Unmanned ship fault diagnosis method based on combination of unsupervised learning and supervised learning

Abstract

The invention discloses an unmanned ship fault diagnosis method based on unsupervised and supervised learning combination, and the method comprises the following steps: 1, building an unmanned ship thrust prediction model, the unmanned ship thrust prediction model comprises an unsupervised neural network, a supervised neural network and a bridging network for connecting the unsupervised neural network and the supervised neural network; 2, training the unmanned ship thrust prediction model to obtain a trained unmanned ship thrust prediction model; 3, collecting real-time state data of the unmanned ship, and inputting the real-time state data into the trained unmanned ship thrust prediction model to obtain an unmanned ship thrust prediction result; and 4, obtaining a final residual error according to a thrust prediction result of the unmanned ship, calculating a T2 statistical magnitude by using the final residual error, and determining whether the unmanned ship has a fault according to the T2 statistical magnitude.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned boats, and in particular to an unmanned boat fault diagnosis method based on a combination of unsupervised and supervised learning. Background Art

[0002] Unmanned aerial vehicles (UAVs), as important tools for modern ocean monitoring and autonomous navigation, are widely used in ocean patrols, target search, environmental exploration, and other fields. The operation of UAVs depends on multiple subsystems (such as power systems, navigation systems, and sensor systems). Failures in these systems can lead to mission failure or even accidents.

[0003] Traditional fault diagnosis methods are mainly divided into several types, including model-based methods, signal processing methods, data-driven methods, and hybrid methods. However, due to the variable operating conditions of unmanned vehicles in complex marine environments, traditional methods have certain limitations in practical applications. In particular, when data is scarce and annotation is difficult, the accuracy and robustness of fault diagnosis cannot be guaranteed. Summary of the Invention

[0004] Purpose of the invention: In order to solve the problem that existing fault diagnosis methods are not applicable to the fault diagnosis of unmanned boats operating in complex marine environments, the present invention proposes an unmanned boat fault diagnosis method based on the combination of unsupervised and supervised learning.

[0005] Technical solution: A fault diagnosis method for unmanned boats based on a combination of unsupervised and supervised learning, including the following steps:

[0006] Step 1: Building an unmanned boat thrust prediction model, which includes an unsupervised neural network, a supervised neural network, and a bridge network for connecting the unsupervised neural network and the supervised neural network;

[0007] Step 2: Train the unmanned boat thrust prediction model to obtain a trained unmanned boat thrust prediction model;

[0008] Step 3: Collect the real-time status data of the unmanned boat, input the real-time status data into the trained unmanned boat thrust prediction model, and obtain the unmanned boat thrust prediction result;

[0009] Step 4: According to the prediction results of the UAV thrust, the final residual is obtained and the T is calculated using the final residual. 2 Statistics, according to T 2 Statistics are used to determine whether the unmanned boat has a fault.

[0010] Furthermore, the unsupervised neural network is based on H unsp (Θ) is obtained by constructing the neural network architecture. unsp (Θ) refers to the neural network architecture with hyperparameter Θ.

[0011] Furthermore, the supervised neural network is based on H sp (Θ) is obtained by constructing the neural network architecture. sp (Θ) refers to the neural network architecture with hyperparameters.

[0012] Furthermore, the bridge network for connecting the unsupervised neural network and the supervised neural network is expressed as:

[0013]

[0014]

[0015] in I represents the identity matrix, H(·) is the variation function used to change the auxiliary variable a u Under the action of , the inverse conversion between supervised and unsupervised is realized.

[0016] Furthermore, the unmanned boat fault diagnosis model is trained to obtain the trained unmanned boat fault diagnosis model, and the specific operations include:

[0017] Use the unit time delay operator z -1 , define a unit time delay operator on the input u(k+s) and output y(k+s) to obtain the input of the unsupervised neural network:

[0018] u(k+s-1)=z -1 u(k+s)

[0019] …

[0020]

[0021] y(k+s-1)=z -1 y(k+s)

[0022] …

[0023]

[0024] Where k is the discrete time, y(·) represents the output variable driven by data at time s, s represents any time, and s p Indicates the length of time delay;

[0025] Thus we get the composite vector z of the past time-delayed input and output p , expressed as:

[0026]

[0027] Where,

[0028]

[0029] And get the output variable y at the current time s s (k), expressed as:

