AUV propeller fault diagnosis method based on wavelet entropy and apn

Abstract

The application discloses an AUV propeller fault diagnosis method based on wavelet entropy and APN, and aims to solve the problems of large signal noise interference, inaccurate fault feature extraction and weak generalization ability in the AUV propeller fault diagnosis. The method sequentially comprises the following steps: S1, a discrete multi-layer wavelet is used to decompose an original signal, a wavelet Shannon entropy is used to determine an optimal reconstruction scale, and a single branch reconstruction is used to enhance a signal; S2, the enhanced signal is input into a CNN containing an attention module, and key fault features are focused; and S3, a prototype network optimized by a joint loss is constructed, unlabeled samples are combined with a pseudo-label mechanism, the generalization ability of small samples is improved, and efficient and accurate diagnosis is realized. The application can effectively filter out noise interference, accurately extract fault features, realize efficient and accurate diagnosis of the AUV propeller fault under the condition of small samples, and provide reliable technical support for the safe and stable operation of the AUV propelling system.

Description

Technical Field

[0001] This invention relates to the field of underwater robot fault diagnosis technology, specifically to an AUV thruster fault diagnosis method based on wavelet entropy and APN. Background Technology

[0002] AUVs are indispensable unmanned platforms in marine engineering, scientific research, and national defense. Their propulsion systems, as critical power components, operate under high loads and complex marine environments for extended periods, making them one of the most vulnerable parts of an AUV. Any failure in the propulsion system can lead to performance degradation, mission failure, or even equipment loss. Therefore, real-time and accurate health monitoring and fault diagnosis are of great significance.

[0003] Underwater acoustic signals contain rich information about equipment operating status, making them an effective means of fault diagnosis. However, the inherent strong noise interference in the underwater environment (such as water currents and marine life) often obscures early, weak fault characteristic signals, posing a serious challenge to the accuracy and sensitivity of diagnosis. Traditional signal processing methods often perform poorly in such low signal-to-noise ratio environments.

[0004] In recent years, deep learning has achieved significant success in the field of fault diagnosis, but its performance is highly dependent on massive amounts of precisely labeled training data. In practical applications of AUV thruster fault diagnosis, obtaining labeled samples is extremely costly, often requiring AUV retrieval, disassembly, and verification by domain experts—a time-consuming and labor-intensive process. This results in a very small number of labeled samples for each fault category, creating a typical few-sample problem. This data predicament makes traditional deep learning models difficult to train adequately, prone to overfitting, and exhibiting poor generalization ability, failing to meet the needs of practical applications. Therefore, developing a new method capable of reliable fault diagnosis under conditions of scarce samples and severe noise interference is a crucial technological bottleneck that urgently needs to be overcome in this field. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing AUV thruster fault diagnosis technologies when faced with the dual constraints of small sample sizes and strong noise, and mainly addresses the following technical problems:

[0006] 1. Problem of strong underwater noise interference: To address the challenge of severe underwater noise significantly affecting fault feature extraction, this invention first employs a signal preprocessing technique aimed at enhancing and separating the signal components most relevant to the fault from a strong noise background.

[0007] 2. Problem of extremely scarce fault samples: For scenarios with only a very small number of labeled samples for each type of fault, this invention proposes a diagnostic framework based on a prototype network, which enables it to learn the discriminative features of fault modes from extremely limited data, avoiding the performance degradation of traditional deep learning models due to insufficient data.

[0008] 3. Insufficient feature discrimination: Under small sample conditions, the features learned by the model often have low discriminative power. This invention introduces an attention mechanism to guide the model to automatically focus on key feature regions in the signal, thereby improving the discriminative power of the features and enhancing the accuracy of diagnosis.

[0009] To address the shortcomings of existing technologies, this invention proposes an AUV thruster fault diagnosis method based on wavelet entropy and APN, which is a high-precision fault diagnosis method applicable to autonomous underwater vehicle (AUV) propulsion systems in situations with scarce fault samples and strong noise. This method combines wavelet entropy optimal reconstruction scale signal enhancement technology with a prototype network integrating attention modules to construct an end-to-end AUV thruster fault diagnosis method specifically designed for small sample and strong noise environments.

[0010] To achieve the above objectives, this invention provides an AUV thruster fault diagnosis method based on wavelet entropy and APN. This method diagnoses AUV thruster faults through signal preprocessing and enhancement, attention feature extraction, and small-sample prototype network model construction. The steps include:

[0011] Step S1: Signal preprocessing and enhancement based on wavelet optimal reconstruction scale.

