Fault diagnosis method and system for novel oil-electricity hybrid power offshore operation platform

Abstract

The invention discloses a novel fault diagnosis method and system for an oil-electricity hybrid power offshore operation platform, and the method is based on a deep residual convolutional neural network of enhanced discriminant feature learning, forms a residual network through introducing a residual learning unit, can avoid the problem of low accuracy in a conventional convolutional neural network, and improves the fault diagnosis accuracy of the oil-electricity hybrid power offshore operation platform. The method has strong learning and analysis capabilities. According to the novel fault diagnosis system of the oil-electricity hybrid power offshore operation platform, potential features of samples are learned through multi-layer nonlinear network training, so that the classification or prediction capability is improved, and the fault diagnosis system has advantages in the aspect of system-level complex fault diagnosis. Besides, the oil-electricity hybrid power system provided by the invention comprises a multi-task convolutional neural network, can process multiple tasks at the same time, allows different tasks to share certain layers, and reserves a specific layer for each task, thereby effectively improving the efficiency and accuracy of fault diagnosis of the oil-electricity hybrid power system of the offshore operation platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis of marine platform power systems, and particularly relates to a fault diagnosis method and system for a new type of oil-electric hybrid offshore operation platform. Background Art

[0002] In order to actively respond to the call and promote the green development of the industry, offshore operation platforms use oil-electric hybrid as the driving energy. With the gradual application of the hybrid power system, the automation level and complexity of the platform power system are continuously increasing. The platform power system faces a higher probability of failure in a complex working condition environment. At the same time, the high system complexity also increases the difficulty of fault diagnosis and fault location of offshore platforms.

[0003] In terms of fault diagnosis of power equipment, at present, the method of comparing operation data with pre-set normal parameters is mainly used, which has the advantages of being fast and direct, etc., but it cannot accurately judge and locate complex system faults. Knowledge-based fault diagnosis is to summarize the historical diagnosis experience of experts and perform diagnosis through reasoning. The process is simple and fast, but there are problems such as knowledge acquisition bottlenecks, and it is not applicable to the fault diagnosis of complex systems or inexperienced systems; in terms of the intelligent operation of the power system, the current monitoring system mostly uses data in an ideal state as the reference basis and performs automatic control operations with pre-set control coefficients, which is difficult to meet the needs of real-time evaluation, prediction and adjustment of complex systems in a dynamically changing operation environment. At the same time, due to the complex equipment in the power cabin of the offshore platform, there are many monitoring items, and the causes of faults and troubleshooting are of great complexity and difficulty, which makes it easy for staff to make missed judgments or wrong judgments.

[0004] Therefore, in order to improve the reliability of the power system operation, reduce the probability of accidents, and avoid unnecessary maintenance costs, a systematic method that can timely detect abnormal conditions in the operation of power equipment and eliminate faults is needed. Summary of the Invention

[0005] Objective of the Invention: The objective of the present invention is to provide a fault diagnosis method and system for a new type of oil-electric hybrid offshore operation platform. Based on a deep residual convolutional neural network with enhanced discriminative feature learning, by introducing a residual learning unit to form a residual network, it can avoid the problem of low accuracy in traditional convolutional neural networks (CNNs), and has powerful learning and analysis capabilities. Through training with a multi-layer non-linear network, the potential features of the samples can be learned, thereby improving the classification or prediction ability. Therefore, it has advantages in system-level complex fault diagnosis. In addition, different from single-task fault diagnosis methods, the oil-electric hybrid system proposed in the present invention includes a multi-task convolutional neural network (MTCNN), which can process multiple tasks simultaneously, allows sharing of certain layers between different tasks, and retains specific layers for each task, effectively improving the efficiency and accuracy of fault diagnosis for the oil-electric hybrid system.

[0006] Technical Solution: A fault diagnosis method for a new type of oil-electric hybrid offshore operation platform of the present invention includes the following steps:

[0007] Step 1: Collect power system signals through sensors, divide the time series data of the signals into multiple segments according to the sampling rate and motor speed, and then divide the data segments into a training set and a test set;

[0008] Step 2: Convert the one-dimensional time series data of the data segments into two-dimensional grayscale images, and convert domain knowledge into information graphs. Combine the two-dimensional grayscale images and the information graphs to construct a fused information image;

[0009] Step 3: Using the fused information image obtained in Step 2 as the input, construct a multi-task convolutional neural network (MTCNN) model based on discriminative feature attention (channel attention module (CAM) and non-local attention module (NLAM)), and use multiple classifiers for multi-task diagnosis.

