Underwater propulsion motor fault diagnosis method, device, equipment and storage medium

Abstract

The present invention discloses a method, apparatus, device, and storage medium for underwater propulsion motor fault diagnosis, relating to the technical field of underwater propulsion motor fault diagnosis. The method comprises the following steps: preprocessing multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set; training a Markov decision framework to obtain an optimal fault frequency band selection strategy, and obtaining a labeled optimal feature subset and an unlabeled optimal feature subset based on the optimal fault frequency band selection strategy; applying the optimal fault frequency band selection strategy to the preprocessed vibration and current signals of the underwater propulsion motor to be diagnosed to obtain corresponding labeled optimal feature subsets and unlabeled optimal feature subsets, which are input into a fault diagnosis model to determine the corresponding fault type. The present invention achieves noise reduction of fault features through fault frequency band selection through the Markov decision process.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater propulsion motor fault diagnosis, and in particular to an underwater propulsion motor fault diagnosis method, device, equipment and storage medium. Background Art

[0002] The propulsion motor is an important power component of underwater mechanical equipment. It is prone to various faults under various complex working conditions and long-term operation. The expansion of the fault will induce other possible faults, causing greater economic losses and safety accidents. Therefore, it is of great significance to detect underwater propulsion motor faults in a timely manner.

[0003] Fault diagnosis of underwater propulsion motors faces three major challenges. First, long-term offshore operation leads to high recall and maintenance costs, resulting in little fault data, most of which is unlabeled data. Second, the complex internal electromagnetic environment means that sensors can only be distributed outside the motor, which is affected by complex factors such as wave impact and propulsion environment stress, resulting in a low signal-to-noise ratio of collected samples. Third, various types of faults are prone to occur during operation in complex underwater environments, including mechanical faults, electrical faults, and demagnetization faults. There are many types of faults, and it is difficult to subjectively select fault characteristics.

[0004] In response to the small sample scenarios faced by underwater propulsion motor fault diagnosis, existing technologies often use generative methods such as generative adversarial networks to expand the data volume, that is, directly inputting the original fault data of the underwater propulsion motor into the generative adversarial network. However, due to the low signal-to-noise ratio of the original data, it often affects the stability of the generative adversarial network training process and the quality of the final generated samples, thereby affecting the performance of the subsequent fault diagnosis model. Summary of the Invention

[0005] The present invention provides a method, device, equipment and storage medium for underwater propulsion motor fault diagnosis, which solves the problem that the existing technology often uses generative methods such as generative adversarial networks to expand the data volume, and directly inputs the original fault data with low signal-to-noise ratio of the underwater propulsion motor into the generative adversarial network, which often affects the stability of the training process and the quality of the generated samples, thereby affecting the performance of the subsequent fault diagnosis model.

[0006] The present invention provides a method for diagnosing underwater propulsion motor faults, comprising the following steps:

[0007] Collect multiple vibration and current signals of underwater propulsion motors under different environmental conditions, including labeled and unlabeled signals. Labeled signals indicate known fault types, while unlabeled signals indicate unknown fault types.

[0008] Preprocessing multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set;

[0009] The Markov decision framework is trained through the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy. Based on the optimal fault frequency band selection strategy, the optimal labeled feature subset and the optimal unlabeled feature subset are selected from the labeled frequency band feature set and the unlabeled frequency band feature set.

[0010] The fault diagnosis model is obtained by training the semi-supervised feature generation adversarial network through the labeled optimal feature subset and the unlabeled optimal feature subset.

[0011] The optimal fault frequency band selection strategy is applied to the preprocessed vibration signal and current signal of the underwater propulsion motor to be diagnosed, and the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed are obtained. The labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed are input into the fault diagnosis model to obtain the corresponding fault type.

[0012] Preferably, preprocessing the multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set includes the following steps:

[0013] The frequency domain features of labeled and unlabeled data in multiple vibration and current signals are extracted by time-frequency conversion method to obtain a set of feature vectors.

[0014] Compress each p adjacent frequencies in the feature vector set as a frequency band value to obtain a feature set, where the size of p represents the resolution of the frequency domain feature;

[0015] The different sensor signals are spliced ​​together to obtain the frequency band feature set {F i ,Y i |i∈m} and {f i |i∈m}, F, Y represent labeled features and their labels respectively, and f represents unlabeled features.