[0012] The purpose of this step is to adaptively denoise and improve the signal-to-noise ratio of the original thruster signal while preserving the main fault characteristics, so as to provide high-quality input for subsequent diagnostic models.

[0013] 1.1 Discrete Multilevel Wavelet Decomposition

[0014] For the collected length is Discretized original time-domain signal conduct Layered discrete wavelet transform. This transform decomposes the signal into subbands of different scales (frequency). For arbitrary decomposition scales... The calculation process is as follows:

[0015] ;

[0016] ;

[0017] in, This represents the original AUV time-domain signal sequence; Indicates the first Approximation coefficients (low-frequency components) at the decomposition scale. Indicates the first Detail coefficients (high-frequency components) at the decomposition scale. Represents the wavelet scaling function; Describe the wavelet basis functions; Indicates the decomposition scale or number of layers; This indicates the position of the coefficient shift over time; This indicates the inner product operation.

[0018] 1.2 Determination of the optimal reconstruction scale based on wavelet Shannon entropy

[0019] The occurrence of faults typically concentrates signal energy within a specific frequency band, leading to increased sparsity of the wavelet coefficient sequence corresponding to that band, i.e., a decrease in information entropy. Using this principle, the optimal scale containing the richest fault information is determined by calculating the Shannon entropy of the detail coefficients at each scale.

[0020] First, each scale detail coefficient sequence Processed into probability distribution sequences

[0021] ;

[0022] In the formula, Indicates the first The first scale Wavelet detail coefficients; This represents the total length of the detail coefficients at this scale; Indicates the first The energy proportion of each coefficient can be regarded as its probability distribution value.

[0023] Then, according to information theory, the Shannon entropy at this scale is calculated.

[0024] ;

[0025] This represents the wavelet entropy value at the current decomposition scale. The smaller the entropy value, the stronger the sparsity of the signal at that scale, and the more deterministic information (i.e., fault characteristics) it contains.

[0026] 1.3 Wavelet Single-Branch Reconstruction

[0027] By calculating all Find the scale with the smallest wavelet entropy at each scale. :

[0028] ;

[0029] This scale The optimal reconstruction scale is determined. Once the optimal reconstruction scale is selected, coefficients from all other scales are discarded, and only the detail coefficients from that scale are used. The enhanced time-domain signal is obtained by reconstructing the signal using inverse discrete wavelet transform. This signal filters out a large amount of noise from irrelevant frequency bands and highlights key fault characteristics.

[0030] Step S2: Deep fault feature extraction based on attention mechanism.

[0031] Will enhance the signal Input a deep convolutional neural network (CNN) as a feature extractor To enable the network to automatically learn and focus on the most discriminative parts of the signal, an attention module is embedded after the key convolutional layers of the CNN. The attention module evaluates the importance of each neuron in the feature map by defining an energy function. For a target neuron... and other neurons in the same channel Its energy function A closed-form solution can be derived by minimizing the energy function:

[0032] ;

[0033] in, Represents the current neuron; This represents the local mean, excluding the target neuron, centered on the target neuron. The average activation value of other neurons within the spatial neighborhood; In addition to the target neuron, the focus is on the target neuron. The variance of other neurons within the spatial neighborhood; for The total number of neurons in the spatial neighborhood; This is a regularization hyperparameter, whose main function is to prevent the denominator from being zero and to ensure the stability of the calculation. This is a variance scaling hyperparameter used to adjust the influence of neighborhood variance on the energy function.

[0034] Energy value The lower the value, the higher the distinction between this neuron and its surrounding neurons, and the more important the information it contains. Therefore, the importance weight of a neuron can be defined as... Ultimately, the attention module processes the original feature map... Element-wise scaling is performed to obtain the enhanced feature map. :

[0035] ;

[0036] in, Represents the original feature map of the input; This represents the energy values ​​of all neurons. The energy diagram is composed of; This indicates element-wise multiplication; This is the activation function used to smooth the weights.

[0037] Step S3: Construct a prototype network diagnostic model based on joint loss optimization.

[0038] A prototype network is used to classify attention-enhanced features in few samples, and the model training is jointly optimized.

[0039] 3.1 Category Prototype Construction and Metric Learning

[0040] The core idea of ​​this invention is to learn a metric space in which similar samples are aggregated and dissimilar samples are separated. For each fault category... Given a containing Support set of labeled samples Using a feature extractor Will Each sample in Mapped to In a dimensional metric space, its feature embedding is obtained. The prototype of this category The mean value of the feature embeddings of its support set samples is obtained by calculating:

[0041] ;

[0042] in, For category The prototype vector is the central representation of this class in the metric space. For category Support set.

