breaks the co-dependency between neurons by randomly deactivating neurons, forcing the model to learn more general feature patterns. Experimental data show that after adding the Dropout layer, the generalization error of the model is reduced by 25% under complex working conditions such as temperature fluctuations (-10℃~60℃) and oil sludge pollution (particle concentration ≤500ppm), which is 18% better than the traditional DNN model without Dropout.

[0202] The Sparrow Search algorithm simulates the foraging and alertness behaviors of sparrow flocks, enabling a global search within the parameter space. This avoids the local optima problem caused by improper initial parameter or learning rate selection in traditional gradient descent algorithms. In a test of a 3MW onshore wind turbine gearbox, the model optimized by Sparrow Search achieved an accuracy of 97.2%, a 2.7% improvement over the model trained solely by the Adam optimizer (94.5%).

[0203] The Sparrow Search algorithm dynamically adjusts parameters based on the fitness of the validation set, enabling it to automatically adapt to the nonlinear distribution of oil state data (such as the non-monotonic changes in impedance amplitude and phase difference). In a test of an offshore wind turbine (high salt spray, high humidity), the model achieved an accuracy of 95.8% in identifying degradation caused by salt spray pollution, a significant improvement over the DNN model with a fixed learning rate (89.3%).

[0204] The optimized model has only 60% of the parameters of a traditional deep model (due to the reduction of redundant neurons in the Dropout layer), with a computation latency of ≤50ms. It can be directly embedded into wind turbine online monitoring systems, meeting real-time requirements. Field tests show that after model deployment, the adoption rate of evaluation results by maintenance personnel reached 98%, effectively improving equipment maintenance efficiency.

[0205] This method constructs a DNN model with a Dropout layer and combines it with a sparrow search algorithm to globally optimize parameters, achieving high accuracy and strong robustness in the oil condition assessment model. It provides key technical support for predictive maintenance of wind turbine gearboxes and has significant engineering application value and innovation.

[0206] As a preferred embodiment, see Figure 6 The steps for finding the optimal solution for the DNN network weights and biases using the sparrow search algorithm during the DNN model training process are as follows:

[0207] S421. Set the maximum number of iterations, safety threshold, and population size for the sparrow search algorithm;

[0208] S422. Initialize a population of n sparrows: , where g represents the dimension of the variable to be optimized, and the position of each sparrow in the population corresponds to a hyperparameter vector composed of the weights and biases of the DNN model;

[0209] S423. Substitute the weights and bias hyperparameter vectors corresponding to each sparrow into the DNN model, perform cross-validation on the fuzzy comprehensive evaluation results, and obtain the cross-validation accuracy.

[0210] S424. Using the accuracy of the cross-validation as the initial fitness value f0, the fitness values ​​of all sparrows in the population are obtained as follows:

[0211] ;

[0212] S425. Sort all sparrows according to their fitness values ​​and divide them into discoverer sparrows and joiner sparrows according to a preset ratio.

[0213] S426. Randomly select several sparrows as guard sparrows;

[0214] S427. Update the position of the discoverer sparrow in the sparrow search algorithm using the discoverer update formula. Specifically, the discoverer update formula is as follows: Where t represents the number of iterations. It is the maximum number of iterations. Let i be the position of the i-th sparrow in the j-th dimension. Let be the position of the i-th sparrow in the j-th dimension after t+1 iterations. Let Q be a random number following a normal distribution, L be a 1×d row matrix with all elements equal to 1, and R² and ST represent the warning value and the safety value, respectively. , ;

[0215] S428. Update the position of the newest sparrow in the sparrow search algorithm using the newest member update formula. Specifically, the newest member update formula is as follows: ,in, It is the position of the s-th sparrow in the v-th dimension. , These represent the positions of the s-th sparrow in dimension v after the t-th and t+1-th iterations, respectively. Indicates the optimal position of the discoverer. It is the optimal position of the discoverer after t+1 iterations. The position of the worst global position; A is a 1×d row matrix with elements of 1 or -1. ;

[0216] S429. Update the position of the vigilant sparrow in the sparrow search algorithm using the vigilant update formula. Specifically, the vigilant update formula is as follows: ,in, It is the position of the a-th sparrow in dimension d. , These represent the positions of the a-th sparrow in the d-th dimension after the t-th and t+1-th iterations, respectively. It is the current global optimal position, and β is the step size control parameter. It is a random number. It is the fitness value of an individual sparrow in the population. The current global best fitness value, The current global worst fitness value is represented by ε, which is a minimum constant to avoid zero in the denominator.