[0016] Preferably, the Markov decision framework includes an agent, an environment, an action space, a state space, a policy function, a reward, and a Q-table, wherein the agent is used to traverse each feature in a set of labeled frequency band features, and the agent interacts with the environment; the action space indicates whether to select the current frequency band, the state space indicates the set of selection states obtained for the selection action of each frequency band, the policy function indicates the basis for the agent to take action through the Softmax algorithm, the reward indicates the reward calculated according to a given reward function, the Q-table is used to record the average reward for taking actions in each state, and the optimal feature subset selected is determined based on the Q-table.

[0017] Preferably, the training of the Markov decision framework by using the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy includes the following steps:

[0018] In one time step, based on the current state of the environment, the agent selects an action from the Q-table according to the policy function and executes it to obtain the next state of the environment and calculate the reward;

[0019] Based on the next state and the corresponding reward, the agent updates the Q value in the Q-table through the Q-learning algorithm;

[0020] The Q value in the Q-table is updated through multiple time steps to obtain a converged Q-table. The optimal fault frequency band selection strategy, that is, a set of states with the highest Q value, is obtained through the converged Q-table.

[0021] Preferably, selecting the optimal labeled feature subset and the optimal unlabeled feature subset from the labeled frequency band feature set and the unlabeled frequency band feature set according to the optimal fault frequency band selection strategy includes the following steps:

[0022] According to the set of states with the highest Q value, the corresponding selected features are obtained, that is, the optimal feature subset with labels;

[0023] The optimal labeled feature subset is mapped to the unlabeled frequency band feature set with the same resolution to obtain the optimal unlabeled feature subset.

[0024] Preferably, the classifier of the semi-supervised feature generation adversarial network is trained by the labeled optimal feature subset and the unlabeled optimal feature subset, and the generator of the semi-supervised feature generation adversarial network is trained by the labeled optimal feature subset.

[0025] Preferably, during the training process of the semi-supervised feature generation adversarial network, the generator and classifier are fixed separately and the parameters are updated in a game-playing manner to complete the training until the training loss reaches the standard.

[0026] A device for diagnosing underwater propulsion motor faults, comprising:

[0027] An acquisition module is used to collect multiple vibration and current signals of the underwater propulsion motor under different environmental conditions, including labeled signals and unlabeled signals. A labeled signal indicates that the signal is a known fault type signal, while an unlabeled signal indicates that the signal is an unknown fault type signal;

[0028] A processing module, configured to pre-process the multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set;

[0029] An acquisition module is used to train the Markov decision framework using the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy, and select the labeled optimal feature subset and the unlabeled optimal feature subset from the labeled frequency band feature set and the unlabeled frequency band feature set according to the optimal fault frequency band selection strategy;

[0030] A training module is used to train a semi-supervised feature generation adversarial network using the labeled optimal feature subset and the unlabeled optimal feature subset to obtain a fault diagnosis model;

[0031] The diagnostic module is used to apply the optimal fault frequency band selection strategy to the preprocessed vibration signal and current signal of the underwater propulsion motor to be diagnosed, obtain the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed, input the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed into the fault diagnosis model, and obtain the corresponding fault type.

[0032] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned underwater propulsion motor fault diagnosis method when executing the program.

[0033] A computer-readable storage medium stores a computer program, which implements the above-mentioned underwater propulsion motor fault diagnosis method when executed by a processor.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention first collects multiple vibration signals and current signals of the underwater propulsion motor under different environmental working conditions, and processes the multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set. The Markov decision framework is trained through the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy, and then the labeled optimal feature subset and the unlabeled optimal feature subset are obtained. The semi-supervised feature generation adversarial network is trained through the labeled optimal feature subset and the unlabeled optimal feature subset to obtain a fault diagnosis model. The present invention realizes the fault diagnosis of the underwater propulsion motor in a small sample scenario efficiently and accurately through the Markov decision process and the semi-supervised feature generation adversarial network. The fault frequency band selection through the Markov decision process realizes the denoising of the fault feature, provides high-quality input for the semi-supervised feature generation adversarial network, and ensures the quality of the generated features; then, the semi-supervised feature generation adversarial network is used to train the fault diagnosis model using a large amount of unlabeled data, thereby improving the generalization ability and robustness of the model. This method makes up for the shortcomings of traditional motor fault diagnosis methods, such as low fault diagnosis accuracy in small sample scenarios and low quality of generated samples under generative models. It can be better applied to the complex and changeable operating environment of underwater propulsion motor systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 This is a flow chart of a method for diagnosing underwater propulsion motor faults according to the present invention;

