Ship key component health state prediction method based on MPSO-RBF neural network

Abstract

The invention discloses a ship key component health state prediction method based on an MPSO-RBF neural network, and belongs to the technical field of prediction and health management, and the method comprises the steps: carrying out the preprocessing of a multi-source heterogeneous monitoring parameter, selecting a monitoring parameter related to the health state of a component as a state feature, and constructing a known state signal feature sequence; a known state signal feature sequence is input into a maturely trained MPSO-RBF neural network model, an output value serves as a subsequent state feature prediction value of a key component, and the MPSO-RBF neural network model introduces an improved PSO algorithm on the basis of an RBF neural network model to conduct self-adaptive selection on network parameters; and determining and evaluating the health state of the current equipment according to the state feature predicted value and the corresponding relationship between the signal feature value and the health state established by priori knowledge. According to the method, the precision of equipment prediction is improved by using the combination of the improved PSO and the RBF neural network.

Description

Technical Field

[0001] The present invention belongs to the field of prediction and health management (PHM) technology, and more specifically, relates to a method for predicting the health status of key components of a ship based on an MPSO-RBF neural network. Background Art

[0002] With the development of technology, ships are becoming increasingly complex, and the demand for health status monitoring and prediction of key components such as sliding bearings and rolling bearings is becoming more and more urgent. Most of the existing key component health status prediction methods rely on limited historical data and simple statistical models, with low prediction accuracy and difficulty in accurately predicting the future status of components. Most existing methods only consider the influence of single factors such as vibration, noise, temperature, and pressure, ignoring the interaction between different factors. The lack of comprehensive evaluation of the influence of multiple factors limits the reliability and accuracy of the prediction results. Given the complexity and uncertainty of the state changes of key ship components, traditional prediction methods may find it difficult to guarantee the effectiveness of their detection and prediction.

[0003] Although there are some intelligent learning algorithms based on failure and degradation modeling or data-driven for diagnosis and prediction of large mechanical equipment, there are few reports on health status prediction methods for key ship components, and the analysis and research in this application field are relatively scarce. Therefore, drawing on the mature development of intelligent diagnosis and prediction methods in the mechanical field, a health status prediction method based on MPSO-RBF (Modified PSO-RBF) neural network is proposed to meet the health status assessment needs of key ship components, which is of great significance to improving the safety and reliability of ship equipment. 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 predicting the health status of key ship components, which can be used for the prediction and evaluation of the status of key components with multiple monitoring parameters such as generators, gearboxes, and propulsion shaft systems of ships, so as to solve the current situation of relative scarcity and underdevelopment of existing intelligent detection methods for ships. The combination of improved PSO and RBF neural network improves the accuracy of equipment prediction and has higher reliability than the traditional single neural network prediction algorithm.

[0005] To achieve the above object, the present invention provides a method for predicting the health status of key components of a ship based on an MPSO-RBF neural network, comprising:

[0006] After preprocessing the multi-source heterogeneous monitoring parameters, the monitoring parameters related to the health status of the components are selected as state features to construct a known state signal feature sequence;

[0007] The known state signal feature sequence is input into the mature MPSO-RBF neural network model, and the output value is used as the subsequent state feature prediction value of the key component. The MPSO-RBF neural network model introduces an improved PSO algorithm to adaptively select network parameters based on the RBF neural network model, and is trained using a large amount of labeled historical sample data of key ship equipment.

[0008] According to the predicted value of the state feature and the correspondence between the signal feature value and the health state established with prior knowledge, the health state of the current device is determined and evaluated.

[0009] In some optional embodiments, the known state signal feature sequence is input into a maturely trained MPSO-RBF neural network model, and the output value is used as the subsequent state feature prediction value of the key component, including:

[0010] From the known state signal feature sequence, the last MN feature values ​​are selected to form the input value, which is input into the mature MPSO-RBF neural network to obtain a vector O containing N output values. This vector O is the state prediction value obtained.

[0011] Continue to select the last M-2N eigenvalues ​​from the known state signal feature sequence and combine them with O to form a new input F, input the input F into the mature MPSO-RBF neural network, and obtain a new vector O′ containing N output values. This vector O′ is the state prediction value obtained in this prediction.

[0012] The last M-3N eigenvalues ​​are selected and combined with O to form a new input F, and so on, until the preset prediction length is reached as the final prediction output eigenvalue result.

[0013] Each time a prediction is made, a predicted value of length N will be added, which will be put back into the input to predict the next segment. To ensure that the input length is always M, the original feature value will be reduced by N in sequence until the predicted length reaches the preset length.

