Simulation data enhanced controllable-pitch propeller small sample fault diagnosis meta-learning method

Abstract

The invention discloses a simulation data enhanced controllable-pitch propeller small sample fault diagnosis meta-learning method, and relates to the field of fault diagnosis, and the method comprises the steps: carrying out the simulation modeling and fault injection based on the working mechanism of a controllable-pitch propeller, and building a system fault model; constructing a fault simulation enhancement data set; constructing a variable-length time series data classification network model, wherein the model is connected with a multimode encoder and a triple loss function in combination with a residual error; training the variable-length time series data classification network model by utilizing a three-stage meta-learning training process, wherein the three-stage meta-learning training process comprises a meta-learning training process by utilizing an open source data set and a fault simulation enhancement stage by utilizing a fault simulation enhancement data set; and applying the trained variable-length time series data classification network model to an actually collected fault small sample data set for small sample task training. According to the method, the model is pre-trained by using the historical related data set, so that the current task data is reduced, training samples are reduced, and higher diagnosis precision is achieved.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and particularly to a meta - learning method for small - sample fault diagnosis of controllable - pitch propellers with enhanced simulation data. Background Art

[0002] The controllable - pitch propeller is a core component in the ship power transmission system. By adjusting the blade angle, it can adjust the propulsion direction and power of the ship without changing the main shaft speed. This device usually uses a hydraulic system to execute the pitch - adjusting command. There are many components, complex structures, and high costs in the system. Once a fault occurs, it is difficult to troubleshoot, and in severe cases, it may lead to major accidents. The high complexity of the hydraulic system makes each component prone to failure in a harsh working environment, thus posing a major threat to the safe operation of the ship. In addition, due to the sporadic and diverse types of controllable - pitch propeller faults, in practical applications, there is often a lack of sufficient fault - condition samples in the monitoring data, which troubles the model during training due to the small - sample problem.

[0003] Meta - learning is an effective learning method to address the small - sample problem, and is widely used to improve the learning ability of models under limited sample conditions. Meta - learning can significantly reduce the computational cost caused by retraining due to frequent parameter tuning and task switching. It endows the model with the ability to "learn to learn", enabling the model to quickly adapt to new tasks based on existing knowledge. Specifically, meta - learning obtains a better hyperparameter configuration by training on multiple tasks and then optimizes for a specific task.

[0004] Meta - learning generally includes two main stages: the training stage and the testing stage. In the training stage, the training samples are divided into a support set and a query set. The support set is used to train the model's parameters, while the query set is used to evaluate the model's performance on a specific task. In the testing stage, the model needs to quickly adapt to new tasks based on the knowledge obtained in the training stage and perform efficient reasoning and decision - making.

[0005] Therefore, those skilled in the art are committed to developing a meta - learning method for small - sample fault diagnosis of controllable - pitch propellers with enhanced simulation data. Summary of the Invention

[0006] In view of the above - mentioned defects of the prior art, the technical problem to be solved by the present invention is how to achieve small - sample fault diagnosis of the controllable - pitch propeller device.

[0007] To achieve the above object, the present invention provides a meta - learning method for small - sample fault diagnosis of controllable - pitch propellers with enhanced simulation data, and the method includes the following steps:

[0008] S101: Based on the working mechanism of the controllable - pitch propeller, perform simulation modeling and fault injection to establish a system fault model;

[0009] S103: Construct a fault simulation enhanced dataset, where the dataset covers data offsets, different configuration parameters, and sensor noise data in actual working conditions;

[0010] S105: Construct a variable-length time series data classification network model, which is configured as a classification network model that combines a residual connection multi-modal encoder and a triplet loss function;

[0011] S107: Train the variable-length time series data classification network model using a three-stage meta-learning training process, where the three-stage meta-learning training process includes a meta-learning training process using an open-source dataset and a fault simulation enhancement stage using the fault simulation enhanced dataset;

[0012] S109: Apply the trained variable-length time series data classification network model to a small sample dataset of faults collected in actuality for small sample task training.

[0013] Further, when performing simulation modeling in step S101, the following steps are included:

[0014] S1011: Construct a system framework, determine the interactions between subsystems in the system and the structural relationships of components in the system;

[0015] S1012: Establish the dynamic equation of the hydraulic system, taking into account the physical properties of hydraulic oil in the dynamic equation;

[0016] S1013: Determine component parameters, where the components include hydraulic pumps, hydraulic cylinders, and check valve pressure sensors;

[0017] S1014: Establish a control system model according to the control strategy of the controllable pitch propeller;

[0018] S1015: Conduct mechanical system modeling, establish the kinematic equation and dynamic equation of the mechanical system, where the mechanical system includes the rotor mechanism of the controllable pitch propeller.

[0019] Further, in step S103, when constructing the fault simulation enhanced dataset, the following steps are included:

[0020] S1031: Establish the change curves of the density and viscosity of hydraulic oil at different temperatures according to the physical properties of hydraulic oil;

[0021] S1032: During the simulation process, adjust the density and viscosity parameters of the hydraulic oil according to the set random temperature to simulate the changing performance of the hydraulic system under actual working conditions;

[0022] S1033: Simulate the hardware setting differences caused by manual operations by randomly setting the working pressure of the check valve within a predetermined range;

[0023] S1034: Generate a large number of fault data sets by introducing the sensor noise and considering the hydraulic oil parameters at different temperatures and the system parameter changes under different personnel operations.

