Ship motor rolling bearing fault diagnosis method based on combination of adaptive multi-branch bidirectional long-short-term memory network and knowledge graph

Abstract

The invention discloses a ship motor rolling bearing fault diagnosis method based on combination of a self-adaptive multi-branch bidirectional long-short-term memory network and a knowledge graph. Comprising the steps of 1, obtaining data, and performing rule extraction and entity construction; 2, constructing a self-adaptive multi-branch bidirectional long-short-term memory network model; 3, training the model by using the collected vibration acceleration signal of the rolling bearing; 4, storing the model after the result of the training model meets the requirement, classifying the rolling bearing, outputting the result classified by the model, and constructing a triple; and step 5, constructing a knowledge graph. The invention provides a ship motor rolling bearing fault diagnosis method based on combination of a self-adaptive multi-branch bidirectional long-short-term memory network and a knowledge graph, overcomes the defect that mechanism knowledge and vibration data of a ship motor bearing cannot be combined, and provides a novel ship motor bearing diagnosis method. The method can better help to understand the cause and influence of the fault, can give a corresponding solution when the fault occurs, and improves the fault diagnosis efficiency.

Description

Technical Field

[0001] This invention pertains to a method for detecting faults in rolling bearings of marine motors, specifically a method for diagnosing faults in rolling bearings of marine motors based on an adaptive multi-branch bidirectional long short-term memory network combined with a knowledge graph. Background Technology

[0002] Marine motors are core components of ship propulsion and power supply systems, and their stability directly impacts the safety and efficiency of ship navigation. Among these motors, rolling bearings are crucial, playing a vital role in supporting the rotor and transmitting loads. However, the complex operating conditions at sea and the harsh working environment of ship bearings result in a high probability of rolling bearing failure. When a rolling bearing fails, it not only affects the performance of the motor itself but can also severely impact the operation of the entire ship, potentially leading to even more serious consequences. Therefore, improving the efficiency and accuracy of rolling bearing fault diagnosis for marine motors is extremely important for ensuring normal ship operation, reducing maintenance costs, and preventing risks.

[0003] In 2012, Google proposed the concept of knowledge graphs. Knowledge graphs are used to store and represent large-scale semantic information, centralizing vast amounts of semantic data and providing search functionality, making them widely used in search engine optimization. Essentially, a knowledge graph is a complex network structure composed of triples. Each triple contains two entities connected by relationships. In the knowledge graph, entities represent nodes, and relationships are the edges connecting them, revealing the relationships between different entities. This approach allows for efficient management of large amounts of information.

[0004] Constructing a knowledge graph for bearing faults requires the collection of vibration signals as a crucial component. Vibration signals effectively reflect the bearing's operating state, indicating whether it is faulty or functioning correctly. With the rapid development of artificial intelligence, increasingly advanced technologies are being applied to bearing fault diagnosis. These algorithms extract sensitive features from large amounts of complex data to identify the corresponding bearing condition and provide accurate diagnostic results.

[0005] Combining artificial intelligence algorithms and knowledge graphs for fault diagnosis of marine motor rolling bearings is an innovative and efficient technical solution. It not only fully utilizes vibration data but also incorporates current bearing information, improving diagnostic accuracy and providing interpretable results. The knowledge graph contains professional knowledge such as bearing failure modes, failure types, and maintenance history, which helps in understanding the causes and effects of faults and provides corresponding solutions, thus improving the efficiency of fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to address the above-mentioned problems by providing a fault diagnosis method for marine motor rolling bearings based on an adaptive multi-branch bidirectional long short-term memory network combined with a knowledge graph. The adaptive multi-branch structure processes data in parallel through multiple convolutional branches of different scales, enabling it to simultaneously capture global and detailed feature information and extract deep features from rolling bearing vibration acceleration signal data. The bidirectional long short-term memory network structure is a modification of LSTM, which can better combine forward and backward propagation hidden layers, simultaneously identifying information from both directions and improving the ability to extract time-series features. The knowledge graph can combine bearing-related information, manage large amounts of information, and provide corresponding auxiliary decision-making in the event of a fault. Combining these three elements solves the problem of the adaptive multi-branch structure lacking temporal correlation, the difficulty of BiLSTM in extracting deep features, and the problem of large and complex knowledge sets, effectively improving the accuracy and stability of bearing fault diagnosis.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for diagnosing rolling bearing faults in marine motors based on an adaptive multi-branch bidirectional long short-term memory network combined with a knowledge graph, comprising the following steps: Step 1: Use an accelerometer to obtain the vibration acceleration signal of the rolling bearing, and then perform rule extraction and entity construction based on the structural data of the rolling bearing and the collected vibration acceleration signal; Furthermore, because rolling bearing data may exhibit differences in magnitude and units, directly inputting the data into the model would negatively impact training performance. Therefore, to ensure the accuracy and reliability of the results and improve model performance, data normalization is necessary. The normalization method used is standardization. The standardization formula is as follows: in It is the raw data; It is the mean of the feature data; This is the standard deviation of the feature data. Standardized data follows a standard normal distribution, which helps subsequent algorithms converge better.

