Part service life analysis method based on load data driving

CN118917008AInactive Publication Date: 2024-11-08HEFEI MEIANGXING ELECTRONIC TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410919212.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-10
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

Smart Images

  • Figure CN118917008A_ABST
    Figure CN118917008A_ABST
Patent Text Reader

Abstract

The invention discloses a part service life analysis method based on load data driving. The part service life analysis method comprises the following steps that S1, load data are collected in real time through a sensor; s2, preprocessing the load data; s3, fusing the preprocessed load data to form a multi-source data fusion model; s4, inputting the data of the multi-source data fusion model into a convolutional neural network, and extracting spatial features; s5, inputting the spatial feature data into an automatic encoder, and performing dimensionality reduction and feature extraction; s6, inputting the data after dimensionality reduction and feature extraction into a recurrent neural network, and capturing time sequence features; s7, further processing the time sequence features through a long short-term memory network; s8, constructing a deep learning model; s9, introducing a self-adaptive online learning mechanism; s10, carrying out life prediction; and S11, when abnormity occurs or the predicted service life is close to the limit, an early warning signal is automatically sent out. According to the method, deep learning and multi-source data fusion are utilized to realize high-precision part service life prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of mechanical engineering and deep learning technology, and in particular to a component life analysis method driven by load data. Background Art

[0002] In the field of mechanical engineering, the life prediction of components has always been an important research direction. Traditional methods for predicting the life of components mainly rely on mechanical stress analysis and empirical formulas. These methods are usually based on static or simplified load models and cannot accurately reflect the complex dynamic loads borne by components under actual working conditions, resulting in large errors in the life prediction results.

[0003] Traditional mechanical stress analysis methods mainly use finite element analysis to simulate the stress distribution and strain state of parts during operation. Although these methods can provide accurate stress and strain distribution, their high computational complexity places high demands on the real-time performance and flexibility of practical applications. In addition, finite element analysis relies on idealized boundary conditions and load conditions and cannot fully reflect the dynamic changes in the actual working environment.

[0004] The empirical formula method is another common life prediction method. This method estimates the life of parts by summarizing a large amount of experimental data and actual usage to establish an empirical formula. This method is simple and easy to use, but its accuracy depends on the richness of the empirical data and the scope of application of the formula. In practical applications, due to the diversity and complexity of the working environment, empirical formulas are often difficult to provide accurate prediction results.

[0005] In recent years, data-driven analysis methods have been gradually introduced to predict the life of components under complex dynamic load conditions. These methods use sensors to collect real-time load data of components during operation, and predict the life of components through data analysis and modeling. Data-driven methods can reflect the actual load conditions borne by components under actual working conditions and have high prediction accuracy and flexibility.

[0006] However, existing data-driven life prediction methods still have some shortcomings. First, most methods rely only on a single type of sensor data, such as stress data or vibration data, which cannot fully reflect the actual working status of components. Secondly, traditional data analysis methods often face problems such as low computational efficiency and insufficient feature extraction when processing large-scale and complex load data, making it difficult to meet real-time requirements. Finally, existing methods lack adaptive capabilities and cannot dynamically adjust the prediction model according to changes in actual working conditions, resulting in insufficient robustness and accuracy of the prediction results.

[0007] Therefore, how to provide a component life analysis method driven by load data is an urgent problem that technicians in this field need to solve. Summary of the invention

[0008] One purpose of the present invention is to propose a component life analysis method driven by load data. The present invention utilizes deep learning and multi-source data fusion, which not only solves the problem that traditional methods are difficult to accurately predict component life under complex dynamic load conditions, but also optimizes various links such as data acquisition, preprocessing, fusion, analysis and prediction. It has the characteristics of high precision, high real-time performance and high adaptability, and provides strong technical support for component life analysis and maintenance.

