Robot multi-pose operation chatter monitoring method and system

Abstract

This invention belongs to the field of intelligent robot operation state monitoring, and particularly relates to a method and system for monitoring chatter during multi-pose robot operations. It includes collecting vibration signals from the robot under different operating poses to construct a two-dimensional time-frequency representation sample; constructing a LODN19-GRU-MHA chatter monitoring network integrating lightweight Darknet-19, gated recurrent units (GRUs), and multi-head attention (MHA); pre-training the LODN19-GRU-MHA chatter monitoring network using source domain data to obtain source domain model parameters; transferring the source domain model parameters to the target domain, and training the target domain model using a parameter freeze-fine-tuning strategy to obtain a trained target domain model, thereby achieving the identification of chatter states during multi-pose robot operations. This invention can improve the adaptability and multi-pose generalization performance of the monitoring model under limited target pose sample conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent robot operation status monitoring, and particularly relates to a method and system for monitoring flutter during multi-pose robot operations. Background Technology

[0002] During multi-pose robot operations, different joint configurations cause changes in system stiffness and dynamic response characteristics, resulting in significant drifts in the amplitude, frequency, and time-frequency distribution of flutter characteristics in vibration signals. Traditional signal processing methods for robot flutter monitoring rely heavily on manually designed feature indices and discrimination thresholds. When the robot's pose changes, the applicability of these original features and discrimination boundaries decreases. Furthermore, existing deep learning monitoring models typically have complex network structures and large parameter scales, leading to high computational costs and hindering online robot status monitoring.

[0003] Furthermore, existing deep learning monitoring models are mostly trained based on single pose data. The resulting models are easy to learn pose-related features, but their recognition accuracy and generalization ability tend to decrease when applied to new robot operation poses. Re-collecting a large number of labeled samples and training the model for new robot poses requires high data acquisition and model training costs, making it difficult to meet the rapid adaptation requirements under multi-pose operation conditions. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a method and system for monitoring flutter during multi-pose robot operations. By reducing the structural complexity and computational overhead of the monitoring model, it fully extracts the time-frequency and temporal dynamic features of the vibration signal and utilizes the common feature information between different robot operation poses to achieve model knowledge transfer. This reduces the impact of differences in flutter feature distribution caused by pose changes on state recognition performance, thereby improving the adaptability and multi-pose generalization performance of the monitoring model under limited target pose sample conditions.

[0005] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for monitoring flutter during multi-pose robot operations.

[0006] A method for monitoring chatter during multi-pose robot operations includes the following steps: Vibration signals of the robot under different working postures are collected and preprocessed. Time-frequency analysis of vibration signals is performed to construct two-dimensional time-frequency characterization samples; Construct a flutter monitoring network that integrates lightweight Darknet-19, gated recurrent unit (GRU), and multi-head attention (MHA); The LODN19-GRU-MHA flutter monitoring network was pre-trained using source domain data to obtain source domain model parameters. The parameters of the source domain model are transferred to the target domain, and the target domain model is trained using a parameter freeze-fine-tuning strategy to obtain a trained target domain model. By inputting the time-frequency samples to be monitored in the target domain into the trained target domain model, the flutter state of the robot in multi-pose operation can be identified.

[0007] The second aspect of the present invention provides a flutter monitoring system for multi-pose operation of robots.

[0008] A chatter monitoring system for multi-pose robot operations includes: The vibration signal acquisition module is configured to: collect vibration signals of the robot in different working postures and perform signal preprocessing; The two-dimensional time-frequency characterization module is configured to perform time-frequency analysis on vibration signals and construct two-dimensional time-frequency characterization samples. The network building module is configured to: build a LODN19-GRU-MHA flutter monitoring network that integrates lightweight Darknet-19, gated recurrent unit (GRU) and multi-head attention (MHA); The source domain training module is configured to: pre-train the LODN19-GRU-MHA flutter monitoring network using source domain data to obtain source domain model parameters; The target domain transfer module is configured to: transfer the parameters of the source domain model to the target domain, and train the target domain model using a parameter freeze-fine-tuning strategy to obtain a trained target domain model; The state recognition module is configured to input the time-frequency samples to be monitored in the target domain into the trained target domain model to realize the recognition of the flutter state of the robot in multi-pose operation.

[0009] The above one or more technical solutions have the following beneficial effects: This invention provides a method and system for monitoring flutter in multi-pose robot operations. By constructing a lightweight feature extraction network by pruning some convolutional layers of Darknet-19, the complexity of the model structure and computational overhead are reduced while maintaining effective flutter feature extraction capabilities, thus improving the computational efficiency and online deployment capability of the flutter monitoring model. Furthermore, a GRU is introduced to model the temporal dependencies of flutter features, and MHA is used to adaptively weight and enhance key flutter features from multiple feature subspaces, fully characterizing the dynamic evolution of flutter states and improving the model's ability to extract and identify complex time-frequency features. In addition, a transfer learning strategy combining source domain pre-training and target domain freeze-fine-tuning is employed. This retains the low-level general flutter features learned by the source domain model and uses limited target domain samples to adjust high-level feature representations and classification decision boundaries, reducing the dependence on a large number of labeled samples and retraining the model under new pose conditions, thereby improving the model's target pose adaptation efficiency.

[0010] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0012] Figure 1 This is a schematic diagram of the overall process of Embodiment 1 of the present invention.

[0013] Figure 2 This is a schematic diagram of the Darknet-19 feature extraction network structure according to Embodiment 1 of the present invention.

[0014] Figure 3 This is a schematic diagram of the gated loop unit (GRU) structure according to Embodiment 1 of the present invention.

[0015] Figure 4 This is a schematic diagram of the multi-head attention (MHA) structure according to Embodiment 1 of the present invention.

[0016] Figure 5 This is a schematic diagram of the transfer learning strategy between the source domain and the target domain in Embodiment 1 of the present invention.

[0017] Figure 6 This is a comparison chart of the recognition performance of different flutter monitoring models on datasets D1 and D2 in Embodiment 1 of the present invention.

[0018] Figure 7 This is the state recognition confusion matrix of the LODN19-GRU-MHA model in Embodiment 1 of the present invention on the D1 dataset.