[0008] After normalization, the data is re-encoded into One-Hot encoding, which maps each category to a unique binary number, eliminating hidden relationships between categories. This makes the data easier to process and train in the model.

[0009] The acquired rolling bearing acceleration signal is augmented using a sliding time window. The signal interval length is defined as *m*, which is the length of the segment extracted from the original data each time. The sliding step size is defined as *n*, which is the distance the window moves backward from its starting position after each slide. The processing steps begin with the first data point in the dataset and slide *n* times until the end of the dataset is reached.

[0010] Furthermore, entity construction involves building relevant entities based on the relevant mechanistic knowledge of the data. This relevant mechanistic knowledge includes the type of motor, the motor load, the motor speed, the fault type, the fault location, etc. Step 2: Construct an adaptive multi-branch bidirectional long short-term memory network model. This adaptive multi-branch bidirectional long short-term memory network model includes the following components: an Inception module and a BiLSTM module. Step 2 includes the following specific contents: Gating mechanisms, such as input gates, forget gates, and output gates, are already present in traditional LSTMs to control the flow of information. In the adaptive multi-branch bidirectional long short-term memory network model described above, similar gating units are added to each branch. These gating units determine which branches' information should be utilized more. For some parts where global information is more important, branches with larger kernels will have a greater weight, and vice versa. For other parts where more detail is needed, branches with smaller kernels will have a greater weight. Assuming we have B branches, and the output of each branch is... Where b = 1, 2, ..., B. The branch selection gate is calculated using the following formula: in, It is the weight matrix of branch b; It is the bias vector of branch b; It is an activation function, usually the sigmoid function, used to generate values ​​between 0 and 1; X t This is the input for the current time step; h t-1 This is the hidden state of the previous time step. The final hidden state h t This can be obtained by weighted summation of the outputs from all branches: The Inception module is defined to process data in parallel using multiple convolutional kernels of different scales, extracting multi-scale feature information. Specifically, the Inception module contains four branches: a 1x1 convolutional branch, which uses 1x1 convolutional kernels for dimensionality reduction, reducing computation while preserving important feature information.

[0011] 3x3 convolution branch: First, dimensionality is reduced by 1x1 convolution, and then local features are extracted by 1x3 convolution.

[0012] 5x5 convolution branch: First, dimensionality is reduced by 1x1 convolution, and then a larger range of features is extracted by 1x5 convolution.

[0013] Pooling branch: Extract global features through max pooling operation, and then adjust the number of channels through 1x1 convolution.

[0014] This involves ensuring that the output dimension of each branch is the same, and finally concatenating them along the channel dimension to form the final output.

[0015] A gating unit is added after the output of each Inception branch. These gating units determine the importance of each branch, thus determining its contribution to the final output. Furthermore, the importance of each branch is determined using a Sigmoid activation function to generate weights between 0 and 1, and these weights are multiplied by the corresponding branch output.

[0016] Define the BiLSTM module. This includes the following parts: Initialization function, Inception module, pooling layer and activation function, BiLSTM layer and fully connected layer.

[0017] Further initialization functions include the number of channels in the input data, the size of the hidden layer dimension of the BiLSTM, the number of layers in the BiLSTM, the number of categories in the final classification, and the Dropout layer.

