A cross-size underwater thruster fault diagnosis method

Abstract

The present application relates to the technical field of underwater equipment fault diagnosis, and in particular to a cross-size underwater propeller fault diagnosis method. The method includes: acquiring and converting the acoustic signal data of the source domain sample data and the target domain sample data, extracting the acoustic frequency domain features; extracting the key frequency features of the acoustic frequency domain features through a feature extractor to obtain the acoustic features, and mapping the speed signal to the speed features through a speed encoder, and then fusing the acoustic features with the speed features to obtain multi-dimensional fault features; inputting the multi-dimensional fault features into a multi-task learning framework to simultaneously train the fault classification model, domain discriminator, and speed prediction model; using the trained multi-task learning framework to identify the acoustic signal data of the underwater propeller in the target domain to obtain its corresponding fault type and confidence. More efficient, accurate, and reliable underwater propeller fault diagnosis can be achieved through this application.

Description

Technical Field

[0001] The present application relates to the technical field of underwater equipment fault diagnosis, and in particular to a cross-size underwater thruster fault diagnosis method. Background Art

[0002] The propulsion system of an underwater robot is its core component, which directly affects the maneuverability, stability and efficiency of the underwater robot, and its reliability directly affects the efficiency and safety of the entire marine system. However, the existing underwater propeller fault diagnosis technology has significant limitations when facing propellers of different sizes and variable speed conditions.

[0003] Traditional methods mainly rely on current and vibration sensors to monitor the operating status of the thruster. As the size and speed of the thruster change, the data collected by these sensors often undergo domain shifts, resulting in a significant decrease in the accuracy of the diagnostic results. Although acoustic sensors have unique advantages in underwater environments, existing technologies still face challenges in processing complex underwater acoustic signals, especially in capturing and retaining key spectral features.

[0004] To overcome these difficulties, researchers began to explore new domain adaptation methods in order to achieve cross-scale knowledge transfer and maintain consistent diagnostic performance under different speed conditions. However, existing domain adaptation methods are not effective in dealing with cross-scale problems and are difficult to meet the fault diagnosis needs of thrusters of different sizes, especially in accurately diagnosing underwater thruster faults under different sizes and speed conditions, and coping with complex and changing working conditions. Summary of the invention

[0005] The embodiment of the present application provides a cross-size underwater thruster fault diagnosis method, which effectively integrates a multi-task learning framework such as fault diagnosis, domain adaptation and speed prediction, has strong adaptability and generalization performance, and can achieve more efficient, accurate and reliable underwater thruster fault diagnosis, providing strong guarantees for the stable operation of underwater robots and the safe development of marine systems.

[0006] To achieve the above objectives, the present application provides a cross-size underwater thruster fault diagnosis method, comprising:

[0007] A fault sample data step, collecting acoustic signal data of underwater thrusters of different sizes controlled by different speed signals and marking their health states, and dividing the acoustic signal data into source domain sample data and target domain sample data according to the size of the underwater thrusters;

[0008] In the signal preprocessing step, the acoustic signal data of the source domain sample data and the target domain sample data are converted and extracted by fast Fourier transform to obtain acoustic frequency domain features;

[0009] The feature extraction step is to extract the key frequency features of the acoustic frequency domain features through a feature extractor to obtain the acoustic features and maps the speed signal into a speed feature through a speed encoder Then the acoustic features With the speed characteristics Fusion to obtain multi-dimensional fault features , , thereby capturing multi-dimensional information of underwater thruster failures;

[0010] The model training step is to input the multi-dimensional fault features into a multi-task learning framework to simultaneously train a fault classification model, a domain discriminator, and a speed prediction model. The fault classification model is used to predict the health status of the underwater propeller; the domain discriminator is used to reduce the distribution difference between the source domain and the target domain, that is, to improve the cross-scale migration capability of the underwater propeller; the speed prediction model is used to predict the rotation speed of the underwater propeller, and the multi-task learning framework is used to jointly optimize the fault classification, domain confrontation and speed prediction loss, so as to further enhance the adaptability of the model to different rotation speeds;

[0011] In the model prediction step, the acoustic signal data of the underwater thruster in the target domain is collected and processed through the signal preprocessing step and the feature extraction step to obtain the multi-dimensional fault features, which are then input into the trained multi-task learning framework to identify the corresponding fault type and confidence.

