Marine diesel engine vibration fault early warning method based on improved VMD

Abstract

The invention discloses a marine diesel engine vibration fault early warning method based on an improved VMD, and the method comprises the steps: improving the VMD through combining a frost ice algorithm RIME, carrying out the hyper-parameter optimization, carrying out the VMD signal decomposition of an optimized data set based on an optimal combination, and obtaining a plurality of IMF values, so as to reconstruct vibration signal sequence data. Performing time domain and frequency domain feature extraction on the reconstructed vibration signal sequence data to obtain a time domain feature parameter and a frequency domain feature parameter; calculating and acquiring a Pearson coefficient between each feature parameter in the mixed domain feature set and the early warning value, and selecting the feature parameter of which the Pearson coefficient meets a preset Pearson coefficient threshold as an optimized feature parameter; and obtaining a residual error between a predicted root-mean-square value and an actual root-mean-square value based on the optimal BiGRU algorithm model so as to obtain a residual error threshold range under the condition of fault-free operation of the marine diesel engine, building an RIME-VMD-BIGRU model for vibration fault early warning of the marine diesel engine, and realizing automatic prediction of the vibration fault of the marine diesel engine. According to the method, sailors can be reminded of possible faults of the ship, the dependence on expert experience is reduced, and meanwhile, compared with other early warning algorithms, the method has higher diagnosis precision.

Description

Technical Field

[0001] The invention relates to the technical field of ship diesel engine fault diagnosis, and in particular to a ship diesel engine vibration fault early warning method based on improved VMD. Background Art

[0002] As a reciprocating internal combustion engine, marine diesel engines are widely used in marine power plants and are an important source of marine power output. However, due to the complex working environment and long-term use, marine diesel engines often fail. In addition, due to their compact and complex structure, some moving parts are usually accompanied by large mechanical and thermal loads, and inherent characteristics such as variable working conditions, the cause-effect relationship of the fault symptoms is easily complicated. At present, there are mainly the following methods for predicting faults in marine diesel engines:

[0003] 1. Fault diagnosis based on physical models: Fault prediction and early warning methods based on physical models often require researchers to build accurate mathematical or physical models to describe the working process of the research object. By analyzing the difference between the actual behavior of the system and the expected behavior of the model, it is possible to detect whether the system is in an abnormal state, locate the fault of the system, and issue early warnings to relevant managers. Although the fault prediction and early warning methods based on physical models can accurately predict and warn equipment failures, the prerequisite is to establish a very accurate mathematical or physical model for the monitored research object. Due to the complex structure of the marine diesel engine system, the extremely harsh working environment, and the changeable working conditions, it is difficult to establish a mathematical model of the marine diesel engine. Therefore, the fault prediction and early warning methods based on physical models are obviously not suitable for the fault warning research of marine diesel engines.

[0004] 2. Fault prediction and early warning methods based on empirical models: Fault diagnosis methods based on empirical models need to be based on the knowledge and experience of experts when diagnosing equipment fault conditions, mainly including fault tree-based and expert system-based diagnostic methods. Methods based on empirical models do rely more on the expertise and practical engineering experience of experts, and are also affected to a certain extent by personal experience and preferences. Therefore, if there is a large deviation between the experience of experts and the actual situation, the accuracy of judgment of abnormal conditions and fault prediction and early warning will drop significantly, which may lead to consequences that are difficult to estimate. Therefore, methods based on empirical models need to use the experience of most experts to build an expert knowledge base for the entire field, which can reduce the subjective influence of individuals on abnormal conditions through data fusion.

[0005] 3. Reliability model-based methods: Reliability models usually use mathematical statistics knowledge to analyze past failure history data to predict possible future equipment failures. These models can help relevant personnel develop effective maintenance plans, thereby reducing the probability of equipment failures and reducing losses. The reliability-based failure prediction and early warning method is essentially to use probability statistics knowledge to use a series of different probability density functions to express the time of occurrence of various types of failures or the probability of failures within a certain period of time. Its basis is also to analyze the data information collected under different equipment states, so some experts and scholars also classify this type of method as a data-driven method.

[0006] 4. Data-driven methods: Data-driven fault prediction and early warning methods have been proposed and widely used in the industrial field. This method does not require the establishment of an accurate physical model for the research object. Instead, it uses the collected data to perform various data analysis and artificial intelligence algorithm processing to mine the implicit information in the data and predict the future state of the research object. Compared with model-based and expert knowledge-based fault prediction and early warning technologies, data-driven methods avoid their shortcomings. Therefore, data-driven fault prediction and early warning methods have received increasing attention.

[0007] In summary, traditional diesel engine fault warning methods usually rely on expert experience, and maintenance personnel need to passively discover faults based on maintenance experience or existing fault parameter records. However, in the era of unmanned cabins and smart shipping, the timeliness and accuracy of this prediction method are often insufficient. Summary of the invention

[0008] The present invention provides a marine diesel engine vibration fault early warning method based on improved VMD to overcome the above technical problems.

