A method for identifying marine propeller blade faults based on deep neural network

Abstract

The present invention discloses a method for identifying marine blade faults based on a deep neural network, the method comprising the following steps: 1) installing a vibration sensor group at a stern shaft structure, a gearbox and a diesel engine connected to a ship propeller; 2) collecting data through a vibration sensor, arranging the data according to the vibration sensor number, and extracting a vibration detection data sequence; 3) collecting a large number of classified fault and normal data sequence sets of ship propeller blades, and expanding them into a data training set and a data verification set through data augmentation; 4) using a VGGNet network as a marine blade fault identification model network to train the data set, and obtain a corresponding marine blade fault identification model. The method of the present invention provides a method for extracting data features by performing a deep neural network on the vibration detection data of a marine propeller connection structure, the method requires a small amount of computation, and has a high marine blade fault identification rate.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning technology and vibration detection, and in particular to a method for identifying marine propeller blade faults based on a deep neural network. Background Art

[0002] Ships rely on propeller blades to rotate in the water, converting the diesel engine's rotational power into thrust to navigate in the water. As the main power equipment of ships, if the propeller is damaged, the normal operation of the ship will be seriously affected. During the long-term use of propellers, phenomena such as blade deformation, fracture, and cavitation will occur, which seriously threatens the safety of ship navigation.

[0003] Normally, there are two ways to diagnose whether a propeller blade is faulty. One is to rely on divers to conduct underwater surveys, and the other is to dock the ship for inspection. Both of these traditional methods are too expensive and inefficient for propeller blade inspection, and the accuracy of the diagnosis depends largely on the level of the operator. The marine propeller blade fault identification method based on deep neural networks solves and improves the above problems by collecting vibration information of structures such as the stern shaft connected to the propeller. It can perform propeller blade inspections without stopping the ship, and it is highly economical. Summary of the invention

[0004] The present invention mainly aims at the hot issues in the prior art field and provides a method for identifying marine propeller blade faults based on a deep neural network.

[0005] The technical solution of the present invention to solve the corresponding technical problem is: a method for identifying marine propeller blade faults based on a deep neural network, comprising the following steps:

[0006] 1) Install the vibration sensor group on the stern shaft structure connected to the ship's propeller, the gearbox and the diesel engine;

[0007] 2) Collecting data through the vibration sensor, arranging the data according to the vibration sensor number, and extracting the vibration detection data sequence;

[0008] 3) Collect a large number of classified fault and normal data series of ship propeller blades, and expand them into data training sets and data verification sets through data augmentation;

[0009] 4) Use the VGGNet network as the marine blade fault recognition model network to train the data set and obtain the corresponding marine blade fault recognition model.

[0010] The step 1) installs the vibration sensor group on the stern shaft structure connected to the ship propeller, the gearbox and the diesel engine, and specifically includes:

[0011] 2.1) The vibration sensor used in the vibration sensor group is an acceleration sensor, the sensor operating frequency is 4096Hz, the range is (0, 50g), and the installation method is bolt fixing;

[0012] 2.2) Install a vibration sensor, numbered 1, at the connection bearing of the stern shaft intermediate shaft;

[0013] 2.3) Install vibration sensors at the front and rear bearings of the intermediate shaft, numbered 2 and 3 respectively.

[0014] 2.3) Install vibration sensors numbered 4 and 5 at the gearbox output shaft and input shaft bearings.

[0015] 2.3) Install vibration sensors, numbered 6 and 7, at both ends of the diesel engine power shaft.

[0016] The step 2) collects data through the vibration sensor, arranges the data according to the vibration sensor number, and extracts the vibration detection data sequence, specifically including:

[0017] 3.1) All vibration sensors start collecting vibration data at the same time. According to the sensor number, the corresponding position data is recorded as g1, g2, g3, g4, g5, g6, g7;

[0018] 3.2) Arrange the vibration data within any collection time period to form sequence data as follows:

[0019]

[0020] Where: D represents the vibration detection data sequence within any acquisition period, t0, t n Respectively represent the starting point and end point of any collection time period.

[0021] The step 3) collects a large number of classified fault and normal data series of ship propeller blades, and expands them into data training sets and data verification sets through data augmentation, specifically including:

[0022] 4.1) Collect a large number of vibration detection data sequences of ship propeller connection structures. The detection sequence data can be within any time period;

[0023] 4.2) The collected vibration detection data sequence is marked with data, and the data tags are divided into 6 categories; the label results are vectorized and expressed as a 6-dimensional vector, and the vector elements correspond to the 6 vibration detection results of normal, blade deformation, blade interference friction, blade sound, blade cavitation and blade fracture, respectively. If the corresponding existence is recorded as 1, otherwise it is recorded as 0;

