A method and system for controlling the inspection of fan component materials

Abstract

The application relates to the technical field of equipment control and discloses an equipment control method and system for fan part material maintenance, which quantitatively fuses fan operation parameters and interactive data into a unified feature vector, then dynamically corrects model parameters based on the fused feature vector, adjusts a convolution layer weight in real time according to the correlation between vibration features and working condition correction values, compensates short-term fluctuations of a rotating speed feature through a residual correction amount output by a pooling layer, and avoids misjudgment of a fixed parameter model under a sudden working condition; through a residual fitting algorithm containing a disturbance mechanism, a random disturbance term related to confidence and time attenuation is introduced, errors caused by microscopic fluctuations are reduced, and system deviation is compensated through interactive correction of residuals; finally, a closed-loop iteration mechanism continuously calculates a prediction deviation and updates model parameters, continuously reduces the difference between theoretical prediction and actual service life, so that the prediction error of the remaining service life is reduced, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the technical field of wind turbine parts maintenance management platform, and discloses an equipment control method and system for wind turbine parts material maintenance. Background Technology

[0002] As a core component of new energy power generation, wind turbines' key parts (such as bearings, gears, and blades) are subjected to complex conditions of high load, strong vibration, and drastic temperature and humidity fluctuations for extended periods. These components are prone to failure due to material fatigue, corrosion, or mechanical wear, severely impacting the unit's operational safety and economic efficiency. Currently, the maintenance and control of wind turbine components mainly relies on two types of technical solutions:

[0003] Passive maintenance based on sensor monitoring: This approach collects operating parameters by deploying sensors for temperature, vibration, and rotational speed, and uses fixed algorithms (such as threshold judgment and simple regression models) to predict the remaining lifespan of components. However, this method relies solely on the equipment's own operating data and does not incorporate interactive information such as manually detected data (e.g., measured bearing clearances) or temporary operating condition adjustments (e.g., load reduction commands). This results in poor adaptability of the model to sudden operating conditions; for example, when a wind turbine temporarily reduces its load for maintenance, sudden changes in sensor data can easily be misinterpreted as a fault.

[0004] Predictive maintenance based on static models: Some solutions introduce machine learning models (such as convolutional neural networks) for lifetime prediction. However, the model parameters (such as convolutional layer weights and pooling layer thresholds) are fixed once trained, and cannot dynamically respond to changes in material aging characteristics (such as fatigue limit decay of metals) or micro-fluctuations (such as the randomness of fiber wear in composite materials), leading to a decline in long-term prediction accuracy. At the same time, the residual fitting of existing models does not consider the timeliness of interactive data (such as the weight of manual feedback from 24 hours ago should be lower than that of real-time data), further amplifying the prediction bias.

[0005] The above approach leads to large errors in remaining life prediction, delayed maintenance plans or excessive maintenance, and increases operation and maintenance costs and downtime risks. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents, and such simplifications or omissions do not limit the scope of the invention.

[0007] To address the aforementioned technical problems, the main objective of this invention is to provide a device control method based on the material inspection of wind turbine components, comprising:

[0008] The system acquires interactive data such as temperature, vibration frequency, speed operating parameters of wind turbine components, and temporary operating condition adjustment commands, quantifies them to form a first feature vector, and fuses them to obtain a second feature vector.

[0009] Based on the second feature vector, the parameters of the wind turbine component monitoring model are adjusted to obtain the corrected model parameters. The parameter adjustment includes: correcting the vibration feature weights and temperature feature biases of the convolutional layer according to the second feature vector, compensating for the rotational speed feature output of the pooling layer, optimizing the prediction results of the remaining life of the component, and fitting the prediction residual.

[0010] The wind turbine component monitoring model is iteratively updated based on the corrected model parameters until the deviation between the predicted remaining life and the actual life is less than a preset threshold. Then, equipment control commands including maintenance timing and spare parts replacement are output to complete the material maintenance control of wind turbine components.

[0011] Optionally, quantizing the interaction data into a first feature vector and fusing it to obtain a second feature vector includes:

[0012] Temporary working condition adjustment instructions are encoded as working condition correction values, the difference between manually detected data and model predicted values ​​is encoded as measured deviation values, the reliability of environmental anomaly feedback is encoded as confidence values, and the data acquisition time is encoded as time decay values, resulting in multiple interactive features; the weights of each interactive feature are calculated, and a first feature vector is formed based on the interactive features and their corresponding weights.

[0013] Calculate the correlation between the temperature and vibration frequency features in the original operating feature vector and the operating condition correction value and measured deviation value in the first feature vector, and then fuse the two based on the correlation to obtain the second feature vector.

[0014] Optionally, calculating the weights of each interaction feature includes:

[0015] Obtain the measured deviation value, confidence value, and sensitivity coefficient; perform exponential operations on the values ​​of each interaction feature to obtain the calculated value; calculate the ratio of the calculated value of a single interaction feature to the sum of the calculated values ​​of all interaction features, and use this ratio as the weight of the interaction feature, where the weight of the measured deviation value is positively correlated with the confidence value.

[0016] Optionally, the parameter adjustment of the wind turbine component monitoring model based on the second feature vector includes:

[0017] Based on the vibration frequency characteristics, working condition correction values, and interaction feature weights in the second feature vector, the weight correction coefficients of the vibration characteristics of the convolutional layer are calculated; based on the temperature characteristics, measured deviation values, and interaction feature weights, the bias value correction coefficients of the temperature characteristics of the convolutional layer are calculated.

[0018] The original vibration weights of the convolutional layer are corrected based on the weight correction coefficient to obtain the corrected vibration weights, and the original temperature bias of the convolutional layer is corrected based on the bias correction coefficient to obtain the corrected temperature bias.

[0019] Based on the rotational speed feature, time decay value and interaction feature weight in the second feature vector, the residual correction amount of the pooling layer is calculated. The residual correction amount is then superimposed on the original rotational speed output obtained by the pooling layer based on the corrected vibration weight and the corrected temperature bias value to obtain the compensated rotational speed output.

[0020] Based on the second feature vector and the compensated rotational speed output, the optimized remaining life prediction value is obtained by adjusting the model parameters through a multi-dimensional objective function that includes the remaining life prediction bias loss and the interaction feature bias loss.

[0021] Based on the second feature vector, the optimized remaining lifetime prediction value and confidence value, and the time decay value, a residual algorithm with a perturbation mechanism is used to fit the prediction residual to obtain the fitted residual.

[0022] The corrected model parameters include the corrected vibration weight, the corrected temperature bias value, the compensated rotational speed output, the optimized remaining life prediction value, and the fitted residual.

[0023] Optionally, adjusting the model parameters through a multi-dimensional objective function includes:

[0024] The first loss term is calculated based on the deviation between the actual remaining lifespan of the wind turbine components and the predicted value obtained based on the compensated speed output.

[0025] The second loss term is calculated based on the deviation between the measured deviation value and the model-predicted deviation value in the interaction features.

[0026] The first and second loss terms are integrated to form a multi-dimensional objective function. Based on this objective function, the lifetime prediction coefficients of the model are adjusted to obtain the optimized remaining lifetime prediction value.

[0027] Optionally, the residual fitting algorithm using a perturbation mechanism includes:

[0028] The basic residual is calculated based on the difference between the optimized remaining lifetime prediction and the actual remaining lifetime.

[0029] The basic residual is corrected based on the working condition correction value in the second feature vector to obtain the interactive correction residual;

[0030] Based on the confidence value and time decay value in the second feature vector, a random perturbation value is generated to simulate the micro-fluctuations of material fatigue as a perturbation term, wherein the higher the confidence value and the smaller the time decay value, the smaller the amplitude of the perturbation value;

[0031] The fitted residuals are obtained by integrating the basic residuals, the interactively corrected residuals, and the perturbation term.

