Method and device for predicting residual life of unmanned aerial vehicle motor

Abstract

The embodiment of the application provides a kind of unmanned aerial vehicle motor remaining life prediction method and device, its method includes: obtaining the early vibration signal of target domain unmanned aerial vehicle motor, and using the early vibration signal of target domain unmanned aerial vehicle motor and the support vector regression SVR model trained in advance, obtain the early health label series of target domain unmanned aerial vehicle motor;Using the early health label series of target domain unmanned aerial vehicle motor, fine-tuning training is carried out to the source domain LSTM prediction model trained in advance, to obtain the trained target domain LSTM prediction model;Obtain the real-time health label sequence of target domain unmanned aerial vehicle motor, and by inputting the real-time health label series of target domain unmanned aerial vehicle motor into the trained target domain LSTM prediction model, obtain the target domain unmanned aerial vehicle motor remaining life prediction result.

Description

Technical Field

[0001] This application relates to the field of remaining life prediction technology, and in particular to a method and apparatus for predicting the remaining life of a drone motor. Background Technology

[0002] As modern intelligent equipment, drones (UAVs) are characterized by their mobility, adaptability, and low operating costs, making them essential tools in fields such as aerial surveying, logistics, and agricultural plant protection. Their complex working environments and high-intensity operations place stringent demands on system reliability. The UAV power system is the core of its safe and reliable operation, and the motor, as a key actuator, directly contributes to thrust loss, increased vibration, and even crashes due to performance degradation. Therefore, accurate health assessment and lifespan prediction of motors are of great significance. Existing motor health management solutions mainly fall into two categories:

[0003] One type is based on physical models, which relies on establishing accurate motor dynamics models or bearing damage evolution models. However, the complex electromagnetic-mechanical coupling within the motor and the uncertainty of the operating environment make it extremely difficult and costly to construct high-fidelity physical models, and they also have poor generalization ability, making them difficult to widely apply in practical engineering.

[0004] One type is data-driven, which is the current mainstream technology direction. It extracts features from collected data and then uses statistics or machine learning to obtain health indicators and predict lifespan. However, current data-driven drone health assessment and lifespan prediction methods suffer from several long-standing core pain points in engineering and large-scale implementation: First, insufficient feature utilization and delayed early warning. Many solutions rely solely on single or limited features such as the root mean square (RMS) value of vibration signals for simple threshold alarms. While the RMS value can effectively reflect overall vibration energy, it is insensitive to early, minor faults (such as initial pitting of bearings), leading to delayed early warning and susceptibility to changes in operating conditions, resulting in a high false alarm rate. Second, strong data dependence. High-precision prediction models based on deep learning heavily rely on a large amount of full lifecycle data, especially labeled data containing fault states, for training. For newly produced models or new batches of motors, the lack of historical fault data makes it impossible to build models or achieve the required prediction accuracy, resulting in high deployment costs and long cycles for PHM technology. Third, poor model generalization ability. Models trained under specific models and operating conditions may learn data distributions that differ from those of new products (domain drift). When existing models are directly applied to new devices, their performance degrades significantly, making it difficult to achieve the goal of "develop once, deploy everywhere" for large-scale application.

[0005] In summary, existing drone motor health management solutions have two major bottlenecks: First, reliance on a single feature leads to insufficient early fault warning capabilities and a high false alarm rate; second, data-driven models heavily depend on the full lifecycle data of specific devices, failing to address the challenges of data scarcity for new devices and model generalization. Summary of the Invention

[0006] In view of this, embodiments of this application provide a method for predicting the remaining lifespan of a drone motor. One or more embodiments of this application also relate to a device for predicting the remaining lifespan of a drone motor, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0007] According to a first aspect of the embodiments of this application, a method for predicting the remaining lifespan of a drone motor is provided, comprising:

[0008] The early vibration signal of the UAV motor in the target domain is obtained, and the early health label series of the UAV motor in the target domain is obtained by using the early vibration signal of the UAV motor in the target domain and the pre-trained support vector regression (SVR) model.

[0009] The pre-trained source domain LSTM prediction model is fine-tuned using the early health label series of the target domain UAV motor to obtain the trained target domain LSTM prediction model.

[0010] The real-time health tag sequence of the target domain UAV motor is obtained, and the remaining lifespan prediction result of the target domain UAV motor is obtained by inputting the real-time health tag sequence of the target domain UAV motor into the trained target domain LSTM prediction model.

[0011] Preferably, the pre-trained SVR model includes:

[0012] The original vibration signal of the UAV motor in the source domain is obtained, and the original vibration signal is preprocessed and framed sequentially to obtain the framed signal of each sample. Multiple time-domain features are calculated on the framed signal of each sample to obtain the multidimensional feature vector of each sample. The multidimensional feature vectors of all samples are used to form a multidimensional feature matrix.

[0013] Obtain the health and failure benchmarks of the source domain UAV motor, and use the health and failure benchmarks to generate a health label vector for each sample;

[0014] An SVR model is constructed by using the multidimensional feature vector of each sample as the input of the SVR model and the health label vector of each sample as the output of the SVR model. The SVR model is then trained to obtain a trained SVR model.

[0015] Preferably, the step of obtaining the early health label series of the UAV motor in the target domain using the early vibration signal of the UAV motor in the target domain and the pre-trained support vector regression (SVR) model includes:

[0016] By sequentially preprocessing and framing the early vibration signal of the UAV motor in the target domain, the framing signal of each sample of the UAV motor in the target domain is obtained. Furthermore, by performing multiple time-domain feature calculations on the framing signal of each sample of the UAV motor in the target domain, the multidimensional feature vector of each sample of the UAV motor in the target domain is obtained.

[0017] By inputting the multidimensional feature vector of each sample of the target domain UAV motor into the pre-trained support vector regression (SVR) model, the early health label series of the target domain UAV motor is obtained.