[0024] Further, in the step S105, the residual connection multi-modal encoder is a multi-modal encoder based on temporal residual connection. The convolutional kernels in the multi-modal encoder are in multi-modal parallel connection, and the length of the inter-layer features remains unchanged. The last layer of the multi-modal encoder performs global average pooling to map the variable-length temporal data to a fixed-length global feature vector.

[0025] Further, the global feature vector averages the feature vectors of each channel after extraction by the convolutional network along the temporal dimension through the global average method:

[0026]

[0027] X = {x 1 , x 2 , …, x T}

[0028] where, is the global feature vector, x t represents the feature vector at time step t, T is the total number of time steps, and X is the time series.

[0029] Further, in the step S105, the triplet loss function is used to evaluate the mapping effect of the classification network model for different classes, and the triplet loss function is:

[0030] L triplet (a, p, n) = max(||f(a) - f(p)|| - ||f(a) - f(n)|| + α, 0)

[0031] where, a is the anchor sample, p is the positive sample, belonging to the same class as the anchor sample, n is the negative sample, belonging to a different class from the anchor sample, f(a) is the feature vector of the anchor sample a, generated by the classification network model, f(p) is the feature vector of the positive sample p, f(n) is the feature vector of the negative sample n, and α is the learning rate of the meta-learner.

[0032] Further, in the step S107, the meta-learning training process using the open-source data set constructs a few-shot task meta-training task with the data in the open-source data set, and improves the ability of the variable-length temporal data classification network model to extract common features in the temporal data and the feature of the pressure and flow detection point data through training.

[0033] Further, in the step S107, the open-source data set includes the UCR Archive data set, the ZeMA hydraulic bench data set, and the Rolf Isermann centrifugal pump data set. Among them, the UCR Archive data set is a widely used and publicly available time series data set; the ZeMA hydraulic bench data set contains multiple sensor data and has multiple sampling rates; the Rolf Isermann centrifugal pump data set is a data set specifically for analyzing the performance and failure modes of centrifugal pumps, including the pressure difference, rotational speed, motor and pump torque, and pump outlet flow data of the centrifugal pump.

[0034] Further, in the step S107, the few-shot task meta-training task regards the network model as a function f with parameters θ θ , the task support set and the task query set of the inner loop training belong to the same distribution. During the training process, the task learner and the meta-learner are iteratively trained and share parameters.

[0035] Further, the few-shot task meta-training task includes the following steps:

[0036] S1071: During the training process, the parameters of the task learner are updated through the task support set:

[0037]

[0038] S1072: Input the task query set T i q into the task learner to calculate the loss;

[0039] S1073: Take the loss as and pass it to the meta-learner;

[0040] S1074: After accumulating one or more , the parameters of the meta-learner are updated as:

[0041]

[0042] where θ is the parameter, f θ is a function with parameters θ, is the task support set, is the task query set, p(T) is the distribution, β is the learning rate of the task learner, α is the learning rate of the meta-learner, i is the task index, representing the current task number, representing each task sample in the meta-learning process, and L is the loss function, which is used to measure the difference between the predicted value output by the task learner and the true label.

[0043] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention uses a historical-related dataset to pre-train the model, thereby reducing the increase in the amount of data for the current task and reducing the training samples required for the fault diagnosis model.

[0045] 2. The present invention uses a multi-stage meta-learning method to increase the relevance and diversity of the historical dataset, thereby improving the gap between the pre-trained model and the current task and enhancing the accuracy and convergence speed of the fault diagnosis model.

[0046] 3. The present invention uses a fault simulation enhanced dataset to improve the adaptability of the model parameters to the target task while ensuring the amount of training data.

[0047] The following will further illustrate the concept, specific structure and technical effects of the present invention with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a schematic flow chart of a meta-learning method for small-sample fault diagnosis of controllable pitch propellers based on simulation data augmentation training according to a preferred embodiment of the present invention;

[0049] Figure 2 is a schematic framework diagram of a meta-learning method for small-sample fault diagnosis of controllable pitch propellers based on simulation data augmentation training according to a preferred embodiment of the present invention;

[0050] Figure 3 is the training process of a single sample of the network trained by enhanced training according to a preferred embodiment of the present invention;

[0051] Figure 4 is the training process of a double sample of the network trained by enhanced training according to a preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The following introduces multiple preferred embodiments of the present invention with reference to the accompanying drawings of the specification to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0053] In the accompanying drawings, components with the same structure are denoted by the same numeral labels, and components with similar structures or functions are denoted by similar numeral labels. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. To make the illustration clearer, the thickness of some components in the drawings is appropriately exaggerated.