[0024] 4.3) performing data augmentation operation on the collected vibration detection data sequence, randomly setting the elements of a row of the vibration detection data sequence to 0, and adding the elements to the vibration detection data sequence set;

[0025] 4.4) The detection data sequence set after the vibration detection data augmentation operation is divided into the model training data set STRA and the model verification data set STES according to the ratio ε, and the value of ε is as follows:

[0026]

[0027] The step 4) uses the VGGNet network as the marine propeller blade fault identification model network to train the data training set to obtain the corresponding fault identification model, specifically including:

[0028] 5.1) Set the activation function of the VGGNet network to LReLU, as follows:

[0029]

[0030] 5.2) Set the pooling window of the VGGNet network to a 1×2 dimensional matrix, the step size to 2, and the pooling method to average pooling;

[0031] 5.3) Set the convolution layer V1, use 64 convolution kernels of size 1×2 to perform one convolution operation with a step size of 2, then activate each neuron and perform pooling downsampling;

[0032] 5.4) Set the convolution layer V2, use 128 convolution kernels of size 1×2 to perform one convolution operation with a step size of 2, and then activate each neuron;

[0033] 5.5) Set the convolution layer V3, use 256 convolution kernels of size 1×2 to perform 2 convolution operations with a step size of 2, then activate each neuron and perform pooling downsampling;

[0034] 5.6) Set the convolution layer V4, use 512 convolution kernels of size 1×2 to perform 2 convolution operations with a step size of 2, and then activate each neuron;

[0035] 5.7) Set the convolution layer V5 to perform 2 convolution operations using 512 convolution kernels of size 1×2 with a stride of 2;

[0036] Set the pooling method after convolution layer V5 to spatial pyramid pooling, the pooling window to a 1×n-dimensional matrix, the step size to n, and the pooling method to maximum pooling, where n is the total number of frames in the detection data sequence;

[0037] 5.8) Set the fully connected layer FC6 with a convolution kernel size of 1×1×1024;

[0038] 5.9) Set the fully connected layer FC7 with a convolution kernel size of 1×1×1024;

[0039] 5.10) Set the output layer with an output size of 1×6, corresponding to the detection data result labeling result classification;

[0040] 5.11) Use the stochastic gradient descent optimization algorithm to update the parameters of the network convolution kernel;

[0041] The parameters of the convolution kernel are initialized using a normal distribution function with a standard deviation of 0.1, and all bias items of the convolution layer are initialized to 0;

[0042] 5.12) Use the model training data set STRA to train the VGGNet network for one round, update the network parameters, and obtain the marine propeller blade fault recognition model;

[0043] 5.13) Use the model verification data set STES to verify the recognition model obtained in 5.12). When the recognition accuracy is greater than or equal to 0.95, the training is completed and a marine propeller blade fault recognition model with correct parameters is obtained. Otherwise, proceed to step 5.12).

[0044] The beneficial effects produced by the present invention are:

[0045] 1. The production tools required by the present invention are easy to obtain, and only a computer, a vibration sensor and a collection device are needed to realize the fault diagnosis of non-water-separated marine blades, which is highly economical;

[0046] 2. The present invention uses a data augmentation method to expand the data set, providing rich samples for model training, while also restoring the phenomenon of bad points of vibration sensors under real working conditions;

[0047] 3. The present invention adopts spatial pyramid pooling as the last layer pooling method of the network, which can process vibration detection data sequence input of any size during VGGNet training. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below with reference to the accompanying drawings, in which:

[0049] Figure 1 This is a flow chart of a method for identifying marine propeller blade faults based on a deep neural network according to the present invention;

[0050] Figure 2 This is a schematic diagram of the installation of the vibration sensor;

[0051] Figure 3 It is a schematic diagram of the vibration detection data sequence label result conversion designed by the present invention;

[0052] Figure 4 It is a diagram of the VGGNet network parameter settings used in the present invention. DETAILED DESCRIPTION

[0053] In order to make the technical solution and problem-solving ideas of the present invention clearer, the present invention is further described in detail below with reference to examples. It should be understood that the specific examples described here are only used to explain the present invention and are not used to limit the present invention.

[0054] like Figure 1 As shown, a method for identifying marine propeller blade faults based on a deep neural network includes the following steps:

[0055] Step STEP1, install the vibration sensor group on the stern shaft structure connected to the ship propeller, the gearbox and the diesel engine, such as Figure 2 As shown, specifically including;

[0056] a) The vibration sensor used in the vibration sensor group is an acceleration sensor, the sensor operating frequency is 4096Hz, the range is (0, 50g), and the installation method is bolt fixing;

[0057] b) Install a vibration sensor, numbered 1, at the connection bearing of the stern shaft intermediate shaft;

[0058] c) Install vibration sensors at the front and rear bearings of the intermediate shaft, numbered 2 and 3 respectively.

