Robot joint wear assessment method and system based on neural networks

Abstract

This application relates to the field of intelligent assessment, specifically disclosing a method and system for assessing robot joint wear based on neural networks. It synchronously acquires the original vibration time series and operating condition data time series of the robot joint, and constructs a feature extraction network with the operating condition data as a dynamic condition to decouple and extract pure degradation feature codes related only to the intrinsic wear state. Furthermore, based on this pure degradation feature code sequence, a degradation trend temporal modeling network is driven to accurately capture its long-term monotonic evolution law and generate a temporal hidden state representing the true wear degree of the joint. Finally, by decoding and predicting this temporal hidden state, an accurate assessment of the remaining service life of the robot joint is achieved. This ensures accurate tracking of the true degradation state of the joint under a wide range of changing operating conditions, thereby realizing precise and intelligent health assessment of robot joints.

Description

Technical Field

[0001] This application relates to the field of intelligent assessment, and more specifically, to a method and system for assessing robot joint wear based on neural networks. Background Technology

[0002] With the deep integration of industrial automation and intelligent manufacturing, industrial robots are playing an increasingly crucial role on high-precision, high-load production lines. As key moving components of robots, joint units are subjected to complex dynamic loads and friction over long periods, making wear and degradation an inevitable physical process. Progressive wear of joints directly leads to decreased robot motion accuracy and stability, potentially causing sudden failures, production interruptions, and economic losses. Therefore, accurately assessing the health status of robot joints and predicting their remaining service life is crucial for implementing predictive maintenance and ensuring production continuity and safety. Traditional methods based on physical models or signal processing often face limitations in modeling difficulties and insufficient generalization ability when dealing with degradation problems in highly nonlinear, strongly coupled systems like robot joints. In recent years, data-driven methods, represented by neural networks, have provided new insights into solving this problem due to their powerful nonlinear mapping and feature self-learning capabilities, becoming a research hotspot in this field.

[0003] However, in existing technologies, when robots perform diverse tasks, their operating parameters, such as motion speed, joint load, and execution trajectory, undergo drastic and frequent changes. These changes introduce strong, non-trend-based disturbances into monitoring signals such as joint vibration and torque, with amplitude and frequency characteristics often far exceeding those caused by early, minor wear. Most existing neural network models are trained on relatively stable operating data. When faced with real, changing working environments, they struggle to effectively isolate the strong interference from these changes, often misinterpreting fluctuations as deterioration of health, or having their learned degradation features overwhelmed by operating noise. How to effectively decouple and extract subtle degradation features that truly reflect wear status from dynamically changing, strong operating disturbances has become a key bottleneck restricting the accuracy of existing assessment technologies. This deep coupling between operating conditions and wear features significantly limits the accuracy and reliability of models in real industrial scenarios, making it difficult to accurately capture the inherent patterns reflecting long-term degradation trends.

[0004] Therefore, an optimized robot joint wear assessment scheme is needed to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a neural network-based method for assessing robot joint wear. This method synchronously acquires the original vibration time series and operating condition data time series of the robot joint, and constructs a feature extraction network with the operating condition data as a dynamic condition to decouple and extract pure degradation feature codes that are only related to the intrinsic wear state. Furthermore, based on this pure degradation feature code sequence, a degradation trend temporal modeling network is driven to accurately capture its long-term monotonic evolution law and generate a temporal hidden state representing the true wear degree of the joint. Finally, by decoding and predicting this temporal hidden state, an accurate assessment of the remaining service life of the robot joint is achieved. This ensures that the true degradation state of the joint is accurately tracked under a wide range of changing operating conditions, thereby realizing precise and intelligent health assessment of the robot joint.

[0006] According to one aspect of this application, a neural network-based method for evaluating robot joint wear is provided, comprising:

[0007] Obtain the original vibration time series and working condition data time series of the robot joints;

[0008] Hybrid feature extraction based on condition-aware feature co-mapping is performed on the original vibration time series of robot joints and the working condition data time series to obtain the time series of working condition modulated vibration time series feature encoding vectors;

[0009] Degradation trend time series modeling is performed on the time series of the time series feature encoding vector of the vibration under working condition to obtain the hidden state of joint wear time series;

[0010] The remaining lifetime is predicted based on the hidden state of joint wear time series.