[0018] Preferably, the pre-trained source domain LSTM prediction model includes:

[0019] The original vibration signals of the Yuanyu UAV motor throughout its entire life cycle are obtained, and the original vibration signals of the Yuanyu UAV motor throughout its entire life cycle and the pre-trained support vector regression (SVR) model are used to obtain a series of health labels for the Yuanyu UAV motor throughout its entire life cycle.

[0020] Using the source domain UAV motor full life cycle health tag series, a training sample pair is constructed, which includes a health tag sequence of length L as input and the next health tag as output.

[0021] Construct a source domain LSTM prediction model and train the source domain LSTM prediction model using the training samples to obtain a trained source domain LSTM prediction model.

[0022] Preferably, the LSTM prediction model consists of an input layer, a first LSTM layer, a first Dropout layer, a second LSTM layer, a second Dropout layer, and an output layer, from bottom to top.

[0023] Preferably, the step of fine-tuning the pre-trained source domain LSTM prediction model using the early health label series of the target domain UAV motors to obtain the trained target domain LSTM prediction model includes:

[0024] Using the previous health tag series of the target domain UAV motor, construct micro-training sample pairs containing a health tag sequence of length L as input and the next health tag as output of the input sequence;

[0025] The pre-trained source domain LSTM prediction model is fine-tuned using the micro-training samples to obtain a trained target domain LSTM prediction model.

[0026] Preferably, the step of fine-tuning the pre-trained source domain LSTM prediction model using the micro-training samples to obtain a trained target domain LSTM prediction model includes:

[0027] During the fine-tuning training of the pre-trained source domain LSTM prediction model using the micro-training samples, the network weights of the first LSTM layer and the first Dropout layer in the pre-trained source domain LSTM prediction model are frozen, while the network weights of the second LSTM layer, the second Dropout layer, and the output layer in the pre-trained source domain LSTM prediction model are fine-tuned using a low learning rate, thus obtaining the trained target domain LSTM prediction model.

[0028] According to a second aspect of the embodiments of this application, a device for predicting the remaining lifespan of a drone motor is provided, comprising:

[0029] The acquisition module is configured to acquire the early vibration signal of the target domain UAV motor and use the early vibration signal of the target domain UAV motor and a pre-trained support vector regression (SVR) model to obtain the early health label series of the target domain UAV motor.

[0030] The model training module is configured to fine-tune the pre-trained source domain LSTM prediction model using the early health label series of the target domain UAV motors to obtain the trained target domain LSTM prediction model.

[0031] The prediction module is configured to acquire the real-time health tag sequence of the target domain UAV motors, and obtain the remaining life prediction result of the target domain UAV motors by inputting the real-time health tag sequence of the target domain UAV motors into the trained target domain LSTM prediction model.

[0032] According to a third aspect of the embodiments of this application, a computing device is provided, comprising:

[0033] Memory and processor;

[0034] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement any of the steps of the UAV motor remaining life prediction method.

[0035] According to a fourth aspect of the present application, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of any one of the unmanned aerial vehicle motor remaining life prediction methods.

[0036] According to a fifth aspect of the present application, a computer program is provided, wherein when the computer program is executed in a computer, the computer is instructed to perform the steps of the above-described method for predicting the remaining lifespan of a drone motor.

[0037] The UAV motor remaining life prediction scheme provided in this application embodiment obtains the early vibration signal of the UAV motor in the target domain, and uses the early vibration signal of the UAV motor in the target domain and a pre-trained support vector regression (SVR) model to obtain the early health label series of the UAV motor in the target domain; it then uses the early health label series of the UAV motor in the target domain to fine-tune the pre-trained source domain LSTM prediction model to obtain a trained target domain LSTM prediction model; finally, it obtains the real-time health label sequence of the UAV motor in the target domain, and inputs the real-time health label series of the UAV motor in the target domain into the trained target domain LSTM prediction model to obtain the remaining life prediction result of the UAV motor in the target domain. By combining multi-dimensional feature extraction and fusion, health indicator generation, and transfer learning prediction, it effectively solves the problems of data scarcity for new equipment and model generalization, while significantly improving the fault warning capability and life prediction accuracy of UAV motors, reducing maintenance costs, and meeting the health management needs of UAVs in large-scale applications. Attached Figure Description

[0038] Figure 1 This is a flowchart of a method for predicting the remaining lifespan of a drone motor according to an embodiment of this application;

[0039] Figure 2 This is a flowchart illustrating a method for assessing and predicting the health of a drone motor based on transfer learning and multi-feature fusion, as described in this embodiment.

[0040] Figure 3 This is a schematic diagram of a drone motor remaining life prediction device provided in one embodiment of this application;

[0041] Figure 4 This is a structural block diagram of a computing device provided in one embodiment of this application. Detailed Implementation

[0042] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0043] This application proposes a method for health assessment and prediction of UAV motors based on transfer learning and multi-feature fusion. Addressing the shortcomings of existing health monitoring technologies, it constructs a multi-dimensional feature vector matrix, generates health labels, trains an SVR health indicator generator, builds a source domain LSTM prediction model, and employs a model transfer learning strategy to achieve rapid and efficient deployment of the model on new equipment and prediction of the remaining service life in the target domain. By combining multi-dimensional feature extraction and fusion, health indicator generation, and transfer learning prediction, this method effectively solves the problems of data scarcity and model generalization in new equipment. This application can significantly improve the fault warning capability and life prediction accuracy of UAV motors, reduce maintenance costs, and meet the health management needs of UAVs in large-scale applications.

[0044] Figure 1 A flowchart of a method for predicting the remaining lifespan of a drone motor according to an embodiment of this application is shown, which specifically includes the following steps.

[0045] Step S101: Obtain the early vibration signal of the UAV motor in the target domain, and use the early vibration signal of the UAV motor in the target domain and the pre-trained support vector regression (SVR) model to obtain the early health label series of the UAV motor in the target domain.

