LSTM-Attention Fusion Gearbox Oil Temperature Prediction and Early Warning Method and System
By constructing an LSTM-attention fusion network and an early warning threshold setting mechanism, the robustness and accuracy of gearbox oil temperature prediction in existing technologies are solved, achieving more efficient fault early warning and improved interpretability.
Patent Information
- Application Number
- CN202610334328.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-30
AI Technical Summary
Existing LSTM-based gearbox oil temperature prediction methods lack robustness and adaptability in practical applications, especially in complex wind power scenarios where prediction accuracy is insufficient and interpretation is difficult.
An LSTM-attention fusion network is constructed. Through data cleaning, feature selection and attention mechanism, combined with the exponential weighted moving average algorithm to set the warning threshold, the model's noise resistance and prediction accuracy are improved, and intuitive fault decision-making basis is provided.
It improves the prediction accuracy under varying operating conditions, enhances the sensitivity to subtle fault characteristics, and improves the interpretability and robustness of the early warning system.
Smart Images

Figure CN122310168A_ABST
Abstract
Description
Technical Field
[0001] The embodiments disclosed herein belong to the field of gearbox oil temperature prediction and early warning technology, specifically relating to an LSTM-attention fusion gearbox oil temperature prediction and early warning method and system. Background Technology
[0002] The gearbox of a wind turbine generator is a critical component in the transmission chain, and its operating status directly affects the reliability and power generation efficiency of the entire unit. Gearbox oil temperature is a core parameter reflecting its health status; abnormal temperature rise usually indicates lubrication failure, component wear, or cooling system malfunction.
[0003] Existing methods for oil temperature prediction based on LSTM models can learn long-term dependencies and effectively transmit long-term information. However, in terms of data engineering, these models often rely heavily on a large amount of clean and balanced labeled data and lack robustness to problems such as strong noise that are common in practice. Secondly, in terms of engineering generalization, many prediction schemes have weak adaptability to the complex scenarios of wind power conditions. Summary of the Invention
[0004] The embodiments disclosed herein aim to at least solve one of the technical problems existing in the prior art, and provide an LSTM-attention fusion gearbox oil temperature prediction and early warning method and system.
[0005] One aspect of this disclosure provides an LSTM-attention fusion method for predicting and warning gearbox oil temperature, the method comprising: Collect SCADA data of wind turbine gearbox under normal operating conditions, and perform data cleaning, normalization and feature selection on the SCADA data to obtain a health dataset; An offline LSTM-attention fusion network is constructed, consisting of an input layer, a first LSTM layer, a second LSTM layer, an attention layer, and a fully connected output layer. The offline LSTM-attention fusion network is trained using the health dataset to obtain the optimal prediction model; The monitoring data of the wind turbine gearbox is collected in real time, and after data cleaning, normalization and feature selection, it is input into the optimal prediction model to output the predicted value of gearbox oil temperature. Calculate the residual between the predicted value and the corresponding actual oil temperature value, set a warning threshold based on the residual, and determine whether the gearbox has malfunctioned based on the comparison result between the residual and the warning threshold.
[0006] Furthermore, the data cleaning includes: Outliers are removed using percentiles, isolated points are marked using the DBSCAN algorithm, and outliers are deleted and corrected using moving averages or interpolation.
[0007] Furthermore, the feature selection includes: The random forest algorithm is used to select relevant features for oil temperature prediction from the SCADA data; wherein, The relevant characteristics include at least one or more of the following: wind speed, power, ambient temperature, nacelle temperature, bearing temperature, gear oil inlet temperature, pressure, and rotational speed.
[0008] Furthermore, the first LSTM layer receives the input sequence and outputs the hidden state sequence of all time steps, and the second LSTM layer receives the hidden state of the last time step of the first LSTM layer as the initial input and outputs the hidden state sequence of all time steps to the attention layer. The attention layer is a multi-head self-attention layer, used to weight and integrate the hidden state sequence output by the second LSTM layer, and output the context vector to the fully connected output layer.
[0009] Furthermore, the attention layer employs a scaled dot product attention mechanism, as shown in the following equation:
[0010] In the formula, , , These are query, key, and value matrices, respectively. Represents the transpose of a matrix. denoted as the dimension of the key matrix.
[0011] Furthermore, the step of setting an early warning threshold based on the residual includes: The residuals are smoothed by a moving window moving average to obtain the mean and standard deviation of the residuals; Based on the mean and standard deviation of the residuals, an exponentially weighted moving average algorithm is used to calculate the dynamic upper limit and lower limit of residual control, and the upper limit and lower limit of residual control are used as early warning thresholds.