[0011] According to another aspect of this application, a neural network-based robot joint wear assessment system is provided, comprising:

[0012] The multi-source time-series data acquisition module is used to acquire the original vibration time series and working condition data time series of the robot joints;

[0013] The hybrid feature extraction module is used to perform hybrid feature extraction based on condition-aware feature collaborative mapping on the original vibration time series of robot joints and the working condition data time series to obtain the time series of working condition modulated vibration time series feature encoding vectors.

[0014] The degradation trend time series modeling module is used to perform degradation trend time series modeling on the time series of the time series feature encoding vector of the vibration under working conditions to obtain the hidden state of joint wear time series.

[0015] The remaining lifetime prediction module is used to predict the remaining lifetime based on the hidden state of joint wear time series.

[0016] Compared with existing technologies, the neural network-based robot joint wear assessment method and system provided in this application simultaneously acquires the original vibration time series and working condition data time series of the robot joint, and constructs a feature extraction network with the working condition data as a dynamic condition to decouple and extract pure degradation feature codes that are only related to the intrinsic wear state. Then, based on this pure degradation feature code sequence, a degradation trend temporal modeling network is driven to accurately capture its long-term monotonic evolution law and generate a temporal hidden state representing the true wear degree of the joint. Finally, by decoding and predicting this temporal hidden state, an accurate assessment of the remaining service life of the robot joint is achieved. This ensures that the true degradation state of the joint is accurately tracked under a wide range of changing working conditions, thereby realizing precise and intelligent health assessment of the robot joint. Attached Figure Description

[0017] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0018] Figure 1 This is a flowchart of a neural network-based robot joint wear assessment method according to an embodiment of this application.

[0019] Figure 2 This is a data flow diagram of a neural network-based robot joint wear assessment method according to an embodiment of this application.

[0020] Figure 3 This is a flowchart of sub-step S2 of the neural network-based robot joint wear assessment method according to an embodiment of this application.

[0021] Figure 4 This is a flowchart of sub-step S23 of the neural network-based robot joint wear assessment method according to an embodiment of this application.

[0022] Figure 5 This is a block diagram of a neural network-based robot joint wear assessment system according to an embodiment of this application. Detailed Implementation

[0023] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0024] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0025] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0026] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0027] Figure 1 This is a flowchart of a neural network-based robot joint wear assessment method according to an embodiment of this application. Figure 2 This is a data flow diagram of a neural network-based robot joint wear assessment method according to an embodiment of this application. Figure 1 and Figure 2 As shown, the robot joint wear assessment method based on neural networks includes the following steps: S1, acquiring the original vibration time series and working condition data time series of the robot joint; S2, performing hybrid feature extraction based on condition-aware feature co-mapping on the original vibration time series and working condition data time series of the robot joint to obtain the time series of working condition modulated vibration time series feature encoding vector; S3, performing degradation trend time series modeling on the time series of working condition modulated vibration time series feature encoding vector to obtain the joint wear time series hidden state; S4, predicting the remaining life based on the joint wear time series hidden state to obtain the predicted remaining life.

[0028] In the aforementioned neural network-based robot joint wear assessment method, step S1 involves acquiring the original vibration time series and operating condition data time series of the robot joint. It should be understood that, because industrial robotic arms operate in precision assembly lines, their joints need to maintain high-frequency, high-precision reciprocating motion for extended periods. The mechanical actions within the joints, such as gear meshing, bearing rolling, and shaft transmission, are prone to abnormal vibrations due to component wear, changes in assembly clearances, or lubrication failure. The real-time operating condition data of the joint (such as load torque, motor speed, running time, and motion angle) directly determine the joint's stress intensity and motion state. Based on this, this application constructs a complete data system that includes mechanical dynamic response and external operating conditions by simultaneously collecting the original vibration signals of the joint and the corresponding operating parameters during the operation of the robotic arm joint. This fills the gap in incomplete information under a single data dimension and provides comprehensive data support for the subsequent accurate identification of the joint's health status. It ensures that the acquired data covers both the dynamic characteristics of the joint's mechanical movement, such as vibration amplitude and frequency distribution, and the external factors affecting these dynamic characteristics, such as the vibration change pattern under high load. This allows the health assessment model built on this data to effectively distinguish between vibration changes caused by normal operating condition fluctuations and abnormal vibrations caused by faults, significantly improving the accuracy of fault warnings, avoiding production line interruptions caused by misjudgments, reducing the amplification of joint faults caused by missed judgments, ensuring the continuity of assembly operations and product assembly accuracy, and reducing equipment maintenance costs and production losses.