[0046] In one specific embodiment of this application, the pre-trained SVR model includes:

[0047] The original vibration signal of the UAV motor in the source domain is obtained, and the original vibration signal is preprocessed and framed sequentially to obtain the framed signal of each sample. Multiple time-domain features are calculated on the framed signal of each sample to obtain the multidimensional feature vector of each sample. The multidimensional feature vectors of all samples are used to form a multidimensional feature matrix.

[0048] Obtain the health and failure benchmarks of the source domain UAV motor, and use the health and failure benchmarks to generate a health label vector for each sample;

[0049] An SVR model is constructed by using the multidimensional feature vector of each sample as the input of the SVR model and the health label vector of each sample as the output of the SVR model. The SVR model is then trained to obtain a trained SVR model.

[0050] In one specific embodiment of this application, the step of obtaining the early health label series of the UAV motor in the target domain using the early vibration signal of the UAV motor in the target domain and the pre-trained support vector regression (SVR) model includes:

[0051] By sequentially preprocessing and framing the early vibration signal of the UAV motor in the target domain, the framing signal of each sample of the UAV motor in the target domain is obtained. Furthermore, by performing multiple time-domain feature calculations on the framing signal of each sample of the UAV motor in the target domain, the multidimensional feature vector of each sample of the UAV motor in the target domain is obtained.

[0052] By inputting the multidimensional feature vector of each sample of the target domain UAV motor into the pre-trained support vector regression (SVR) model, the early health label series of the target domain UAV motor is obtained.

[0053] Step S102: Use the early health label series of the target domain UAV motor to fine-tune the pre-trained source domain LSTM prediction model to obtain the trained target domain LSTM prediction model.

[0054] In one specific embodiment of this application, the pre-trained source domain LSTM prediction model includes:

[0055] The original vibration signals of the Yuanyu UAV motor throughout its entire life cycle are obtained, and the original vibration signals of the Yuanyu UAV motor throughout its entire life cycle and the pre-trained support vector regression (SVR) model are used to obtain a series of health labels for the Yuanyu UAV motor throughout its entire life cycle.

[0056] Using the source domain UAV motor full life cycle health tag series, a training sample pair is constructed, which includes a health tag sequence of length L as input and the next health tag as output.

[0057] Construct a source domain LSTM prediction model and train the source domain LSTM prediction model using the training samples to obtain a trained source domain LSTM prediction model.

[0058] In one specific embodiment of this application, the LSTM prediction model consists of an input layer, a first LSTM layer, a first Dropout layer, a second LSTM layer, a second Dropout layer, and an output layer, from bottom to top.

[0059] In one specific embodiment of this application, the step of fine-tuning the pre-trained source domain LSTM prediction model using the early health label series of the target domain UAV motors to obtain the trained target domain LSTM prediction model includes:

[0060] Using the previous health tag series of the target domain UAV motor, construct micro-training sample pairs containing a health tag sequence of length L as input and the next health tag as output of the input sequence;

[0061] The pre-trained source domain LSTM prediction model is fine-tuned using the micro-training samples to obtain a trained target domain LSTM prediction model.

[0062] In one specific embodiment of this application, the step of fine-tuning the pre-trained source domain LSTM prediction model using the micro-training samples to obtain a trained target domain LSTM prediction model includes:

[0063] During the fine-tuning training of the pre-trained source domain LSTM prediction model using the micro-training samples, the network weights of the first LSTM layer and the first Dropout layer in the pre-trained source domain LSTM prediction model are frozen, while the network weights of the second LSTM layer, the second Dropout layer, and the output layer in the pre-trained source domain LSTM prediction model are fine-tuned using a low learning rate, thus obtaining the trained target domain LSTM prediction model.

[0064] Step S103: Obtain the real-time health tag sequence of the target domain UAV motor, and input the real-time health tag sequence of the target domain UAV motor into the trained target domain LSTM prediction model to obtain the prediction result of the remaining lifespan of the target domain UAV motor.

[0065] The UAV motor remaining life prediction scheme provided in this application combines multi-dimensional feature extraction and fusion, health indicator generation, and transfer learning prediction to effectively solve the problems of scarce data for new equipment and model generalization. At the same time, it can significantly improve the fault warning capability and life prediction accuracy of UAV motors, reduce maintenance costs, and meet the health management needs of UAVs in large-scale applications.

[0066] Figure 2 This application illustrates a method for predicting the remaining lifespan of a drone motor, the core process of which includes two main stages: health indicator construction and migration prediction. The method includes the following steps:

[0067] S1: Multidimensional feature extraction and dataset construction. Collect the vibration signal of the UAV motor, and after preprocessing, extract the time-domain features including kurtosis, peak factor, root mean square value and waveform factor to construct a multidimensional feature matrix.

[0068] Raw vibration signals are acquired using an accelerometer mounted on the motor housing. After DC component removal and high-speed filtering for noise reduction, the acquired raw vibration signals are segmented into frames. For each frame, multiple time-domain statistical features are calculated in parallel, including: kurtosis (for sensitive detection of early impact faults); crease factor (to evaluate the peak characteristics of the signal); root mean square (RMS) value (characterizing the average energy level of the signal); and shape factor (reflecting the difference between the signal waveform and a normal distribution). These features collectively constitute a multidimensional feature vector describing the current state of the motor. A multidimensional feature matrix is ​​constructed by combining the feature vectors of all samples. .

[0069] In the specific implementation process, in step S1, the vibration data of the UAV motor is collected and preprocessed, and key time-domain features are extracted to construct a multi-dimensional feature matrix X.

