An industrial robot autonomous fault diagnosis method under zero fault samples

Abstract

The application relates to an industrial robot autonomous fault diagnosis method under a zero-failure sample, and belongs to the field of intelligent abnormality detection of mechanical equipment. The method comprises the following steps: (1) constructing a continuous wavelet transform knowledge base, performing multidimensional time-frequency analysis on normal state monitoring data of an industrial robot, and constructing a multimodal feature set of the monitoring data of the industrial robot; (2) establishing a monitoring data state representation extractor based on a deep network model; (3) constructing a multimodal feature comparison matrix by using the multimodal feature set, constructing a comparison learning loss function capable of determining the optimization direction of the network model based on the above two angles, and optimizing the parameters of the deep network model by using the comparison learning loss function; and (4) using the deep network model obtained through training to perform state representation on the multimodal features of real-time monitoring data of the industrial robot, taking the multimodal features of the normal state monitoring data as a benchmark, comparing and analyzing the state representation changes of the real-time monitoring data, and performing fault state discrimination.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent abnormal detection of mechanical equipment, and particularly relates to an industrial robot autonomous fault diagnosis method under zero fault samples. BACKGROUND

[0002] Industrial robots have become an indispensable important part of production and manufacturing due to high automation, long running time and high action completion quality. With the rapid development of new energy automobile, electronic semiconductor, lithium battery / photovoltaic, aerospace, home appliance manufacturing and other industries, there is an urgent demand for intelligent industrial robots. Autonomous state perception, as an important part of intelligence, is the key to realizing efficient and healthy management of industrial robots. However, with the increasing complexity of the structure and system of industrial robots, it becomes more and more difficult to detect fault states with high accuracy. Existing fault state detection methods include two categories, namely model-driven and data-driven. Since the model-driven method needs to be modeled according to structural dynamics and kinematics knowledge, it is very complex in practice and has a narrow application range, so in recent years, data-driven methods have received widespread attention. The key of this kind of method is to establish the mapping relationship between the fault state of mechanical equipment and the monitoring data.

[0003] At present, the data-driven industrial robot anomaly detection method is difficult to construct a large and complete fault sample set for deep intelligent model training, and it is difficult to collect enough industrial robot fault samples in engineering practice. Accordingly, there is a technical demand in the field to construct an autonomous anomaly detection of industrial robots under zero fault samples. SUMMARY

[0004] The purpose of the present application is to overcome the limitations of existing data-driven anomaly detection technology, and provide an industrial robot autonomous fault diagnosis method under zero fault samples, which combines multiple continuous wavelet analysis and deep network models to realize autonomous anomaly detection of industrial robots under zero fault samples.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows: an industrial robot autonomous fault diagnosis method under zero fault samples, comprising:

[0006] A continuous wavelet transform knowledge base is constructed according to multiple mother wavelet functions, multi-dimensional time-frequency analysis is performed on the normal state monitoring data of the industrial robot, and a multi-modal feature set of the normal state monitoring data of the industrial robot is constructed;

[0007] A monitoring data state representation extractor based on a deep network model is established, and the state representation extractor is used to perform state representation on the modal features of the normal state monitoring data of the industrial robot;

[0008] The normal state monitoring data of the industrial robot has time sequence and different modal characteristics, a multi-modal feature comparison matrix is constructed by using a multi-modal feature set, a comparison learning loss function capable of determining the optimization direction of the deep network model is constructed, and the parameters of the deep network model are optimized by using the comparison learning loss function.

[0009] The multi-modal features of the real-time monitoring data of the industrial robot are represented by using the deep network model obtained by training, the multi-modal features of the normal state monitoring data of the industrial robot are taken as a benchmark, and the state representation change of the real-time monitoring data is compared and analyzed to determine the fault state.

[0010] In an embodiment of the present application, the plurality of mother wavelet functions include mexican hat mother wavelet function mexh, complex gaussian B-spline wavelet function fbsp, morlet wavelet function, harr wavelet function, gaus wavelet function, etc.