[0009] In order to achieve the above object, the technical solution of the present invention is:

[0010] A marine diesel engine vibration fault early warning method based on improved VMD comprises the following steps:

[0011] S1: Obtain the fault-free historical vibration signal sequence data of the ship diesel engine, and perform normalization preprocessing to obtain the optimized data set;

[0012] S2: Based on the frost optimization algorithm RIME, the optimal combination of the number of decomposition layers and penalty factors of the VMD algorithm is optimized according to the optimization data set;

[0013] S3: Based on the optimal combination, VMD signal decomposition is performed on the optimized data set to obtain several IMF values ​​to reconstruct the vibration signal sequence data;

[0014] Based on the frequency domain and time domain analysis method, the time domain and frequency domain features of the reconstructed vibration signal sequence data are extracted to obtain the time domain feature parameters and frequency domain feature parameters;

[0015] And obtaining a mixed domain feature set according to the time domain feature parameters and the frequency domain feature parameters;

[0016] S4: Obtain the root mean square value of the original vibration signal sequence data as a warning value;

[0017] Calculate and obtain the Pearson coefficient between each feature parameter and the warning value in the hybrid domain feature set, and select the feature parameter whose Pearson coefficient meets the preset Pearson coefficient threshold as the optimized feature parameter;

[0018] S5: randomly dividing the optimized feature parameters into a training set and a test set according to a preset ratio;

[0019] S6: pre-constructing a BiGRU algorithm model, taking the optimized feature parameters in the training set as feature data, taking the warning value corresponding to the feature data as label data as input of the BiGRU algorithm model, and training the BiGRU algorithm model based on the loss function to obtain the optimal BiGRU algorithm model; and then obtaining a RIME-VMD-BIGRU model for vibration fault warning of a ship diesel engine;

[0020] Based on the RIME-VMD-BIGRU model, the residual between the predicted RMS value and the actual RMS value is calculated and obtained according to the test set;

[0021] S7: according to the residual, a residual threshold range [A, B] is set under the condition of trouble-free operation of the ship diesel engine; and A is the minimum residual value obtained according to the test set, and B is the maximum residual value obtained according to the test set;

[0022] S8: Input the acquired current vibration signal sequence data of the ship diesel engine into RIME-VMD-

[0023] BIGRU model, obtains the residual of the current ship diesel engine, and determines whether the residual of the current ship diesel engine meets the residual threshold range [A, B]; if the residual of the current ship diesel engine meets the residual threshold range [A, B], it is confirmed that the vibration operation of the current ship diesel engine is normal; otherwise, it is confirmed that the vibration operation of the current ship diesel engine is faulty, and outputs the vibration fault warning of the current ship diesel engine.

[0024] Furthermore, the S2 specifically includes the following steps:

[0025] S21: Set the maximum number of iterations and randomly initialize the initial frost particle population of the frost optimization algorithm RIME;

[0026] The parameter combination [k, α] of the VMD algorithm's decomposition level and penalty factor is defined as frost ice particles;

[0027] S22: performing iterative signal decomposition on the fault-free historical vibration signal sequence data according to the VMD algorithm corresponding to each frost particle of the initial frost particle population, and obtaining several intrinsic mode function IMF values;

[0028] The sample entropy or envelope entropy is used as the fitness function to determine the fitness value of each intrinsic mode function IMF value, and the minimum fitness value is used as the current optimal fitness of the optimal frost ice particle;

[0029] S23: Obtaining the current adhesion coefficient E of the frost ice particles according to the current number of iterations and the maximum number of iterations;

[0030] S24: define a first random number r2 and a second random number r3;

[0031] And determine the size of the first random number r2 and the current adhesion coefficient E;

[0032] If the first random number r2 is greater than or equal to the current adhesion coefficient E, the current number of iterations is updated and step S23 is repeated;

[0033] Otherwise, the position of the frost ice particle is updated to obtain an updated frost ice particle population, and each frost ice particle in the updated frost ice particle population and its corresponding optimal fitness are obtained;

[0034] According to each frost ice particle in the updated frost ice particle population and its corresponding optimal fitness, confirm the current normalized fitness value of the updated frost ice particle population;

[0035] S25: Determine the magnitude of the second random number r3 and the current normalized fitness value;

[0036] If the second random number r3 is greater than or equal to the current normalized fitness value, then the current number of iterations is updated and steps S23 to S24 are repeated;

[0037] Otherwise, the position of the updated frost ice particle is updated to obtain the latest frost ice particle population, and the latest optimal frost ice particle in the latest frost ice particle population and its corresponding latest optimal fitness are obtained;

[0038] S26: Determine the size of the latest optimal fitness and the current optimal fitness;

[0039] If the latest optimal fitness is greater than or equal to the current optimal fitness, the current number of iterations is updated and steps S23 to S25 are repeated;

[0040] Otherwise, continue to execute step S27;

[0041] S27: Replace the latest optimal fitness with the current optimal fitness, and determine whether the current number of iterations reaches the maximum number of iterations;

[0042] If not, update the current number of iterations and repeat steps S23 to S26;

[0043] If so, the frost ice particle corresponding to the current optimal fitness is used as the final frost ice particle, that is, the parameter combination of the VMD algorithm corresponding to the final frost ice particle is used as the optimal combination.

[0044] Furthermore, the time domain characteristic parameters and frequency domain characteristic parameters obtained in S3 are specifically:

[0045] The time domain characteristic parameters include mean value, root square amplitude, root mean square, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor and pulse index;

[0046] The frequency domain characteristic parameters include average spectrum amplitude, variance of spectrum amplitude, skewness of spectrum amplitude, kurtosis of spectrum amplitude, centroid frequency, frequency mean square error, root mean square frequency, fourth root frequency, ratio of root mean square frequency to square root of spectrum amplitude, ratio of frequency mean square error to centroid frequency, frequency skewness and frequency kurtosis.

[0047] Furthermore, the loss function for training the BiGRU algorithm model in S6 is

[0048]

[0049] Where: n represents the number of feature data for optimizing feature parameters; y i Represents the true value of the i-th feature data; Represents the predicted value of the i-th feature data;

[0050] The residual error between the predicted RMS value and the actual RMS value is expressed as

[0051]

[0052] Where: y i Represents the true value of the i-th feature data; Represents the predicted value of the i-th feature data.