[0029] In the specific implementation process, the first step is to deploy and configure the data acquisition hardware. High-frequency accelerometers are fixedly installed at key monitoring positions in the radial and axial directions of the robot joint motor housing. The sampling frequency of these sensors matches the characteristic frequency range of the joint vibration signal, and are used to acquire the original vibration time series during joint operation in real time. Simultaneously, the internal data bus of the robot controller is accessed through a standard industrial communication interface (such as an EtherCAT interface). Operating condition data corresponding to the time dimension of the vibration signal is read from the bus, specifically including the real-time angular velocity of the joint, the joint load torque output by the controller, and joint motion trajectory parameters. Next, a data synchronization mechanism is established. Using the sampling clock of the high-frequency accelerometer as a global time reference, the sampling time of the operating condition data is precisely aligned with the sampling time of the vibration data at the millisecond level through the controller's built-in timestamp synchronization protocol, ensuring that each set of vibration data sampling points corresponds to the operating condition parameters at the same moment. Finally, the analog vibration signal output by the accelerometer is converted into a digital signal through a data acquisition card and transmitted along with the operating condition digital signal read from the controller to the data storage unit. This data is stored sequentially as a continuous time-series data file, completing the acquisition of the original vibration time series and operating condition data time series of the robot joint.

[0030] In the aforementioned neural network-based robot joint wear assessment method, step S2 involves extracting hybrid features from the original vibration time series and the working condition data time series of the robot joint using condition-aware feature co-mapping to obtain a time series of working condition-modulated vibration time series feature encoding vectors. It should be understood that the original vibration time series of the robot joint contains strong interference signals caused by changes in working conditions. The amplitude and frequency characteristics of such interference often mask the weak signals of early wear. Simply extracting vibration features or working condition features cannot decouple the two, leading to misjudgments of the wear state by the subsequent assessment model. Therefore, this application further performs hybrid feature extraction based on condition-aware feature co-mapping on the two types of time series data. This dynamically guides the vibration feature extraction process through working condition data, specifically removing working condition interference. This accurately separates the influence of working condition changes and inherent wear on the vibration signal, obtaining a time series of working condition-modulated vibration time series feature encoding vectors that are only related to the joint wear state. This provides high-purity feature input for subsequent degradation trend time series modeling, avoids assessment bias caused by working condition fluctuations, and ensures the accuracy of wear state tracking under a wide range of working conditions. Figure 3 This is a flowchart of sub-step S2 of the neural network-based robot joint wear assessment method according to an embodiment of this application. Figure 3 As shown, step S2 includes the following steps: S21, inputting the original vibration time series of the robot joint into the sensor feature extraction branch to obtain the local temporal feature sequence of joint vibration; S22, inputting the working condition data time series into the working condition feature extraction branch to obtain the local temporal feature sequence of working condition; S23, performing feature co-mapping on each group of corresponding local temporal features of working condition and joint vibration in the local temporal feature sequence of working condition and the local temporal feature sequence of joint vibration to obtain the time series of the working condition modulated vibration temporal feature encoding vector.

[0031] Specifically, in step S21, the original vibration time series of the robot joint is input into the sensor feature extraction branch to obtain a local temporal feature sequence of joint vibration. It should be understood that since the original vibration time series of the robot joint is high-frequency continuous data, it contains a large amount of environmental noise and redundant information. Directly using it for feature co-mapping would increase computational costs and would fail to focus on wear-related local signal details, such as impact pulses and frequency shifts. Therefore, this application further inputs the original vibration time series into the sensor feature extraction branch to filter and enhance key local features in the vibration signal through the temporal feature extraction structure within the branch. In a specific example of this application, the sensor feature extraction branch includes a one-dimensional convolutional neural network model, which can strip away invalid noise from the high-frequency redundant original vibration data and extract local temporal features reflecting joint wear, such as peak intervals and energy changes, forming a structured local temporal feature sequence of joint vibration. This reduces the computational complexity of subsequent feature co-mapping and provides a high-quality vibration feature foundation for accurate matching with working condition features, ensuring the effectiveness of subsequent interference decoupling.