[0010] Step 4: Introduce a residual learning unit and train the multi-task convolutional neural network model with the training set;

[0011] Step 5: Test the multi-task convolutional neural network model obtained in Step 4 with the data of the test set. If the accuracy of the output data reaches the standard, proceed to Step 6; otherwise, if the accuracy of the output data does not reach the standard, return to Step 3;

[0012] Step 6: Apply the multi-task convolutional neural network model to the new type of oil-electric hybrid offshore operation platform for fault diagnosis.

[0013] Further, Step 1 specifically includes the following steps:

[0014] Step 1.1: Collect the operation signal data of the motor, thruster, and other power equipment through the sensors on each device;

[0015] Step 1.2: Make a preliminary judgment on the obtained signal data and clear the signal data with obvious large errors;

[0016] Step 1.3: Estimate the sample length according to the sampling frequency and the motor rotation speed;

[0017] Step 1.4: Divide the processed data into a training set and a test set according to the sample length calculated in Step 1.3.

[0018] Further, Step 2 specifically includes the following steps:

[0019] Step 2.1: Convert the one-dimensional time series data obtained by the sensor into a two-dimensional image. Based on this conversion program, the segments of the original data are converted into grayscale images of a standard size;

[0020] For each data sample segment, apply a sliding window to the original data. The grayscale pixels of the first sample segment are calculated according to formula (1) and used as the first row of the grayscale image. Based on this conversion program, the segments of the original data are converted into grayscale images of size M×M. The pixel intensity values of the image are obtained by normalizing the original data to the range of (0, 255). Define I(i, j) as the original time series data, and the pixel intensity value G(i, j) of the image is obtained by formula (1):

[0021]

[0022] where i = 1, …, M; j = 1, …, M; M is the coordinate of the image, I max and I min represent the maximum and minimum values of the original data I respectively, and round(·) is the rounding function;

[0023] Step 2.2: Construct an information graph from the grayscale image obtained in Step 2.1;

[0024] Step 2.3: Construct an operation information graph, which represents the operation condition information in grayscale levels and is obtained by calculation through a formula;

[0025] For each fault mode, use the original data from N sensors; construct an empty information graph of size M×M×N. The background definition of this graph is the operation condition information represented in grayscale levels and is estimated as follows:

[0026] o ij = k p (2)

[0027] where Oij is the gray value of the background at the coordinate (i, j), k is the discretized rotational speed or load level, p = 1 when k is the rotational speed level, and p = 2 when k is the load level.

[0028] Furthermore, step 3 specifically includes the following steps:

[0029] Step 3.1: Initialize according to the input information and task, and then use the dynamic training strategy to train the fused information image;

[0030] Step 3.2: Compress the features through the global average pooling (GAP) layer, and obtain the channel statistical vector through formula calculation;

[0031] The fault features are compressed through the global average pooling (GAP) layer, x i is the convolution map obtained from the i-th convolution channel, c is the number of convolution kernels. The size of the pooling filter is H×W, where H and W are the width and height of the convolution map x c respectively. Obtain the channel statistical vector y c :

[0032]

[0033] Step 3.3: Recalibrate the feature map obtained in step 3.2, and obtain the recalibrated feature map through formula calculation;

[0034] The channel recalibration weight vector y′ is defined as:

[0035] y′ = σ(F″(δ(F′(y))

[0036] where δ is the rectified linear unit ReLU activation function, σ is the Sigmoid function, F′ and F″ are convolution operations, the convolution kernel size is 1×1, the output vector is regarded as the channel-level dependency relationship, and its size is 1×1×c, which is used to indicate the importance of each convolution map channel. The recalibrated feature map M is obtained as follows:

[0037] M = x·y′ = [x1y′1, x2y′2, …, x c y′ c (5)

[0038] X is the combination of convolution channels, X = [x1, x2, …, x c , where x i is the convolution map obtained from the i-th convolution channel, c is the number of convolution kernels, and y′ is the channel recalibration weight vector.

[0039] Step 3.4: For the long-range dependence relationship exhibited by the features, calculate the feature interaction between any two positions through non-local operations to capture the long-range dependence;

[0040] The long-range dependence is captured based on appearance similarity and estimated by calculating the feature interaction between any two positions. The non-local operation is defined as:

[0041]

[0042] where i is the feature index of the output position for which the response is to be calculated, j is the index enumerating all possible positions, f is the function for calculating the relationship between y i and all y j , and C(y) is the normalization factor.

[0043] Step 3.5: Obtain the attention weight vector of the non-local channel and the output of the non-local attention module through formula calculation.