[0018] The further Inception module uses a three-word Inception module to extract features at different levels: The first Inception module takes the same number of input channels as the initialization function and outputs 64 channels; the second Inception module takes 64 input channels and outputs 128 channels; the third Inception module takes 128 input channels and outputs 256 channels. Each module extracts multi-scale features and further enhances their expressiveness through batch normalization and activation functions.

[0019] Further pooling layers and activation functions are used. Max pooling layers are employed to reduce the feature map size without losing the most important feature information. The feature map width is halved after each pooling. The ReLU activation function is used to introduce non-linearity and enhance the model's expressive power.

[0020] Furthermore, a global gating unit, used to control the output of the entire feature map, is implemented by learning an additional parameter matrix that matches the input size of the LSTM.

[0021] Furthermore, the BiLSTM layer. The input size of BiLSTM is equal to the number of output channels of the convolutional layer multiplied by the width of the pooled feature map. Simultaneously using bidirectional LSTM allows for the capture of temporal dependencies from both the forward and backward directions.

[0022] Further fully connected layers. These consist of several fully connected layers with different input and output dimensions. To prevent overfitting during training, Dropout is introduced into these fully connected layers. For a given layer of neurons, Dropout randomly discards some neurons according to the Dropout probability, while ensuring that the number of neurons in the input and output layers remains constant. Then, the parameters are updated according to a corresponding method, and in subsequent iterations, some neurons are randomly discarded again using the same method until training ends.

[0023] Furthermore, the Dropout probability is 0.5.

[0024] Furthermore, in order to input the data after convolution and pooling operations into the BiLSTM layer, a flattening layer needs to be introduced.

[0025] Step 3: Divide the collected rolling bearing vibration acceleration signal into a training set and a test set according to the ratio. Train the adaptive multi-branch bidirectional long short-term memory network model on the divided training set, carry out multiple rounds of training, and find the optimal model hyperparameters. Furthermore, in step 3, the rolling bearing data is divided into a training set and a test set according to a certain ratio, ensuring that the test set accounts for 30% of the total dataset. The training set is used to learn the adaptive multi-branch bidirectional long short-term memory network model. After the model is learned, the test set is used to observe the effect, and the test set is used to verify the final model.

[0026] Furthermore, the loss function is calculated during training. This combines two steps: converting the model output into a probability distribution and calculating the difference between the predicted probability and the true label. The specific formula for converting the model output into a probability distribution is as follows: Here, it is assumed that the output of the model is an unnormalized score vector, denoted as z, where z i p represents the score of the i-th class; C is the number of classes; p i It is the probability of class i; EXP(z) i ) is z i The exponential function is used. Furthermore, the negative log-likelihood loss is calculated to measure the difference between the probability distribution predicted by the model and the true label.

[0027] Step 4: After the training model meets the requirements, save the model, classify the rolling bearings, and output the classification results. Furthermore, using the model to classify the data is the extraction of relations in knowledge graph construction. Based on the hierarchical results in step 1, the data is classified, relations are extracted, and triples are obtained.

[0028] Step 5: After constructing triples based on the relationships and entities obtained from the classification, use Python to import the constructed triples into the Neo4j database to build a bearing knowledge graph and realize fault diagnosis auxiliary decision-making.

[0029] Furthermore, knowledge graph construction methods include top-down, bottom-up, and a combination of both. This knowledge graph construction adopts a combined approach. First, by analyzing the general mechanism of ship bearings, the various entities and their relationships are determined based on the mechanism, constructing the first schema layer. Then, based on relevant data such as fault types, motor speeds, and fault locations, the initial schema layer of the knowledge graph is designed using a top-down approach. Next, guided by the schema layer, a bottom-up approach is used, based on the construction of mechanistic entities and an adaptive multi-branch bidirectional long short-term memory network model, to extract relations and form high-quality knowledge representations. Finally, the obtained relations and entities are added as attribute nodes to the schema layer to ensure its updates, thus completing the construction of the knowledge graph.