[0012] In some of the embodiments, the input ends of the fault classification model, the domain discriminator, and the speed prediction model are commonly connected to the multi-dimensional fault features, wherein a gradient reversal layer GRL is arranged in front of the domain discriminator, and the gradient reversal layer GRL is used to not change the input during forward propagation and to invert the gradient during back propagation to achieve domain adversarial training. The embodiments of the present application introduce the gradient reversal layer and the domain discriminator into the multi-task learning framework to prompt the model to learn domain invariant features, thereby improving the adaptability and generalization ability of the model for underwater thrusters of different sizes.

[0013] In some of the embodiments, the fault classification model includes a linear layer and a softmax layer.

[0014] In some of the embodiments, the domain discriminator includes a fully connected layer, a batch normalization layer, an activation function and an output layer connected in sequence, and a gradient reversal layer is connected to the front end of the fully connected layer to obtain multi-dimensional fault features and output them to the fully connected layer.

[0015] In some of the embodiments, the speed prediction model includes a fully connected layer for outputting a predicted rotational speed value as an auxiliary task to improve the adaptability of the model to different operating conditions. The addition of the speed prediction model enables the method to better and more directly adapt to fault diagnosis under different rotational speeds, more comprehensively understand the working state of the underwater thruster, and further improve the practicality and robustness of the model.

[0016] In some embodiments, the feature extractor includes a one-dimensional residual network ResNet1D, and the one-dimensional residual network ResNet1D includes a plurality of residual blocks, each of which is connected to a spectral attention module.

[0017] In some of these embodiments, the spectral attention module includes a global average pooling layer, a first fully connected layer, a second fully connected layer, and corresponding activation functions.

[0018] In some embodiments, the speed encoder includes a fully connected layer fc1 and a fully connected layer fc2, and the encoding process thereof is represented by the following computational model:

[0019] ,

[0020] in, is the input speed signal value, , They are the weights of the fully connected layer fc1 and the fully connected layer fc2, and ReLU is the activation function.

[0021] In some embodiments, the computational model of the spectral attention module is expressed as:

[0022] ,

[0023] in, , X is the original feature map of the input acoustic frequency domain features, GAP is the global average pooling, ReLU is the activation function of the first fully connected layer, and is the weight of the fully connected layer, Used to represent the Sigmoid activation function, which is the activation function of the second fully connected layer.

[0024] In some embodiments, the model training step further includes: constructing a total loss function based on fault classification loss, domain adversarial loss and speed prediction loss for evaluating the training process to ensure that the multi-task learning framework has good cross-scale diagnostic capabilities, and the total loss function is represented by the following calculation model:

[0025] ,

[0026] in, is the fault classification loss, is the domain adversarial loss, Predict loss for speed, and is an adjustable weight parameter.

[0027] In some embodiments, the fault classification loss is calculated based on the following cross entropy loss function model:

[0028] ,

[0029] in, represents the number of source domain sample data, Indicates the number of fault categories, is the true label, is the predicted probability of the fault category.

[0030] In some embodiments, the domain adversarial loss It is calculated based on the following binary cross entropy loss function model:

[0031] ,

[0032] in, , represents the number of samples in the target domain, Indicates the domain label to which the sample belongs, The sample belongs to the source domain, The sample belongs to the target domain, represents the predicted probability of the domain discriminator.

[0033] In some embodiments, the speed prediction loss It is calculated based on the following mean square error loss function model:

[0034] ,

[0035] in, is the speed value of the speed signal, The speed prediction model predicts the output speed value.

[0036] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 is a flow chart of a fault diagnosis method according to an embodiment of the present application;

[0039] Figure 2 It is a principle logic diagram of the fault diagnosis method according to an embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0041] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0042] Unless otherwise defined, the technical terms or scientific terms used in this application shall have the usual meanings understood by persons with ordinary skills in the technical field to which this application belongs. The words "one", "a", "the" and the like used in this application do not indicate a quantitative limitation and may indicate the singular or plural. The terms "first", "second", "third" and the like used in this application are merely used to distinguish similar objects and do not represent a specific ordering of the objects.