[0070] S11: Sensor Deployment and Signal Acquisition. An ICP-type accelerometer (such as the PCB Piezotronics 352C33) with a frequency response range of at least 5kHz was selected and securely mounted above the radial bearing seat on the motor housing using a magnetic base, ensuring the sensor axis is aligned with the radial vibration direction of the motor. A 24-bit high-precision data acquisition card (such as the NI USB-4431) was used for signal acquisition, with a sampling frequency set to 25.6kHz to satisfy the Nyquist sampling theorem and effectively capture the high-frequency impact components caused by bearing failure. Complete vibration data of the motor from start-up to shutdown was continuously recorded during each flight mission.

[0071] S12: Signal Preprocessing Flow. The raw vibration signal is first subjected to DC component removal to eliminate signal baseline drift. Then, a 4th-order Butterworth high-pass digital filter is used with a cutoff frequency set to 10Hz to eliminate low-frequency environmental noise interference. The preprocessed continuous signal is divided into non-overlapping analysis frames, each with a fixed length of 4096 data points, corresponding to a time length of approximately 0.16 seconds, ensuring both frequency resolution and computational efficiency.

[0072] S13: Time-domain feature calculation. For each frame of signal. (in , The system computes four core time-domain features in parallel: kurtosis, root mean square (RMS) value, peak factor, and waveform factor. Kurtosis is highly sensitive to impact-related faults and can provide effective early warnings. The RMS value, as the most stable representation of vibration energy, is a key indicator for assessing the overall degradation trend. The peak factor reflects the ratio of peak value to RMS value in the signal and has a good indicative effect on local defects. The waveform factor describes the degree of deviation of the signal waveform from the normal distribution.

[0073] S14: Feature Matrix Construction. The four feature values ​​of each sample are combined in a fixed order (kurtosis, root mean square value, peak factor, waveform factor) to form a 4-dimensional feature vector. The feature vectors of all samples together constitute the feature matrix. ,in The total number of samples.

[0074] S2: Health label generation, the root mean square value is subjected to minimum-maximum normalization to generate a health label vector in the range of [0, 1];

[0075] In the health label generation process, the root mean square (RMS) value, the most sensitive feature parameter, is selected as the measurement indicator. Further, the RMS value calculated for each sample in step S1 is subjected to minimum-maximum normalization. The maximum value required for normalization is... Taken from the stable RMS average value during the initial healthy operation of the motor, minimum value Taken from known or preset motor failure thresholds. (Using the formula...) Calculate and generate a health label vector in the interval [0, 1]. , where 1 represents healthy and 0 represents invalid.

[0076] In the specific implementation process, in step S2, by determining and The RMS value is normalized and used as a health label vector.

[0077] S21: Normalization baseline determination. Based on the complete historical data of source domain motor A, determine the normalization parameters:

[0078] : Select all RMS values ​​during the stable operation phase (excluding start-up and shutdown transients) of the first 100 flight missions in the initial stage of motor use, and calculate their 95th percentile as the health benchmark.

[0079] : Determine the failure threshold through accelerated life testing, or select the RMS value corresponding to the recognized functional failure of this model of motor.

[0080] S22: Label Generation and Quality Control. For each sample, the initial label is calculated using the formula:

[0081]

[0082] To prevent the influence of extreme values, the calculation results are cropped:

[0083] Ensure that all labels fall strictly within the range [0,1], where 1 represents complete health and 0 represents complete failure.

[0084] S3: Training the SVR health indicator generator involves standardizing the feature matrix, training it with a support vector regression model with a radial basis function kernel, and optimizing its hyperparameters through grid search and cross-validation to obtain an SVR model that maps multidimensional features to a single health indicator.

[0085] Multidimensional feature matrix and health tag vector The dataset was divided into training and testing sets. First, the feature matrix of the training feature set was standardized using the Z-score method to eliminate the influence of dimensions. Then, a Support Vector Regression (SVR) model was trained using the Radial Basis Function (RBF) as the kernel function. To obtain optimal performance, GridSearchCV and 5-fold cross-validation were used to test the key hyperparameter of SVR—the regularization parameter. Kernel coefficient Optimization is performed with the objective of minimizing the negative mean squared error. After training, the optimal SVR model is obtained and used as a health indicator (HI) generator. This model maps new multidimensional feature vectors to a fused, robust health indicator value, which serves as input for subsequent time-series predictions. Its performance can be validated on a test set.

[0086] In the specific implementation process, in step S3, the multidimensional feature vectors are Z-score standardized, a support vector regression model with radial basis function kernels is selected, and its hyperparameters are optimized using grid search combined with a 5-fold cross-validation system. Finally, an SVR health indicator generator that can accurately map multidimensional features to a single, robust health indicator is trained.

[0087] S31: Data Standardization. Before training the SVR model, the feature matrix must be standardized. Standardization is performed to transform the data for each feature dimension into a standard normal distribution with a mean of 0 and a standard deviation of 1. This step is crucial because the optimization objective function of the SVR model includes a regularization term that penalizes all feature weights equally. If the feature scales differ significantly (e.g., kurtosis values ​​are typically between 3 and 5, while RMS values ​​can reach 10-100), the model will over-rely on the larger-scale features, leading to unstable training and degraded prediction performance. Standardization eliminates the influence of feature scales, ensuring that each feature receives a fair weight during model training. Specifically, the Z-score standardization method is used.

[0088]

[0089] Where X is the characteristic matrix. The standardized feature matrix, This is the mean vector of each feature column. This is the standard deviation vector.

[0090] S32: SVR Model Selection and Kernel Function Construction. This application selects the Support Vector Regression (SVR) model because it exhibits excellent generalization ability in small-sample, high-dimensional scenarios and can effectively handle nonlinear relationships. Specifically, the Radial Basis Function (RBF) is chosen as the kernel function, which can map the original feature space to an infinite-dimensional space, thereby capturing the complex nonlinear relationship between features and health labels. The expression for the RBF kernel function is:

[0091]

[0092] in The kernel function coefficients control the influence range of a single training sample; The Euclidean distance between the two samples; Let be the RBF kernel function for two samples. Using the kernel function, SVR can construct a linear decision function in a high-dimensional feature space without explicitly computing the high-dimensional mapping.

