A fault diagnosis method for industrial robot joints using few-label vibration data

Abstract

The present invention provides a fault diagnosis method for the few-label vibration data of the joints of an industrial robot, which includes data preprocessing the collected original vibration data; using the preprocessed data samples to construct graph structure data; performing graph convolution, BatchNorm, ReLU, and another layer of graph convolution operations on the graph structure data in sequence to obtain an encoded feature matrix; performing decoding operations on the feature matrix to obtain a reconstructed adjacency matrix; calculating a loss function; performing back propagation according to the loss function L to train an encoder; obtaining a trained encoder; using the trained encoder to encode the graph data to obtain extracted fault features; using the extracted fault features as new samples, adding all labels, and performing supervised training; obtaining a fault diagnosis result. The present invention is mainly used in the construction of graph data and supervised fault feature extraction, and can be used for fault diagnosis of few-label samples and fault feature extraction of unlabeled samples.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault diagnosis method for low-label vibration data of industrial robot joints based on graph neural networks. Background Art

[0002] In recent years, with the rapid development of the machinery industry, industrial robots have been widely used in various mechanical equipment. In particular, harmonic reducers, which are widely used in the joints of lightweight robots and are core precision components, are mainly targeted at samples with no or few labels, and the data type is the vibration signal collected by a single sensor. At the same time, this technical method can also be applied to other components such as bearings. The common states of harmonic reducers mainly include one normal state and three fault states, a total of four states, among which the fault states are mainly flexspline pitting fault, wave generator jamming fault and output end bearing misalignment fault. At present, the common intelligent fault diagnosis methods are mainly supervised methods, which require a large number of labeled data samples for learning. As for supervised methods, most of them are traditional deep learning methods, such as methods based on Auto-encoder and VariationalAuto-encoder. As for fault diagnosis based on graph learning, the biggest difficulty at present is still how to process traditional sample data and build a suitable graph data structure. Since the sensor signal data used in fault diagnosis do not have a natural association graph relationship like a social network, how to construct the association graph between data is extremely important, which greatly affects the quality and upper limit of the subsequent graph learning effect. At present, most of the graph data structure construction methods are mainly based on the K-nearest neighbor (KNN) algorithm and the construction method based on prior knowledge, and most of them deal with systems or multi-source data with complex node relationships, while the construction methods for a single type of vibration signal data are limited. Using the KNN method to construct graph data: First, extract a large number of time-frequency features of a given data set; then, through feature selection, obtain a small number of time-frequency features that can reflect the characteristics of the system; then use the nearest neighbor (KNN) method to find the K nearest neighbors of a given node based on these features, and connect them by edges; finally, the construction of association graph data is realized. Using prior knowledge to construct association graph data: First, collect data in a targeted manner according to the specific research object, especially multi-source data; then, based on your full understanding of the research object and data collection method, based on deep expert knowledge, analyze the association relationship between data through traditional mathematical methods; then establish edge connections based on the association relationship to realize the construction of graph data.

[0003] Disadvantages of existing technologies: supervised learning methods have a high demand for labeled samples, but in actual industrial environments, there may not be sufficient labeled data, which does not meet this condition. At the same time, it takes a lot of time and money to collect additional labeled data or label existing data, which will increase the company's budget. The supervised method based on traditional deep learning will regard the relationship between samples as independent and identically distributed, and cannot fully utilize the correlation between data samples.

[0004] Using the KNN method to construct graph data: The KNN method is not reliable enough, and the selection of parameter K greatly affects the performance of the KNN algorithm. This method has certain requirements on the significance of the fault characteristics contained in the data itself. If the difference between the sample data is not large and the fault characteristics are not significant enough, it is easy to lead to poor KNN effect, resulting in more erroneous connections between nodes on the graph, making the graph data quality poor, affecting subsequent learning and classification effects.

[0005] Using prior knowledge to construct association graph data: Prior knowledge requires sufficient understanding of the research object and the collected content, and needs to be supported by rich and deep expert knowledge, which is quite difficult. At the same time, the introduction of prior knowledge significantly increases the training time, requiring not only knowledge about the system mechanism, but also data information.