[0009] The component life analysis method based on load data driving according to an embodiment of the present invention comprises the following steps:

[0010] S1. Collecting load data of components in real time during operation through sensors, wherein the sensors include stress sensors, accelerometers and temperature sensors, and the load data include stress data, vibration data and temperature data;

[0011] S2. Preprocessing the load data, including data cleaning, missing value filling and standardization operations;

[0012] S3, fusing the preprocessed load data to form a multi-source data fusion model;

[0013] S4, inputting the data of the multi-source data fusion model into a convolutional neural network to extract spatial features of the data;

[0014] S5, inputting the spatial feature data into an automatic encoder for dimensionality reduction and feature extraction;

[0015] S6, inputting the data after dimension reduction and feature extraction into a recurrent neural network to capture the time series characteristics of the data;

[0016] S7, further processing the time series features through a long short-term memory network;

[0017] S8, combining the convolutional neural network, autoencoder, recurrent neural network and long short-term memory network to construct a deep learning model;

[0018] S9. Introducing an adaptive online learning mechanism to adjust the parameters of the deep learning model according to the load data acquired in real time;

[0019] S10, using the deep learning model to perform life prediction and generate life prediction results of parts;

[0020] S11. When the load data is abnormal or the predicted life is close to the limit, an early warning signal is automatically issued.

[0021] Furthermore, the S3 specifically includes:

[0022] S31, obtaining the pre-processed stress data, vibration data and temperature data, and marking them as stress data sequence D s ={d s1 ,d s2 ,…,d sn}, vibration data sequence D v ={d v1 ,d v2 ,…,d vn}, temperature data series D t ={d t1 ,d t2 ,…,d tn}, where n represents the number of time steps of data acquisition;

[0023] S32, normalizing the stress data sequence, the vibration data sequence and the temperature data sequence using a normalization formula:

[0024]

[0025] Among them, d ij Represents the value of the i-th category data at the j-th time step, i∈{s,v,t}, j∈{1,2,…,n}, d′ ij Represents the normalized data, max(D i ) represents the maximum value of the i-th category data, min(D i ) represents the minimum value of the i-th category data;

[0026] S33, aligning the normalized stress data sequence, vibration data sequence and temperature data sequence according to time steps to form a time series matrix M:

[0027]

[0028] S34, performing data enhancement processing on the time series matrix M, generating more sample data through interpolation and smoothing operations, and expanding the data set;

[0029] S35, perform data fusion processing on the expanded time series matrix M, and use a linear weighted fusion method to fuse multi-source data into a comprehensive feature vector:

[0030] F j =w s ·d′ sj +w v ·d′vj +w t ·d′ tj ;

[0031] Among them, F j represents the comprehensive feature vector of the jth time step, w s Represents the weight coefficient of stress data, w v Represents the weight coefficient of vibration data, w t Represents the weight coefficient of temperature data and satisfies:

[0032] w s +w v +w t =1;

[0033] S36, adjusting the weight coefficient w through an iterative optimization algorithm s , w v and w t , build and optimize multi-source data fusion models.

[0034] Furthermore, the S4 specifically includes:

[0035] S41, converting the data of the multi-source data fusion model into the input format of a convolutional neural network, assuming that the input data is a three-dimensional tensor X, wherein the dimension of X is [N, H, W, C], N represents the number of samples, H represents the height of the data, W represents the width of the data, and C represents the number of channels;

[0036] S42, normalizing the three-dimensional tensor X to ensure that the data value is within the range of [0, 1] or [-1, 1] to improve the training effect and stability of the convolutional neural network;

[0037] S43, inputting the normalized data into the convolution layer, and extracting spatial features through convolution operation, the formula of the convolution operation is:

[0038]

[0039] Among them, Y i,j,k represents the convolution result, X i,j Represents the pixel value of the input data, W m,n,k represents the weight of the kth convolution kernel, b k represents the bias of the kth convolution kernel, M represents the height of the convolution kernel, N represents the width of the convolution kernel, i represents the row index of the output feature map, j represents the column index of the output feature map, and k represents the channel index;

[0040] S44, activate the convolution result, using the nonlinear activation function ReLU:

[0041] f(x)=max(0,x);

[0042] Where x represents the input value;

[0043] S45, input the activated feature map into the pooling layer for downsampling operation. The formula of pooling operation is:

[0044]

[0045] Among them, P i,j,k Represents the pooling result, using the maximum pooling operation, M represents the height of the pooling window, N represents the width of the pooling window, i represents the row index of the pooling result, j represents the column index of the pooling result, and k represents the channel index;

[0046] S46, repeat the above convolution and pooling operations to extract higher-level spatial features layer by layer to form a multi-layer feature map;

[0047] S47, flatten the feature map of the last pooling layer, convert it into a one-dimensional vector, and input it into the fully connected layer for further feature extraction and classification operations;

[0048] S48. Output the finally extracted spatial feature data.