[0093] S33: Hyperparameter Optimization Strategies and Implementation. The performance of SVR models is highly dependent on hyperparameter settings, mainly including:

[0094] Regularization parameters : Controlling the model's tolerance for error, balancing margin maximization with classification error;

[0095] kernel coefficient : Controlling the width of the RBF kernel affects the distribution of samples in the feature space.

[0096] To find the optimal combination of hyperparameters, this application employs a systematic method combining grid search and 5-fold cross-validation:

[0097] In 5-fold cross-validation, the training set is randomly divided into 5 mutually exclusive subsets. Each subset is approximately the same size. For each set of candidate hyperparameter combinations Five rounds of training and verification were conducted. In the first... wheel( In this context, the training and validation process is as follows:

[0098] Use except The model was trained using four subsets of data outside the parameter range. Training an SVR model This model is temporary and is only used for this round of evaluation;

[0099] model Applied to the validation set The predicted values ​​of all samples are obtained through the decision function of SVR. .

[0100]

[0101] Among them, the number of support vectors Lagrange multipliers and bias terms All of these were determined during the initial temporary training in step 1. Represents the upper boundary of the range, This represents the lower boundary of the range. In steps S31 and S32, the model (linear SVR) and kernel function (RBF) are defined. The goal of model training is to find the optimal weights, which mathematically involves solving a constrained convex optimization problem. Typically, the original constrained optimization problem is transformed into an unconstrained Lagrangian function. This is an extremum problem. Therefore, in each round of "temporary training" during cross-validation:

[0102] Solver (such as SMO) determines all Lagrange multipliers by solving a convex quadratic programming problem. The optimal value;

[0103] Based on these multipliers, support vectors are automatically identified, and their numbers are counted. .

[0104] Using the KKT conditions and the identified support vectors, the bias term is calculated. .

[0105] Ultimately, the support vector set The corresponding Lagrange multiplier pairs and bias terms These three factors together uniquely determine the current SVR model. This model can then be used for prediction.

[0106] Calculate the negative mean square error (NMSE) for this round:

[0107]

[0108] in For the validation set The number of samples, For the true value, These are predicted values.

[0109] The final performance score of this candidate parameter combination is the average of 5 rounds of validation:

[0110]

[0111] The grid search method systematically traverses the predefined hyperparameter space. and All 16 combinations. By comparing the combinations Value, selection The combination of hyperparameters with the largest value is selected as the optimal configuration. This process ensures that the selected parameters maintain stable and excellent performance across different subsets of the training data, effectively avoiding overfitting and enhancing the model's generalization ability.

[0112] S34: Model Validation. After obtaining the optimal combination of hyperparameters, the final model is validated on an independent test set. Evaluation metrics include the coefficient of determination. Root mean square error (RMSE):

[0113] Requirements for the test set RMSE ≤ 0.05, ensuring the model has sufficient prediction accuracy.

[0114] S35: The new feature vector after standardization in S31 When input into a pre-trained, optimal SVR model, the model will automatically execute its decision function. The output value of this model is a robust health indicator HI, which is between 0 and 1 and incorporates multi-dimensional information.

[0115] S4: Pre-training of the source domain LSTM prediction model. Select source domain motor A, use the model obtained in S3 to generate its health index time series, and use this to construct training samples. Train a deep network containing two layers of LSTM and Dropout layers to obtain the pre-trained model Model_A.

[0116] A motor with complete lifecycle data is selected as the source domain A. The SVR model trained with S3 is used to process all vibration data of motor A, generating its health index time series. Based on this sequence, using the sliding window method (window length...) =30) Construct supervised learning samples, that is, use past The HI value at one time step is used to predict the HI value at the next time step. A deep LSTM network is constructed, whose structure includes an input layer, a Dropout layer inserted between two LSTM layers to prevent overfitting, a Dropout layer, and a fully connected output layer. The mean squared error (MSE) is used as the loss function. The LSTM network is trained on the source domain data, and an early stopping mechanism is used to monitor the loss on the validation set. When the loss does not decrease for 15 consecutive rounds, the network is automatically stopped and the best model Model_A for motor A is saved.

[0117] In the specific implementation process, in step S4, based on the health index sequence of the source domain motor throughout its entire life cycle, a pre-trained model Model_A is obtained by constructing supervised learning samples and training a deep LSTM network, ultimately capable of learning and predicting the general degradation law of the motor.

[0118] S41: Source Domain HI Sequence Generation. The SVR health indicator generator trained in S3 is invoked to process all vibration data frames from the entire lifecycle of source domain motor A in chronological order, generating a one-dimensional health indicator time series. This sequence is the raw material for constructing supervised learning samples.

[0119] S42: Supervised Learning Sample Construction. LSTM is a sequence model, and the above HI sequence needs to be constructed into a form that it can learn, i.e., "predicting future values ​​from historical sequences." The specific process is as follows:

[0120] Define the learning task: The learning task of the model is defined as a single-step prediction problem, that is, predicting the HI value at the next moment based on a continuous historical HI value in the past.

[0121] Time window definition: Set a fixed history window length. This means that the model needs to observe the context of the past 30 time points when making predictions;

[0122] Constructing sample pairs: A large number of training samples are generated from the HI sequence using the sliding window method. Each sample consists of two parts:

[0123] Input sequence A length of The vector, i.e. ;

[0124] Output Labels The next HI value immediately following the input sequence, i.e. .

[0125] Data reshaping: all input sequences Combined into a dimension The three-dimensional tensor. This is exactly the input shape required by the LSTM network: (number of samples, time step, number of features), where there is only one feature at each time step, namely the HI value.