[0006] In addition, the above two graph data construction methods mostly deal with systems or multi-source data with complex node relationships, and there are limited ways to construct graphs for a single type of vibration signal data. Moreover, for different research objects or different situations of the same research object, specific analysis may be required, which is more complex and difficult and has poor versatility. When the research object or data changes slightly, it may take a lot of time to re-analyze and train. Summary of the invention

[0007] In order to solve the problems existing in the existing fusion methods, the purpose of the present invention is to provide a fault diagnosis method for sparsely labeled vibration data of industrial robot joints based on graph neural networks. This method is mainly used for the construction of graph data and supervised fault feature extraction, and can be used for fault diagnosis of sparsely labeled samples and fault feature extraction of unlabeled samples.

[0008] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0009] A fault diagnosis method for industrial robot joints based on few-label vibration data, the method comprising the following steps:

[0010] Step S1, collecting original vibration data of the industrial robot joints;

[0011] Step S2, performing data analysis and preprocessing on the collected raw vibration data;

[0012] Step S3, using the preprocessed data samples, constructing graph structure data based on the supervised learning method of GNN;

[0013] Step S4, performing graph convolution, BatchNorm, ReLU, and another layer of graph convolution operations on the graph structure data in sequence to obtain the encoded feature matrix Z;

[0014] Step S5, performing a decoding operation on the feature matrix Z to obtain a reconstructed adjacency matrix A;

[0015] Step S6, calculating the loss function L according to the adjacency matrix A and the feature matrix Z;

[0016] Step S7, performing back propagation according to the loss function L to train the encoder;

[0017] Step S8, repeating steps S4-S7 multiple times to obtain a trained encoder;

[0018] Step S9, using the trained encoder to encode the graph data to obtain the extracted fault features;

[0019] Step S10, using the extracted fault features as new samples, adding all labels, and performing supervised training;

[0020] Step S11, obtaining a fault diagnosis result.

[0021] According to a fault diagnosis method for the few-label vibration data of an industrial robot joint provided by the present invention, before constructing the graph structure data, a sliding window method is selected to divide the data samples to standardize the length of the sample data, and a Z-Score Normalization operation is used. In the Z-Score normalization, the operation shown in formula (1) is performed on each data point:

[0022]

[0023] Among them, x is the original data point, μ is the mean of the data set, σ is the standard deviation of the data set, and z is the normalized data point.

[0024] According to a fault diagnosis method for few-label vibration data of an industrial robot joint provided by the present invention, in step S3, when constructing graph structure data, each preprocessed sample is taken as a node, and then connections between nodes are established, and then a vector representation of each node is obtained through a supervised learning method of a GNN;

[0025] The vibration signal of each preprocessed sample is divided into several parts according to the time sequence, and K parts are selectively taken, so that each sample is divided into K new samples, where the K new samples are all vibration signals of the same length;

[0026] The K new samples are regarded as K nodes, and the vibration signal of each sample is the attribute feature of the node. Then, an undirected complete graph is constructed, in which the weight of each edge is the cosine distance between the two nodes it connects. Assuming that there is an edge connection between nodes u and v, the calculation expression of the cosine distance dist(u,v) between them is shown in formula (2):

[0027]

[0028] Among them, Z u and Z v are two vectors, representing the attribute features of node u and node v respectively, representing the dot product of the vectors, ||Z u || and ||Z v || respectively represent vector Z u and Z v The norm of .

[0029] According to a fault diagnosis method for sparsely labeled vibration data of an industrial robot joint provided by the present invention, a complete graph obtained with K nodes is used as a subgraph, and each preprocessed sample can also obtain a subgraph. Finally, all subgraphs are combined together to form complete graph structure data.

[0030] According to a fault diagnosis method for few-label vibration data of an industrial robot joint provided by the present invention,

[0031] In step S5, when the reconstructed adjacency matrix A is obtained, it is expressed as formula (3):

[0032] A=sigmoid(ZZ T ) (3)

[0033] According to a fault diagnosis method for few-label vibration data of an industrial robot joint provided by the present invention,

[0034] In step S7, the loss function L is expressed as formula (4):

[0035]

[0036] Among them, L_recon represents the reconstruction error loss, N represents the number of nodes in the graph, and A ij The elements of the adjacency matrix representing the graph data, A ij Represents the elements of the reconstructed adjacency matrix, L_kl corresponds to the KL divergence loss, Z ijRepresents the elements of the feature matrix obtained by the encoder, h represents the dimension of the hidden layer of the encoder, and α is a constant coefficient used to balance the proportion of reconstruction error loss and KL divergence loss.