[0012] Furthermore, the calculation of the dynamic residual control upper limit and residual control lower limit is shown in the following formula:
[0013] In the formula, , These are the upper and lower limits for residual control, respectively. , These are the mean and standard deviation of the residuals, respectively. , As a preset constant, For time.
[0014] Another aspect of this disclosure provides an LSTM-attention fusion gearbox oil temperature prediction and early warning system, the system comprising: The data acquisition and processing module is used to acquire SCADA data of the wind turbine gearbox under normal operating conditions, and to perform data cleaning, normalization and feature selection on the SCADA data to obtain a health dataset. The model building module is used to construct an offline LSTM-attention fusion network that sequentially includes an input layer, a first LSTM layer, a second LSTM layer, an attention layer, and a fully connected output layer. The model training module is used to train the offline LSTM-attention fusion network using the health dataset to obtain the optimal prediction model; The oil temperature prediction module is used to collect monitoring data of the wind turbine gearbox in real time, perform data cleaning, normalization and feature selection, and then input the data into the optimal prediction model to output the predicted value of the gearbox oil temperature. The fault warning module is used to calculate the residual between the predicted value and the corresponding actual oil temperature value, set a warning threshold based on the residual, and determine whether the gearbox has malfunctioned based on the comparison result between the residual and the warning threshold.
[0015] Another aspect of this disclosure provides an electronic device, comprising: At least one processor; and, A memory communicatively connected to the at least one processor is used to store one or more programs that, when executed by the at least one processor, enable the at least one processor to implement the LSTM-attention fusion gearbox oil temperature prediction and early warning method described above.
[0016] Another aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the LSTM-attention fusion gearbox oil temperature prediction and early warning method described above.
[0017] This disclosure presents an LSTM-attention fusion method and system for predicting and warning gearbox oil temperature. By constructing a hierarchical collaborative architecture, the LSTM layer is specifically responsible for modeling complex temporal dynamics that conform to physical mechanisms, effectively improving the prediction accuracy under varying operating conditions. The attention layer adaptively focuses on key early fault features, enhancing the model's sensitivity to weak signals and its noise resistance. Finally, the model provides intuitive decision-making basis through the allocation of attention weights, significantly improving the interpretability of the warning system, thereby solving the problems of insufficient accuracy, poor robustness, and difficulty in interpretation of existing methods. Attached Figure Description
[0018] Figure 1This is a flowchart illustrating an LSTM-attention fusion gearbox oil temperature prediction and early warning method according to an embodiment of this disclosure; Figure 2 This is a schematic diagram of the structure of an LSTM-attention fusion gearbox oil temperature prediction and early warning system according to another embodiment of this disclosure; Figure 3 This is a schematic diagram of the structure of an electronic device according to another embodiment of the present disclosure. Detailed Implementation
[0019] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this disclosure.
[0020] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.
[0021] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.
[0022] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of this disclosure. As used in this disclosure, the term "and / or" includes all combinations of any and more of the associated listed items.
[0023] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of exemplary embodiments, and the modules or processes in the drawings are not necessarily necessary for implementing this disclosure, and therefore cannot be used to limit the scope of protection of this disclosure.
[0024] like Figure 1As shown, one embodiment of this disclosure provides an LSTM-attention fusion gearbox oil temperature prediction and early warning method, including: Step S1: Collect SCADA data of the wind turbine gearbox under normal operating conditions, and perform data cleaning, normalization and feature selection on the SCADA data to obtain a health dataset.
[0025] Specifically, SCADA data of the wind turbine gearbox is collected, including timestamps, gearbox oil temperature, active power, generator speed, front and rear bearing temperatures, coolant temperature, ambient temperature, and nacelle temperature. The data is then preprocessed, including: (1) Data cleaning: for imbalanced data, a method with fewer labels is used to establish a normal behavior model; outliers are removed using the percentile method (interquartile range method), and isolated points are marked using the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm; for outliers, obviously erroneous records (such as constant values and values exceeding physical limits) are deleted and replaced with moving average or interpolation (linear, spline). (2) Normalization: the data is scaled to a uniform scale. (3) Feature selection: Using random forest and correlation coefficient methods, different features are arranged according to importance, and common important features (i.e., features related to oil temperature prediction) are selected. These may include: average wind speed per unit time, average power per unit time, gear pump pressure, gear inlet pressure, ambient temperature, engine room temperature, low-speed shaft bearing temperature, generator stator temperature, generator bearing temperature, gear oil inlet temperature, gearbox bearing temperature, generator speed, gearbox speed, main shaft speed, etc. This ultimately forms a healthy data sample set.