[0126] S43: Deep LSTM Network Architecture Design. Construct a deep learning model capable of learning long-term dependencies from HI sequences. Its specific structure is as follows:

[0127] Input layer: Receiver shape is The three-dimensional tensor, i.e., 32 samples, There are 1 time step, and each time step has 1 feature (HI value). The number of training samples can be changed according to needs.

[0128] First LSTM layer: Configured with 100 neurons and set return_sequences=True. This setting causes this layer to output not only the hidden state of the last time step, but also all the hidden states. The complete hidden state sequence at each time step transforms the input of shape (32, L, 1) into a richer temporal feature representation of shape (32, L, 100). This complete sequence output forms the basis for subsequent layers to perform deeper temporal analysis.

[0129] First Dropout layer: Immediately following the first LSTM layer, a portion of the 100-dimensional features output by the first LSTM layer is randomly discarded with a probability of 0.2 to prevent overfitting;

[0130] The second LSTM layer has 50 neurons and sets `return_sequences=False`. This layer receives the complete sequence (32, L, 100) output from the first LSTM layer. It also processes the sequence step-by-step, but its ultimate goal is to "condense" or "encode" the information from the entire sequence into a single, information-dense vector. Because `return_sequences=False`, it only returns the hidden state of the last time step. This 50-dimensional vector can be understood as the network's summary of the "degradation trend implied by the HI sequence over the past L time points".

[0131] Second Dropout layer: Dropout is applied to the 50-dimensional summary vector output by the second LSTM layer to further improve generalization ability;

[0132] Output layer: A fully connected layer containing one neuron that maps the 50-dimensional summary vector generated by the second LSTM layer to a single scalar value, namely the predicted HI value for the next time step. .

[0133] S44: Model Training Configuration. To efficiently train the model and obtain the optimal weights, the following configuration is required:

[0134] Model compilation: The Adam optimizer is used, with an initial learning rate set to 0.001. The loss function is specified as mean squared error (MSE), i.e. Because it is very sensitive to large errors in regression problems, it can effectively drive the model to learn the correct degradation trend.

[0135] Early stopping mechanism: Configure the callback function for the early stopping mechanism. Its parameters are set as follows: monitor the loss on the validation set, and the patience value indicates that if the validation set loss no longer decreases for 15 consecutive training epochs, training will automatically terminate, and the model weights will automatically revert to the state corresponding to the minimum validation set loss. This effectively prevents overfitting and saves training time.

[0136] S45: Model training and saving.

[0137] Input the training set data into the model for training. Set the batch size to 32 and the training epochs to 150 (the actual number of epochs is determined by the early stopping mechanism).

[0138] During training, the model iteratively adjusts its internal parameters (including weights and biases in the LSTM unit) with the goal of minimizing the MSE between the predicted HI value and the true HI value.

[0139] After training is completed, the model that performs best on the validation set and is saved by the early stopping mechanism is designated as the pre-trained source domain prediction model Model_A, which encapsulates the general temporal degradation rules learned from the source domain data.

[0140] S5: Target domain LSTM model transfer learning and fine-tuning. Select target domain motor B, load the Model_A, freeze its bottom network layer, use a small amount of early health index data of motor B, and fine-tune the unlocked layer with an extremely low learning rate to obtain the adapted model Model_B.

[0141] We use a target domain B consisting of different motor models with scarce data. We load a pre-trained model, Model_A. We employ a layer-freezing strategy for network fine-tuning: freezing the bottom LSTM layer (responsible for extracting general temporal features) and unlocking the top LSTM layer and output layer (responsible for specific decisions). Using a small amount of early HI data for motor B, we fine-tune the model with an extremely low learning rate. The unlocking layer is trained. This process allows the model to fine-tune its decision-making part to adapt to the specific degradation mode of motor B while retaining general knowledge of the source domain, ultimately resulting in a prediction model Model_B that is fully adapted to motor B.

[0142] In the target domain model transfer learning and fine-tuning process: a small amount of vibration data from the life cycle of the target domain motor B is collected to obtain the training samples required for fine-tuning; the pre-trained model model_A is called, and then a layer freezing strategy is implemented to use the data of motor B for model fine-tuning training; thus, a dedicated prediction model Model_B adapted to the target domain motor B is obtained.

[0143] In the specific implementation process, in step S5, the source domain pre-trained model is loaded and its underlying network weights are frozen. Then, the high-level network is fine-tuned using an extremely low learning rate. This allows the model to quickly adapt to the specific degradation mode of the target domain motor while retaining general temporal knowledge, thus enabling the construction of a high-precision prediction model Model_B using only a small amount of data.

[0144] S51: Target Domain Data Preparation. Vibration data of motor B in the target domain during the first 10% of its lifespan (or the first 50 flights) is collected, and its early HI sequence is generated using the same procedures as S1-S3. ,in The sequence length is much smaller than that of the source domain. Following this sequence, with the same window length as in S42... Construct the training samples needed for fine-tuning ;

[0145] S52: Model Loading and Freezing. The pre-trained model `model_A` saved in S45 is loaded, and then a layer freezing strategy is implemented. This is a core operation of transfer learning.

[0146] The Freeze Theory: Deep learning models learn different levels of features at different levels. In sequence models, the lower layers (such as the first LSTM layer) typically learn general, low-level temporal patterns (e.g., trends, cycles, short-term dependencies), which are universal across different layers. The higher layers (such as the second LSTM layer and the output layer) learn more complex combinations of decisions relevant to a specific task.

[0147] Operation: Set the trainable attribute of the first few layers of the model (specifically the first LSTM layer and the subsequent Dropout layers) to False. This means that during subsequent fine-tuning training, the weight parameters of these frozen layers will not be updated based on new target domain data. They are "locked" to the general temporal feature extraction state learned from the source domain.