[0217] S430. Calculate the update fitness value of the sparrows after all update positions;

[0218] S431. Compare the updated fitness value of each sparrow with the optimal initial fitness value among the initial fitness values. If the updated fitness value is greater than or equal to the optimal initial fitness value, then the updated fitness value is used as the new fitness value of the corresponding sparrow; if the updated fitness value is less than the optimal initial fitness value, then the optimal initial fitness value is used as the new fitness value of the corresponding sparrow.

[0219] S432. Update the global optimal position based on the new fitness value: until the maximum number of iterations is reached, output the global optimal position. The coordinates of the global optimal position are the optimal solution for the weights and biases of the DNN model.

[0220] The sparrow search algorithm simulates the foraging and alertness behavior of sparrows, enabling it to perform a global search in the parameter space. This avoids the local optima problem caused by improper selection of initial parameters or learning rate in traditional gradient descent algorithms.

[0221] To further illustrate the inventive concept of this invention and the accuracy of the evaluation results, the oil state evaluation model provided by this invention was trained, and the training results are as follows: Figure 11 As shown, Figure 11 Figure (a) shows the confusion matrix of the sample prediction results. Figure 11 Figure (b) shows the early warning accuracy. Figure 11 Figure (c) in the figure shows the predicted loss;

[0222] Specifically, the oil state assessment model was trained using 750 sets of samples with a learning rate of 0.01 and 1000 training iterations. The sparrow search algorithm was used to optimize the weights and biases of the DNN model with a Dropout layer. The highest-precision DNN model was tested using 250 sets of samples, and a comparative analysis was conducted between the DNN model without a Dropout layer and the DNN model with a Dropout layer without optimization. The results show that adding a Dropout layer to the first hidden layer of the DNN model reduces the fitting degree of the DNN by preventing the interaction of feature detectors and reducing the redundancy of input information, thus significantly improving the model's accuracy and reducing the loss. Furthermore, the sparrow search algorithm for optimizing the model's weights and biases can quickly find the optimal solution for the hyperparameters, increasing the convergence speed during model training and improving the model's prediction accuracy.

[0223] For the optimized oil condition assessment model, another 500 sets of measured data samples were used for testing and verification; the test results are as follows. Figure 12 As shown, Figure 12 Figure (a) shows the confusion matrix of the sample prediction results. Figure 12 Figure (b) shows the early warning accuracy. Figure 12 Figure (c) shows the prediction loss. The results show that the optimized oil condition assessment model has a prediction loss of 0.0026 for the measured data samples, achieving a prediction accuracy of 95%. This indicates that the optimized oil condition assessment model has a high ability to identify the deterioration state / degree of the fan lubricating oil. Only 24 out of 500 test samples were misclassified, with all 100 samples in state v1 being correctly classified. This demonstrates that the optimized oil condition assessment model can effectively identify abnormal deterioration of the lubricating oil. The misclassification probability for other states is less than 10%, further proving that the optimized oil condition assessment model has a high accuracy rate in identifying the degree of lubricating oil deterioration.

