Ship mechanism element fault residual available boundary quantification method

Abstract

The invention discloses a ship mechanism element fault residual available boundary quantification method, and belongs to the technical field of combat readiness state control, and the method comprises the steps: obtaining a full-life data set of ship mechanism elements, and carrying out the preprocessing of the data in the full-life data set, and obtaining an input data set; constructing a time convolutional neural network model, and performing residual service life prediction uncertainty quantitative representation based on a Bayesian theory to construct a Bayesian-TCN model; a Bayesian-TCN model is trained by an input data set, and the trained Bayesian-TCN model is used to carry out RUL prediction with uncertainty quantization. By adjusting the posterior distribution of the network parameters and regularizing the hyper-parameter weight, the probability of over-fitting is reduced, the probability explanation of the network model can be provided, the RUL prediction result with uncertainty representation can be generated, the information about the prediction accuracy can be provided, and the method has more instructive significance for the actual maintenance decision.

Description

Technical Field

[0001] The present invention belongs to the technical field of combat readiness control, and more specifically, relates to a method for quantifying the residual available margin of a ship mechanism element failure. Background Art

[0002] As an important carrier of maritime combat forces, ships have an outstanding strategic position. On the one hand, with the continuous advancement of science and technology, the mechanical components of modern ships have become increasingly complex. This not only improves the combat effectiveness of ships, but also brings challenges in maintenance and management. Especially in the quantification of the remaining available boundary, the traditional methods based on experience or simple models still have limitations in accuracy when facing complex ship mechanical components, resulting in unreasonable quantification boundaries. On the other hand, the performance degradation characteristics of the mechanical components of large and complex equipment are easily submerged, and the mapping relationship between the acquired data signals and the equipment performance degradation status is unclear, making the quantification of its remaining available boundary more difficult.

[0003] In this context, it is urgent to propose a method to quantify the remaining available margin of ship mechanical component failures, so as to guide ship maintenance support and equipment decision-making. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention proposes a method for quantifying the remaining available boundary of ship mechanism element failures, which helps to quantify and analyze the remaining service life of ship mechanical components under failure conditions, and can be used to guide the maintenance and installation decisions of ship mechanical components.

[0005] To achieve the above object, the present invention provides a method for quantifying the remaining available margin of a ship mechanism element failure, comprising:

[0006] Acquire the full life data set of ship structure elements, and pre-process the data in the full life data set to obtain an input data set;

[0007] Construct a temporal convolutional neural network model, and quantitatively characterize the uncertainty of remaining useful life prediction based on Bayesian theory to construct a Bayesian-TCN model;

[0008] The Bayesian-TCN model is trained by the input data set, and the trained Bayesian-TCN model is used to perform RUL prediction with uncertainty quantification.

[0009] In some optional implementation schemes, preprocessing the data in the full life data set to obtain the input data set includes:

[0010] The data in the full life data set are normalized, standardized, processed with sliding time windows, and RUL labels are constructed.

[0011] In some optional implementation schemes, in the construction of the temporal convolutional neural network model, causal convolution operations are used to process time series data, and an expanded convolution algorithm is introduced to remove the pooling layer.

[0012] In some optional implementation schemes, the uncertainty quantification characterization of the remaining useful life prediction based on Bayesian theory includes:

[0013] In the Bayesian framework, the distribution function of uncertainty, namely the prior distribution, is established based on historical information;

[0014] Construct a Bayesian update model of the distribution function of uncertainty through Bayesian theory;

[0015] According to the whole life cycle data of ship structure elements that are continuously obtained, the distribution function of the uncertainty quantity, namely the posterior distribution, is updated and corrected.

[0016] In some optional embodiments, for the Bayesian-TCN model, a probability distribution is placed on the output of the Bayesian network to model aleatoric uncertainty.

[0017] In some optional implementation schemes, for the Bayesian-TCN model, solving the minimum KL divergence value of the variational distribution and the posterior distribution is transformed into solving the maximum ELBO problem.