[0019] Figure 8 This is the state recognition confusion matrix of the LODN19-GRU-MHA model in Embodiment 1 of the present invention on the D2 dataset. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] Example 1 Industrial robots, characterized by their large workspace, high flexibility, and strong adaptability, have been increasingly applied in the processing of complex components and intelligent manufacturing. However, because industrial robots typically employ a series-jointed multi-joint structure, their overall stiffness is relatively low, and the joint configuration, structural stiffness, and dynamic response characteristics vary significantly under different working postures. This makes them prone to vibration instability and chattering during high-load dynamic operations. Chattering not only reduces robot operational stability and workpiece surface quality but can also lead to accelerated tool wear and decreased operational accuracy. Therefore, accurate monitoring of the chattering state of robots under different working postures is crucial for ensuring robot operational quality and stability.

[0024] Regarding the method of identifying flutter using monitoring models, how to improve the ability to characterize flutter features while reducing model complexity, and enhance the adaptability and generalization ability of the monitoring model to different robot working poses, are urgent problems to be solved in existing robot flutter monitoring technologies.

[0025] This invention aims to overcome the problems of complex network structures, insufficient temporal feature representation, and drift in flutter feature distribution and decreased model generalization ability caused by changes in robot pose in existing robot flutter monitoring methods. It proposes a robot multi-pose operation flutter monitoring method based on a lightweight temporal attention transfer network. This method combines lightweight feature extraction, temporal feature modeling, multi-head attention feature enhancement, and transfer learning to achieve accurate identification of flutter states under different robot operation poses and rapid adaptation of the monitoring model.

[0026] A method for monitoring chatter during multi-pose robot operations includes the following steps: Vibration signals of the robot under different working postures are collected and preprocessed. Time-frequency analysis of vibration signals is performed to construct two-dimensional time-frequency characterization samples; Construct a flutter monitoring network that integrates lightweight Darknet-19, gated recurrent unit (GRU), and multi-head attention (MHA); The LODN19-GRU-MHA flutter monitoring network was pre-trained using source domain data to obtain source domain model parameters. The parameters of the source domain model are transferred to the target domain, and the target domain model is trained using a parameter freeze-fine-tuning strategy to obtain a trained target domain model. By inputting the time-frequency samples to be monitored in the target domain into the trained target domain model, the flutter state of the robot in multi-pose operation can be identified.

[0027] The method proposed in this embodiment first collects vibration signals of the robot under different working poses and working conditions. The vibration signals are preprocessed by sliding window segmentation, mean removal and normalization, and a two-dimensional time-frequency representation is constructed based on short-time Fourier transform. Then, a LODN19-GRU-MHA flutter monitoring network is constructed, which integrates lightweight Darknet-19, gated recurrent unit (GRU) and multi-head attention (MHA). The two-dimensional time-frequency spatial features are extracted by lightweight Darknet-19, the temporal dependency of flutter features is modeled by GRU, and key flutter information in different feature subspaces is adaptively weighted and enhanced by MHA.

[0028] Furthermore, the robot's working poses with sufficient samples are used as the source domain for model pre-training, and the parameters of the source domain model are transferred to the target domain. By freezing the low-level general feature extraction module of the target domain model, the high-level feature fusion module and classification module of the target domain model are fine-tuned using limited target domain samples, so that the model can adapt to the drift of flutter feature distribution caused by changes in robot pose. Finally, the transferred flutter monitoring model is used to realize the flutter state recognition of the robot under multi-pose working conditions.

[0029] This invention can reduce the model complexity and target pose sample requirements of the monitoring network, improve the flutter feature representation ability, cross pose adaptability and state recognition stability, solve the problems of vibration response and flutter feature distribution changes caused by robot pose changes, and the insufficient cross pose adaptability and generalization ability of existing monitoring models. It can be used for flutter state monitoring of robot multi-pose operation process.

[0030] The following will combine Figure 1 The method of this embodiment will be explained in detail.

[0031] Figure 1The overall data processing flow of this invention embodiment is shown, which mainly includes three stages: offline modeling, transfer learning, and online verification. The arrows in the figure indicate the direction of data, feature, or model parameter transfer.

[0032] In the offline modeling stage, vibration signals during robot operation are first collected, and signal preprocessing is completed through sliding window segmentation, mean removal, and normalization. Then, the vibration signals are converted into two-dimensional time-frequency representations using short-time Fourier transform (STFT). The multiple time-frequency graphs in the figure represent the time-frequency features corresponding to different vibration samples. The two-dimensional time-frequency samples are input into the lightweight Darknet-19 for source domain pre-training to obtain the source domain model and its feature extraction parameters.

[0033] In the transfer learning stage, the source domain and target domain in the figure represent the source robot's operational pose data domain and the target robot's operational pose data domain to be adapted, respectively. Through parameter transfer and feature sharing, the general features learned in the source domain model are transferred to the target domain, and the model is adapted using the target domain samples with freezing and fine-tuning strategies to obtain the target domain transfer model.

[0034] In the online verification stage, vibration signals in the target domain are acquired in real time. After signal segmentation and preprocessing, they are converted into continuous two-dimensional time-frequency samples and input into the transfer model for feature extraction. The extracted high-level features are sequentially input into the GRU time-series modeling module and the MHA feature enhancement module to obtain the temporal dependency of vibration features and enhance key flutter features. Finally, the system is input into the state recognition module, which outputs the robot's current operating state as "stable" or "fluttering".

[0035] Step S1, combined Figure 1 and Figure 6 The system collects vibration signals from the robot under different working postures and conditions, and performs sliding window segmentation and signal preprocessing.

[0036] During robot task execution, different working postures correspond to different joint configurations, causing changes in the structural stiffness and dynamic response characteristics of the robot system. This, in turn, leads to differences in the amplitude, frequency, and time-frequency distribution of the vibration response under different postures. Therefore, vibration signals are collected under multiple robot working postures and different working conditions to obtain raw vibration data that can characterize the dynamic response characteristics of the robot under different postures.

