Robot fault diagnosis method and apparatus suitable for special dangerous environment

Abstract

Disclosed in the present application is a robot fault diagnosis method and apparatus suitable for special dangerous environments. The method comprises: performing fault prediction by using a kurtosis measurement-based fault prediction method, so as to obtain anomaly points; arranging measured vibration signals to obtain aligned target vibration signals; performing feature extraction processing on a source domain to obtain original fault features; inputting the original fault features into a preset model, and performing forward propagation on the preset model to obtain a degenerative feature layer; predicting the remaining useful lives for samples in the source domain, and extracting a highly mixed sample of the source domain and a target domain; acquiring a prediction error of the remaining useful lives of the samples in the source domain as a first training objective; acquiring a distribution difference error of the samples in the source domain and the target domain as a second training objective; and constructing an overall training objective, and performing training according to the overall training objective so as to obtain a target model. The present application can improve the accuracy of predicting the remaining useful lives for robots, and can be widely used in the technical field of robot fault diagnosis.

Description

A robot fault diagnosis method and device suitable for special dangerous environments Technical Field

[0001] The present application relates to the technical field of robot fault diagnosis, and in particular to a robot fault diagnosis method and device suitable for special dangerous environments. Background Art

[0002] Ensuring the safety and reliability of robots is crucial in hazardous environments. Robots are increasingly being used in hazardous environments for tasks such as handling hazardous materials, automated sample processing, and automated analysis. However, robots can malfunction under long periods of operation and high-intensity working conditions, which can lead to interruptions in the experimental process, data loss, or safety risks. To mitigate these potential problems, existing robot fault diagnosis methods exist to help monitor and identify robot failures and take timely measures to fix them. However, traditional data-driven methods for device fault prediction assume that the source and target domains have the same feature distribution. In reality, the devices themselves may differ, and their operating conditions may differ. This makes the fault prediction effect of models trained based on device failure data in the source domain worse or even no longer applicable in the target domain.

[0003] Summary of the Invention

[0004] In view of this, an embodiment of the present application provides a robot fault diagnosis method suitable for special hazardous environments to improve the accuracy of predicting the remaining service life of the robot.

[0005] One aspect of an embodiment of the present application provides a robot fault diagnosis method applicable to special hazardous environments, the method comprising:

[0006] Determine the machine's operating cycle in the source and target domains, and obtain the first prediction time and monitoring points for fault prediction;

[0007] Using a fault prediction method based on kurtosis measurement to predict faults at the monitoring points, and obtaining abnormal points;

[0008] Arranging the measured vibration signals according to the first predicted time and the abnormal point to obtain an aligned target vibration signal;

[0009] Performing feature extraction processing on the source domain according to the target vibration signal to obtain original fault features;

[0010] Inputting the original fault feature into a preset model, performing forward propagation on the preset model to obtain a degradation feature layer;

[0011] Predicting the remaining useful life of samples in the source domain according to the degradation characteristic layer, and extracting highly mixed samples from the source domain and the target domain;

[0012] Obtaining a remaining useful life prediction error of samples in the source domain according to the predicted remaining useful life as a first training target;

[0013] Obtaining, based on the highly mixed samples, a sample distribution difference error between the source domain and the target domain as a second training target;

[0014] An overall training objective is constructed according to the first training objective and the second training objective, and a target model is obtained by training according to the overall training objective, and fault prediction data is obtained through the target model.

[0015] Optionally, the fault prediction method using kurtosis measurement to perform fault prediction on the monitoring point to obtain an abnormal point includes:

[0016] Obtaining a mean value of the kurtosis measure and a standard deviation of the kurtosis measure;

[0017] setting a preset interval based on the standard deviation of the kurtosis measure;

[0018] According to the average value of the kurtosis measurement, monitoring points exceeding the preset interval are identified as abnormal points.

[0019] Optionally, the measured vibration signal includes a first vibration signal and a second vibration signal, and the arranging the measured vibration signals according to the first prediction time and the abnormal point to obtain the aligned target vibration signal includes:

[0020] Acquire a first vibration signal of the source domain data set and a second vibration signal of the target domain data set according to the first prediction time and the abnormal point;

[0021] The first vibration signal and the second vibration signal are arranged to obtain an aligned target vibration signal.

[0022] Optionally, the original fault feature includes a first fault feature and a second fault feature, and the performing feature extraction processing on the source domain according to the target vibration signal to obtain the original fault feature includes:

[0023] Acquire a vibration subsequence of the target vibration signal, and obtain a plurality of first fault features based on time series according to the vibration subsequence;

[0024] The second fault feature based on the time-frequency spectrum is obtained by continuous wavelet transform.