[0054] Such as Figure 1 and Figure 2As shown, in view of the problem that it is impossible to effectively achieve small-sample fault diagnosis of controllable pitch propeller devices in the prior art, the present invention provides a method for combining meta-learning and data augmentation to solve the small-sample fault diagnosis of controllable pitch propeller monitoring data. The present invention introduces a fault simulation data augmentation stage in two stages of meta-learning. By using the mechanism model of the controllable pitch propeller and injecting faults to form a fault simulation augmented data set, while ensuring the amount of training data, it improves the adaptability of model parameters to the target task. Finally, when facing a small number of fault samples of the target task, it can successfully extract features based on prior knowledge, significantly improving the accuracy and efficiency of small-sample fault diagnosis.

[0055] As Figure 1 shown, a meta-learning method for small-sample fault diagnosis of a controllable pitch propeller with simulation data augmentation provided by an embodiment of the present invention includes the following steps:

[0056] S101: Conduct simulation modeling and fault injection based on the working mechanism of the controllable pitch propeller to establish a system fault model.

[0057] When conducting simulation modeling of the ship power transmission system in this embodiment, it includes the following steps:

[0058] S1011: Construct a system framework to determine the interaction between each subsystem in the system and the structural relationship of each component. Determine the interaction between hydraulic, control, mechanical and other subsystems based on the schematic diagram of the system, as well as the structural relationship of components such as pump sets, valves, oil distributors and pipelines.

[0059] S1012: Establish the dynamic equation of the hydraulic system, and consider the physical properties of hydraulic oil in the dynamic equation, including considering physical properties such as the viscosity and density of hydraulic oil.

[0060] S1013: Determine component parameters, and the components include components such as hydraulic pumps, hydraulic cylinders and one-way valve pressure sensors.

[0061] S1014: According to the control strategy of the controllable pitch propeller, establish a control system model to complete the modeling of the control system.

[0062] S1015: Mechanical system modeling, establish the kinematic equation and dynamic equation of the mechanical system, and the mechanical system includes mechanical components such as the swashplate mechanism of the controllable pitch propeller.

[0063] S103: Construct a fault simulation augmented data set, and the data set covers data offsets in actual working conditions, different configuration parameters and sensor noise data.

[0064] When constructing the fault simulation augmented data set, it includes the following steps:

[0065] S1031: According to the physical properties of hydraulic oil, establish the change curves of the density and viscosity of hydraulic oil at different temperatures.

[0066] S1032: During the simulation process, adjust the density and viscosity parameters of the hydraulic oil according to the set random temperature to simulate the performance of the hydraulic system changing under actual working conditions.

[0067] S1033: Simulate the hardware setting differences caused by manual operation by randomly setting the working pressure of the check valve within a predetermined range.

[0068] S1034: Generate a large number of fault data sets by introducing sensor noise and considering the changes in hydraulic oil parameters at different temperatures and system parameters under different personnel operations.

[0069] S105: Construct a variable-length time series data classification network model, and the model is configured as a classification network model that combines a residual connection multi-modal encoder and a triplet loss function.

[0070] In this embodiment, the residual connection multi-modal encoder is a multi-modal encoder based on time series residual connection. The convolutional kernels in the multi-modal encoder are in multi-modal parallel, and the inter-layer feature length is kept unchanged. The last layer of the multi-modal encoder performs global average pooling to map the variable-length time series data to a fixed-length global feature vector.

[0071] The global feature vector averages the feature vectors of each channel extracted by the convolutional network along the time series dimension through the global average method. The global feature vector is calculated in the following way:

[0072]

[0073] X = {x 1 , x 2 , …, x T}

[0074] Wherein, is the global feature vector, x t represents the feature vector at time step t, T is the total number of time steps, and X is the time series.

[0075] In this embodiment, the triplet loss function is used to evaluate the mapping effect of the classification network model on different classes. The triplet loss function is:

[0076] L triplet( a, p, n) = max(||f(a) - f(p)|| - ||f(a) - f(n)|| + α, 0)

[0077] Among them, a is the anchor sample, p is the positive sample belonging to the same category as the anchor sample, n is the negative sample belonging to a different category from the anchor sample, f(a) is the feature vector of the anchor sample a generated by the classification network model, f(p) is the feature vector of the positive sample p, f(n) is the feature vector of the negative sample n, and α is the learning rate of the meta-learner.

[0078] Compared with the cross-entropy loss function, the triplet loss function has obvious advantages in training the encoder. The core of the triplet loss function lies in focusing on the relative distance between samples rather than the specific class labels, so it does not depend on the number of classes. This feature enables the triplet loss function to perform effective learning without explicit class information. In addition, due to the design of the triplet loss function, when the model switches between different tasks, there is no need to retrain the model parameters for each new task. In this way, the model can quickly adapt to new tasks by adjusting the relative distance relationship based on the existing tasks, greatly improving its adaptability and transfer ability. Therefore, the triplet loss function is particularly suitable for application in the training process of meta-learning, can efficiently transfer between tasks, and accelerate the learning of the model.

[0079] S107: Train the variable-length time-series data classification network model using the three-stage meta-learning training process, which includes the meta-learning training process using the open-source dataset and the fault simulation enhancement stage using the fault simulation enhanced dataset.

[0080] When training the variable-length time-series data classification network model, in this embodiment, by introducing simulation data based on the mechanism model in the two-stage training process of meta-learning, the adaptability and accuracy of the model are significantly improved. The key to this strategy is to add fault nodes to the mechanism model and simulate the offset of actual working condition data, thereby generating a large amount of high-quality training data highly relevant to the actual working conditions. Through this method, the simulation data not only maintains the quality and relevance of the data but also avoids the training dilemma caused by insufficient data in actual operations.