[0044] Furthermore, Step 3.5 is specifically as follows: In the non-local attention (NLAM) module, the attention weight φ is obtained through F′ and the softmax function. The attention feature map is obtained by performing matrix multiplication of the weight φ and the input Y of the NLAM module. The transformed features are aggregated to the input at each position through element-wise addition; the input feature map Y is flattened using a 1×1 convolutional kernel. The attention weight vector φ of the non-local channel is defined as:

[0045] φ = λ(F′(Y)) (7)

[0046] where λ is the softmax function, F′ is the convolutional operation with a kernel size of 1×1. The enhanced non-local fault-related features are obtained as follows:

[0047] Z = Y + F″(Y · φ) (8)

[0048] where Z is the output of the attention module, F″ is the convolutional operation with a kernel size of 1×1, and the size of the obtained attention feature map is converted to the size of the NLAM module input Y.

[0049] Furthermore, Step 4 is specifically as follows:

[0050] Step 4.1: Obtain the output value of the residual learning unit through formula calculation; when constructing the residual network, a shortcut connection is adopted to directly connect the input and output of the stacked layers. Under this structure, the integrated rectified linear unit (RLU) is defined as:

[0051] y = F(x, W i ) + x (9)

[0052] Among them, x and y are the input and output of the residual structure respectively, F is the residual function representing the residual mapping to be learned, and F(x, W i ) + x is obtained by shortcut connection and element-wise addition;

[0053] Step 4.2: Continuously train the neural network based on the Adam optimization algorithm, introduce a learning rate scheduling based on exponential decay to achieve dynamic adjustment of the learning rate; the dynamic learning rate is defined as:

[0054]

[0055] where β0 is the initial learning rate, t is the number of iterations, is a hyperparameter, and β0 and are obtained from empirical estimates of the initial training performance;

[0056] Step 4.3: Judge whether the network trained in Step 4.2 converges. If it does, continue to execute; otherwise, return to the training module.

[0057] The present invention also discloses a fault diagnosis system for a new type of oil-electric hybrid offshore operation platform, including an inverter, a generator, a motor, a thruster, a battery pack, a host computer, corresponding sensors, and a communication network. The fault diagnosis system obtains signal data such as rotational speed and frequency through sensors connected to the motor. The engine is connected to both the generator and the thruster at the same time. According to different operation modes of the engine, it can drive the generator to supply power to the battery pack or drive the thruster to provide power for the system. The battery pack supplies power to the motor through a power inverter, and the motor drives the thruster to provide power for the system. Each sensor transmits the operation signal data of devices such as the motor, the thruster, and the battery pack to the host computer through a signal bus, and the staff can monitor the operation status of the hybrid operation platform in real time through the host computer.

[0058] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of the method of the present invention.

[0059] The present invention also discloses a computer-readable storage medium, on which a computer program / instructions is stored, and when the computer program / instructions is executed by a processor, the steps of the method of the present invention are implemented.

[0060] The present invention also discloses a computer program product, including a computer program / instructions, and when the computer program / instructions is executed by a processor, the steps of the method of the present invention are implemented.

[0061] Advantages: Compared with the prior art, the present invention has the following remarkable advantages: The present invention proposes a fault diagnosis system and method for a new type of oil-electric hybrid offshore operation platform. This method is based on a deep residual convolutional neural network with enhanced discriminative feature learning. By introducing a residual learning unit to form a residual network, it can avoid the problem of low accuracy in traditional convolutional neural networks (CNNs) and has powerful learning and analysis capabilities. Through multi-layer non-linear network training, the potential features of the samples can be learned, thereby improving the classification or prediction ability. Therefore, it has advantages in system-level complex fault diagnosis. This system continuously trains the neural network based on the Adam optimization algorithm, introduces a learning rate scheduling based on exponential decay, improves the dynamic adjustment of the learning rate, accelerates the balanced convergence speed, and enhances the robustness of the system to noise and unbalanced data at the same time. The adaptive gradient normalization of Adam (such as the RMSProp feature) can suppress the gradient influence of noise samples, and the global learning rate adjustment based on exponential decay further reduces the interference of noise in the later stage of training.

[0062] In addition, different from single-task fault diagnosis methods, the oil-electric hybrid system proposed by the present invention includes a multi-task convolutional neural network (MTCNN), which can process multiple tasks simultaneously, allows sharing of certain layers between different tasks, and retains specific layers for each task, effectively improving the efficiency and accuracy of fault diagnosis for the oil-electric hybrid system. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of the hardware connection of the fault diagnosis system based on the multi-task convolutional neural network;

[0064] Figure 2 Schematic diagram of the software design of the fault diagnosis system based on the multi-task convolutional neural network; [[ID=1s5]]

[0065] Figure 3 Flowchart of the training of the fault diagnosis system for the hybrid operation platform;

[0066] Figure 4 Flowchart of data conversion and processing;

[0067] Figure 5 Schematic diagram of the structure of the channel attention module (CAM);

[0068] Figure 6 Schematic diagram of the structure of the non-local attention module (NLAM);

[0069] Figure 7 Schematic diagram of the residual learning unit;

[0070] Figure 8 Method framework diagram of the fault diagnosis system for the hybrid system. DETAILED DESCRIPTION OF THE INVENTION

[0071] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings.