[0032] Optionally, the iterative update of the wind turbine component monitoring model based on the corrected model parameters includes:

[0033] Calculate the deviation value corresponding to the corrected model parameters, where the deviation value is the difference between the optimized remaining lifetime prediction value and the actual remaining lifetime;

[0034] When the deviation exceeds the preset threshold, the feature extraction layer parameters of the fan component monitoring model are updated based on the corrected vibration weight and the compensated speed output.

[0035] Repeat the update steps until the deviation value is less than or equal to the preset threshold. Determine that the model output meets the preset conditions and output the device control command.

[0036] Optionally, the deviation values ​​corresponding to the calculated corrected model parameters include:

[0037] Calculate the first deviation between the actual remaining life of the wind turbine components and the optimized predicted remaining life;

[0038] Calculate the second deviation between the measured deviation value and the model-predicted deviation value in the interaction features;

[0039] The maximum value of the first and second deviations is taken as the deviation value corresponding to the corrected model parameters.

[0040] Optionally, the obtained multiple interaction features include:

[0041] Based on the wind turbine's temporary load reduction and speed increase commands, the system generates operating condition correction values ​​that reflect load changes;

[0042] The difference between the manually measured bearing clearance value and the bearing clearance value predicted by the model is used to obtain the measured deviation value;

[0043] The confidence level value reflecting the reliability of the data collection personnel is determined based on their qualifications in manual testing.

[0044] The time decay value, which reflects the effectiveness, is calculated based on the interval between the collection time of the interactive data and the current time.

[0045] The present invention also provides an equipment control system based on the material inspection of wind turbine components, comprising:

[0046] The feature processing module acquires interactive data such as temperature, vibration frequency, speed operating parameters of wind turbine parts and temporary working condition adjustment commands, quantifies them to form a first feature vector and fuses them to obtain a second feature vector.

[0047] The model adjustment module corrects the convolutional layer weights, compensates the pooling layer output, optimizes the remaining lifetime prediction results and fitting residuals based on the second feature vector, and obtains the corrected model parameters.

[0048] The control module iteratively updates the monitoring model based on the corrected model parameters until the remaining life prediction deviation meets the threshold, and outputs control commands including maintenance timing to complete the maintenance control.

[0049] The beneficial effects of this invention are as follows: First, this invention quantifies and fuses wind turbine operating parameters (temperature, vibration frequency, speed, etc.) with interactive data (temporary operating condition commands, manual detection values, environmental feedback, etc.) into a unified feature vector. Then, based on the fused feature vector, the model parameters are dynamically corrected. The weights of the convolutional layer are adjusted in real time according to the correlation between vibration features and operating condition correction values. The pooling layer output compensates for short-term fluctuations in speed characteristics through residual correction, avoiding misjudgments by fixed-parameter models under sudden operating conditions. Through a residual fitting algorithm with a perturbation mechanism, random perturbation terms related to confidence and time decay are introduced to reduce errors caused by micro-fluctuations. At the same time, system bias is compensated through interactive correction of residuals. Finally, the closed-loop iterative mechanism continuously calculates the prediction bias and updates the model parameters, constantly narrowing the gap between theoretical prediction and actual lifespan, thereby reducing the remaining lifespan prediction error and improving prediction accuracy. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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 these drawings without creative effort. Wherein:

[0051] Figure 1 This invention provides a method for equipment control of material inspection of wind turbine components, and a flowchart of the wind turbine component monitoring model algorithm.

[0052] Figure 2 This is a system composition diagram of an equipment control method for the material inspection of wind turbine parts according to the present invention. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] This application provides an equipment control method based on the material inspection of wind turbine components, such as... Figure 1 As shown, it includes:

[0056] S101: Acquire interactive data such as temperature, vibration frequency, speed operating parameters of fan parts and temporary working condition adjustment commands, quantify to form the first feature vector and fuse to obtain the second feature vector;

[0057] S102: Adjust the parameters of the wind turbine component monitoring model based on the second feature vector to obtain the corrected model parameters. The parameter adjustment includes: correcting the vibration feature weights and temperature feature biases of the convolutional layer according to the second feature vector, compensating for the rotational speed feature output of the pooling layer, optimizing the prediction results of the remaining life of the component, and fitting the prediction residual.

[0058] S103: Iteratively update the wind turbine component monitoring model based on the corrected model parameters until the deviation between the predicted remaining life value and the actual life value output by the model is less than the preset threshold. Output equipment control instructions including maintenance timing and spare parts replacement to complete the material maintenance control of wind turbine components.

[0059] This application provides an equipment control method based on the material inspection of wind turbine components. This method achieves precise control of wind turbine component inspection by integrating multi-source information and a dynamic optimization model. The specific process is as follows:

[0060] The first step is data acquisition and feature vector generation.

[0061] In this embodiment, two types of core data need to be acquired: First, the operating parameters of the wind turbine components, including temperature, vibration frequency, and rotational speed. These parameters are collected by sensors deployed at key locations on the components. Temperature parameters can be monitored in real time using contact temperature sensors. Vibration frequency is obtained by collecting vibration signals from the components using piezoelectric vibration sensors and performing spectrum analysis. Rotational speed is calculated by recording the rotation cycle of the components using photoelectric encoders. Second, interactive data. For example, interactive data includes temporary operating condition adjustment commands (such as load reduction or speed increase commands due to changes in grid load), manual inspection data (such as component dimensional deviations measured by maintenance personnel using professional measuring tools), and environmental anomaly feedback (such as alarm information for excessive humidity or abnormal concentration of corrosive gases). This type of data can be collected through the command interface of the wind turbine control system, manual input terminals, or environmental monitoring modules.

[0062] It should be noted that the above data needs to be quantized to form feature vectors. Specifically, the operating parameters are transformed into original operating feature vectors through feature extraction (e.g., signal filtering, dimensionality reduction), which represent the objective operating state of the parts. Interactive data is structured into a first feature vector, where the encoding process converts unstructured information into computable values, for example, converting the adjustment range of instructions into corresponding coefficients. Subsequently, a feature fusion algorithm merges the original operating feature vector and the first feature vector into a second feature vector. The purpose of fusion is to ensure that the feature vector simultaneously includes both the equipment's own state and external intervention information, improving the model's adaptability to complex operating conditions. This application is not limited to this; the fusion algorithm can employ weighted fusion, feature concatenation, or attention mechanisms, depending on the data correlation.

[0063] It should be noted that the operating parameters of wind turbine components include temperature, vibration frequency, and rotational speed. These parameters are collected by sensors deployed on or near the surface of the components—temperature parameters can be collected in real time using thermocouple sensors, vibration frequency can be obtained using accelerometers, and rotational speed can be measured using magnetoelectric sensors. Interactive data includes temporary operating condition adjustment commands (such as load reduction / increase commands issued by the dispatching system), manually detected data (such as component wear recorded by maintenance personnel), and environmental anomaly feedback (such as humidity exceeding alarms). This type of data can be collected through the control system interface or manual input terminals.

[0064] In this embodiment, the above data is quantified: the running parameters are transformed into the original running feature vector V0 through a feature extraction model (such as principal component analysis); the interaction data is encoded into the first feature vector V1, and the encoding process needs to be implemented through the mapping function f(x), that is, V1 = f(interaction data). Then, V0 and V1 are fused through the fusion function g(x,y) to obtain the second feature vector V2, and the fusion formula is:

[0065] V2=g(V0,V1)=V0*α+V1*(1-α)

[0066] Where α is the fusion coefficient, which adjusts the weights of the two types of features. The value of α is determined by the correlation between the running parameters and the interactive data.