[0148] S53: Fine-tuning the training configuration. Reconfigure the model to suit the fine-tuning task:

[0149] Recompiling: The key to successful fine-tuning is specifying an extremely low learning rate, which is set here to... The low learning rate ensures that the adjustment is very small when updating the weights that are not frozen, so that it can adapt to the characteristics of new data without drastically destroying or "forgetting" the valuable general knowledge learned from the source domain pre-training.

[0150] The Adam optimizer and mean squared error (MSE) loss function are used. The Adam optimizer ensures robust and efficient parameter updates even under strict constraints of low learning rates; MSE ensures that when the model adapts to new data, its core task remains to accurately capture the core degradation trend of health indicators, rather than overfitting noise.

[0151] S54: Fine-tuning execution and weight updates. Using a small amount of data from target domain B. The model is trained. During this process, only the weights of the unfrozen layers (the second LSTM layer, the Dropout layer, and the output layer) are updated.

[0152] S55: Model Saving. After a limited number of fine-tuning training rounds (e.g., 100 rounds), the model's high-level networks have adapted to the specific degradation patterns of the target domain B (e.g., potentially having different degradation rates or initial healthy baselines) while retaining general knowledge of the source domain. Finally, the fine-tuned model is saved, becoming the dedicated prediction model Model_B that is fully adapted to the target domain motor B.

[0153] S6: Remaining life prediction. Input the real-time health index sequence of motor B into Model_B to predict its future trend. Estimate the remaining useful life by comparing the predicted sequence with a preset failure threshold.

[0154] In the online monitoring of motor B, vibration signals are acquired in real time, and real-time HI values ​​are generated through the process from S1 to S3, maintaining a length of [missing information]. The latest HI sequence window is then input into the fine-tuned Model_B to predict its HI value sequence over a future period. By analyzing this predicted sequence, when the HI value is lower than the preset failure threshold (e.g., HI=0.2) or intersects the threshold line through trajectory extrapolation, the time corresponding to the intersection point is calculated. The difference between this time and the current time is the estimated remaining useful life (RUL), thus achieving early warning.

[0155] This application generates robust health indicators that outperform single features through multi-feature fusion, and overcomes the bottleneck of scarce data for new devices by utilizing transfer learning, achieving rapid model deployment and excellent generalization, and significantly improving the accuracy and engineering practicality of UAV motor health management.

[0156] In the remaining lifetime prediction: the HI sequence of the current window of motor B is input into the fine-tuned model Model_B to obtain the HI value prediction sequence for the next k steps. Set failure thresholds based on engineering experience. When the predicted HI value remains below this threshold, the function is deemed to have failed; the first value that satisfies this threshold is then searched in the predicted sequence. time point It can calculate the remaining lifespan. By combining the current HI value, predicted trend, and RUL, the health status of motor B can be graded and assessed.

[0157] In the specific implementation process, in step S6, the remaining useful life is accurately predicted by inputting the real-time health index sequence of the target domain motor into its dedicated prediction model Model_B:

[0158] S61: Real-time HI sequence generation. In the online monitoring system of motor B, the S1-S3 process is executed in real time: vibration signal acquisition → preprocessing → feature extraction → SVR model inference, generating real-time HI values, and maintaining a sliding window of length 30. .

[0159] S62: Health Trend Prediction. (This will change the current window...) Input the fine-tuned model Model_B to obtain the HI value prediction sequence for the next k steps. Recursive prediction or sequence-to-sequence prediction can be used.

[0160] S63: Failure threshold setting. Set the failure threshold based on engineering experience. When the predicted HI value remains below this threshold, the function is deemed to be in failure.

[0161] S64: Remaining useful life calculation. Find the first sequence that satisfies this condition. time point Then the remaining useful life is:

[0162]

[0163] in, It is the current time point. The sampling interval for the HI sequence (e.g., the time interval between each flight). The system triggers a maintenance alarm when the RUL falls below a preset safety margin (e.g., 10 flights).

[0164] Figure 3 A schematic diagram of a drone motor remaining life prediction device according to an embodiment of this application is shown. Figure 3 As shown, the device includes:

[0165] The acquisition module is configured to acquire the early vibration signal of the target domain UAV motor and use the early vibration signal of the target domain UAV motor and a pre-trained support vector regression (SVR) model to obtain the early health label series of the target domain UAV motor.

[0166] The model training module is configured to fine-tune the pre-trained source domain LSTM prediction model using the early health label series of the target domain UAV motors to obtain the trained target domain LSTM prediction model.

[0167] The prediction module is configured to acquire the real-time health tag sequence of the target domain UAV motors, and obtain the remaining life prediction result of the target domain UAV motors by inputting the real-time health tag sequence of the target domain UAV motors into the trained target domain LSTM prediction model.

[0168] The above is a schematic scheme of a drone motor remaining life prediction device according to this embodiment. It should be noted that the technical solution of this drone motor remaining life prediction device and the technical solution of the drone motor remaining life prediction method described above belong to the same concept. For details not described in detail in the technical solution of the drone motor remaining life prediction device, please refer to the description of the technical solution of the drone motor remaining life prediction method described above.

[0169] Figure 4 A structural block diagram of a computing device 400 according to an embodiment of this application is shown. The components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.

[0170] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0171] In one embodiment of this application, the aforementioned components of the computing device 400 and Figure 4 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 4 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this application. Those skilled in the art can add or replace other components as needed.

[0172] The computing device 400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 400 can also be a mobile or stationary server.

[0173] The processor 420 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described method for predicting the remaining lifespan of the UAV motor.

[0174] An embodiment of this application also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described method for predicting the remaining lifespan of a UAV motor.

[0175] An embodiment of this application also provides a computer program, wherein when the computer program is executed in a computer, the computer is instructed to perform the steps of the above-described method for predicting the remaining lifespan of a drone motor.

[0176] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the present invention.