[0011] In an embodiment of the present application, the specific implementation mode of the multi-dimensional time-frequency analysis on the normal state monitoring data of the industrial robot and the construction of the multi-modal feature set of the normal state monitoring data of the industrial robot is as follows:

[0012] Selecting connected data samples from the normal state monitoring data of the industrial robot 、 , constructing a multi-modal feature set for the two samples by using a continuous wavelet transform knowledge base:

[0013]

[0014] Among them, and are modal features, indicates a total of n different continuous wavelet transforms, m indicates the m data sample.

[0015] In an embodiment of the present application, the specific representation form of the state representation of the modal features of the normal state monitoring data of the industrial robot by using the state representation extractor is as follows:

[0016]

[0017] Among them, F () is a feature extraction operation, and are the results after feature extraction.

[0018] In an embodiment of the present application, the representation form of the construction of the multi-modal feature comparison matrix by using the multi-modal feature set is as follows:

[0019]

[0020] Where M is the comparison matrix.

[0021] In one embodiment of the present invention, the specific implementation of constructing the contrastive learning loss function capable of determining the optimization direction of the network model is as follows:

[0022] right , Correlation calculations were performed to obtain the comparative analysis coefficient matrix. :

[0023]

[0024] in, cc() This is an operation for calculating cosine similarity.

[0025] Loss function 1 is constructed based on the maximum similarity assumption:

[0026]

[0027] Loss function 2 is constructed based on the minimum similarity assumption:

[0028]

[0029] As can be seen from the symmetry relationship, when At that time, matrix No. i In the line The maximum value should be taken, which is represented as:

[0030]

[0031] Through with By comparison, we construct cross-entropy and obtain loss function 3:

[0032]

[0033] in, express T The Middle r row element, express The Middle r row element, R for T The number of Bank of China branches;

[0034] Constructing the contrastive learning loss function:

[0035] .

[0036] In an embodiment of the present application, the deep network model obtained by training is used to perform state representation on the multi-modal features of the real-time monitoring data of the industrial robot, and the multi-modal features of the normal state monitoring data of the industrial robot are taken as a benchmark, and by comparing and analyzing the state representation changes of the real-time monitoring data, the specific implementation mode of the fault state discrimination is as follows:

[0037] Selecting the normal state monitoring data of the industrial robot as a benchmark, constructing a multi-modal feature set and extracting state representation, obtaining state representation ;

[0038] Extracting state representation of real-time monitoring data of the industrial robot ;

[0039] Constructing an anomaly detection matrix :

[0040]

[0041] Normalizing each row element of the anomaly detection matrix to the interval [-0.5, 0.5], and using the argmax operation to obtain the index of the maximum value in each row element, denoted as , and finally obtaining the detection label :

[0042]

[0043] Designing a threshold value Threshold , when satisfies , the detection result is a fault, when , and , the detection result is normal; wherein abs() is the absolute value, and sum() is the sum.

[0044] Compared with the prior art, the present application has the following beneficial effects:

[0045] 1. The introduction of a variety of wavelet mother functions to construct a continuous wavelet transform knowledge base for multi-modal feature extraction of industrial robot monitoring data can reveal the time-frequency information of industrial robot monitoring data from different angles, providing a basis for accurately sensing the running state of the industrial robot.

[0046] 2. The powerful feature extraction capability of the deep network model makes it possible to obtain representative state representation of the industrial robot, providing support for high-precision anomaly detection.

[0047] 3. The state representation of normal state monitoring data is used to build a contrast learning loss function, which is used to optimize the parameters of the deep network, so that the deep network model is free from the dependence on labeled samples, breaks through the limitation of the deep network model caused by the scarcity of real fault samples in engineering practice, and improves the application range of the deep network model.

[0048] 4. The state representation of normal state monitoring data of the industrial robot is extracted by the deep learning, and the real-time monitoring data can be monitored by the state representation extractor based on the network model. If the state representation of the real-time monitoring data is found to be abnormal, the running state of the industrial robot can be determined. Such state detection not only solves the problem of not being able to extract the deep state representation of the monitoring data, but also makes the deep network model training free from the dependence on fault samples.