[0025] Optionally, inputting the original fault feature into a preset model and performing forward propagation on the preset model to obtain a degradation feature layer includes:

[0026] Inputting the original fault feature into a first channel and feeding it into a plurality of consecutive fully connected layers to obtain a first output feature;

[0027] Inputting the original fault feature into the second channel, and processing the original fault feature input into the second channel using multiple consecutive strided convolutional layers to obtain a second output feature;

[0028] Obtaining a third output feature of a sub-network based on a multi-layer perceptron according to the first output feature;

[0029] Flattening the second output feature to obtain a fourth output feature of a sub-network based on a deep convolutional neural network;

[0030] The third output feature and the fourth output feature are connected to form a degenerate feature layer.

[0031] Optionally, predicting the remaining useful life of samples in the source domain according to the degradation characteristic layer includes:

[0032] connecting the degradation characteristic layer and the remaining useful life prediction layer;

[0033] The remaining useful life of the samples in the source domain is predicted by using an activation function to obtain a remaining useful life percentage.

[0034] Optionally, obtaining a remaining useful life prediction error of samples in the source domain based on the predicted remaining useful life as a first training target includes:

[0035] Obtaining a batch size of samples in the source domain, a marked remaining useful life percentage value of the samples in the source domain, and a predicted remaining useful life percentage value of the samples in the source domain;

[0036] The difference between the marked remaining useful life percentage value and the predicted remaining useful life percentage value is projected into the reproducing kernel Hilbert space, and the remaining useful life prediction error is obtained in combination with the batch size of the samples in the source domain.

[0037] Optionally, obtaining, based on the highly mixed samples, a sample distribution difference error between the source domain and the target domain as a second training objective includes:

[0038] The distribution difference of the highly mixed samples is measured by a multi-kernel maximum mean difference method to obtain the sample distribution difference error.

[0039] The present application also provides a robot fault diagnosis device suitable for use in special hazardous environments, including:

[0040] The first module is used to determine the machine's operating cycle in the source domain and the target domain, and obtain the first prediction time and monitoring points for fault prediction;

[0041] The second module is used to predict the faults of the monitoring points by using a fault prediction method based on kurtosis measurement to obtain abnormal points;

[0042] A third module is configured to arrange the measured vibration signals according to the first prediction time and the abnormal point to obtain an aligned target vibration signal;

[0043] A fourth module is configured to perform feature extraction processing on the source domain according to the target vibration signal to obtain original fault features;

[0044] A fifth module is configured to input the original fault features into a preset model, perform forward propagation on the preset model, and obtain a degradation feature layer;

[0045] A sixth module is configured to predict the remaining useful life of samples in the source domain according to the degradation characteristic layer, and extract highly mixed samples from the source domain and the target domain;

[0046] A seventh module is configured to obtain a remaining useful life prediction error of samples in the source domain according to the predicted remaining useful life, as a first training target;

[0047] An eighth module is configured to obtain, based on the highly mixed samples, a sample distribution difference error between the source domain and the target domain as a second training target;

[0048] The ninth module is used to construct an overall training objective based on the first training objective and the second training objective, and to obtain a target model through training according to the overall training objective, and to obtain fault prediction data through the target model.

[0049] An embodiment of the present application also provides an electronic device, which includes a processor and a memory; the memory stores a program; the processor executes the program to execute the aforementioned robot fault diagnosis method suitable for special hazardous environments; the electronic device has the function of carrying and running the business data processing software system provided in the embodiment of the present application, such as a personal computer (PC), a mobile phone, a smart phone, a personal digital assistant (PDA), a wearable device, a handheld computer PPC (Pocket PC), a tablet computer, a vehicle-mounted terminal, etc.

[0050] An embodiment of the present application also provides a computer-readable storage medium, which stores a program, and the program is executed by a processor to implement the aforementioned robot fault diagnosis method applicable to special dangerous environments.

[0051] The present application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned robot fault diagnosis method applicable to special hazardous environments.

[0052] An embodiment of the present application determines the operating cycle of a machine in a source domain and a target domain, obtains the first prediction time and monitoring points for fault prediction; uses a fault prediction method based on kurtosis measurement to perform fault prediction on the monitoring points to obtain abnormal points; arranges the measured vibration signals according to the first prediction time and the abnormal points to obtain aligned target vibration signals; performs feature extraction processing on the source domain according to the target vibration signals to obtain original fault features; inputs the original fault features into a preset model, and performs forward propagation on the preset model to obtain a degradation feature layer; based on the degradation feature layer, predicts the remaining service life of samples in the source domain, and extracts highly mixed samples from the source domain and the target domain; based on the predicted remaining service life, obtains the remaining service life prediction error of the samples in the source domain as a first training target; based on the highly mixed samples, obtains the sample distribution difference error of the source domain and the target domain as a second training target; constructs an overall training target based on the first training target and the second training target, and obtains a target model through training based on the overall training target, and obtains fault prediction data through the target model. The present application obtains a target model through training, and based on the target model, learns domain-invariant degradation features, effectively reduces the distribution difference between the source domain and the target domain, realizes fault prediction, and ensures the accuracy of the remaining service life prediction in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0054] FIG1 is a schematic diagram of the steps of a fault prediction method according to an embodiment of the present application;