[0053] Further, the BiGRU algorithm model pre-constructed in S6 includes an input layer, a first GRU unit, a second GRU unit, and a fusion output layer;

[0054] The input layer is used to transmit input data to the first GRU unit and the second GRU unit respectively;

[0055] The first GRU unit is used to perform a time forward processing operation on the output of the input layer to obtain forward processing data; the second GRU unit is used to perform a time reverse processing operation on the output of the input layer to obtain reverse processing data;

[0056] The fusion output layer is used to merge and output the forward-order processed data and the reverse-order processed data.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. Strong applicability: By normalizing the fault-free historical vibration signal sequence data of the ship diesel engine, the influence of different dimensions and orders of magnitude of each diagnostic parameter on the diagnostic result is eliminated;

[0059] 2. Efficient and accurate: The RIME algorithm is used to optimize and obtain the optimal combination of the decomposition layer number and penalty factor of the VMD algorithm, avoiding the problem of inaccurate manual selection, fully mining the information in the data, reducing the need for artificial features, and improving the efficiency and accuracy of diagnosis;

[0060] 3. Strong adaptability: By introducing the Pearson coefficient, it is possible to effectively select strongly correlated features, reduce feature dimensions, reduce the waste of computing resources, and improve the accuracy and stability of predictions;

[0061] 4. Good real-time performance: By adopting online data processing and real-time diagnosis, BIGRU can be used to accurately predict the root mean square characteristics and set the threshold line. If the threshold line range is exceeded, the on-duty personnel can be reminded to check the fault in time, reducing maintenance costs.

[0062] In general, the ship diesel engine vibration fault early warning method based on improved VMD can automatically predict ship diesel engine vibration faults and quickly remind the crew of impending problems, reducing dependence on expert experience. At the same time, compared with the single BiGRU, it has higher accuracy early warning and can be widely promoted in the field of ship engine engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0064] Figure 1 It is a flow chart of the marine diesel engine vibration fault early warning method based on improved VMD of the present invention;

[0065] Figure 2Flow chart of the RIME-VMD algorithm in this embodiment;

[0066] Figure 3 Schematic diagram of the BiGRU algorithm model in this embodiment;

[0067] Figure 4 This is a decomposition diagram of the simulated vibration signal in this embodiment;

[0068] Figure 5 This is a comparison diagram of the original simulated vibration signal and the reconstructed signal in this embodiment;

[0069] Figure 6 This is the original sequence diagram of the ship vibration signal in this embodiment;

[0070] Figure 7 It is an iterative convergence diagram of the envelope entropy ship vibration signal in this embodiment;

[0071] Figure 8 It is the iterative convergence diagram of the sample entropy ship vibration signal in this embodiment;

[0072] Fig. 9 This is the envelope entropy ship vibration signal IMF spectrum decomposition diagram in this embodiment;

[0073] Fig.10 This is the IMF spectrum decomposition diagram of the sample entropy ship vibration signal in this embodiment;

[0074] Fig.11 This is a comparison diagram of the original vibration signal and the reconstructed signal of the ship in this embodiment;

[0075] Fig.12 This is a comparison diagram of the predictions using VMD decomposition and without VMD decomposition in this embodiment;

[0076] Fig.13 Graph showing the prediction results of the RIME-VMD-BiGRU model in this embodiment;

[0077] Fig.14 Graph showing the prediction results of the RIME-VMD-GRU model in this embodiment;

[0078] Fig.15 This is the prediction result diagram of the RIME-VMD-LSTM model in this embodiment;

[0079] Fig.16 Graph showing the prediction results of the RIME-VMD-RNN model in this embodiment;

[0080] Fig.17 is the residual threshold map in this embodiment;

[0081] Fig.18 is the residual threshold fault diagram in this embodiment;

[0082] Fig.19 It is the core block diagram of the ship diesel engine vibration fault early warning method based on improved VMD in this embodiment. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] This embodiment provides a marine diesel engine vibration fault early warning method based on improved VMD, such as Figure 1 and Fig.19 As shown, the following steps are included:

[0085] S1: Obtain the fault-free historical vibration signal sequence data of the ship diesel engine, and perform normalization preprocessing to obtain the optimized data set;

[0086] Specifically, the sample data of this embodiment, i.e., the fault-free historical vibration signal sequence data of the ship diesel engine, comes from the new teaching and training dual-purpose ship "Xin Hong Zhuan" of Dalian Maritime University. Three groups of original vibration data were collected with a sampling frequency of 4000Hz. The data were tested under five load conditions of 25%, 50%, 75%, 85% and 100%, respectively. The tests were conducted at the following eight measuring points, as shown in Table 1: 1# cylinder head, 8# cylinder head, supercharger, engine body side (between 7\8 cylinders), engine body side (between 1\2 cylinders), base (1 cylinder side), base (8 cylinder side), generator bearing cover (non-drive end) for collection. This embodiment mainly analyzes the data under 85% load conditions. In the first group of data, each measuring point recorded 20,000 data points, with a total duration of 5 seconds; in the second and third groups of data, each measuring point recorded 16,000 data points, with a total duration of 4 seconds;

[0087] In order to eliminate the influence of different diagnostic parameters and magnitudes on the diagnostic results, the sample data is normalized. The normalization formula of the sample data is:

[0088]

[0089] Where: X norm represents the data after linear normalization; X represents the original data; X max and X min Respectively represent the maximum and minimum value of a feature in the original data;

[0090] Table 1. Number of data columns and measurement points

[0091] Number of columns Measuring point 1 1# cylinder head 2 8# cylinder head 3 Supercharger 4 Engine side (between cylinders 7 and 8) 5 Engine side (between cylinders 1 and 2) 6 Base (1 cylinder side) 7 Base (8 cylinder side) 8 Generator bearing cover (non-drive end)

[0092] S2: Based on the frost optimization algorithm RIME, the optimal combination of the number of decomposition layers and penalty factors of the VMD algorithm is optimized according to the optimization data set;

[0093] like Figure 2 As shown, the specific steps include:

[0094] S21: Initialize various algorithm parameters of the frost optimization algorithm RIME and the VMD algorithm, set the maximum number of iterations H, and randomly initialize the initial frost particle population of the frost optimization algorithm RIME;

[0095] The parameter combination [k, α] of the VMD algorithm's decomposition level and penalty factor is defined as frost ice particles;

[0096] S22: performing iterative signal decomposition on the fault-free historical vibration signal sequence data according to the VMD algorithm corresponding to each frost particle of the initial frost particle population, and obtaining several intrinsic mode function IMF values;

[0097] The sample entropy or envelope entropy is used as the fitness function to determine the fitness value of each intrinsic mode function IMF value, and the minimum fitness value is used as the current optimal fitness of the optimal frost ice particle;

[0098] Among them, the decomposition process of the VMD algorithm is mainly divided into two parts: the construction and solution of the variational constraint problem:

[0099] 1) Construct variational problem: Assume that the original signal is composed of several band-limited intrinsic mode functions (IMFs), and construct the IMF component μ k (t) The formula to minimize the sum of the total bandwidth, and find the formula that makes all μ k The solution with the minimum sum of estimated bandwidths of (t) is:

[0100]

[0101] Where: is the time derivative; δ(t) is the pulse signal function; j is an imaginary number; * is the symbol of the convolution calculation; ω k is the assumed frequency; f(t) is the original signal;

[0102] 2) Solution of variational problem: After constructing the variational problem, the quadratic penalty factor α and the Lagrange multiplier λ(t) are introduced to transform each constrained variational problem into an unconstrained variational problem:

[0103]

[0104] In addition, the influence of Gaussian noise is greatly reduced by the quadratic penalty term. After the variational problem is transformed, the alternating direction method ADMM of the multiplier is used to update alternately. Until the variational problem is solved, several intrinsic mode function IMF values ​​are obtained;

[0105] S23: Obtain the current adhesion coefficient E of the frost ice particles according to the current iteration number h and the maximum iteration number H, and the expression of the current adhesion coefficient E is:

[0106]

[0107] S24: define a first random number r2 and a second random number r3;

[0108] And determine the size of the first random number r2 and the current adhesion coefficient E;

[0109] If the first random number r2 is greater than or equal to the current adhesion coefficient E, the current number of iterations is updated and step S23 is repeated;

[0110] Otherwise, the position of the frost ice particle is updated to obtain an updated frost ice particle population, and each frost ice particle in the updated frost ice particle population and its corresponding optimal fitness are obtained;

[0111] According to each frost ice particle in the updated frost ice particle population and its corresponding optimal fitness, confirm the current normalized fitness value of the updated frost ice particle population;

[0112] The method for obtaining the normalized fitness value is as follows:

[0113] Define the RIME frost particle population as R, then the frost particle population can pass through x ij Indicates, i represents the sequence number of the frost crystal in the frost particle population, j represents the sequence number of the frost particle in the frost particle population, and its expression is

[0114]

[0115] Then the fitness value of all frost particles is F(S i )for

[0116]

[0117] According to the fitness value F(S i ) Calculate and obtain the normalized fitness value F n (S i ), whose expression is

[0118]

[0119] S25: Determine the magnitude of the second random number r3 and the current normalized fitness value;

[0120] If the second random number r3 is greater than or equal to the current normalized fitness value, then the current number of iterations is updated and steps S23 to S24 are repeated;

[0121] Otherwise, the position of the updated frost ice particle is updated to obtain the latest frost ice particle population, and the latest optimal frost ice particle in the latest frost ice particle population and its corresponding latest optimal fitness are obtained;

[0122] S26: Determine the size of the latest optimal fitness and the current optimal fitness;

[0123] If the latest optimal fitness is greater than or equal to the current optimal fitness, the current number of iterations is updated and steps S23 to S25 are repeated;

[0124] Otherwise, continue to execute step S27;

[0125] S27: Replace the latest optimal fitness with the current optimal fitness, and determine whether the current number of iterations reaches the maximum number of iterations;

[0126] If not, update the current number of iterations and repeat steps S23 to S26;

[0127] If so, the frost ice particle corresponding to the current optimal fitness is used as the final frost ice particle, that is, the parameter combination of the VMD algorithm corresponding to the final frost ice particle is used as the optimal combination.

[0128] The updating formula of the number of iterations is h=h+1;

[0129] The formula for updating the position of the frost particles is:

[0130] R i,j =R best +r1·cosθ·β·(δ·(Ub ij -Lb ij )+Lb ij ),r2<E

[0131]

[0132] Where: R i,j Indicates the new position of the frost particles after the update; R best represents the best frost particle in the frost population; r1 represents a random number of (-1,1); r1 and cosθ jointly control the direction of particle movement; β represents the environmental factor; δ represents a random number of (0,1), which controls the center distance of the frost particle; h represents the current number of iterations; H represents the maximum number of iterations; w represents the number of segments of the control step function; E represents the adhesion coefficient; Ubij With Lb ij They represent the upper and lower bounds of the escape space of frost and ice particles respectively;

[0133] S3: Based on the optimal combination, VMD signal decomposition is performed on the optimized data set to obtain several IMF values ​​to reconstruct the vibration signal sequence data;

[0134] Specifically, in the preset Python script implementation example, first use the pd.read-csv command under the pands library to read the preprocessed data set csv file into the algorithm. There are three groups of data in total. The first group is 20,000 groups of 8-dimensional data, and the second and third groups are 16,000 groups of 8-dimensional data. They are respectively input into the RIME-VMD algorithm, and the sample entropy is used as the fitness function. The specific parameters are tau: smoothing factor, the default is 0. DC: whether to perform DC component processing, the default is 0. init: initialization method, the default is 1. tol: tolerance, the default is 1e-7. The population size is 10, the upper bound of each dimension is [4000,12], the lower bound of each dimension is [100,3], and the maximum number of iterations is 20. After three inputs, good convergence effects were obtained, and the reconstructed data after VMD decomposition were recombined to form a new data set;

