Ship equipment fault detection method, equipment, medium and product

Abstract

The invention discloses a ship equipment fault detection method and device, a medium and a product, and relates to the field of electric digital data processing, and the method comprises the steps: introducing a bidirectional gating circulation unit and a variational auto-encoder into a teacher-student network model, and obtaining an initial detection model; respectively inputting a fault-free signal in the first sample data set and a corresponding signal in the second sample data set into a student network and a teacher network, training an initial detection model to obtain a ship equipment detection model, and inputting a to-be-detected ship equipment signal to obtain an actual detection result; in the training process, a BM method and an extreme value theory are adopted to obtain GEV distribution so as to determine a fault detection threshold value, and then ship equipment fault detection is achieved in combination with an actual detection result. According to the method, the accuracy, the robustness and the practicability of detection are remarkably improved, the technical bottlenecks in the aspects of data dependence, priori knowledge utilization, detection precision and the like in the prior art are broken through, and the detection requirements of complex and changeable statistical characteristics in a real environment are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric digital data processing, and in particular to a ship equipment fault detection method, device, medium and product. BACKGROUND

[0002] The rapid development of modern science and technology and its wide application in ship equipment make the ship equipment present the development trend of technology-intensive, complex structure and powerful function. The traditional fault detection method is not suitable for the requirements of rapidity, visibility and accuracy of equipment detection under the condition of information-based war because of low information utilization rate, slow detection response speed, long time and high detection cost. In order to maintain the combat readiness of modern ship equipment and fully exert its technical performance, it is urgent to develop a ship equipment fault detection method with higher resource utilization rate and lower cost.

[0003] The priori knowledge guided teacher-student self-supervised learning model can be trained and fault detected by using only fault-free samples. The teacher-student model is composed of a teacher and a student. The teacher network is pre-trained and has certain feature extraction capability. For the training of the student network, the features extracted by the student network are as close as possible to the features extracted by the teacher network under the condition of using only fault-free samples. When the test involves incorrect data, due to the heterogeneity of the teacher and student networks, there will be a certain feature distance between the feature pairs. The feature distance within the fault-free pair is closer, while the feature distance within the fault pair is farther. When the feature distance exceeds the set threshold, it is determined that a fault is detected.

[0004] The bidirectional gated recurrent unit (Gated Recurrent Unit, GRU) is a variant of recurrent neural network (Recurrent Neural Network, RNN) and aims to solve the gradient disappearance problem in standard RNN. GRU effectively controls the flow of information by introducing a gating mechanism, allowing the network to capture long-term dependencies while reducing computational complexity. It has two key components: update gate and reset gate. The update gate determines the degree of influence of the previous hidden state on the current state, while the reset gate determines the degree to which the information of the previous state is written into the current candidate hidden state. This design makes GRU more concise in structure, usually with fewer parameters, and more efficient in training. The application fields of GRU are very wide, including natural language processing (such as text classification, sentiment analysis, machine translation), speech recognition, time series prediction, etc. In these tasks, GRU can handle the time sequence characteristics in the data and provide accurate prediction and classification.

[0005] A multilayer perceptron (MLP) is a type of feedforward artificial neural network composed of an input layer, one or more hidden layers, and an output layer, each consisting of multiple neurons. MLPs can learn and model complex nonlinear relationships through the connections between neurons and weights in these layers, making them widely applicable in tasks such as classification, regression, and pattern recognition. The working principle of MLP involves two main processes: forward propagation and backpropagation. In forward propagation, input data is passed through each layer, where neurons perform a weighted sum of inputs and introduce nonlinearity through an activation function, ultimately generating a predicted output. Backpropagation, on the other hand, involves calculating the gradient of the loss function with respect to network parameters and using these gradients to update the weights and biases of the network to minimize prediction error. Activation functions play a crucial role in MLPs, with common examples including Sigmoid, ReLU (Rectified Linear Unit), and Tanh, which determine the activation state of neurons and provide the network with nonlinearity mapping capabilities. Training of MLPs typically employs gradient descent or its variants, such as Stochastic Gradient Descent (SGD) and the Adam optimizer, which iteratively adjust network parameters to achieve better performance.

[0006] Existing ship equipment fault detection methods generally require a large amount of labeled fault data for training. However, in practical applications, due to the complexity and high safety requirements of ship systems, fault data is extremely scarce and difficult to obtain, making it difficult for these methods to effectively generalize and apply. At the same time, existing methods often lack effective use of equipment operation mechanisms and prior knowledge, and cannot model the essence of fault changes combined with expert experience, resulting in insufficient model robustness and explainability. Some methods use deep learning models for feature extraction, but often ignore the temporal characteristics of signals and are prone to performance degradation in complex noise environments, making it difficult to achieve accurate detection. In addition, traditional methods often rely on manually set or fixed threshold values based on simplified distribution assumptions in fault discrimination, making it difficult to adapt to the complex and variable statistical characteristics of real-world environments. SUMMARY

[0007] The purpose of the present application is to provide a ship equipment fault detection method, device, medium and product that can significantly improve the accuracy, robustness and practicality of detection, breaking through the technical bottlenecks of existing methods in data dependency, prior knowledge utilization and detection accuracy, and thus can adapt to the complex and variable statistical characteristics of real-world environments.