[0049] Furthermore, the S6 specifically includes:

[0050] S61, sorting the data after dimension reduction and feature extraction in chronological order to form a time series data set;

[0051] S62, defining the input format of the time series data as:

[0052] X IN ={x1,x2,…,x T};

[0053] Where T represents the time step, x t Represents the feature data vector corresponding to time step t;

[0054] S63, inputting the time series data into a recurrent neural network, wherein the hidden state h of the recurrent neural network t Calculated recursively by the following formula:

[0055] h t =f(W xh x t +W hh h t-1 +b h );

[0056] Among them, h t represents the hidden state at time step t, x trepresents the feature data vector corresponding to time step t, W xh Represents the weight matrix input to the hidden state, W hh Represents the weight matrix from the previous hidden state to the current hidden state, b h represents the bias term, and f represents the activation function;

[0057] S64, normalizing the output of the recurrent neural network to ensure data stability of the output data, the normalization formula is:

[0058]

[0059] in, represents the normalized hidden state, h t represents the hidden state of time step t, μ represents the mean, and σ represents the standard deviation;

[0060] S65. Output the final time series feature data.

[0061] Furthermore, the S7 specifically includes:

[0062] S71, inputting the time series feature data into a long short-term memory network for processing;

[0063] S72, performing refinement processing on the time series data through an input gate, a forget gate, and an output gate in the long short-term memory network;

[0064] The calculation formula of the input gate is:

[0065] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0066] Among them, i t represents the state of the input gate, σ represents the activation function, W i represents the weight matrix, h t-1 represents the hidden state of the previous time step, x t represents the input of the current time step, b i represents the bias vector;

[0067] The calculation formula of the forget gate is:

[0068] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0069] Among them, ft Represents the state of the forget gate, W f represents the weight matrix, b f represents the bias vector;

[0070] The calculation formula of the output gate is:

[0071] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0072] Among them, t Indicates the state of the output gate, W o represents the weight matrix, b o represents the bias vector;

[0073] S73. Calculate the new cell state C at each time step t and the hidden state h t :

[0074] The cell state update formula is:

[0075]

[0076] Among them, C t represents the cell state at the current time step, C t-1 represents the cell state at the previous time step, Represents the candidate cell state, and the calculation formula is:

[0077]

[0078] Among them, tanh represents the hyperbolic tangent activation function, W C represents the weight matrix, b C represents the bias vector;

[0079] The update formula of the hidden state is:

[0080] h t =o t tanh(C t );

[0081] Among them, h t represents the hidden state of the current time step;

[0082] S74. The long short-term memory network captures the long-term dependencies of time series features and generates input data for a deep learning model.

[0083] The beneficial effects of the present invention are:

[0084] (1) Comprehensive load data collection: Real-time collection of stress data, vibration data, and temperature data through pressure sensors, accelerometers, and temperature sensors can fully reflect the actual working status of components. The combination of multiple sensor data overcomes the limitations of traditional single data sources and improves the comprehensiveness and accuracy of data.

[0085] (2) Efficient data preprocessing: Data cleaning, missing value filling and standardization operations are performed on the collected load data to ensure data quality and consistency. The preprocessing step effectively removes noise and outliers, completes incomplete data, and eliminates dimensional differences between different data sources through standardization operations, laying a good foundation for subsequent data analysis.

[0086] (3) Multi-source data fusion: The pre-processed data from various sensors are fused to form a multi-source data fusion model, which comprehensively analyzes various data such as stress, vibration, and temperature. The multi-source data fusion model effectively improves the comprehensiveness and accuracy of life prediction and makes up for the shortcomings of a single data source.