[0135] Based on the frequency domain and time domain analysis method, the time domain and frequency domain features of the reconstructed vibration signal sequence data are extracted to obtain the time domain feature parameters and frequency domain feature parameters;

[0136] And obtain the hybrid domain feature set according to the time domain feature parameters and the frequency domain feature parameters, as shown in Table 2;

[0137] The time domain characteristic parameters include mean value, root square amplitude, root mean square, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor and pulse index;

[0138] The frequency domain characteristic parameters include average spectrum amplitude, variance of spectrum amplitude, skewness of spectrum amplitude, kurtosis of spectrum amplitude, centroid frequency, frequency mean square error, root mean square frequency, fourth root frequency, ratio of root mean square frequency to square root of spectrum amplitude, ratio of frequency mean square error to centroid frequency, frequency skewness and frequency kurtosis;

[0139] Table 2. Marine diesel engine vibration prediction characteristic parameters

[0140]

[0141] S4: Obtain the root mean square value of the original vibration signal sequence data as a warning value;

[0142] Calculate and obtain the Pearson coefficient between each feature parameter and the warning value in the hybrid domain feature set, and select the feature parameter whose Pearson coefficient meets the preset Pearson coefficient threshold as the optimized feature parameter;

[0143] Specifically, the mixed domain feature is concentrated into the domain feature parameters, and the Pearson coefficient is used for feature selection. The Pearson coefficient is a statistic that measures the strength and direction of the linear relationship between two variables. The calculation formula of the Pearson coefficient is:

[0144]

[0145] Where X i ,Y i represents different feature parameter variables in the mixed domain feature set, is the characteristic parameter variable X i ,Y i The mean of the characteristic Pearson coefficient is calculated as shown in Table 3

[0146] Table 3. Pearson coefficients of characteristics

[0147]

[0148] In this embodiment, the correlation coefficient in the Pearson correlation coefficient is 0.4-0.6, which is already relevant. Therefore, the feature parameters with a correlation coefficient greater than 0.5 are selected as the input of the BiGRU algorithm, that is, the root mean square, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor, pulse index, pulse index, and average spectrum amplitude are used as feature inputs to complete the feature dimensionality reduction process, thereby avoiding the phenomenon of long calculation time and waste of computing resources;

[0149] S5: randomly dividing the optimized feature parameters into a training set and a test set according to a preset ratio;

[0150] S6: pre-constructing a BiGRU algorithm model, taking the optimized feature parameters in the training set as feature data, taking the warning value corresponding to the feature data as label data as input of the BiGRU algorithm model, and training the BiGRU algorithm model based on the loss function to obtain the optimal BiGRU algorithm model; and then obtaining a RIME-VMD-BIGRU model for vibration fault warning of a ship diesel engine;

[0151] Specifically, Figure 3 As shown, the pre-built BiGRU algorithm model includes an input layer, a first GRU unit, a second GRU unit, and a fusion output layer;

[0152] The input layer is used to transmit input data to the first GRU unit and the second GRU unit respectively;

[0153] The first GRU unit is used to perform a time forward processing operation on the output of the input layer to obtain forward processing data; the second GRU unit is used to perform a time reverse processing operation on the output of the input layer to obtain reverse processing data;

[0154] The fusion output layer is used to merge and output the forward sequence processed data and the reverse sequence processed data;

[0155] The BiGRU algorithm model in this embodiment has a stronger memory capacity and can achieve efficient use of input information because it can simultaneously consider the impact of historical states and future states on the current state. The input feature data is divided into training set and test set at 0.8:0.2. After multiple manual debugging, when the number of hidden layer nodes is 128 and the number of layers is 2, the maximum R2 is 0.9956, and the minimum MAE and MAPE are 0.0022 and 0.3578% respectively, and the model effect is the best; in addition to the above parameters, the optimizer uses Adam, the loss function uses MSEloss, the batch size is 100, and the number of training rounds is 100 rounds;

[0156] The loss function for training the BiGRU algorithm model is

[0157]

[0158] Where: n represents the number of feature data for optimizing feature parameters; y i Represents the true value of the i-th feature data; Represents the predicted value of the i-th feature data;

[0159] Based on the RIME-VMD-BIGRU model, the residual between the predicted RMS value and the actual RMS value is calculated and obtained according to the test set;

[0160] Get the residual error between the predicted RMS value and the actual RMS value, which is expressed as

[0161]

[0162] Where: y i Represents the true value of the i-th feature data; Represents the predicted value of the i-th feature data;

[0163] S7: according to the residual, a residual threshold range [A, B] is set under the condition of trouble-free operation of the ship diesel engine; and A is the minimum residual value obtained according to the test set, and B is the maximum residual value obtained according to the test set;

[0164] S8: Input the acquired current vibration signal sequence data of the ship diesel engine into RIME-VMD-

[0165] BIGRU model, obtains the residual of the current ship diesel engine, and determines whether the residual of the current ship diesel engine meets the residual threshold range; if the residual of the current ship diesel engine meets the residual threshold range, it is confirmed that the vibration operation of the current ship diesel engine is normal; otherwise, it is confirmed that the vibration operation of the current ship diesel engine is faulty, and outputs the vibration fault warning of the current ship diesel engine.