[0038] Figure 2 A schematic diagram of feature selection by establishing a Markov decision process according to the present invention;

[0039] Figure 3 This is a schematic diagram of the diagnostic process using the trained fault diagnosis model in the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] The present invention provides a method for diagnosing underwater propulsion motor faults, referring to Figure 1 , including the following steps:

[0042] Step 1: Signal processing, collecting vibration signals of underwater propulsion motor under different environmental conditions {A i |i∈m} and current signal {C i |i∈m}, where m represents the environmental condition. After performing mean zeroing and normalization on the vibration and current signals, time-frequency conversion methods (such as Fourier transform) are used to extract the frequency domain features of the labeled and unlabeled data, resulting in a set of feature vectors. The label refers to the fault type of the signal. Labeled data refers to the fault type of the acquired fault signal, which is known, while unlabeled data refers to the fault type of the signal, which is unknown. A small portion of the acquired vibration and current signals are acquired when the fault type is known, while the majority are of unknown fault type.

[0043] Each p adjacent frequencies in the feature vector set are taken as a frequency band value, that is, the frequency domain feature vector is compressed to obtain the feature set {F i 1 ,Y i 1 |i∈m} and {f i 1 |i∈m}, the size of p represents the resolution of the frequency domain feature. The frequency band feature set {F i ,Y i |i∈m} and {f i |i∈m}, F, Y represent labeled features and their labels respectively, and f represents unlabeled features.

[0044] Step 2: Establish a Markov decision framework, define the action space A = {1,0} to indicate whether to select the current frequency band, define the agent to traverse each feature of the feature set, and define the state space X to represent the selection state x obtained by the selection action for each frequency band. k We define a policy function π, which represents the basis for the agent's actions provided by the Softmax algorithm. We define a reward r, which represents the reward calculated based on the given reward function. We also define a Q-table, which records the average reward for actions taken in each state and is used to determine the final selected frequency band subset. The Softmax algorithm is used as the selection strategy. Compared to the commonly used greedy algorithm, the Softmax strategy has the advantage of being able to explore the state space to a certain extent. Even actions with low Q values ​​have a chance of being selected, which allows the agent to explore more smoothly and avoid excessive concentration on a single optimal action.

[0045] The selection state obtained by the selection action for each frequency band is, for example, a state xk ={1,1,0,…,0,1}, where the element value 1 represents the feature F corresponding to the position i The combination of all state vectors forms the state space X.

[0046] When the agent takes action, it follows the policy function π. Here, the Softmax algorithm is selected as the policy, which specifically means:

[0047] π(x k ,a k )=rand(A)P(k)

[0048]

[0049] Where: rand(A) represents the agent in state x k Under this condition, an action is randomly selected in the action space A with a probability of P(k); P(k) is a probability distribution function based on the Boltzmann distribution; where Q π (x k ,A) Record the average reward of the current action, τ>0 is the temperature parameter, which is used to control the randomness of the selected action, and τ is selected as 0.1.

[0050] Calculate the reward for the agent's actions. The reward function is:

[0051] r k =xa k -MAXxa k

[0052] where x k is the cross-validation classification accuracy of the feature subset in each state, MAXxa k The maximum classification accuracy in each action step. The feature subset selected each time is used as the input of the SVM model for cross-validation to calculate the classification accuracy.

[0053] Step 3: Fault frequency band selection, using the above labeled frequency band feature set {F i ,Y i |i∈m} trains the established Markov decision framework. Figure 2 , the agent selects action a from the Q-table based on the current state in the environment according to the Softmax algorithm k And execute to get the next state x k , and calculate the reward r k , update the Q value in the Q-table according to the Q-learning algorithm. The agent obtains the optimal fault frequency band selection strategy x by continuously interacting with the environment and optimizing the Q-table. s, get the optimal feature subset {F x ,Y x Apply the fault band selection strategy to the unlabeled band feature set {f i |i∈m}, we get the optimal unlabeled feature subset f x .