[0037] In one specific embodiment, three different robot working poses are set. Specifically, three different working positions within the robot's workspace are selected as typical working poses, denoted as P1, P2, and P3, respectively. These three working poses correspond to different robot joint configurations, wherein: The six joint angles of the robot in pose P1 are -24.86°, -72.46°, 144.80°, -27.80°, -73.51° and -83.77°. The six joint angles in pose P2 are -0.40°, -75.71°, 150.29°, -2.19°, -73.84° and -91.32°, respectively. The six joint angles in pose P3 are 23.43°, -72.84°, 145.46°, 22.76°, -73.14° and -98.54°.

[0038] By setting the robot's working poses with different joint configurations, vibration response data under different pose conditions are obtained. Specifically, a three-axis accelerometer is placed near the robot's end effector to collect vibration response signals during robot operation; the vibration signal sampling frequency is set to 10kHz (kilohertz), and raw vibration data under different working poses and working conditions are obtained through a data acquisition device.

[0039] To convert continuously acquired vibration signals into discrete samples suitable for flutter monitoring model training and state recognition, a sliding window approach is used to segment the original vibration signals.

[0040] In one specific embodiment, the sliding window length is set to 5000 sampling points, and the overlap rate between adjacent sliding windows is set to 50%, thereby constructing continuous vibration samples while preserving local temporal information of the vibration signal.

[0041] The vibration signal samples after sliding window segmentation are sequentially subjected to mean removal and normalization. Mean removal reduces the impact of the DC component on vibration feature extraction; normalization reduces amplitude scale differences between samples, ensuring a relatively uniform data scale for each vibration sample, thereby improving the stability of subsequent time-frequency representation construction and network training. After these processes, preprocessed robot multi-pose vibration signal samples are obtained.

[0042] Furthermore, the vibration samples are labeled according to the actual operating state of the robot. In one specific embodiment, the state is determined by combining the workpiece surface condition and whether obvious flutter frequency components appear in the vibration spectrum, and each vibration sample is marked as either stable or fluttering, thereby constructing a multi-pose vibration sample set of the robot with state labels. The experimental data used resulted in a total of 2400 vibration samples, including 1400 stable state samples and 1000 fluttering state samples.

[0043] Step S2, combined Figure 1Based on the short-time Fourier transform, the preprocessed vibration signal obtained in step S1 is subjected to time-frequency analysis to construct a two-dimensional time-frequency characterization sample.

[0044] When a robot experiences flutter, its vibration response not only manifests as changes in time-domain amplitude but also as energy accumulation within a specific frequency range. Furthermore, due to differences in system stiffness and dynamic response characteristics under different robot working postures, the main flutter frequencies and spectral energy distribution also change accordingly. Therefore, time-frequency analysis is used to simultaneously characterize the variations of vibration signals in both the time and frequency domains to improve the characterization capability of flutter features under different robot working postures.

[0045] This invention employs Short-Time Fourier Transform (STFT) to perform time-frequency analysis on the preprocessed vibration signal obtained in step S1. Specifically, by segmenting the vibration signal into local time segments and calculating the spectral distribution corresponding to each time segment, the one-dimensional time-domain vibration signal is converted into a two-dimensional time-frequency representation that simultaneously contains time information, frequency information, and vibration energy distribution information, thereby obtaining two-dimensional time-frequency samples corresponding to different robot working poses and working states.

[0046] Furthermore, the obtained two-dimensional time-frequency representation is subjected to uniform size processing to meet the input requirements of the subsequent flutter monitoring network. In one specific embodiment, the two-dimensional time-frequency representation is uniformly set to 128×128×3 and used as the input of the LODN19-GRU-MHA flutter monitoring network. After the above processing, the original one-dimensional vibration signal is converted into a two-dimensional time-frequency sample with uniform data dimensions, providing input data for subsequent lightweight feature extraction, temporal dependency modeling, and attention feature enhancement.

[0047] Step S3: Construct the LODN19-GRU-MHA flutter monitoring network.

[0048] Step S3, combined Figure 2 , Figure 3 and Figure 4 A flutter monitoring network LODN19-GRU-MHA integrating lightweight Darknet-19, gated recurrent unit (GRU), and multi-head attention (MHA) was constructed to sequentially complete two-dimensional time-frequency spatial feature extraction, temporal dependency modeling, key feature enhancement, and operation status classification.

[0049] 1. Lightweight Darknet-19 feature extraction.

[0050] First, the Darknet-19 network is used as the basic feature extraction backbone, and the original network is lightweighted by pruning some convolutional layers to construct a lightweight Darknet-19, namely the LODN19 feature extraction module.

[0051] The 128×128×3 two-dimensional time-frequency representation obtained in step S2 is input into the LODN19 network. The input sample first undergoes a 3×3 convolution operation to map the input features into a 128×128×32 feature map. Then, a 2×2 max-pooling layer is used for downsampling to obtain a 64×64×32 feature representation. Subsequently, multiple levels of convolution and pooling operations are used to gradually reduce the spatial size of the time-frequency feature map and increase the number of feature channels, thereby extracting flutter features at different scales.

[0052] In some convolutional modules, 1×1 convolutions are combined with 3×3 convolutions to reduce parameter overhead while achieving channel feature compression, expansion, and fusion. After multi-level convolutional feature extraction, the feature maps undergo spatial information aggregation through global average pooling layers, followed by feature recombination to obtain the feature sequence required by the subsequent GRU module. The feature size of LODN19 is successively compressed to 128×128, 64×64, 32×32, 16×16, 8×8, and 4×4, and finally, high-level feature representations are obtained through global average pooling.

[0053] like Figure 2 As shown, the purple blocks represent multi-channel feature maps output from different network layers, the gray arrows indicate the feature transfer process between adjacent network layers, and the red blocks represent fully connected layers at the end of the network. The boxes also show the basic feature extraction units consisting of convolution, batch normalization, ReLU activation, and max pooling. The overall feature extraction process includes: First, a 2D time-frequency representation of size 128×128×3 is input into the network. Shallow time-frequency features are extracted through 3×3 convolution, batch normalization, and ReLU activation. Max pooling is then used for downsampling to obtain a 64×64×32 feature map. Subsequently, features are extracted and compressed layer by layer through multi-level convolution and max pooling, reducing the feature map size sequentially from 64×64, 32×32, 16×16, 8×8 to 4×4, while the number of feature channels gradually increases from 32 to 64, 128, 256, and 512. As the number of network layers increases, the receptive field gradually expands, transforming shallow local textures and frequency energy distributions into higher-level features with stronger flutter state discrimination capabilities.

