Search

CN-121997240-A - Industrial robot joint abnormal vibration detection method and system based on deep learning

CN121997240ACN 121997240 ACN121997240 ACN 121997240ACN-121997240-A

Abstract

The invention discloses an industrial robot joint abnormal vibration detection method and system based on deep learning, and relates to the technical field of robot fault diagnosis. The method comprises the steps of collecting robot joint vibration signals, constructing a training data set, obtaining an enhanced sample through autocorrelation analysis and decoupling of a projection matrix, constructing a deformable convolution feature extraction network based on a deformable convolution block and physical phase codes, obtaining a deep time sequence feature sequence, obtaining a classification result of the deep time sequence feature sequence through a time sequence classification head, constructing a loss function based on cross entropy classification loss and dynamic time regularity regularization items, carrying out back propagation on the model, and updating parameters. The method and the device can improve the visibility of the abnormal characteristics in the input process, enable the model to accurately capture the local abnormal waveform related to the rotation phase in the period, and improve the accuracy of model abnormal detection.

Inventors

  • XU DEYONG
  • SUN XIAOYAN
  • YANG ZHICHENG

Assignees

  • 青岛浩海网络科技股份有限公司

Dates

Publication Date
20260508
Application Date
20260408

Claims (10)

  1. 1. The method for detecting abnormal vibration of the joint of the industrial robot based on deep learning is characterized by comprising the following steps of: S1, acquiring an original vibration signal in the running process of an industrial robot, wherein the original vibration signal comprises an acceleration signal and an angular velocity signal to form a training data set with class labels; S2, estimating a dominant rotation period of each sample through autocorrelation analysis, and mapping each sample to a uniform standardized period length through interpolation resampling to obtain a vibration signal after phase alignment; constructing a decoupling projection matrix based on the instantaneous angular velocity, decomposing the vibration signal after phase alignment into a dynamic component and a residual vibration component, applying self-adaptive enhancement to the residual vibration component and splicing the residual vibration component with the dynamic component to obtain an enhanced sample; S3, constructing a deformable convolution feature extraction network fused with a physical phase code, inputting the enhanced sample into the deformable convolution feature extraction network, and extracting a deep time sequence feature sequence, wherein the physical phase code is generated based on a sample corresponding angular velocity signal and is used for guiding sampling offset prediction of deformable convolution; s4, constructing a time sequence classification head based on one-dimensional convolution and global average pooling, inputting a deep time sequence feature sequence into the time sequence classification head, compressing and mapping to obtain a predicted class probability distribution; S5, constructing a similar feature consistency constraint based on dynamic time warping, inputting a deep time sequence feature sequence to obtain a dynamic time warping regular term, constructing a total loss function based on cross entropy classification loss and the dynamic time warping regular term, and executing back propagation and parameter updating on the model by using the total loss function to obtain a trained detection model; s6, inputting a detection model after training is completed after preprocessing and characteristic enhancement of the vibration signal to be detected, which is acquired in real time, in the step S2, and outputting an abnormal vibration detection result.
  2. 2. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 1, wherein S2 specifically comprises: S201, acquiring an original vibration signal of a sample, and executing average value removal processing on each characteristic channel of the original vibration signal; S202, presetting a delay search interval, calculating the similarity between a sequence after mean removal and a delay sequence of each candidate delay amount in the delay search interval to obtain an autocorrelation function, taking the corresponding delay amount when the autocorrelation function takes the maximum value as the estimated rotation period length of the sample, and intercepting a complete period section from an original vibration signal according to the estimated rotation period length; s203, mapping each complete period segment to a unified standardized time axis to obtain vibration signals with aligned phases; S204, dividing the vibration signals with the aligned phases into acceleration sub-signals and angular velocity sub-signals according to the sensor types, wherein the acceleration sub-signals comprise three-axis acceleration signals and three-axis gyroscope signals which are used as six-dimensional input vectors; s205, constructing a decoupling projection matrix based on the instantaneous angular velocity vector, performing projection operation on the six-dimensional input vector by using the decoupling projection matrix, extracting dynamic components dominated by rigid rotation from the mixed signal, and separating residual vibration components from the input signal after phase alignment by using a complementary relation; S206, generating an adaptive gain factor corresponding to time according to the ratio between the energy of the dynamic component and the energy of the residual vibration component, wherein the adaptive gain factor is specifically expressed as follows: in the formula, Representing the adaptive gain factor(s) of the adaptive gain, The enhanced intensity coefficient is used for controlling the gain variation amplitude, and the value is 2.0; Representing kinetic components Energy of (2); Representing residual vibration components Is a function of the energy of the (c), Representing a numerical stability term; enhancing the residual vibration component time by utilizing the self-adaptive gain factor to obtain an enhanced residual vibration component; S207, traversing all time indexes in a standardized period to obtain an enhanced residual vibration component sequence, and splicing the dynamic component sequence and the enhanced residual vibration component sequence along a characteristic dimension to obtain an enhanced sample.
  3. 3. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 1, wherein S1 specifically comprises: S101, the industrial robot moves according to a preset periodic movement track, continuous signals are respectively collected for each operation state, the continuous signals are truncated in a sliding window mode, and a plurality of samples with fixed time window lengths are constructed; S102, according to the actual running state corresponding to the samples, assigning a category label to each sample, wherein the category label is specifically marked as one of four categories of normal, slight abrasion, serious abrasion or poor lubrication; S103, all samples with category labels jointly form a training data set.