[0087] (4) Advanced deep learning model: Convolutional neural network is used to extract spatial features, autoencoder is used for dimensionality reduction and feature extraction, recurrent neural network is used to capture time series features, and long short-term memory network is used to further process time series features. The deep learning model can process both spatial and time series features of load data, thus improving the accuracy of life prediction.

[0088] (5) Adaptive online learning: An adaptive online learning mechanism is introduced to dynamically adjust the parameters of the deep learning model according to the load data obtained in real time, so that the model can be updated in real time and maintain adaptability to actual working conditions. Adaptive online learning improves the robustness and prediction accuracy of the model and ensures efficient prediction under complex dynamic load conditions.

[0089] (6) High-precision life prediction: The deep learning model is used to predict the life of the components, and the prediction accuracy is significantly improved. Compared with the traditional mechanical stress analysis and empirical formula methods, the method of the present invention can more accurately reflect the remaining life of the components under actual working conditions.

[0090] (7) Intelligent early warning system: When the load data is abnormal or the predicted life is close to the limit, an early warning signal is automatically issued to remind the user to perform maintenance. The intelligent early warning system can effectively prevent the sudden failure of parts during operation, thereby improving the safety and reliability of equipment operation.

[0091] (8) Improve maintenance efficiency: Provide specific maintenance suggestions based on early warning signals to help users formulate maintenance plans, reduce unplanned downtime and maintenance costs, and improve the overall operating efficiency and economic benefits of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0093] Figure 1 A flow chart of a component life analysis method based on load data drive proposed by the present invention;

[0094] Figure 2 This is a structural diagram of the multi-source data fusion model of the load data-driven component life analysis method proposed in the present invention. DETAILED DESCRIPTION

[0095] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0096] refer to Figure 1 and Figure 2 , the component life analysis method based on load data drive includes the following steps:

[0097] S1. Collecting load data of components in real time during operation through sensors, wherein the sensors include stress sensors, accelerometers and temperature sensors, and the load data include stress data, vibration data and temperature data;

[0098] S2. Preprocessing the load data, including data cleaning, missing value filling and standardization operations;

[0099] S3, fusing the preprocessed load data to form a multi-source data fusion model;

[0100] S4, inputting the data of the multi-source data fusion model into a convolutional neural network to extract spatial features of the data;

[0101] S5, inputting the spatial feature data into an automatic encoder for dimensionality reduction and feature extraction;

[0102] S6, inputting the data after dimension reduction and feature extraction into a recurrent neural network to capture the time series characteristics of the data;

[0103] S7, further processing the time series features through a long short-term memory network;

[0104] S8, combining the convolutional neural network, autoencoder, recurrent neural network and long short-term memory network to construct a deep learning model;

[0105] S9. Introducing an adaptive online learning mechanism to adjust the parameters of the deep learning model according to the load data acquired in real time;

[0106] S10, using the deep learning model to perform life prediction and generate life prediction results of parts;

[0107] S11. When the load data is abnormal or the predicted life is close to the limit, an early warning signal is automatically issued.

[0108] In this implementation, S3 specifically includes:

[0109] S31, obtaining the pre-processed stress data, vibration data and temperature data, and marking them as stress data sequence D s ={d s1 ,d s2 ,…,d sn}, vibration data sequence D v ={d v1 ,d v2 ,…,d vn}, temperature data series D t ={d t1 ,d t2 ,…,d tn}, where n represents the number of time steps of data acquisition;

[0110] S32, normalizing the stress data sequence, the vibration data sequence and the temperature data sequence using a normalization formula:

[0111]

[0112] Among them, d ij Represents the value of the i-th category data at the j-th time step, i∈{s,v,t}, j∈{1,2,…,n}, d′ ij Represents the normalized data, max(D i ) represents the maximum value of the i-th category data, min(D i ) represents the minimum value of the i-th category data;

[0113] S33, aligning the normalized stress data sequence, vibration data sequence and temperature data sequence according to time steps to form a time series matrix M:

[0114]

[0115] S34, performing data enhancement processing on the time series matrix M, generating more sample data through interpolation and smoothing operations, and expanding the data set;

[0116] S35, perform data fusion processing on the expanded time series matrix M, and use a linear weighted fusion method to fuse multi-source data into a comprehensive feature vector:

[0117] F j =w s ·d′ sj +w v ·d′ vj +w t ·d′ tj ;

[0118] Among them, F j represents the comprehensive feature vector of the jth time step, w s Represents the weight coefficient of stress data, w v Represents the weight coefficient of vibration data, w t Represents the weight coefficient of temperature data and satisfies:

[0119] w s +w v +w t =1;

[0120] S36, adjusting the weight coefficient w through an iterative optimization algorithm s , w v and w t , build and optimize multi-source data fusion models.

[0121] In this implementation manner, the S4 specifically includes:

[0122] S41, converting the data of the multi-source data fusion model into the input format of a convolutional neural network, assuming that the input data is a three-dimensional tensor X, wherein the dimension of X is [N, H, W, C], N represents the number of samples, H represents the height of the data, W represents the width of the data, and C represents the number of channels;

[0123] S42, normalizing the three-dimensional tensor X to ensure that the data value is within the range of [0, 1] or [-1, 1] to improve the training effect and stability of the convolutional neural network;

[0124] S43, inputting the normalized data into the convolution layer, and extracting spatial features through convolution operation, the formula of the convolution operation is:

[0125]

[0126] Among them, Y i,j,k represents the convolution result, X i,j Represents the pixel value of the input data, W m,n,k represents the weight of the kth convolution kernel, b krepresents the bias of the kth convolution kernel, M represents the height of the convolution kernel, N represents the width of the convolution kernel, i represents the row index of the output feature map, j represents the column index of the output feature map, and k represents the channel index;

[0127] S44, activate the convolution result, using the nonlinear activation function ReLU:

[0128] f(x)=max(0,x);

[0129] Where x represents the input value;

[0130] S45, input the activated feature map into the pooling layer for downsampling operation. The formula of pooling operation is:

[0131]

[0132] Among them, P i,j,k Represents the pooling result, using the maximum pooling operation, M represents the height of the pooling window, N represents the width of the pooling window, i represents the row index of the pooling result, j represents the column index of the pooling result, and k represents the channel index;

[0133] S46, repeat the above convolution and pooling operations to extract higher-level spatial features layer by layer to form a multi-layer feature map;

[0134] S47, flatten the feature map of the last pooling layer, convert it into a one-dimensional vector, and input it into the fully connected layer for further feature extraction and classification operations;

[0135] S48. Output the finally extracted spatial feature data.

[0136] In this implementation manner, S6 specifically includes:

[0137] S61, sorting the data after dimension reduction and feature extraction in chronological order to form a time series data set;

[0138] S62, defining the input format of the time series data as:

[0139] X IN ={x1,x2,…,x T};

[0140] Where T represents the time step, x t Represents the feature data vector corresponding to time step t;

[0141] S63, inputting the time series data into a recurrent neural network, wherein the hidden state h of the recurrent neural network t Calculated recursively by the following formula:

[0142] ht =f(W xh x t +W hh h t-1 +b h );

[0143] Among them, h t represents the hidden state at time step t, x t represents the feature data vector corresponding to time step t, W xh Represents the weight matrix input to the hidden state, W hh Represents the weight matrix from the previous hidden state to the current hidden state, b h represents the bias term, and f represents the activation function;

[0144] S64, normalizing the output of the recurrent neural network to ensure data stability of the output data, the normalization formula is:

[0145]

[0146] in, represents the normalized hidden state, h t represents the hidden state of time step t, μ represents the mean, and σ represents the standard deviation;

[0147] S65. Output the final time series feature data.