[0224] Example 2

[0225] See Figure 7 The present invention also provides a wind turbine gearbox lubrication condition assessment device, comprising an impedance spectrum monitoring unit, a data preprocessing unit, a fuzzy comprehensive assessment unit, an assessment model training unit, and an online condition assessment unit connected in sequence. Each unit works together to complete real-time monitoring and intelligent assessment of the gear oil lubrication condition, as detailed below:

[0226] The impedance spectrum monitoring unit is used to apply multiple preset frequency small-amplitude sinusoidal current disturbance signals to the oil in the gearbox of the wind turbine under test, and simultaneously collect the electrochemical response signal and real-time temperature data of the oil through the electrochemical impedance spectroscopy detection module. Specifically, in this embodiment, it is used to generate five preset frequency small-amplitude sinusoidal current disturbance signals to monitor and collect the electrochemical response data and real-time temperature data of the oil. The impedance spectrum monitoring unit includes an impedance spectrum sensor and additional circuitry. The five preset frequency small-amplitude sinusoidal current disturbance signals are generated by the sensor's built-in signal generator, and the real-time temperature data is measured by the sensor's built-in temperature probe. The impedance spectrum sensor is a ZXEIS impedance spectrum oil sensor, with five preset frequencies of 0.15Hz, 1Hz, 10Hz, 100Hz, and 1000Hz. The sampling time interval between the impedance spectrum sensor and the oil electrochemical response data and real-time temperature data is 5 minutes.

[0227] The data preprocessing unit is used to extract features from the acquired electrochemical response signal to obtain the impedance amplitude |Z| and phase difference θ; and based on a preset polynomial surface regression model, combined with oil temperature data, to perform temperature drift correction on the impedance amplitude |Z| and phase difference θ to generate the temperature-compensated equivalent impedance amplitude |Z'| and equivalent phase difference θ'; and to map |Z'| and θ' to the [0,1] interval using a normalization method to obtain the normalized equivalent impedance spectrum features.

[0228] The fuzzy comprehensive evaluation unit is used to construct a fuzzy evaluation index set based on the preset deterioration level standard of gear oil; it uses the maximum membership method to classify the deterioration level of the impedance spectrum characteristics after temperature compensation, and outputs the fuzzy comprehensive evaluation results (level labels and membership vectors); specifically, the operation flow of the fuzzy comprehensive evaluation unit is as follows:

[0229] Constructing a fuzzy comprehensive evaluation factor set using normalized equivalent impedance spectrum eigenvalues ,in These are the impedance amplitudes at five preset frequencies. These are the phase differences at five preset frequencies;

[0230] Fuzzy Comprehensive Evaluation Set ,in These correspond to five gear oil degradation levels: new oil, mild degradation, moderate degradation, severe degradation, and scrap.

[0231] Constructing a set of fuzzy comprehensive evaluation factors The i-th factor in the fuzzy comprehensive evaluation set Membership degree of the j-th state Composition of factor evaluation matrix ;

[0232] Based on the fuzzy comprehensive evaluation factor set Based on the importance of each factor and expert knowledge, a weight vector for each factor is defined. In this embodiment, For elements all of 1 Row vectors;

[0233] Based on factor evaluation matrix and weight vector Calculate the fuzzy vector ,in, Let be the probability of belonging to the i-th state. The evaluation result is the state corresponding to the maximum probability. In order to comprehensively evaluate the synthesis operator, in this embodiment, Set it to general matrix multiplication;

[0234] The evaluation model training unit is used to construct a deep neural network (DNN) model with a Dropout layer as a regularization component as the initial oil condition evaluation model; the initial model is trained using fuzzy comprehensive evaluation results as supervision labels, and the hyperparameters of the model are optimized using a sparrow search algorithm to obtain the optimized oil condition evaluation model; specifically, the operation flow of the evaluation model training unit is as follows:

[0235] A DNN model is constructed, and a Dropout layer is added before the first hidden layer. In this embodiment, the DNN model consists of an input layer, an output layer, and three hidden layers. The number of neurons in the input layer, the first hidden layer, the second hidden layer, the third hidden layer, and the output layer are 10, 10, 12, 8, and 5, respectively.

[0236] The result of the fuzzy comprehensive evaluation is input into the DNN model for model training;

[0237] During the training of the DNN model, the Sparrow Search algorithm is used to find the optimal solution for the weights and biases of the DNN network;

[0238] The optimal solution of weights and biases is input into the DNN model to obtain the optimized oil state assessment model.