[0166] The experiments and results of this embodiment are analyzed as follows:

[0167] (1) Analysis of RIME-VMD simulation data results

[0168] After the RIME-VMD model is built, in order to verify the reliability of the model, a simulation signal is first constructed for verification. Three pure harmonic signals of different frequencies are used to form a third harmonic composite signal. The center frequencies of the component signals of the composite signal are 8Hz, 68Hz, and 788Hz respectively. In order to better represent the actual signal, Gaussian white noise ζ with a noise intensity of 0.2 is added to the composite signal. The simulation signal calculation formula is:

[0169] f(t)=cos(16πt)+0.5cos(136πt)+0.05cos(1576πt)+ζ

[0170] In this embodiment, the analog signal is decomposed by RIME-VMD, and the composite signal and its IMF decomposition diagram are shown as follows. Figure 4 As shown. The simulation data is decomposed by the optimal parameter combination and arranged in order from low frequency to high frequency. The RIME optimization of VMD parameters can adaptively decompose the simulation signal sequence into 3 subsequences. The volatility of each subsequence is significantly reduced compared to the simulation signal sequence, and there is no obvious modal aliasing phenomenon, which shows the feasibility of RIME-VMD adaptive decomposition. Among the 3 subsequences after decomposition, the volatility of each subsequence is different. Among them, the subsequence IMF1 has the most gentle fluctuation, the smoothest waveform, and the largest average amplitude, reflecting the overall change trend. From subsequence IMF1 to IMF3, the volatility and randomness of the subsequences gradually become stronger, and the average amplitude becomes smaller, reflecting the influence of random factors. It can be seen from the figure that the decomposed IMF can well show the center frequency of each sub-signal, and there is no modal aliasing phenomenon. As shown Figure 5 As shown, the decomposed signal is reconstructed, and compared with the original signal, the noise in the simulated signal is better removed, and the denoised signal is similar to the original composite signal waveform. Therefore, it is proved that the RIME-VMD method proposed in this embodiment is feasible.

[0171] (2) Analysis of actual data results of RIME-VMD model

[0172] Figure 6 is the original sequence diagram of ship vibration signal. This implementation adopts frost algorithm to optimize variational mode decomposition, and uses envelope entropy and sample entropy as fitness functions for optimization. The optimization results of the two fitness functions are compared. Figure 7 , the algorithm shows a fast convergence rate. After the first few iterations, the fitness value quickly stabilizes, indicating that the algorithm quickly finds a local optimal solution. However, due to the large changes in the fitness value, this may indicate that there is a large volatility and risk of local optimality in the optimization process; in contrast, when using sample entropy as the fitness function, Figure 8 ,The algorithm shows continuous improvement in multiple iterations. Although the fitness value changes slightly, it gradually decreases during the iteration process and finally converges to a relatively stable value.,This shows that the sample entropy fitness function can explore the optimization space in more ,details and has the potential to find a better solution.

[0173] In summary, sample entropy as a fitness function may be more suitable for deep optimization problems because it can maintain stability while continuously looking for further optimization possibilities. When compared with the variational mode decomposition (VMD) optimized by the frost algorithm using envelope entropy and sample entropy as fitness functions, sample entropy shows obvious advantages in many aspects, making its application in signal analysis more attractive. Fig. 9 After sample entropy optimization, the number of IMFs is small, usually including only a few main frequency components, which makes the decomposition result more concise. This simplicity not only helps to interpret the main features of the signal more easily, but also avoids the complexity and noise interference that may be introduced by too many IMFs. Fig.10 , although envelope entropy can provide more detailed decomposition and generate more IMFs, some of the modes may contain overlapping frequency components or noise components, resulting in unclear decomposition results. Secondly, the spectrum diagram after sample entropy optimization shows that the frequency distribution of each IMF is extremely clear and concentrated. This sharp frequency peak indicates that each mode represents an independent frequency component in the frequency domain, which greatly improves the accuracy of signal decomposition. This clarity is crucial in frequency analysis because it can accurately locate and analyze specific frequency components in the signal. However, the IMF decomposed by envelope entropy has a wider frequency range and a more dispersed distribution of frequency components, which may result in more unclear frequency bands in the spectrum diagram, reducing the accuracy of frequency analysis.

[0174] In summary, the application of sample entropy in optimizing VMD decomposition can provide a clearer, more concise and easier-to-interpret spectrum diagram, which significantly improves the accuracy of signal decomposition and the operability of practical applications. Therefore, it is superior to envelope entropy in many scenarios.

[0175] Fig.11 This is a comparison chart between the original vibration signal of the ship and the reconstructed signal. As can be seen from the figure, using sample entropy as the fitness function, it can be seen that the noise in the simulation signal is well removed, and the denoised signal is similar to the original composite signal waveform. The denoising method proposed in this chapter accurately reconstructs the noise-free composite signal. The best k values ​​after three outputs are 4.9, 4.5, and 4.83, and the best α values ​​are 4000, 2898.11, and 4000, respectively; because the k value is usually an integer, the rounding method is used to take 4 as the k value for three times. In order to prove the superiority of RIME-VMD, the real ship data after RIME-VMD optimization and before RIME-VMD optimization are input into the BiGRU algorithm for comparison, as shown in Figure 12;

[0176] Table 4. Model performance comparison table

[0177] Prediction Model <![CDATA[R 2 ]]> MAE MAPE RIME-VMD-BiGRU 0.9956 0.0022 0.3578％ BiGRU 0.9924 0.1551 9.3803％

[0178] From Table 4, we can see that, first, from R 2 From the value of 2 is 0.9956, close to 1, indicating that the model fits the data very well. 2 It is 0.9924, which is relatively high, but slightly lower than RIME-VMD-BiGRU, indicating that its fitting effect is slightly inferior. In terms of MAE, the RIME-VMD-BiGRU model performs better, with a MAE of only 0.0022, which means that the average error between the predicted value and the true value of the model is extremely small, while the MAE of the BiGRU model is 0.1551, which is relatively large, indicating that its prediction accuracy is not as good as RIME-VMD-BiGRU. The MAPE indicator reflects the relative percentage of prediction error. The MAPE of RIME-VMD-BiGRU is only 0.3578%, showing that the error of the model is very low when predicting, which is almost negligible. In contrast, the MAPE of BiGRU is 9.3803%, and the error is relatively large, indicating that in practical applications, the prediction results of this model may deviate a lot from the actual values.