[0054] The update process of Q-table is:

[0055] Q(x k ,a k )=(1-α)×Q(x k ,a k )+α×(r k +γ×maxQ(x k+1 ,a k+1 ))

[0056] Record each state x k The average reward obtained for each action in the state is calculated, and the discount γ is selected to update the reward value. The learning rate α controls the update speed of the Q value in each state. The final feature selection strategy x is obtained by updating the Q-table after convergence.

[0057] Step 4: Replace {F x ,Y x}、f x Input semi-supervised feature generation adversarial network fault diagnosis model for training, with labeled feature samples {F x ,Y x} Participate in the training of generator and classifier at the same time. Unlabeled feature sample f x This classifier is only involved in classifier training. The classifier distinguishes between real samples and generated pseudo samples, while also distinguishing the fault type of the samples. During training, the generator and classifier are fixed separately, updating parameters in a game-like manner until the training loss reaches the target. Through training, the generator can produce realistic pseudo samples, and the classifier can serve as a fault classification model for subsequent diagnosis.

[0058] Step 5: Reference Figure 3 , use the newly collected data samples as the test set to verify the fault diagnosis model: First, repeat the first step to obtain the test sample {F i ,Y i |i∈m} and {f i |i∈m}, then the optimal fault frequency band selection strategy x obtained in step 3 is s Applied to the test sample, the feature subset {F x ,Y x} and f x Finally, the feature subset is input into the trained classifier D for fault diagnosis to obtain the fault type corresponding to the sample.

[0059] Based on the same concept, the present invention also provides an underwater propulsion motor fault diagnosis device, which includes an acquisition module, a processing module, an acquisition module, a training module and a diagnosis module.

[0060] The acquisition module is used to collect multiple vibration signals and current signals of the underwater propulsion motor under different environmental working conditions, including labeled signals and unlabeled signals. The labeled signal indicates that the signal is a known fault type signal, and the unlabeled signal indicates that the signal is an unknown fault type signal.

[0061] The processing module is used to pre-process multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set.

[0062] The acquisition module is used to train the Markov decision framework through the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy, and select the labeled optimal feature subset and the unlabeled optimal feature subset from the labeled frequency band feature set and the unlabeled frequency band feature set according to the optimal fault frequency band selection strategy.

[0063] The training module is used to train the semi-supervised feature generation adversarial network through the labeled optimal feature subset and the unlabeled optimal feature subset to obtain a fault diagnosis model.

[0064] The diagnostic module is used to apply the optimal fault frequency band selection strategy to the preprocessed vibration signal and current signal of the underwater propulsion motor to be diagnosed, obtain the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed, input the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed into the fault diagnosis model, and obtain the corresponding fault type. The present invention also provides a computer-readable storage medium, the storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned underwater propulsion motor fault diagnosis method.

[0065] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0066] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for diagnosing underwater propulsion motor faults, characterized in that: The following steps are involved: Collect multiple vibration and current signals of underwater propulsion motors under different environmental conditions, including labeled and unlabeled signals. Labeled signals indicate known fault types, while unlabeled signals indicate unknown fault types. Preprocessing multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set; The Markov decision framework is trained through the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy. Based on the optimal fault frequency band selection strategy, the optimal labeled feature subset and the optimal unlabeled feature subset are selected from the labeled frequency band feature set and the unlabeled frequency band feature set. The fault diagnosis model is obtained by training the semi-supervised feature generation adversarial network through the labeled optimal feature subset and the unlabeled optimal feature subset. The optimal fault frequency band selection strategy is applied to the preprocessed vibration signal and current signal of the underwater propulsion motor to be diagnosed, and the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed are obtained. The labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed are input into the fault diagnosis model to obtain the corresponding fault type.