[0054] Furthermore, the 4×4×512 feature map is subjected to high-level convolutional mapping to obtain a 4×4×1000 high-dimensional feature representation, and feature compression and output are completed through end-point feature mapping and a fully connected structure. Through the above-mentioned stepwise convolution, activation and pooling process, multi-scale extraction of flutter-related features in two-dimensional time-frequency representation is realized, providing a feature foundation for subsequent GRU time-series feature modeling and MHA feature enhancement.

[0055] Through the above lightweight design, while reducing the model parameter size and computational overhead, the network retains its ability to effectively extract the two-dimensional time-frequency features of robot vibration.

[0056] 2. GRU temporal feature modeling.

[0057] Combination Figure 3 A gated recurrent unit (GRU) is introduced at the output of LODN19 to model the temporal dependency of the extracted feature sequences.

[0058] Because robot vibration signals exhibit significant temporal continuity, the formation and evolution of flutter typically manifest as a dynamic process. Therefore, simply using convolutional networks to extract two-dimensional time-frequency spatial features is insufficient to fully describe the dynamic characteristics of flutter state changes over time. This invention employs a GRU to jointly model current features and historical state information.

[0059] like Figure 3 As shown, let the current input feature be... The hidden state at the previous moment was GRU first determines the degree to which historical information is retained during the calculation of the current candidate state by using a reset gate.

[0060] (1) in, This indicates the output of the reset gate, used to adjust the degree of influence of historical hidden states on the current feature calculation; This represents the trainable weight matrix of the reset gate, used to perform a linear transformation on the concatenated historical hidden state and the current input features; the subscript r indicates that the weight belongs to the reset gate. This represents the trainable bias vector of the reset gate, used to adjust the bias of the linear transformation result of the reset gate. (·) represents the Sigmoid activation function, which maps the calculation result to the range of 0-1, thus forming the gating coefficient.

[0061] Subsequently, the retention ratio between historical state information and newly generated information is determined by updating the gate.

[0062] (2) in, This represents the update gate output, used to achieve a dynamic balance between historical states and current feature information; This represents the trainable weight matrix of the update gate, used to update the concatenated [ h t-1, x t Perform a linear transformation. This represents the trainable bias vector of the update gate, used to adjust the bias of the linear transformation result of the update gate.

[0063] Calculate the candidate hidden states based on the reset gate and the current input information. .

[0064] (3) in, represents a candidate hidden state; ⊙ represents element-wise multiplication; tanh represents the hyperbolic tangent activation function. This represents the trainable weight matrix used in the computation of candidate hidden states. This represents the trainable bias vector in the computation of candidate hidden states.

[0065] like Figure 3 As shown, the gated recurrent unit (GRU) mainly consists of four parts: reset gate, update gate, candidate hidden state calculation, and current hidden state update. Let the input feature at the current time be... x t The hidden state at the previous moment was h t-1 First, x t and h t-1 Input a reset gate, and obtain the reset gate output through a Sigmoid activation function. r t This is used to control the degree to which the hidden state from the previous time step is retained in the calculation of the current candidate hidden state; when r t When the size is small, reduce the impact of historical information on the current state. r t When the size is larger, more historical feature information is retained.

[0066] Then, x t and h t-1 Input the update gate, and obtain the update gate output through the Sigmoid activation function. z t This is used to adjust the retention ratio between historical hidden states and newly generated information. Further, the historical information filtered by the reset gate is fused with the current input features, and a non-linear mapping is performed using the tanh activation function to obtain candidate hidden states. h t .

[0067] After obtaining the candidate hidden states, output according to the update gate. z t and its supplement 1- z tThe hidden state of the previous moment is respectively h t-1 and candidate hidden state We perform weighted summation and fuse information through element-wise addition to obtain the hidden state at the current moment. h t .

[0068] Figure 3 middle, The symbol represents the Sigmoid activation function, tanh represents the hyperbolic tangent activation function, and "(1-)" indicates that the complement of the update gate output is calculated. This represents element-wise multiplication. This indicates element-wise addition.

[0069] Through the synergistic effect of reset and update gates, GRU can selectively forget, retain, and update historical flutter features, thereby capturing the dynamic information of vibration features changing over time and providing high-level feature representations with temporal dependencies for subsequent multi-head attention (MHA) modules.

[0070] Through the synergistic effect of the reset gate and the update gate, GRU can selectively forget, retain and update historical flutter features, thereby capturing the dynamic changes of vibration features in the time dimension and providing high-level feature representations with temporal relationships for subsequent multi-head attention modules.

[0071] 3. Enhanced MHA multi-head attention features.

[0072] Furthermore, a multi-head attention (MHA) module is set at the output of the GRU to enhance the network's ability to extract key flutter information in different feature subspaces.

[0073] like Figure 4 As shown, for the temporal feature sequence input to the MHA module, the query matrix Q, key matrix K, and value matrix V are first obtained through different trainable linear mappings. For the i-th attention head, its feature mapping process is expressed as: (4) in, , and Let represent the trainable mapping matrix corresponding to the i-th attention head; , , They represent the first i The query matrix, key matrix, and value matrix obtained after linear mapping of each attention head, where X represents the feature input to MHA multi-head attention feature enhancement, i.e., the high-level temporal feature output by the GRU module.

[0074] After obtaining the query matrix, key matrix, and value matrix, the correlation between different features and the corresponding attention weights are calculated using scaled dot product attention: (5) in, Indicates the first i The output features of the attention head are specifically the first attention head. i The feature representation is obtained by performing correlation calculation and weighted aggregation on the input features using an attention head. This represents the Softmax normalization function, which normalizes the correlation scores between different features into attention weights, so that each weight is in the range of 0-1 and the sum is 1. Key matrix K i The feature dimension of the key vector. The superscript T indicates matrix transpose.