[0179] In summary, RIME-VMD-BiGRU outperforms BiGRU in all indicators, showing stronger prediction ability and higher accuracy, and is suitable for prediction tasks with high accuracy requirements. According to the evaluation criteria of the three parameters, data processed by RIME-VMD is more superior for fault prediction.

[0180] To further prove the superiority of RIME-VMD-BiGRU, we compared it with other algorithms and conducted ablation experiments. The comparison results are as follows: Figures 13 to 16 As shown;

[0181] Table 5. Comparison of results of different prediction models

[0182] Prediction Model <![CDATA[R 2 ]]> MAE MAPE RIME-VMD-BiGRU 0.9956 0.0022 0.3578％ RIME-VMD-GRU 0.9723 0.0056 0.9185％ RIME-VMD-LSTM 0.9697 0.0058 0.9579％ RIME-VMD-RNN 0.9896 0.0037 0.6122％

[0183] As can be seen from Table 5, the four different prediction models are shown in terms of three key performance indicators (R 2 , MAE, and MAPE), which are RIME-VMD-BiGRU, RIME-VMD-GRU, RIME-VMD-LSTM, and RIME-VMD-RNN respectively. Through these indicators, we can deeply analyze the prediction accuracy and fitting ability of each model.

[0184] First, from R 2 The R values ​​of all models are 2 All of them are above 0.96, indicating that they fit the data very well. 2 The highest is 0.9956, which shows its strong fitting ability. 2 The value is 0.9896, which is close behind, also showing excellent fitting ability. In contrast, the R 2 The values ​​are 0.9723 and 0.9697 respectively, which are slightly lower than the previous two, but still at a high level, indicating that their performance in fitting is also very reliable.

[0185] In terms of MAE (mean absolute error), RIME-VMD-BiGRU once again showed a clear advantage, with a MAE value of only 0.0022, which means that the average error between its predicted value and the actual value is extremely small, and it has the smallest error among all models. RIME-VMD-RNN has a MAE of 0.0037, which also performs well. In contrast, the MAEs of RIME-VMD-GRU and RIME-VMD-LSTM are 0.0056 and 0.0058, respectively, with slightly larger errors, indicating that their prediction accuracy is slightly inferior.

[0186] MAPE (mean absolute percentage error) is an important indicator to measure the relative percentage of prediction error. In this indicator, RIME-VMD-BiGRU also performs well, with a MAPE of only 0.3578%, showing that its prediction error in practical applications is very low and almost negligible. The MAPE of RIME-VMD-RNN is 0.6122%, which also shows a relatively low error. The MAPE of RIME-VMD-GRU and RIME-VMD-LSTM are 0.9185% and 0.9579% respectively, with relatively large errors, which indicates that in actual predictions, they may have relatively larger deviations.

[0187] In summary, the RIME-VMD-BiGRU model outperforms other models in all indicators, showing excellent prediction ability and extremely high accuracy, and is the most ideal prediction model in current research. Although RIME-VMD-RNN is slightly inferior, it still has strong fitting ability and high prediction accuracy. In contrast, although RIME-VMD-GRU and RIME-VMD-LSTM also perform well to a certain extent, their accuracy and fitting effect are slightly inferior to the former two. In application scenarios with high requirements for prediction accuracy, RIME-VMD-BiGRU is undoubtedly the most advantageous choice. According to the evaluation criteria of the three parameters, RIME-VMD-BIGRU has more advantages than other prediction models.

[0188] (3) Threshold setting

[0189] In the application of the prediction model, this embodiment first analyzes the residual between the model prediction value and the actual value, and sets the maximum and minimum residual values ​​as thresholds. Fig.17 As shown in the figure, under normal working conditions, the residuals between the predicted values ​​and the actual values ​​of the model show a good distribution state, and all residuals remain within the preset upper and lower thresholds. This shows that when the system is operating without faults, the prediction accuracy of the model is high and the false alarm phenomenon is effectively avoided. It can be seen that the set threshold can effectively constrain the alarm behavior of the system under normal conditions and reduce unnecessary interference.

[0190] In order to further verify the effectiveness of the proposed method, since the data set of this embodiment does not contain fault data, an artificial linear adjustment method is adopted to adjust the data set to lower or increase the root mean square value, thereby simulating the potential fault hazards of the diesel engine. Fig.18As shown in the figure, the results show that when the predicted value deviates significantly from the actual value and exceeds the set upper and lower thresholds, the system successfully triggers a fault alarm. This result strongly proves that the early warning method proposed in this embodiment can timely and effectively detect abnormal conditions in the operation of the diesel engine, especially when the root mean square value fluctuates abnormally, the system can accurately issue a fault warning to prompt the on-duty personnel to perform necessary inspections and maintenance. At the same time, the method also successfully avoids the possibility of false alarms under normal conditions, ensuring the stability and reliability of the system operation.