[0059] d) Install vibration sensors, numbered 4 and 5, at the gearbox output shaft and input shaft bearings.

[0060] e) Install vibration sensors, numbered 6 and 7, at both ends of the diesel engine power shaft.

[0061] Step STEP2, collecting data through the vibration sensor, arranging the data according to the vibration sensor number, and extracting the vibration detection data sequence, specifically including:

[0062] a) All vibration sensors start collecting vibration data at the same time. According to the sensor number, the corresponding position data is recorded as g1, g2, g3, g4, g5, g6, g7;

[0063] b) Arrange the vibration data within any collection period to form sequence data, as follows:

[0064]

[0065] Where: D represents the vibration detection data sequence within any acquisition period, t0, t n Respectively represent the starting point and end point of any collection time period.

[0066] Step STEP3, collecting a large number of classified fault and normal data sequence sets of ship propeller blades, and expanding them into data training sets and data verification sets through data augmentation, specifically including;

[0067] a) Collect a large number of vibration detection data sequences of ship propeller connection structures, and the detection sequence data can be within any time period;

[0068] b) The collected vibration detection data sequence is marked with data, and the data tags are divided into 6 categories; the label results are vectorized and represented as a 6-dimensional vector, such as Figure 3 As shown in the figure, the vector elements correspond to the six vibration detection results of normal, blade deformation, blade interference friction, blade sound, blade cavitation and blade breakage. If the corresponding result exists, it is recorded as 1, otherwise it is recorded as 0.

[0069] c) performing a data augmentation operation on the collected vibration detection data sequence, randomly setting an element of a row of the vibration detection data sequence to 0, and adding the element to the vibration detection data sequence set;

[0070] d) The detection data sequence set after the vibration detection data augmentation operation is divided into the model training data set STRA and the model verification data set STES according to the proportion ε, and the value of ε is as follows:

[0071]

[0072] Step STEP4, using the VGGNet network as the marine propeller blade fault identification model network to train the data training set to obtain the corresponding fault identification model, specifically including;

[0073] a) Set the activation function of the VGGNet network to LReLU, as follows:

[0074]

[0075] b) Set the pooling window of the VGGNet network to a 1×2 dimensional matrix, the step size to 2, and the pooling method to average pooling;

[0076] c) Set the convolution layer V1, use 64 convolution kernels of size 1×2 to perform one convolution operation with a step size of 2, then activate each neuron and perform pooling downsampling;

[0077] d) Set the convolution layer V2, use 128 convolution kernels of size 1×2 to perform one convolution operation with a step size of 2, and then activate each neuron;

[0078] e) Set the convolution layer V3, use 256 convolution kernels of size 1×2 to perform 2 convolution operations with a step size of 2, then activate each neuron and perform pooling downsampling;

[0079] f) Set the convolution layer V4, use 512 convolution kernels of size 1×2 to perform 2 convolution operations with a step size of 2, and then activate each neuron;

[0080] g) Set the convolution layer V5 to perform two convolution operations using 512 convolution kernels of size 1×2 with a stride of 2;

[0081] Set the pooling method after convolution layer V5 to spatial pyramid pooling, the pooling window to a 1×n-dimensional matrix, the step size to n, and the pooling method to maximum pooling, where n is the total number of frames in the detection data sequence;

[0082] h) Set the fully connected layer FC6 with a convolution kernel size of 1×1×1024;

[0083] i) Set the fully connected layer FC7 with a convolution kernel size of 1×1×1024;

[0084] j) Set the output layer with an output size of 1×6, corresponding to the detection data result labeling result classification;

[0085] k) Update the parameters of the network convolution kernel using the stochastic gradient descent optimization algorithm;

[0086] The parameters of the convolution kernel are initialized using a normal distribution function with a standard deviation of 0.1, and all bias items of the convolution layer are initialized to 0;

[0087] l) Use the model training data set STRA to train the VGGNet network for one round, update the network parameters, and obtain the marine propeller blade fault recognition model;

[0088] m) Use the model verification data set STES to verify the recognition model obtained in 5.12). When the recognition accuracy is greater than or equal to 0.95, the training is completed and a marine propeller blade fault recognition model with correct parameters is obtained. Otherwise, proceed to step 1).