[0067] The second step is to dynamically adjust the model parameters.

[0068] In this embodiment, the parameters of the wind turbine component monitoring model are adjusted based on the second feature vector to make the model output more closely match the actual state of the components. The adjustments include:

[0069] Correcting the vibration feature weights and temperature feature biases of the convolutional layer: The vibration feature weights adjust the importance of the vibration signal in feature extraction, and the temperature feature biases compensate for systematic errors in temperature measurement. Both are dynamically updated based on the relevant features in the second feature vector.

[0070] Output rotational speed characteristics of the compensated pooling layer: Since rotational speed is easily affected by instantaneous disturbances, a residual correction is added to offset short-term fluctuations, making the output rotational speed characteristics more stable.

[0071] Optimize the remaining life prediction results of parts: combine multi-dimensional information to correct the initial prediction value and reduce the deviation caused by a single factor;

[0072] Fitting and predicting residuals: By introducing a perturbation term to cover the fluctuations in the microscopic properties of the material (such as the randomness of fatigue crack propagation), the completeness of the residual estimation is improved.

[0073] It should be noted that the core logic of parameter adjustment is to establish a dynamic relationship between feature vectors and model parameters, so that the model can respond in real time to changes in the state of the parts and the external environment.

[0074] In this embodiment of the application, parameter adjustment is performed on the wind turbine component monitoring model (such as a convolutional neural network model), specifically including:

[0075] Correcting convolutional layer parameters: Adjusting the vibrational feature weights ω of the convolutional layer based on the vibrational frequency and temperature characteristics in V2. v and temperature characteristic bias value b t ;

[0076] Compensating for pooling layer output: Adding residual correction to the rotation speed characteristics of the pooling layer output to offset the impact of short-term fluctuations;

[0077] Optimize remaining lifetime prediction results: minimize prediction bias using an objective function;

[0078] Fitting and predicting residuals: A perturbation mechanism is introduced to cover micro-fluctuations, and the fitted residuals are obtained.

[0079] It should be noted that the above adjustment process must ensure that each parameter is dynamically correlated with V2 so that the model output can reflect the changes in the state of the part in real time.

[0080] The third step is to iteratively update the model and output control commands.

[0081] In this embodiment, the wind turbine component monitoring model is iteratively updated based on the corrected model parameters. The model performance is evaluated by calculating the deviation between the predicted remaining life and the actual life. When the deviation is less than a preset threshold, the model prediction accuracy is deemed to meet the requirements, and equipment control instructions containing maintenance timing (such as the optimal maintenance time determined based on the predicted remaining life) and spare parts replacement (such as a recommended parts replacement list based on the degree of wear) are output. If the deviation does not meet the requirements, the parameter adjustment and model update steps are repeated until the preset conditions are met.

[0082] For example, determining the timing of maintenance requires comprehensive consideration of the remaining lifespan of the parts and the wind turbine's operating schedule to avoid unplanned downtime; spare parts replacement instructions must match the part model and inventory information to ensure maintenance feasibility. This application is not limited to this, and the specific content of the control instructions can be flexibly adjusted according to the wind turbine type (such as wind turbines, industrial fans) and the characteristics of the parts (such as bearings, gearboxes).

[0083] In this embodiment, the monitoring model is continuously updated iteratively based on the corrected model parameters, using a deviation function. Calculate the remaining life prediction value The deviation from the actual lifespan y, i.e. When e is less than the preset threshold, the model converges and outputs control commands including maintenance timing and spare parts replacement; otherwise, the parameter adjustment and model update steps are repeated.

[0084] This application is not limited to this. The specific content of the control command can be dynamically adjusted according to the fan model and part type. For example, the logic for determining the maintenance timing of bearings and gearboxes can be set with differentiated thresholds.

[0085] Optionally, quantizing the interaction data into a first feature vector and fusing it to obtain a second feature vector includes:

[0086] Temporary working condition adjustment instructions are encoded as working condition correction values, the difference between manually detected data and model predicted values ​​is encoded as measured deviation values, the reliability of environmental anomaly feedback is encoded as confidence values, and the data acquisition time is encoded as time decay values, resulting in multiple interactive features; the weights of each interactive feature are calculated, and a first feature vector is formed based on the interactive features and their corresponding weights.

[0087] Calculate the correlation between the temperature and vibration frequency features in the original operating feature vector and the operating condition correction value and measured deviation value in the first feature vector, and then fuse the two based on the correlation to obtain the second feature vector.

[0088] In this embodiment of the application, the process of quantizing interactive data into a first feature vector and fusing it to obtain a second feature vector is key to realizing the collaborative utilization of multi-source information, and specifically includes:

[0089] First, the generation of interactive features encodes temporary operating condition adjustment instructions into operating condition correction values: Temporary operating condition adjustment instructions reflect changes in the wind turbine's operating load (such as load reduction or speed increase). When encoding, the load adjustment ratio in the instruction needs to be converted into a correction coefficient. For example, a load reduction instruction corresponds to a coefficient less than 1, and a speed increase instruction corresponds to a coefficient greater than 1, thus correcting the model's response sensitivity to load changes.

[0090] The difference between manually measured data and model predictions is encoded as the measured deviation value: the manually measured data refers to the part state parameters (such as gaps and amplitudes) measured by maintenance personnel using professional tools (such as micrometers and vibration meters). The difference between these and the model predictions is the measured deviation value. This value characterizes the degree of deviation between the model prediction and the actual state, providing a basis for model correction.

[0091] The reliability of environmental anomaly feedback is coded as a confidence value: The reliability of environmental anomaly feedback depends on the accuracy of the detection equipment and the data acquisition method. For example, the feedback reliability of laboratory-grade detection equipment is higher than that of manual visual inspection. When coding, it is characterized by a coefficient (such as a value between 0 and 1). A higher confidence value corresponds to a more reliable environmental impact assessment.

[0092] Encode the data acquisition time as a time decay value: The timeliness of interactive data decreases over time (e.g., detection data from 24 hours ago has low reference value for the current state). This timeliness is quantified by the time decay value. The longer the time interval between the acquisition time and the current time, the smaller the decay value, thus reducing the impact of outdated data on the model.

[0093] It should be noted that the above-mentioned interactive features together constitute a feature set. The first feature vector is formed by calculating the weight of each feature (the weight reflects the degree of influence of the feature on the model). The weight calculation needs to take into account the differences in the importance of the features (for example, the measured deviation value usually has a more direct impact on model correction than environmental feedback).

[0094] Secondly, the eigenvector fusion involves calculating the correlation between the original operating eigenvector (containing objective operating features such as temperature and vibration frequency) and the first eigenvector (containing interactive features such as operating condition correction values ​​and measured deviation values). The correlation measure the relationship between the two types of features (e.g., a high correlation between vibration frequency and operating condition correction values ​​indicates a significant impact of load changes on vibration), and can be calculated using statistical analysis methods (such as correlation analysis). Based on the correlation, the two types of eigenvectors are fused to obtain a second eigenvector. During fusion, features with high correlation are given higher weights, ensuring that the second eigenvector reflects both the component's own operating state and the influence of external interventions and environmental factors.

[0095] The method for calculating correlation can be selected according to the data type. For example, Pearson correlation coefficient is used for continuous features, and chi-square test is used for discrete features.

[0096] Specifically, in this embodiment of the application, the process of quantizing the interactive data into a first feature vector and fusing it to obtain a second feature vector is as follows:

[0097] First, generate interactive features.