[0043] This embodiment provides a cross-size underwater thruster fault diagnosis method. Figure 1-Figure 2 is a flow chart and logic diagram of a cross-size underwater thruster fault diagnosis method according to an embodiment of the present application, such as Figure 1-Figure 2 As shown, the process includes the following steps:

[0044] Fault sample data step S101, using a hydrophone to collect acoustic signal data of underwater propellers of different sizes controlled by different speed signals and mark their health status, and generate corresponding speed signals based on the required rotation speed. For example but not limited to, the hydrophone is set at 0.3 meters from the large-sized and small-sized underwater propellers, the sampling frequency is 64000 Hz, and the sampling time for each rotation speed is set to 3 minutes. The health status of the acoustic signal data at each rotation speed includes the acoustic signal data of healthy, slightly faulty, and seriously faulty underwater propellers, and the acoustic signal data is divided into source domain sample data and target domain sample data according to the size of the underwater propeller. The target domain sample data includes a test set and a verification set, and the rotation speed can be set incrementally between 300 and 900, wherein: the source domain can be used to represent a small-sized underwater propeller, and the target domain refers to a large-sized underwater propeller, wherein small size and large size can further include different sizes, and the underwater propeller size is classified based on a size threshold, or it can refer to a specific size, and the size represented by the source domain and the target domain can also be adjusted;

[0045] In the signal preprocessing step S102, the acoustic signal data of the source domain sample data and the target domain sample data are divided into time windows of length 1 second, and then converted and extracted by fast Fourier transform (FFT) to obtain acoustic frequency domain features. Each time window contains 64,000 data points. The first 512 frequency points are extracted as acoustic frequency domain features, and their speed signals are marked accordingly.

[0046] Feature extraction step S103, extracting the key frequency features of the acoustic frequency domain features through a feature extractor to obtain acoustic features and maps the speed signal into a speed feature through a speed encoder Then the acoustic features With the speed characteristics Fusion, i.e. feature superposition, obtains multi-dimensional fault features , , thereby capturing multi-dimensional information of underwater thruster failures;

[0047] In the model training step S104, the multi-dimensional fault features are input into a multi-task learning framework to simultaneously train a fault classification model, a domain discriminator, and a speed prediction model. The fault classification model is used to predict the health status of the underwater propeller; the domain discriminator is used to reduce the distribution difference between the source domain and the target domain, that is, to improve the cross-scale migration capability of the underwater propeller; the speed prediction model is used to predict the rotation speed of the underwater propeller, and the multi-task learning framework is used to jointly optimize the fault classification, domain confrontation and speed prediction loss, so as to further enhance the adaptability of the model to different rotation speeds;

[0048] In the model prediction step S105, the acoustic signal data of the underwater thruster in the target domain is collected and processed through the signal preprocessing step and the feature extraction step to obtain the multi-dimensional fault features, which are then input into the trained multi-task learning framework to identify the corresponding fault type and confidence. The multi-dimensional feature fusion method can more comprehensively capture the fault information of the underwater thruster, improve the accuracy of diagnosis, and provide reliable decision support for the maintenance and management of the underwater thruster.

[0049] The embodiments of the present application solve the problem of inconsistent data distribution caused by differences in thruster size and changes in speed through comprehensive data collection and labeling, sophisticated signal preprocessing, targeted feature extraction and fusion, application of a multi-task learning framework, and efficient and reliable model prediction, significantly improve the accuracy and generalization capability of cross-size fault diagnosis, and provide strong support for the reliable operation of the underwater propulsion system.

[0050] In some of the embodiments, the input ends of the fault classification model, domain discriminator, and speed prediction model are commonly connected to the multi-dimensional fault features, and each has an independent fully connected layer for prediction of specific tasks, wherein a gradient reversal layer GRL is arranged in front of the domain discriminator, and the gradient reversal layer GRL is used to not change the input during forward propagation and to invert the gradient during back propagation to achieve domain adversarial training. The embodiments of the present application introduce the gradient reversal layer and the domain discriminator into the multi-task learning framework to prompt the model to learn domain invariant features, thereby improving the adaptability and generalization ability of the model for underwater thrusters of different sizes.

[0051] In some of the embodiments, the fault classification model includes a linear layer and a softmax layer.

[0052] In some embodiments, the domain discriminator includes a fully connected layer, a batch normalization layer, an activation function, and an output layer connected in sequence.

[0053] In some of the embodiments, the speed prediction model includes a fully connected layer for outputting a predicted rotational speed value as an auxiliary task to improve the adaptability of the model to different operating conditions. The addition of the speed prediction model enables the method to better and more directly adapt to fault diagnosis under different rotational speeds, more comprehensively understand the working state of the underwater thruster, and further improve the practicality and robustness of the model.