  4. 4. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 1, wherein S3 specifically comprises: S301, a deformable convolution block simultaneously comprises an offset prediction module, a deformable convolution, a nonlinear activation and normalization processing feature extraction unit, and a plurality of deformable convolution blocks are sequentially stacked to form a deformable convolution feature extraction network; s302, before the input feature diagram is input into the deformable convolution block, constructing a physical phase coding sequence matched with the time length of the input feature diagram; S303, splicing the input feature map and the physical phase coding sequence along the channel dimension to obtain fusion features; s304, a fusion characteristic input offset prediction module obtains an offset set, and deformable convolution is carried out on an input characteristic diagram according to the offset set; S305, inputting the output obtained by the deformable convolution into a nonlinear activation layer and a batch normalization layer in sequence to obtain the output of the deformable convolution block; S306, repeating the process until all deformable convolution blocks are processed, and obtaining the deep time sequence characteristic sequence.
  5. 5. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 3, wherein the physical phase coding sequence construction comprises: S311, extracting a triaxial angular velocity sequence from an original angular velocity channel corresponding to the enhanced sample, calculating an angular velocity module length, and carrying out accumulated summation on the angular velocity module length from the starting point of the standardized period to the current moment to obtain accumulated angular displacement; s312, performing full-period accumulation on the angular velocity modular length in the whole standardized period to obtain total angular displacement, wherein the total angular displacement is expressed as: in the formula, Representing the total angular displacement over a standard period, Representing the angular velocity module length at each instant The sampling interval is represented by the number of samples, Representing the length of the modulo length sequence; s313, dividing the accumulated angular displacement at the current moment by the total angular displacement to obtain a normalized rotation progress, and respectively taking a sine value and a cosine value of the normalized rotation progress to form a two-dimensional physical phase code vector; S314, arranging the two-dimensional physical phase code vectors corresponding to all the time indexes in time sequence to obtain a physical phase code sequence.
  6. 6. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 1, wherein S4 specifically comprises: S401, receiving a deep time sequence feature sequence by a time sequence classification head, applying one-dimensional convolution processing to the deep time sequence feature sequence, and integrating context information; S402, performing global average pooling on one-dimensional convolution output to obtain a fixed-length feature vector; S403, inputting the fixed-length feature vector into a full-connection layer to obtain a category score vector, normalizing the category score vector to obtain category probability distribution, and taking the category with the maximum probability value in the category probability distribution as the final prediction category.
  7. 7. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 1, wherein S5 specifically comprises: S501, inputting a plurality of samples in a batch mode, dividing the samples into a plurality of class sets, and collecting all sample indexes belonging to the class in the current batch for each class to form the sample set; S502, constructing a cost matrix for any two sample feature sequences in the same sample set, and obtaining a dynamic time warping distance by utilizing dynamic programming search; S503, averaging the dynamic time warping distances corresponding to the samples in the class set to obtain intra-class dynamic time warping costs of the class, and averaging the intra-class dynamic time warping costs of all the classes to obtain intra-batch dynamic time warping regular terms; s504, calculating cross entropy classification loss according to the real class labels and the predicted class probability distribution of the samples, and obtaining an overall loss function by weighting and summing the cross entropy classification loss and the intra-batch dynamic time warping regular term; s505, performing back propagation and parameter updating on the model by using the total loss function, and storing the optimal parameters to obtain a trained detection model.
  8. 8. The method for detecting abnormal vibration of an industrial robot joint based on deep learning according to claim 1, wherein S6 specifically comprises: Intercepting the collected continuous real-time signals into samples to be tested with fixed time window length in a sliding window mode; Pretreating and enhancing the characteristics of the sample to be tested by S2 to obtain an enhanced sample; the enhanced data is input into a trained detection model to obtain probability distribution of four categories of normal, slight abrasion, serious abrasion and poor lubrication, and the category corresponding to the maximum value in the probability distribution is taken as an abnormal vibration detection result.
  9. 9. An abnormal joint vibration detection system of an industrial robot based on deep learning, which performs the method according to any one of claims 1 to 7, and is characterized by comprising: the data acquisition module is used for acquiring vibration signals of the joints of the industrial robot and constructing a training data set; the data enhancement module is used for enabling different samples to strictly correspond to each other in rotation phase to obtain enhanced samples; the feature extraction module is used for extracting local abnormal waveforms in the enhanced samples and constructing a deep time sequence feature sequence; the time sequence classification module is used for constructing a time sequence classification head to obtain final output class probability; The model training module is used for constructing an overall loss function based on cross entropy classification loss and dynamic time warping regular term, and executing back propagation and parameter updating on the model; And the online detection module is used for carrying out classification prediction on vibration signals to be detected acquired in real time and outputting abnormal vibration detection results.
  10. 10. A computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, the computer program being executable by a processor to implement the method for detecting abnormal vibration of an industrial robot joint based on deep learning as set forth in any one of claims 1 to 7.

