A multi-factor ship health dynamic prediction method

Abstract

The present invention belongs to the technical field of ship management and control methods, and in particular relates to a multi-factor ship health dynamic prediction method. Specifically, the method comprises the following steps: establishing various influencing factors associated with the ship's health status, intercepting a set of historical parameter indicators, calculating characteristic variables, constructing a change feature label for a multi-dimensional feature vector, establishing a CNN-LSTM-AM prediction model, training and validating the model, and inputting the original data to be analyzed into the trained model to predict the results; the present application analyzes historical element data in combination with time series information, uses a sliding time window to extract the phased change characteristics of the ship's health status as it changes with time series, and uses this as an entry point to analyze and predict the phased data. This method does not rely on the continuous analysis process of a single data, and provides a new approach for further optimizing and improving the analysis and prediction of the ship's health status.

Description

Technical Field

[0001] The present invention belongs to the technical field of ship management and control methods, and in particular relates to a multi-factor ship health dynamic prediction method. Background Art

[0002] Ship health management is an important part of ensuring the long-term stable operation of ships. The ship's health status is a very broad concept, involving the current health status and expected change trends of various systems and equipment on the ship. The specific health status and change trends of each system and equipment involve a large number of influencing factors. Therefore, the current assessment and analysis of the ship's health status is mainly applied to the independent analysis of local parts of the ship, such as specific equipment or systems, and lacks sufficient explanatory power for the overall health of the ship. The above-mentioned equipment or systems are often interdependent and have a very complex impact on the overall health status of the ship. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for better and more comprehensive ship health status prediction by better realizing full-factor analysis and processing and integrating multi-factor correlation influence characteristics based on time series information.

[0004] To achieve the above objectives, the present invention adopts the following technical solutions.

[0005] A multi-factor ship health dynamic prediction method specifically includes the following steps:

[0006] (1) Confirm and establish the various factors affecting the ship's health status, analyze the ship's own attributes and determine the parameter index set G = {g1, g2... g a ...g A}, where g a Represents the ath indicator element after normalization of the parameter indicator set;

[0007] Obtain the historical operating data of the ship to be analyzed and determine its historical parameter index set G time ;

[0008] G time ={G1,G2...G t ...G T};

[0009] G t ={g 1,t ,g 2.t ...g a,t ...g A,t};

[0010] Among them, G t represents the parameter index set obtained in the tth collection cycle, ga,t Represents the ath indicator element in the parameter indicator set obtained in the tth acquisition cycle;

[0011] For the historical parameter indicator set G time ,After pre-processing the data, invalid data is cleared and erroneous data is corrected;

[0012] (2) Using the time window L win Intercept the historical parameter indicator set G in order along the time series time , we get several acquisition cycles as historical parameter indicator sequences {G nL} for analysis, where:

[0013]

[0014] Among them, L win =k 2 t, k are positive integers, G nL Indicates the period L obtained by the nth interception win The parameter index set within, n = 1 ~ L;

[0015] (3) Calculate the historical parameter indicator sequence {G nL}, and obtain the characteristic variable sequence of the adjacent index sequence by taking the difference of the characteristic values ​​of the adjacent index sequence

[0016] (4) Calculate the time window L win Internal characteristic variable G Δn The value of the statistical characteristic variable G Δn The change of the multidimensional feature vector is used to construct the change feature label:

[0017]

[0018] Where r1 refers to the calibration value in the characteristic variable sequence The number of elements of Refers to the characteristic variable calibration value that maintains the ship's health status unchanged; where r2 refers to the calibration value in the characteristic variable sequence The number of elements; where r3 refers to the calibration value in the characteristic variable sequence The number of elements; where q1 refers to the calibration value in the feature variable sequence The accumulated value of the elements; where q2 refers to the calibration value in the characteristic variable sequence The accumulated value of the elements; where q3 refers to the calibration value in the characteristic variable sequence The accumulated value of the elements of ; ω1 and ω2 are preset balancing weights and ω1+ω2=1;

[0019] Determine the historical parameter indicator sequence {G nLCorresponding to the characteristic variables, determine the parameter indicator sequence whose characteristic variable sequence meets the corresponding change characteristic pre-threshold value, and obtain three types of health status data sets: health status deterioration, health status improvement, and health status stability;