[0055] FIG2 is a schematic diagram of the operation procedure of the vibration signal arrangement according to an embodiment of the present application;

[0056] FIG3 is a schematic diagram of the improved domain adaptive Bayesian dual network structure of an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0058] Before describing the embodiments of the present application in detail, it is necessary to explain the terms that may be involved in the embodiments of the present application:

[0059] Remaining Useful Life (RUL) refers to the remaining service life of a system after it has been running for a period of time.

[0060] Remaining Useful Life Percentage (RULP) refers to the percentage of the remaining useful life of a system after it has been running for a period of time to the total life of the system.

[0061] Monte Carlo (MC) refers to a family of statistical methods that are essentially used to find solutions to problems, using randomly drawn samples to calculate the solution to a given problem.

[0062] Variational inference (VI) is a deterministic approximate inference method. Approximate inference is a method used to approximate a computationally complex distribution or at least obtain some statistics of the target distribution.

[0063] A confidence interval (CI) is an estimated interval for a population parameter constructed from a sample statistic. In statistics, a confidence interval for a probability sample is an interval estimate of a population parameter for that sample.

[0064] Continuous wavelet transform (CWT) refers to selecting a center frequency, obtaining some center frequencies through scaling transformation, and then obtaining a series of basis functions in different intervals through time shift. The products are respectively multiplied with a certain section of the original signal (corresponding to the interval of the basis function) and then integrated. The frequency corresponding to the generated extreme value is the frequency contained in this interval of the original signal.

[0065] Deep learning (DL) is a new research direction in machine learning. It studies the inherent patterns and representational hierarchies of sample data. Its ultimate goal is to enable machines to have human-like analytical and learning capabilities, enabling them to recognize data such as text, images, and sounds. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition that far surpass previous related technologies.

[0066] The Rectified Linear Unit (ReLU), also known as the rectified linear unit, is an activation function in deep learning. This function is mainly used in the convolution and pooling of CNN convolutional neural networks. The values ​​of the image matrix after convolution and pooling operations are both positive and negative. After passing through the ReLU function, only positive values ​​are left in the image matrix.

[0067] The Multilayer Perceptron (MLP) is a classic neural network model consisting of an input layer, hidden layers, and an output layer. The input layer receives input data, the hidden layers learn feature representations, and the output layer generates the final predictions. Each neuron in the hidden and output layers has an activation function to introduce nonlinear mapping. Its structure and functionality make it a crucial component in deep learning.

[0068] A Deep Convolutional Neural Network (DCNN) is a computational model composed of multiple layers of neurons designed to mimic the way the human brain processes images. "Depth" refers to the diversity of layers in the network, while "convolution" refers to a specific way neurons are connected that effectively captures local features of an image.

[0069] Maximum Mean Discrepancy (MMD) is the most widely used loss function in transfer learning, especially in domain adaptation. It is mainly used to measure the distance between the distributions of two different but related random variables.

[0070] Multiple Kernel Maximum Mean Discrepancy (MK-MMD) is developed based on the original MMD and proposes to use multiple kernels to construct a total kernel.

[0071] Reproducing kernel Hilbert space (RKHS) is a Hilbert space with the reproducing property. Among them, the complete inner product space is called Hilbert space.

[0072] Bias is the error between the predicted value and the true value.

[0073] The first prediction time (FPT) refers to the time when the system first predicts a machine failure.

[0074] There are existing robot fault diagnosis methods to help monitor and identify robot failures and take timely measures to repair problems. However, because traditional data-driven methods for equipment fault prediction assume that the source domain and target domain have the same feature distribution, while in reality the equipment itself may have differences and operating conditions, the fault prediction effect of the model trained based on equipment failure data in the source domain becomes worse or even no longer applicable in the target domain.