[0026] Step S2: Construct an offline LSTM-attention fusion network that sequentially includes an input layer, a first LSTM layer, a second LSTM layer, an attention layer, and a fully connected output layer.
[0027] Specifically, a sequence network is constructed, consisting of an input layer, two LSTM layers, an attention layer, and a fully connected output layer. L2 regularization and dropout layers are added to prevent overfitting. The first LSTM layer passes the complete sequence to the second LSTM layer, and the second LSTM layer also returns the complete sequence. An attention mechanism is introduced after the second LSTM layer to output the final encoding. The output layer is a single neuron used to output the normalized predicted future oil temperature of the gearbox.
[0028] Step S3: Train the offline LSTM-attention fusion network using the health dataset to obtain the optimal prediction model.
[0029] Specifically, the data in the health dataset is divided into training, validation, and test sets according to chronological order. Using the training set data, with mean squared error as the loss function, an optimizer is employed to train the constructed offline LSTM-attention fusion network. During training, the validation set is used to monitor model performance, and the model weights with the minimum loss on the validation set are saved to obtain the optimal prediction model. The detailed process is as follows: The cleaned, normalized, and feature-selected health data is used as input. The data is constructed into a three-dimensional tensor format (number of samples, time step, number of features) using a sliding window approach. For each sample, the LSTM process is as follows: Initialize the memory and hidden states as zero vectors, and input a sample of shape (1, a, b) (assuming a batch size of 1). During training, each sample contains a time steps, and each time step has b features. For each sample, the LSTM processes those a time steps sequentially. At each time step t (t = 0, 1, 2, …, a), the LSTM receives the input... (b-dimensional), then update the cell state. and hidden state The updated formula is as follows:
[0030]
[0031]
[0032] In memory Before being officially recognized, one must first become a candidate memory element. Its calculation uses the tanh function as the activation function; Weight matrix, Bias. Input gate Forgotten Gate Output gate The calculations are shown in the following formulas:
[0033]
[0034]
[0035] The calculations for the three gates are roughly the same, but each uses a specific weight matrix. With bias .
[0036] The first LSTM layer takes b-dimensional input and outputs a 64-dimensional hidden state for each of the a time steps, and the cell state is propagated across a time steps. The first LSTM layer outputs the hidden states for all time steps (time steps 0 to a-1, each 64-dimensional), and passes the hidden state of the last time step to the second LSTM layer. The second LSTM layer receives the 64-dimensional hidden state of the last time step from the first LSTM layer as input, and then outputs the 64-dimensional hidden states for all time steps, using these as input to the attention module.
[0037] Further assume the input hidden state sequence is , among them For time steps ( =36), For model dimensions ( = =16).
[0038] The attention layer uses scaled dot product attention. For a single attention head, the query, key, and value are first computed: , , .in, , , , These are the dimensions of a single head.
[0039] The formula for calculating scaled dot product attention is as follows:
[0040] In the formula, , , These are query, key, and value matrices, respectively. Represents the transpose of a matrix. denoted as the dimension of the key matrix.
[0041] For h heads (assuming h=8), each head has an independent projection matrix. The query, key, and value matrices of the i-th head (i=1, …, h) are respectively... , , .
[0042] Attention is calculated independently for each head, as shown in the following formula:
[0043] The outputs of each header are concatenated:
[0044] in, It is the output projection matrix.
[0045] The complete formula for self-attention is:
[0046] After assigning attention weights to all hidden states, the context vector is summed and connected to the output layer (i.e., the fully connected layer). The output layer uses a linear activation function to obtain the oil temperature prediction. The entire sequence (a time steps) in the sample is used to predict the oil temperature at the (a+1)th time step. The input of each sample is the features of a time steps, and the output is the oil temperature at the (a+1)th time step.
[0047] Step S4: Collect real-time monitoring data of the wind turbine gearbox, perform data cleaning, normalization and feature selection, input the data into the optimal prediction model, and output the predicted value of gearbox oil temperature.