[0054] In some embodiments, the feature extractor includes a one-dimensional residual network ResNet1D, the one-dimensional residual network ResNet1D includes a plurality of residual blocks, the number of which is flexibly adjustable, and each residual block is connected to a spectral attention module, having The overall process of feature extraction can be expressed as:

[0055] ,in Indicates The spectral attention module of the layer, Indicates residual blocks, The acoustic frequency domain characteristics Layer feature maps, based on which the spectral attention module enables feature extraction which can adaptively emphasize fault-related frequency components while suppressing less important or noisy frequencies.

[0056] In the above embodiment, the feature extractor specifically includes: a convolutional layer, a batch normalization layer, a ReLU activation function, a maximum pooling layer, the above residual block and an average pooling layer.

[0057] In some embodiments, the spectral attention module includes a global average pooling layer, a first fully connected layer, a second fully connected layer and a corresponding activation function, wherein the first fully connected layer is a dimension reduction fully connected layer, and its activation function is a ReLU activation function, and the second fully connected layer is a dimension increase fully connected layer, and its activation function is a sigmoid activation function, and the spectral attention module is used to calculate the attention weight , the attention weight is applied to the original feature map X of the acoustic frequency domain feature by element-by-element multiplication to obtain .

[0058] In some embodiments, the speed encoder includes a fully connected layer fc1 and a fully connected layer fc2, and the encoding process thereof is represented by the following computational model:

[0059] ,

[0060] in, is the input speed signal value, , They are the fully connected layer fc1 and the fully connected layer fc2, ReLU is the activation function, and the scalar speed value is mapped to the high-dimensional feature space through the speed encoder. Specifically, the 1-dimensional speed signal value is output as a 64-dimensional feature through the fully connected layer fc1, and then the 64-dimensional feature is output as a 512-dimensional feature through the fully connected layer fc2. The activation function is set between the fully connected layer fc1 and the fully connected layer fc2.

[0061] In some embodiments, the computational model of the spectral attention module is expressed as:

[0062] ,

[0063] in, , X is the original feature map of the input acoustic frequency domain features, GAP is the global average pooling, ReLU is the activation function of the first fully connected layer, and is the weight of the fully connected layer, Used to represent the Sigmoid activation function, which is the activation function of the second fully connected layer.

[0064] In some embodiments, the model training step further includes: constructing a total loss function based on fault classification loss, domain adversarial loss and speed prediction loss, for jointly optimizing the loss function to evaluate the training process, so as to ensure that the multi-task learning framework has good cross-scale diagnostic capabilities, and the total loss function is expressed as the following calculation model:

[0065] ,

[0066] in, is the fault classification loss, is the domain adversarial loss, Predict loss for speed, and The total loss function is used to guide the training process of the multi-task learning framework, ensuring that the model achieves good results in fault classification, cross-size migration and adaptability to different speeds.

[0067] In some embodiments, the fault classification loss is calculated based on the following cross entropy loss function model:

[0068] ,

[0069] in, represents the number of source domain sample data, Indicates the number of fault categories, is the true label, For the predicted probability of the fault category, the cross entropy loss function is used to calculate the loss based on the true label and the predicted probability. The fault classification model is continuously optimized during the training process to make the fault classification more accurate.

[0070] In some embodiments, the domain adversarial loss It is calculated based on the following binary cross entropy loss function model:

[0071] ,

[0072] in, , represents the number of samples in the target domain, Indicates the domain label to which the sample belongs, The sample belongs to the source domain, The sample belongs to the target domain, It represents the prediction probability of the domain discriminator. By calculating the domain adversarial loss of the target domain samples, the domain discriminator learns how to reduce the distribution difference between the source domain and the target domain, thereby enhancing the adaptability to underwater propellers of different sizes. The binary cross entropy loss function is used to calculate the loss according to the domain label to which the sample belongs and the prediction probability of the domain discriminator, so as to encourage the model to learn domain invariant features and improve the generalization ability.

[0073] In some embodiments, the speed prediction loss It is calculated based on the following mean square error loss function model:

[0074] ,

[0075] in, is the speed value of the speed signal, To predict the output speed value of the speed prediction model, the mean square error loss function is used to calculate the loss according to the true speed value and the predicted speed value. The auxiliary model directly learns the relationship between the speed and the fault, thereby improving the accuracy and robustness of the diagnosis.

[0076] In the above embodiment, the training uses the Adam optimizer to update parameters, the initial learning rate is set to 0.01, the batch size is 64, and the number of training rounds is 100 rounds; after each training round, the model performance is evaluated on the validation set, and the model parameters with the best performance are selected; finally, the selected best model is used to perform performance evaluation on the test set, and indicators such as accuracy, precision, recall and F1 score are calculated to comprehensively evaluate the performance of the model in cross-scale fault diagnosis tasks.