[0030] Furthermore, a knowledge graph is constructed using triples. A triple is a representation of a knowledge graph. A triple typically contains extracted entities and relations, usually two entities and a pair of relations (two nodes and one relation). The previously extracted entity nodes and relations are imported into the Neo4j database using Python to complete the construction of the rolling bearing knowledge graph.

[0031] Furthermore, the constructed knowledge graph can display detailed information about the faulty node and clearly show all related nodes under the faulty node.

[0032] Furthermore, in step 5, knowledge graphs can be used to assist in fault diagnosis decision-making. After constructing the knowledge graph, Cypher queries can be used to query nodes, retrieving the required node and all nodes connected to it. Simultaneously, knowledge graphs can be used to perform corresponding operations, such as knowledge-based question answering. Given existing information, the knowledge graph will return a related entity, aiding in decision-making and improving the accuracy and efficiency of fault diagnosis.

[0033] The beneficial effects of this invention are as follows: This invention proposes a fault diagnosis method for rolling bearings of marine motors based on an adaptive multi-branch bidirectional long short-term memory network combined with a knowledge graph. It makes full use of the bearing vibration signal and organically combines the bearing vibration signal with the corresponding knowledge rules to complete the construction of the bearing knowledge graph. This method differs from previous bearing diagnosis methods, keeps pace with the times, and provides a new approach to bearing diagnosis. Attached Figure Description

[0034] Figure 1 This is the overall framework and flowchart for bearing fault diagnosis in this invention; Figure 2 This is the adaptive multi-branch bidirectional long short-term memory network in this invention; Figure 3 This is one of the first-level training loss and training accuracy in an embodiment of the present invention; Figure 4 This is the second of the first-level training loss and training accuracy in one embodiment of the present invention; Figure 5 This is one of the second-level training loss and training accuracy in an embodiment of the present invention; Figure 6 This is the second level of training loss and training accuracy in one embodiment of the present invention; Figure 7 This is one of the third-level training loss and training accuracy in an embodiment of the present invention; Figure 8 This is the second of the third-level training loss and training accuracy in one embodiment of the present invention; Figure 9 This is one of the fourth-level training loss and training accuracy in an embodiment of the present invention; Figure 10 This is the second of the fourth-level training loss and training accuracy in one embodiment of the present invention; Figure 11 This is one of the fifth-level training loss and training accuracy in an embodiment of the present invention; Figure 12 This is the second of the fifth-level training loss and training accuracy in one embodiment of the present invention; Figure 13 This is a knowledge graph constructed according to an embodiment of the present invention. Detailed Implementation Plan like Figure 1 This is the overall framework and flowchart for bearing fault diagnosis in this invention. The method includes the following steps: Obtain the data of the rolling bearing, then construct the entity based on this data. After construction, input the data into the model, train it using the training dataset, and check if the results meet the requirements. When the results meet the requirements, save the model, output the classification results, construct triples based on the classification results, complete the extraction of knowledge graph relationships, and then complete the construction of the knowledge graph based on the above results.

[0035] The following is a detailed explanation of the specific implementation of each step: Step 1: Use an accelerometer to obtain the vibration acceleration signal of the rolling bearing, and then perform rule extraction and entity construction based on the structural data of the rolling bearing and the collected vibration acceleration signal; Furthermore, because rolling bearing data may exhibit differences in magnitude and units, directly inputting the data into the model would negatively impact training performance. Therefore, to ensure the accuracy and reliability of the results and improve model performance, data normalization is necessary. The normalization method used is standardization. The standardization formula is as follows: Where x is the original data; It is the mean of the feature data; This is the standard deviation of the feature data. Standardized data follows a standard normal distribution, which helps the subsequent algorithm converge better. The diagnostic method in this embodiment is based on bearing vibration data. It uses the publicly available bearing dataset from Case Western Reserve University (CWRU). This dataset contains various information, such as the motor load and speed at the time of data collection, the type of fault, the location of the fault, and the magnitude of the fault. Based on this knowledge, entity extraction is performed, mainly including: The data is divided into four categories: normal data, driver-end data with sampling frequencies of 12kHz and 48kHz, and fan-end fault data with a sampling frequency of 12kHz. The data in the above four categories were generated under different load conditions. The loads were 0 HP, corresponding to a speed of 1797 r / min; 1 HP, corresponding to a speed of 1772 r / min; 2 HP, corresponding to a speed of 1750 r / min; and 3 HP, corresponding to a speed of 1730 r / min. The fault sizes at the drive end with a sampling frequency of 12kHz are 0.007 inches, 0.014 inches, 0.021 inches, and 0.028 inches, respectively. The fault sizes at the drive end with a sampling frequency of 48kHz and the fan end with a sampling frequency of 12kHz are 0.007 inches, 0.014 inches, and 0.021 inches, respectively. Bearing failures at the drive end and fan end include rolling element failures, inner ring failures, and outer ring failures. Outer ring failures are further divided into three categories based on their location: failures at the 3, 6, and 12 o'clock positions.