[0037] According to a fault diagnosis method for few-label vibration data of an industrial robot joint provided by the present invention, the encoder is a fault feature extractor, which includes a first layer of graph convolution, a ReLU activation layer, a batch normalization layer and a second layer of graph convolution.

[0038] A fault diagnosis method for few-label vibration data of an industrial robot joint provided by the present invention further includes the step of constructing a mathematical model of the industrial robot, including:

[0039] The harmonic reducer fault test system is used as an experimental platform, and sensor data containing the operating status information of the industrial robot is selected for fault diagnosis;

[0040] The harmonic reducer fault test system includes a test bench with a faulty harmonic reducer installed, an acceleration sensor, an NI data acquisition device, and a fault signal acquisition software;

[0041] The test bench includes a control system, a drive motor, a load motor, a torque tachometer, a planetary reduction gearbox, a dynamic torque sensor, and a control panel.

[0042] According to a fault diagnosis method for the few-label vibration data of an industrial robot joint provided by the present invention, when conducting an experiment, the faulty part is first manufactured, and then the harmonic reducer with different types of faults can be replaced by replacing the faulty part of the harmonic reducer on the test bench;

[0043] Next, the control panel is used to display various parameter settings of the test bench, among which the input rated torque, load torque, input speed, output speed and other data are monitored in real time through the panel according to the torque tachometer and dynamic torque sensor;

[0044] When the test bench is running, the drive motor provides the rated input torque, and the load torque is provided by the motor at the load end to simulate the load operation of the harmonic reducer under real working conditions;

[0045] During data collection, the input speed and load torque are controlled through a digital control panel to collect data for various working conditions, where the load torque is measured as a percentage of the input rated torque.

[0046] According to a fault diagnosis method for sparsely labeled vibration data of an industrial robot joint provided by the present invention, after step S9, the acquired fault features can be individually taken out and added with the original labels as new sample data for executing all downstream tasks.

[0047] It can be seen that, compared with the prior art, the present invention is based on graph neural networks, fully utilizes the correlation between samples, and proposes a simple, feasible and reliable graph data construction method to solve the problem that it is difficult to construct a suitable graph data structure for single vibration signal data of non-multi-source information and non-complex systems under supervised conditions, so that it is suitable for vibration signal data collected by a single sensor. At the same time, compared with the currently common graph composition method, the graph composition method provided by the present invention can reduce the demand for expert knowledge, and has certain versatility and stability, and can obtain better harmonic reducer fault diagnosis effect.

[0048] Furthermore, the present invention can be used for supervised graph embedding learning and supervised fault feature extraction.

[0049] Furthermore, the present invention is applicable to intelligent fault diagnosis under the condition of few labeled samples, which is more in line with the actual industrial environment and helps save costs.

[0050] Furthermore, this method can construct a suitable graph data structure for single vibration signal data of non-multi-source information and non-complex systems in a simple and reliable way, without requiring too much expert knowledge and with a certain degree of stability.

[0051] Furthermore, the supervised feature extraction part of the present invention can be used alone as a feature extraction technology, has certain scalability, and can be combined with the KNN composition method or the prior knowledge composition method to obtain better feature extraction effect or fault diagnosis accuracy.

[0052] Furthermore, the present invention is derived from an experimental model combined with a harmonic reducer, and the proposed method can be further applied to other industrial components with good reliability and versatility.

[0053] The present invention also provides an electronic device, comprising:

[0054] a memory storing computer executable instructions;

[0055] a processor configured to execute the computer executable instructions,

[0056] Wherein, when the computer executable instructions are executed by the processor, the steps of any one of the above-mentioned fault diagnosis methods for the sparsely labeled vibration data of the industrial robot joints are implemented.

[0057] The present invention also provides a storage medium having a computer program stored thereon, and when the computer program is executed by a processor, it is used to implement the steps of any of the above-mentioned fault diagnosis methods for sparsely labeled vibration data of industrial robot joints.

[0058] It can be seen that the present invention also provides an electronic device and storage medium for a fault diagnosis method of vibration data with few labels of an industrial robot joint, which includes: one or more memories and one or more processors. The memory is used to store program codes and intermediate data generated during program operation, storage of model output results, and storage of models and model parameters; the processor is used for processor resources occupied by code operation and multiple processor resources occupied when training models.

[0059] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a flow chart of an embodiment of a fault diagnosis method for low-label vibration data of an industrial robot joint according to the present invention.

[0061] Figure 2 It is a schematic diagram of an embodiment of a fault diagnosis method for low-label vibration data of an industrial robot joint of the present invention.