[0098] The temporary operating condition adjustment command is encoded as the operating condition correction value c: if the command is "reducing load", c takes a value less than 1; if it is "increasing load", c takes a value greater than 1. The encoding formula is c = 1 + k * Δ load rate, where k is the proportional coefficient and Δ load rate is the percentage of load change.

[0099] The difference between manually detected data and model predictions is encoded as the measured deviation value d: d = manually detected value - model prediction value. This value reflects the deviation between the model and the actual state.

[0100] The reliability of environmental anomaly feedback is encoded as a confidence value r: the value is assigned based on the credibility of the feedback source. For example, the feedback from laboratory-level testing equipment has a confidence value r close to 1, while the feedback from manual visual inspection has a confidence value r close to 0.

[0101] The data acquisition time is encoded as a time decay value t: t = exp(-λ*Δt), where Δt is the interval between the acquisition time and the current time, λ is the decay coefficient, and t decreases as Δt increases, reflecting the timeliness of the data.

[0102] It should be noted that the above-mentioned interaction features together constitute the feature set {c,d,r,t}, and the weights w of each feature are calculated. i (i represents the feature index), forming the first feature vector V1 = [w c *c,w d *d,w r *r,w t *t], where w c +w d +w r +w t =1.

[0103] Secondly, feature vectors are fused.

[0104] Calculate the correlation degree k between the original operating feature vector V0 (containing temperature feature t and vibration frequency feature f) and the first feature vector V1 (containing c and d). The correlation degree is calculated using the Pearson coefficient.

[0105]

[0106] The second feature vector V2 = V0 + k * V1 is obtained by fusing based on the correlation degree k, so that V2 simultaneously contains the device operating status and external interaction information.

[0107] Optionally, calculating the weights of each interaction feature includes:

[0108] Obtain the measured deviation value, confidence value, and sensitivity coefficient; perform exponential operations on the values ​​of each interaction feature to obtain the calculated value; calculate the ratio of the calculated value of a single interaction feature to the sum of the calculated values ​​of all interaction features, and use this ratio as the weight of the interaction feature, where the weight of the measured deviation value is positively correlated with the confidence value.

[0109] In this embodiment, the calculation of interaction feature weights distinguishes the importance of different interaction features in the model, ensuring that key features (such as measured deviation values) play a greater role. The specific process is as follows:

[0110] Obtain the numerical values ​​and sensitivity coefficients of the interaction features: The numerical values ​​of the interaction features are the working condition correction values, measured deviation values, etc. obtained by the above encoding; the sensitivity coefficient is a preset parameter that adjusts the degree of influence of the feature on the weight. For example, a higher sensitivity coefficient can be set for the measured deviation value, which is more critical to the model correction, so that the feature is more easily highlighted in the weight calculation.

[0111] Perform exponential operations on the values ​​of each interactive feature: Exponential operations can amplify the differences between features, making the values ​​of important features more significant after the operation, and avoiding insufficient weight discrimination caused by similar value ranges.

[0112] The ratio of the calculated value of a single feature to the sum of the calculated values ​​of all features is used as the weight of that feature. This process uses normalization to ensure that the sum of the weights is 1, making the weights of each feature comparable. It should be noted that the weight of the measured deviation value is positively correlated with the confidence value. This is because when the confidence level of manually tested data is high (e.g., collected by experienced technicians), the measured deviation value more accurately reflects model defects. Therefore, it is necessary to enhance its role in model correction through weight adjustment. Specifically, this can be achieved by incorporating the influence of the confidence value into the sensitivity coefficient (e.g., the sensitivity coefficient increases with the increase of the confidence value).

[0113] In this embodiment of the application, calculating the weights of each interaction feature includes the following steps:

[0114] Obtain numerical values ​​of characteristics such as measured deviation d, confidence level r, and sensitivity coefficient s. i (s i Adjusting the degree of influence of features on weights, such as s of d d It can be set to s greater than r r ).

[0115] The numerical values ​​of each interaction feature are exponentially calculated to obtain the calculated value e. i =exp(s i *x i ), where x i The original numerical values ​​are for the features.

[0116] The ratio of the calculated value of a single feature to the sum of the calculated values ​​of all features is used as the weight of that feature.

[0117]

[0118] Where n is the total number of interactive features.

[0119] It should be noted that the weight w of the measured deviation value d d It is positively correlated with the confidence value r because when r is high (data is reliable), d has a more significant effect on model correction. Therefore, by using s d The influence factor of r is introduced in the implementation, i.e., s d =s d0 *r(s d0 (Based sensitivity coefficient).

[0120] Optionally, the parameter adjustment of the wind turbine component monitoring model based on the second feature vector includes:

[0121] Based on the vibration frequency characteristics, working condition correction values, and interaction feature weights in the second feature vector, the weight correction coefficients of the vibration characteristics of the convolutional layer are calculated; based on the temperature characteristics, measured deviation values, and interaction feature weights, the bias value correction coefficients of the temperature characteristics of the convolutional layer are calculated.

[0122] The original vibration weights of the convolutional layer are corrected based on the weight correction coefficient to obtain the corrected vibration weights, and the original temperature bias of the convolutional layer is corrected based on the bias correction coefficient to obtain the corrected temperature bias.

[0123] Based on the rotational speed feature, time decay value and interaction feature weight in the second feature vector, the residual correction amount of the pooling layer is calculated. The residual correction amount is then superimposed on the original rotational speed output obtained by the pooling layer based on the corrected vibration weight and the corrected temperature bias value to obtain the compensated rotational speed output.

[0124] Based on the second feature vector and the compensated rotational speed output, the optimized remaining life prediction value is obtained by adjusting the model parameters through a multi-dimensional objective function that includes the remaining life prediction bias loss and the interaction feature bias loss.

[0125] Based on the second feature vector, the optimized remaining lifetime prediction value and confidence value, and the time decay value, a residual algorithm with a perturbation mechanism is used to fit the prediction residual to obtain the fitted residual.

[0126] The corrected model parameters include the corrected vibration weight, the corrected temperature bias value, the compensated rotational speed output, the optimized remaining life prediction value, and the fitted residual.

[0127] In this embodiment of the application, adjusting the parameters of the wind turbine component monitoring model based on the second feature vector is the core step in improving the model's prediction accuracy, specifically including:

[0128] Based on the vibration frequency feature, operating condition correction value, and interaction feature weights in the second feature vector, the weight correction coefficient of the vibration feature of the convolutional layer is calculated; based on the temperature feature, measured deviation value, and interaction feature weights, the bias correction coefficient of the temperature feature of the convolutional layer is calculated. The weight correction of the vibration frequency feature needs to be combined with the change of operating conditions (such as the vibration intensity decreasing when the load is reduced, the weight should be reduced accordingly), and the bias correction of the temperature feature needs to consider the measured deviation (such as when the difference between the actual temperature and the predicted temperature is large, the bias value needs to be increased to compensate for the deviation).

[0129] The original vibration weights of the convolutional layer are corrected based on the weight correction coefficient to obtain the corrected vibration weights. The original temperature bias of the convolutional layer is corrected based on the bias correction coefficient to obtain the corrected temperature bias. The original weights and biases are the initial training parameters of the model. The corrected values ​​enable the convolutional layer to extract lifespan-related features (such as abnormal vibrations and temperature rises) more accurately.