[0036] Node class Label Node description Underlying CWRU_ball bearing Experimental equipment First stage DE48 Drive end fault with sampling frequency of 48 kHz Second stage DE48_PH0 Drive end_load of 0 PH fault with sampling frequency of 48 kHz Third stage DE48_PH0_size0.007 Drive end_load of 0 PH_size of 0.007 inches fault with sampling frequency of 48 kHz Fourth stage DE48_PH0_size0.007_OR Drive end_load of 0 PH_size of 0.007 inches_outer race fault with sampling frequency of 48 kHz Fifth stage DE48_PH0_size0.007_OR_clock3 Drive end_load of 0 PH_size of 0.007 inches_outer race_3 o'clock fault with sampling frequency of 48 kHz Step 2: Construct an adaptive multi-branch bidirectional long short-term memory network model, which includes the following parts: the Inception module and the BiLSTM module.

[0037] Gating mechanisms, such as input gates, forget gates, and output gates, are already present in traditional LSTMs to control the flow of information. In the adaptive multi-branch bidirectional long short-term memory network model described above, similar gating units are added to each branch. These gating units determine which branches' information should be utilized more. For some parts where global information is more important, branches with larger kernels will have a greater weight, and vice versa. For other parts where more detail is needed, branches with smaller kernels will have a greater weight. Assuming we have B branches, and the output of each branch is... Where b = 1, 2, ..., B. The branch selection gate is calculated using the following formula: in, It is the weight matrix of branch b; It is the bias vector of branch b; It is an activation function, usually the Sigmoid function, used to generate values ​​between 0 and 1; This is the input for the current time step; This is the hidden state of the previous time step. The final hidden state. This can be obtained by weighted summation of the outputs from all branches: The Inception module is defined to process data in parallel using multiple convolutional kernels of different scales, extracting multi-scale feature information. Specifically, the Inception module contains four branches: 1x1 convolution branch: Use 1x1 convolution kernels for dimensionality reduction, which reduces computation while preserving important feature information.

[0038] 3x3 convolution branch: First, dimensionality is reduced by 1x1 convolution, and then local features are extracted by 1x3 convolution.

[0039] 5x5 convolution branch: First, dimensionality is reduced by 1x1 convolution, and then a larger range of features is extracted by 1x5 convolution.

[0040] Pooling branch: Extract global features through max pooling operation, and then adjust the number of channels through 1x1 convolution.

[0041] This involves ensuring that the output dimension of each branch is the same, and finally concatenating them along the channel dimension to form the final output.

[0042] A gating unit is added after the output of each Inception branch. These gating units determine the importance of each branch, thus determining its contribution to the final output.

[0043] Furthermore, the importance of each branch is determined by using the Sigmoid activation function to generate weights between 0 and 1, and these weights are then multiplied by the corresponding branch output.

[0044] Define the BiLSTM module. This includes the following parts: Initialization function, Inception module, pooling layer and activation function, BiLSTM layer and fully connected layer.

[0045] Further initialization functions include the number of channels in the input data, the size of the hidden layer dimension of the BiLSTM, the number of layers in the BiLSTM, the number of categories in the final classification, and the Dropout layer.

[0046] The further Inception module uses a three-word Inception module to extract features at different levels: The first Inception module takes the same number of input channels as the initialization function and outputs 64 channels; the second Inception module takes 64 input channels and outputs 128 channels; the third Inception module takes 128 input channels and outputs 256 channels. Each module extracts multi-scale features and further enhances their expressiveness through batch normalization and activation functions.