[0018] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0019] Aiming at the problems that the cognition and accidental uncertainty of the quantitative evaluation process will affect the prediction results, and the reliability of the quantification method based on traditional neural network is low, this paper proposes an uncertainty quantitative evaluation model combining Bayesian network and time convolutional neural network, which can output a reasonable prediction interval of the remaining available boundary of the ship mechanism element failure / abnormality. This method can accurately extract the fault / abnormal depth characteristics and obtain the quantitative evaluation interval value of the remaining available boundary under the influence of uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of a residual available boundary quantization process provided by an embodiment of the present invention;

[0021] Figure 2 It is a schematic diagram of sliding time window processing of a multi-sensor time series provided by an embodiment of the present invention;

[0022] Figure 3 is a schematic diagram of dilated causal convolution provided by an embodiment of the present invention;

[0023] Figure 4 is a residual connection schematic diagram provided by an embodiment of the present invention;

[0024] Figure 5 is a schematic diagram of uncertainty classification provided by an embodiment of the present invention;

[0025] Figure 6 It is a schematic diagram of a remaining life prediction process based on Bayesian-TCN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0027] The present invention is suitable for quantifying the residual available margin of faults of ship mechanical components, and its prediction results have good stability and reliability. The residual available margin is quantified by using a method based on a time convolutional neural network, considering the quantitative characterization of multi-source uncertainty, constructing a quantitative evaluation model of the residual available margin considering uncertainty, and obtaining the residual available margin evaluation interval.

[0028] When traditional neural networks quantify the residual boundaries of complex equipment, there are problems such as being unable to represent the uncertainty of the model and simplifying the feature information, resulting in low accuracy. The present invention reduces the probability of overfitting by adjusting the posterior distribution of network parameters and regularizing the hyperparameter weights. At the same time, it can provide a probabilistic interpretation of the network model, generate RUL prediction results with uncertainty representation, and provide information on the accuracy of the prediction, which is more instructive for actual maintenance decisions.

[0029] like Figure 1 As shown, a method for quantizing a remaining available boundary provided by an embodiment of the present invention includes the following aspects:

[0030] 1. Data sample preprocessing

[0031] Data sample preprocessing mainly includes normalization, standardization, sliding time window processing, and RUL label construction.

[0032] 1) Normalization

[0033] In order to reduce the impact of the dimension of sensor data on the prediction performance, the sensor parameters are normalized so that the parameter values ​​are limited to the range of [0,1].

[0034]

[0035] 2) Standardization

[0036] The method used for data standardization is the Z-score method, which transforms the data to a range with a mean of 0 and a standard deviation of 1, thereby eliminating the dimensional relationship between variables and avoiding the influence of data outliers and extreme values.

[0037]

[0038] 3) Sliding time window processing

[0039] For a multi-sensor time series of length n, the input data is segmented using a time window of size T sliding one time step along the time axis each time, and n-T+1 time windows can be obtained, such as Figure 2 shown.

[0040] 4) URL label structure

[0041] The RUL tag is constructed using a piecewise linear function. The lifetime threshold set by the present invention is [T imax / 2], that is, the maximum life span T of the i-th mechanism imax The quotient when divided by 2 is rounded to an integer.

[0042]

[0043] Where, T imax is the maximum lifespan of the ith mechanism, T ij is the jth lifespan value of the ith institution.

[0044] 2. Temporal Convolutional Neural Network Model Construction

[0045] 1) One-dimensional convolution algorithm

[0046] One-dimensional convolution operation is simple and much faster. The formula is as follows:

[0047]

[0048] The length of time series data will become shorter after one-dimensional convolution operation. In order to ensure that the length of the time series before and after convolution remains unchanged, the data needs to be padded: according to the time sequence of the time series data from left to right, the data 0 is usually padded at the far left of the sequence to ensure that the input and output lengths are the same.

[0049] 2) Causal Convolution Algorithm

[0050] The observations are x1, x2, …, x t , the predicted values ​​are y1, y2, …, y t , the filter is F=(f1,f2,...,f K ), the time series is X=(x1,x2,...,x T ), in xt The causal convolution at is:

[0051]

[0052] Where K is the number of convolution kernels. Assume that the last two nodes of the input layer are x t-1 , x t , the filter function is F = (f1, f2), and the prediction value of the last node of the first hidden layer is y t =f1x t-1 +f2x t .