[0008] To achieve the above purpose, the present application provides the following solutions:

[0009] In a first aspect, the application provides a ship equipment fault detection method, comprising:

[0010] Obtaining a fault-free signal of the ship equipment to form a first sample data set;

[0011] According to prior knowledge, the fault-free signal in the first sample data set is transformed into signals of multiple different modes to form a second sample data set;

[0012] Under the framework of the teacher-student network model, a bidirectional gate recurrent unit and a variational autoencoder are introduced into the teacher network to construct an initial detection model;

[0013] The fault-free signal in the first sample data set is used as the input of the student network in the initial detection model, and the signals of different modes corresponding to the fault-free signal input into the student network in the second sample data set are used as the input of the teacher network to train the initial detection model until a set condition is reached, obtaining a trained initial detection model;

[0014] In the training process of the initial detection model, the BM method and the extreme value theory are used to obtain the DEV distribution;

[0015] The trained initial detection model is used as the ship equipment detection model;

[0016] Obtaining a signal of a ship equipment to be detected;

[0017] The signal of the ship equipment to be detected is input into the teacher network and the student network of the ship equipment detection model respectively to obtain a ship equipment detection result;

[0018] The fault detection threshold is determined based on the GEV distribution;

[0019] The ship equipment fault detection is realized based on the ship equipment detection result and the fault detection threshold.

[0020] Optionally, according to prior knowledge, the fault-free signal in the first sample data set is transformed into signals of multiple different modes to form a second sample data set, comprising:

[0021] According to prior knowledge, each fault-free signal in the first sample data set is subjected to multiple signal transformations to obtain signals of multiple different modes;

[0022] In the signal transformation process, pseudo-labels corresponding to the signals of different modes are generated;

[0023] The fault-free signal, the signals of multiple different modes, and the pseudo-labels corresponding to the signals of different modes are stored to obtain the second sample data set.

[0024] Optionally, the multiple signal transformations include amplitude transformation, frequency transformation, and waveform transformation.

[0025] Optionally, the variational autoencoder comprises a teacher encoding network and a teacher decoding network.

[0026] The teacher encoding network is connected with the teacher decoding network and the bidirectional gated recurrent unit respectively; and the teacher decoding network is connected with the bidirectional gated recurrent unit.

[0027] The teacher encoding network is configured to encode an input signal to obtain an encoded signal; the teacher decoding network is configured to decode the encoded signal to obtain a decoded signal; and the bidirectional gated recurrent unit is configured to obtain a fault detection and classification result based on the encoded signal and the decoded signal.

[0028] Optionally, the construction process of the teacher encoding network and the teacher decoding network comprises:

[0029] An initial teacher encoding network and an initial teacher decoding network are constructed by using a recurrent neural network.

[0030] A hidden layer is constructed by replacing neurons of the recurrent neural network in the initial teacher encoding network and the initial teacher decoding network with a memory network, so as to obtain the teacher encoding network and the teacher decoding network.

[0031] Optionally, an over-layer perceptron is used as a student network in the initial detection model.

[0032] Optionally, the process of determining the fault detection threshold based on the GEV distribution comprises:

[0033] A GEV distribution is determined by using An extreme quantile of the GEV distribution is determined, and the determined extreme quantile is used as the fault detection threshold.

[0034] In the formula, z represents an extreme quantile when the significance level is α, ξ represents a shape parameter, σ represents a size parameter, and μ represents a location parameter. α In the formula, z represents an extreme quantile when the significance level is α, ξ represents a shape parameter, σ represents a size parameter, and μ represents a location parameter.

[0035] In a second aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the ship equipment fault detection method provided above.

[0036] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the steps of the ship equipment fault detection method provided above.

[0037] In a fourth aspect, the present application provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the ship equipment fault detection method provided above.

[0038] According to the specific embodiments provided in the present application, the present application has the following technical effects:

[0039] The present application provides a ship equipment fault detection method, device, medium and product, which combines prior knowledge, uses a first sample data set formed by a fault-free signal and a second sample data set formed by signals of multiple different modes obtained by transforming the fault-free signal in the first sample data set as training data sets for model training, without fault data, solving the problem of data acquisition. By introducing a bidirectional gated recurrent unit and a variational autoencoder in the framework of the teacher-student network model, the feature extraction and time series modeling capabilities can be enhanced, the robustness of the model to noise can be enhanced, and the fault detection can be more accurate. Moreover, the determination of the dynamic fault detection threshold based on the GEV distribution can significantly improve the accuracy, robustness and practicality of the detection, thereby breaking through the technical bottlenecks of existing methods in data dependency, utilization of prior knowledge and detection accuracy, etc. to adapt to the detection needs of complex and variable statistical characteristics in real environments. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0041] Figure 1 A flowchart of a ship equipment fault detection method provided by an embodiment of the present application is shown in the figure.