[0030] y s (k) = γ x L p z p (k)+γ u u s (k)+Υ e e s (k)

[0031] Where, γ x for The compound operation of γ u for e is the random noise variable inside and outside the system, γ e for L p is the parameter for adjusting the error;

[0032] By minimizing the loss function L of the unsupervised neural network unsp Get the optimal hyperparameter Θ * :

[0033]

[0034] Where u s represents the current input variable, Θ z Neural network parameters related to past inputs and outputs, used to extract state features from input data or model state dynamics, Θ u Neural network parameters related to the current input variables, used to process control signals or external inputs, Θ y The neural network parameters associated with the current output variable, used to generate output predictions;

[0035] Represents an unsupervised joint transformation function with parameters Θ z 、Θ u 、Θ y The constructed neural network architecture is used to generate a joint representation of the system behavior by fusing information about input, state, and output;

[0036]

[0037] Use H unsp (Θ * ) to obtain the unsupervised residual signal, which is expressed as:

[0038]

[0039] By minimizing the loss function L of the supervised neural network sp Get the optimal hyperparameter Θ * :

[0040]

[0041] Use H sp (Θ * ) to obtain the supervised residual signal

[0042]

[0043] By minimizing the loss function L of the bridge network ext , get the trained bridge network:

[0044]

[0045] Use the trained bridge network to output thrust prediction values:

[0046]

[0047] Furthermore, the final residual is expressed as:

[0048]

[0049] Where Y s Represents the actual output obtained by the sensor;

[0050] The final residual is used to calculate T 2 Statistics:

[0051]

[0052] Where, represents the transpose of the final residual, and E(·) represents the expectation;

[0053] According to T 2 Statistics are used to determine whether the unmanned boat has a fault. The specific operations include:

[0054] like If it is fault-free, it means there is a fault, otherwise, J th Indicates T 2 The threshold value of the statistic.

[0055] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0056] (1) Improve the accuracy of fault diagnosis: By integrating the advantages of unsupervised learning and supervised learning, the present invention can effectively extract fault features and perform accurate diagnosis even when data annotation is insufficient. Experiments show that the fault diagnosis accuracy of the present invention can reach over 95%, which is significantly higher than that of traditional methods.

[0057] (2) Enhanced robustness of fault diagnosis: The unsupervised learning module can adapt to unknown fault types, while the supervised learning module can improve the diagnostic accuracy of known faults. The combination of the two makes the present invention more robust in complex marine environments and can cope with various interference factors (such as waves, wind speed changes, etc.);

[0058] (3) Reduce false alarm rate: Through the residual fusion module and dynamic threshold setting, the method of the present invention can effectively distinguish normal noise from real faults and reduce the false alarm rate. Experiments show that the false alarm rate of the present invention is less than 5%, which is significantly better than the traditional method.

[0059] (4) Reduce data annotation costs: The unsupervised learning module does not rely on fault label data and can perform fault diagnosis even when data annotation is insufficient, significantly reducing data annotation costs.

[0060] (5) Real-time fault detection: The method of the present invention can monitor the operating status of the unmanned boat in real time and complete the detection within 10 seconds after the fault occurs, ensuring the safe operation of the unmanned boat. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a schematic diagram of fault injection;

[0062] Figure 2 The output difference of the unmanned boat thruster in normal and faulty states and the fault detection result obtained by the method of the present invention;

[0063] Figure 3 It is a schematic diagram of the change of T2 statistics on the time axis;

[0064] Figure 4 This is a schematic diagram of the neural network regression results;

[0065] Figure 5 A schematic diagram comparing the fault detection rates of different models;

[0066] Figure 6 is the ROC curve;

[0067] Figure 7 Schematic diagram for comparison of residuals of different models;

[0068] Figure 8 Schematic diagram of the false alarm rate of different models. DETAILED DESCRIPTION

[0069] The technical solution of this embodiment will now be further described with reference to the accompanying drawings.

[0070] This embodiment proposes a fault diagnosis method for unmanned underwater vehicles based on a combination of unsupervised and supervised learning. By integrating the advantages of unsupervised and supervised learning, the accuracy and robustness of fault diagnosis are improved. In particular, when data annotation is insufficient, the method can effectively detect and diagnose faults of unmanned underwater vehicles. The method specifically includes the following steps:

[0071] Step 1: Collect the state value u and output result y of the system. In this embodiment, the state value u is a combination of vibration signal, current, input signal, and thrust.