[0130] Based on the rotational speed feature, time decay value, and interaction feature weight in the second feature vector, the residual correction amount of the pooling layer is calculated. The residual correction amount is then superimposed on the original rotational speed output obtained by the pooling layer based on the corrected vibration weight and the corrected temperature bias value to obtain the compensated rotational speed output. The calculation of the residual correction amount needs to consider the timeliness of the rotational speed data (e.g., the correction amount for recent data is larger). The compensated rotational speed output can filter out instantaneous fluctuations and more realistically reflect the long-term operating status of the parts.

[0131] Based on the second feature vector and the compensated rotational speed output, the optimized remaining life prediction value is obtained by adjusting the model parameters through a multi-dimensional objective function that includes the remaining life prediction bias loss and the interaction feature bias loss. The multi-dimensional objective function focuses on both the accuracy of life prediction and the consistency of interaction data (such as the prediction accuracy of measured bias values) to avoid the model overfitting a single data type.

[0132] Based on the second feature vector, the optimized remaining lifetime prediction value, the confidence value, and the time decay value, a residual algorithm with a perturbation mechanism is used to fit the prediction residual to obtain the fitted residual: the perturbation mechanism simulates the micro-fluctuations of the material (such as the randomness of metal fatigue), the confidence value and the time decay value determine the perturbation intensity (such as the perturbation corresponding to high confidence data is smaller), and the fitted residual can improve the model's adaptability to uncertain factors.

[0133] It should be noted that the corrected model parameters integrate the results of feature extraction, output compensation, prediction optimization, and residual fitting, providing a comprehensive parameter basis for model iterative updates.

[0134] In this embodiment of the application, the process of adjusting the parameters of the wind turbine component monitoring model based on the second feature vector is as follows:

[0135] (1) Correcting convolutional layer parameters

[0136] Based on the vibration frequency characteristic f, the working condition correction value c, and the interaction characteristic weight w in V2, the vibration characteristic weight correction coefficient k is calculated. v :

[0137] k v =1+w d *d+w c *(c-1)

[0138] Based on the temperature characteristic t, the measured deviation value d, and the weight w, the temperature characteristic bias correction coefficient k is calculated. t :

[0139] k t =w t *t+w r *r

[0140] Corrected vibration weight ω v '=ω v *k v Corrected temperature bias value b t '=b t +k t .

[0141] (2) Compensation pooling layer output

[0142] Based on the rotational speed characteristic n, time decay value t, and weight w in V2, calculate the residual correction Δn:

[0143] Δn=w t *t*(n max -n)

[0144] Where n max This represents the rated rotational speed of the part. Δn is then superimposed onto the original output n of the pooling layer. pool After compensation, the speed output n' = n pool +Δn.

[0145] (3) Optimize remaining lifetime prediction results

[0146] The model parameters are adjusted using a multi-dimensional objective function L, and the formula for calculating L is:

[0147] L=λ1*L1+λ2*L2

[0148] Where L1 is the remaining lifetime prediction bias loss. L2 is the interactive feature bias loss ( λ1 and λ2 are the deviation values ​​predicted by the model, and λ1 and λ2 are the loss weights. The lifetime prediction coefficients θ are adjusted based on the L minimization principle to obtain the optimized remaining lifetime prediction value.

[0149] (4) Fitting and predicting residuals

[0150] Using a residual algorithm with a perturbation mechanism, the formula for calculating the fitted residual Res is as follows:

[0151] Res=Res0+Res1+ξ

[0152] Res0 is the basic residual. Res1 is the interactively corrected residual (Res1 = w) c *c*Res0), ξ is the perturbation term (ξ=σ*(1-r)*(1-t)*N(0,1), σ is the perturbation cardinality, N(0,1) is a standard normal distribution random number).

[0153] It should be noted that the corrected model parameters include

[0154] Optionally, adjusting the model parameters through a multi-dimensional objective function includes:

[0155] The first loss term is calculated based on the deviation between the actual remaining lifespan of the wind turbine components and the predicted value obtained based on the compensated speed output.

[0156] The second loss term is calculated based on the deviation between the measured deviation value and the model-predicted deviation value in the interaction features.

[0157] The first and second loss terms are integrated to form a multi-dimensional objective function. Based on this objective function, the lifetime prediction coefficients of the model are adjusted to obtain the optimized remaining lifetime prediction value.

[0158] In this embodiment, the model parameters are adjusted through a multi-dimensional objective function to simultaneously optimize remaining lifetime prediction and interaction feature matching. The specific process is as follows:

[0159] The first loss term is calculated by comparing the actual remaining lifespan of the wind turbine components (such as the actual operating time from the current state to failure) with the predicted value obtained based on the compensated speed output. The deviation between the two is calculated using a loss function (such as mean square error or absolute error), which is the first loss term. This loss term directly reflects the accuracy of the lifespan prediction.

[0160] The second loss term is calculated based on the measured deviation value (the difference between manual detection and the model's initial prediction) and the model's predicted deviation value (the model's prediction result for this difference) in the interaction features. The deviation between the two is calculated using the same type of loss function, which is the second loss term. This loss term measures the model's adaptability to the interaction data.

[0161] The first and second loss terms are integrated to form a multi-dimensional objective function: the two losses are integrated by setting weight coefficients (e.g., allocating proportions according to the importance of the two types of losses). The weight coefficients can be adjusted according to actual needs (e.g., increasing the weight of the first loss term when more emphasis is placed on lifetime prediction). Based on this objective function, the lifetime prediction coefficients of the model are adjusted using parameter optimization algorithms (e.g., gradient descent, genetic algorithm) (e.g., correcting the proportional coefficient of the initial prediction value) to minimize the objective function value, ultimately obtaining the optimized remaining lifetime prediction value.

[0162] It should be noted that the core function of the multi-dimensional objective function is to balance equipment state prediction with external information adaptation, and to avoid the model from reducing overall performance due to excessive focus on a single indicator.

[0163] In this embodiment of the application, the process of adjusting model parameters through a multi-dimensional objective function is as follows:

[0164] Calculate the first loss term L1: the predicted value obtained based on the actual remaining lifespan y of the wind turbine components and the compensated speed output n'. (Square loss function).

[0165] Calculate the second loss term L2: the deviation between the measured bias value d in the interaction features and the model prediction.

[0166] Integrate L1 and L2 to form the objective function L = α*L1 + (1-α)*L2 (where α is the weighting coefficient). Adjust the lifetime prediction coefficients θ (initial value of θ is 1) of the model using gradient descent until L is minimized, thus obtaining the optimized remaining lifetime prediction value.

[0167] Optionally, the residual fitting algorithm using a perturbation mechanism includes:

[0168] The basic residual is calculated based on the difference between the optimized remaining lifetime prediction and the actual remaining lifetime.

[0169] The basic residual is corrected based on the working condition correction value in the second feature vector to obtain the interactive correction residual;

[0170] Based on the confidence value and time decay value in the second feature vector, a random perturbation value is generated to simulate the micro-fluctuations of material fatigue as a perturbation term, wherein the higher the confidence value and the smaller the time decay value, the smaller the amplitude of the perturbation value;

[0171] The fitted residuals are obtained by integrating the basic residuals, the interactively corrected residuals, and the perturbation term.

[0172] In this embodiment, a residual algorithm with a perturbation mechanism is used to fit the residual in order to cover the random uncertainty in the prediction process. The specific process is as follows:

[0173] The basic residual is calculated based on the difference between the optimized remaining lifetime prediction and the actual remaining lifetime: the basic residual reflects the prediction deviation of the model at the macro level and is the basic data for residual fitting.

[0174] The basic residual is corrected based on the working condition correction value in the second feature vector to obtain the interactive correction residual: the working condition correction value reflects the impact of load changes on the life of the parts (such as heavy load will accelerate aging). By correcting the basic residual with this value, the residual can be made to better fit the deviation law under actual working conditions (such as the correction range of the residual is larger under heavy load).