[0048] Specifically, gearbox monitoring data collected in real time and preprocessed using the same steps as in step S1 is input into the optimal prediction model trained in step S3 to obtain predicted gearbox oil temperature values T' for one or more future time steps. Then, the residual between the predicted oil temperature T' and the actual oil temperature t' is calculated. This residual is used as a starting point to set a warning threshold. To prevent extreme values in the residual and reduce the possibility of false alarms, a moving window moving average method is used to smooth the residual. The warning threshold is then set based on the mean and standard deviation of the processed residual. The upper control limit (UCL) and lower control limit (LCL) of the residual are calculated using the EWMA (Exponential Weighted Moving Average) algorithm. These residual control limits are used as the warning threshold to achieve dynamic warning threshold setting based on the residual. The specific calculation method is shown in the following formula:
[0049] In the formula, , These are the upper and lower limits for residual control, respectively. , These are the mean and standard deviation of the residuals, respectively. , The preset constants determine the positional factors of the UCL and LCL; For time.
[0050] Step S5: Calculate the residual between the predicted value and the corresponding actual oil temperature value, set a warning threshold based on the residual, and determine whether the gearbox has malfunctioned based on the comparison result between the residual and the warning threshold.
[0051] Specifically, the residual is compared with the warning threshold. If the residual is greater than the upper limit of residual control (UCL) or less than the lower limit of control (LCL), the system generates a warning, indicating that a gearbox failure may occur at the predicted time point.
[0052] This disclosure discloses an LSTM-attention fusion gearbox oil temperature prediction and early warning method. By constructing a hierarchical collaborative architecture, the LSTM layer is specifically responsible for modeling complex temporal dynamics that conform to physical mechanisms, effectively improving the prediction accuracy under varying operating conditions. The attention layer adaptively focuses on key early fault features, enhancing the model's sensitivity to weak signals and its noise resistance. Finally, the model provides intuitive decision-making basis through the allocation of attention weights, significantly improving the interpretability of the early warning system, thereby solving the problems of insufficient accuracy, poor robustness, and difficulty in interpretation of existing methods.
[0053] like Figure 2 As shown, another embodiment of this disclosure provides an LSTM-attention fusion gearbox oil temperature prediction and early warning system, the system comprising: The data acquisition and processing module 210 is used to acquire SCADA data of the wind turbine gearbox under normal operating conditions, and to perform data cleaning, normalization and feature selection on the SCADA data to obtain a health dataset. The model building module 220 is used to build an offline LSTM-attention fusion network that sequentially includes an input layer, a first LSTM layer, a second LSTM layer, an attention layer, and a fully connected output layer; The model training module 230 is used to train the offline LSTM-attention fusion network using the health dataset to obtain the optimal prediction model; The oil temperature prediction module 240 is used to collect monitoring data of the wind turbine gearbox in real time, perform data cleaning, normalization processing and feature selection, and then input the data into the optimal prediction model to output the predicted value of the gearbox oil temperature. The fault warning module 250 is used to calculate the residual between the predicted value and the corresponding actual oil temperature value, set a warning threshold based on the residual, and determine whether the gearbox has malfunctioned based on the comparison result between the residual and the warning threshold.
[0054] Specifically, an LSTM-attention fusion gearbox oil temperature prediction and early warning system according to an embodiment of this disclosure is used to implement the LSTM-attention fusion gearbox oil temperature prediction and early warning method described in the above embodiments. The specific implementation process has been described in detail in the above embodiments and will not be repeated here.
[0055] like Figure 3 As shown, another embodiment of this disclosure provides an electronic device, including: At least one processor 301; and a memory 302 communicatively connected to the at least one processor 301 for storing one or more programs that, when executed by the at least one processor 301, enable the at least one processor 301 to implement the LSTM-attention fusion gearbox oil temperature prediction and early warning method described above.
[0056] The memory 302 and processor 301 are connected via a bus, which can include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 301 and memory 302 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 301 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 301.
[0057] Processor 301 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 302 can be used to store data used by processor 301 during operation.
[0058] Another embodiment of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the LSTM-attention fusion gearbox oil temperature prediction and early warning method described above.
[0059] The computer-readable storage medium may be included in the systems or electronic devices disclosed herein, or it may exist independently.
[0060] Computer-readable storage media can be any tangible medium that contains or stores a program, and can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, optical fibers, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.
[0061] Computer-readable storage media may also include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code, specific examples of which include, but are not limited to, electromagnetic signals, optical signals, or any suitable combination thereof.
[0062] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of this disclosure, and this disclosure is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this disclosure, and these modifications and improvements are also considered to be within the scope of protection of this disclosure.