[0081] The simulation data is generated based on the mechanism model of the target system, ensuring its high consistency with actual applications. Combining the general feature learning in the public time-series dataset, the domain knowledge in the professional hydraulic fault dataset, and the simulation data based on the mechanism model of the target system, the proposed solution in this paper can effectively address the small-sample problem and provide strong support for the fault diagnosis task. Through this multi-stage and multi-source data training strategy, not only the generalization ability of the model is improved, but also it is ensured that the model can achieve efficient and accurate fault diagnosis in actual applications, significantly improving the reliability and applicability of the fault diagnosis system.

[0082] The meta-learning training process based on open-source datasets uses the data in the open-source datasets to construct few-shot task meta-training tasks, and improves the ability of the variable-length time-series data classification network model to extract common features and pressure-flow detection point data features from time-series data through training.

[0083] Among them, the open-source datasets include the UCR Archive dataset, the ZeMA hydraulic bench dataset, and the Rolf Isermann centrifugal pump dataset.

[0084] 1) The UCR Archive dataset is a widely used and publicly available time-series dataset;

[0085] 2) The ZeMA hydraulic bench dataset contains multiple sensor data and has multiple sampling rates;

[0086] 3) The Rolf Isermann centrifugal pump dataset is a dataset specifically for analyzing the performance and fault modes of centrifugal pumps, including the pressure difference, rotational speed, motor and pump torque, and pump outlet flow data of centrifugal pumps.

[0087] In this embodiment, the few-shot task meta-training task regards the network model as a function f with parameters θ θ , and the task support set and task query set of the inner-loop training belong to the same distribution. During the training process, the task learner and the meta-learner are iteratively trained and share parameters.

[0088] The few-shot task meta-training task includes the following specific steps:

[0089] S1071: During the training process, the parameters of the task learner are updated through the task support set:

[0090]

[0091] S1072: Input the task query set T i q into the task learner to calculate the loss;

[0092] S1073: Take the loss as and pass it to the meta-learner;

[0093] S1074: After accumulating one or more , the parameters of the meta-learner are updated to:

[0094]

[0095] where θ is the parameter, f θ is a function with parameters θ, T i s is the task support set, and T iq Let \(T\) be the task query set, \(p(T)\) be the distribution, \(\beta\) be the learning rate of the task learner, \(\alpha\) be the learning rate of the meta-learner, \(i\) be the task index representing the current task number, which represents each task sample in the meta-learning process, and \(L\) be the loss function used to measure the difference between the predicted value output by the task learner and the true label.

[0096] S109: Apply the trained variable-length time-series data classification network model to the small sample fault data set collected in actuality for small sample task training.

[0097] The model parameters after three-step meta-training are applied to the small sample task training of bench data by the meta-learning method, significantly improving the stability and accuracy of the training results. In contrast, the results of directly training using the target data set highly depend on the random selection of training samples. Especially in single-sample training, the accuracy of the validation set fluctuates greatly during each training. After meta-training with the public data set and data augmentation, the sensitivity of the model to the randomness of training samples is significantly reduced, the accuracy of the validation set is steadily improved, showing stronger robustness and generalization ability.

[0098] As Figure 3 and Figure 4 shown, it demonstrates the performance of the enhanced training network during single-sample and double-sample training processes. By using the monitoring data collected from the simulation of a certain pitch-changing propeller bench under fault conditions to construct a small sample fault diagnosis task, comparing the directly trained classification network with the network enhanced by the method of this embodiment, it can be clearly seen that the method of this embodiment significantly improves the diagnostic performance in the small sample fault diagnosis of pitch-changing propeller monitoring data. The advantage of this method lies in effectively utilizing the combined effect of meta-training and data augmentation, reducing the impact of the small sample problem, and providing technical support for the efficient fault diagnosis of key equipment.

[0099] A meta-learning method for small sample fault diagnosis of pitch-changing propellers based on simulation data augmentation provided by the present invention combines meta-learning and data augmentation techniques, aiming to solve the small sample fault diagnosis problem existing in pitch-changing propeller monitoring data. Specifically, the present invention analyzes the possible fault conditions of each component of the pitch-changing propeller through mechanism simulation modeling and fault injection of the pitch-changing propeller, and establishes a systematic fault model. On this basis, by simulating the offset of actual working condition data, adding sensor noise, considering the influence of temperature change on hydraulic oil parameters, and the possible changes in system parameters under different personnel operations, the diversity of the data set is further enhanced.

[0100] In addition, by setting multiple fault classifications and pitch change inputs, the present invention generates a fault simulation enhanced dataset under multiple fault conditions, ensuring the richness and high quality of the data. To solve the problem of fault diagnosis for variable-length time series data, this paper proposes a variable-length time series data classification network that combines a residual connection multi-modal encoder and a triplet loss function, which can effectively perform transfer training between tasks.