[0075] In response to the problems existing in the prior art, with reference to FIG1 , the present application provides a robot fault diagnosis method applicable to special hazardous environments, comprising:

[0076] S100: Determine the operation cycle of the machine in the source domain and the target domain, and obtain the first prediction time and monitoring point of the fault prediction;

[0077] Specifically, as shown in Figure 2, the first prediction time of the machine in the source domain is obtained, the sampling frequency of the monitoring point is 25.6kHz, the sampling time is 0.1s, and the sampling interval is 10s; the first prediction time of the machine in the target domain is obtained, the sampling frequency of the monitoring point is 25.6kHz, the sampling time is 1.28s, and the sampling interval is 60s. Among them, the PHM2012 dataset is a public dataset for bearing remaining life prediction. This dataset is obtained by conducting controlled experiments on a type of bearing widely used in mechanical systems. The XJTU-SY dataset contains the full life cycle vibration signals of 15 rolling bearings under three working conditions. EoL (End of Life) represents the end of life time.

[0078] S200, using a fault prediction method based on kurtosis measurement to perform fault prediction on the monitoring point to obtain an abnormal point;

[0079] Optionally, in some embodiments, step S200 specifically includes the following steps:

[0080] S210, obtaining the average value of the kurtosis measurement and the standard deviation of the kurtosis measurement;

[0081] S220, setting a preset interval according to the standard deviation of the kurtosis measurement;

[0082] S230, identifying monitoring points exceeding the preset interval as abnormal points based on the average value of the kurtosis measurement;

[0083] Specifically, the set of outliers is expressed as: {|k t+j |-μ>3σ} j＝1,2,3,4

[0084] Where μ represents the mean of the kurtosis measure; σ represents the standard deviation of the kurtosis measure; k represents the kurtosis measure obtained from the original vibration subsequence; t represents the first prediction time for triggering the fault prediction algorithm; and j represents the sequence index of the original vibration subsequence.

[0085] S300, arranging the measured vibration signals according to the first prediction time and the abnormal point to obtain an aligned target vibration signal;

[0086] Optionally, in some embodiments, step S300 specifically includes the following steps:

[0087] S310: Acquire a first vibration signal of the source domain data set and a second vibration signal of the target domain data set according to the first prediction time and the abnormal point;

[0088] S320: Arrange the first vibration signal and the second vibration signal to obtain an aligned target vibration signal;

[0089] Specifically, the measured vibration signal includes a first vibration signal and a second vibration signal. The first vibration signal and the second vibration signal are arranged, that is, the vibration signals are aligned to maintain the same sampling duration and interval of the measured vibration signals in the source domain and the target domain, so as to characterize the degradation information of different bearings in the source domain and the target domain within the same scale and granularity, and overcome the adverse effect of extracting original fault features from vibration signals of different machines across domains on the accuracy of the final transmission fault prediction.

[0090] S400: performing feature extraction processing on the source domain according to the target vibration signal to obtain original fault features;

[0091] Optionally, in some embodiments, step S400 specifically includes the following steps:

[0092] S410: Acquire a vibration subsequence of the target vibration signal, and obtain a plurality of first fault features based on time series according to the vibration subsequence;

[0093] S420, obtaining the second fault feature based on the time-frequency spectrum through continuous wavelet transform;

[0094] Specifically, the original fault feature includes a first fault feature and a second fault feature. The source domain includes a plurality of samples. The samples include the first fault feature and the second fault feature of dual-channel input. The expression of the source domain is:

[0095] Where,

[0096] In particular,

[0097] Among them, D S represents the source domain; χ S The feature space representing the source domain; Samples representing the source domain; represents the remaining useful life value of the marker in the source domain, corresponding to the ith target vibration signal alignment subsequence in the source domain; y S Represents the data space where the remaining service life value of the mark in the source domain is located; i represents the sequence number of the target vibration signal alignment subsequence; N S The number of samples representing the source domain; represents the first fault feature set; Represents 25 first fault features based on time series; Represents the second fault characteristic;

[0098] In addition, the expression of the target domain is:

[0099] Where,

[0100] Among them, D T represents the target domain; The sample representing the target domain; T Represents the feature space of the target domain; N T represents the number of samples in the target domain; i represents the sample number;

[0101] The remaining useful life value marked in the target domain is not available. We can use the knowledge learned from the source domain in a supervised manner through the prediction function to estimate the remaining useful life value corresponding to the test sample in the target domain. The remaining useful life value corresponding to the test sample in the target domain is:

[0102] in, Test samples representing the target domain; represents the remaining useful life value corresponding to the test sample in the target domain; f(·) represents the prediction function.