[0072] As Figure 1 shown is a schematic diagram of the hardware connection of a fault diagnosis system based on a multi-task convolutional neural network. The hardware circuit of a novel oil-electric hybrid offshore operation platform fault diagnosis system proposed by the present invention mainly consists of a generator, a motor, a thruster, a battery pack, a host computer, etc. The system obtains signal data such as rotational speed and frequency through sensors connected to the motor. The engine is connected to both the generator and the thruster at the same time. According to different operation modes of the engine, it can drive the generator to supply power to the battery pack or drive the thruster to provide power for the system. The battery pack supplies power to the motor through a power inverter, and the motor drives the thruster to provide power for the system. Each sensor transmits the operation signal data of devices such as the motor, the thruster, and the battery pack to the host computer through a signal bus. Staff can monitor the operation status of the system in real time through the host computer, discover faults in time and take measures.

[0073] As Figure 2 shown is a schematic diagram of the software design of a fault diagnosis system based on a multi-task convolutional neural network. The software design circuit of a novel oil-electric hybrid offshore operation platform fault diagnosis system mainly consists of a data preprocessing module, a judgment feature attention module, a deep residual convolutional neural network module, and a fault diagnosis model. The main function of data preprocessing is to construct an information image, convert the 1D time series data obtained by the sensor into a 2D grayscale image, and combine it with the information map formed by domain knowledge to construct an information image, which is used as the input of the judgment feature attention module after convolution operation. The judgment feature attention module consists of a channel attention module and a non-local attention module. In terms of the feature learning network for fault diagnosis based on a multi-task convolutional neural network (MTCNN), it mainly consists of several alternating convolutional layers, batch normalization layers, and rectified linear unit (ReLU) activation layers. The present invention introduces a residual learning unit (RLUs) by connecting two shortcut connections, and also introduces a dropout layer to avoid overfitting problems. For the fault diagnosis model of MTCNN, in order to achieve multi-task classification through a single network, multiple classifiers are set behind the fully connected layer. S1 and S2 in the figure are samples of two classifiers. For each task, there is an independent fully connected layer and a classifier, and the independent part can be trained separately based on the features of the shared structure.

[0074] Figure 3It is a training flowchart of a fault diagnosis system for a hybrid operation platform. The data preprocessing module includes the generation of 2D images and the establishment of information images. The 1D time series data directly obtained by the system through various sensors cannot be recognized and used by the system. Therefore, the data needs to be dimensionally transformed first. The 1D time series data is converted into 2D grayscale images, and then combined with expert knowledge to construct information images, which are used as the input of the proposed multi-task CNN network model. The specific implementation steps are as follows:

[0075] (1) Signal acquisition

[0076] First, collect the operation signal data of motors, thrusters, and other power equipment through the sensors on each device, and make a preliminary judgment on the obtained signal data to confirm that it is the signal data of the corresponding device. Then, estimate the sample length according to the sampling frequency and the motor rotation speed. The principle is to ensure that all sampling data is included in one rotation. According to the calculated sample length, divide the processed data into a training set and a test set.

[0077] (2) Data preprocessing

[0078] Convert the monitoring data segments from different sensors into two-dimensional (2D) grayscale images, and transform the domain knowledge into information images. Combine the two-dimensional (2D) grayscale images with the information images to construct a fused information image, which is used as the input of the proposed multi-task CNN network model.

[0079] Data conversion and combination, as Figure 4 shown in the data conversion and processing flowchart. The original system monitoring data is 1D time series data. To meet the input requirements of the network, the primary task is to convert the 1D data into 2D images. During the conversion process, the original data is randomly divided into a training set and a test set with a length of M 2 , where M is the length and width of the grayscale image after conversion. Estimate the sample length according to the sampling rate and the motor rotation speed. The principle is to ensure that all sampling data is included in one rotation.

[0080] For each data sample segment, apply a sliding window to the original data. The grayscale pixels of the first sample segment can be calculated according to formula (1) and used as the first row of the grayscale image. Based on this conversion procedure, the segments of the original data can be converted into grayscale images with a size of M×M, and the pixel intensity values of the images can be obtained by normalizing the original data to the range of (0, 255). Define I(i,j) as the original time series data, and the pixel intensity value G(i,j) of the image can be obtained by formula (1):

[0081]

[0082] where \(i = 1,\ldots,M\); \(j = 1,\ldots,M\); \(M\) is the coordinate of the image, \(I\) max and \(I\) min represent the maximum and minimum values of the original data \(I\) respectively, and \(round(\cdot)\) is the rounding function.