[0148] In this implementation manner, the S7 specifically includes:

[0149] S71, inputting the time series feature data into a long short-term memory network for processing;

[0150] S72, performing refinement processing on the time series data through an input gate, a forget gate, and an output gate in the long short-term memory network;

[0151] The calculation formula of the input gate is:

[0152] i t =σ(W i ·[h t-1 ,x i ]+b i );

[0153] Among them, i t represents the state of the input gate, σ represents the activation function, W i represents the weight matrix, h t-1 represents the hidden state of the previous time step, x t represents the input of the current time step, b i represents the bias vector;

[0154] The calculation formula of the forget gate is:

[0155] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0156] Among them, f t Represents the state of the forget gate, W f represents the weight matrix, b f represents the bias vector;

[0157] The calculation formula of the output gate is:

[0158] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0159] Among them, t Indicates the state of the output gate, W o represents the weight matrix, b o represents the bias vector;

[0160] S73. Calculate the new cell state C at each time step t and the hidden state h t :

[0161] The cell state update formula is:

[0162]

[0163] Among them, C t represents the cell state at the current time step, C t-1 represents the cell state at the previous time step, Represents the candidate cell state, and the calculation formula is:

[0164]

[0165] Among them, tanh represents the hyperbolic tangent activation function, W C represents the weight matrix, b C represents the bias vector;

[0166] The update formula of the hidden state is:

[0167] h t =o t tanh(C t );

[0168] Among them, h trepresents the hidden state of the current time step;

[0169] S74. The long short-term memory network captures the long-term dependencies of time series features and generates input data for a deep learning model.

[0170] Embodiment 1:

[0171] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the life prediction of the gearbox, a key component of a wind turbine generator set in a large wind farm. The gearbox plays an important transmission role in the wind turbine generator set, and its working state directly affects the operating efficiency and reliability of the entire power generation system. However, since the gearbox is subjected to complex dynamic loads during operation, including stress fluctuations, mechanical vibrations, and temperature changes caused by wind speed changes, traditional mechanical stress analysis and empirical formulas are difficult to accurately predict its life. Based on this, we apply the method of the present invention to achieve high-precision prediction of its life by real-time collection and analysis of the load data of the gearbox, and timely issue maintenance recommendations through the intelligent early warning system.

[0172] A variety of sensors are installed on the gearbox of the wind turbine, including stress sensors, accelerometers and temperature sensors, which are used to collect stress data, vibration data and temperature data of the gearbox during operation. The sensors collect data in real time and transmit it to the data processing center.

[0173] First, the load data collected by the sensor is preprocessed. Data preprocessing includes data cleaning, missing value filling and standardization operations. During the data cleaning process, the noise and outliers that may exist in the collection process are removed; the missing data is filled by interpolation method; finally, the data is standardized to eliminate the dimensional differences between different data sources.

[0174] The pre-processed load data is fused with multi-source data to form a comprehensive multi-source data fusion model. This model can comprehensively analyze the stress, vibration and temperature data from different sensors to fully reflect the actual working status of the gearbox.

[0175] Based on the multi-source data fusion model, the convolutional neural network (CNN) is used to extract the spatial features of the data. CNN has significant advantages in processing images and spatial data. Its convolution and pooling operations can effectively extract the spatial features in the load data. Then, the extracted spatial feature data is input into the autoencoder for dimensionality reduction and feature extraction. The autoencoder compresses the data dimension, extracts important features in the data, and reduces data redundancy through the encoding and decoding process.

[0176] The data after dimensionality reduction and feature extraction is input into a recurrent neural network (RNN) to capture the time series characteristics of the data. RNN performs well in processing sequence data and can remember and use historical information to model time series data. Furthermore, the long short-term memory network (LSTM) is used to process time series features to enhance the model's ability to remember long-term dependencies. LSTM can solve the gradient vanishing and explosion problems of ordinary RNN when processing long sequences, and improve the ability to capture time series features.

[0177] A deep learning model is constructed by combining convolutional neural networks, autoencoders, recurrent neural networks, and long short-term memory networks. This model can simultaneously process the spatial and time series characteristics of load data and provide high-precision life prediction.