[0042] Figure 2 An implementation framework diagram of a ship equipment fault detection method provided by an embodiment of the present application is shown in the figure.

[0043] Figure 3 A schematic diagram of a command and control system provided by an embodiment of the present application is shown in the figure.

[0044] Figure 4 A structure diagram of a variational autoencoder provided by an embodiment of the present application is shown in the figure.

[0045] Figure 5 A structure diagram of a bidirectional gated recurrent unit provided by an embodiment of the present application is shown in the figure.

[0046] Figure 6A structure schematic diagram of a multi-layer perceptron provided by an embodiment of the present application is shown in the figure.

[0047] Figure 7 A structure schematic diagram of a computer device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0049] In order to make the above objectives, characteristics and advantages of the present application more apparent, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0050] In an exemplary embodiment, the present application provides a ship equipment fault detection method, which is executed by a computer device, specifically, can be executed by a terminal or a server alone or by a terminal and a server together. In the embodiments of the present application, the method is taken as an example applied to a server. Figure 1 As shown in the figure, the method comprises:

[0051] Step 100, obtaining a fault-free signal of a ship equipment to form a first sample data set.

[0052] Step 101, transforming the fault-free signal in the first sample data set into signals of multiple different modes according to prior knowledge to form a second sample data set.

[0053] Step 102, introducing a bidirectional gate recurrent unit and a variational autoencoder into a teacher network under the framework of a teacher-student network model to construct an initial detection model.

[0054] Step 103, taking the fault-free signal in the first sample data set as an input of a student network in the initial detection model, and taking the signals of different modes corresponding to the fault-free signal input into the student network in the second sample data set as an input of the teacher network, to train the initial detection model until a set condition is reached, to obtain a trained initial detection model.

[0055] Step 104, obtaining a DEV distribution by using a BM method and an extreme value theory in the training process of the initial detection model.

[0056] Step 105, taking the trained initial detection model as a ship equipment detection model.

[0057] Step 106, obtaining a ship equipment signal to be detected.

[0058] Step 107, inputting the ship equipment signal to be detected into a teacher network and a student network of a ship equipment detection model respectively to obtain a ship equipment detection result.

[0059] Step 108, determining a fault detection threshold based on a GEV distribution.

[0060] Step 109, implementing ship equipment fault detection based on the ship equipment detection result and the fault detection threshold.

[0061] By implementing the above steps 100-109, the present application can greatly reduce the data acquisition cost, improve the efficiency, adaptability, accuracy and robustness of fault detection, and further break through the technical bottlenecks of existing methods in data dependency, prior knowledge utilization and detection precision, to adapt to the detection requirements of complex and variable statistical characteristics in real environment.

[0062] The existing ship equipment system includes detection and perception system, command and control system (i.e. command and control system), weapon system, communication system, etc. The command and control system as a performance multiplier of ship equipment plays a more and more key role in the modern ship equipment system, based on which the following embodiments are described taking the command and control system as an example. As shown in Figure 3 The command and control system includes intelligence reconnaissance subsystem, command decision subsystem, communication subsystem and execution subsystem. Each subsystem contains relevant different components, wherein the intelligence reconnaissance subsystem includes radar system, sonar system and unmanned reconnaissance machine. The command decision subsystem includes command center and battlefield management system. The communication subsystem includes radio communication system, satellite communication system and network management system. The execution subsystem includes information processing device, display device and execution device.

[0063] Based on the above description of steps 100-109 of the present application, in the fault detection data acquisition process, it is difficult to collect large-scale labeled fault data sets, and the fully supervised method is not practical in real scenarios. Therefore, the present application introduces a prior knowledge guided teacher-student self-supervised learning model (i.e. teacher-student network model) in detection, which can be trained only using fault-free samples and perform fault detection. The prior knowledge guided teacher-student self-supervised learning model consists of two sub-models, as shown in Figure 2 The teacher network and the student network respectively. The teacher network is pre-trained and has certain feature extraction capability. The student network is trained only using fault-free samples, and the feature Z is the output of the teacher network, is the approximate value of the output of the student network. When the test involves error data, Z and Due to the heterogeneity of the teacher network and the student network, a large gap exceeding the set threshold will be generated to detect faults. The two features in the fault-free pair are close, while the two features in the fault pair are far apart.