[0053] When the time series data is too long and the network needs to remember more distant information, the number of convolution layers must be increased, which greatly increases the complexity of the model, leading to a series of problems such as gradient disappearance and training difficulties. In order to solve this problem, the present invention uses dilated convolution in conjunction with causal convolution.

[0054] 3) Dilated Convolution Algorithm

[0055] Compared with the traditional convolutional neural network (CNN), the dilated convolution does not use a pooling layer, so the feature information is not simplified during the training process, so that all historical information can be retained to obtain more stable and reliable prediction data.

[0056] The theoretical basis of dilated convolution is: the observations are x1, x2, ..., x t , the predicted values ​​are y1, y2, …, y t , the filter function is F=(f1,f2,...,f K ), the time series is X=(x1,x2,...,x T ), in x t The dilated convolution with a dilation factor of d at is:

[0057]

[0058] Where K is the number of convolution kernels.

[0059] Assume that the last five nodes of the first hidden layer are x t-4 ,x t-3 ,x t-2 ,x t-1 ,x t , the last node of the second hidden layer is y t , the filter function is F = (f1, f2, f3), according to the formula y t =f1x t-2d +f2x t-d +f3x t , at this time the expansion factor d=2.

[0060] like Figure 3 As shown in the figure, as the expansion factor increases, the range of historical data that can be obtained by the dilated convolution becomes wider. In the input layer, the expansion factor d = 1 is a normal convolution operation, and in the hidden layer, the expansion factor increases exponentially, that is, d = 2. i (i=1,2…,n), for example Figure 3 In this case, d is 1, 2, and 4, and the receptive field is expanded accordingly, that is, (K-1)d+1, where K is the size of the convolution kernel. Therefore, as the dilation factor d increases layer by layer, the required hidden layers decrease, thereby reducing the complexity of the TCN network structure and improving the network's operational efficiency. Due to the use of dilated convolution, the receptive field is expanded. In order to avoid insufficient sequence length, padding (usually 0) is required for each layer. The padding size is (K-1)d, where K is the size of the convolution kernel.

[0061] 4) Residual Connection

[0062] In order to solve the "vanishing gradient" problem of deep neural networks, the present invention introduces a residual network, which uses residual blocks to stack and deepen the network layers. Each residual block contains a branch to derive a series of conversion functions F, and the output of the conversion function is added to the input of the next residual block, as shown in the following formula:

[0063] ο=Activation(X+F(X)) (7)

[0064] Where: ο is the output value of the residual block; X is the input value of the residual block; F(·) is the identity mapping function.

[0065] In the residual module, it consists of two layers of dilated causal convolution kernel nonlinear mapping functions, and weight normalization and Dropout layers are added to each layer to regularize the network, such as Figure 4 Adding the Dropout layer can effectively solve the problem of overfitting, reduce the common dependencies between neurons, and make the model more adaptable to noise interference.

[0066] 3. Quantification of uncertainty in remaining useful life prediction

[0067] When modeling ship mechanical components for remaining life prediction, it is necessary to consider both accidental uncertainty and epistemic uncertainty, such as Figure 5 As shown, the former is caused by the noise in the observation data, and the latter is caused by the limited observation data and thus the limited cognition, which makes it impossible to know the full picture of the model.

[0068] The present invention uses the output f of the Bayesian network model to ωA probability distribution is placed on (x) to model aleatoric uncertainty, using a Gaussian distribution:

[0069]

[0070] In the above formula, the variance σ n 2 Set it as a constant. If the diversity of noise in the training data is taken into account, the heteroscedastic form σ in which the variance varies with the input data can also be used. n (x i ) 2 .

[0071] The cognitive uncertainty is simulated by probabilistically modeling the parameters of the Bayesian network model. The model parameter is ω, and the prior distribution of the parameter is:

[0072]

[0073] In addition to the standard Gaussian distribution, the prior of ω can also be in the form of mixed Gaussian. When the observed sample is D, the expectation is to obtain the posterior distribution of the parameter p(ω|D). During the training process, the posterior distribution is continuously updated based on the prior knowledge and training samples, and variational inference is used to solve the approximate distribution. After the training is completed, the parameter ω in the model is also in the form of probability distribution. For the new test set sample x * , the distribution of network predictions is expressed as:

[0074] p(y * ∣x * )=E p(ω∣D) [p(y * ∣x * ,ω)] (10)

[0075] 4. Uncertainty Quantification Modeling Based on Bayesian-TCN

[0076] In data-driven RUL prediction, and Collect data for the equipment system status and the corresponding remaining service life, x (i) and (i) is the data of the i-th sample in the data set, and N represents the number of training samples. The remaining life prediction based on the Bayesian network can be expressed as a y (i) In x (i) This method mainly consists of two parts: a regression model based on Bayesian networks and variational inference.