[0178] An adaptive online learning mechanism is introduced to dynamically adjust the parameters of the deep learning model according to the load data obtained in real time. Adaptive online learning enables the model to be updated in real time, maintains adaptability to actual working conditions, and improves prediction accuracy and robustness.

[0179] The life prediction of the gearbox is generated through the deep learning model. When the load data is abnormal or the predicted life is close to the limit, the intelligent early warning system will automatically send out an early warning signal to remind the maintenance personnel to carry out maintenance. The system will also provide specific maintenance suggestions based on the early warning information to help formulate a maintenance plan to avoid sudden failure of the gearbox during operation.

[0180] We have put this method into practical use in a large wind farm and conducted a comparative analysis of the experimental data of traditional mechanical stress analysis and the method of the present invention.

[0181] Table 1 Data comparison between traditional mechanical stress analysis and the method of the present invention

[0182]

[0183] It can be seen from the data in Table 1 that the method of the present invention can achieve high-precision prediction of component life under complex dynamic load conditions, has the advantages of strong real-time performance, high adaptability and intelligent early warning, and provides a strong guarantee for the operation and maintenance of wind turbines.

[0184] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A component life analysis method based on load data drive, characterized in that: The steps include: S1. Collecting load data of components in real time during operation through sensors, wherein the sensors include stress sensors, accelerometers and temperature sensors, and the load data include stress data, vibration data and temperature data; S2. Preprocessing the load data, including data cleaning, missing value filling and standardization operations; S3, fusing the preprocessed load data to form a multi-source data fusion model; S4, inputting the data of the multi-source data fusion model into a convolutional neural network to extract the spatial features of the data; S5, inputting the spatial feature data into an automatic encoder for dimensionality reduction and feature extraction; S6, inputting the data after dimension reduction and feature extraction into a recurrent neural network to capture the time series characteristics of the data; S7, further processing the time series features through a long short-term memory network; S8, combining the convolutional neural network, autoencoder, recurrent neural network and long short-term memory network to construct a deep learning model; S9. Introducing an adaptive online learning mechanism to adjust the parameters of the deep learning model according to the load data acquired in real time; S10, using the deep learning model to perform life prediction and generate life prediction results of parts; S11. When the load data is abnormal or the predicted life is close to the limit, an early warning signal is automatically issued.

2. The component life analysis method based on load data drive according to claim 1 is characterized in that: The S3 specifically includes: S31, obtaining the pre-processed stress data, vibration data and temperature data, and marking them as stress data sequence D s ={d s1 ,d s2 ,…,d sn }, vibration data sequence D v ={d v1 ,d v2 ,…,d vn }, temperature data series D t ={d t1 ,d t2 ,…,d tn }, where n represents the number of time steps of data acquisition; S32, normalizing the stress data sequence, the vibration data sequence and the temperature data sequence using a normalization formula: Among them, d ij Represents the value of the i-th category data at the j-th time step, i∈{s,v,t}, j∈{1,2,…,n}, d′ ij Represents the normalized data, max(D i ) represents the maximum value of the i-th category data, min(D i ) represents the minimum value of the i-th category data; S33, aligning the normalized stress data sequence, vibration data sequence and temperature data sequence according to time steps to form a time series matrix M: S34, performing data enhancement processing on the time series matrix M, generating more sample data through interpolation and smoothing operations, and expanding the data set; S35, perform data fusion processing on the expanded time series matrix M, and use a linear weighted fusion method to fuse multi-source data into a comprehensive feature vector: F j =w s ·d′ sj +w v ·d′ vj +w t ·d′ tj ; Among them, F j represents the comprehensive feature vector at the jth time step, w s Represents the weight coefficient of stress data, w v Represents the weight coefficient of vibration data, w t Represents the weight coefficient of temperature data and satisfies: In s +in v +in t =1; S36, adjusting the weight coefficient w through an iterative optimization algorithm s , w v and w t , build and optimize multi-source data fusion models.