[0075] Through the above calculations, the network can automatically adjust its weights according to the correlation between different features and flutter states, thereby highlighting key features with high state discrimination capabilities.

[0076] Different attention heads are learned in parallel in different feature subspaces, and the outputs of all attention heads are concatenated along the feature dimensions: (6) Where h represents the number of attention heads in the multi-head attention mechanism, and Concat represents the feature concatenation operation; This represents the multi-head attention aggregation feature obtained by concatenating all attention heads; This represents the output of the first attention head; This indicates the output of the second attention head; This represents the output of the h-th attention head, which is the output of the last attention head.

[0077] After concatenating features from multiple attention heads, a linear mapping is performed on the resulting features to obtain the final output of the MHA module: (7) Among them, W O Y represents the output mapping matrix of the multi-head attention module; Y represents the final output feature of the multi-head attention module obtained by concatenating the features of each attention head output and then linearly transforming the output mapping matrix.

[0078] Through the above multi-head parallel attention calculation, the robot flutter-related information can be selectively enhanced from different feature subspaces, thereby improving the network's ability to express and identify key flutter features. MHA completes Q, K, V mapping, scaled dot product attention, multi-head splicing, and final linear mapping through equations (4)-(7).

[0079] After LODN19 spatial feature extraction, GRU temporal feature modeling, and MHA key feature enhancement, the obtained high-level features are input into the fully connected classification module to obtain the classification result of the current robot operation state.

[0080] Step S4: Source domain model pre-training.

[0081] Step S4: Use the two-dimensional time-frequency samples in the source domain to pre-train the LODN19-GRU-MHA flutter monitoring network established in step S3 to obtain a source domain model with general flutter feature extraction capability.

[0082] Because changes in robot pose alter the robot's structural stiffness and dynamic response, the characteristic probability distributions of vibration signals from different poses differ. Therefore, this embodiment introduces a transfer learning framework, defining the data domain with sufficiently labeled samples as the source domain and the new robot operational pose data requiring model adaptation as the target domain.

[0083] The robot's working pose with sufficient training samples is selected as the source domain, and the two-dimensional time-frequency representation of the source domain obtained in step S2 is input into the LODN19-GRU-MHA network established in step S3 for training.

[0084] During training, local spatial and high-level semantic features in two-dimensional time-frequency samples are first extracted using the LODN19 network; then, the temporal dependencies in the feature sequences are modeled using GRU; further, the MHA module is used to highlight key features closely related to the flutter state in multiple feature subspaces; finally, the classification module is used to complete the state prediction, and the network parameters are updated through backpropagation using the loss function.

[0085] In one specific embodiment, the model uses the Adam optimizer, with a learning rate of 0.0001, a batch size of 60, a dropout rate of 0.3, and a cross-entropy loss function to optimize the network. The maximum number of training epochs is set to 200.

[0086] After training with source domain samples, the source domain LODN19-GRU-MHA pre-trained model and corresponding network parameters are obtained, providing an initialization basis for subsequent target domain parameter transfer.

[0087] Step S5: Source domain parameter migration and target domain freezing - fine-tuning.

[0088] Step S5, combined Figure 5 The source domain model parameters obtained in step S4 are transferred to the target domain, and the target domain model is adapted using a freeze-fine-tuning strategy to reduce the impact of flutter feature distribution drift caused by robot pose changes on state recognition performance.

[0089] Figure 5 The upper part shows the source domain data samples, the feature extraction and classification tasks of the source domain model, and the final state classification results.

[0090] Figure 5 The lower half shows the target domain data sample, the feature extraction and classification tasks of the target domain model, and the final state classification results.

[0091] First, the network parameters obtained from the source domain pre-training are transferred to the target domain and used as the initialization parameters for the target domain LODN19-GRU-MHA model.

[0092] (8) in, This represents the model parameters obtained by training with source domain data. Represents the initial parameters of the target domain model. (Symbol) This indicates that the source domain model parameters Assigning initial parameters to the target domain model That is, the target domain model is initialized using the pre-trained parameters of the source domain.

[0093] Since the low-level feature extraction modules in deep networks are mainly used to extract basic features such as local structure, edges and general time-frequency patterns, while the high-level network mainly extracts discriminative features related to specific state recognition tasks, this invention adopts a parameter transfer strategy of freezing the low-level network and fine-tuning the high-level network.

[0094] Specifically, during the training of the target domain model, the low-level feature extraction module in LODN19 is frozen to keep the low-level network parameters unchanged, so as to retain the general flutter feature representation ability obtained during the source domain training. At the same time, the high-level feature fusion module and classification module are set to a trainable state, and the network parameters are fine-tuned using a limited number of target domain samples.

[0095] The optimization process for the target domain model parameters is expressed as follows: (9) in, Represents the target domain model parameters. This represents the loss function corresponding to the target domain. This represents the optimized parameters of the target domain model.

[0096] By backpropagating and updating the high-level network parameters using limited labeled samples in the target domain, the high-level discriminative features and classification boundaries of the network are adapted to the data distribution in the target domain, while the underlying general time-frequency features are preserved.

[0097] This forms a transfer learning process of "source domain pre-training - model parameter transfer - low-level feature freezing - high-level feature and classification module fine-tuning", which reduces the number of training samples required for new robot poses while improving the model's adaptability to different robot operation poses.

[0098] Step S6: Identify the flutter state of the robot during multi-pose operation.

[0099] Step S6, combined Figure 1 By inputting the two-dimensional time-frequency samples of the target domain into the LODN19-GRU-MHA flutter monitoring model after migration and fine-tuning, the flutter state recognition of the robot in multi-pose operation can be realized.

[0100] In the actual monitoring process, the vibration signal of the robot under the current working posture is first collected in real time, and the samples are segmented in the same sliding window method as in step S1, and then the mean removal and normalization processes are performed in sequence.

[0101] Subsequently, following step S2, a short-time Fourier transform is performed on the preprocessed vibration signal to convert the one-dimensional vibration signal into a two-dimensional time-frequency representation of uniform size, which is then input into the LODN19-GRU-MHA network after the target domain fine-tuning is completed.