[0083] When the motor is running, the speed and load have a significant impact on the signal, amplitude, frequency, and noise of the monitoring data. These effects pose challenges to fault diagnosis, especially for systems operating under varying conditions and loads. Therefore, it is very necessary to integrate information on operating conditions during the diagnosis process. Deep learning, based on its powerful feature learning and fusion capabilities, provides a solution for integrating complex heterogeneous data during the fault diagnosis process.

[0084] In the present invention, an operating information graph is constructed with the aim of integrating domain knowledge and two-dimensional images of information on operating conditions into the network input. To keep the data format of the operating information graph consistent with the format of the acquired data images, different operating conditions are represented by different gray levels. Using data from multiple sensors can improve the diagnostic performance. In the present invention, for each fault mode, the original data from \(N\) sensors is used. First, an empty information graph of size \(M\times M\times N\) is constructed, and the background definition of this graph is the operating condition information represented by gray levels, estimated as follows:

[0085] o ij = k p (2)

[0086] where \(O\) ij is the gray value of the background at coordinate \((i,j)\), \(k\) is the discretized speed or load level, \(p = 1\) when \(k\) is the speed level, and \(p = 2\) when \(k\) is the load level. To distinguish the operating conditions represented by different backgrounds, a reasonable choice of the value of \(k\) can avoid the overlap of different gray backgrounds.

[0087] (3) Construct a discriminative feature attention mechanism based on the channel attention module and the non-local attention module

[0088] The judgment feature attention module, as Figure 2 shown, the judgment feature attention module consists of a channel attention module and a non-local attention module. Feature information is extracted through them for training the constructed deep residual convolutional neural network and fault diagnosis of the new oil-electric hybrid offshore operation platform. Figure 5 Shown is the structural schematic diagram of the channel attention module (CAM), Figure 6The following figure shows the schematic structure of the Non-Local Attention Module (NLAM). The specific implementation steps for constructing a discriminative feature attention mechanism based on the Channel Attention Module (CAM) and the Non-Local Attention Module (NLAM) are as follows: 1. Construction of the Channel Attention Module (CAM). The Channel Attention Module (CAM) can identify fault features from the input image or the feature map from the previous layer through data processing with different convolutional kernels at different scales. The new oil-electric hybrid offshore operation platform provides rich feature information. However, due to the existence of interference signals, some information features are irrelevant to the fault itself and may even convey incorrect information. To solve this problem, the present invention introduces CAM to enhance the feature extraction ability of the fault diagnosis system. Its main idea is to distinguish different features in different channels under convolutional operations and improve the sensitivity of the network by explicitly modeling the importance of channels. In this Figure 5 where X represents the combination of convolutional channels, denoted as X = [x1, x2, ···, x c , where x i is the convolutional map obtained through the i-th convolutional channel, and c is the number of convolutional kernels. It should be noted that the convolutional map X is obtained by applying a convolutional operation to the input image, which is a combination of system monitoring data, operating conditions, and domain knowledge. The convolutional operation is used to extract features at different scales from the input information image.

[0089] The fault features are compressed through the Global Average Pooling (GAP) layer. The size of the pooling filter is H × W, where H and W are the width and height of the convolutional map x c respectively. Thus, the channel statistical vector y c can be obtained as follows:

[0090]

[0091] The channel recalibration weight vector y′ is defined as:

[0092] y′ = σ(F″(δ(F′(y))) (4)

[0093] where δ is the Rectified Linear Unit (ReLU) activation function, σ is the Sigmoid function, and F′ and F″ are convolutional operations with a convolutional kernel size of 1×1. The output vector can be regarded as the channel-level dependency relationship, with a size of 1×1×c, used to indicate the importance of each convolutional map channel. The recalibrated feature map M can be obtained as follows:

[0094] M = x · y′ = [x1y′1, x2y′2, …, x c y′ c (5)

[0095] X is the combination of convolutional channels, X = [x1, x2, …, xc , where x i is the convolution map obtained from the i-th convolution channel, c is the number of convolution kernels, and y′ is the channel recalibration weight vector.

[0096] In the obtained recalibrated feature map M, discriminative feature information is highlighted by assigning the weight vector y′. In this module, the GAP operation is performed on all convolution channels of X, but it causes the loss of effective information. To retain the complete original information, the present invention introduces a residual connection, improving the effectiveness of the design.