[0175] Based on the confidence value and time decay value in the second feature vector, a random perturbation value is generated to simulate the micro-fluctuations of material fatigue as a perturbation term: the material fatigue process has micro-randomness (such as uneven wear at grain boundaries), and the perturbation term simulates this randomness; the higher the confidence value (the more reliable the interactive data) and the smaller the time decay value (the fresher the data), the smaller the amplitude of the perturbation value, so as to avoid interfering with the effective information; conversely, the perturbation amplitude is increased to enhance the model's ability to cover uncertainty.

[0176] By integrating the basic residuals, interactive correction residuals, and disturbance terms, the fitted residuals are obtained. The fitted residuals integrate macroscopic deviations, operating condition corrections, and microscopic disturbances, enabling the model to more comprehensively assess prediction errors and providing a more accurate basis for subsequent parameter adjustments.

[0177] This application is not limited to this. The method for generating the disturbance value may be random sampling (such as based on normal distribution or uniform distribution), and the specific distribution type shall be selected according to the material properties.

[0178] In this embodiment of the application, the process of fitting the residual using a residual algorithm with a perturbation mechanism is as follows:

[0179] Calculate the basic residual Res0: Res0 = Predicted remaining lifetime after optimization - Actual remaining lifetime.

[0180] Generate interactive correction residual Res1: Correct Res_0 based on the working condition correction value c in the second feature vector, Res1 = k * c * Res0, where k is the correction coefficient.

[0181] Generate the perturbation term ξ: Based on the confidence value r and the time decay value t, ξ = μ*(1-r)*(1-t)*∈, where μ is the base of the perturbation amplitude and ∈ is a random number in the interval [-1,1]; the higher r and the smaller t (the more reliable and fresh the data), the smaller the absolute value of ξ.

[0182] The fusion yields the fitted residual Res = Res0 + Res1 + ξ.

[0183] In a preferred embodiment, a random disturbance term Δξ and a micro-fluctuation coefficient are added to the original residual correction formula, as shown in the following formula:

[0184]

[0185] In this embodiment, Δξ is used to simulate random disturbances at the microscopic level (such as random fluctuations in internal stress of the part material and minute noise collected by the sensor), and the calculation formula is as follows:

[0186]

[0187] Parameter description:

[0188] α3: Confidence of interactive data (taken from interactive feature vector I, α3 approaches 1 for high confidence data and approaches 0 for low confidence data);

[0189] i4: Time decay factor (timeliness of interactive data; i4 approaches 1 for recent data and approaches 0 for outdated data);

[0190] σ: Disturbance amplitude coefficient (used to simulate the increase in microscopic instability caused by material aging);

[0191] (0,1): Standard normal distribution random number (mean 0, variance 1, simulating the probability characteristics of random disturbance).

[0192] It should be noted that the above parameters were obtained by fitting empirical data, and this application will not elaborate on them.

[0193] Through the above-mentioned random perturbation mechanism, since the actual loss of wind turbine parts has micro-randomness, the residual fitting based on random perturbation not only matches the macro data, but also covers the micro-fluctuation range, thus avoiding the model overfitting to "perfect data".

[0194] Optionally, the iterative update of the wind turbine component monitoring model based on the corrected model parameters includes:

[0195] Calculate the deviation value corresponding to the corrected model parameters, where the deviation value is the difference between the optimized remaining lifetime prediction value and the actual remaining lifetime;

[0196] When the deviation exceeds the preset threshold, the feature extraction layer parameters of the fan component monitoring model are updated based on the corrected vibration weight and the compensated speed output.

[0197] Repeat the update steps until the deviation value is less than or equal to the preset threshold. Determine that the model output meets the preset conditions and output the device control command.

[0198] In this embodiment, iteratively updating the wind turbine component monitoring model based on the corrected model parameters is key to continuously optimizing prediction accuracy. The specific process is as follows:

[0199] Calculate the deviation value corresponding to the corrected model parameters: This deviation value is the difference between the optimized remaining lifetime prediction value and the actual remaining lifetime, and directly evaluates the predictive performance of the model.

[0200] When the deviation exceeds the preset threshold, the feature extraction layer parameters of the wind turbine component monitoring model are updated based on the corrected vibration weight and the compensated rotational speed output. The feature extraction layer is responsible for extracting key features (such as abnormal frequency in vibration and trend changes in temperature) from the original data. Updating parameters (such as convolution kernel weight and pooling threshold) can make the feature extraction more in line with the current component status and reduce the deviation.

[0201] Repeat the update steps until the deviation value is less than or equal to the preset threshold. If the model output meets the preset conditions, output the device control command. The iterative update process is essentially the model's self-calibration. By continuously feeding back the deviation and adjusting the parameters, the model gradually adapts to the aging characteristics of the parts and changes in working conditions, and finally achieves stable prediction accuracy.

[0202] It should be noted that the preset threshold setting should be combined with maintenance requirements (such as the maximum allowable prediction error) to ensure that the output control commands can meet the balance between safe operation and economical maintenance of the equipment.

[0203] In this embodiment of the application, the process of iteratively updating the wind turbine component monitoring model based on the corrected model parameters is as follows:

[0204] Calculate the deviation values ​​corresponding to the corrected model parameters:

[0205] e: e = |Optimized remaining lifetime prediction - Actual remaining lifetime|.

[0206] When e exceeds the preset threshold, based on the corrected vibration weight ω v The compensated rotational speed output n' updates the model's feature extraction layer parameters; for example, the convolutional kernel weights are replaced with ω. v ', update the pooling layer output baseline value to n'.

[0207] Repeat the update step until e is less than or equal to a preset threshold, then determine if the model output meets the preset conditions. The output includes the maintenance timing (based on...). Equipment control commands for determining wear and replacement of spare parts (based on the absolute value of Res to determine the degree of wear).

[0208] Optionally, the deviation values ​​corresponding to the calculated corrected model parameters include:

[0209] Calculate the first deviation between the actual remaining life of the wind turbine components and the optimized predicted remaining life;

[0210] Calculate the second deviation between the measured deviation value and the model-predicted deviation value in the interaction features;

[0211] The maximum value of the first and second deviations is taken as the deviation value corresponding to the corrected model parameters.

[0212] In this embodiment, calculating the deviation value corresponding to the corrected model parameters is the basis for evaluating model performance and triggering updates. The specific process is as follows:

[0213] Calculate the first deviation: the difference between the actual remaining life of the wind turbine parts and the predicted remaining life after optimization. This deviation directly reflects the accuracy of the life prediction and is an indicator for evaluating the core performance of the model.

[0214] Calculate the second bias: the difference between the measured bias value in the interaction features and the bias value predicted by the model. This bias reflects the model's ability to predict interaction data and indirectly affects the model's utilization of external information.

[0215] The maximum value of the first and second deviations is used as the deviation value corresponding to the corrected model parameters: using the maximum value can ensure that the model does not exceed the acceptable range in both types of deviations, avoid ignoring the situation where the other type of deviation exceeds the standard because a single deviation is acceptable, and fully protect the model performance.

[0216] It should be noted that the maximum value was chosen as the overall bias because either of the two types of bias exceeding the limit may lead to unreliable model output (e.g., a large bias in life prediction may lead to maintenance delays, and a large bias in interactive data prediction may lead to model correction errors).

[0217] In this embodiment of the application, the process of calculating the deviation value corresponding to the corrected model parameters is as follows:

[0218] Calculate the first deviation:

[0219] e1: e1 = |Actual remaining life of wind turbine parts - Optimized predicted remaining life|.