[0101] Introducing a time series data meta-learning framework, in the meta-training stage, the present invention uses the UCR time series dataset and an open-source hydraulic system dataset to construct few-shot training tasks, and trains the initial parameters of the network through the meta-learning framework. In the fault simulation enhancement stage, training is carried out using the enhanced fault dataset, which not only ensures the sufficiency of the data volume but also greatly improves the adaptability of the model parameters to the target task.

[0102] Compared with other traditional data-driven fault diagnosis methods, the present invention only requires a small number of fault training samples and can significantly improve the accuracy of fault diagnosis, demonstrating stronger robustness and generalization ability. This method can not only effectively solve the few-shot problem but also provide a more efficient and accurate fault diagnosis scheme for the controllable pitch propeller hydraulic system in practical applications.

[0103] The following will describe the present invention in detail in conjunction with the preferred embodiments of the present invention.

[0104] As Figure 1 、 Figure 2 shown, aiming at the problem that the few-shot fault diagnosis of the controllable pitch propeller device cannot be effectively achieved in the prior art, the present invention provides a meta-learning method for few-shot fault diagnosis of a controllable pitch propeller with simulation data enhancement. The present invention introduces a fault simulation data enhancement stage in two stages of meta-learning, uses the mechanism model of the controllable pitch propeller to form a fault simulation enhanced dataset through fault injection, improves the adaptability of the model parameters to the target task while ensuring the training data volume, and finally, when facing a small number of fault samples of the target task, can smoothly extract features based on prior knowledge, significantly improving the accuracy and efficiency of few-shot fault diagnosis.

[0105] The meta-learning method for few-shot fault diagnosis of a controllable pitch propeller with simulation data enhancement provided in this embodiment specifically includes the following steps:

[0106] (1) Based on the actual structure and working principle of the controllable pitch propeller, construct a system simulation model

[0107] Step 1: Build a system framework, and determine the interaction of subsystems such as hydraulic, control, and mechanical, and the structural relationship of components such as pump sets, valves, oil distributors, and pipelines based on the schematic diagram.

[0108] Step 2: Considering physical properties such as the viscosity and density of hydraulic oil, establish the dynamic equation of the hydraulic system.

[0109] Step 3: Determine component parameters, including components such as hydraulic pumps, hydraulic cylinders, check valves, and pressure sensors.

[0110] Step 4: Control system modeling: Establish a control system model according to the control strategy of the controllable pitch propeller.

[0111] Step 5: Mechanical system modeling: Include mechanical components such as the controllable pitch propeller's swashplate mechanism, and establish its kinematic and dynamic equations.

[0112] (2) Fault injection based on mechanism and system fault models

[0113] Step 6: Simulate fault conditions with different fault locations and failure modes by modifying the parameters in the mechanism model or adding fault nodes.

[0114] Table 1 shows the specific injection components, fault modes, and modeling methods. In addition to the normal condition, this embodiment simulated a total of four fault conditions, including pump leakage fault condition, distributor leakage fault condition, proportional valve delay fault condition, and servo cylinder leakage fault condition. Among them, the pump leakage situation was simulated by modifying the volumetric efficiency of the pump; in the distributor leakage condition, the oil leakage of the oil pipe was simulated by adding a damping hole to the oil delivery pipe and grounding it; the proportional valve delay fault was simulated by increasing the control time delay of the proportional valve; in the servo cylinder leakage condition, the servo cylinder leakage fault was simulated by modifying the leakage area.

[0115] Table 1 Fault simulation enhanced dataset: Fault modes and their modeling methods

[0116] Component Fault mode Modeling method PTO pump Pump leakage Modify volumetric efficiency parameter Oil distributor Oil leakage Add leakage port to oil distribution pipeline Proportional valve Proportional valve delay Add delay to control signal Servo cylinder Servo cylinder leakage Add leakage port at servo cylinder

[0117] (3) Simulation dataset based on the physical properties of hydraulic oil and check valves

[0118] Step 7: According to the physical properties of hydraulic oil, establish the change curves of the density and viscosity of hydraulic oil at different temperatures.

[0119] During the simulation process, adjust the density and viscosity parameters of the hydraulic oil according to the set random temperature to simulate the changing performance of the hydraulic system under actual working conditions. For example, in a high-temperature environment, the viscosity of the hydraulic oil decreases, and the system response speed may become faster; in a low-temperature environment, the viscosity of the hydraulic oil increases, and the system response speed may become slower.

[0120] Step 8: Hardware setting differences caused by manual operation: The check valve works based on the principle of balancing the liquid pressure acting on the valve core and the spring force inside the valve. Its maximum pressure is set by manually operating the spring knob each time it runs, so there will be differences in the maximum pressure each time it runs. Simulate the operating characteristics of different operators by randomly setting the working pressure of the check valve within a 5% range.

[0121] (4) Simulate and enhance the dataset with sensor noise

[0122] Step 9: By introducing sensor noise and considering the hydraulic oil parameters at different temperatures and the system parameter changes under different personnel operations, generate a large number of fault datasets. Introduce sensor noise into the simulation model to simulate data offset in actual working conditions: According to the characteristics of piezoresistive pressure sensors, establish a noise model that includes thermal noise and flicker noise. Thermal noise is mainly caused by the thermal motion of the resistor, while flicker noise is related to the microstructure of the sensor material. During the simulation process, add a certain proportion of thermal noise and flicker noise to the sensor output data to simulate the sensor output under real working conditions.