[0102] Two-dimensional time-frequency samples are first input into the LODN19 module, which extracts spatial time-frequency features at different scales through multi-layer convolution and pooling operations. After global average pooling and feature recombination, the high-level features are input into the GRU module, which selectively fuses historical and current information through reset and update gates, thereby extracting the dynamic features of flutter state changes over time.

[0103] The GRU output is further input into the MHA module, where multiple attention heads perform parallel analysis of temporal features from different feature subspaces. Attention weights are assigned based on the correlation between each feature and the job status, thereby enhancing key features with higher flutter discrimination capabilities.

[0104] After completing the MHA feature enhancement, the obtained features are input into the fully connected classification module to obtain the prediction result of the robot's current working state, thereby realizing online identification of flutter state under different robot working pose conditions.

[0105] Next, we will verify the flutter characteristics under different robot poses, the effectiveness of the LODN19-GRU-MHA network, and the effectiveness of the transfer learning strategy.

[0106] (1) Verification of flutter characteristics under different robot poses.

[0107] To verify the influence of robot pose changes on flutter feature distribution, in a specific embodiment, three different robot working poses, P1, P2 and P3, were selected for experiments, and vibration response signals under the corresponding working states were obtained respectively.

[0108] Experimental results show that the main frequency components and energy distribution of the vibration signal change significantly when flutter occurs under different robot poses. For example, under one pose, the flutter energy is mainly concentrated around 3500-4000 Hz; after changing the robot pose, the main flutter frequency range changes to 3500-4500 Hz; under another pose, the main flutter frequency is further distributed in the 3000-5000 Hz range.

[0109] The above results show that even when the robot is in the same type of flutter state, different working poses will still cause significant changes in the vibration amplitude and spectral energy distribution. That is, there is a flutter characteristic distribution drift caused by changes in robot pose. Therefore, a fixed model obtained by training based on a single robot pose is difficult to be directly applied to other robot poses.

[0110] (2) Validation of LODN19-GRU-MHA network.

[0111] To verify the effectiveness of the lightweight temporal attention network structure in step S3, a comparative test was conducted on the complete LODN19-GRU-MHA network, the network after removing the GRU module, and the network after removing the MHA module.

[0112] In one specific embodiment, the complete LODN19-GRU-MHA model achieved classification accuracies of 92.0% and 90.0% on two different pose datasets, respectively. Removing the GRU module reduced the accuracies to 88.9% and 86.3%, respectively; removing the MHA module reduced the accuracies to 90.7% and 88.5%.

[0113] The results show that the GRU module can improve the model's ability to describe the dynamic evolution of flutter by modeling the temporal dependencies of vibration features; the MHA module can enhance key flutter features from multiple feature subspaces and improve the ability to identify state features.

[0114] Therefore, the synergistic combination of LODN19, GRU and MHA can balance feature extraction efficiency, temporal dynamic information representation and key feature enhancement.

[0115] (3) Validation of the effectiveness of transfer learning strategies.

[0116] To verify the effectiveness of the transfer learning strategy in steps S4 and S5 for monitoring flutter during multi-pose robot operations, the data obtained under different robot poses were used to construct the source domain, target domain, and a new pose test domain that was not involved in model training. The state recognition performance of the model was compared with and without the use of transfer learning.

[0117] In one specific embodiment, the LODN19-GRU-MHA model is pre-trained using the P1 pose dataset D1 as the source domain, and freeze-fine-tuning is performed using the P2 pose dataset D2 as the target domain. Simultaneously, the model's generalization performance on new robot poses not previously used in training is verified using the P3 pose dataset D3. Furthermore, P1 and P2 are combined into a hybrid pose dataset D4, and P1, P2, and P3 are combined into a multi-pose hybrid dataset D5 to further verify the model's flutter monitoring performance under complex multi-pose conditions.

[0118] After source domain pre-training and target domain freeze-fine-tuning, the transfer model achieved recognition accuracies of 92.0%, 90.0%, 80.2%, 91.0%, and 85.2% on D1, D2, D3, D4, and D5, respectively.

[0119] Compared to models trained only on a single pose data, transfer models can not only maintain a high state recognition capability on source domain data, but also improve recognition performance on target domain and new pose data that were not trained. This shows that the "source domain pre-training + target domain freezing - fine-tuning" approach can reduce the dependence of a single pose model on the distribution of specific pose data.

[0120] Furthermore, on a hybrid dataset consisting of multiple robot poses, the transferred LODN19-GRU-MHA model still maintains good recognition stability.

[0121] The results show that this embodiment can effectively improve the model's adaptability to different robot working poses and complex multi-pose conditions by preserving the general flutter feature knowledge in the source domain model and making targeted adjustments to the high-level discrimination features in the target domain.

[0122] right Figure 6 , Figure 7 and Figure 8 The detailed content is as follows: To verify the recognition performance of different flutter monitoring models, six models, M1 to M6, were set up for comparison. Figure 6As shown, the state recognition accuracy of different flutter monitoring models on datasets D1 and D2 is compared. M1 represents the LODN19-GRU-MHA model used in this invention; M2 represents the model after removing the GRU module; M3 represents the model after removing the MHA module; M4, M5, and M6 represent the PSD+LSTM, OVMD+MT+CEI, and CWT-DCNN comparison models, respectively.

[0123] PSD+LSTM refers to a flutter monitoring model based on power spectral density (PSD) and long short-term memory (LSTM) networks. It extracts the power spectral density features of vibration signals and inputs them into an LSTM network for state identification.

[0124] OVMD+MT+CEI refers to a flutter identification method based on Optimized Variational Mode Decomposition (OVMD), Multi-band Information Fusion and Compression Technology (MT), and Cross Entropy of Image (CEI). Its general process is: vibration signal → OVMD decomposition → Multi-band Information Fusion and Compression → CEI calculation → flutter determination.

[0125] CWT-DCNN refers to a flutter monitoring model based on Continuous Wavelet Transform (CWT) and Deep Convolutional Neural Network (DCNN). It converts vibration signals into two-dimensional time-frequency representations through continuous wavelet transform and uses deep convolutional neural networks for feature extraction and state recognition.