[0049] 5. The proposed autonomous anomaly detection method is widely applicable to various mechanical equipment, and can quickly establish an anomaly detection model at each stage of the life cycle of the mechanical equipment, which shows that the method can adapt to various operating conditions of the mechanical equipment. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of the autonomous anomaly detection method of the industrial robot under zero fault samples provided by the present application;

[0051] Figure 2 is Figure 1 the change process of the loss function constructed in the training process of the autonomous anomaly detection model of the industrial robot under zero fault samples in DETAILED DESCRIPTION

[0052] In order to make the ideas, technical solutions and main innovations of the present application clearer, the details of the present application will be described in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0053] Please refer to Figure 1 , the present application provides an autonomous fault diagnosis method of the industrial robot under zero fault samples, mainly including the following steps:

[0054] Step 1, a continuous wavelet transform knowledge base is constructed according to a plurality of mother wavelet functions, multi-dimensional time-frequency analysis is performed on the normal state monitoring data of the industrial robot, and a multi-modal feature set of the monitoring data of the industrial robot is constructed.

[0055] Specifically, a continuous wavelet knowledge base is constructed using existing wavelet mother functions, wherein each continuous wavelet transform has a different time-frequency analysis angle for the industrial robot monitoring data. Then, acceleration sensors are installed at key positions of the industrial robot, such as different motion joints, to collect state data. At the same time, the vibration signals collected under normal conditions are used as training data, and training samples are obtained from the time series vibration signals through sliding windows. The continuous wavelet knowledge base is used to perform multiple continuous wavelet transforms on the training samples to obtain a multi-modal feature set of the normal state data samples.

[0056] Step two, considering that the industrial robot monitoring data contains rich state representations, a monitoring data state representation extractor based on a deep network model is established.

[0057] Specifically, to further obtain the state representations contained in the industrial robot monitoring data, deep feature extraction is performed on the multi-modal feature set of the vibration data samples. A deep network model is established, and the model used in this embodiment is a Vision Transformer (ViT) network. The ViT network used includes 3 Transformer blocks and 8 multi-head attention layers, and the optimizer is an Adam method.

[0058] Step three, from the time series nature of the industrial robot monitoring data and the mutual difference between different modal features, a multi-modal feature comparison matrix is constructed using the multi-modal feature set. A comparison learning loss function that can determine the optimization direction of the network model is constructed based on the above two angles, and the network parameters are optimized using the loss function.

[0059] Specifically, considering that the vibration signals collected by the acceleration sensor have strong time series characteristics, i.e., the vibration signals collected in consecutive time have high similarity. At the same time, since different wavelet mother functions are used for continuous wavelet transform to obtain time-frequency features with different focuses, the different modal features have mutual differences. Based on the above two priors, let two consecutive samples be 、 , respectively, perform multiple continuous wavelet transforms to obtain corresponding multi-modal features such as and . Then, input into the ViT to obtain state representations and . Then, according to the two priors, a comparison relationship is constructed, and based on the similarity analysis results, a comparison matrix is designed:

[0060]

[0061] By analyzing the characteristics of the comparison matrix, three loss functions are constructed and combined as the final loss function:

[0062]

[0063] Finally, by obtaining the vibration signal sample, the ViT is optimized in combination with the designed loss function.

[0064] Step four, using the network model obtained by training to perform state representation on the multi-modal features of the real-time monitoring data of the industrial robot, and taking the multi-modal features of the normal state monitoring data as a benchmark, comparing and analyzing the state representation changes of the real-time monitoring data to perform abnormal state discrimination.