[0103] S500, inputting the original fault feature into a preset model, performing forward propagation on the preset model, and obtaining a degradation feature layer;

[0104] Specifically, the preset model is a Bayesian dual neural network. As shown in Figure 3, the Bayesian dual neural network contains three modules, namely feature extraction, RUL prediction and domain adaptation; the feature extraction module automatically extracts discriminative high-level fusion degradation features from labeled source domain samples and unlabeled target domain samples, namely highly mixed samples; then, the learned high-level fusion degradation features are further fed into the RUL prediction module constructed by the linear regression layer to predict the remaining useful life of the samples; finally, the distribution difference of the high-level fusion features between the source domain and the target domain is calculated in the domain adaptation module; in addition, Monte Carlo dropout can be implemented in both the offline model training stage and the online model testing stage; then, multiple predictions with Monte Carlo dropout enabled in the online stage can be obtained based on the Bernoulli distribution, and the uncertainty of the model prediction can also be quantified from the posterior distribution;

[0105] Optionally, in some embodiments, step S500 specifically includes the following steps:

[0106] S510: Input the original fault feature into a first channel and feed it into a plurality of consecutive fully connected layers to obtain a first output feature;

[0107] S520: Input the original fault feature into the second channel, and process the original fault feature input into the second channel using multiple consecutive strided convolutional layers to obtain a second output feature;

[0108] S530: Obtain a third output feature of a sub-network based on a multi-layer perceptron according to the first output feature;

[0109] S540: performing a flattening process on the second output feature to obtain a fourth output feature of a sub-network based on a deep convolutional neural network;

[0110] S550, connecting the third output feature and the fourth output feature to form a degradation feature layer;

[0111] Specifically, the original fault feature is input into the first channel and fed into three consecutive fully connected layers to obtain the first output feature. The three consecutive fully connected layers are FC1, FC2 and FC3 respectively. The expression of the first output feature is:

[0112] Where, X 1D =(x S_1D ,x T_1D )

[0113] Among them, r and r-1 represent the number of layers; W FC Represents the weight matrix connecting layer r and layer r-1; bFC Represents the bias vector connecting layer r and layer r-1; Represents the rth layer of input data; represents the r-1th layer of input data; F(·) represents the hyperbolic tangent function; X 1D Represents the first channel of input data; x S_1D Represents the first channel of the source domain data; x T_1D The first channel representing the target domain data; Represents the third output feature of the multilayer perceptron-based subnetwork; Represents the 4th layer of input data;

[0114] In addition, three consecutive strided convolutional layers are used to process the original fault feature of the second input channel to obtain a second output feature, and the second output feature is flattened to obtain a fourth output feature of the subnetwork based on the deep convolutional neural network. The three consecutive strided convolutional layers are Conv1, Conv2 and Conv3, respectively. The expression of the second output feature is:

[0115] The expression of the fourth output feature is:

[0116] in, Represents the j-th output feature map of the r-th layer; represents the i-th input feature of the r-1th layer; j represents the feature map number; i represents the feature number of the input data in the convolutional layer; Represents the weight matrix of the convolution kernel; represents the bias vector of the convolution kernel; * represents the strided convolution operation; g(·) represents the linear rectification function; Represents the fourth output feature of the sub-network based on deep convolutional neural network; represents the second channel output feature map of the third strided convolutional layer; flatten(·) represents the flattening operator; C3 represents the number of convolution kernels in Conv3;

[0117] Connect the third output feature of the sub-network based on the multi-layer perceptron and the fourth output feature of the sub-network based on the deep convolutional neural network to form a degraded feature layer, then we have:

[0118] Among them, FL represents the degenerate feature layer; h FL Represents the high-level fusion degraded feature layer formed by connecting the output features of the MLP-based and DCNN-based dual sub-networks.

[0119] In some optional embodiments, a remaining useful life prediction model based on a deep learning model is trained using training samples from the source domain and the remaining useful life values ​​labeled in the training samples from the source domain, and then combined with the training samples from the target domain to obtain a generalized remaining useful life prediction model; wherein, by learning domain-invariant degradation features based on the deep learning model, the distribution difference between the source domain and the target domain is effectively reduced, transmission fault prediction is achieved, and the accuracy of the remaining useful life prediction model in the target domain is improved;

[0120] The remaining useful life prediction model based on the deep learning model is:

[0121] The generalized remaining useful life prediction model is:

[0122] Where f(·) represents the prediction function to be learned; Training samples representing the source domain; Training samples representing the target domain; represents the remaining useful life value corresponding to the training sample in the source domain; Θ represents the parameter set; Θ * Represents the updated optimal parameter set to be learned;

[0123] In addition, the traditional deep learning model-based transmission fault prediction method under different operating conditions can be formulated as various deep learning model-based regression models; the predicted remaining useful life is a deterministic continuous value and cannot represent the confidence interval of its corresponding prediction; and the improved Bayesian dual network with domain adaptation is basically a probabilistic extension of the traditional deep learning model; the parameter set is set as a random variable on the probability distribution rather than a fixed value through Monte Carlo dropout; compared with random dropout enabled only in the training phase, Monte Carlo dropout is used to expand the traditional neural network into a Bayesian neural network in both training and testing phases.