Description

Industrial robot joint abnormal vibration detection method and system based on deep learning Technical Field The invention relates to the technical field of robot fault diagnosis, in particular to an industrial robot joint abnormal vibration detection method and system based on deep learning. Background Along with the continuous improvement of the industrial automation level, the industrial robot plays an increasingly important role in key manufacturing links such as welding, assembly, carrying and the like. The joints are used as core moving parts of the industrial robot, and the health state of the joints directly determines the running precision, stability and service life of the whole robot. Under the complex working conditions of long-term high speed, heavy load and reciprocating motion, the key transmission parts in the joint, such as a harmonic reducer, a bearing, a gear and the like, inevitably have the degradation phenomena of abrasion, looseness, poor lubrication and the like. The abnormal states are firstly represented as changes of joint vibration characteristics, if the abnormal states are not found and processed in time, the abnormal vibrations are gradually aggravated, and finally, the joint is possibly blocked, the accuracy is lost, even the whole machine is stopped, so that serious economic loss and production safety accidents are caused. Therefore, the method for detecting the abnormal vibration and identifying the state of the industrial robot joint in real time and accurately is an important technical means for realizing predictive maintenance of equipment and guaranteeing efficient and stable operation of a production line. The existing method directly analyzes the original signal intercepted by a fixed time window, ignores joint rotation speed fluctuation, causes incomplete rotation period and inconsistent phase covered by different samples, and causes the same physical position of similar faults to be disordered on a time axis, thereby affecting the stable learning of the model on periodic rules and local abnormal characteristics. Secondly, the existing method simply splices the triaxial acceleration and the triaxial gyroscope signals, and cannot utilize the kinematic relation to carry out structural decomposition on the multimode signals, so that the rigid rotation component and the abnormal vibration component are mixed, the model is difficult to distinguish the background from the abnormality, and the feature extraction difficulty is increased. And moreover, the conventional convolutional neural network adopts a regular convolutional kernel with a fixed sampling position, is difficult to adaptively capture non-rigid time sequence deformation such as local waveform stretching, compression, mutation or tiny dislocation in an abnormal vibration signal, and has insufficient capability of extracting fine-granularity abnormal characteristics related to a rotation phase. Finally, the existing method only depends on cross entropy loss optimization class discrimination capability, and the inherent consistency of similar abnormal samples in time sequence morphology is ignored. When the local impact time is slightly deviated or stretched, the anomaly similar in nature is easily misjudged as larger in difference, and the robustness of time dimension deformation is lacking. Disclosure of Invention In order to improve the visibility of abnormal features in input, the invention accurately captures local abnormal waveforms related to rotation phases in a period, and provides the following technical scheme: The invention provides an industrial robot joint abnormal vibration detection method based on deep learning, which comprises the following steps: S1, acquiring an original vibration signal in the running process of an industrial robot, wherein the original vibration signal comprises an acceleration signal and an angular velocity signal to form a training data set with class labels; S2, estimating a dominant rotation period of each sample through autocorrelation analysis, and mapping each sample to a uniform standardized period length through interpolation resampling to obtain a vibration signal after phase alignment; constructing a decoupling projection matrix based on the instantaneous angular velocity, decomposing the vibration signal after phase alignment into a dynamic component and a residual vibration component, applying self-adaptive enhancement to the residual vibration component and splicing the residual vibration component with the dynamic component to obtain an enhanced sample; S3, constructing a deformable convolution feature extraction network fused with a physical phase code, inputting the enhanced sample into the deformable convolution feature extraction network, and extracting a deep time sequence feature sequence, wherein the physical phase code is generated based on a sample corresponding angular velocity signal and is used for guiding sampling offset prediction of deformable c