[0220] Calculate the second bias e2: e2 = |Measured bias value in interactive features - Predicted bias value in model|.

[0221] The maximum value of e1 and e2 is used as the deviation value e = max(e1, e2) corresponding to the corrected model parameters, ensuring that the deviation assessment covers the consistency of device status and interaction data.

[0222] Optionally, the obtained multiple interaction features include:

[0223] Based on the wind turbine's temporary load reduction and speed increase commands, the system generates operating condition correction values ​​that reflect load changes;

[0224] The difference between the manually measured bearing clearance value and the bearing clearance value predicted by the model is used to obtain the measured deviation value;

[0225] The confidence level value reflecting the reliability of the data collection personnel is determined based on their qualifications in manual testing.

[0226] The time decay value, which reflects the effectiveness, is calculated based on the interval between the collection time of the interactive data and the current time.

[0227] In this embodiment of the application, the generation of interaction features is the key to transforming unstructured interaction data into computable metrics, specifically including:

[0228] Based on the temporary load reduction and speed increase commands of the wind turbine, the operating condition correction value is generated to reflect the load change: the temporary load reduction and speed increase commands are usually generated by the wind turbine control system according to external conditions (such as wind speed and grid demand). The load change rate (the ratio of load reduction / speed increase) is obtained through command parsing. The operating condition correction value is calculated by the load change rate and the conversion coefficient (converting the ratio into a coefficient that the model can recognize), which quantifies the impact of load change on the life of the components.

[0229] The difference between the manually measured bearing clearance value and the model-predicted bearing clearance value is used to obtain the measured deviation value: Bearing clearance is a key indicator reflecting the degree of bearing wear. Manual measurement is carried out using tools such as feeler gauges and laser diameter gauges. The bearing clearance value predicted by the model is obtained based on historical operating data and wear models. The difference between the two is the measured deviation value, which is used to correct the wear prediction algorithm of the model.

[0230] The confidence level value reflecting the reliability of the data collection personnel is determined based on their qualifications: the qualifications of the personnel include professional level, operating experience, etc. The higher the qualifications, the higher the reliability of the data collection. The confidence level is determined by the qualification level and the preset correspondence (such as a high value for senior technicians), which adjusts the weight of the manual data in the model.

[0231] The time decay value, which reflects the timeliness of data, is calculated based on the interval between the collection time of the interactive data and the current time: the interval time is calculated by the difference of timestamps, and the time decay value is calculated by the decay function (such as exponential decay or linear decay). The longer the interval, the smaller the decay value, which reduces the impact of old data on the model and ensures that the model is more dependent on recent information.

[0232] For example, the decay function can be selected according to the data type. For instance, environmental data decays faster than manually detected data because its timeliness is shorter. This application is not limited to this; the specific form of the decay function can be adjusted according to the actual data characteristics.

[0233] In this embodiment of the application, the process of obtaining multiple interactive features is as follows:

[0234] The generated operating condition correction value c is based on the temporary load reduction and speed increase instructions of the wind turbine. c = 1 + δ * load change rate, where δ is the conversion coefficient. The load change rate is negative when the load is reduced and positive when the load is increased.

[0235] Generate the measured deviation value d: Calculate the manually measured bearing clearance value G 人工 The bearing clearance value G predicted by the model 模型 The difference, d = G 人工 -G 模型 .

[0236] Generate confidence value r: based on the qualifications of the personnel collecting the manual inspection data. For example, r is 0.9 to 1.0 for data collected by senior technicians and 0.6 to 0.8 for data collected by junior maintenance personnel.

[0237] The generation time decay value t is based on the interval ΔT between the acquisition time of the interactive data and the current time, t=1 / (1+γ*ΔT), where γ is the decay factor, and the larger ΔT is, the smaller t is.

[0238] This invention also provides an equipment control system based on the material inspection of wind turbine components, such as... Figure 2 As shown, it includes:

[0239] Feature processing module 10 acquires interactive data such as temperature, vibration frequency, speed operating parameters and temporary working condition adjustment commands of fan parts, quantifies them to form a first feature vector and fuses them to obtain a second feature vector;

[0240] The model adjustment module 20 corrects the convolutional layer weights, compensates the pooling layer output, optimizes the remaining lifetime prediction results and fitting residuals based on the second feature vector, and obtains the corrected model parameters.

[0241] The control module 30 iteratively updates the monitoring model based on the corrected model parameters until the remaining life prediction deviation meets the threshold, and outputs control commands including maintenance timing to complete the maintenance control.

[0242] In this embodiment, the system achieves intelligent control of wind turbine component maintenance through module collaboration. The functions of each module are as follows:

[0243] Feature Processing Module: Responsible for data acquisition and feature vector generation. It collects operating parameters such as temperature, vibration frequency, and rotational speed of fan components through sensor interfaces (e.g., RS485, Ethernet); receives temporary operating condition adjustment commands through the control system bus (e.g., CAN bus); collects manually detected data through manual input terminals (e.g., touchscreen, keyboard); and obtains environmental anomaly feedback through the environmental monitoring interface. It quantizes and encodes the interactive data (e.g., command to coefficient conversion, difference calculation) to form the first feature vector, extracts features from the operating parameters to form the original operating feature vector, and then uses a fusion algorithm (e.g., weighted fusion) to obtain the second feature vector, providing input for subsequent model adjustments.

[0244] Model adjustment module: Dynamically optimizes model parameters based on the second feature vector. It corrects the vibration feature weights and temperature feature biases of the convolutional layer using a weight correction algorithm, compensates for the rotational speed feature output of the pooling layer using a residual compensation algorithm, adjusts the remaining life prediction results using a multi-dimensional optimization algorithm, obtains the fitted residuals using a residual fitting algorithm with a perturbation mechanism, and finally outputs the integrated and corrected model parameters to ensure the model accurately reflects the part's condition.

[0245] Control Module: Implements iterative model updates and outputs control commands. Based on the corrected model parameters, it calculates the deviation between the predicted remaining lifespan and the actual lifespan. When the deviation exceeds a preset threshold, it triggers a model update (e.g., adjusting feature extraction layer parameters). This update is repeated until the deviation meets the requirements. Based on the optimized remaining lifespan prediction, it determines the maintenance timing, generates spare parts replacement suggestions based on component status, and outputs control commands through execution interfaces (e.g., relay outputs, communication commands) to drive maintenance actuators (e.g., alarm devices, robotic arms) to complete the maintenance operation.

[0246] It should be noted that each module can be implemented using hardware (such as industrial control boards, FPGAs) or software (such as embedded programs, server scripts). Information is transmitted between modules via data buses (such as PCIe, Ethernet) to ensure data real-time performance and reliability. For example, for large wind turbine clusters, the system can adopt a distributed architecture, with feature processing and model adjustment completed at edge nodes, and control commands coordinated by a central server. This application is not limited to this; the system architecture can be flexibly designed according to the application scenario (such as single-machine or cluster).