3. The component life analysis method based on load data drive according to claim 1 is characterized in that: The S4 specifically includes: S41, converting the data of the multi-source data fusion model into the input format of a convolutional neural network, assuming that the input data is a three-dimensional tensor X, wherein the dimension of X is [N, H, W, C], N represents the number of samples, H represents the height of the data, W represents the width of the data, and C represents the number of channels; S42, normalizing the three-dimensional tensor X to ensure that the data value is within the range of [0, 1] or [-1, 1] to improve the training effect and stability of the convolutional neural network; S43, inputting the normalized data into the convolution layer, and extracting spatial features through convolution operation, the formula of the convolution operation is: Among them, Y ,j,k represents the convolution result, X i,j Represents the pixel value of the input data, W m,n,k represents the weight of the kth convolution kernel, b k represents the bias of the kth convolution kernel, M represents the height of the convolution kernel, N represents the width of the convolution kernel, i represents the row index of the output feature map, j represents the column index of the output feature map, and k represents the channel index; S44, activate the convolution result, using the nonlinear activation function ReLU: f(x)=max(0,x); Where x represents the input value; S45, input the activated feature map into the pooling layer for downsampling operation, and the formula of pooling operation is: Among them, P i,j,k Represents the pooling result, using the maximum pooling operation, M represents the height of the pooling window, N represents the width of the pooling window, i represents the row index of the pooling result, j represents the column index of the pooling result, and k represents the channel index; S46, repeat the above convolution and pooling operations to extract higher-level spatial features layer by layer to form a multi-layer feature map; S47, flatten the feature map of the last pooling layer, convert it into a one-dimensional vector, and input it into the fully connected layer for further feature extraction and classification operations; S48. Output the finally extracted spatial feature data.

4. The component life analysis method based on load data drive according to claim 1 is characterized in that: The S6 specifically includes: S61, sorting the data after dimension reduction and feature extraction in chronological order to form a time series data set; S62, defining the input format of the time series data as: X IN ={x1,t2,…,x T }; Where T represents the time step, x t Represents the feature data vector corresponding to time step t; S63, inputting the time series data into a recurrent neural network, wherein the hidden state h of the recurrent neural network t Calculated recursively by the following formula: h t =f(W xh x t +W hh h t-1 +b h ); Among them, h t represents the hidden state at time step t, x t represents the feature data vector corresponding to time step t, W xh Represents the weight matrix input to the hidden state, W hh Represents the weight matrix from the previous hidden state to the current hidden state, b h represents the bias term, and f represents the activation function; S64, normalizing the output of the recurrent neural network to ensure data stability of the output data, the normalization formula is: in, represents the normalized hidden state, h t represents the hidden state of time step t, μ represents the mean, and σ represents the standard deviation; S65. Output the final time series feature data.

5. The component life analysis method based on load data drive according to claim 1 is characterized in that: The S7 specifically includes: S71, inputting the time series feature data into a long short-term memory network for processing; S72, performing refinement processing on the time series data through an input gate, a forget gate, and an output gate in the long short-term memory network; The calculation formula of the input gate is: i t =σ(W i ·[h t-1 ,x t ]+b i ); Among them, i t represents the state of the input gate, σ represents the activation function, W i represents the weight matrix, h t-1 represents the hidden state of the previous time step, x t represents the input of the current time step, b i represents the bias vector; The calculation formula of the forget gate is: f t =σ(W f ·[h t-1 ,x t ]+b f ); Among them, f t Represents the state of the forget gate, W f represents the weight matrix, b f represents the bias vector; The calculation formula of the output gate is: the t =σ(W o ·[h t-1 ,x t ]+b o ); Among them, t Indicates the state of the output gate, W o represents the weight matrix, b o represents the bias vector; S73. Calculate the new cell state C at each time step t and the hidden state h t : The cell state update formula is: Among them, C t represents the cell state at the current time step, C t-1 represents the cell state at the previous time step, Represents the candidate cell state, and the calculation formula is: Among them, tanh represents the hyperbolic tangent activation function, W C represents the weight matrix, b C represents the bias vector; The update formula of the hidden state is: h t =o t ·tanh(C t ); Among them, h t represents the hidden state of the current time step; S74. The long short-term memory network captures the long-term dependencies of time series features and generates input data for a deep learning model.