[0077] 1) Building a Bayesian Neural Network (BNN)

[0078] Given training samples X and Y, BNN consists of a prior distribution p(ω) and a Bayesian regression likelihood function The output y = f ω (x), usually defined as l(y (i) |f ω (x (i) )) is a Gaussian distribution. The model parameters ω are independent. The BNN model is trained by jointly learning X and Y, and then the posterior distribution of the model parameters is obtained by Bayes' theorem:

[0079]

[0080] Based on p(ω|X,Y), BNN model y=f ω (x) can be used for subsequent reasoning of uncertainty quantification. Assuming that the newly observed state detection data At time t, the device termination time is The formula can be Make predictions.

[0081] 2) Variational Inference

[0082] Variational inference is used to solve the problem that the posterior distribution p(ω|X,Y) is difficult to calculate in Bayesian deep learning. The key idea is to approximate the actual posterior distribution through a probability distribution that is easy to evaluate and further infer. This distribution is called a variational distribution. There are usually two steps to implement variational inference: the first step is to select the probability distribution family q(ω|θ) parameterized by θ as the variational distribution family; the second step is to minimize the KL divergence of q(ω|θ) and p(ω|X,Y) with respect to θ to find the optimal variational distribution. The KL divergence formula is as follows:

[0083]

[0084] Then substitute p(ω|X,Y) into the above formula:

[0085]

[0086] Where log p(X,Y) is the logarithmic marginal likelihood function, which does not depend on the parameter θ, so it can be regarded as a constant term; the first and second terms are the variational free energy F(X,Y,θ) defined as follows:

[0087] F(X,Y,θ)=KL(q(ω|θ)||p(ω))-E q(ω|θ) logp(X,Y|ω) (14)

[0088] In the formula, the first term is the KL divergence value of the variational distribution q(ω|θ) and the prior distribution p(ω), also known as the complex cost function; the second term is the likelihood expectation value of the variational distribution, also known as the likelihood cost function. After transforming the KL divergence term, the free energy function is as follows:

[0089] F(X,Y,θ)=E q(ω|θ) logq(ω|θ)-E q(ω|θ) logp(ω)-E q(ω|θ) logp(X,Y|ω) (15)

[0090] The three terms in the above formula are all about the expected value of the variational distribution q(ω|θ), so the free energy loss function can be approximated by the parameter ω (i) function.

[0091]

[0092] The KL divergence formula can be written as:

[0093] KL(q(ω|θ)||p(ω|X,Y))=F(X,Y,θ)+logp(X,Y) (17)

[0094] Since logp(X,Y) is independent of the parameter θ, to minimize the KL divergence value, we only need to minimize the free energy loss function F(X,Y,θ). Transforming the above formula yields:

[0095] -F(X,Y,θ)=logp(X,Y)-KL(q(ω|θ)||p(ω|X,Y)) (18)

[0096] Since the negative free energy function is called the evidence lower limit (ELBO) L(X,Y,θ), the above formula can be written as: L(X,Y,θ)=logp(X,Y)-KL(q(ω|θ)||p(ω|X,Y)).

[0097] Since KL divergence is non-negative, the problem is transformed from solving the minimum KL divergence value of the variational distribution and the posterior distribution to solving the maximum ELBO problem. Use the optimizer to obtain the best parameter θ, that is, minimize KL(q(ω|θ)||p(ω|X,Y)), and then use the variational distribution q(ω|θ) to approximate the posterior distribution p(ω|X,Y), and train it through a deep Bayesian network. For a given The predicted value can be solved by the following formula:

[0098]

[0099] 5. Remaining available boundary prediction

[0100] like Figure 6 As shown in the figure, the test samples are input into the constructed time convolution-based neural network to obtain the RUL prediction results with uncertainty quantification, so its mean is used as the point estimate of the RUL prediction, and the mean and variance are used to construct the RUL prediction interval.