[0032] In the specific implementation process, the original vibration time series is first preprocessed by using a moving average filter to eliminate high-frequency environmental noise, and then standardized to unify the data dimensions. Next, a fixed-length sliding window (e.g., a window length of 1024 sampling points and a step size of 256 sampling points) is used to segment the long time series data, generating multiple data samples. Each sample has a dimension of [1024, 1], where 1 indicates that the vibration signal is a single channel. Subsequently, a sensor feature extraction branch is constructed. This branch is based on a one-dimensional convolutional neural network and contains four one-dimensional convolutional layers. The first two layers use a larger convolutional kernel size (e.g., 64) to capture global and local features, while the last two layers use a smaller convolutional kernel size (e.g., 16) to focus on fine wear features. Each convolutional layer is followed by a ReLU activation function and an average pooling layer. The pooling layer is used to compress the feature dimension and retain key features. Finally, the segmented data samples are input into this branch in batches. After layer-by-layer processing of convolution, activation and pooling, for each input [1024,1] sample, this branch outputs a local temporal feature sequence of joint vibration with a fixed dimension. For example, after network downsampling, the output feature sequence has a dimension of [8,64], where 8 represents the sequence length and 64 represents the feature dimension of each time step.

[0033] Specifically, in step S22, the working condition data time series is input into the working condition feature extraction branch to obtain the working condition local time series feature sequence. It should be understood that since the working condition data time series contains multi-dimensional discrete or continuous parameters such as joint velocity, load torque, and running trajectory, these parameters are dispersed in form and have large differences in dimensions. Directly using them to modulate vibration features cannot achieve accurate matching and cannot reflect the temporal correlation of the working condition parameters. Therefore, this application further inputs the working condition data time series into the working condition feature extraction branch to integrate multi-dimensional working condition parameters through the branch's feature processing structure and establish their temporal correlation. In a specific example of this application, the working condition feature extraction branch includes a one-dimensional convolutional neural network model, which can transform the dispersed working condition parameters into a working condition local time series feature sequence that is consistent with the time dimension and matches the feature dimension of the vibration local time series features. This effectively represents the dynamic change law of the working condition, provides accurate working condition basis for subsequent feature co-mapping, avoids interference decoupling failure caused by mismatch in the form of working condition features, and ensures the pertinence of vibration feature co-mapping.

[0034] In the specific implementation process, the time series of operating condition data is first cleaned by removing abnormal operating condition data using the 3σ criterion and supplementing missing data using linear interpolation. Similarly, a sliding window (window length 1024, step size 256) synchronized with the vibration signal is used to segment the multidimensional operating condition data, generating data samples that correspond one-to-one with the vibration samples. Assuming three operating condition parameters are collected, the dimension of each sample is [1024, 3]. Subsequently, an operating condition feature extraction branch is built, which adopts a combination structure of one-dimensional convolutional layers and fully connected layers. The one-dimensional convolutional layer is used to extract the temporal local correlation features of the operating condition data. After convolution, the two-dimensional temporal features are transformed into a one-dimensional feature vector through a flattening layer, and then the features are integrated and the dimensions are adjusted through two fully connected layers. Finally, the segmented working condition data samples are input into this branch. After layer-by-layer processing of convolution, flattening and full connection, for each input [1024,3] sample, this branch outputs a working condition local time series feature sequence with the same dimension as the vibration feature sequence, that is, the dimension is also [8,64]. This ensures that in the subsequent feature co-mapping steps, the two sets of feature sequences can correspond precisely at each time step.