[0064] Based on the above description, the priori knowledge guided teacher-student self-supervised learning model is divided into two stages, wherein, Figure 2 Steps 1-3 in the above are the training stage, and step 4 is the test stage. First, since it is difficult to collect a large-scale labeled fault dataset, first, according to the priori knowledge, taking the example that the fault caused signal will change in amplitude, frequency and waveform, the fault-free signal X n is converted into multiple signals. In step 2, the signal is first input into the teacher encoding network, and then the encoded signal is input into the teacher decoding network and the bidirectional GRU network respectively. The above teacher encoding and decoding network constitutes a VAE network, which aims to denoise the signal. The following bidirectional gated recurrent unit (GRU) classifies and identifies the signal. Step 3 transfers the fault-free knowledge to the student network, so that the student network only learns fault-free samples in the training process, so that it outputs similar features to the teacher network when it is fault-free. The test stage corresponds to Figure 2 Step 4 of the above. Since the student network only has the fault-free knowledge transferred by the teacher network, the residual error between the student network and the teacher network output of the fault-free sample is small. On the contrary, when the fault sample is input, the cognitive bias of the student network and the teacher network is larger, that is, the output residual error is larger.

[0065] Based on the above description, the teacher network has feature extraction capability after pre-training, the student network learns fault-free features in training, and the fault is judged by feature residual error in testing, which can greatly reduce the data acquisition cost and improve the efficiency and adaptability of fault detection.

[0066] In another exemplary embodiment of the present application, in order to train only using fault-free samples to solve the data acquisition problem, the implementation process of the above step 101 can include:

[0067] Step 11, according to the priori knowledge, a plurality of signal transformations are performed on each fault-free signal in the first sample dataset to obtain signals of different modes. The plurality of signal transformations include: amplitude transformation, frequency transformation and waveform transformation.

[0068] Step 12, in the signal transformation process, pseudo labels corresponding to signals of different modes are generated.

[0069] Step 13, store the fault-free signal, the plurality of signals of different modes and the pseudo labels corresponding to the signals of different modes to obtain the second sample dataset.

[0070] Based on the description of steps 11-13 above, taking the radar system fault detection of the intelligence reconnaissance subsystem in the prosecution system as an example, after the radar system detection signal is known to be fault-free, the fault-free signal is extracted as shown Figure 2 The signal is constructed based on the prior knowledge of the change of the fault signal. By constructing the signal, the feature extraction capability of the teacher network for the fault signal is further enhanced. Based on this, when a fault occurs in the radar system during detection, the detection waveform will change in some specific ways. These changes can usually be detected by analyzing the amplitude, shape, reflection pattern, etc. of the waveform. Here are some common situations and corresponding signal waveform changes:

[0071] (1) Amplitude change: Radar system failure may cause echo intensity to change abruptly, for example, test signal power becomes smaller or test channel parameter loss increases, which may cause the power of the calibration test signal injected by the receiver front end to become smaller, resulting in an abrupt change in echo intensity due to online correction. Increased transmit power measurement error may also cause abnormal echo intensity, such as a smaller in-machine power measurement value, which results in stronger echo intensity, and vice versa.

[0072] (2) Frequency change: The failure of the radar system mainly affects the position and shape of the echo, but it may also indirectly affect the frequency characteristics of the signal. For example, when the signal processor has no master clock signal failure, it may cause a series of alarms in the transmitter amplifier chain circuit and the receiver main channel parameters, indirectly affecting the signal frequency.

[0073] (3) Wave peak shape change: Radar system failure may also cause echo shape to change, such as filamentous, V-shaped notch echo or abnormal detection range. Filamentous echo may be caused by unstable operation in the azimuth of the radar system, while V-shaped notch echo may be caused by unstable operation of the servo structure in the radar system or abnormality in the azimuth control circuit.

[0074] (4) Integrated waveform change: In addition to the possibility of no echo or abnormal echo intensity caused by transmitter failure, characteristic echo notches may also be formed, which start from the center of the station and have clear boundaries. In addition, abnormal radar echo data can be divided into four categories according to the impact: echo shape change, echo position change, echo intensity and range change, etc.

[0075] Based on this, when the radar system fails, its detection waveform will reflect different amplitude, reflection pattern and time delay characteristics. By analyzing these waveform characteristics, the radar system can be fault detected and distinguished, for example:

[0076] A set of fault-free training data (i.e. the first sample data set) where n denotes the corresponding sample sliding window, and N denotes the total number of training samples. Then seven signal transformations are performed for each sample, and the corresponding seven pseudo labels are automatically labeled. The pseudo-labeled dataset is denoted as:

[0077]

[0078] where X n denotes the sample signal, denotes the sample signal after the i-th transformation, denotes the pseudo label corresponding to the i-th transformation.

[0079] These signal transformations can be divided into four categories: no change, amplitude change, frequency change, and waveform change, corresponding to i = 1, 2, 3, 4, respectively. First, the teacher network learns the fault-free signal using the same transformation. In this transformation, no processing is performed on the sample signal X n , so it still represents a fault-free signal. The sample signal after the same transformation can be represented as

[0080] Since in the detection process, once a fault is detected, the signal will be transformed (such as amplitude transformation, frequency transformation, waveform transformation, etc.).