[0247] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device control method based on the material inspection of wind turbine components, characterized in that, include: The temperature, vibration frequency, speed operating parameters of the fan parts and the interaction data of temporary working condition adjustment commands are obtained, quantified to form the first feature vector and fused to obtain the second feature vector. The parameters of the wind turbine component monitoring model are adjusted based on the second feature vector to obtain the corrected model parameters. The parameter adjustment includes: correcting the vibration feature weights and temperature feature bias values ​​of the convolutional layer, compensating for the rotational speed feature output of the pooling layer, optimizing the remaining life prediction results of the components, and fitting the prediction residuals based on the corrected model parameters; iteratively updating the wind turbine component monitoring model based on the corrected model parameters until the deviation between the predicted remaining life value and the actual life value output by the model is less than a preset threshold, and outputting equipment control commands including maintenance timing and spare parts replacement to complete the material maintenance control of wind turbine components. The parameter adjustment of the wind turbine component monitoring model based on the second feature vector includes: Based on the vibration frequency characteristics, working condition correction values, and interaction feature weights in the second feature vector, the weight correction coefficients of the vibration characteristics of the convolutional layer are calculated; based on the temperature characteristics, measured deviation values, and interaction feature weights, the bias value correction coefficients of the temperature characteristics of the convolutional layer are calculated. The original vibration weights of the convolutional layer are corrected based on the weight correction coefficient to obtain the corrected vibration weights, and the original temperature bias of the convolutional layer is corrected based on the bias correction coefficient to obtain the corrected temperature bias. Based on the rotational speed feature, time decay value and interaction feature weight in the second feature vector, the residual correction amount of the pooling layer is calculated. The residual correction amount is then superimposed on the original rotational speed output obtained by the pooling layer based on the corrected vibration weight and the corrected temperature bias value to obtain the compensated rotational speed output. Based on the second feature vector and the compensated rotational speed output, the optimized remaining life prediction value is obtained by adjusting the model parameters through a multi-dimensional objective function that includes the remaining life prediction bias loss and the interaction feature bias loss. Based on the second feature vector, the optimized remaining lifetime prediction value and confidence value, and the time decay value, a residual algorithm with a perturbation mechanism is used to fit the prediction residual to obtain the fitted residual. The corrected model parameters include the corrected vibration weight, the corrected temperature bias value, the compensated rotational speed output, the optimized remaining life prediction value, and the fitted residual.

2. The method according to claim 1, characterized in that, The step of quantizing the interactive data into a first feature vector and fusing it to obtain a second feature vector includes: Temporary working condition adjustment instructions are encoded as working condition correction values, the difference between manually detected data and model predicted values ​​is encoded as measured deviation values, the reliability of environmental anomaly feedback is encoded as confidence values, and the data acquisition time is encoded as time decay values, resulting in multiple interactive features; the weights of each interactive feature are calculated, and a first feature vector is formed based on the interactive features and their corresponding weights. Calculate the correlation between the temperature and vibration frequency features in the original operating feature vector and the operating condition correction value and measured deviation value in the first feature vector, and then fuse the two based on the correlation to obtain the second feature vector.

3. The method according to claim 2, characterized in that, The calculation of the weights for each interaction feature includes: Obtain the measured deviation value, confidence value, and sensitivity coefficient; perform exponential operations on the values ​​of each interaction feature to obtain the calculated value; calculate the ratio of the calculated value of a single interaction feature to the sum of the calculated values ​​of all interaction features, and use this ratio as the weight of the interaction feature, where the weight of the measured deviation value is positively correlated with the confidence value.

4. The method according to claim 1, characterized in that, The adjustment of model parameters through a multi-dimensional objective function includes: The first loss term is calculated based on the deviation between the actual remaining lifespan of the wind turbine components and the predicted value obtained based on the compensated speed output. The second loss term is calculated based on the deviation between the measured deviation value and the model-predicted deviation value in the interaction features. The first and second loss terms are integrated to form a multi-dimensional objective function. Based on this objective function, the lifetime prediction coefficients of the model are adjusted to obtain the optimized remaining lifetime prediction value.

5. The method according to claim 1, characterized in that, The residual fitting algorithm using a perturbation mechanism includes: The basic residual is calculated based on the difference between the optimized remaining lifetime prediction and the actual remaining lifetime. The basic residual is corrected based on the working condition correction value in the second feature vector to obtain the interactive correction residual; Based on the confidence value and time decay value in the second feature vector, a random perturbation value is generated to simulate the micro-fluctuations of material fatigue as a perturbation term, wherein the higher the confidence value and the smaller the time decay value, the smaller the amplitude of the perturbation value; The fitted residuals are obtained by integrating the basic residuals, the interactively corrected residuals, and the perturbation term.

6. The method according to claim 1, characterized in that, The iterative update of the wind turbine component monitoring model based on the corrected model parameters includes: Calculate the deviation value corresponding to the corrected model parameters, where the deviation value is the difference between the optimized remaining lifetime prediction value and the actual remaining lifetime; When the deviation exceeds the preset threshold, the feature extraction layer parameters of the fan component monitoring model are updated based on the corrected vibration weight and the compensated speed output. Repeat the update steps until the deviation value is less than or equal to the preset threshold. Determine that the model output meets the preset conditions and output the device control command.

7. The method according to claim 6, characterized in that, The deviation values ​​corresponding to the calculated and corrected model parameters include: Calculate the first deviation between the actual remaining life of the wind turbine components and the optimized predicted remaining life; Calculate the second deviation between the measured deviation value and the model-predicted deviation value in the interaction features; The maximum value of the first and second deviations is taken as the deviation value corresponding to the corrected model parameters.

8. The method according to claim 2, characterized in that, The obtained multiple interaction features include: Based on the wind turbine's temporary load reduction and speed increase commands, operating condition correction values ​​reflecting load changes are generated; The difference between the manually measured bearing clearance value and the bearing clearance value predicted by the model is used to obtain the measured deviation value; The confidence level value reflecting the reliability of the data collection personnel is determined based on their qualifications in manual testing. The time decay value, which reflects the effectiveness, is calculated based on the interval between the collection time of the interactive data and the current time.

9. A device control system based on the material inspection of wind turbine parts, characterized in that, include: The feature processing module acquires the temperature, vibration frequency, speed operating parameters of the fan parts and the interaction data of temporary working condition adjustment commands, quantifies them to form a first feature vector and fuses them to obtain a second feature vector. The model adjustment module corrects the convolutional layer weights, compensates the pooling layer output, optimizes the remaining lifetime prediction results and fitting residuals based on the second feature vector, and obtains the corrected model parameters. The control module iteratively updates the monitoring model based on the corrected model parameters until the remaining life prediction deviation meets the threshold, and outputs control commands including maintenance timing to complete the maintenance control. The model adjustment module is specifically used for: Based on the vibration frequency characteristics, working condition correction values, and interaction feature weights in the second feature vector, the weight correction coefficients of the vibration characteristics of the convolutional layer are calculated; based on the temperature characteristics, measured deviation values, and interaction feature weights, the bias value correction coefficients of the temperature characteristics of the convolutional layer are calculated. The original vibration weights of the convolutional layer are corrected based on the weight correction coefficient to obtain the corrected vibration weights, and the original temperature bias of the convolutional layer is corrected based on the bias correction coefficient to obtain the corrected temperature bias. Based on the rotational speed feature, time decay value and interaction feature weight in the second feature vector, the residual correction amount of the pooling layer is calculated. The residual correction amount is then superimposed on the original rotational speed output obtained by the pooling layer based on the corrected vibration weight and the corrected temperature bias value to obtain the compensated rotational speed output. Based on the second feature vector and the compensated rotational speed output, the optimized remaining life prediction value is obtained by adjusting the model parameters through a multi-dimensional objective function that includes the remaining life prediction bias loss and the interaction feature bias loss. Based on the second feature vector, the optimized remaining lifetime prediction value and confidence value, and the time decay value, a residual algorithm with a perturbation mechanism is used to fit the prediction residual to obtain the fitted residual. The corrected model parameters include the corrected vibration weight, the corrected temperature bias value, the compensated rotational speed output, the optimized remaining life prediction value, and the fitted residual.

Images