[0124] S600: Predicting the remaining useful life of samples in the source domain according to the degradation characteristic layer, and extracting highly mixed samples from the source domain and the target domain;

[0125] Optionally, in some embodiments, step S600 specifically includes the following steps:

[0126] S610, connecting the degradation characteristic layer and the remaining service life prediction layer;

[0127] S620: Predict the remaining useful life of the samples in the source domain using an activation function to obtain a remaining useful life percentage;

[0128] Specifically, the expression for the remaining useful life percentage is:

[0129] Among them, W R b represents the weight matrix connecting the degradation feature layer and the remaining service life prediction layer; R represents the deviation vector connecting the degradation feature layer and the remaining service life prediction layer; σ(·) represents the sigmoid function; h FL Represents the high-level fusion degradation feature layer formed by connecting the output features of the MLP-based and DCNN-based dual sub-networks; Represents the predicted remaining useful life in percentage,

[0130] In some optional embodiments, as shown in FIG3 , in the domain adaptation module, high-level fusion degradation features (denoted as FL_S and FL_T) in the source domain and the target domain are extracted and denoted as h FL_S and h FL_T .h FL_S and h FL_T The difference in distribution between is measured by the MK-MMD metric, which is a multi-core variant of the classic MMD metric. The square formula of the MMD metric is defined as:

[0131] where ||·|| represents the reproducing kernel Hilbert space, N S and N T are the number of samples in the source domain and the target domain respectively; ψ(·):h FL_S ,h FL_T →H is the feature map function, which can project the degenerate features of high-level fusion into the reproducing kernel Hilbert space. In order to simplify the tedious calculation, the kernel trick can be used to calculate the feature map in the reproducing kernel Hilbert space, which is described as follows:

[0132] Where <·,·> is the inner product operator; ψ(·) represents the feature map in the reproducing kernel Hilbert space; K(·,·) represents the feature kernel associated with the feature map in the reproducing kernel Hilbert space;

[0133] MK-MMD is defined as:

[0134] Among them, K r (·,·),r=1,…,m represent different kernels, which are chosen as Gaussian kernels with different bandwidths.

[0135] S700: Obtaining a remaining useful life prediction error of samples in the source domain according to the predicted remaining useful life as a first training target;

[0136] Optionally, in some embodiments, step S700 specifically includes the following steps:

[0137] S710: Obtain a batch size of samples in the source domain, a marked remaining useful life percentage value of the samples in the source domain, and a predicted remaining useful life percentage value of the samples in the source domain;

[0138] S720: Projecting the difference between the marked remaining useful life percentage value and the predicted remaining useful life percentage value into the reproducing kernel Hilbert space, and combining the batch size of the source domain samples to obtain the remaining useful life prediction error;

[0139] Specifically, the remaining useful life prediction error expression is:

[0140] Where, Θ=(θ f ,θ p ) g(Θ,ε)={diag(ε)θ f ,θ p}

[0141] Among them, N batch represents the batch size of samples from the source domain; y i Represents the remaining service life of the marker; l p The remaining useful life prediction error term represents the difference between the labeled remaining useful life value and the predicted remaining useful life value of the sample in the source domain; represents the predicted output of the Bayesian dual network with Monte Carlo dropout for the remaining useful life; Θ represents the parameter set of the feature extractor and the remaining useful life predictor; θ f represents the parameter set of the feature extractor; θ p represents the parameter set of the remaining useful life predictor; g(Θ,ε)={diag(ε)θ f ,θ p} and diag(ε) represent diagonal matrices, whose diagonal elements are binary vectors ε~q(ε), q(ε) is the Bernoulli distribution product related to the discarding method; the feature extractor is used to obtain fault features from the robot, and the remaining service life predictor is used to predict the remaining service life of the machine based on the fault features.

[0142] S800: Obtain, based on the highly mixed samples, a sample distribution difference error between the source domain and the target domain as a second training target;

[0143] Optionally, in some embodiments, step S800 specifically includes the following steps:

[0144] S810 , measuring the distribution difference of the highly mixed samples by a multi-kernel maximum mean difference method to obtain the sample distribution difference error.

[0145] Specifically, the expression of the distribution difference error is:

[0146] In order to reduce the complexity of the equation calculation and improve the training efficiency of the small batch stochastic gradient descent algorithm, an unbiased estimate of the equation is proposed:

[0147] Where K(·,·) represents a multi-core operator; the batch sizes of the source and target domains are as well as Represents high-level fused degradation features in the source domain; as well as represents the high-level fusion degradation features in the target domain; κ r (·,·), r=1,…,m represents different kernels; j represents the target domain feature map number; i represents the source domain feature map number.