[0081] Taking amplitude increase as an example, the amplitude increase transformation is used to simulate the increase in the amplitude of the detection signal caused by the fault. Assume that a uniform distribution in the interval [1.1, 1.5] is obeyed, denoted as A ~ U[1.1, 1.5], a n ∈ R 1×1 . Then the transformed signal A can be represented as:

[0082]

[0083] where a n denotes the amplitude increase factor.

[0084] Taking frequency increase as an example. Assume that C ~ U[0.5, 0.9], c n ∈ R 1×1 . This transformation compresses the length of the original signal in unit time to c n times, and then the signal is recovered by obtaining a segment of the compressed signal and padding.

[0085] When simulating changes in signal waveform, assume that E ~ U[1.5, 3.0], e n ∈ R 1×1 , the transformed signal is represented as:

[0086]

[0087] In the formula, E represents an energy coefficient, e n represents a dynamic index vector, e n When greater than 1, the peak of the original signal can be made sharp.

[0088] Finally, the corresponding pseudo-labels will be automatically added to the seven transformed signal samples. All the above signal conversion processes can be automatically realized by program code without manual marking.

[0089] In another exemplary embodiment of the present application, in order to solve the problem that the prior art is difficult to effectively capture the timing characteristics and noise interference when processing complex signals, the teacher network provided by the present application combines a variational auto-encoder (VAE) and a bidirectional gated recurrent unit (BiGRU). The VAE is used for signal denoising and feature extraction, and the BiGRU captures the timing characteristics of the signal through bidirectional information transmission. This combination can improve the feature extraction capability of the fault signal, enhance the robustness of the model to noise, and make the fault detection more accurate. Based on this, the teacher network constructed includes two parallel tasks, which are respectively a VAE composed of a teacher encoding network and a teacher decoding network to realize signal denoising, and a teacher encoding network and a BiGRU to realize signal classification and identification.

[0090] Compared with ordinary autoencoders, VAE can learn the approximate probability distribution of hidden variables and generate new data through sampling. Specifically, if there is a random vector Y in a continuous or discrete space V that obeys an unknown distribution P θ (Y), the VAE can learn a parameterized model to approximate P θ (Y) through a series of observable samples, and then use the model to generate new samples that are the same distribution as the real samples. Assuming that the real data in the training data set is x and the hidden variable is z, the probability distribution of z can be inferred by the Bayes formula as follows:

[0091]

[0092] In the formula, p(z|x) is the probability distribution of z given x, which can be calculated by observing a large number of data samples to calculate the distribution of the hidden variable. p(x|z) is the probability distribution of z given x, which can be calculated by observing a large number of data samples to calculate the distribution of the hidden variable. p(z) is the probability distribution of z. p(x) is the probability distribution of x.

[0093] In practical applications, real data x is generally high-dimensional continuous variables, which is difficult to directly solve the distribution of z by formula. Therefore, VAE uses neural network to train a simple distribution to approximate the unknown distribution P θ (z|x), and the closeness of the two is measured by KL (Kullback-Leibler) divergence, as shown in the following formula (5).

[0094]

[0095] In the formula, is the KL divergence, which measures the difference between the approximate distribution and the real posterior distribution p θ (z|x). is the expectation operator, which represents the expectation of the expression in the brackets under the condition that the latent variable z follows the approximate distribution.

[0096] Substituting formula (4) into formula (5) can obtain the following objective function:

[0097]

[0098] In the formula, p θ (x) is the marginal likelihood, p θ (z) is the prior distribution, and p θ (x|z) is the likelihood function.

[0099] Let where is called the variational lower bound, then the parameter estimation problem of P θ (z|x) is converted into the optimization problem of the lower bound . θ is the generation model parameter, and φ is the inference model parameter.

[0100] Since any complex distribution can be mapped by a simple distribution through a function, in order to simplify the solving process, it is generally assumed that follows a Gaussian distribution, and the latent variable z follows a standard normal distribution, that is, P θ (z|x) ~ N(0, 1). m and v represent the mean and variance, which can be learned through the network. In order to be able to use gradient descent method to optimize the network loss of VAE, when performing sampling operation, VAE uses reparameterization method to combine μ and σ 2 into latent variable z through a Gaussian noise e, that is, z = μ + e × σ.

[0101] The network structure of VAE is as follows Figure 4As shown, mainly contains the latent variable encoding part (i.e. teacher encoding network) and decoding generation part (i.e. teacher decoding network), the optimization goal of VAE is to make the learned latent variable distribution as close to the standard normal distribution as possible, and constantly reduce the reconstruction error between the generated sample and the original real sample, and its loss function is shown in equation (7). The first term on the right side of equation (7) represents the KL divergence value between the latent variable distribution and the standard normal distribution, and the second term is the squared error between the original data X and the generated data .