[0020] (5) Establish a CNN-LSTM-AM prediction model, and train the CNN prediction model based on the change feature label of the multidimensional feature vector determined in step (4) as the output; establish a sub-model based on the AM-LSTM network model, and set a fully connected layer separately between the hidden layer and the output layer of the sub-model, and use the three types of health status data sets as the training and verification data of the CNN model and the AM-LSTM model to train and verify the model;

[0021] Input the original data to be analyzed into the trained model to obtain the change feature threshold output and the sub-model prediction output, and multiply the change feature threshold output with the sub-model prediction result to determine the prediction result;

[0022] In a further improvement or preferred implementation of the aforementioned multi-factor ship health dynamic prediction method, the selection of various basic parameter indicators depends on the analysis of the current ship's own properties and characteristics, including selection based on different types of ships targeted by the health prediction, or based on different functional systems targeted by the prediction, including but not limited to various indicator parameters that directly or indirectly affect the ship or the health status of the ship's various functional components.

[0023] A further improvement or preferred embodiment of the aforementioned multi-factor ship health dynamic prediction method is to construct a CNN-LSTM-AM prediction model, which includes determining the convolution unit parameters: determining the weights and parameters of the convolution unit in the convolutional neural network based on the forward propagation calculation; wherein the unit state transfer coefficient f t =σ(w f ·[h t-1 ,x t ]+b f ), w f Refers to the unit state transfer weight matrix, b f is the unit state transfer bias term, σ is the sigmoid function, h t-1 Refers to the hidden layer output at the previous moment, and is the input x at the current moment t ;Unit input transfer coefficient i t =σ(w i ·[h t-1 ,x t ]+b i ), w f Refers to the unit input transfer weight matrix, b i is the unit input transfer bias term; the current state coefficient c t =ft ⊙c t-1 +i t ⊙tanh(w c [h t-1 ,x t ]+b c ), w c Refers to the current state weight matrix, b c is the current state bias term; ⊙ refers to element-wise multiplication; the current unit output coefficient o t =σ(w o ·[h t-1 ,x t ]+b o ); where w o Refers to the unit output weight matrix at the current moment, b o is the current moment unit output transfer bias term; hidden layer output h t =o t tanh(c t ); Based on the back-propagation algorithm, the aforementioned weight matrix and bias terms are optimized and trained, and the parameters of each unit are determined by forward propagation calculation through initialization parameters.

[0024] A further improvement or preferred embodiment of the aforementioned multi-factor ship health dynamic prediction method further includes a step of further optimizing the prediction results based on a genetic algorithm, specifically:

[0025] For the comprehensive prediction value goal determined in step (5) i , the sub-model weight μ i value as the optimization object, with the real value goal real and comprehensive predicted value goal i The minimum deviation is used as the optimization goal for search optimization; where: goal i represents the comprehensive prediction result of the i-th time, μ i represents the weight of the i-th sub-model, represents the predicted output of the jth sub-model during the i-th prediction;

[0026] The optimization process can be expressed as:

[0027]

[0028] where μ min is the lower limit of weight, μ min is the lower limit of the weight, Aeq*μ i =beq ​​is the weight linear constraint, A*μ i ≤b is the linear inequality constraint of weights.

[0029] A further improvement or preferred implementation of the aforementioned multi-factor ship health dynamic prediction method, the optimization step includes: based on a genetic algorithm, generating an initialized population consisting of the weights of the sub-models, performing genetic changes to generate offspring, performing fast non-dominated sorting on the first-generation sub-population, and then generating a second-generation sub-population through selection, crossover, and mutation, combining the two populations to form a new population, performing fast non-dominated sorting on the new population, and calculating the crowding degree of individuals in each non-dominated layer at the same time, selecting suitable individuals to form a new parent population based on the non-dominated relationship and the crowding degree of the individuals, generating a new offspring population through genetic algorithm operations, and repeating the above operations until the new population meets the requirements or the iteration ends.