[0077] It has been verified experimentally that the embodiments of the present application can effectively realize fault diagnosis of underwater thrusters of different sizes and maintain stable diagnostic performance under different rotation speed conditions. In the migration task from small to large thrusters, the diagnostic accuracy can reach 95.6%; in the migration task from large to small thrusters, the diagnostic accuracy can reach 84.4%. This method is significantly superior to existing domain adaptation methods and provides reliable technical support for predictive maintenance and safe operation of underwater propulsion systems.

[0078] It should be noted that the steps shown in the above process or the flowchart of the attached figure can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here. For example, steps S101 to S104 can also be pre-trained and run separately S105.

[0079] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A cross-size underwater propeller fault diagnosis method, characterized in that: include: A fault sample data step, collecting acoustic signal data of underwater thrusters of different sizes controlled by different speed signals and marking their health states, and dividing the acoustic signal data into source domain sample data and target domain sample data according to the size of the underwater thrusters; The signal preprocessing step converts and extracts the acoustic signal data of the source domain sample data and the target domain sample data to obtain acoustic frequency domain features; The feature extraction step is to extract the key frequency features of the acoustic frequency domain features through a feature extractor to obtain the acoustic features and maps the speed signal into a speed feature through a speed encoder Then the acoustic features With the speed characteristics Fusion, obtain multi-dimensional fault features; A model training step, inputting the multi-dimensional fault features into a multi-task learning framework to simultaneously train a fault classification model, a domain discriminator, and a speed prediction model; In the model prediction step, the acoustic signal data of the underwater thruster in the target domain is collected and processed through the signal preprocessing step and the feature extraction step to obtain the multi-dimensional fault features, which are then input into the trained multi-task learning framework to identify the corresponding fault type and confidence.

2. The cross-size underwater propeller fault diagnosis method according to claim 1 is characterized in that: The input ends of the fault classification model, the domain discriminator, and the speed prediction model are commonly connected to the multi-dimensional fault features, wherein a gradient reversal layer GRL is arranged before the domain discriminator.

3. The cross-size underwater propeller fault diagnosis method according to claim 1 or 2, characterized in that: The feature extractor includes a one-dimensional residual network ResNet1D, which includes multiple residual blocks, and each residual block is connected to a spectral attention module.

4. The cross-size underwater propeller fault diagnosis method according to claim 3 is characterized in that: The spectral attention module includes a global average pooling layer, a first fully connected layer, a second fully connected layer and corresponding activation functions.

5. The cross-size underwater propeller fault diagnosis method according to claim 4 is characterized in that: The speed encoder includes a fully connected layer fc1 and a fully connected layer fc2, and its encoding process is represented by the following calculation model: ， in, is the input speed signal value, , They are the weights of the fully connected layer fc1 and the fully connected layer fc2, and ReLU is the activation function.

6. The cross-size underwater propeller fault diagnosis method according to claim 4, characterized in that: The computational model of the spectral attention module is expressed as: ， in, , X is the original feature map of the input acoustic frequency domain features, GAP is the global average pooling, ReLU is the activation function of the first fully connected layer, and is the weight of the fully connected layer, Used to represent the Sigmoid activation function, which is the activation function of the second fully connected layer.

7. The cross-size underwater propeller fault diagnosis method according to claim 1 or 4, characterized in that: The model training step further includes: constructing a total loss function based on fault classification loss, domain adversarial loss and speed prediction loss for evaluating the training process, wherein the total loss function is represented by the following calculation model: ， in, is the fault classification loss, is the domain adversarial loss, Predict loss for speed, and is an adjustable weight parameter.

8. The cross-size underwater propeller fault diagnosis method according to claim 7, characterized in that: The fault classification loss is calculated based on the following model: ， in, represents the number of source domain sample data, Indicates the number of fault categories, is the true label, is the predicted probability of the fault category.

9. The cross-size underwater propeller fault diagnosis method according to claim 8, characterized in that: The domain adversarial loss Calculated based on the following model: ， in, , represents the number of samples in the target domain, Indicates the domain label to which the sample belongs, The sample belongs to the source domain, The sample belongs to the target domain, represents the predicted probability of the domain discriminator.

10. The cross-size underwater propeller fault diagnosis method according to claim 9, characterized in that: The speed prediction loss Calculated based on the following model: ， in, is the speed value of the speed signal, The speed prediction model predicts the output speed value.

Images