[0035] Specifically, step S23 involves performing feature co-mapping on each set of corresponding local time-series features of the working condition and joint vibration in the local time-series feature sequence of the working condition and joint vibration to obtain a time series of the working condition modulated vibration time-series feature encoding vector. It should be understood that since the joint vibration local time-series features still retain working condition interference, and the working condition local time-series features can only characterize the working condition state, using only the two types of features cannot achieve interference removal. It is necessary to establish a dynamic relationship between the two through feature co-mapping, and to correct the vibration features based on the working condition features. Therefore, this application further performs feature co-mapping on the corresponding features in the two sets of feature sequences to dynamically adjust the feature components of the vibration local time-series features using the working condition local time-series features, strengthening wear information and suppressing working condition interference. In this way, through feature value interaction, intensity modeling, and differential enhancement, the influence of working condition changes on vibration features can be accurately eliminated, resulting in a time series of the working condition modulated vibration time-series feature encoding vector that only reflects the internal wear state of the joint. This provides a clean and effective feature input for subsequent degradation trend time-series modeling (such as LSTM modeling), ensuring the accuracy and reliability of remaining life prediction. Figure 4 This is a flowchart of sub-step S23 of the neural network-based robot joint wear assessment method according to an embodiment of this application. Figure 4 As shown, step S23 includes the following steps: S231, performing fine-grained feature mapping fusion on the local temporal features of the working condition and the local temporal features of joint vibration to obtain a fine-grained feature mapping fusion encoding vector for the working condition and joint vibration; S232, performing correlation topological quantization modeling on the feature values ​​at each position in the fine-grained feature mapping fusion encoding vector for the working condition and joint vibration to obtain a set of correlation topological quantization coefficients for the working condition and joint vibration; S233, based on the set of correlation topological quantization coefficients for the working condition and joint vibration, performing interactive importance weighting on the fine-grained feature mapping fusion encoding vector for the working condition and joint vibration to obtain a temporal feature encoding vector for the working condition modulated vibration.

[0036] More specifically, step S231 involves performing fine-grained feature mapping fusion on the local temporal features of the working condition and the local temporal features of joint vibration to obtain a fine-grained feature mapping fusion encoding vector for the working condition and joint vibration. It should be understood that the local temporal features of the working condition and the local temporal features of joint vibration exist independently, without establishing a deep correlation at the underlying feature dimension. This makes it impossible to directly reflect the synergistic changes or differences between the two under different working conditions. This leads to difficulties in accurately distinguishing between vibration signal changes caused by joint wear and vibration signal fluctuations caused by working condition fluctuations during subsequent processing, thus failing to provide an effective basis for wear assessment. Therefore, this application further performs element-level multiplication, addition, and subtraction operations on the two types of local temporal features to explore the synergistic, positive, and differential correlations between them at each feature dimension. Step S231 is expressed by the formula:

[0037]

[0038] in, In time step The working condition-joint vibration fine-grained feature mapping fusion encoding vector, In time step The local timing characteristics of the operating conditions, In time step The local temporal characteristics of joint vibration, This indicates element-wise multiplication.

[0039] Specifically, the feature dimensions of the collected local time-series features of the working conditions and the local time-series features of joint vibration are first verified. If the two dimensions are inconsistent, feature dimensionality reduction or dimensionality increase operations (such as principal component analysis or zero-filling) are performed to make their feature dimensions match perfectly. Subsequently, for the two types of features after matching, element-wise multiplication is performed for each feature dimension to obtain the co-correlation results, element-wise addition is performed to strengthen the positive correlation information, and element-wise subtraction is performed to capture the difference features. Finally, the results of these three types of operations are added one by one according to the feature dimensions to generate a fine-grained feature mapping fusion encoding vector of working condition-joint vibration containing multi-type correlation information. This ensures that each dimension of the vector can provide sufficient correlation pattern information for subsequent correlation topology quantitative modeling. It can capture the synchronous change patterns of working condition and vibration features in each dimension through element-wise multiplication, strengthen the positive correlation information between the two through addition, and highlight the difference features between the two through subtraction. The final fine-grained feature mapping fusion encoding vector of working condition-joint vibration carries multi-type correlation information of working condition and vibration features in each feature dimension. This provides a comprehensive and detailed underlying data foundation for subsequent judgment of the importance of each dimension to wear assessment, avoids wear signal identification deviation caused by missing correlation patterns, and ensures that subsequent quantitative analysis can accurately locate the feature dimensions that are valuable for wear assessment.