[0014] In some optional embodiments, the training process of the MPSO-RBF neural network model includes:

[0015] Obtain key component monitoring signals and obtain a series of equally spaced data samples {x1, x2, …, x n}, x n is the nth sample data, n is the number of samples;

[0016] Select state features and find the sequence {x1,x2,…,x n} corresponding to the eigenvalues ​​{X1,X2,…,X n};

[0017] An initial RBF neural network is established, with the number of output nodes of the network set to N, and the number of input nodes set to M - N, where N < M; the center value c of the RBF network i , width δ i and weight ω i are encoded as a particle, and the total dimension of the particle is D = h(I + O + 1), where I represents the RBF input dimension and O is the output dimension; the fitness function of the particle swarm is set, the particle swarm size m and the maximum number of iterations t are set, and the particle swarm is initialized;

[0018] In the eigenvalue sequence {X1, X2, …, X n}, a new set of n - M + 1 groups of consecutive eigenvalue sequences {Y1, Y2, …, Y n-M+1} is obtained in a recursive manner. Among them, starting from X1, M consecutive eigenvalues are taken to form a new sample Y1 = [X1, X2, …, X M , then starting from X2, another M consecutive eigenvalues are taken to form the second new sample Y2 = [X2, X3, …, X M+1 , then the third new sample Y3 = [X3, X4, …, X M+2 , and so on. The last new sample is Y n-M+1 = [X n-M+1 , X n-M+2 , …, X n ;

[0019] Each sequence in the new eigenvalue sequence set {Y1, Y2, …, Y n-M+1} is divided into two parts. The first part contains the first M - N eigenvalues of the sequence, and the second part contains the remaining N eigenvalues, thus obtaining two corresponding new sets. Let the two new sets obtained be {P1, P2, …, P n-M+1} and {T1, T2, …, T n-M+1}, then there is Y i = [P i T i , P i contains M - N eigenvalues, and T i contains N eigenvalues, i = 1, 2, …, n - M + 1;

[0020] Taking {P1, P2, …, P n-M+1} as the input value of the MPSO - RBF neural network and the corresponding {T1, T2, …, T n-M+1} as the output value, the initial MPSO - RBF neural network is trained, the fitness of each particle is calculated, the individual extreme value and global extreme value of the particle are updated, and the position of the particle is updated;

[0021] Determine whether the number of iterations meets the maximum number of iterations. If not, repeat the process to set {P1, P2, …, P n-M+1} as the input value of the MPSO-RBF neural network, and the corresponding {T1,T2,…,T n-M+1} as the output value, train the initial MPSO-RBF neural network, calculate the fitness of each particle, update the individual extreme value and global extreme value of the particle, and update the position of the particle. If satisfied, map the optimal particle position vector to the various parameters of the neural network, and finally obtain a mature MPSO-RBF neural network prediction model.

[0022] In some optional implementation schemes, a nonlinear inertia weight w(t) is introduced into the MPSO-RBF neural network, and the original basic particle swarm algorithm formula is modified as follows:

[0023]

[0024]

[0025] Among them, t is the current iteration number, c1 and c2 are learning factors, r1 and r2 are random numbers, and w max and w min are the maximum and minimum values ​​of w, t max is the maximum number of iterations, t p Deployment for global extreme evolution stagnation, t g is the global extreme stagnation deployment threshold that needs to be disturbed, U(0,1) is a random number uniformly distributed in the interval (0,1), represents the flying speed of the i-th particle in the d-th dimension in the t-th iteration, x pb represents the best historical position of the individual particle, x gb Represents the global optimal particle position.

[0026] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0027] The present invention provides a method for predicting the health status of key ship components based on MPSO-RBF neural network, aiming to improve the safety of the operation of key ship components. By introducing improved PSO into the RBF neural network to optimize the network parameters such as the center and width of the hidden layer basis function and the connection weight between the hidden layer and the output layer, it has higher fitting than the traditional RBF model and the RBF model guided by PSO, improves the reliability of the prediction results, and finally realizes the accurate prediction of the health status of the components, which can provide strong support for early fault warning to a certain extent, thereby reducing the occurrence of faults, extending the service life of equipment, improving maintenance efficiency, and optimizing the maintenance decision-making process. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of a state prediction process based on an MPSO-RBF neural network model provided by an embodiment of the present invention;

[0029] Figure 2 It is a MPSO-RBF neural network training principle diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] 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.

[0031] The following is a detailed description of the health status prediction of key ship components of the present invention. The execution subject of this method is the ship's health management system, such as Figure 1 As shown, the following steps are included:

[0032] 1. Data collection and preprocessing

[0033] (1) Select sensor type: Based on the common faults and key performance indicators of key ship components, select corresponding sensors and place them around the firmware of key components associated with the health status;

[0034] (2) Various types of sensors (such as temperature, pressure, vibration, etc.) installed on key components of the ship are used to collect monitoring parameters of key components of the ship, and the collected data are cleaned, normalized, calibrated, filtered and converted to remove noise interference.