[0047] Further pooling layers and activation functions are used. Max pooling layers are employed to reduce the feature map size without losing the most important feature information. The feature map width is halved after each pooling. The ReLU activation function is used to introduce non-linearity and enhance the model's expressive power.

[0048] Furthermore, a global gating unit, used to control the output of the entire feature map, is implemented by learning an additional parameter matrix that matches the input size of the LSTM.

[0049] Furthermore, the BiLSTM layer. The input size of BiLSTM is equal to the number of output channels of the convolutional layer multiplied by the width of the pooled feature map. Simultaneously using bidirectional LSTM allows for the capture of temporal dependencies from both the forward and backward directions.

[0050] Further fully connected layers. These consist of several fully connected layers with different input and output dimensions. To prevent overfitting during training, Dropout is introduced into these fully connected layers. For a given layer of neurons, Dropout randomly discards some neurons according to the Dropout probability, while ensuring that the number of neurons in the input and output layers remains constant. Then, the parameters are updated according to a corresponding method, and in subsequent iterations, some neurons are randomly discarded again using the same method until training ends.

[0051] Furthermore, the Dropout probability is 0.5.

[0052] Furthermore, in order to input the data after convolution and pooling operations into the BiLSTM layer, a flattening layer needs to be introduced.

[0053] Step 3: Divide the collected rolling bearing vibration acceleration signal into a training set and a test set according to the ratio. Train the adaptive multi-branch bidirectional long short-term memory network model on the divided training set, carry out multiple rounds of training, and find the optimal model hyperparameters. Based on the first-level classification, the data was divided into four categories. After labeling, the data was further divided according to a 30% weighting for the test set. The model was trained on the training set using gradient descent and backpropagation, and then tested using the test set. The training set was used to learn the adaptive multi-branch bidirectional long short-term memory network model. After training, the performance was observed using the test set, which served as the final validation set for the model.

[0054] Specifically, the Case Western Reserve University bearing dataset was labeled according to the first-level classification categories, totaling 4 categories, and divided into training and test sets.

[0055] Training accuracy Test accuracy First stage 97.74% 95.13% Second stage 94.89% 91.22% Third stage 97.00% 91.31% Fourth stage 95.84% 90.31% Fifth stage 95.12% 93.15% Step 4: After the training model meets the requirements, save the model, classify the rolling bearing data, and output the classification results. Specifically, after the training results meet the requirements, the model is saved and used to classify the rolling bearing data. When using the first-level model, the rolling bearing data is input into the model, and the first-level model automatically divides the data into four categories: DE12, DE48, FE12, and normal. These four categories are different entities in the knowledge graph, and the relationship between these four categories and the entities in the zero-level model is "belong". When using the second-level model, the rolling bearing data is input into the model, and the second-level model automatically divides the data into 16 categories: DE12_PH0, DE12_PH1, DE12_PH2, DE48_PH0, FE12_PH0, etc. These 16 categories are different entities in the knowledge graph, and the relationship between them and the four categories of entities in the previous level is "belong". When using the third-level model, the rolling bearing data is input into the model, and the third-level model will automatically divide the data into 44 categories, namely DE12_PH0_size0.007, DE12_PH1_size0.007, DE12_PH2_size0.007, DE48_PH0_size0.007, FE12_PH0_size0.007, etc. These 44 categories are different entities in the knowledge graph, and their relationship with the 16 categories of entities in the previous level is "belong". When using the fourth-level model, the rolling bearing data is input into the model, and the fourth-level model will automatically divide the data into 121 categories, namely DE12_PH0_size0.007_OR, DE12_PH1_size0.007_OR, DE12_PH2_size0.007_OR, DE48_PH0_size0.007_OR, FE12_PH0_size0.007_OR, etc. These 121 categories are different entities in the knowledge graph, and their relationship with the 44 categories of entities in the previous level is "belong". When using the Level 5 model, the rolling bearing data is input into the model. The Level 5 model automatically divides the data into 161 categories, namely DE12_PH0_size0.007_OR_clock3, DE12_PH1_size0.007_OR_clock3, DE12_PH2_size0.007_OR_clock3, DE48_PH0_size0.007_OR_clock3, FE12_PH0_size0.007_OR_clock3, etc. These 161 categories are different entities in the knowledge graph, and their relationship with the 121 categories of entities in the previous level is "belong".