[0148] S900: constructing an overall training objective according to the first training objective and the second training objective, obtaining a target model through training according to the overall training objective, and obtaining fault prediction data through the target model;

[0149] Specifically, the overall training objective expression is: total =l p +λl d

[0150] Among them, l total Represents the total loss value, that is, the total optimization goal; λ is a hyperparameter that balances the trade-off between the remaining useful life prediction error term and the distribution difference error term;

[0151] By minimizing l total To optimize the parameter set of the preset model, the final expression of the overall training objective is: total (Θ)=min l p (θ f ,θ p )+λl d (θ f )

[0152] Then, the Adam optimizer is used to update the parameter set, and the updated parameter set is:

[0153] Wherein, η represents the learning rate; a loop termination criterion is preset to determine whether the loop termination criterion is met. If the loop termination criterion is not met, the process returns to the step of inputting the original fault feature into the preset model, forward propagating the preset model, and obtaining the degradation feature layer until the loop termination criterion is triggered, and the domain-invariant degradation features of the machine in the source domain and the target domain are learned; the loop termination criterion is generally that the change in the optimized variable is small enough. In practice, it is generally necessary to specify a sufficiently small number in advance, and the loop is terminated when the difference before and after the variable update is less than this amount.

[0154] Referring to Figure 1, during the online prediction stage, the system can use the trained target model to perform fault prediction on the vibration signal of the target domain to obtain fault prediction data.

[0155] An embodiment of the present application also provides an electronic device, which includes a processor and a memory; the memory stores a program; the processor executes the program to execute the aforementioned robot fault diagnosis method suitable for special hazardous environments; the electronic device has the function of carrying and running the business data processing software system provided in the embodiment of the present application, such as a personal computer (PC), a mobile phone, a smart phone, a personal digital assistant (PDA), a wearable device, a handheld computer PPC (Pocket PC), a tablet computer, a vehicle-mounted terminal, etc.

[0156] An embodiment of the present application also provides a computer-readable storage medium, which stores a program, and the program is executed by a processor to implement the aforementioned robot fault diagnosis method applicable to special dangerous environments.

[0157] The present application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned robot fault diagnosis method applicable to special hazardous environments.

[0158] In summary, the robot fault diagnosis method and device applicable to special hazardous environments according to the embodiments of the present application have the following advantages:

[0159] 1. This application uses a deep learning model to learn domain-invariant degradation features, which can effectively reduce the distribution difference between the source domain and the target domain, realize the prediction of transmission failures, and improve the accuracy of remaining service life prediction.

[0160] 2. This application is also applicable to robot fault diagnosis in special hazardous environments, improving the reliability, safety and operational efficiency of robots in special hazardous environments.

[0161] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flow chart of the present application are provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0162] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present application. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the routine skills of an engineer. Therefore, a person skilled in the art can implement the present application as set forth in the claims using ordinary techniques without undue experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0163] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, 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. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. 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.

[0164] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0165] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0166] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0167] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0168] Although the embodiments of the present application have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and intent of the present application, and that the scope of the present application is defined by the claims and their equivalents.

[0169] The above is a specific description of the preferred implementation of the present application, but the present application is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present application, and these equivalent modifications or substitutions are all included in the scope defined by the claims of the present application.

Claims

1. A robot fault diagnosis method suitable for special dangerous environments, characterized in that: include: Determine the operation cycle of the machine in the source domain and the target domain, and obtain the first prediction time and monitoring points for fault prediction; Using a fault prediction method based on kurtosis measurement to predict faults at the monitoring points, and obtaining abnormal points; Arranging the measured vibration signals according to the first prediction time and the abnormal point to obtain an aligned target vibration signal; According to the target vibration signal, feature extraction processing is performed on the source domain to obtain original fault features; Inputting the original fault feature into a preset model, performing forward propagation on the preset model, and obtaining a degradation feature layer; According to the degradation characteristic layer, the remaining service life of the samples in the source domain is predicted, and highly mixed samples of the source domain and the target domain are extracted; According to the predicted remaining useful life, obtaining a remaining useful life prediction error of the sample in the source domain as a first training target; According to the highly mixed samples, obtaining a sample distribution difference error between the source domain and the target domain as a second training target; An overall training objective is constructed according to the first training objective and the second training objective, and a target model is obtained by training according to the overall training objective, and fault prediction data is obtained through the target model.