Claims
1. An LSTM-attention fusion gearbox oil temperature prediction and early warning method, characterized in that, The method includes: Collect SCADA data of wind turbine gearbox under normal operating conditions, and perform data cleaning, normalization and feature selection on the SCADA data to obtain a health dataset; An offline LSTM-attention fusion network is constructed, consisting of an input layer, a first LSTM layer, a second LSTM layer, an attention layer, and a fully connected output layer. The offline LSTM-attention fusion network is trained using the health dataset to obtain the optimal prediction model; The monitoring data of the wind turbine gearbox is collected in real time, and after data cleaning, normalization and feature selection, it is input into the optimal prediction model to output the predicted value of gearbox oil temperature. Calculate the residual between the predicted value and the corresponding actual oil temperature value, set a warning threshold based on the residual, and determine whether the gearbox has malfunctioned based on the comparison result between the residual and the warning threshold.
2. The LSTM-attention fusion gearbox oil temperature prediction and early warning method according to claim 1, characterized in that, The data cleaning includes: Outliers are removed using percentiles, isolated points are marked using the DBSCAN algorithm, and outliers are deleted and corrected using moving averages or interpolation.
3. The LSTM-attention fusion gearbox oil temperature prediction and early warning method according to claim 1, characterized in that, The feature selection includes: The random forest algorithm is used to select relevant features for oil temperature prediction from the SCADA data; wherein, The relevant characteristics include at least one or more of the following: wind speed, power, ambient temperature, nacelle temperature, bearing temperature, gear oil inlet temperature, pressure, and rotational speed.
4. The LSTM-attention fusion gearbox oil temperature prediction and early warning method according to claim 1, characterized in that, The first LSTM layer receives the input sequence and outputs the hidden state sequence of all time steps. The second LSTM layer receives the hidden state of the last time step of the first LSTM layer as the initial input and outputs the hidden state sequence of all time steps to the attention layer. The attention layer is a multi-head self-attention layer, used to weight and integrate the hidden state sequence output by the second LSTM layer, and output the context vector to the fully connected output layer.
5. The LSTM-attention fusion gearbox oil temperature prediction and early warning method according to claim 4, characterized in that, The attention layer employs a scaled dot product attention mechanism, as shown in the following equation: In the formula, , , These are query, key, and value matrices, respectively. Represents the transpose of a matrix. denoted as the dimension of the key matrix.
6. The LSTM-attention fusion gearbox oil temperature prediction and early warning method according to claim 1, characterized in that, The setting of the early warning threshold based on the residual includes: The residuals are smoothed by a moving window moving average to obtain the mean and standard deviation of the residuals; Based on the mean and standard deviation of the residuals, an exponentially weighted moving average algorithm is used to calculate the dynamic upper limit and lower limit of residual control, and the upper limit and lower limit of residual control are used as early warning thresholds.
7. The LSTM-attention fusion gearbox oil temperature prediction and early warning method according to claim 6, characterized in that, The calculation of the upper limit and lower limit of the residual control for dynamic residuals is shown in the following formula: In the formula, , These are the upper and lower limits for residual control, respectively. , These are the mean and standard deviation of the residuals, respectively. , As a preset constant, For time.
8. An LSTM-attention fusion gearbox oil temperature prediction and early warning system, characterized in that, The system includes: The data acquisition and processing module is used to acquire SCADA data of the wind turbine gearbox under normal operating conditions, and to perform data cleaning, normalization and feature selection on the SCADA data to obtain a health dataset. The model building module is used to construct an offline LSTM-attention fusion network that sequentially includes an input layer, a first LSTM layer, a second LSTM layer, an attention layer, and a fully connected output layer. The model training module is used to train the offline LSTM-attention fusion network using the health dataset to obtain the optimal prediction model; The oil temperature prediction module is used to collect monitoring data of the wind turbine gearbox in real time, perform data cleaning, normalization and feature selection, and then input the data into the optimal prediction model to output the predicted value of the gearbox oil temperature. The fault warning module is used to calculate the residual between the predicted value and the corresponding actual oil temperature value, set a warning threshold based on the residual, and determine whether the gearbox has malfunctioned based on the comparison result between the residual and the warning threshold.
9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor is used to store one or more programs that, when executed by the at least one processor, enable the at least one processor to implement the LSTM-attention fusion gearbox oil temperature prediction and early warning method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the LSTM-attention fusion gearbox oil temperature prediction and early warning method as described in any one of claims 1 to 7.