[0040] More specifically, step S232 involves performing correlation topology quantization modeling on the feature values ​​at each position in the fine-grained feature mapping fusion encoding vector of the working condition-joint vibration to obtain a set of correlation topology quantization coefficients for the working condition-joint vibration. It should be understood that although the fine-grained feature mapping fusion encoding vector of the working condition-joint vibration contains multi-dimensional correlation information, the contribution of each dimension to joint wear assessment varies significantly. Some dimensions may only reflect working condition fluctuation noise unrelated to wear. If the importance of these dimensions is not quantified and distinguished, subsequent feature enhancement operations will lack a clear direction, causing key wear features to be submerged by noise, ultimately affecting the accuracy of wear assessment. Therefore, this application further uses a two-layer fully connected network containing ReLU and Sigmoid activation functions to perform nonlinear transformation and normalization processing on the feature values ​​at each position of the interactive encoding vector to quantify the correlation strength of each feature dimension. Step S232 is expressed by the formula:

[0041]

[0042] in, In time step The working condition-joint vibration correlation topology quantization coefficient vector, that is, the set of working condition-joint vibration correlation topology quantization coefficients. This represents the Sigmoid activation function. This represents a modified linear unit, introducing nonlinearity. and Here are the learnable weight matrix and bias vector for the first fully connected layer. and This represents the learnable weight matrix and bias vector for the second fully connected layer.

[0043] Specifically, a network model for correlation quantization is first constructed: the learnable weight matrix of the first fully connected layer has dimensions [d, h] (d is the dimension of the interactive encoding vector, and h is the dimension of the hidden layer), the bias vector has dimensions [h], and the activation function is ReLU; the learnable weight matrix of the second fully connected layer has dimensions [h, d], the bias vector has dimensions [d], and the activation function is Sigmoid. Then, the fine-grained feature mapping fusion encoding vector of the working condition-joint vibration is input into this network, first passing through the first fully connected layer and then activated by ReLU to obtain the hidden layer feature vector. The hidden layer feature vector is then input into the second fully connected layer and activated by Sigmoid to obtain the correlation strength quantization results for each dimension. Finally, this result is extracted as a set of working condition-joint vibration correlation topology quantization coefficients, ensuring that each coefficient accurately represents the correlation importance of the i-th feature dimension. In this way, the nonlinear correlation patterns between features can be extracted through the first fully connected layer and the ReLU activation function, and the correlation strength of each dimension can be quantized to the range of 0 to 1 through the second fully connected layer and the Sigmoid activation function. The resulting set of working condition-joint vibration correlation topology quantization coefficients can accurately reflect the importance of the corresponding feature dimension to wear assessment, providing an interpretable and stable basis for subsequent interaction importance weighting enhancement operations. This ensures that the enhancement process can focus on high-contribution wear correlation features and filter out low-contribution noise features.

[0044] More specifically, step S233 involves performing interactive importance weighting on the fine-grained feature mapping fusion encoding vector of the working condition-joint vibration based on the set of correlation topology quantization coefficients to obtain the temporal feature encoding vector of the working condition modulated vibration. It should be understood that the signal intensity of each dimension in the fine-grained feature mapping fusion encoding vector of the working condition-joint vibration does not reflect its importance difference in wear assessment. This means that feature signals crucial to wear assessment may be masked by irrelevant working condition noise signals, and directly inputting them into subsequent models will affect the model's accuracy in capturing joint wear degradation trends. Therefore, this application further performs element-wise weighting on the interactive encoding vector through correlation strength quantization coefficients and superimposes it with the original interactive encoding vector to enhance the wear correlation features of high-importance dimensions while retaining basic feature information. Step S233 is expressed by the formula:

[0045]

[0046] in, This is the encoding vector for the timing features of vibration under operating conditions. This is a dot product based on position.