[0062] Figure 3 It is a schematic diagram of the visualization effect in an embodiment of a fault diagnosis method for low-label vibration data of an industrial robot joint of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0065] See also Figures 1 to 3 The present invention provides a fault diagnosis method for the few-label vibration data of an industrial robot joint, the method comprising the following steps:

[0066] Step S1, collecting original vibration data of the industrial robot joints;

[0067] Step S2, performing data analysis and preprocessing on the collected raw vibration data;

[0068] Step S3, using the preprocessed data samples, constructing graph structure data based on the supervised learning method of GNN;

[0069] Step S4, performing graph convolution, BatchNorm, ReLU, and another layer of graph convolution operations on the graph structure data in sequence to obtain the encoded feature matrix Z;

[0070] Step S5, performing a decoding operation on the feature matrix Z to obtain a reconstructed adjacency matrix A;

[0071] Step S6, calculating the loss function L according to the adjacency matrix A and the feature matrix Z;

[0072] Step S7, performing back propagation according to the loss function L to train the encoder;

[0073] Step S8, repeating steps S4-S7 multiple times to obtain a trained encoder;

[0074] Step S9, using the trained encoder to encode the graph data to obtain the extracted fault features;

[0075] Step S10, using the extracted fault features as new samples, adding all labels, and performing supervised training;

[0076] Step S11, obtaining a fault diagnosis result.

[0077] In the above step S2, before constructing the graph structure data, the sliding window method is selected to divide the data samples to standardize the length of the sample data, and the Z-Score Normalization operation is used. In the Z-Score normalization, the operation shown in formula (1) is performed on each data point:

[0078]

[0079] Among them, x is the original data point, μ is the mean of the data set, σ is the standard deviation of the data set, and z is the normalized data point. By standardizing the data to a distribution with a mean of 0 and a standard deviation of 1, it helps to improve the performance of the model and the stability of data processing.

[0080] For the data samples of vibration signals, it is assumed that they have undergone preliminary data analysis and preprocessing operations (such as normalization operations, Fourier transform operations, etc.), and each sample obtained in this embodiment is temporarily referred to as a "preprocessed sample". Next, it is necessary to construct graph data. Since this embodiment needs to perform supervised learning based on GNN, fault features are extracted, that is, representation learning is performed. Therefore, this is a task at the node level, a task called graph embedding or node embedding. Therefore, in step S3, when constructing graph structure data, each preprocessed sample (that is, a vibration signal) is taken as a node, and then connections between nodes are established, and then the vector representation of each node is obtained through the supervised learning method of GNN.

[0081] Then, the vibration signal of each preprocessed sample is divided into several parts in time order, and K parts are selectively taken, so that each sample is divided into K new samples. It should be noted that these K new samples are vibration signals of the same length.

[0082] Next, these K new samples can be regarded as K nodes, and the vibration signal of each sample is the attribute feature of the node. Then, an undirected complete graph is constructed, in which the weight of each edge is the cosine distance between the two nodes it connects. Assuming that there is an edge connection between nodes u and v, the calculation expression of the cosine distance dist(u,v) between them is shown in formula (2):

[0083]

[0084] Among them, Z u and Z v are two vectors, representing the attribute features of node u and node v respectively, representing the dot product of the vectors, ||Z u || and ||Z v || respectively represent vector Z u and Z v The norm of .

[0085] Taking the complete graph obtained with K nodes as a subgraph, each preprocessed sample can also get a subgraph. Finally, all subgraphs are combined together (there are no edges connecting the subgraphs) to form a complete graph structure data.

[0086] In step S5, when the reconstructed adjacency matrix A is obtained, it is expressed as formula (3):

[0087] A=sigmoid(ZZ T ) (3)

[0088] In step S7, the loss function L is expressed as formula (4):

[0089]

[0090] Among them, L_recon represents the reconstruction error loss, N represents the number of nodes in the graph, and A ij The elements of the adjacency matrix representing the graph data, A ij Represents the elements of the reconstructed adjacency matrix, L_kl corresponds to the KL divergence loss, Z ij Represents the elements of the feature matrix obtained by the encoder, h represents the dimension of the hidden layer of the encoder, and α is a constant coefficient used to balance the proportion of reconstruction error loss and KL divergence loss.

[0091] In the above steps S7-S9, the encoder is a fault feature extractor, which includes a first layer of graph convolution, a ReLU activation layer, a batch normalization layer and a second layer of graph convolution.