[0123] In this embodiment, on the mechanism model of a certain controllable pitch propeller, two normal working conditions are simulated: running at a normal speed of 218 rpm and running at a low speed of 50 rpm; four fault conditions: pump leakage fault, distributor leakage fault, proportional valve delay of 3 s fault, servo cylinder leakage fault. The final dataset includes six working conditions: 1) normal speed of 218 rpm working condition; 2) low speed of 50 rpm working condition; 3) pump leakage fault; 4) distributor leakage fault; 5) proportional valve delay of 3 s fault; 6) servo cylinder leakage fault. During the process of simulating the data required for the experiment, collect data by adding pressure and flow measurement points in the Simulink simulation model, obtain the historical experimental data of six different working conditions, set two operation modes from full ahead to full astern and from full astern to full ahead, and the running time of each sample is 100 s. After simulation, there are 285 normal data, 140 pump leakage data, 120 proportional valve delay of 3 s data, 116 distributor leakage data, and 126 servo cylinder leakage data. Table 2 shows the simulated and enhanced dataset.

[0124] Table 2 Simulated fault dataset

[0125]

[0126] (5) Multi-modal encoder based on temporal residual connection

[0127] Step 10: For the temporal convolutional network with residual connection, the convolutional kernels are in multi-modal parallel, and 0 padding is used to keep the feature length between layers unchanged.

[0128] Step 11: The last layer is global average pooling to map the temporal data to the feature space.

[0129] Using a temporal convolutional layer that keeps the length of the features between layers unchanged and an encoder that finally takes the mean along the temporal dimension, there is no requirement for the length of the temporal data, and data of different lengths can be mapped to the feature space with the same process. In a temporal convolutional network, the correlation and dependence between time steps are very crucial. Therefore, a strategy of keeping the length of the features between layers unchanged is adopted. Keeping the length unchanged can better capture this information. On the feature map after convolutional processing, by taking the mean along the temporal dimension (i.e., global average pooling), the variable-length temporal data can be mapped to a fixed-length feature vector. This method is simple and effective, capable of smoothing the local variations in the temporal data and obtaining global features. This network structure maps data of different lengths to the feature space in the same way, which is particularly suitable for applying transfer learning between tasks with different temporal lengths and constructing a data encoder.

[0130] Suppose the time series data is represented as, X = {x 1 , x 2 , …, x T} where x t represents the feature vector at time step t. The global average method takes the average of the feature vectors of each channel along the temporal dimension after being extracted by the convolutional network to obtain the global feature vector

[0131]

[0132] The method of convolution with invariant feature length and taking features by global average is not affected by the length of the input data. The length of the running data to be intercepted can be set as needed, and the evaluation results can be given in real time.

[0133] (6) Construct the model training objective of the triplet loss function

[0134] Step 12: Use the triplet loss function to evaluate the mapping effect of the network on different classes:

[0135] L triplet (a, p, n) = max(||f(a) - f(p)|| - ||f(a) - f(n)|| + α, 0)

[0136] Step 13: Online update the triplet to update the encoder parameters

[0137] Each training combination includes three samples: an anchor sample a, a positive sample p (belonging to the same class as the anchor sample a), and a negative sample n (belonging to a different class from the anchor sample a). The training objective is to cluster the samples of the same class together in the feature space by reducing the loss function, while separating the samples of different classes in the feature space.

[0138] The triplet loss function has an advantage over the cross-entropy loss function in training the encoder: the triplet loss function focuses on the relative distance between samples rather than the specific class labels, so it does not depend on the number of classes. Due to the design of the triplet loss function, when switching between different tasks, there is no need to retrain the model parameters for each new task, which greatly improves the adaptability and transfer ability of the model. Therefore, it is particularly suitable for application in the meta-training process between different tasks.

[0139] (7) Add the three-stage meta-learning training process in the fault simulation increase stage

[0140] Step 14, Open-source time series data meta-training: Use the UCR time series dataset to construct a large number of few-shot task meta-training tasks.

[0141] Regard the network model as a function f with parameters θ θ , the task support set for the inner-loop training and the task query set T i q Both belong to the distribution p(T).

[0142] During the training process, the parameters of the task learner are updated through the support set:

[0143]

[0144] Among them, β is the learning rate of the task learner, i is the task index, representing the current task number, representing each task sample in the meta-learning process, and L is the loss function, which is used to measure the difference between the predicted value output by the task learner and the true label.

[0145] After that, the query set T i q is input into the task learner to calculate the loss. Then, the loss is used as passed to the meta-learner.

[0146] After accumulating one or more , the parameters of the meta-learner are updated as follows:

[0147]

[0148] Among them, α is the learning rate of the meta-learner.

[0149] During the training process, the two learners are iteratively trained and share parameters.

[0150] Step 15, Hydraulic system data meta-training: Use the monitoring dataset of the open-source hydraulic system to construct few-shot tasks.

[0151] In the meta-training stage, by exposing to a large number of training tasks, the widely used UCR Archive dataset, ZeMA hydraulic bench dataset, and Rolf Isermann centrifugal pump dataset are adopted, thereby improving the model's ability to extract common features from time series data and extract the feature of pressure and flow measurement point data.