[0065] Specifically, the trained ViT is used as a state representation extractor of the real-time monitoring vibration signal of the industrial robot to obtain the state representation of the multi-modal features of the real-time vibration signal , and the state representation of the multi-modal features of the vibration signal under the normal state is taken as a benchmark. The correlation between the two state representations is analyzed, and an abnormal detection matrix is constructed:

[0066]

[0067] Each row in the abnormal detection matrix is normalized to map to the interval [-0.5, 0.5]. Set a threshold value, if there is an element in the normalized comparison matrix that exceeds the threshold value, then the real-time vibration signal of the industrial robot is abnormal at this time, and the detection result is that the running state of the industrial robot is abnormal; if none of the elements exceeds the threshold value, and the maximum value of each row is on the diagonal line, then the real-time vibration signal of the industrial robot is normal at this time, otherwise the running state of the industrial robot is abnormal.

[0068] In order to further illustrate the present application, the industrial robot fault test data is used to verify the present method. The test industrial robot includes six motion joints, each joint is realized by a connecting shaft. The test industrial robot is accelerated by loading 150% load for degradation, and in the later stage of accelerated degradation, the connecting shaft 1 appears abnormal braking action, the connecting shafts 2 and 3 have abnormal noise, and the connecting shafts 5 and 6 reducer oil leaks. In order to obtain single-axis running fault data, the servo motor of the 1-axis and 2-axis braking fault, the 3-axis fault reducer and the 5-axis fault reducer are replaced respectively, and the fault test is carried out one by one in the single-axis running state. The vibration acceleration sensor with a sampling frequency of 10 kHz is used to collect the running state data during the test. A total of 1 set of normal state data and 4 sets of fault state data are obtained. Each type of state data is sampled by sliding window, 2000 samples of normal state samples are obtained, and 1000 samples of fault state data are obtained.

[0069] Firstly, considering that the time-frequency analysis angles of continuous wavelet transform based on different mother wavelet functions are different, 15 kinds of mainstream mother wavelet functions are screened, considering that the Mexican hat function can well localize the time domain and frequency information of the input signal, the complex Gaussian B-spline function has good time-frequency smooth segmentation characteristics, and the Gaussian function has good resolution in time and frequency, the three continuous wavelet transforms have different emphases, therefore, the above three wavelet functions are used to construct the continuous wavelet transform knowledge base. Then, the ViT network model is constructed on the Pytorch platform, and the parameters are as shown in Table 1. Then, based on the two assumptions that the vibration signal samples collected under the normal state have high similarity and the samples are different after being transformed by different continuous wavelet transforms, the connected normal state samples are selected as a training sample to optimize the ViT parameters. After inputting the normal state sample, the change process of the network loss function of the ViT is as shown in Figure 2 (a), and the test accuracy in the training process is as shown in Figure 2 (b). Then, the proposed method is verified, the multi-modal time-frequency features of the fault state samples are extracted by using the continuous wavelet transform knowledge base, and the time-frequency features are represented by using the trained ViT. Finally, the comparison matrix is constructed according to the state representation of the normal state sample and the state representation of the fault state sample, and the threshold is set to 0.2, and the abnormal detection of the fault state sample is carried out, and the detection result is as shown in Table 2.

[0070] Table 1 Details of the constructed ViT parameters

[0071] Input format 128x128 Number of transformer modules 3 Number of multi-headed attentions 8 Output format 128

[0072] Table 2 Detection results of ten times of each state

[0073] Test sample Normal state Fault state 1 Fault state 2 Fault state 3 Fault state 4 Detection result (%) 100±0 98.35±0.45 98.60±0.6 100±0 100±0

[0074] As shown in Table 2, the experiment of the application obtains an accuracy of more than 97.90% in the detection verification results of 10 times of fault state, and realizes the accurate detection of the fault state. The feasibility of the method is verified.

[0075] The industrial robot autonomous anomaly detection method under zero fault samples provided by the application combines the continuous wavelet transform knowledge and the Transformer network, constructs a comparison learning loss function that can train a deep network model by the characteristics of the result industrial robot real-time monitoring data and the existing wavelet transform knowledge, and effectively breaks through the limitation of zero fault samples on the training of a deep network model. The proposed method can accurately detect various fault states.