[0092] Specifically, this embodiment also provides a method and steps for constructing a mathematical model of an industrial robot, including:

[0093] The harmonic reducer fault test system is used as an experimental platform, and sensor data containing the operating status information of the industrial robot is selected for fault diagnosis;

[0094] The harmonic reducer fault test system includes a test bench with a faulty harmonic reducer installed, an acceleration sensor, an NI data acquisition device, and a fault signal acquisition software;

[0095] The test bench includes a control system, a drive motor, a load motor, a torque tachometer, a planetary reduction gearbox, a dynamic torque sensor, and a control panel.

[0096] When conducting the experiment, the faulty parts are manufactured first, and then the harmonic reducers with different types of faults can be replaced by replacing the faulty parts of the harmonic reducer on the test bench.

[0097] Next, the control panel is used to display various parameter settings of the test bench. Among them, the input rated torque, load torque, input speed, output speed and other data are monitored in real time through the panel based on the torque tachometer and dynamic torque sensor.

[0098] When the test bench is running, the drive motor provides the rated input torque, and the load torque is provided by the motor at the load end to simulate the load operation of the harmonic reducer under real working conditions.

[0099] During data collection, the input speed and load torque are controlled through a digital control panel to collect data for various working conditions, where the load torque is measured as a percentage of the input rated torque.

[0100] In addition, the faulty parts can be replaced by detaching the coupling, and under the action of the lower slide rail, the half platform can be conveniently moved with a hand wheel to facilitate the replacement of the faulty parts.

[0101] The experimental data of the industrial robot mathematical model are shown in Table (1):

[0102] (1)

[0104] Based on the above-mentioned harmonic reducer test bench, this embodiment produced several typical types of faulty parts, conducted experiments and data collection, and obtained three types of fault type data and one type of normal state data. As shown in the above table (1), the four states are referred to as "normal", "pitting", "misalignment" and "stuck". During the data collection process, the sampling frequency is always fixed at 25600HZ, the speed is divided into 600rpm, 1200rpm and 1500rpm, and the load is set to 0%, 10%, 20% and 30% respectively. The state, speed and load are randomly combined to obtain a total of 48 different types of sub-data sets, all of which are vibration signals measured by the same acceleration sensor.

[0105] The detailed operation of the fault diagnosis method for unlabeled vibration signal data based on graph neural network proposed in this embodiment in experimental research (supplementary explanation of specific details) is as follows:

[0106] Before constructing the graph data, this embodiment selects to use the sliding window method to divide the data samples to standardize the length of the sample data, and uses the Z-Score Normalization operation. In the Z-Score normalization, each data point is performed as shown in the above formula (1).

[0107] Next, in order to make the graph neural network learn better, while taking into account the purpose of being as end-to-end as possible and reducing the need for expert knowledge, the time domain signal data is transformed by FFT, and no other signal processing operations are required. Then, only the sample data under the four conditions of 600rpm and 0 load are selected, and the undirected graph data can be constructed according to the above method to obtain the input graph of GAE. The graph has a total of 1800 nodes, including 180 CompleteSubgraphs described above. Each subgraph contains 10 nodes, and the length of the attribute features of each node is 1024. There is no edge connection between the nodes of the subgraphs. After that, it is input into the GAE model for training, and the extracted features can be obtained by encoding using the trained model.

[0108] In the specific experimental operation, combined with the PyG framework, this embodiment designs the parameters of the GAE model according to the specific conditions of the experiment as shown in the following table (2):

[0109] (2)

[0111] The experimental parameter settings are shown in Table (3):

[0112] (3)

[0114] Then, after the training is completed, a trained encoder is obtained, which can be regarded as the feature extractor of the graph data. By using the encoder to encode the graph data, a feature matrix can be obtained. These features are regarded as the new sample features after feature extraction. Then, the labels of each new sample are added correspondingly using the obtained new sample features to obtain a new data set. 6% of the data in the new data set is used as the training set for supervised training of the classifier to simulate the actual industrial situation with fewer labeled samples. In addition, the remaining 94% of the new samples are used as the test set. In order to fully verify the quality of the new sample data, based on scikit-learn, only some basic and common machine learning classification models are selected here, and the new and old samples are used for training and testing respectively to compare the diagnosis results. During model training, in order to fully reflect the quality of sample features, the parameters of the classifier are selected more arbitrarily, and there is no need to deliberately adjust the parameters to reduce the influence of the classifier. To avoid contingency, five common classifiers are selected here, namely support vector machine classifier (SVC), random forest classifier (RFC), XGBoost classifier (XGB), K nearest neighbor classifier (KNN) and multi-layer perceptron (MLP).