2. The underwater propulsion motor fault diagnosis method according to claim 1, characterized in that: Preprocessing multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set includes the following steps: The frequency domain features of labeled and unlabeled data in multiple vibration and current signals are extracted by time-frequency conversion method to obtain a set of feature vectors. Compress each p adjacent frequencies in the feature vector set as a frequency band value to obtain a feature set, where the size of p represents the resolution of the frequency domain feature; The different sensor signals are spliced ​​together to obtain the frequency band feature set {F i ,Y i |i∈m} and {f i |i∈m}, F, Y represent labeled features and their labels respectively, and f represents unlabeled features.

3. The underwater propulsion motor fault diagnosis method according to claim 2, characterized in that: The Markov decision framework includes an agent, an environment, an action space, a state space, a policy function, a reward, and a Q-table, wherein the agent is used to traverse each feature in a labeled frequency band feature set, and the agent interacts with the environment; the action space indicates whether to select the current frequency band, the state space indicates the set of selection states obtained for the selection action of each frequency band, the policy function indicates the basis for the agent to take action through the Softmax algorithm, the reward indicates the reward calculated according to a given reward function, the Q-table is used to record the average reward for taking actions in each state, and the optimal feature subset for selection is determined based on the Q-table.

4. The underwater propulsion motor fault diagnosis method according to claim 3, characterized in that: The Markov decision framework is trained by using a set of labeled frequency band features to obtain an optimal fault frequency band selection strategy, including the following steps: In one time step, based on the current state of the environment, the agent selects an action from the Q-table according to the policy function and executes it to obtain the next state of the environment and calculate the reward; Based on the next state and the corresponding reward, the agent updates the Q value in the Q-table through the Q-learning algorithm; The Q value in the Q-table is updated through multiple time steps to obtain a converged Q-table. The optimal fault frequency band selection strategy, that is, a set of states with the highest Q value, is obtained through the converged Q-table.

5. The underwater propulsion motor fault diagnosis method according to claim 4, characterized in that: Selecting the optimal labeled feature subset and the optimal unlabeled feature subset from the labeled frequency band feature set and the unlabeled frequency band feature set according to the optimal fault frequency band selection strategy includes the following steps: According to the set of states with the highest Q value, the corresponding selected features are obtained, that is, the optimal feature subset with labels; The optimal labeled feature subset is mapped to the unlabeled frequency band feature set with the same resolution to obtain the optimal unlabeled feature subset.

6. The underwater propulsion motor fault diagnosis method according to claim 1, characterized in that: The classifier of the semi-supervised feature generative adversarial network is trained by the labeled optimal feature subset and the unlabeled optimal feature subset, and the generator of the semi-supervised feature generative adversarial network is trained by the labeled optimal feature subset.

7. The underwater propulsion motor fault diagnosis method according to claim 6, characterized in that: During the training process of the semi-supervised feature generation adversarial network, the generator and classifier are fixed separately and the parameters are updated in a game-theory manner to complete the training until the training loss reaches the standard.

8. An underwater propulsion motor fault diagnosis device, characterized in that: include: An acquisition module is used to collect multiple vibration and current signals of the underwater propulsion motor under different environmental conditions, including labeled signals and unlabeled signals. A labeled signal indicates that the signal is a known fault type signal, while an unlabeled signal indicates that the signal is an unknown fault type signal; A processing module, configured to pre-process the multiple vibration signals and current signals to obtain a labeled frequency band feature set and an unlabeled frequency band feature set; An acquisition module is used to train the Markov decision framework using the labeled frequency band feature set to obtain the optimal fault frequency band selection strategy, and select the labeled optimal feature subset and the unlabeled optimal feature subset from the labeled frequency band feature set and the unlabeled frequency band feature set according to the optimal fault frequency band selection strategy; A training module is used to train a semi-supervised feature generation adversarial network using the labeled optimal feature subset and the unlabeled optimal feature subset to obtain a fault diagnosis model; The diagnostic module is used to apply the optimal fault frequency band selection strategy to the preprocessed vibration signal and current signal of the underwater propulsion motor to be diagnosed, obtain the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed, input the labeled optimal feature subset and the unlabeled optimal feature subset of the vibration signal and current signal of the underwater propulsion motor to be diagnosed into the fault diagnosis model, and obtain the corresponding fault type.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for diagnosing underwater propulsion motor faults as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the underwater propulsion motor fault diagnosis method according to any one of claims 1 to 7 is implemented.

Images