[0076] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for autonomous fault diagnosis of an industrial robot with zero failure samples, characterized by, The method comprises the following steps: A continuous wavelet transform knowledge base is constructed according to a plurality of mother wavelet functions, multi-dimensional time-frequency analysis is performed on the normal state monitoring data of the industrial robot, and a multi-modal feature set of the normal state monitoring data of the industrial robot is constructed; A state representation extractor based on a deep network model is established, and the state representation extractor is used to perform state representation on the modal features of the normal state monitoring data of the industrial robot; Based on the time sequence and different modal features of the normal state monitoring data of the industrial robot, a multi-modal feature comparison matrix is constructed using the multi-modal feature set, a comparison learning loss function capable of determining the optimization direction of the deep network model is constructed, and the parameters of the deep network model are optimized using the comparison learning loss function; The multi-modal features of the real-time monitoring data of the industrial robot are represented using the trained deep network model, the multi-modal features of the normal state monitoring data of the industrial robot are used as a reference, and the state representation of the real-time monitoring data is compared and analyzed to determine the fault state.

2. The method according to claim 1, wherein, The plurality of mother wavelet functions include a mexican hat mother wavelet function mexh, a complex gaussian B-spline wavelet function fbsp, a morlet wavelet function, a harr wavelet function, and a gaus wavelet function.

3. The method according to claim 1, wherein, The multi-dimensional time-frequency analysis of the normal state monitoring data of the industrial robot is performed, and the multi-modal feature set of the normal state monitoring data of the industrial robot is constructed in the following specific manner: Selecting connected data samples s from normal state monitoring data of industrial robots m-1 、 m , Constructing multi-modal feature sets for the two samples respectively by using continuous wavelet transform knowledge base: wherein, with are modal features, CWT n denotes a total of n different successive wavelet transforms, m denotes the mth data sample.

4. The method according to claim 3, wherein, The state representation extractor is used to perform state representation on the modal features of the normal state monitoring data of the industrial robot in the following specific manner: wherein F() is a feature extraction operation, with w x are the results after feature extraction, respectively.

5. The method according to claim 4, wherein, The comparison matrix is constructed using the multi-modal feature set in the following manner: Wherein, M is the comparison matrix.

6. The method according to claim 5, wherein, The comparison loss function capable of determining the optimization direction of the network model is constructed in the following specific manner: w x , correlation calculation, a comparative analysis coefficient matrix M cc is obtained Wherein, cc() is a cosine similarity calculation operation; Loss function 1 is constructed based on the maximum similarity assumption: Loss function 2 is constructed based on the minimum similarity assumption: From the symmetry relation, when i = j, the matrix M cc The maximum should be taken in the i-th row, denoted as: The maximum should be taken in the i-th row, denoted as: By comparing with M cc The cross-entropy is constructed to obtain the loss function 3: wherein T r represents the rth row element in T, P r represents the rth row element in M cc and R is the number of rows in T; The loss function is constructed as follows: loss = loss2 - loss1 + loss3.

7. The method according to claim 6, wherein, The multi-modal features of the real-time monitoring data of the industrial robot are represented using the trained deep network model, the multi-modal features of the normal state monitoring data of the industrial robot are used as a reference, and the state representation of the real-time monitoring data is compared and analyzed to determine the fault state in the following specific manner: The normal state monitoring data of the industrial robot is selected as a benchmark, a multi-modal feature set is constructed, state representation is extracted, and state representation is obtained Extracting industrial robot real-time monitoring data state representation Construction of anomaly detection matrix : The anomaly detection matrix is obtained by Each row element is normalized to the interval [-0.5, 0.5], and the index of the maximum value in each row element is obtained by using the argmax operation, denoted as , and finally the detection label p x is obtained. A threshold value Threshold is designed, when satisfies , the detection result is fault, when and P x = [0, 1, …, n], the detection result is normal; wherein, abs() is to take absolute value, sum() is to sum.

Images