[0102]

[0103] Further, in gradient calculation, when the time step is large or small, the gradient of recurrent neural network (RNN) is prone to the problems of disappearance and explosion, so the memory network is used to replace the neurons in RNN to construct the hidden layer, which is GRU. GRU introduces reset gate and update gate, and modifies the calculation method of hidden state in recurrent neural network. Among them, the reset gate helps to capture the short-term dependence relationship in time series. The update gate helps to capture the long-term dependence relationship in time series, both of which enable GRU to pass information across multiple time windows. The calculation of reset gate and update gate is as follows:

[0104] z t =φ(W z x t +U z h t-1 +b z ) (8)

[0105] r t =φ(W r x t +U r h t-1 +b r ) (10)

[0106] The candidate hidden state of time step t is calculated as:

[0107]

[0108] Finally, the calculation of the hidden state of time step t uses the update gate of the current time step to combine the hidden state of the previous time step and the candidate hidden state of the current time step, which has:

[0109]

[0110] In the formula, z t and r t represent the update gate and the reset gate respectively. x t represents the input feature vector of time step t. h *The output of the hidden layer at the time step t. represents the candidate output. z r c z r are trainable parameter matrices. b z r c represent the bias of the update gate, the reset gate and the output gate respectively. φ(·) is the sigmoid function.

[0111] Although the unidirectional GRU helps to deal with gradient disappearance and gradient explosion, it only considers the previous data information for prediction, which leads to the neglect of the information of early historical data. The BiGRU is an improved version of the unidirectional GRU, which is composed of a forward standard GRU and a reverse standard GRU. The addition of the reverse standard GRU makes it achieve the effect of extracting global data features. At this time, the BiGRU considers the parameter factors of both past and future directions, which greatly improves the accuracy of algorithm prediction. The BiGRU is shown in Figure 5

[0112] Based on the above description, the is taken as the input signal of the teacher network. In the identification task, the purpose is to enable the teacher network to identify different categories of prior fault signal conversion. Therefore, the cross-entropy function is used to calculate the classification loss L C , which is:

[0113]

[0114] In the formula, 1{} is an indication function, is the feature extracted by the teacher encoding network, is the cth pseudo label obtained by the ith conversion of the nth sample of the predicted feature . is the real fault label.

[0115] Based on the above description, the construction process of the teacher encoding network and the teacher decoding network used in the present application can include:

[0116] Step 1. Adopting a recurrent neural network to construct an initial teacher encoding network and an initial teacher decoding network.

[0117] Step 2. Adopting a memory network to replace the neurons of the recurrent neural network in the initial teacher encoding network and the initial teacher decoding network to construct a hidden layer, so as to obtain the teacher encoding network and the teacher decoding network.

[0118] ​​​​​​​​In another example embodiment of the present application, in order to realize the knowledge transfer to the student network, the parameters of the teacher network are fixed and only the parameters of the student network are updated. Based on this, during the training phase of the student network, only the same transformed signal With the noise being input into the teacher network and the student network. Therefore, the student network only learns how to extract the features in the same transformed signal (i.e. the fault-free signal). The training loss function L S may be written as:

[0119]

[0120] wherein, and respectively represent the output of the encoder in the teacher network and the output of the student network, K is and the total size in the hidden space.

[0121] In this embodiment, a multilayer perceptron (MLP) can be used as the student network. The MLP is also called an artificial neural network (ANN), and in addition to the input and output layers, it can have multiple hidden layers. The simplest MLP contains only one hidden layer, i.e. a three-layer structure, as shown in Figure 6 The layers of the multilayer perceptron are fully connected. The bottom layer of the multilayer perceptron is the input layer, the middle is the hidden layer, and the last is the output layer. Figure 6 wherein Layer L1 is the layer in which the network receives the original features, x1-x3 represent three components of the input vector, and +1 is used to increase the model expression capability, and has the same effect as a constant term. Layer L2 is the hidden layer, which receives the linear combination of the input layer and applies a nonlinear activation, is the activation value of the hidden layer neuron, i.e. the activation output of the second layer (hidden layer). Layer L3 is used to output the predicted value. h w,b (x) represents the predicted output of the entire neural network.

[0122] Through training, the student network can have the feature learning capability of the fault-free signal.

[0123] In another example embodiment of the present application, since the student network does not have prior knowledge of the fault, when the fault sample is input, the teacher network will produce a large cognitive bias, i.e. the extracted features have a large residual error. Based on this, the residual error is calculated using the mean squared error (MSE), which has:

[0124]

[0125] where r is the residual, and k is the feature dimension.

[0126] Therefore, whether there is a fault defect can be detected by setting a threshold of the residual. Based on this, the specific generation process of the fault detection threshold adopted in the present application is described.