[0152] Among them, the UCR Time Series Classification Archive is a widely used and publicly available collection of time series datasets, mainly for the classification research of time series data. This dataset covers multiple industries and fields, including healthcare, environmental monitoring, robot perception, image recognition, etc., and contains more than a hundred different dataset tasks, providing rich training tasks for time series data classification research.

[0153] The ZeMA hydraulic system dataset contains data from multiple sensors with different sampling rates. This hydraulic system periodically repeats a constant load cycle (duration of 60 seconds) and records data such as pressure, vibration, and temperature through 14 real sensor channels and 3 virtual sensor channels. This dataset contains 2,205 samples, and each sample may have 0 to 4 faults. The faulty components include coolers, valves, hydraulic pumps, and accumulators. Each component has different fault levels. Based on the ZeMA dataset, two composite fault diagnosis tasks with different difficulties are constructed in the experiment, namely a single-fault dataset with 4 labels and a composite-fault dataset with 14 labels.

[0154] The Rolf Isermann centrifugal pump fault dataset is a dataset specifically for analyzing the performance and fault modes of centrifugal pumps. This dataset contains data collected from measurement points such as the pressure difference, rotational speed, motor and pump torque, and outlet flow rate of the centrifugal pump. The sampling frequency of these data is 10 Hz, and the dataset is divided into two categories. The first category of data includes three fault types: normal operation, excessive clearance, and insufficient clearance, with 50 groups of data for each type, and the duration of each group of data is 201 seconds. The second category of data covers the normal state and 9 different fault types, and each category also includes 50 groups of data, with the duration of each group of data being 55 seconds. This dataset contains a wide range of fault scenarios, providing rich cases for studying the fault modes of centrifugal pumps and fault classification based on monitoring data.

[0155] Step 16: Enhance the meta-training of the simulation fault dataset: Construct enhanced few-shot training tasks based on the simulation dataset.

[0156] In the training process of two-stage meta-learning, introducing simulation data based on the mechanism model is the key to improving the adaptability and accuracy of the model. By injecting fault nodes into the mechanism model and simulating the offset of the actual working condition data, a large amount of high-quality training data can be generated while ensuring data quality and relevance. These simulation data not only make up for the dilemma of insufficient data in actual operation, but also have a high relevance to the actual application scenario because they are generated based on the mechanism model of the target system. Therefore, while ensuring sufficient data volume, the simulation data can still maintain a high relevance to the target task.

[0157] By combining general feature learning from public time series datasets, domain knowledge from professional hydraulic fault datasets, and simulation data based on the target system mechanism model, this solution can effectively solve challenges such as insufficient sample size in small sample fault diagnosis. Specifically, this multi-stage, multi-source data training strategy not only enhances the generalization ability of the model, but also enables the model to adapt to different tasks and working conditions, thereby achieving efficient and accurate fault diagnosis in practical applications. The introduction of simulation data provides a powerful training support for the model. Combining the advantages of meta-learning and simulation data, it can greatly improve the performance of the model in complex and variable tasks, ensuring the accuracy and robustness of fault diagnosis.

[0158] Step 17: Apply the model to the existing small sample data set of faults actually collected.

[0159] The model parameters after the three-step meta-training are used for the small sample task training of the bench data according to the meta-learning method. The results after training are as follows Figure 3 and Figure 4 As shown in the figure, the stability and accuracy of the results are improved. The results of direct training with the target dataset are seriously affected by the randomness of the training sample selection, especially the validation set accuracy of single-sample training varies greatly from one training to another. The results after meta-training with public datasets and meta-training with data augmentation are less affected by the randomness of the training sample selection, and both can achieve a high validation set accuracy.

[0160] like Figure 3 and Figure 4 The figure shows the single sample and double sample training process of the network after enhanced training. The small sample fault diagnosis task is constructed by collecting monitoring data from a certain controllable pitch propeller test bench simulating fault operation. By comparing the training results of the classification network directly trained in this task and enhanced by the method proposed in this patent, it can be seen that the method of this patent has significantly improved the small sample fault diagnosis performance of the controllable pitch propeller based on monitoring data.

[0161] Compared with the prior art, the simulation data-enhanced controllable pitch propeller small sample fault diagnosis meta-learning method provided in the embodiment of the present invention has the following beneficial effects:

[0162] 1. Aiming at the problem of the small amount of data in the adjustable pitch propeller system fault dataset in the prior art, the present invention uses the meta-learning method to pre-train the model with the historically relevant dataset, thereby reducing the increase in the amount of data for the current task and reducing the training samples required by the fault diagnosis model.

[0163] 2. Aiming at the problem of the low training efficiency of the two-stage meta-learning method, the present invention uses the multi-stage meta-learning method to increase the relevance and diversity of the historical dataset, thereby improving the gap between the pre-trained model and the current task and enhancing the accuracy and convergence speed of the fault diagnosis model.

[0164] 3. The present invention uses the fault simulation enhanced dataset to improve the adaptability of the model parameters to the target task while ensuring the amount of training data.