[0072] Use the unit time delay operator z -1 , define the unit time delay operator on u(k+s) and y(k+s) to obtain the input of the unsupervised neural network:

[0073] u(k+s-1)=z -1 u(k+s)

[0074] …

[0075]

[0076] y(k+s-1)=z -1 y(k+s)

[0077] …

[0078]

[0079] Where k is the discrete time, y(·) represents the output variable driven by data at time s, s represents any time, and s p Indicates the length of the time delay.

[0080] The meaning of the above formula is that since the current state x is unknown, it needs to be estimated using the past.

[0081] Thus we get the composite vector z of the past time-delayed input and output p , expressed as:

[0082]

[0083] Where,

[0084]

[0085] And get the output variable y at the current time s s (k), which is composed of past input and output, current input variables and residuals, is expressed as:

[0086] y s (k) = Υ x L p z p (k)+γ u u s (k)+Υ e e s (k)

[0087] Where, x for Compound operation, Υ u for e is the random noise variable inside and outside the system, Υ e for The purpose is to unify the dimensions of different variables, L p is the parameter for adjusting the error.

[0088] Step 2: H unsp (Θ) is an unsupervised neural network constructed by the neural network architecture. unsp (Θ) refers to the neural network architecture with hyperparameters Θ, and the subscripts of Θ refer to their corresponding outputs.

[0089] Constructing the loss function L of the unsupervised neural network unsp for:

[0090]

[0091] Where u s represents the current input variable, Θ z Neural network parameters related to past inputs and outputs, used to extract state features from input data or model state dynamics, Θ u Neural network parameters related to the current input variables, used to process control signals or external inputs, Θ y The neural network parameters associated with the current output variable, used to generate output predictions;

[0092] Represents an unsupervised joint transformation function with parameters Θ z 、Θ u 、Θ y The constructed neural network architecture is used to generate a joint representation of the system behavior by fusing information about input, state, and output;

[0093] By minimizing L unsp ,Right now:

[0094]

[0095] Get the optimal Θ * ;

[0096] Use H unsp (Θ * ) to obtain the unsupervised residual signal, expressed as

[0097]

[0098] Step 3: Build a H sp (Θ) supervised neural network, H sp (Θ) is a parameterized supervised model used to fuse system dynamic information and generate residual signals, ultimately achieving fault diagnosis of nonlinear dynamic systems. When , the loss function formula is as follows:

[0099]

[0100] By minimizing L sp Optimization parameters:

[0101]

[0102] Generate the following supervised residual signal

[0103]

[0104] Step 4: Define the bridge network H(Θ ext ) and auxiliary converter a u , used to connect unsupervised and supervised residuals, and the mapping relationship is:

[0105]

[0106] in I represents the identity matrix, H(·) is the variation function used to change the auxiliary variable a u Under the action of , the inverse conversion between supervised and unsupervised is realized;

[0107] By training the bridge network H(Θ ext ), minimize the loss function L ext :

[0108]

[0109] Using the trained bridge network H(Θ ext ), output predicted thrust Generate the final residuals:

[0110]

[0111] Where Y s Represents the actual output obtained by the sensor;

[0112] Step 5: Calculate the improved T according to the following formula 2 Statistics:

[0113]

[0114] In the formula, in the formula, represents the transpose of the final residual, E(·) represents the expectation, and is the calculation formula of the characteristic matrix.

[0115] Determine T 2 The threshold of the statistic is expressed as:

[0116]

[0117] Where, J th Indicates T 2 The threshold of the statistic, Pr(·) represents the probability; the overall meaning of this formula is: the probability of the part exceeding the threshold is α, α is a small probability set, so the threshold corresponding to different α can be obtained. Here, T 2 The probability that the statistic is less than the threshold is 1-α, and it can be considered that no fault has occurred (no fault). If yes, it means there is no fault, otherwise there is a fault.

[0118] The unsupervised neural network, supervised neural network, and bridge network mentioned above are all trained according to the following steps:

[0119] Simulation data, including normal and fault states, was obtained. The data was divided into training, validation, and test sets according to a 6:2:2 ratio. The fault mode was derived using the following procedures: a time series length of 100 seconds, a sampling interval of 0.1 seconds, and a fault injection period of 30 to 60 seconds. During the fault phase, the thrust coefficient decreased and the contact resistance increased, which resulted in current anomalies. Uncertainty was added to the inference signal to simulate thrust failures. Random and pulse signals were added to the current signal to simulate transient current shocks. An exponentially decaying sinusoidal signal was added to the vibration signal to simulate vibration shocks caused by the fault.