[0030] Its beneficial effects are:

[0031] The multi-factor ship health dynamic prediction method of this application fully considers the current needs of ship health status analysis and prediction. By combining the historical element data of time series information for analysis, a sliding time window is used to extract the phased change characteristics of the ship health status with time series changes, and this is used as the entry point for analysis and prediction of the phased data. It does not rely on the continuous analysis process of a single data, and can better analyze and extract the phased impact of multiple factors on the ship health status, and then achieve a more comprehensive health status analysis based on this. This method provides a new way to further optimize and improve the analysis and prediction of ship health status. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart of multi-factor ship health dynamic prediction;

[0033] Figure 2 It is a schematic diagram of the multi-factor ship health dynamic prediction optimization process. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to specific embodiments.

[0035] The present application proposes a multi-factor ship health dynamic prediction method, which is mainly used to improve the limitations of existing ship health status analysis and prediction methods. It proposes to analyze and predict the ship health status based on the phased changes of factor factors under time series changes. It is not limited to the single point changes of a certain data in a certain continuous period. It can better start from the overall situation and establish a new way of ship health status prediction and analysis with the phased multi-factor change characteristics.

[0036] The basic process of the multi-factor ship health dynamic prediction method is as follows: Figure 1 As shown;

[0037] First, in order to more comprehensively analyze and obtain the health status of the ship, this application considers a global analysis based on multiple influencing factors of the ship's health status, and provides a solution for effective health prediction based on the large amount of multi-factor data. First, the various influencing factors associated with the ship's health status should be identified and established, and the ship's own attributes should be analyzed to determine the parameter index set G = {g1, g2... g a ...g A}, where g a Represents the ath indicator element after normalization of the parameter indicator set;

[0038] In the process of establishing the aforementioned parameter indicator set, the selection of various basic parameter indicators depends on the analysis of the current ship's own attributes and characteristics, including selection based on the different types of ships targeted by the health prediction, or based on the different functional systems targeted by the prediction, including but not limited to various indicator parameters that directly or indirectly affect the health status of the ship or its functional components;

[0039] Obtain the historical operating data of the ship to be analyzed and determine its historical parameter index set G time ;

[0040] G time ={G1,G2...G t ...G T};

[0041] G t ={g 1,t ,g 2.t ...g a,t ...g A,t};

[0042] Among them, G t represents the parameter index set obtained in the tth collection cycle, g a,t Represents the ath indicator element in the parameter indicator set obtained in the tth acquisition cycle;

[0043] For the historical parameter indicator set G time ,After pre-processing the data, invalid data is cleared and erroneous data is corrected;

[0044] In addition to directly affecting the health status of the ship, the aforementioned basic parameter indicators may also have known or unknown correlations with each other. In order to retain the correlation effects between these parameter indicator sets and reduce the number of basic parameter indicators that need to be directly processed, the time window L is used. win Intercept the historical parameter indicator set G in order along the time series time , we get several acquisition cycles as historical parameter indicator sequences {G nL} for analysis, where:

[0045]

[0046] Among them, L win =k 2 t, k are positive integers, G nL Indicates the period L obtained by the nth interception win The parameter index set within, n = 1 ~ L;

[0047] (1) Calculate the historical parameter indicator sequence {G nL}, and obtain the characteristic variable sequence of the adjacent index sequence by taking the difference of the characteristic values ​​of the adjacent index sequence

[0048] (2) Calculate the time window L win Internal characteristic variable G Δn The value of the statistical characteristic variable G Δn Construct a multidimensional feature vector of change feature labels based on the changes in

[0049]

[0050] Where r1 refers to the calibration value in the characteristic variable sequence The number of elements of Refers to the characteristic variable calibration value that maintains the ship's health status unchanged; where r2 refers to the calibration value in the characteristic variable sequence The number of elements; where r3 refers to the calibration value in the characteristic variable sequence The number of elements; where q1 refers to the calibration value in the feature variable sequence The accumulated value of the elements; where q2 refers to the calibration value in the characteristic variable sequence The accumulated value of the elements; where q3 refers to the calibration value in the characteristic variable sequence The accumulated value of the elements of ; ω1 and ω2 are preset balancing weights and ω1+ω2=1;

[0051] Determine the historical parameter indicator sequence {G nL Corresponding to the characteristic variables, determine the parameter indicator sequence whose characteristic variable sequence meets the corresponding change characteristic pre-threshold value, and obtain three types of health status data sets: health status deterioration, health status improvement, and health status stability;