[0191] 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 it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A marine diesel engine vibration fault early warning method based on improved VMD, characterized in that: The specific steps include: S1: Obtain the fault-free historical vibration signal sequence data of the ship diesel engine, and perform normalization preprocessing to obtain the optimized data set; S2: Based on the frost optimization algorithm RIME, the optimal combination of the number of decomposition layers and penalty factors of the VMD algorithm is optimized according to the optimization data set; S3: Based on the optimal combination, VMD signal decomposition is performed on the optimized data set to obtain several IMF values ​​to reconstruct the vibration signal sequence data; Based on the frequency domain and time domain analysis method, the time domain and frequency domain features of the reconstructed vibration signal sequence data are extracted to obtain the time domain feature parameters and frequency domain feature parameters; And obtaining a mixed domain feature set according to the time domain feature parameters and the frequency domain feature parameters; S4: Obtain the root mean square value of the original vibration signal sequence data as a warning value; Calculate and obtain the Pearson coefficient between each feature parameter and the warning value in the hybrid domain feature set, and select the feature parameter whose Pearson coefficient meets the preset Pearson coefficient threshold as the optimized feature parameter; S5: randomly dividing the optimized feature parameters into a training set and a test set according to a preset ratio; S6: pre-constructing a BiGRU algorithm model, taking the optimized feature parameters in the training set as feature data, taking the warning value corresponding to the feature data as label data and as input of the BiGRU algorithm model, training the BiGRU algorithm model based on the loss function to obtain the optimal BiGRU algorithm model; and then obtaining a RIME-VMD-BIGRU model for vibration fault warning of a ship diesel engine; Based on the RIME-VMD-BIGRU model, the residual between the predicted RMS value and the actual RMS value is calculated and obtained according to the test set; S7: according to the residual, a residual threshold range [A, B] is set under the condition of trouble-free operation of the ship diesel engine; and A is the minimum residual value obtained according to the test set, and B is the maximum residual value obtained according to the test set; S8: Input the obtained current vibration signal sequence data of the ship diesel engine into the RIME-VMD-BIGRU model, obtain the residual of the current ship diesel engine, and determine whether the residual of the current ship diesel engine meets the residual threshold range [A, B]; if the residual of the current ship diesel engine meets the residual threshold range [A, B], it is confirmed that the vibration operation of the current ship diesel engine is normal; otherwise, it is confirmed that the vibration operation of the current ship diesel engine is faulty, and the vibration fault warning of the current ship diesel engine is output.

2. A marine diesel engine vibration fault early warning method based on improved VMD according to claim 1, characterized in that: The S2 specifically includes the following steps: S21: Set the maximum number of iterations and randomly initialize the initial frost particle population of the frost optimization algorithm RIME; The parameter combination [k, α] of the VMD algorithm's decomposition level and penalty factor is defined as frost ice particles; S22: performing iterative signal decomposition on the fault-free historical vibration signal sequence data according to the VMD algorithm corresponding to each frost particle of the initial frost particle population, and obtaining several intrinsic mode function IMF values; The sample entropy or envelope entropy is used as the fitness function to determine the fitness value of each intrinsic mode function IMF value, and the minimum fitness value is used as the current optimal fitness of the optimal frost ice particle; S23: Obtaining the current adhesion coefficient E of the frost ice particles according to the current number of iterations and the maximum number of iterations; S24: define a first random number r2 and a second random number r3; And determine the size of the first random number r2 and the current adhesion coefficient E; If the first random number r2 is greater than or equal to the current adhesion coefficient E, the current number of iterations is updated and step S23 is repeated; Otherwise, the position of the frost ice particle is updated to obtain an updated frost ice particle population, and each frost ice particle in the updated frost ice particle population and its corresponding optimal fitness are obtained; According to each frost ice particle in the updated frost ice particle population and its corresponding optimal fitness, confirm the current normalized fitness value of the updated frost ice particle population; S25: Determine the magnitude of the second random number r3 and the current normalized fitness value; If the second random number r3 is greater than or equal to the current normalized fitness value, then the current number of iterations is updated and steps S23 to S24 are repeated; Otherwise, the position of the updated frost ice particle is updated to obtain the latest frost ice particle population, and the latest optimal frost ice particle in the latest frost ice particle population and its corresponding latest optimal fitness are obtained; S26: Determine the size of the latest optimal fitness and the current optimal fitness; If the latest optimal fitness is greater than or equal to the current optimal fitness, the current number of iterations is updated and steps S23 to S25 are repeated; Otherwise, continue to execute step S27; S27: Replace the latest optimal fitness with the current optimal fitness, and determine whether the current number of iterations reaches the maximum number of iterations; If not, update the current number of iterations and repeat steps S23 to S26; If so, the frost ice particle corresponding to the current optimal fitness is used as the final frost ice particle, that is, the parameter combination of the VMD algorithm corresponding to the final frost ice particle is used as the optimal combination.

3. A marine diesel engine vibration fault early warning method based on improved VMD according to claim 1, characterized in that: The time domain characteristic parameters and frequency domain characteristic parameters obtained in S3 are specifically: The time domain characteristic parameters include mean value, root square amplitude, root mean square, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, waveform factor and pulse index; The frequency domain characteristic parameters include average spectrum amplitude, variance of spectrum amplitude, skewness of spectrum amplitude, kurtosis of spectrum amplitude, centroid frequency, frequency mean square error, root mean square frequency, fourth root frequency, ratio of root mean square frequency to square root of spectrum amplitude, ratio of frequency mean square error to centroid frequency, frequency skewness and frequency kurtosis.

4. The method for early warning of vibration fault of marine diesel engine based on improved VMD according to claim 1, characterized in that: The loss function for training the BiGRU algorithm model in S6 is Where: n represents the number of feature data for optimizing feature parameters; y i Represents the true value of the i-th feature data; Represents the predicted value of the i-th feature data; The residual error between the predicted RMS value and the actual RMS value is expressed as Where: y i Represents the true value of the i-th feature data; Represents the predicted value of the i-th feature data.

5. The method for early warning of vibration fault of marine diesel engine based on improved VMD according to claim 1, characterized in that: The BiGRU algorithm model pre-constructed in S6 includes an input layer, a first GRU unit, a second GRU unit, and a fusion output layer; The input layer is used to transmit input data to the first GRU unit and the second GRU unit respectively; The first GRU unit is used to perform a time forward processing operation on the output of the input layer to obtain forward processing data; the second GRU unit is used to perform a time reverse processing operation on the output of the input layer to obtain reverse processing data; The fusion output layer is used to merge and output the forward-order processed data and the reverse-order processed data.

Images