[0239] The online condition assessment unit is used to process the real-time collected electrochemical impedance data through the data preprocessing unit to complete temperature compensation and normalization, and the fuzzy comprehensive assessment unit to complete the degradation level classification. Then, it is input into the optimized oil condition assessment model for feature learning and condition prediction, and outputs the real-time degradation condition assessment result of the gear oil.

[0240] Specifically, after the electrochemical impedance data samples are processed by the data preprocessing unit and the fuzzy comprehensive evaluation unit, the output results are input into the optimized oil condition assessment model, and the output category corresponding to the test sample is used as the condition assessment result. The condition assessment result is divided into five states: "Excellent", "Good", "Medium", "Poor", and "Scrap". Specifically, "Excellent" corresponds to the "new oil" degradation level, indicating that the oil has not shown a significant performance degradation trend; "Good" corresponds to the "slight degradation period" degradation level, indicating that the oil is in a state of slight performance degradation; "Medium" corresponds to the "moderate degradation period" degradation level, indicating that the oil has shown a relatively obvious performance degradation trend; "Poor" corresponds to the "severe degradation period" degradation level, indicating that the oil is in a relatively serious performance degradation state; and "Scrap" corresponds to the "scrap period" degradation level, indicating that the oil's performance is in an extremely serious degradation state and needs to be replaced in time.

[0241] As a preferred embodiment, see Figure 8 The data preprocessing unit includes:

[0242] The feature extraction subunit is used to extract the impedance amplitude and phase difference at a preset frequency from the oil electrochemical response data;

[0243] The temperature compensation subunit is used to call the polynomial surface regression temperature compensation model to calculate the equivalent impedance spectrum characteristic values ​​at the reference temperature, including the equivalent impedance amplitude and the equivalent phase difference, and to periodically update the polynomial surface regression temperature compensation model; in this embodiment, the reference temperature is set to 60℃.

[0244] Specifically, the update process for the polynomial surface regression temperature compensation model is as follows:

[0245] The model update routine is initiated once every T interval. In this embodiment, T=48h.

[0246] The model update routine automatically records data from N consecutive sampling points. In this embodiment, N=100.

[0247] Extract the temperature data from N sampling points and the difference between the temperature and the reference temperature. For data within ℃, calculate the average value of the impedance spectrum characteristic values ​​and use it as the impedance spectrum reference value;

[0248] Calculate the impedance spectrum characteristic difference between the impedance spectrum characteristic values ​​of N sampling points and the impedance spectrum reference value, and the temperature difference between the temperature data and the reference temperature;

[0249] Substitute the impedance spectrum characteristic difference, impedance spectrum characteristic value, and temperature difference of N data points into a cubic polynomial surface regression model. ,in, Due to poor impedance spectrum characteristics, These are characteristic values ​​of the impedance spectrum. Calculate the parameters of the regression model for the temperature difference. ;

[0250] parameters Input the model and update the polynomial surface regression temperature compensation model;

[0251] Normalization subunit: Normalizes the equivalent impedance spectrum eigenvalues ​​to Interval.

[0252] This device, through multi-unit collaboration, multi-algorithm fusion, and intelligent design, solves the technical bottlenecks of traditional oil condition assessment methods in terms of temperature interference, single features, and insufficient model generalization. It achieves accurate, real-time, and intelligent assessment of gear oil lubrication status, providing key technical support for the efficient and reliable operation of wind turbine units, and has significant engineering application value and innovation.

[0253] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct or indirect applications in other related technical fields, are within the patent protection scope of the present invention.

Claims

1. A method for evaluating the lubrication condition of a fan gearbox, characterized in that, include: S1. Impedance spectroscopy sensor collects electrochemical impedance data and temperature data of wind turbine gear oil; S2. Perform data preprocessing: including extracting impedance spectrum eigenvalues, and performing temperature compensation and normalization; S3. Classify the impedance spectrum characteristics according to the deterioration level of the gear oil and perform fuzzy comprehensive evaluation; S4. Initially construct an oil condition assessment model based on neural networks, input the obtained fuzzy comprehensive assessment results into the oil condition assessment model for training, and perform hyperparameter tuning on the oil condition assessment model to obtain an optimized oil condition assessment model. S5. Collect the electrochemical impedance data and temperature data of the gear oil of the wind turbine in real time online. Repeat steps S2 and S3. Input the obtained real-time fuzzy comprehensive evaluation results into the optimized oil condition evaluation model to obtain the gear oil deterioration condition evaluation results.