[0127] Generally, it is difficult to obtain accurate statistical distribution of a complex system, which is not conducive to the determination of the fault detection threshold. The simple threshold estimation method based on Gaussian distribution may lead to serious precision decline. The extreme value theorem is an effective method to solve the problem of unknown distribution form. In addition, an important prerequisite for the application of the extreme value theorem in fault detection is its additional attention to the extreme value point. Therefore, the block maximum (BM) technique (i.e., the BM method) is introduced in the above step 104 to establish a GEV distribution model of the fault detection statistics, so as to determine the fault detection threshold.

[0128] In the field of probability statistics, the extreme value statistics of independent and identically distributed samples is derived from the population distribution or the degenerate distribution. The GEV distribution is a unified form of three classical extreme value distributions, i.e., Gumbel distribution, Frechet distribution and Weibull distribution, thereby solving the selection problem of the extreme value distribution. Assuming that X1,X2,...,X n are independent and identically distributed sample sequences obeying distribution F, and M n =max(X1,X2,...,X n ). If there exist constant sequences a n >0 and b n , the following equation (15) is established.

[0129]

[0130] The non-degenerate distribution G(z) conforms to the generalized extreme value distribution (GEV distribution), and has:

[0131]

[0132] where ξ is a shape parameter, σ is a size parameter, and μ is a location parameter. These parameters satisfy 1+ξ(z-μ) / σ>0,-∞<ξ<∞ and-∞<μ<∞. In order to obtain the threshold when the significance level is α (i.e., the extreme value quantile z α when the significance level is α), the extreme value quantile of the GEV distribution is estimated by using the inverse equation represented by the following formula (17).

[0133]

[0134] In the fault detection, the extreme value of the training sample data in the normal state is used to model the GEV distribution, thereby determining the fault detection threshold.

[0135] As an extreme value construction method, the BM technique focuses on dividing the data into equal size non-overlapping blocks and calculating the extreme value in each block. Suppose the data Y1, Y2,..., Y s are divided into n' blocks with length l, where the data are the fault detection statistics of the training sample data in the offline training phase. Perform descending order sorting on all blocks to construct n' order statistics z1, z2,..., z n' , calculate the maximum value of n' blocks, and get the extreme value sequence X1, X2,..., X n' In this application, the maximum likelihood estimation (MLE) technique provides a strategy for estimating the unknown parameters of the GEV distribution using the extreme value sequence, where the estimates of the unknown parameters are the maximum log-likelihood function ξ, σ, μ.

[0136] In the BM technique, the number of blocks n' is an important parameter that affects the quality of the fault detection threshold. Too few blocks will result in insufficient extreme values, while too many blocks will cause the estimate to be biased. The number of blocks in the BM technique is determined using the Kolmogorov-Smirnov (KS) hypothesis testing method. This method is robust, widely applicable, and not sensitive to scale, and can measure whether there is a significant difference between two distributions. The KS statistic of the extreme value sequence is calculated as follows:

[0137]

[0138] where KS i' is the Kolmogorov-Smirnov statistic, which measures the maximum deviation of the empirical distribution from the theoretical distribution, F() is the generalized extreme value distribution, and X i ' is the ith' extreme value that meets the independent and identically distributed condition.

[0139] Selecting different n' will result in different values of the KS statistic. The number of blocks corresponding to the minimum value of the KS statistic is selected as the optimal. Then, the fault detection threshold is calculated according to equation (17), the comparison between r and the fault detection threshold is used to determine whether the component (subsystem) has failed, and finally the system is determined to have entered a fault state by judging each component (subsystem) in the control system.

[0140] In summary, the present application has the following advantages over the prior art:

[0141] 1. The present application uses a teacher-student self-supervised learning model guided by prior knowledge, which only uses fault-free samples for training, solving the data acquisition problem. The teacher network has feature extraction capability after pre-training, and the student network learns fault-free features during training. During testing, the fault is judged by the feature residual, which greatly reduces the data collection cost and improves the efficiency and adaptability of fault detection.

[0142] 2、The prior art is difficult to effectively capture the timing characteristics and noise interference when processing complex signals. The teacher network of the present application combines a variational autoencoder (VAE) and a bidirectional gated recurrent unit (BiGRU). The VAE is used for signal denoising and feature extraction, and the BiGRU captures the timing characteristics of the signal through bidirectional information transmission. This combination can improve the feature extraction capability of the fault signal, enhance the robustness to noise, and make the fault detection more accurate.

[0143] 3、The present application introduces the block maximum (BM) technique and the generalized extreme value distribution (GEV) model, determines the optimal block number through maximum likelihood estimation and Kolmogorov-Smirnov hypothesis testing, and accurately models the distribution of the fault detection statistic, so as to more accurately determine the fault threshold and improve the accuracy and reliability of fault detection.