[0165] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art shall fall within the protection scope determined by the claims.

Claims

1. A meta-learning method for small sample fault diagnosis of controllable pitch propellers enhanced by simulation data, characterized in that: The method comprises the following steps: S101: Perform simulation modeling and fault injection based on the working mechanism of the controllable pitch propeller to establish a system fault model; S103: constructing a fault simulation enhanced data set, where the data set covers data offset, different configuration parameters, and sensor noise data in actual working conditions; S105: constructing a variable-length time series data classification network model, wherein the model is configured as a classification network model combining a residual connection multi-mode encoder and a triplet loss function; S107: training the variable-length time series data classification network model using a three-stage meta-learning training process, wherein the three-stage meta-learning training process includes a meta-learning training process using an open source data set and a fault simulation enhancement stage using the fault simulation enhancement data set; S109: Apply the trained variable-length time series data classification network model to the actually collected small sample fault data set to perform small sample task training.

2. The method according to claim 1, characterized in that When the simulation modeling is performed in step S101, the following steps are included: S1011: Build a system framework to determine the interactions among the subsystems in the system and the structural relationships among the components in the system; S1012: establishing a dynamic equation of the hydraulic system, in which the physical properties of the hydraulic oil are considered; S1013: Determine component parameters, where the components include a hydraulic pump, a hydraulic cylinder, and a one-way valve pressure sensor; S1014: Establishing a control system model according to the control strategy of the controllable pitch propeller; S1015: Mechanical system modeling, establishing kinematic equations and dynamic equations of the mechanical system, wherein the mechanical system includes a blade rotating mechanism of the controllable pitch propeller.

3. The method according to claim 2, characterized in that In step S103, when constructing the fault simulation enhanced data set, the following steps are included: S1031: Based on the physical properties of hydraulic oil, establish the change curve of hydraulic oil density and viscosity at different temperatures; S1032: During the simulation, the density and viscosity parameters of the hydraulic oil are adjusted according to the set random temperature to simulate the changing hydraulic system performance under actual working conditions; S1033: Simulating hardware setting differences caused by manual operation by randomly setting the working pressure of the one-way valve within a predetermined range; S1034: Generate a large number of fault data sets by introducing the sensor noise and considering the hydraulic oil parameters at different temperatures and the system parameter changes under different personnel operations.

4. The method according to claim 3, characterized in that In step S105, the residual connection multi-mode encoder is a multi-mode encoder based on time series residual connection. The convolution kernel in the multi-mode encoder adopts multi-mode parallel connection and keeps the inter-layer feature length unchanged. The last layer of the multi-mode encoder performs global average pooling to map the variable-length time series data to a global feature vector of fixed length.

5. The method according to claim 4, characterized in that The global feature vector is averaged along the time dimension by the feature vector of each channel extracted by the convolutional network through the global averaging method: X={x1,x2,…,x T } in, is the global eigenvector, x t Represents the feature vector of time step t, T is the total time step, and X is the time series.

6. The method according to claim 5, characterized in that In step S105, the triplet loss function is used to evaluate the mapping effect of the classification network model on different categories, and the triplet loss function is: L triplet (a,p,n)=max(||f(a)-f(p)||-||f(a)-f(n)||+α,0) Among them, a is the anchor sample, p is the positive sample, n is the negative sample, f(a) is the feature vector of anchor sample a, f(p) is the feature vector of positive sample p, f(n) is the feature vector of negative sample n, and α is the learning rate of the meta-learner.

7. The method according to claim 6, characterized in that In step S107, the meta-learning training process using the open source data set uses the data in the open source data set to construct a small sample task meta-training task, and improves the ability of the variable-length time series data classification network model to extract common features and pressure flow detection point data features in time series data through training.

8. The method according to claim 7, characterized in that In step S107, the open source data sets include the UCRArchive data set, the ZeMA hydraulic bench data set and the Rolf Isermann centrifugal pump data set, wherein the UCRArchive data set is a widely used and public time series data set; the ZeMA hydraulic bench data set contains multiple sensor data and has multiple sampling rates; the Rolf Isermann centrifugal pump data set is a data set specifically for analyzing centrifugal pump performance and failure modes, including the pressure difference, speed, motor and pump torque of the centrifugal pump and the outlet flow data of the pump.

9. The method according to claim 8, characterized in that In step S107, the small sample task meta-training task regards the network model as a function f with parameter θ θ ,The task support set and task query set of the inner loop training belong to the same distribution. During the training process, the task learner and meta-learner are trained iteratively and share parameters.

10. The method according to claim 9, characterized in that The small sample task meta-training task includes the following steps: S1071: During the training process, the parameters of the task learner are updated using the task support set: S1072: The task query set Input the task learner to calculate the loss; S1073: The loss is regarded as passed to the meta-learner; S1074: After accumulating one or more After that, the parameters of the meta-learner are updated as follows: Among them, θ is a parameter, f θ is a function with parameter θ, is the task support set, is the task query set, p(T) is the distribution, β is the learning rate of the task learner, α is the learning rate of the meta-learner, i is the task index, indicating the current task number, representing each task sample in the meta-learning process, and L is the loss function, which is used to measure the difference between the predicted value output by the task learner and the true label.

Images