[0097] The construction of the non-local attention module (NLAM), in the non-local attention module (NLAM), each input information map is transformed and constructed from a segment of system data. Some features in the data exhibit long-range dependencies, including intrinsic motor operation signals, pulse excitation signals related to faults, system periodic signals, etc. Since these signals are not short-lived burst signals but motor data collected over a relatively long time range, the features have long-range dependencies, and different fault modes exhibit different long-range dependencies. In the output of the NLAM, long-range dependent features (such as periodic excitation signals caused by faults and certain operation behaviors) are emphasized, and useless information is suppressed. Enhancing the feature learning ability related to long-range dependencies is crucial for improving the fault diagnosis efficiency of the power system. In addition, motor data is usually accompanied by noise, which also has a significant impact on the judgment performance of the fault diagnosis system.

[0098] To capture the long-range dependencies of features and reduce the impact of data noise in deep learning, the present invention adopts and analyzes different non-local modules. These non-local blocks are very effective in capturing useful information and ignoring the influence of noise. The present invention adopts a simplified NLAM to construct the network, enhancing the fault feature learning ability of the fault diagnosis system with a slightly increased computational cost. The NLAM is designated after the CAM, enhancing the network's learning ability for long-range dependent features.

[0099] The core idea of the NLAM is non-local operation, which calculates the weighted average of all sensor signal points in the input information Y of the NLAM module. In the power system diagnosis task, fault-related features should exist in each sample, and non-local operation is used to capture these fault features that not only exist in some samples. Long-range dependencies can be captured based on appearance similarity, estimated by calculating the feature interaction between any two positions. The general non-local operation can be defined as:

[0100]

[0101] where i is the feature index of the output position for which the response is to be calculated, j is the index enumerating all possible positions, and f is used to calculate y iA function of the relationship with all y j where C(y) is a normalization factor.

[0102] In the simplified NLAM, note that the weight φ is obtained through the F′ and softmax functions. The weight φ is multiplied by the input Y of the NLAM module through matrix multiplication to obtain the attention feature map, and the transformed features are aggregated onto the input at each position through element-wise addition. Residual connections can insert the attention module into any network without destroying or losing any original information, thereby improving the efficiency and reliability of network feature learning.

[0103] As Figure 6 shown in the schematic diagram of the non-local attention module (NLAM) structure, where the input feature map Y is flattened through a 1×1 convolutional kernel, and the attention weight vector φ of the non-local channel is defined as:

[0104] φ = λ(F′(Y)) (7)

[0105] where λ is the softmax function and F′ is a convolutional operation with a convolutional kernel size of 1×1. The enhanced non-local fault-related features can be obtained as follows:

[0106] Z = Y + F″(Y·φ) (8)

[0107] where Z is the output of the attention module, F″ is a convolutional operation with a convolutional kernel size of 1×1, and the size of the obtained attention feature map is converted to the size of the NLAM module input Y.

[0108] (4) Introduce a residual learning unit to strengthen the training of the convolutional neural network

[0109] The deep residual convolutional neural network with a residual learning unit can solve the problem of low accuracy in traditional CNNs. Under varying operating conditions and complex working environments, deeper network structures are usually used to obtain better diagnostic results. However, this makes the training of the network very difficult. Existing research shows that when the depth of the network increases to a certain value, the accuracy will decrease, indicating that a deeper network does not necessarily guarantee better performance.

[0110] The fault diagnosis model of the new oil-electric hybrid offshore operation platform power system forms a residual network by introducing a residual learning unit, which can avoid the problem of low accuracy in traditional convolutional neural networks (CNNs). The residual network has a residual connection function that bypasses certain layers in the CNN structure, and this residual connection provides a simple way for parameters to propagate from the output layer to the previous layers. Based on this structure, a deeper network structure can be designed to achieve the goal of diagnosing complex tasks with high accuracy.

[0111] Taking this into consideration, the present invention adopts a residual learning unit (RLU) composed of two convolutional layers, two batch normalization (BN) layers, and one rectified linear unit (ReLU) activation layer, as Figure 7 shown in the schematic diagram of the residual learning unit. When constructing the residual network, it should be noted that a shortcut link is adopted in the structure, directly connecting the input and output of the stacked layers. Through this shortcut path, the saturation problem faced when training a very deep network can be alleviated. Under this structure, the integrated RLU can be defined as:

[0112] y = F(x, W i ) + x (9)