2. The method for evaluating the lubrication condition of a fan gearbox according to claim 1, characterized in that, The electrochemical impedance data collected for wind turbine gear oil includes: Generate multiple small-amplitude sinusoidal current disturbance signals with preset frequencies; Monitor the electrochemical response of the oil under a preset frequency disturbance and obtain response data.

3. A method for evaluating the lubrication condition of a fan gearbox according to claim 1 or 2, characterized in that, The impedance spectrum characteristic values ​​include the impedance amplitude and phase difference corresponding to the electrochemical response data at multiple preset frequencies; The temperature compensation is achieved by converting the impedance amplitude and phase difference at different temperatures into the equivalent impedance amplitude and equivalent phase difference at the reference temperature through polynomial surface regression. The normalization involves normalizing the equivalent impedance magnitude and the equivalent phase difference at the reference temperature to... Interval.

4. The method for evaluating the lubrication condition of a fan gearbox according to claim 3, characterized in that, The steps for obtaining the equivalent impedance magnitude are as follows: The reference temperature is set according to the working environment and lubrication mechanism of the wind turbine gearbox. Select the impedance spectrum characteristic values ​​and the temperature data from N consecutive sampling points; Extract data from the N sampling points whose temperature difference with the reference temperature is within a preset error range, calculate the average value of the corresponding impedance spectrum characteristic value, and use it as the impedance spectrum reference value. Calculate the impedance spectrum characteristic difference between the impedance spectrum characteristic value of the N sampling points and the impedance spectrum reference value, and the temperature difference between the temperature data and the reference temperature; Substitute the impedance spectrum characteristic difference, impedance spectrum characteristic value, and temperature difference of the N data points into the polynomial surface regression model to calculate the parameters of the polynomial surface regression model. A polynomial surface regression temperature compensation model is constructed using various parameters. Calculate the impedance spectrum characteristic value and the temperature difference at each real-time sampling point, input them into the polynomial surface regression temperature compensation model, and obtain the impedance spectrum characteristic difference; The equivalent impedance spectrum characteristic value of each sampling point is obtained by adding the impedance spectrum characteristic difference to the actual impedance spectrum characteristic value of the sampling point.

5. A method for evaluating the lubrication condition of a fan gearbox according to claim 1 or 2, characterized in that, The steps for classifying impedance spectrum characteristics according to the deterioration level of gear oil and performing fuzzy comprehensive evaluation are as follows: Based on the normalized equivalent impedance magnitude and equivalent phase difference, a fuzzy comprehensive evaluation factor set is established. : ,in These are the impedance amplitudes at n preset frequencies. These are the phase differences at n preset frequencies; Based on the deterioration level of gear oil, a fuzzy comprehensive evaluation set is established. ,in These correspond to five gear oil deterioration levels: new oil, mild deterioration, moderate deterioration, severe deterioration, and scrapping. Construct the set of fuzzy comprehensive evaluation factors The i-th factor in the fuzzy comprehensive evaluation set Membership degree of the j-th state Composition of factor evaluation matrix ; Based on the fuzzy comprehensive evaluation factor set The importance of each factor and expert knowledge are considered, and the weight vector of each factor is assumed to be... ; Based on the factor evaluation matrix and the weight vector Calculate the fuzzy vector ,in, Let be the probability of belonging to the i-th state level. The state level corresponding to the maximum probability is the evaluation result. To comprehensively evaluate the synthetic operators.

6. A method for evaluating the lubrication condition of a fan gearbox according to claim 1 or 2, characterized in that, In step S4, the construction of the oil condition assessment model based on neural networks specifically involves constructing a DNN model including a Dropout layer and using a sparrow search algorithm to perform hyperparameter tuning on the oil condition assessment model.