[0115] The specific settings of various classifiers are shown in Table (4):

[0116] (4)

[0118] When using new and old samples to train and test the classifier respectively, please note that the classifier settings must be exactly as shown in the above table.

[0119] The test results (6% training, 94% testing) are shown in Table (5):

[0120]

[0121] (5)

[0123] It can be seen that for the above five common machine learning classifiers, for each classifier under the same settings, the effect of using new samples for training and testing is significantly higher than that of using old samples. Therefore, it can be judged that the new samples obtained according to the above method are of higher quality and their fault features are more significant. The fault feature extraction in the above method has achieved good results. In addition, the average accuracy of using new samples is above 96%. Therefore, it can be seen that the proposed technical method has achieved a relatively excellent fault diagnosis effect, with a high accuracy, and this is true for multiple classifiers, and has a certain stability.

[0124] In addition, in order to observe the feature extraction effect more intuitively, this embodiment uses the t-SNE algorithm to reduce the dimensions of the original features and the extracted features, and compares the embedding clustering effects before and after extraction. The visualization effect is as follows: Figure 3 shown.

[0125] After step S9, the acquired fault features can be taken out separately, added with the original labels, and used as new sample data for executing all downstream tasks.

[0126] It can be seen that the method provided in this embodiment can be combined with many methods and has good scalability. For example, in steps S1-S2, some existing methods with good effects can be combined to analyze and preprocess the data as long as the fault characteristics can be made more significant.

[0127] In addition, after step S9, the obtained fault features can be taken out separately, and the original labels can be added as new sample data with better quality to perform all downstream tasks, such as training with other supervised learning methods with excellent results. In general, this method can be used for intelligent fault diagnosis with few labels, but it is not limited to the case of few labels. It can also take out the supervised fault feature method and use it separately, and intersperse it with other various fault diagnosis methods to improve the quality of sample data.

[0128] Obviously, compared with the unextracted features, after a series of operations using the above method, the extracted features achieved a better clustering effect, which also intuitively proves that the fault features have been effectively extracted.

[0129] In summary, the present invention proposes a fault diagnosis method for industrial robot joints based on graph neural network with few-label vibration signal data, which mainly focuses on the construction of graph data and supervised fault feature extraction, and can be used for fault diagnosis of few-label samples and fault feature extraction of unlabeled samples. This method can construct appropriate graph structure data for vibration signals collected by a single sensor and perform graph learning. At the same time, the training time is short and the diagnostic accuracy is high. The summary can include the following points:

[0130] 1. Extract fault features from unlabeled samples; 2. Fault diagnosis for samples with few labels; 3. Construct appropriate graph structure data for vibration signals collected by a single sensor for graph learning; 4. No need for complex expert knowledge as a basis, simple and effective; 5. The feature extraction part can be taken out separately as a method to improve sample quality, and combined with other fault diagnosis methods, it has good stability, versatility and scalability.

[0131] In one embodiment, an electronic device is provided, which may be a server. The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the electronic device is used to store data. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a fault diagnosis method for low-label vibration data of an industrial robot joint is implemented.

[0132] Those skilled in the art will understand that the electronic device structure shown in this embodiment is only a partial structure related to the scheme of the present application, and does not constitute a limitation on the electronic device to which the scheme of the present application is applied. The specific electronic device may include more or fewer components than those shown in this embodiment, or combine certain components, or have a different arrangement of components.

[0133] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0134] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0135] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0136] It can be seen that this embodiment also provides an electronic device and storage medium for a fault diagnosis method of vibration data with few labels of an industrial robot joint, which includes: one or more memories and one or more processors. The memory is used to store program codes and intermediate data generated during program operation, storage of model output results, and storage of models and model parameters; the processor is used for processor resources occupied by code operation and multiple processor resources occupied when training models.