[0047] In other words, the first step is to verify the dimensionality consistency between the set of topological quantization coefficients for the working condition-joint vibration correlation and the fusion encoding vector of the working condition-joint vibration fine-grained feature mapping, ensuring a perfect dimensional match. Then, element-wise multiplication is performed on the two vectors to obtain a weighted feature vector. In this vector, signals in high-correlation-strength dimensions are amplified, while signals in low-correlation-strength dimensions are suppressed. Next, element-wise addition is performed between the weighted feature vector and the original working condition-joint vibration fine-grained feature mapping fusion encoding vector, preserving basic feature information and enhancing salient features. Finally, the working condition modulated vibration time-series feature encoding vector is output, ensuring that this vector provides a high-quality input for subsequent time-series models, containing both significant wear correlation features and basic information. In this way, element-wise multiplication of the quantization coefficients and the interactive encoding vector amplifies the feature dimension signals important for wear assessment and suppresses noise signals in irrelevant dimensions; and superposition with the original interactive encoding vector preserves necessary basic feature information and avoids the loss of key information. The final obtained operating condition modulated vibration time-series feature encoding vector not only highlights the significant features highly correlated with joint wear, but also retains the integrity of the basic features, greatly improving the signal-to-noise ratio of the feature signals. This provides accurate and pure feature inputs for subsequent time-series-based modeling of joint wear degradation trends, enabling the model to more sensitively capture subtle changes in wear evolution and significantly improve the accuracy of remaining life prediction.

[0048] In the aforementioned neural network-based robot joint wear assessment method, step S3 involves performing degradation trend time-series modeling on the time series of the working condition-modulated vibration time-series feature encoding vector to obtain the hidden state of joint wear time-series. It should be understood that although the time series of the working condition-modulated vibration time-series feature encoding vector has been stripped of working condition interference and only carries information related to the joint's intrinsic wear, this sequence still possesses dynamic time-series characteristics. The long-term monotonic evolution law of joint wear contained within it cannot be effectively captured by simple feature extraction or static modeling methods. Without a targeted time-series modeling process, it will be difficult to accurately extract the cumulative change trend of the wear state over time, thus affecting the accuracy of subsequent remaining life prediction. In a specific example of this application, step S3 includes: inputting the time series of the working condition-modulated vibration time-series feature encoding vector into a pre-trained LSTM model to obtain the hidden state of joint wear time-series. Specifically, this application further employs a pre-trained LSTM model to perform degradation trend time-series modeling on the time-series feature encoding vector sequence, thereby utilizing the core ability of the LSTM model to capture long-term dependencies in time-series data to deeply mine the evolution pattern and intrinsic correlation of wear features in the time dimension. In this way, a temporal hidden state that accurately represents the true wear degree of the joint can be generated. This hidden state not only integrates the wear feature information of each time step in the sequence, but also clearly presents the cumulative evolution process of the wear state from the early stage to the current stage. This provides high-quality and highly correlated feature input for the subsequent remaining life prediction module, fundamentally ensuring the accuracy and reliability of the remaining life prediction results, and ensuring that it can accurately reflect the actual degradation process of joint wear.

[0049] In the specific implementation process, the first step is to prepare a pre-trained LSTM model: An LSTM network structure is constructed, with three layers and 128, 64, and 32 neurons in each layer, respectively. The tanh activation function is used, and an appropriate forget gate threshold is set to optimize the transmission of long-term temporal information. A time series dataset containing the full lifecycle of the robot joint, from its initial healthy state to its severe wear state, is used to train the LSTM model. During training, the mean squared error loss function is used to optimize the model parameters until the model converges and the wear trend fitting error on the validation set is below a preset threshold. Next, input data adaptation is performed: the format of the time series of the current working condition modulation vibration time series feature encoding vector is adjusted to ensure that its time step and feature dimension are completely consistent with the input specifications of the pre-trained LSTM model, avoiding modeling errors caused by data format mismatch. The model then performs input and computation: the adapted temporal feature encoding vector sequence is input into the pre-trained LSTM model. The model learns the changing patterns of wear features in the sequence step-by-step through iterative computation of neurons in each layer. Through the synergistic effect of forget gates, input gates, and output gates, it retains key long-term wear evolution information and filters out short-term irrelevant fluctuations. Finally, the joint wear temporal hidden state is extracted: after the LSTM model has processed the entire temporal sequence, the hidden state sequence output at all time steps is extracted. This sequence accurately represents the current and historical wear accumulation level of the joint, providing complete degradation process information for subsequent steps.