[0113] where x and y are the input and output of the residual structure respectively, F is the residual function, representing the residual mapping to be learned, and F(x, W i ) + x is obtained through the shortcut connection and element-wise addition. Finally, a fault diagnosis framework based on a multi-task convolutional neural network (MTCNN) is constructed for fault diagnosis. Combining the specific descriptions of Figure 1 and Figure 2 above, the present invention proposes a method framework as Figure 8 shown. The first step of this method is the preprocessing and information fusion of data. The 1D data obtained through sensors is converted into a 2D grayscale image, and then an information image is constructed by combining the information map converted from domain knowledge. This image contains a large amount of fault information and serves as the input quantity for the feature learning part. Then comes the feature learning step. A discriminative feature attention mechanism is constructed through a channel attention module (CAM) and a non-local attention module (NLAM) to extract fault feature information. Next, a fault diagnosis network model based on MTCNN is constructed. The model training process is based on the Adam optimization algorithm. To achieve the goals of high accuracy and fast convergence, a learning rate scheduling based on exponential decay is introduced during the training process. By introducing this learning rate scheduling, the learning rate can be dynamically adjusted, gradually decreasing according to a predefined plan. As the training progresses, the learning rate will decrease to a smaller value. The dynamic learning rate can be defined as:

[0114]

[0115] where β0 is the initial learning rate, t is the number of iterations, is a hyperparameter, and β0 and are obtained through empirical estimation of the initial training performance.

[0116] (5) Model testing

[0117] The multi-task convolutional neural network fault diagnosis model obtained through the above steps is tested with the data in the test set. If the accuracy of the output data reaches the standard, the final model is obtained; otherwise, if the accuracy of the output data does not reach the standard, the data is returned to the training module for continued training.

[0118] (6) Apply the obtained model to fault diagnosis

[0119] Input the real-time data of the power system of the new oil-electric hybrid offshore operation platform into the trained multi-task convolutional neural network fault diagnosis model. If a fault is detected, the model outputs the fault type, fault location, and other information related to the fault, facilitating the staff to promptly discover the fault and take effective measures.

Claims

1. A fault diagnosis method for a new type of oil-electric hybrid offshore operation platform, characterized in that, It includes the following steps: Step 1: Collect power system signals through sensors, divide the time series data of the signals into multiple segments according to the sampling rate and motor speed, and then divide the data segments into a training set and a test set; Step 2: Convert the one-dimensional time series data of the data segments into two-dimensional grayscale images, and convert domain knowledge into information graphs. Combine the two-dimensional grayscale images and the information graphs to construct a fused information image; Step 3: Take the fused information image obtained in Step 2 as the input, construct a multi-task convolutional neural network MTCNN model based on the discriminative feature attention channel attention module CAM and the non-local attention module NLAM, and use multiple classifiers for multi-task diagnosis; Step 4: Introduce a residual learning unit, and train the multi-task convolutional neural network MTCNN model with the training set; Step 5: Test the multi-task convolutional neural network MTCNN model obtained in Step 4 with the data of the test set. If the accuracy of the output data reaches the standard, go to Step 6; otherwise, if the accuracy of the output data does not reach the standard, return to Step 3; Step 6: Apply the multi-task convolutional neural network MTCNN model to a new oil-electric hybrid offshore operation platform for fault diagnosis.

2. The fault diagnosis method of a new type of oil-electric hybrid offshore operation platform according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1: Collect the operation signal data of the motor, propeller and other power equipment through the sensors on each device; Step 1.2: Make a preliminary judgment on the obtained signal data and eliminate the error signal data; Step 1.3: Estimate the sample length according to the sampling frequency and motor rotation speed; Step 1.4: Divide the processed data into a training set and a test set according to the sample length calculated in Step 1.

3.

3. A fault diagnosis method for a new type of oil-electric hybrid offshore operation platform according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Convert the one-dimensional time series data obtained by the sensor into a two-dimensional image. Based on this conversion program, the segments of the original data are converted into grayscale images of standard size; For each data sample segment, apply a sliding window to the original data. The grayscale pixels of the first sample segment are calculated according to formula (1) and used as the first row of the grayscale image. Based on this conversion program, the segments of the original data are converted into grayscale images of size M×M. The pixel intensity values of the image are obtained by normalizing the original data to the range of (0, 255). Define I(i,j) as the original time series data, and the pixel intensity value G(i,j) of the image is obtained by formula (1): where \(i = 1,\ldots,M\); \(j = 1,\ldots,M\); \(M\) is the coordinate of the image, \(I\) max and \(I\) min represent the maximum and minimum values of the original data \(I\) respectively, and \(round(\cdot)\) is the rounding function; Step 2.2: Construct an information graph from the grayscale image obtained in Step 2.1; Step 2.3: Construct an operating information graph, which represents the operating condition information in gray levels and is obtained by calculation using a formula; For each fault mode, use the original data from N sensors; construct an empty information graph of size M×M×N, the background definition of which is the operating condition information represented in gray levels and is estimated as follows: o ij = k p (2) where O ij is the gray value of the background at the coordinate (i, j), k is the discretized rotational speed or load level, p = 1 when k is the rotational speed level, and p = 2 when k is the load level.