[0126] Figure 6 In the diagram, D1 and D2 represent the datasets corresponding to the robot P1 and P2 operation poses, respectively. The horizontal axis represents models M1 to M6, and the vertical axis represents the recognition accuracy. The light blue bars represent the results of dataset D1, and the yellow bars represent the results of dataset D2.

[0127] The recognition accuracy of M1 on the D1 and D2 datasets was 92.0% and 90.0%, respectively, both higher than other comparative models. Meanwhile, the recognition accuracy decreased after removing the GRU or MHA modules, indicating that GRU temporal feature modeling and MHA feature enhancement both help improve the flutter state recognition performance.

[0128] Figure 7 and Figure 8 These are the state recognition confusion matrices of the LODN19-GRU-MHA model on the D1 and D2 datasets, respectively. Figure 7 The confusion matrix represents the recognition results of the M1 model on the D1 dataset, where the recognition accuracy for steady state and flutter state is 92.86% and 90.75%, respectively. Figure 8 The confusion matrix represents the recognition results of the M1 model on the D2 dataset, with recognition accuracies of 90.89% and 88.75%, respectively, indicating that the model has good flutter state recognition ability under different robot pose conditions.

[0129] Specifically: like Figure 7 As shown, the horizontal axis represents the model's predicted state, the vertical axis represents the sample's true state, and the percentage in the square represents the proportion of samples in the corresponding category. The darker the color, the higher the proportion. The main diagonal represents the proportion of correct identification, and the off-diagonal represents the proportion of incorrect identification.

[0130] For true steady-state samples, 92.86% of the samples were correctly identified as steady-state, and 7.14% were misidentified as fluttering states; for true fluttering state samples, 90.75% of the samples were correctly identified as fluttering states, and 9.25% were misidentified as steady-state. The results show that the model of this invention can accurately distinguish between steady-state and fluttering states on the D1 dataset.

[0131] like Figure 8 As shown, the horizontal axis represents the model's predicted state, the vertical axis represents the sample's true state, and the percentage in the square represents the proportion of samples in the corresponding category. The darker the color, the higher the proportion. The main diagonal represents the proportion of correct identification, and the off-diagonal represents the proportion of incorrect identification.

[0132] For true steady-state samples, 90.89% of the samples were correctly identified as steady-state, and 9.11% of the samples were misidentified as fluttering states; for true fluttering state samples, 88.75% of the samples were correctly identified as fluttering states, and 11.25% of the samples were misidentified as steady-state.

[0133] The results show that the model of this invention still maintains a high state recognition ability on the D2 dataset after the robot's working pose changes, and has a certain degree of cross-pose adaptability.

[0134] In summary, this invention constructs a two-dimensional time-frequency representation of vibration signals through short-time Fourier transform, achieves efficient extraction of flutter spatial features through lightweight Darknet-19, captures the time dependence of flutter features through GRU, and uses MHA to adaptively enhance key flutter information in different feature subspaces. Furthermore, it employs a source domain pre-training and target domain freeze-fine-tuning transfer learning strategy to enable the model to adapt to feature distribution drift caused by robot pose changes while retaining general flutter features, thereby achieving flutter state monitoring under multi-pose robot operation conditions.

[0135] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By pruning some convolutional layers of Darknet-19, a lightweight feature extraction network is constructed. While maintaining the effective flutter feature extraction capability, the complexity of the model structure and the computational overhead are reduced, thereby improving the computational efficiency and online deployment capability of the flutter monitoring model.

[0136] (2) GRU is introduced to model the time dependence of flutter features, and MHA is used to adaptively weight and enhance key flutter features from multiple feature subspaces to fully characterize the dynamic evolution characteristics of flutter state and improve the model's ability to extract and identify complex time-frequency features. Ablation results also show that both GRU and MHA make positive contributions to the recognition performance.

[0137] (3) A transfer learning strategy combining source domain pre-training and target domain freeze-fine-tuning is adopted to retain the low-level general flutter features learned by the source domain model and to adjust the high-level feature representation and classification decision boundary using limited target domain samples, thereby reducing the dependence on a large number of labeled samples and retraining the model under new pose conditions and improving the model's target pose adaptation efficiency.

[0138] (4) To address the differences in vibration response and the drift in flutter feature distribution caused by changes in robot pose, cross-pose parameter migration and target domain model fine-tuning are used to reduce the impact of feature distribution changes on flutter state recognition performance, thereby improving the model’s generalization ability, recognition stability and monitoring reliability under different robot working poses and multi-pose mixed conditions.

[0139] Example 2 This embodiment discloses a flutter monitoring system for multi-pose operation of robots.

[0140] A chatter monitoring system for multi-pose robot operations includes: The vibration signal acquisition module is configured to: collect vibration signals of the robot in different working postures and perform signal preprocessing; The two-dimensional time-frequency characterization module is configured to perform time-frequency analysis on vibration signals and construct two-dimensional time-frequency characterization samples. The network building module is configured to: build a LODN19-GRU-MHA flutter monitoring network that integrates lightweight Darknet-19, gated recurrent unit (GRU) and multi-head attention (MHA); The source domain training module is configured to: pre-train the LODN19-GRU-MHA flutter monitoring network using source domain data to obtain source domain model parameters; The target domain transfer module is configured to: transfer the parameters of the source domain model to the target domain, and train the target domain model using a parameter freeze-fine-tuning strategy to obtain a trained target domain model; The state recognition module is configured to input the time-frequency samples to be monitored in the target domain into the trained target domain model to realize the recognition of the flutter state of the robot in multi-pose operation.

[0141] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0142] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for monitoring chatter during multi-pose robot operation, characterized in that, Includes the following steps: Vibration signals of the robot under different working postures are collected and preprocessed. Time-frequency analysis of vibration signals is performed to construct two-dimensional time-frequency characterization samples; Construct a flutter monitoring network that integrates lightweight Darknet-19, gated recurrent unit (GRU), and multi-head attention (MHA); The LODN19-GRU-MHA flutter monitoring network was pre-trained using source domain data to obtain source domain model parameters. The parameters of the source domain model are transferred to the target domain, and the target domain model is trained using a parameter freeze-fine-tuning strategy to obtain a trained target domain model. By inputting the time-frequency samples to be monitored in the target domain into the trained target domain model, the flutter state of the robot in multi-pose operation can be identified.