[0137] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0138] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A fault diagnosis method for industrial robot joints using sparsely labeled vibration data, characterized in that: The method comprises the following steps: Step S1, collecting original vibration data of the industrial robot joints; Step S2, performing data analysis and preprocessing on the collected raw vibration data; Step S3, using the preprocessed data samples, constructing graph structure data based on the supervised learning method of GNN; Step S4, performing graph convolution, BatchNorm, ReLU, and another layer of graph convolution operations on the graph structure data in sequence to obtain the encoded feature matrix Z; Step S5, performing a decoding operation on the feature matrix Z to obtain a reconstructed adjacency matrix A; Step S6, calculating the loss function L according to the adjacency matrix A and the feature matrix Z; Step S7, performing back propagation according to the loss function L to train the encoder; Step S8, repeating steps S4-S7 multiple times to obtain a trained encoder; Step S9, using the trained encoder to encode the graph data to obtain the extracted fault features; Step S10, using the extracted fault features as new samples, adding all labels, and performing supervised training; Step S11, obtaining a fault diagnosis result; In step S3, when constructing graph structure data, each preprocessed sample is taken as a node, and then connections between nodes are established, and then a vector representation of each node is obtained through a supervised learning method of GNN; The vibration signal of each preprocessed sample is divided into several parts according to the time sequence, and K parts are selectively taken, so that each sample is divided into K new samples, where the K new samples are all vibration signals of the same length; The K new samples are regarded as K nodes, and the vibration signal of each sample is the attribute feature of the node. Then, an undirected complete graph is constructed, in which the weight of each edge is the cosine distance between the two nodes it connects. Assuming that there is an edge connection between nodes u and v, the calculation expression of the cosine distance dist(u,v) between them is shown in formula (2): Among them, Z u and Z v are two vectors, representing the attribute features of node u and node v respectively, · represents the dot product of the vectors, ||Z u || and ||Z v || respectively represent vector Z u and Z v The norm of ; Taking the complete graph obtained by K nodes as a subgraph, each preprocessed sample can also get a subgraph, and finally all subgraphs are combined together to form a complete graph structure data; In step S7, the loss function L is expressed as formula (4): L=L_recon+α*L_kl Among them, L_recon represents the reconstruction error loss, N represents the number of nodes in the graph, and A ij The elements of the adjacency matrix representing the graph data, Represents the elements of the reconstructed adjacency matrix, L_kl corresponds to the KL divergence loss, Z ij Represents the elements of the feature matrix obtained by the encoder, h represents the dimension of the hidden layer of the encoder, and α is a constant coefficient used to balance the proportion of reconstruction error loss and KL divergence loss.

2. The method according to claim 1, characterized in that: Before constructing the graph structure data, we choose to use the sliding window method to divide the data samples to standardize the length of the sample data, and use the Z-Score Normalization operation. In the Z-Score normalization, each data point is operated as shown in formula (1): Among them, x is the original data point, μ is the mean of the data set, σ is the standard deviation of the data set, and z is the normalized data point.

3. The method according to claim 1, characterized in that: In step S5, when the reconstructed adjacency matrix A is obtained, it is expressed as formula (3):

4. The method according to claim 1, characterized in that: The encoder is a fault feature extractor, which includes a first layer of graph convolution, a ReLU activation layer, a batch normalization layer, and a second layer of graph convolution.

5. The method according to any one of claims 1 to 4, characterized in that: It also includes steps to build a mathematical model of an industrial robot, including: The harmonic reducer fault test system is used as an experimental platform, and sensor data containing the operating status information of the industrial robot is selected for fault diagnosis; The harmonic reducer fault test system includes a test bench with a faulty harmonic reducer installed, an acceleration sensor, an NI data acquisition device, and a fault signal acquisition software; The test bench includes a control system, a drive motor, a load motor, a torque tachometer, a planetary reduction gearbox, a dynamic torque sensor and a control panel.

6. The method according to claim 5, characterized in that: When conducting the experiment, the faulty parts are manufactured first, and then the harmonic reducers with different types of faults can be replaced by replacing the faulty parts of the harmonic reducer on the test bench; Next, the control panel is used to display various parameter settings of the test bench, among which the input rated torque, load torque, input speed, and output speed data are monitored in real time through the panel according to the torque tachometer and dynamic torque sensor; When the test bench is running, the drive motor provides the rated input torque, and the load torque is provided by the motor at the load end to simulate the load operation of the harmonic reducer under real working conditions; During data collection, the input speed and load torque are controlled through a digital control panel to collect data for various working conditions, where the load torque is measured as a percentage of the input rated torque.

7. The method according to any one of claims 1 to 4, characterized in that: After step S9, the acquired fault features can be taken out separately, added with the original labels, and used as new sample data for executing all downstream tasks.

Images