[0043] 3.2 Probabilistic Classification and Prototype Network Loss Calculation

[0044] For a query sample to be classified Calculate its feature embedding Then, by calculating the embedding with all category prototypes... The squared Euclidean distance between them is calculated, and the Softmax function is applied to determine which category it belongs to. probability distribution:

[0045] ;

[0046] in, Indicates query sample Predicted as category The probability of; Representing vectors and The squared Euclidean distance between them.

[0047] sample The predicted class is the class with the highest probability. The training objective of the prototype network is to maximize the probability that the query sample belongs to its true class, which is equivalent to minimizing its negative log-likelihood loss, i.e., the prototype network loss. :

[0048] ;

[0049] 3.3 Calculation of Total Loss Function

[0050] To further enhance the feature extractor To improve the discriminative ability, a parallel auxiliary classifier is introduced. This classifier is a standard fully connected layer plus a softmax layer. It directly classifies the feature embeddings and uses cross-entropy to calculate its loss, which is the auxiliary classifier loss. :

[0051] ;

[0052] in, It is the predicted probability of the auxiliary classifier. It is a one-hot encoding of the true labels. Ultimately, the model's total loss function... It is a weighted sum of the prototype loss and the auxiliary loss:

[0053] ;

[0054] in, The contribution used to balance the two types of losses.

[0055] During training, this total loss function is minimized using the gradient descent algorithm, and the feature extractor is updated collaboratively. and auxiliary classifier The parameters are used to complete the training of the diagnostic model. For unlabeled samples, if their highest predicted probability exceeds a preset threshold, a pseudo-label is assigned to them, and they are temporarily added to the corresponding support set to assist in updating the prototype, thereby effectively utilizing the information of unlabeled data and improving the model's generalization ability in small sample scenarios.

[0056] Advantages and beneficial effects of the present invention:

[0057] This invention provides a stable and reliable AUV thruster fault diagnosis method under small sample conditions by deeply integrating wavelet entropy signal enhancement technology with a prototype network integrating an attention module. This method not only achieves good diagnostic accuracy in small sample scenarios where only a few labels are available for each type of fault, based on the prototype network, but also significantly enhances the discriminative power of fault features using an attention mechanism. Furthermore, the wavelet entropy optimization preprocessing step at the front end ensures the model's robustness in underwater high-noise environments, while the introduction of an unlabeled learning mechanism efficiently utilizes readily available unlabeled data, significantly reducing the reliance on expensive manual annotation. This makes it both high-precision and economical in practical applications, possessing significant technical advantages and application value in the field of AUV thruster fault diagnosis, as reflected in:

[0058] In the signal preprocessing stage, the optimal reconstruction scale is selected by combining discrete multi-layer wavelet decomposition with wavelet Shannon entropy, which can adaptively filter out irrelevant frequency band noise, effectively improve the signal-to-noise ratio, and provide high-purity data support for subsequent fault diagnosis. In the feature extraction stage, an attention module is innovatively embedded in the key layer of CNN. The importance of neurons is quantified according to the energy function and the feature map is scaled element by element, which strengthens the identification and extraction of key fault features and breaks through the bottleneck of redundant information interference in traditional feature extraction.

[0059] To address the industry pain point of high difficulty and scarcity in collecting AUV fault samples, a prototype network is used to construct category prototype vectors. This enables effective fault classification without the need for massive amounts of labeled samples, significantly reducing the dependence on sample quantity. Simultaneously, a pseudo-label strategy is employed to fully exploit the value of unlabeled data, further enhancing the model's generalization ability in small-sample scenarios. The model training phase utilizes a joint optimization mechanism of "prototype network loss + auxiliary classifier loss" to supervise model training from multiple dimensions, accelerating model convergence, enhancing the model's stability and anti-interference capabilities under complex operating conditions, and ensuring reliable diagnostic results.

[0060] At the engineering application level, all technical aspects adopt lightweight algorithm design, which can run efficiently on conventional embedded hardware, meet the needs of real-time monitoring of AUV thruster faults, identify potential faults in advance and issue early warnings, reserve sufficient time for operation and maintenance, reduce downtime due to faults, reduce operation and maintenance costs, and ensure the safe and stable operation of AUVs in various underwater operations, combining technical practicality and practical application value. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method of the present invention;

[0062] Figure 2 This is a flowchart of the signal preprocessing process of the present invention;

[0063] Figure 3 This is a schematic diagram of the feature extractor network structure of the present invention;

[0064] Figure 4 Flowchart for training the prototype network diagnostic model;

[0065] Figure 5 The original signal waveforms from multiple sources under the "propeller blade wear" condition;

[0066] Figure 6 The scale coefficients after performing 4-level wavelet decomposition on the "propeller blade wear" signal;

[0067] Figure 7 This describes a specific implementation method for the multi-source data fusion process. Detailed Implementation

[0068] The technical solution of the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0069] Example 1

[0070] This embodiment aims to verify the effectiveness of the method described in this invention. The fault data in this embodiment was obtained in the laboratory.