[0101] The present invention has the following advantages by adopting the above technical solution:

[0102] (1) To address the problem that traditional convolutional neural networks are not suitable for quantifying the residual available margins of ship mechanical component failures, the present invention combines causal convolution with dilated convolution algorithms to construct a temporal convolutional neural network.

[0103] The causal convolution operation is used to process time series data to ensure that the network only uses the information before each time step in the sequence, solving the problem of future information "leaking" into the past caused by traditional convolution operations. On the other hand, the present invention introduces an expansion convolution algorithm. Compared with the traditional convolutional neural network CNN, it does not use a pooling layer, so the feature information is not simplified during the training process, so that all historical information can be retained, improving the stability and reliability of the quantitative results of the remaining available boundaries of ship mechanical component failures.

[0104] (2) In view of the fact that various uncertainties in the system will affect the reliability of the quantitative results of the residual available margin of the ship's mechanical component failure, the present invention models the simulated accidental uncertainty and cognitive uncertainty.

[0105] In the Bayesian framework, the uncertainty distribution function, i.e., the prior distribution, is first established based on historical information; then the Bayesian update model of the uncertainty distribution function is constructed through the Bayesian theory; finally, as new information is continuously acquired, the distribution function of the uncertainty, i.e., the posterior distribution, can be continuously updated and corrected, thereby gradually and accurately quantifying its uncertainty. When modeling the remaining service life of ship mechanical components, the present invention considers various types of uncertainty in the system and quantifies the comprehensive impact of these sources of uncertainty on the remaining service life prediction, thereby further improving the reliability of the prediction results.

[0106] (3) In order to solve the problem of complex calculation process of quantifying the remaining available boundary after considering uncertainty, the present invention uses variational inference method to simplify the calculation

[0107] The key problem of remaining useful life prediction is transformed from solving the minimum KL divergence value of the variational distribution and the posterior distribution to solving the maximum ELBO problem. The optimizer is used to obtain the optimal parameter θ, that is, to minimize KL(q(ω|θ)||p(ω|X,Y)), and then the variational distribution q(ω|θ) is used to approximate the posterior distribution p(ω|X,Y). Through deep Bayesian network training, For a given The predicted value can be solved by the following formula:

[0108] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0109] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for quantifying the residual available margin of ship structural element failures, characterized in that: include: Acquire the full life data set of ship structure elements, and pre-process the data in the full life data set to obtain an input data set; Construct a temporal convolutional neural network model, and quantitatively characterize the uncertainty of remaining useful life prediction based on Bayesian theory to construct a Bayesian-TCN model; The Bayesian-TCN model is trained by the input data set, and the trained Bayesian-TCN model is used to perform RUL prediction with uncertainty quantification.

2. The method according to claim 1, characterized in that The step of preprocessing the data in the full life data set to obtain the input data set includes: The data in the full life data set are normalized, standardized, processed with sliding time windows, and RUL labels are constructed.

3. The method according to claim 2, characterized in that In the construction of the temporal convolutional neural network model, causal convolution operations are used to process time series data, and an expanded convolution algorithm is introduced to remove the pooling layer.

4. The method according to claim 3, characterized in that: The uncertainty quantitative characterization of remaining useful life prediction based on Bayesian theory includes: In the Bayesian framework, the distribution function of uncertainty, namely the prior distribution, is established based on historical information; Construct a Bayesian update model of the distribution function of uncertainty through Bayesian theory; According to the whole life cycle data of ship structure elements that are continuously obtained, the distribution function of the uncertainty quantity, namely the posterior distribution, is updated and corrected.

5. The method according to claim 4, characterized in that For the Bayesian-TCN model, a probability distribution is placed on the output of the Bayesian network to model aleatoric uncertainty.

6. The method according to claim 5, characterized in that For the Bayesian-TCN model, solving the minimum KL divergence value of the variational distribution and the posterior distribution is transformed into solving the maximum ELBO problem.

Images