Claims

1. A method for predicting the remaining lifespan of a drone motor, characterized in that, include: The early vibration signal of the UAV motor in the target domain is acquired, and using the early vibration signal of the UAV motor in the target domain and a pre-trained support vector regression (SVR) model, a series of early health labels for the UAV motor in the target domain are obtained, including: By sequentially preprocessing and framing the early vibration signal of the UAV motor in the target domain, the framing signal of each sample of the UAV motor in the target domain is obtained. Furthermore, by performing multiple time-domain feature calculations on the framing signal of each sample of the UAV motor in the target domain, the multidimensional feature vector of each sample of the UAV motor in the target domain is obtained. By inputting the multidimensional feature vector of each sample of the target domain UAV motor into the pre-trained support vector regression (SVR) model, the early health label series of the target domain UAV motor is obtained; The pre-trained source domain LSTM prediction model is fine-tuned using the early health label series of the target domain UAV motor to obtain the trained target domain LSTM prediction model. The real-time health tag sequence of the target domain UAV motor is obtained, and the remaining lifespan prediction result of the target domain UAV motor is obtained by inputting the real-time health tag sequence of the target domain UAV motor into the trained target domain LSTM prediction model. The pre-trained SVR model includes: The original vibration signal of the UAV motor in the source domain is obtained, and the original vibration signal is preprocessed and framed sequentially to obtain the framed signal of each sample. Multiple time-domain features are calculated on the framed signal of each sample to obtain the multidimensional feature vector of each sample. The multidimensional feature vectors of all samples are used to form a multidimensional feature matrix. Obtain the health and failure benchmarks of the source domain UAV motor, and use the health and failure benchmarks to generate a health label vector for each sample; An SVR model is constructed by using the multidimensional feature vector of each sample as the input of the SVR model and the health label vector of each sample as the output of the SVR model. The SVR model is then trained to obtain a trained SVR model.

2. The method according to claim 1, characterized in that, The pre-trained source domain LSTM prediction model includes: The original vibration signals of the Yuanyu UAV motor throughout its entire life cycle are obtained, and the original vibration signals of the Yuanyu UAV motor throughout its entire life cycle and the pre-trained support vector regression (SVR) model are used to obtain a series of health labels for the Yuanyu UAV motor throughout its entire life cycle. Using the source domain UAV motor full life cycle health tag series, a training sample pair is constructed, which includes a health tag sequence of length L as input and the next health tag as output. Construct a source domain LSTM prediction model and train the source domain LSTM prediction model using the training samples to obtain a trained source domain LSTM prediction model.

3. The method according to claim 2, characterized in that, The LSTM prediction model consists of, from bottom to top, an input layer, a first LSTM layer, a first Dropout layer, a second LSTM layer, a second Dropout layer, and an output layer.

4. The method according to claim 3, characterized in that, The step of fine-tuning the pre-trained source domain LSTM prediction model using the early health tag series of the target domain UAV motor to obtain the trained target domain LSTM prediction model includes: Using the previous health tag series of the target domain UAV motor, construct micro-training sample pairs containing a health tag sequence of length L as input and the next health tag as output of the input sequence; The pre-trained source domain LSTM prediction model is fine-tuned using the micro-training samples to obtain a trained target domain LSTM prediction model.

5. The method according to claim 4, characterized in that, The step of fine-tuning the pre-trained source domain LSTM prediction model using the micro-training samples to obtain a trained target domain LSTM prediction model includes: During the fine-tuning training of the pre-trained source domain LSTM prediction model using the micro-training samples, the network weights of the first LSTM layer and the first Dropout layer in the pre-trained source domain LSTM prediction model are frozen, while the network weights of the second LSTM layer, the second Dropout layer, and the output layer in the pre-trained source domain LSTM prediction model are fine-tuned using a low learning rate, thus obtaining the trained target domain LSTM prediction model.

6. A device for predicting the remaining lifespan of a drone motor, characterized in that, include: The acquisition module is configured to acquire the early vibration signal of the target domain UAV motor, and use the early vibration signal of the target domain UAV motor and a pre-trained support vector regression (SVR) model to obtain an early health label series of the target domain UAV motor. This includes: sequentially preprocessing and framing the early vibration signal of the target domain UAV motor to obtain a framed signal for each sample of the target domain UAV motor; calculating multiple time-domain features on the framed signal of each sample of the target domain UAV motor to obtain a multi-dimensional feature vector for each sample of the target domain UAV motor; and inputting the multi-dimensional feature vector of each sample of the target domain UAV motor into the pre-trained SVR model to obtain the early health label series of the target domain UAV motor. The model training module is configured to fine-tune the pre-trained source domain LSTM prediction model using the early health label series of the target domain UAV motors to obtain the trained target domain LSTM prediction model. The prediction module is configured to acquire the real-time health tag sequence of the target domain UAV motor, and obtain the prediction result of the remaining lifespan of the target domain UAV motor by inputting the real-time health tag sequence of the target domain UAV motor into the trained target domain LSTM prediction model. The pre-trained SVR model includes: acquiring the original vibration signal of the UAV motor in the source domain, and obtaining the framed signal of each sample by preprocessing and framing the original vibration signal in sequence; calculating multiple time-domain features of each sample's framed signal to obtain a multi-dimensional feature vector of each sample, and constructing a multi-dimensional feature matrix by combining the multi-dimensional feature vectors of all samples; acquiring the health benchmark and failure benchmark of the UAV motor in the source domain, and generating a health label vector of each sample using the health benchmark and failure benchmark; constructing an SVR model, using the multi-dimensional feature vector of each sample as the input of the SVR model, and using the health label vector of each sample as the output of the SVR model, and training the SVR model to obtain a trained SVR model.

7. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the unmanned aerial vehicle motor remaining life prediction method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the unmanned aerial vehicle motor remaining life prediction method according to any one of claims 1 to 5.

Images