[0071] (1) Fault data acquisition and setting

[0072] Five common propeller states were set up: healthy state, propeller blade fracture, propeller blade deformation, propeller blade wear, and propeller blade entanglement. For each state, three one-dimensional time-series signals were collected: underwater acoustic signals, propeller current signals, and hull vibration signals collected by different sensors. Figure 5 The waveform of the multi-source raw signal under the condition of "propeller blade wear" is shown.

[0073] (2) Signal preprocessing and multi-source sample construction

[0074] The collected multi-channel signals under various states are preprocessed to obtain sample data suitable for small-sample learning. This process strictly follows the technical solution of this invention:

[0075] 1) Signal Enhancement: First, each original one-dimensional time series signal is input into a signal preprocessing module based on the optimal wavelet reconstruction scale. The core of this module is the discrete wavelet transform. For example... Figure 6 The figure shows the scale coefficients obtained after performing four-level wavelet decomposition on the "propeller blade wear" signal. This invention determines the optimal reconstruction scale by calculating the wavelet Shannon entropy of these coefficients and performs single-branch reconstruction to obtain the enhanced signal.

[0076] 2) Standardization: Perform Min-Max standardization on all enhanced signals to map data values ​​to the [0,1] interval.

[0077] 3) Data truncation: The standardized signal is truncated into equal parts using a sliding window method. In this embodiment, the window length is set to 512 data points.

[0078] 4) Data Reconstruction: The three signal segments corresponding to each window position are spliced ​​together to form a data structure with a length of 1536 (512). 3) Multi-source sample data. For example... Figure 7 The data reconstruction process is shown.

[0079] Through the above steps, a total of 2000 multi-source sample data with clear status labels were generated.

[0080] (3) Dataset partitioning

[0081] To simulate a few-shot learning scenario, the 2000 sample data points were divided according to the following proportions:

[0082] 1) Training set: accounting for 70%, of which 5 samples of each type of fault are randomly selected as the support set (a total of 25 labeled samples), and the remaining samples are used as the unlabeled set.

[0083] 2) Validation set: accounting for 15%, used for model tuning.

[0084] 3) Test set: accounting for 15%, used for final performance evaluation.

[0085] (4) Model training and testing

[0086] The pre-divided dataset is input into the semi-supervised prototype network diagnostic model based on joint loss optimization described in this invention for training. The model is prototyped on the support set and semi-supervised optimized using a large amount of unlabeled data. After 200 rounds of iterative training, it is tested on an independent test set.

[0087] (5) Implementation results

[0088] The diagnostic model in this embodiment achieved a fault identification accuracy of 9.5% on the test set, demonstrating the effectiveness and superiority of the method in small sample and multi-source signal fusion scenarios. It can quickly and accurately diagnose various specific fault modes of AUV propellers, such as blade breakage, deformation, and wear.

[0089] This embodiment verifies the effectiveness of the method by collecting underwater acoustic, current, and vibration signals from an AUV propeller in five states (including healthy and four types of faults) in a laboratory setting. The signals were preprocessed according to the invention's scheme: enhanced by a wavelet optimal reconstruction scaling module, then standardized and truncated using a sliding window (512 points). The three truncated segments were then spliced ​​together to form multi-source samples, generating a total of 2000 labeled samples. The dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training set used only five samples from each class as the support set to simulate a small-sample scenario. The data was then fed into a semi-supervised prototype network with joint loss optimization for 200 rounds of training before testing. The test results confirm the significant advantages of this invention: multi-source fusion and wavelet entropy preprocessing ensure data quality; attention feature extraction enhances fault information; the prototype network adapts to small-sample scenarios; and joint loss optimization improves model performance, ultimately achieving efficient fault diagnosis. This demonstrates the effectiveness of the method in small-sample, multi-source scenarios.