[0035] 2. Feature extraction

[0036] (1) At regular intervals, a set of key component monitoring signals are collected to obtain a series of equally spaced key component status signals. The monitoring parameters closely related to the health status of the components are selected as status features, which are set as {x1, x2, …, x n};

[0037] (2) After extracting the eigenvalues ​​of effective monitoring parameters using eigenvalue analysis algorithms (such as mean, root mean square, standard deviation, etc.), a known state signal feature sequence is constructed, which is assumed to be {X1, X2, …, X n}.

[0038] 3. Model construction and training, such as Figure 2 shown.

[0039] (1) Establish an initial RBF neural network, set the number of output nodes of the network to N, the number of input nodes to M - N, where N < M; take the center value c of the RBF network i , width δ i and weight ω i as a particle for encoding. The total dimension of the example is D = h(I + O + 1), where I represents the RBF input dimension and O is the output dimension; set the fitness function of the particle swarm, and select the root mean square error (RMSE) of the RBF neural network as the fitness function. Among them, y i is the true value, y' i is the predicted value, and K is the number of samples; set the particle swarm size m and the maximum number of iterations t, and initialize the particle swarm.

[0040] (2) In the eigenvalue sequence {X1, X2,..., X n}, obtain a new set of n - M + 1 groups of continuous eigenvalue sequences {Y1, Y2,..., Y n-M+1} in a recursive manner, that is, select new sequences according to the following process: Continuously take M eigenvalues starting from X1 to form a new sample Y1, that is, Y1 = [X1, X2,..., X M , then continue to take M continuous eigenvalues starting from X2 to form the second new sample Y2, that is, Y2 = [X2, X3,..., X M+1 , then the third new sample Y3 = [X3, X4,..., X M+2 , and so on, then the last new sample is Y n-M+1 = [X n-M+1 , X n-M+2 ,..., X n ;

[0041] (3) Divide each sequence in the new eigenvalue sequence set {Y1, Y2,..., Y n-M+1} into two parts. The first part contains the first M - N eigenvalues of the sequence, and the second part contains the remaining N eigenvalues, then two corresponding new sets can be obtained. Let the two new sets obtained be {P1, P2,..., P n-M+1} and {T1, T2,..., T n-M+1}, then Y i = [P i T i , P i contains M - N eigenvalues, T i contains N eigenvalues, i = 1, 2,..., n - M + 1;

[0042] (4) For {P1, P2,..., P n-M+1} as the input value of the MPSO-RBF neural network, and the corresponding {T1,T2,…,T n-M+1} as the output value, train the initial MPSO-RBF neural network, calculate the fitness of each particle, update the individual extreme value x and the global extreme value of the particle, and update the position of the particle;

[0043] (5) Determine whether the number of iterations meets the maximum number of iterations. If not, repeat the previous step. If it does, map the optimal particle position vector to the various parameters of the neural network, and finally obtain a mature MPSO-RBF neural network prediction model.

[0044] Further, according to the health status prediction method of key ship components based on MPSO-RBF neural network provided by the present invention, the improvement of PSO algorithm is mainly reflected in the introduction of nonlinear inertia weight ω(t) to modify the original basic particle swarm algorithm formula. The original calculation formula is:

[0045]

[0046] Where t is the current iteration number, c1 and c2 are learning factors, and r1 and r2 are random numbers in (0, 1).

[0047] The revised calculation formula is:

[0048]

[0049]

[0050] In the formula, w max and w min is the maximum and minimum value of w, t max is the maximum number of iterations, t p Deployment for global extreme evolution stagnation, t h is the global extreme stagnation deployment threshold that needs to be disturbed, and U(0,1) is a random number uniformly distributed in the interval (0,1).

[0051] 4. Health status prediction

[0052] (1) The known state signal feature sequence is input into the well-trained MPSO-RBF neural network, and the output value is used as the subsequent state feature prediction value of the key component.

[0053] (2) Select the prediction length according to the requirements, obtain the state characteristic values ​​of the required future time, and evaluate the health status of key components based on the corresponding results and thresholds.

[0054] 5. Result feedback

[0055] The prediction results are fed back to the operation management personnel so that corresponding maintenance measures can be taken.

[0056] 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.