2. The method for monitoring chatter during multi-pose robot operation as described in claim 1, characterized in that, Vibration signals from the robot in different working poses are collected and preprocessed, specifically including: Vibration signals were collected under multiple robot working poses and different working conditions to obtain raw vibration data characterizing the dynamic response characteristics of the robot in different poses. A sliding window is used to segment the raw vibration data into samples. The original vibration data after segmentation is subjected to mean removal and normalization to obtain the preprocessed vibration signal.

3. The method for monitoring chatter during multi-pose robot operation as described in claim 1, characterized in that, Time-frequency analysis of vibration signals is performed to construct two-dimensional time-frequency characterization samples, specifically including: Short-time Fourier transform is used to perform time-frequency analysis on the preprocessed vibration signal: the vibration signal is segmented into local time segments, and the spectral distribution corresponding to each time segment is calculated. The one-dimensional time-domain vibration signal is converted into a two-dimensional time-frequency representation that simultaneously contains time information, frequency information and vibration energy distribution information. The one-dimensional vibration signal is converted into a two-dimensional time-frequency image, and a two-dimensional time-frequency representation sample is constructed.

4. The method for monitoring chatter during multi-pose robot operation as described in claim 1, characterized in that, Constructing a flutter monitoring network integrating lightweight Darknet-19, gated recurrent units (GRUs), and multi-head attention (MHA), specifically including: The Darknet-19 network was used as the basic feature extraction backbone, and the original Darknet-19 network was lightweighted by pruning some convolutional layers to construct a lightweight Darknet-19. A gated recurrent unit (GRU) is introduced at the output of the lightweight Darknet-19. A multi-head attention (MHA) module is set at the output of the gated recurrent unit (GRU).

5. The method for monitoring chatter during multi-pose robot operation as described in claim 4, characterized in that, In the LODN19-GRU-MHA flutter monitoring network: The lightweight Darknet-19 extracts spatial features from time-frequency images through convolution operations, obtaining a flutter feature sequence composed of flutter features at different scales; The gated recurrent unit (GRU) models the time dependencies in the flutter feature sequence; Multi-head attention (MHA) adaptively weights and fuses key flutter features in different feature subspaces.

6. The method for monitoring chatter during multi-pose robot operations as described in claim 5, characterized in that, The gated recurrent unit (GRU) models the time dependencies in the flutter feature sequence, specifically as follows: Let the current input feature be The hidden state at the previous moment was GRU first determines the degree to which historical information is retained during the calculation of the current candidate state by using a reset gate: in, Indicates resetting the gate output; This represents the trainable weight matrix of the reset gate. This represents the trainable bias vector for the reset gate. (·) denotes the Sigmoid activation function; The retention ratio between historical state information and newly generated information is determined by updating the gate: ； in, This indicates an update to the gate output; This represents the trainable weight matrix of the update gate. This represents the trainable bias vector for updating the gate; Calculate the candidate hidden states based on the reset gate and the current input information: ； in, represents a candidate hidden state; ⊙ represents element-wise multiplication; tanh represents the hyperbolic tangent activation function. This represents the trainable weight matrix used in the computation of candidate hidden states. This represents the trainable bias vector in the computation of candidate hidden states.

7. The method for monitoring chatter during multi-pose robot operation as described in claim 5, characterized in that, The lightweight Darknet-19 extracts spatial features from time-frequency images through convolution operations, obtaining a flutter feature sequence composed of flutter features at different scales, specifically including: The input samples are first processed by convolution to map the input features into feature maps; Subsequently, downsampling is performed using a max pooling layer to obtain the feature representation; Subsequently, the spatial size of the time-frequency feature map is gradually reduced and the number of feature channels is increased through multi-level convolution and pooling operations, thereby extracting flutter features at different scales.

8. The method for monitoring chatter during multi-pose robot operation as described in claim 1, characterized in that, The LODN19-GRU-MHA flutter monitoring network was pre-trained using source domain data to obtain source domain model parameters, specifically including: Define robot pose data with sufficient samples as the source domain; The source domain two-dimensional time-frequency representation samples are input into the LODN19-GRU-MHA network for training. The lightweight Darknet-19 is used to extract general flutter features, the gated recurrent unit GRU is used to extract temporal dependency features, and the multi-head attention MHA is used to enhance key features, thereby obtaining the source domain model and corresponding model parameters.

9. The method for monitoring chatter during multi-pose robot operation as described in claim 1, characterized in that, The source domain model parameters are transferred to the target domain, and the target domain model is trained using a parameter freeze-fine-tuning strategy to obtain a trained target domain model. Specifically, this includes: The source domain model parameters are transferred to the target domain and used as the initialization parameters for the target domain model. The target domain model is defined to extract general local features using a low-level feature extraction module, a high-level feature fusion module, and a classification module. Freeze the parameters of the low-level feature extraction module; The parameters of the high-level feature fusion module and the classification module are fine-tuned using target domain samples.

10. A chatter monitoring system for multi-pose robot operations, characterized in that, include: The vibration signal acquisition module is configured to: collect vibration signals of the robot in different working postures and perform signal preprocessing; The two-dimensional time-frequency characterization module is configured to perform time-frequency analysis on vibration signals and construct two-dimensional time-frequency characterization samples. The network building module is configured to: build a LODN19-GRU-MHA flutter monitoring network that integrates lightweight Darknet-19, gated recurrent unit (GRU) and multi-head attention (MHA); The source domain training module is configured to: pre-train the LODN19-GRU-MHA flutter monitoring network using source domain data to obtain source domain model parameters; The target domain transfer module is configured to: transfer the parameters of the source domain model to the target domain, and train the target domain model using a parameter freeze-fine-tuning strategy to obtain a trained target domain model; The state recognition module is configured to input the time-frequency samples to be monitored in the target domain into the trained target domain model to realize the recognition of the flutter state of the robot in multi-pose operation.

Images