4. The fault diagnosis method of a new oil-electric hybrid offshore operation platform according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Initialize according to the input information and tasks, and then use a dynamic training strategy to train the fused information image; Step 3.2: Compress the features through the global average pooling (GAP) layer, and obtain the channel statistical vector through formula calculation; The fault features are compressed through the global average pooling (GAP) layer, and x i is the convolutional map obtained from the i-th convolutional channel, c is the number of convolutional kernels, and the size of the pooling filter is H×W, where H and W are the width and height of the convolutional map x c respectively, to obtain the channel statistical vector y c : Step 3.3: Recalibrate the feature map obtained in Step 3.2, and obtain the recalibrated feature map through formula calculation; The channel recalibration weight vector y′ is defined as: y′ = σ(F″(δ(F′(y)))) (4) where δ is the rectified linear unit (ReLU) activation function, σ is the Sigmoid function, F′ and F″ are convolutional operations with a kernel size of 1×1. The output vector is regarded as the channel-level dependency relationship, and its size is 1×1×c, which is used to indicate the importance of each convolutional map channel. The recalibrated feature map M is obtained in the following way: M = x·y′ = [x1y′1, x2y′2, …, x c y′ c (5) X is a combination of convolutional channels, X = [x1, x2, …, x c , where x i is the convolutional map obtained from the i-th convolutional channel, c is the number of convolutional kernels, and y′ is the channel recalibration weight vector; Step 3.5: For the long-range dependency relationship exhibited by the features, calculate the feature interaction between any two positions through non-local operations to capture the long-range dependency; The long-range dependency is captured based on appearance similarity, estimated by calculating the feature interaction between any two positions. The non-local operation is defined as: where i is the feature index of the output position for which the response is to be calculated, j is the index that enumerates all possible positions, f is the function for calculating the relationship between y i and all y j and C(y) is the normalization factor; Step 3.5: Obtain the attention weight vector of the non-local channel and the output of the non-local attention module through formula calculation.

5. The fault diagnosis method of a new oil-electric hybrid offshore operation platform according to claim 4, characterized in that, Specifically, in the non-local attention (NLAM) module, the attention weight φ is obtained through the convolutional operation F′ and the softmax function. The weight φ is multiplied with the input Y of the NLAM module through matrix multiplication to obtain the attention feature map. The transformed features are aggregated to the input at each position through element-wise addition; the input feature map Y is flattened through a 1×1 convolutional kernel. The attention weight vector φ of the non-local channel is defined as: φ = λ(F′(Y)) (7) where λ is the softmax function, F′ is the convolutional operation with a kernel size of 1×1. The enhanced non-local fault-related features are obtained in the following way: Z = Y + F″(Y·φ) (8) where Z is the output of the attention module, F″ is the convolutional operation with a kernel size of 1×1, and the size of the obtained attention feature map is converted to the size of the NLAM module input Y.

6. The fault diagnosis method of a new type of oil-electric hybrid offshore operation platform according to claim 4, characterized in that, Step 4 is specifically as follows: Step 4.1: Obtain the output value of the residual learning unit through formula calculation; when constructing the residual network, a shortcut connection is adopted to directly connect the input and output of the stacked layers. Under this structure, the integrated RLU is defined as: y = F(x, W i ) + x (9) where x and y are the input and output of the residual structure respectively, F is the residual function representing the residual mapping to be learned, and F(x, W i ) + x is obtained by shortcut connection and element-wise addition; Step 4.2: Continuously train the neural network based on the Adam optimization algorithm, introduce a learning rate schedule based on exponential decay to achieve dynamic adjustment of the learning rate; the dynamic learning rate is defined as: where β0 is the initial learning rate, t is the number of iterations, is a hyperparameter, β0 and are obtained from an empirical estimate of the initial training performance; Step 4.3: Judge whether the network trained in Step 4.2 converges. If it converges, continue to execute; otherwise, return to the training module.

7. A fault diagnosis system for a new type of oil-electric hybrid offshore operation platform, which is used to implement the method described in claim 1, characterized in that, It includes an inverter, a generator, a motor, a thruster, a battery pack, a host computer, corresponding sensors, and a communication network; the fault diagnosis system obtains signal data of rotational speed and frequency through sensors connected to the motor. The engine is connected to both the generator and the thruster at the same time. According to different operation modes of the engine, it drives the generator to supply power to the battery pack and also drives the thruster to provide power for the system; the battery pack supplies power to the motor through a power inverter, and the motor drives the thruster to provide power for the system; each sensor transmits the operation signal data of the motor, thruster, and battery pack devices to the host computer through a signal bus, and the staff monitors the operation status of the hybrid operation platform in real time through the host computer.

8. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method described in claim 1.

9. A computer-readable storage medium having computer programs / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method described in claim 1 are implemented.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method described in claim 1 are implemented.

Images