[0052] Establish a CNN-LSTM-AM prediction model, in which the construction of the convolution layer includes: determining the convolution unit parameters: determining the weights and parameters of the convolution unit in the convolutional neural network based on forward propagation calculation; the unit state transfer coefficient f t =σ(w f ·[h t-1,x t ]+b f ), w f Refers to the unit state transfer weight matrix, b f is the unit state transfer bias term, σ is the sigmoid function, h t-1 Refers to the hidden layer output at the previous moment, and is the input x at the current moment t ;Unit input transfer coefficient i t =σ(w i ·[h t-1 ,x t ]+b i ), w f Refers to the unit input transfer weight matrix, b i is the unit input transfer bias term; the current state coefficient c t =f t ⊙c t-1 +i t ⊙tanh(w c [h t-1 ,x t ]+b c ), w c Refers to the current state weight matrix, b c is the current state bias term; ⊙ refers to element-wise multiplication; the current unit output coefficient o t =σ(w o ·[h t-1 ,x t ]+b o ); where w o Refers to the unit output weight matrix at the current moment, b o is the current moment unit output transfer bias term; hidden layer output h t =o t tanh(c t ); Based on the back-propagation algorithm, the aforementioned weight matrix and bias terms are optimized and trained, and the parameters of each unit are determined by forward propagation calculation through initialization parameters.

[0053] Based on the above, the CNN prediction model is trained using the change feature labels of the multidimensional feature vector determined in step 4 as output. A sub-model is established based on the AM-LSTM network model, and a fully connected layer is separately set between the hidden layer and the output layer of the sub-model. The three health status datasets are used as training and validation data for the CNN model and the AM-LSTM sub-model to optimize the model.

[0054] The original data to be analyzed is input into the trained model. The CNN prediction model obtains the output of the change feature label value as the weight, and the AM-LSTM sub-model prediction output is used as the variable to multiply them respectively. The final result is the prediction result.

[0055] like Figure 2 As shown, the present application also includes the step of further optimizing the prediction results based on the genetic algorithm:

[0056] Specifically, for the comprehensive prediction value goal determined in step (5) i , the sub-model weight μ i value as the optimization object, with the real value goal real and comprehensive predicted value goal i The minimum deviation is used as the optimization goal for search optimization; where:

[0057]

[0058] Among them, goal i represents the comprehensive prediction result of the i-th time, μ i represents the weight of the i-th sub-model, represents the predicted output of the jth sub-model during the i-th prediction;

[0059] The optimization process can be expressed as:

[0060]

[0061] where μ min is the lower limit of weight, μ min is the lower limit of the weight, Aeq*μ i =beq ​​is the weight linear constraint, A*μ i ≤b is the linear inequality constraint of weights;

[0062] The specific implementation process includes:

[0063] Based on the genetic algorithm, an initialization population consisting of the weights of the sub-models is generated, and genetic changes are performed to generate offspring. After a fast non-dominated sort is performed on the first-generation sub-population, a second-generation sub-population is generated through selection, crossover, and mutation. The two populations are combined to form a new population, and a fast non-dominated sort is performed on the new population. At the same time, the crowding degree of individuals in each non-dominated layer is calculated. According to the non-dominated relationship and the crowding degree of the individuals, suitable individuals are selected to form a new parent population. A new offspring population is generated through genetic algorithm operations. The above operations are repeated until the new population meets the requirements or the iteration ends.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A multi-factor ship health dynamic prediction method, characterized in that: The steps include: (1) Identify and establish the various factors affecting the ship's health status, analyze the ship's own attributes and determine the parameter index set associated with the ship's health status ,in Indicates the parameter index set after normalization indicator elements; Obtain the historical operating data of the ship to be analyzed and determine its historical parameter index set ; ； ； in, Indicates the The parameter indicator set obtained during the collection period, Indicates the The first parameter index set obtained in the collection cycle indicator elements; For historical parameter indicator sets ,After pre-processing the data, invalid data is cleared and erroneous data is corrected; (2) Using time windows Intercept the historical parameter indicator set in sequence along the time series , and obtain several collection cycles as historical parameter indicator sequences Perform analysis, where: ； ； in, , is a positive integer, Indicates the The cycle obtained by intercepting The parameter index set within ; (3) Calculate the historical parameter indicator sequence The characteristic variable sequence of the adjacent index sequence is obtained by taking the difference of the characteristic values ​​of the adjacent index sequence , ; (4) Calculation time window Internal characteristic variables The value of the statistical characteristic variable The change of the multidimensional feature vector is used to construct the change feature label: ； in It means that the calibration value in the characteristic variable sequence The number of elements of Refers to the characteristic variable calibration value that maintains the ship's health status unchanged; It means that the calibration value in the characteristic variable sequence The number of elements of It means that the calibration value in the characteristic variable sequence The number of elements of Refers to the calibration value in the characteristic variable sequence The accumulated value of the elements of It means that the calibration value in the characteristic variable sequence The accumulated value of the elements of It means that the calibration value in the characteristic variable sequence The accumulated value of the elements; and is the preset balance weight and ; Determine the historical parameter indicator sequence Corresponding to the characteristic variables, determine the parameter indicator sequence whose characteristic variable sequence meets the corresponding change characteristic pre-threshold, and obtain three types of health status data sets: health status deterioration, health status improvement, and health status stability; (5) Establish a CNN-LSTM-AM prediction model, and train the CNN prediction model based on the change feature label of the multi-dimensional feature vector determined in step (4) as the output; establish a sub-model based on the AM-LSTM network model, and set a fully connected layer separately between the hidden layer and the output layer of the sub-model, and use the three types of health status data sets as the training and verification data of the CNN model and the AM-LSTM model to train and verify the model; The original data to be analyzed is input into the trained model to obtain the change feature threshold output and the sub-model prediction output. The change feature threshold output is used as the weight and multiplied with the sub-model prediction result to determine the prediction result.

2. The multi-factor ship health dynamic prediction method according to claim 1 is characterized in that: The selection of various basic parameter indicators depends on the analysis of the current ship's own properties and characteristics, including the different types of ships targeted by the health prediction, or the different functional systems targeted by the prediction, including various indicator parameters that directly or indirectly affect the health status of the ship or the functional components of the ship.

3. The multi-factor ship health dynamic prediction method according to claim 1 is characterized in that: Building the CNN-LSTM-AM prediction model includes determining the convolution unit parameters: determining the weights and parameters of the convolution unit in the convolutional neural network based on forward propagation calculations; The unit state transfer coefficient , refers to the unit state transfer weight matrix, is the unit state transfer bias term, yes function, Refers to the hidden layer output at the previous moment, Input for the current moment; unit input transfer coefficient , is the unit input transfer weight matrix, is the unit input transfer bias term; Current state coefficient , Refers to the current state weight matrix, is the current state bias; Refers to element-wise multiplication; the unit output coefficient at the current moment is ;in Refers to the unit output weight matrix at the current moment, Is the current moment unit output transfer bias term; hidden layer output ; The aforementioned weight matrix and bias terms are optimized and trained based on the back-propagation algorithm, and the parameters of each unit are determined by forward propagation calculation through initialization parameters.

4. The multi-factor ship health dynamic prediction method according to claim 1 is characterized in that: It also includes the step of further optimizing the prediction results based on the genetic algorithm, specifically: For the comprehensive prediction value determined in step (5), , the sub-model weight Value is used as the optimization object, with the real value and the comprehensive forecast value The minimum deviation is used as the optimization goal for search optimization; where: ; Indicates the The comprehensive prediction results of Indicates the The weight of each sub-model, Indicates the The second prediction The prediction output of each sub-model; The optimization process can be expressed as: ； in is the lower limit of the weight, is the weight linear constraint, is the linear inequality constraint of weights.

5. The multi-factor ship health dynamic prediction method according to claim 4 is characterized in that: The optimization step includes: generating an initial population composed of weights of the sub-models based on a genetic algorithm, performing genetic changes to generate offspring, performing fast non-dominated sorting on the first-generation sub-population, and then generating a second-generation sub-population through selection, crossover, and mutation, combining the two populations to form a new population, performing fast non-dominated sorting on the new population, and calculating the crowding degree of individuals in each non-dominated layer. Based on the non-dominated relationship and the crowding degree of the individuals, suitable individuals are selected to form a new parent population, and a new offspring population is generated through genetic algorithm operations. The above steps are repeated until the new population meets the requirements or the iteration ends.

Images