Claims

1. A method for identifying marine propeller blade faults based on a deep neural network, which is characterized by the following steps: 1) Install the vibration sensor group on the stern shaft structure connected to the ship's propeller, the gearbox and the diesel engine; 2) Collecting data through the vibration sensor, arranging the data according to the vibration sensor number, and extracting the vibration detection data sequence; 3) Collect a large number of classified fault and normal data series of ship propeller blades, and expand them into data training sets and data verification sets through data augmentation; 4) Use the VGGNet network as the marine propeller blade fault recognition model network to train the data set and obtain the corresponding marine propeller blade fault recognition model; The step 2) collects data through the vibration sensor, arranges the data according to the vibration sensor number, and extracts the vibration detection data sequence, specifically including: 3.1) All vibration sensors start collecting vibration data at the same time. According to the sensor number, the corresponding position data is recorded as g1, g2, g3, g4, g5, g6, g7; 3.2) Arrange the vibration data within any collection time period to form sequence data as follows: Where: D represents the vibration detection data sequence within any acquisition period, t0, t n Respectively represent the starting point and end point of any collection time period; The step 3) collects a large number of classified fault and normal data series of ship propeller blades, and expands them into data training sets and data verification sets through data augmentation, specifically including: 4.1) Collect a large number of vibration detection data sequences of ship propeller connection structures. The detection sequence data can be within any time period; 4.2) The collected vibration detection data sequence is marked with data, and the data tags are divided into 6 categories; the label results are vectorized and expressed as a 6-dimensional vector, and the vector elements correspond to the 6 vibration detection results of normal, blade deformation, blade interference friction, blade sound, blade cavitation and blade fracture, respectively. If the corresponding existence is recorded as 1, otherwise it is recorded as 0; 4.3) performing data augmentation operation on the collected vibration detection data sequence, randomly setting the elements of a row of the vibration detection data sequence to 0, and adding the elements to the vibration detection data sequence set; 4.4) The vibration detection data sequence set after the vibration detection data augmentation operation is divided into the model training data set STRA and the model verification data set STES according to the ratio ε, and the value of ε is as follows:

2. A method for identifying marine propeller blade faults based on a deep neural network according to claim 1, characterized in that: The step 1) of installing the vibration sensor group on the stern shaft structure connected to the propeller of the ship, the gearbox and the diesel engine specifically includes: 2.1) The vibration sensor used in the vibration sensor group is an acceleration sensor, the sensor operating frequency is 4096Hz, the range is (0, 50g), and the installation method is bolt fixing; 2.2) Install a vibration sensor, numbered 1, at the connection bearing of the stern shaft intermediate shaft; 2.3) Install vibration sensors at the front and rear bearings of the intermediate shaft, numbered 2 and 3 respectively; 2.4) Install vibration sensors, numbered 4 and 5, at the output shaft and input shaft bearings of the gearbox; 2.5) Install vibration sensors, numbered 6 and 7, at both ends of the diesel engine power shaft.

3. The method for identifying marine propeller blade faults based on a deep neural network according to claim 1, characterized in that: The step 4) uses the VGGNet network as the marine propeller blade fault identification model network to train the data training set to obtain the corresponding fault identification model, specifically including: 5.1) Set the activation function of the VGGNet network to LReLU, as follows: 5.2) Set the pooling window of the VGGNet network to a 1×2 dimensional matrix, the step size to 2, and the pooling method to average pooling; 5.3) Set the convolution layer V1, use 64 convolution kernels of size 1×2 to perform one convolution operation with a step size of 2, then activate each neuron and perform pooling downsampling; 5.4) Set the convolution layer V2, use 128 convolution kernels of size 1×2 to perform one convolution operation with a step size of 2, and then activate each neuron; 5.5) Set the convolution layer V3, use 256 convolution kernels of size 1×2 to perform 2 convolution operations with a step size of 2, then activate each neuron and perform pooling downsampling; 5.6) Set the convolution layer V4, use 512 convolution kernels of size 1×2 to perform 2 convolution operations with a step size of 2, and then activate each neuron; 5.7) Set the convolution layer V5 to perform 2 convolution operations using 512 convolution kernels of size 1×2 with a stride of 2; Set the pooling method after convolution layer V5 to spatial pyramid pooling, the pooling window to a 1×n-dimensional matrix, the step size to n, and the pooling method to maximum pooling, where n is the total number of frames in the detection data sequence; 5.8) Set the fully connected layer FC6 with a convolution kernel size of 1×1×1024; 5.9) Set the fully connected layer FC7 with a convolution kernel size of 1×1×1024; 5.10) Set the output layer with an output size of 1×6, corresponding to the vibration detection data result labeling result classification; 5.11) Use the stochastic gradient descent optimization algorithm to update the parameters of the network convolution kernel, use the normal distribution function with a standard deviation of 0.1 to initialize the parameters of the convolution kernel, and initialize all the bias items of the convolution layer to 0; 5.12) Use the model training data set STRA to train the VGGNet network for one round, update the network parameters, and obtain the marine propeller blade fault recognition model; 5.13) Use the model verification data set STES to verify the recognition model obtained in 5.12). When the recognition accuracy is greater than or equal to 0.95, the training is completed and a marine propeller blade fault recognition model with correct parameters is obtained. Otherwise, proceed to step 5.12).

Images