[0057] 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 predicting the health status of key ship components based on MPSO-RBF neural network, characterized in that: include: After preprocessing the multi-source heterogeneous monitoring parameters, the monitoring parameters related to the health status of the components are selected as state features to construct a known state signal feature sequence; The known state signal feature sequence is input into the mature MPSO-RBF neural network model, and the output value is used as the subsequent state feature prediction value of the key component. The MPSO-RBF neural network model introduces an improved PSO algorithm to adaptively select network parameters based on the RBF neural network model, and is trained using a large amount of labeled historical sample data of key ship equipment. According to the predicted value of the state feature and the correspondence between the signal feature value and the health state established with prior knowledge, the health state of the current device is determined and evaluated.

2. The method according to claim 1, characterized in that The known state signal feature sequence is input into the maturely trained MPSO-RBF neural network model, and the output value is used as the subsequent state feature prediction value of the key component, including: From the known state signal feature sequence, the last MN feature values ​​are selected to form the input value, which is input into the mature MPSO-RBF neural network to obtain a vector O containing N output values. This vector O is the state prediction value obtained. Continue to select the last M-2N eigenvalues ​​from the known state signal feature sequence and combine them with O to form a new input F. Input F into the mature MPSO-RBF neural network to obtain a new vector O' containing N output values. This vector O' is the state prediction value obtained in this prediction. The last M-3N eigenvalues ​​are selected and combined with O to form a new input F, and so on, until the preset prediction length is reached as the final prediction output eigenvalue result.

3. The method according to claim 2, characterized in that The training process of the MPSO-RBF neural network model includes: Obtain key component monitoring signals and obtain a series of equally spaced data samples {x1, x2, …, x n }, x n is the nth sample data, n is the number of samples; Select state features and find the sequence {x1,x2,…,x n } corresponding to the eigenvalues ​​{X1,X2,…,X n }; Establish an initial RBF neural network, set the number of output nodes of the network to N, the number of input nodes to M - N, where N < M; take the center value c i , width δ i and weight ω i as a particle for encoding. The total dimension of the particle is D = h(I + O + 1), where I represents the RBF input dimension and O is the output dimension; set the fitness function of the particle swarm, set the particle swarm size m and the maximum number of iterations t, and initialize the particle swarm; In the eigenvalue sequence {X1,X2,…,X n }, a new set of n-M+1 continuous eigenvalue sequences {Y1,Y2,…,Y n-M+1 }, where starting from X1, M eigenvalues ​​are taken continuously to form a new sample Y1 = [X1, X2, ..., X M ], and then continue to take M consecutive eigenvalues ​​from X2 to form the second new sample Y2 = [X2, X3, ..., X M+1 ], then the third new sample Y3=[X3,X4,…,X M+2 ], and so on, the last new sample is Y n-M+1 =[X n-M+1 ,X n-M+2 ,…,X n ]; The new eigenvalue sequence set {Y1,Y2,…,Y n-M+1 Each sequence in} is divided into two parts, the first part contains the first MN eigenvalues ​​of the sequence, and the second part contains the remaining N eigenvalues, then two new sets with one-to-one correspondence are obtained. Let the two new sets be {P1,P2,…,P n-M+1 } and {T1,T2,…,T n-M+1 }, then Y i =[P i T i ], P i Contains MN eigenvalues, T i Contains N eigenvalues, i = 1, 2, ..., n-M+1; {P1,P2,…,P n-M+1 } as the input value of the MPSO-RBF neural network, and the corresponding {T1,T2,…,T n-M+1 } as the output value, train the initial MPSO-RBF neural network, calculate the fitness of each particle, update the individual extreme value and global extreme value of the particle, and update the position of the particle; Determine whether the number of iterations meets the maximum number of iterations. If not, repeat the process to convert {P1, P2, …, P n-M+1 } as the input value of the MPSO-RBF neural network, and the corresponding {T1,T2,…,T n-M+1 } as the output value, train the initial MPSO-RBF neural network, calculate the fitness of each particle, update the individual extreme value and global extreme value of the particle, and update the position of the particle. If satisfied, map the optimal particle position vector to the various parameters of the neural network, and finally obtain a mature MPSO-RBF neural network prediction model.

4. The method according to claim 3, characterized in that In the MPSO-RBF neural network, the nonlinear inertia weight w(t) is introduced to modify the original basic particle swarm algorithm formula as follows: Among them, t is the current iteration number, c1 and c2 are learning factors, r1 and r2 are random numbers, and w max and w min are the maximum and minimum values ​​of w, t max is the maximum number of iterations, t p Deployment for global extreme evolution stagnation, t g is the global extreme stagnation deployment threshold that needs to be disturbed, U(0,1) is a random number uniformly distributed in the interval (0,1), represents the flying speed of the i-th particle in the d-th dimension in the t-th iteration, x pb represents the best historical position of the individual particle, x gb Represents the global optimal particle position.

Images