[0120] The fault state involved in this embodiment shows a significant reduction in thrust, simulating the dynamic characteristics of an actual fault, current anomalies simulating transient current shocks, and increasing high-frequency vibrations to simulate mechanical vibrations caused by faults. By injecting different types of faults, the system's fault detection capabilities can be tested. In addition, fault injection takes into account complex factors in actual working conditions such as transient response, noise, and pulses, making the simulated data closer to the real situation. Specific fault injection is as follows: Figure 1 shown.

[0121] Figure 2The output difference of the unmanned boat thruster in normal and faulty states and the fault detection results obtained by the method of this embodiment are shown. Figure 2 As shown in the figure, within the time range of 0-100 seconds, the thruster output remains stable in the normal state; however, in the fault state, the output fluctuates significantly at a specific time point (such as around 80 seconds), indicating that the system is subject to abnormal interference. The fault diagnosis simulation using this system is reliable.

[0122] The change of T2 statistic on the time axis is as follows Figure 3 As shown, when the statistic exceeds the preset threshold (e.g., around 80 seconds), the system triggers the "Fault Detected" flag, successfully identifying the fault. The residual signal shows a significant trend away from zero when the fault occurs; the T2 statistic reaches a peak of 11.1967 at 92.1 seconds, far exceeding the threshold, further verifying the effectiveness of this embodiment's method.

[0123] By analyzing the results of each neural network regression task, we found that the model's predicted values on the training set, test set, and validation set were highly consistent with the actual values, with R values close to 1. The slope of the fitting curve was close to 1, indicating a very strong linear relationship between the model's predicted values and the actual values.

[0124] The fault detection rate (FDR) is a measure of the model's ability to correctly identify faults. Experimental results show that the model proposed in this embodiment has the highest FDR, reaching 94.65%, indicating that it has high accuracy in fault detection tasks. The supervised model has an FDR of 77.31%, which is second best, while the unsupervised model has the lowest FDR, which is only 23.88%. This result shows that the model of the present invention can more effectively capture fault features by combining the advantages of unsupervised and supervised learning, thereby significantly improving the fault detection rate. The supervised model relies on labeled data and can learn the explicit features of the fault, but its performance is limited by the quality and quantity of the training data. Due to the lack of labeled information, the unsupervised model only performs fault detection through data reconstruction or clustering methods, and its performance is poor, especially in complex fault modes.

[0125] The supervised model achieved the highest AUC value of 0.876, and its ROC curve was closest to the upper left corner, indicating optimal classification performance in distinguishing faulty and normal data. The proposed model achieved an AUC value of 0.832, slightly lower than the supervised model, but its ROC curve still demonstrated good classification performance. The unsupervised model achieved the lowest AUC value, indicating poor classification performance, approaching random guessing.

[0126] Although supervised models offer the best AUC, they rely on large amounts of labeled data, which can be difficult to obtain in real-world applications. Hybrid models offer better overall performance and are more suitable for complex and changing scenarios.

[0127] In fault diagnosis systems, residual analysis is an important tool for evaluating model performance. Residuals reflect the difference between the model's predicted values and the actual values. Analyzing the distribution and characteristics of residuals provides insight into the model's fit and fault detection capabilities. Smaller residuals indicate better fit. The hybrid model exhibits a narrow residual range, with most residual values close to zero, demonstrating that it combines the advantages of unsupervised and supervised learning and exhibits good fit.

[0128] The false alarm rate (FAR) reflects the probability that a model will misclassify normal data as faulty. Experimental results show that the unsupervised model has the lowest FAR of 0.29%, indicating a low misclassification rate for normal data. The hybrid model and supervised model have slightly higher FARs of 1.15% and 1.01%, respectively. The unsupervised model has a low false alarm rate due to its conservative fault detection strategy, but this also results in a significant decrease in its fault detection rate. While the hybrid and supervised models improve the fault detection rate, the false alarm rate increases slightly. This may be due to the increased sensitivity of the models to fault characteristics, which results in some normal data being misclassified as faulty.