Claims

1. A wavelet entropy and APN-based AUV propeller fault diagnosis method, characterized in that: The method comprises the following steps: S1: signal preprocessing and enhancement based on wavelet optimal reconstruction scale; S2: deep fault feature extraction based on attention mechanism, input the enhanced time domain signal into a deep convolutional neural network (CNN), and embed an attention module after a key convolutional layer of the CNN; the attention module evaluates the importance of neurons in a feature map through an energy function; S3: prototype network diagnosis model construction based on joint loss optimization, specifically comprising: S3.1: Category prototype construction and metric learning; in said step S3.1: Category prototype construction and metric learning, for each failure category , given a support set containing labeled samples ; each sample in the support set is mapped by a feature extractor to a metric space of dimensionality , resulting in its feature embedding ; the prototype of this category is obtained by computing the mean of the support set samples' feature embeddings ​​​ ； wherein, is a prototype vector of the class , is a support set of the class ; S3.2: Probabilistic Classification and Prototype Network Loss Calculation; In step S3.2: probabilistic classification and prototype network loss calculation, for the query samples that need to be classified... Calculate its feature embedding By calculating the embedding with all category prototypes The squared Euclidean distance between the two values ​​is used to apply the Softmax function to determine which category the query sample belongs to. probability distribution: ； wherein, denotes a query sample predicted to be of class ; denotes the squared Euclidean distance between vectors and ; Prototype network loss is the negative log-likelihood loss: ； S3.3: total loss function calculation; in the step S3.3: total loss function calculation, an auxiliary classifier is introduced The auxiliary classifier is a full connection layer plus a Softmax layer, which classifies the feature embedding and calculates the auxiliary classifier loss by cross-entropy : ； where, is the prediction probability of the auxiliary classifier, is the one-hot encoding of the true label; finally, the total loss function of the model is the weighted sum of the prototype loss and the auxiliary loss: ； wherein, contribute to balancing the two losses; the total loss function is minimized by a gradient descent algorithm to update the feature extractor and the auxiliary classifier parameters, complete the diagnosis model training; for unlabeled samples, if their highest prediction probability exceeds a preset threshold, assign a pseudo label to them and temporarily add them to the corresponding support set to update the prototype.

2. The wavelet entropy and APN-based AUV thruster fault diagnosis method according to claim 1, characterized in that: The step S1 specifically comprises: S1.1: discrete multi-layer wavelet decomposition; S1.2: determination of the optimal reconstruction scale based on wavelet Shannon entropy; S1.3: single branch reconstruction of the wavelet.

3. The AUV thruster fault diagnosis method based on wavelet entropy and APN according to claim 2, characterized in that: In the step S1.1: Discrete multi-level wavelet decomposition is performed on the discrete AUV thruster raw time-domain signal with length Layer discrete wavelet transform decomposes the signal into sub-bands of different scales. For any decomposition scale , the calculation process is as follows:​​ ； ； wherein, denotes the original AUV time-domain signal sequence; denotes the approximation coefficients at the decomposition scale denotes the approximation coefficients at the decomposition scale denotes the detail coefficients at the decomposition scale denotes the detail coefficients at the decomposition scale denotes the wavelet scaling function; denotes the wavelet basis function; denotes the decomposition scale or level; denotes the time shift position of the coefficients; denotes the inner product operation; In the step S1.2, based on the optimal reconstruction scale determination of wavelet Shannon entropy, first, the detail coefficient sequence of each scale is processed into a probability distribution sequence: ​ ； In the formula, denotes the first denotes the first wavelet detail coefficient at the scale denotes the total length of the detail coefficients at the scale denotes the energy proportion of the first coefficient; that is, the probability distribution value; and the Shannon entropy of the scale is calculated again : ； is the wavelet entropy value for the current decomposition scale; In the step S1.3: single branch reconstruction of wavelet, by calculating the wavelet entropy of all scales, the scale with the minimum entropy value is found :​ ； The scale That is, the optimal reconstruction scale, discard all other scale coefficients, only use the optimal reconstruction scale detail coefficient , through the inverse discrete wavelet transform of signal single branch reconstruction, get enhanced time domain signal .

4. The AUV thruster fault diagnosis method based on wavelet entropy and APN according to claim 3, characterized in that: In the step S2, the energy function is: ； in, Represents the current neuron; This represents the local mean, that is, the mean centered on the target neuron, excluding the target neuron itself. The average activation value of other neurons within the spatial neighborhood; In addition to the target neuron, the focus is on the target neuron. The variance of other neurons within the spatial neighborhood; for The total number of neurons in the spatial neighborhood; For regularization hyperparameters, The importance weights of neurons are defined as variance scaling hyperparameters. The attention module processes the original feature map. Element-wise scaling is performed to obtain the enhanced feature map. : ； wherein, represents an input raw feature map; represents an energy map composed of energy values of all neurons; represents an input raw feature map; represents an element-wise multiplication; is an activation function.

Images