[0056] Step 5: After constructing triples based on the relationships and entities obtained from the classification, use Python to import the constructed triples into the Neo4j database to build a bearing knowledge graph and realize fault diagnosis auxiliary decision-making.

[0057] Furthermore, knowledge graph construction methods include top-down, bottom-up, and a combination of both. This knowledge graph construction adopts a combined approach. First, by analyzing the general mechanism of ship bearings, the various entities and their relationships are determined based on the mechanism, constructing the first schema layer. Then, based on relevant data such as fault types, motor speeds, and fault locations, the initial schema layer of the knowledge graph is designed using a top-down approach. Next, guided by the schema layer, a bottom-up approach is used, based on the construction of mechanistic entities and an adaptive multi-branch bidirectional long short-term memory network model, to extract relations and form high-quality knowledge representations. Finally, the obtained relations and entities are added as attribute nodes to the schema layer to ensure its updates, thus completing the construction of the knowledge graph.

[0058] First, a first model layer is constructed. The entities in this layer mainly include fault location, fault type, fault size, and fault state. This model layer contains relationships, such as the "belong" relationship between fault type and faulty component. Guided by the first model layer, a bottom-up approach is used to instantiate the model layer. Based on fault type, it can be instantiated as inner race fault, outer race fault, and rolling element fault. Based on fault state, it can be categorized as 0HP, 1HP, and 2HP motor load faults. Based on fault location, it can be categorized as faults at the 3 o'clock, 6 o'clock, and 12 o'clock positions. Finally, the model layer is updated synchronously based on the entity relationships obtained from the specific rolling bearing data, ensuring a high-quality representation of knowledge.

[0059] The extracted entities and relations can be represented using triples, for example...<DE48,belong,CWRU_ballbearing> ,<DE48_PH0,belong,DE48> We use Python programs to import entity nodes and relationships into the Neo4j database to construct a knowledge graph.

[0060] As shown in the figure, Neo4j can display different levels and the number of nodes according to different needs. This figure shows a portion of the knowledge graph nodes, where different colors represent different levels, with the same color indicating the same level. Red represents the bottom layer, green the first level, yellow the second level, and so on. The red bottom layer is CWRU_ball bearing, and the four green nodes represent the first level: normal data (NO), a driver with a sampling frequency of 12kHz (DE12), a driver with a sampling frequency of 48kHz (DE48), and a fan with a sampling frequency of 12kHz (FE12).

[0061] In knowledge graphs, you can use Cypher statements to perform queries. For example, if you want to query all the nodes connected to a certain node, you can enter the relevant command in neo4j, which will display all the nodes connected to that node and show the relevant relationships.

Claims

1. A ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with a knowledge graph, characterized in that The vibration acceleration signal of the rolling bearing is obtained using an acceleration sensor, and then rule extraction and entity construction are performed according to the structural data information of the rolling bearing and the collected vibration acceleration signal, and an adaptive multi-branch bidirectional long short-term memory network model is constructed, the adaptive multi-branch bidirectional long short-term memory network model comprises the following parts: an Inception module and a BiLSTM module, the collected rolling bearing vibration acceleration signal is divided into a training set and a test set according to a proportion, the adaptive multi-branch bidirectional long short-term memory network model is trained on the divided training set, multiple rounds of training are carried out, the optimal model hyperparameters are found, the result of the trained model meets the requirements, the model is saved, the rolling bearing is classified, and the result of the model classification is output, a triple is constructed, after the relationship and the entity obtained by classification are used to construct a triple, the constructed triple is imported into a Neo4j database using python to construct a bearing knowledge graph, and auxiliary decision-making for fault diagnosis is realized.

2. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 1, characterized in that, In step 1, the corresponding rule extraction and entity construction are performed according to the collected original bearing vibration acceleration data, which needs to be carried out according to the specific information of the data.

3. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 2, characterized in that, After the data is layered, the data in the same level is classified according to the corresponding level, and the knowledge graph is constructed using the classification result after the classification is completed, and the fault diagnosis is realized.

4. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 1, characterized in that, The pooling layers in step 2 are all maximum pooling layers.

5. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 4, characterized in that, The adaptive multi-branch bidirectional long short-term memory network model sequentially comprises a gating mechanism, in a traditional LSTM, an input gate, a forgetting gate and an output gate exist to control the flow of information. In the adaptive multi-branch bidirectional long short-term memory network model, similar gating units are added on each branch, which can determine which branch information should be used more. Some parts may be more important globally, so the branch with a larger convolution kernel will have a larger proportion. Conversely, some information may need to pay more attention to the details, so the branch with a smaller convolution kernel will have a larger proportion. Assuming that there are B branches, the output of each branch is wherein b = 1, 2,..., B. The branch selection gate is calculated by the following formula: wherein, is the weight matrix of branch b; is the bias vector of branch b; is an activation function, usually a Sigmoid function, for generating a value between 0 and 1; is the input of the current time step; is the hidden state of the previous time step. The final hidden state can be obtained by weighted summation of all branch outputs: An Inception module is defined, which processes data in parallel through multiple convolution kernels of different scales to extract multi-scale feature information. The specific Inception module contains four branches respectively: 1x1 convolution branch: 1x1 convolution kernel is used for dimension reduction operation, which reduces the amount of calculation while retaining important feature information. 6.3x3 convolution branch: first, dimension reduction is performed through 1x1 convolution, and then local features are extracted through 1x3 convolution. 7.5x5 convolution branch: first, dimension reduction is performed through 1x1 convolution, and then larger range features are extracted through 1x5 convolution.

8. Pooling branch: global features are extracted through maximum pooling operation, and then the channel number is adjusted through 1x1 convolution.

9. Among them, the output dimensions of each branch are ensured to be the same, and finally they are spliced in the channel dimension to form the final output.

10. Define the BiLSTM module. Which includes the following parts: Initialization function, Inception module, pooling layer and activation function, BiLSTM layer and full connection layer.

11. The further initialization function contains the channel number of the input data, the hidden layer dimension size of the BiLSTM, the number of layers in the BiLSTM, the number of categories of the final classification, and the Dropout layer.

12. The further Inception module uses a three-word Inception module to extract different levels of features: The first Inception module has an input channel number of the input data channel number in the initialization function and an output channel of 64; the second Inception module has an input channel number of 64 and an output channel number of 128; the third Inception module has an input channel number of 128 and an output channel number of 256. Each module will extract multi-scale features, and further enhance the expressiveness through batch normalization and activation function.

13. Further pooling layer and activation function. Max pooling layer is used to reduce the feature map size, but will not lose the most important feature information. After each pooling, the feature map width is halved. ReLU activation function is used to introduce nonlinearity, in order to enhance the model expression ability. Further BiLSTM layer. The input size of BiLSTM is equal to the output channel number of convolution layer multiplied by the width of the feature map after pooling. At the same time, using bidirectional LSTM can capture the time dependence from both forward and backward directions.

14. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 4, characterized in that The described The activation function is a sigmoid function.

15. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 4, characterized in that, The full connection layer introduces Dropout.

16. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 5, characterized in that, The dropout probability P of the Dropout is 0.

5.

17. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 1, wherein in step 3, the collected rolling bearing vibration acceleration signals are divided into a training set and a test set according to a proportion, and the proportion is to ensure that the test machine accounts for 30% of the total data set.

18. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 1, characterized in that, In step three, the model is trained on the training set through gradient descent and back propagation, and the test set is used for testing.

19. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 1, characterized in that, In step 4, the extracted nodes and relationships can be expressed by triplets.

20. The ship electric machine rolling bearing fault diagnosis method based on adaptive multi-branch bidirectional long short-term memory network combined with knowledge graph according to claim 8, characterized in that, The previously extracted entity nodes and relationships are imported into the Neo4j database by using python to complete the construction of the rolling bearing knowledge graph.

Images