2. A robot fault diagnosis method suitable for special dangerous environments according to claim 1, characterized in that: The fault prediction method using kurtosis measurement performs fault prediction on the monitoring point to obtain an abnormal point, including: Obtaining a mean value of the kurtosis measure and a standard deviation of the kurtosis measure; Setting a preset interval according to a standard deviation of the kurtosis measure; According to the average value of the kurtosis measurement, monitoring points exceeding the preset interval are identified as abnormal points.

3. A robot fault diagnosis method suitable for special dangerous environments according to claim 1, characterized in that: The measured vibration signal includes a first vibration signal and a second vibration signal, and the measured vibration signal is arranged according to the first prediction time and the abnormal point to obtain an aligned target vibration signal, including: Acquire a first vibration signal of the data set of the source domain and a second vibration signal of the data set of the target domain according to the first prediction time and the abnormal point; The first vibration signal and the second vibration signal are arranged to obtain an aligned target vibration signal.

4. A robot fault diagnosis method suitable for special dangerous environments according to claim 1, characterized in that: The original fault feature includes a first fault feature and a second fault feature. The feature extraction process is performed on the source domain according to the target vibration signal to obtain the original fault feature, including: Obtain a vibration subsequence of the target vibration signal, and obtain a plurality of time series-based the first fault characteristic; The second fault feature based on the time-frequency spectrum is obtained by continuous wavelet transform.

5. A robot fault diagnosis method suitable for special dangerous environments according to claim 1, characterized in that: The step of inputting the original fault feature into a preset model and forward-propagating the preset model to obtain a degradation feature layer includes: Input the original fault feature into a first channel, feed it into a plurality of consecutive fully connected layers, and obtain a first output feature; Inputting the original fault feature into the second channel, and processing the original fault feature input into the second channel using a plurality of continuous strided convolutional layers to obtain a second output feature; According to the first output feature, obtaining a third output feature of a sub-network based on a multi-layer perceptron; Flattening the second output feature to obtain a fourth output feature of a sub-network based on a deep convolutional neural network; The third output feature and the fourth output feature are connected to form a degradation feature layer.

6. A robot fault diagnosis method suitable for special dangerous environments according to claim 1, characterized in that: The predicting of the remaining service life of the samples in the source domain according to the degradation characteristic layer includes: connecting the degradation characteristic layer and the remaining useful life prediction layer; The remaining useful life of the samples in the source domain is predicted by using an activation function to obtain a remaining useful life percentage.

7. A robot fault diagnosis method suitable for special dangerous environments according to claim 6, characterized in that: The obtaining, according to the predicted remaining useful life, a remaining useful life prediction error of the sample in the source domain as a first training target, includes: Obtaining a batch size of samples in the source domain, a marked remaining useful life percentage value of the samples in the source domain, and a predicted remaining useful life percentage value of the samples in the source domain; The difference between the marked remaining useful life percentage value and the predicted remaining useful life percentage value is projected into the reproducing kernel Hilbert space, and combined with the batch size of the samples in the source domain, the remaining useful life prediction error is obtained.

8. The robot fault diagnosis method suitable for special dangerous environments according to claim 1, characterized in that: The step of obtaining, according to the highly mixed samples, a sample distribution difference error between the source domain and the target domain as a second training target includes: The distribution difference of the highly mixed samples is measured by a multi-kernel maximum mean difference method to obtain the sample distribution difference error.

9. A robot fault diagnosis device suitable for special dangerous environments, characterized in that: include: The first module is used to determine the operation cycle of the machine in the source domain and the target domain, and obtain the first prediction time and monitoring points of fault prediction; The second module is used to predict the faults of the monitoring points by using the fault prediction method of kurtosis measurement to obtain abnormal points; A third module is used to arrange the measured vibration signals according to the first prediction time and the abnormal point to obtain an aligned target vibration signal; The fourth module is used to perform feature extraction processing on the source domain according to the target vibration signal to obtain original fault features; A fifth module is used to input the original fault feature into a preset model, perform forward propagation on the preset model, and obtain a degradation feature layer; A sixth module is used to predict the remaining useful life of the samples in the source domain according to the degradation characteristic layer, and extract highly mixed samples of the source domain and the target domain; A seventh module is used to obtain a remaining useful life prediction error of the sample in the source domain according to the predicted remaining useful life as a first training target; An eighth module is used to obtain, according to the highly mixed samples, a sample distribution difference error between the source domain and the target domain as a second training target; The ninth module is used to construct an overall training target according to the first training target and the second training target, and obtain a target model through training according to the overall training target, and obtain fault prediction data through the target model.

10. An electronic device, characterized in that: including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method according to any one of claims 1 to 8.

Images