[0144] Further, the above-mentioned scheme provided by the present application can automatically realize signal conversion and pseudo-label generation through program code, simulate the amplitude, frequency and waveform changes of the fault signal, not only can improve the efficiency of data preparation, but also can reduce human errors, to further improve the performance of the fault detection model.

[0145] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and its internal structure diagram can be as shown in Figure 7 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store ship equipment fault detection data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a ship equipment fault detection method.

[0146] Those skilled in the art can understand that Figure 7 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0147] In an exemplary embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above method embodiments when executing the computer program.

[0148] In an exemplary embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program implementing the steps in the above method embodiments when executed by a processor.

[0149] In an exemplary embodiment, a computer program product is provided, comprising a computer program, and the computer program implementing the steps in the above method embodiments when executed by a processor.

[0150] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0151] It can be understood by those skilled in the art that all or part of the processes in the above method embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (Resistive Random Access Memory, RRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

[0152] The database involved in each of the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, and the like, without being limited thereto. The processor involved in each of the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, and the like, without being limited thereto.

[0153] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but it should be considered that any combination of the technical features is within the scope of the present disclosure, as long as there is no contradiction.

[0154] The principles and implementation modes of the present application are described by applying specific examples herein, and the above embodiments are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation modes and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present description should not be understood as a limitation of the present application.

Claims

1. A method of ship equipment failure detection, characterized by, The method comprises the following steps: Obtaining the fault-free signals of the ship equipment to form a first sample data set; According to prior knowledge, the fault-free signals in the first sample data set are transformed into signals of multiple different modes to form a second sample data set; In the framework of the teacher-student network model, a bidirectional gated recurrent unit and a variational autoencoder are introduced into the teacher network to construct an initial detection model; The fault-free signals in the first sample data set are used as the input of the student network in the initial detection model, and the signals of different modes corresponding to the input fault-free signals of the student network in the second sample data set are used as the input of the teacher network to train the initial detection model until a set condition is reached, and a trained initial detection model is obtained; In the training process of the initial detection model, the BM method and the extreme value theory are used to obtain the DEV distribution; The trained initial detection model is used as the ship equipment detection model; Obtaining the signals of the ship equipment to be detected; The signals of the ship equipment to be detected are input into the teacher network and the student network of the ship equipment detection model to obtain the detection results of the ship equipment; Based on the GEV distribution, a fault detection threshold is determined; Based on the detection results of the ship equipment and the fault detection threshold, the fault detection of the ship equipment is realized.

2. The ship equipment failure detection method according to claim 1, characterized in that, According to prior knowledge, the fault-free signals in the first sample data set are transformed into signals of multiple different modes to form a second sample data set, which comprises the following steps: According to prior knowledge, each fault-free signal in the first sample data set is subjected to multiple signal transformations to obtain signals of multiple different modes; In the signal transformation process, pseudo-labels corresponding to the signals of different modes are generated; The fault-free signals, the signals of multiple different modes, and the pseudo-labels corresponding to the signals of different modes are stored to obtain the second sample data set.

3. The ship equipment failure detection method according to claim 2, characterized in that, The multiple signal transformations include amplitude transformation, frequency transformation, and waveform transformation.

4. The ship equipment failure detection method according to claim 1, characterized in that, The variational autoencoder comprises a teacher encoding network and a teacher decoding network; The teacher encoding network is connected with the teacher decoding network and the bidirectional gated recurrent unit, and the teacher decoding network is connected with the bidirectional gated recurrent unit; The teacher encoding network is used to encode the input signals to obtain encoded signals, and the teacher decoding network is used to decode the encoded signals to obtain decoded signals; The bidirectional gated recurrent unit is used to obtain fault detection and classification results based on the encoded signals and the decoded signals.

5. The ship equipment failure detection method according to claim 4, characterized in that, The construction process of the teacher encoding network and the teacher decoding network comprises the following steps: An initial teacher encoding network and an initial teacher decoding network are constructed by using a recurrent neural network; A hidden layer is constructed by using a memory network to replace the neurons of the recurrent neural network in the initial teacher encoding network and the initial teacher decoding network to obtain the teacher encoding network and the teacher decoding network.

6. The ship equipment failure detection method according to claim 1, characterized in that, An over-layer perceptron is used as the student network in the initial detection model.

7. The ship equipment failure detection method according to claim 1, characterized in that, The process of determining the fault detection threshold based on the GEV distribution comprises the following steps: Adopting determining the extreme quantile of the GEV distribution, and taking the determined extreme quantile as the fault detection threshold; where z α denotes the extreme quantile for a significance level of a, ξ denotes the shape parameter, σ denotes the scale parameter, and μ denotes the location parameter.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the ship equipment fault detection method in any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, which is executed by a processor, implements the method for detecting a failure of a ship equipment according to any one of claims 1-7.

10. A computer program product comprising a computer program, characterized in that, The computer program, which is executed by a processor, implements the method for detecting a failure of a ship equipment according to any one of claims 1-7.

Images