[0129] Combining the experimental results of FDR, FAR, and AUC values, it can be seen that the hybrid model performs best in terms of fault detection rate (FDR), indicating that it can effectively combine the advantages of unsupervised and supervised learning and is suitable for scenarios with high requirements for fault detection accuracy.

Claims

1. A method for fault diagnosis of unmanned boats based on a combination of unsupervised and supervised learning, characterized by: The following steps are involved: Step 1: Building an unmanned boat thrust prediction model, which includes an unsupervised neural network, a supervised neural network, and a bridge network for connecting the unsupervised neural network and the supervised neural network; Step 2: Train the unmanned boat thrust prediction model to obtain a trained unmanned boat thrust prediction model; Step 3: Collect the real-time status data of the unmanned boat, input the real-time status data into the trained unmanned boat thrust prediction model, and obtain the unmanned boat thrust prediction result; Step 4: According to the prediction results of the unmanned boat thrust, the final residual is obtained and the T is calculated using the final residual. 2 Statistics, according to T 2 Statistics are used to determine whether the unmanned boat has a fault.

2. The method for diagnosing faults of an unmanned boat based on a combination of unsupervised and supervised learning according to claim 1, characterized in that: The unsupervised neural network is based on H unsp (Θ) is obtained by constructing the neural network architecture. unsp (Θ) refers to the neural network architecture with hyperparameter Θ.

3. The method for diagnosing faults of an unmanned boat based on a combination of unsupervised and supervised learning according to claim 2, characterized in that: The supervised neural network is based on H sp (Θ) is obtained by constructing the neural network architecture. sp (Θ) refers to the neural network architecture with hyperparameters.

4. The method for diagnosing faults of an unmanned boat based on a combination of unsupervised and supervised learning according to claim 3, characterized in that: The bridge network for connecting the unsupervised neural network and the supervised neural network is expressed as: in I represents the identity matrix, H(·) is the variation function used to change the auxiliary variable a u Under the action of , the inverse conversion between supervised and unsupervised is realized.

5. The method for diagnosing faults of an unmanned boat based on a combination of unsupervised and supervised learning according to claim 4, characterized in that: The training of the unmanned boat fault diagnosis model to obtain the trained unmanned boat fault diagnosis model specifically includes the following operations: Use the unit time delay operator z -1 , define a unit time delay operator on the input u(k+s) and output y(k+s) to obtain the input of the unsupervised neural network: Where k is the discrete time, y(·) represents the output variable driven by data at time s, s represents any time, and s p Indicates the length of time delay; Thus we get the composite vector z of the past time-delayed input and output p , expressed as: Where, And get the output variable y at the current time s s (k), expressed as: y s (k)=Y x L p z p (k)+Y u you s (k)+Y e e s (k) Where, Compound operation, Υ u for e is the random noise variable inside and outside the system, Υ e for L p is the parameter for adjusting the error; By minimizing the loss function L of the unsupervised neural network unsp Get the optimal hyperparameter Θ * : Where u s represents the current input variable, Θ z Neural network parameters related to past inputs and outputs, used to extract state features from input data or model state dynamics, Θ u Neural network parameters related to the current input variables, used to process control signals or external inputs, Θ y The neural network parameters associated with the current output variable, used to generate output predictions; Represents an unsupervised joint transformation function with parameters Θ z 、Θ u 、Θ y The constructed neural network architecture is used to generate a joint representation of the system behavior by fusing information about input, state, and output; Use H unsp (Θ * ) to obtain the unsupervised residual signal, which is expressed as: By minimizing the loss function L of the supervised neural network sp Get the optimal hyperparameter Θ * : Use H sp (Θ * ) to obtain the supervised residual signal By minimizing the loss function L of the bridge network ext , get the trained bridge network: Use the trained bridge network to output thrust prediction values:

6. The method for diagnosing faults of an unmanned boat based on a combination of unsupervised and supervised learning according to claim 5, characterized in that: The final residual is expressed as: Where Y s Represents the actual output obtained by the sensor; The final residual is used to calculate T 2 Statistics: Where, represents the transpose of the final residual, and E(·) represents the expectation; According to T 2 Statistics are used to determine whether the unmanned boat has a fault. The specific operations include: like If it is fault-free, it means there is a fault, otherwise, J th Indicates T 2 The threshold value of the statistic.

Images