7. The method for evaluating the lubrication condition of a fan gearbox according to claim 6, characterized in that, The steps to obtain the optimized oil condition assessment model are as follows: Add a Dropout layer before the first hidden layer of the DNN model to construct a DNN model that includes a Dropout layer. The result of the fuzzy comprehensive evaluation is input into the DNN model for model training; During the training of the DNN model, the Sparrow Search algorithm is used to find the optimal solution for the weights and biases of the DNN network; The optimal solution of weights and biases is input into the DNN model to obtain the optimized oil state assessment model.

8. The method for evaluating the lubrication condition of a fan gearbox according to claim 7, characterized in that, The steps for constructing a DNN model that includes a Dropout layer are as follows: A multilayer perceptron-based DNN model is constructed by stacking multiple hidden layers between the input and output layers. Each layer performs a nonlinear transformation through a nonlinear activation function, progressively extracting features and abstracting representations from the input data. The calculation formula for a single neuron in the DNN model is as follows: ， in, It is the output of the previous neuron. For the weight vector, For bias, Represents a node. It is an activation function; A Dropout layer is added before the first hidden layer of the DNN model; the formula for calculating a single neuron in a DNN with a Dropout layer is: , in, Let be the Bernoulli probability vector. It is the probability in the Bernoulli function. This is the result of the Dropout layer calculation.

9. The method for evaluating the lubrication condition of a fan gearbox according to claim 7, characterized in that, The steps for finding the optimal solution for the DNN network weights and biases using the sparrow search algorithm during the DNN model training process are as follows: The weights and bias hyperparameter vectors corresponding to each sparrow are substituted into the DNN model, and the fuzzy comprehensive evaluation results are cross-validated to obtain the cross-validation accuracy. Using the accuracy of the cross-validation as the initial fitness value, the fitness values ​​of all sparrows in the population are obtained. Sort all sparrows according to their fitness values, and divide them into discoverer sparrows and joiner sparrows according to a preset ratio; randomly select a number of sparrows as guard sparrows; The position of the discoverer sparrow in the sparrow search algorithm is updated using the discoverer update formula; The position of the newest sparrow in the sparrow search algorithm is updated using the newest sparrow update formula. The position of the vigilant sparrow in the sparrow search algorithm is updated using the vigilant update formula; Calculate the update fitness value of the sparrow after all update positions; The updated fitness value of each sparrow is compared with the best initial fitness value among the initial fitness values ​​to obtain the new fitness value of the corresponding sparrow; Update the global optimal position based on the new fitness value: until the maximum number of iterations is reached, output the global optimal position, and the coordinates of the global optimal position are the optimal solution for the weights and biases of the DNN model.

10. A device for assessing the lubrication condition of a fan gearbox, characterized in that, include: Impedance spectroscopy monitoring unit: used to generate multiple preset frequency small amplitude sinusoidal current disturbance signals to monitor and collect the electrochemical response data and real-time temperature data of the oil; Data preprocessing unit: used to extract impedance amplitude and phase difference from electrochemical response data, and to perform temperature compensation on impedance spectrum characteristics using polynomial surface regression algorithm based on real-time oil temperature, generating equivalent impedance amplitude and equivalent phase difference, and normalizing equivalent impedance spectrum characteristics. Fuzzy comprehensive evaluation unit: used to classify impedance spectrum characteristics according to the preset deterioration level of gear oil and perform fuzzy comprehensive evaluation; Evaluation model training unit: used to construct a DNN model containing a Dropout layer as the initial oil condition evaluation model, use the fuzzy comprehensive evaluation result as input to train the model, and use the sparrow search algorithm to fine-tune the hyperparameters of the model to obtain the optimized oil condition evaluation model. Online condition assessment unit: This unit is used to input the real-time electrochemical impedance data, after passing it through the data preprocessing unit and the fuzzy comprehensive assessment unit, into the optimized oil condition assessment model to obtain the gear oil deterioration condition assessment results.

Images