[0050] In the aforementioned neural network-based robot joint wear assessment method, step S4 involves predicting the remaining lifespan based on the hidden state of joint wear time series to obtain the predicted remaining lifespan. It should be understood that since the hidden state of joint wear time series is a high-dimensional abstract feature generated after processing by a pre-trained LSTM model, it only implicitly contains the long-term evolutionary law of the joint from health to wear, and does not directly reflect a quantifiable remaining lifespan value that can be used for maintenance decisions. Furthermore, the dimensionality of this hidden state does not match the requirement for "single-value quantification" of remaining lifespan in industrial scenarios, and cannot directly provide an effective output for robot joint health assessment. Therefore, in a specific example of this application, step S4 includes: inputting the hidden state of joint wear time series into a RUL regression head to obtain the predicted remaining lifespan, wherein the RUL regression head is one or more fully connected layers, and the last fully connected layer in the one or more fully connected layers is a linear activation layer with a single neuron, and the predicted remaining lifespan is output by the linear activation layer with a single neuron. This enables a precise mapping from latent wear patterns to explicit lifespan values, ensuring that the output remaining lifespan is a continuous and reasonable quantitative result. It avoids deviations of predicted values ​​from the actual wear process caused by nonlinear activation. At the same time, the linear activation layer of a single neuron can guarantee the interpretability and stability of the prediction results, meeting the need for accurate quantification of the remaining lifespan of robot joints in industrial scenarios and providing direct and reliable data support for predictive maintenance.

[0051] In the specific implementation process, a RUL regression head for remaining lifespan prediction is first constructed. This regression head adopts a fully connected layer structure. The number of fully connected layers and the neuron configuration are determined according to the dimension of the joint wear time-series hidden state. The last fully connected layer is fixed as a linear activation layer with a single neuron to ensure that the output is continuous and conforms to the quantitative result of the physical meaning of remaining lifespan. The preceding fully connected layer is used to compress the dimension of the hidden state and integrate the features. Then, the joint wear time-series hidden state is preprocessed. If the hidden state is a multi-time-step feature sequence output by the temporal modeling network, it is transformed into a single vector feature through a global average pooling operation to ensure that it is adapted to the input dimension of the RUL regression head. Next, the preprocessed joint wear time-series hidden state is input into the RUL regression head. The preceding fully connected layer performs dimensionality reduction and information filtering layer by layer on the high-dimensional hidden state, retaining the feature components that are strongly correlated with remaining lifespan. The processed features are then fed into the last single-neuron linear activation layer. After linear calculation, the predicted value of the remaining lifespan of the corresponding joint is directly output. Finally, during the model training phase, the parameters of the RUL regression head are iteratively optimized using the true remaining lifespan of the device throughout its entire lifecycle as the label. This enables the regression head to stably learn the mapping relationship between the hidden state and the remaining lifespan, ensuring that reliable remaining lifespan prediction results can be accurately output based on the input joint wear time-series hidden state during the inference phase.

[0052] In summary, the neural network-based robot joint wear assessment method based on the embodiments of this application is explained. It simultaneously acquires the original vibration time series and operating condition data time series of the robot joint, and constructs a feature extraction network with the operating condition data as a dynamic condition to decouple and extract pure degradation feature codes related only to the intrinsic wear state. Furthermore, based on this pure degradation feature code sequence, a degradation trend temporal modeling network is driven to accurately capture its long-term monotonic evolution law and generate a temporal hidden state representing the true wear degree of the joint. Finally, by decoding and predicting this temporal hidden state, an accurate assessment of the remaining service life of the robot joint is achieved. This ensures accurate tracking of the true degradation state of the joint under a wide range of changing operating conditions, thereby achieving precise and intelligent health assessment of the robot joint.

[0053] Furthermore, a robot joint wear assessment system based on neural networks is also provided.

[0054] Figure 5 This is a block diagram of a neural network-based robot joint wear assessment system according to an embodiment of this application. Figure 5 As shown, the robot joint wear assessment system 100 based on a neural network according to an embodiment of this application includes: a multi-source time-series data acquisition module 110, used to acquire the original vibration time series and working condition data time series of the robot joint; a hybrid feature extraction module 120, used to perform hybrid feature extraction based on condition-aware feature co-mapping on the original vibration time series and working condition data time series of the robot joint to obtain the time series of working condition modulated vibration time-series feature encoding vector; a degradation trend time-series modeling module 130, used to perform degradation trend time-series modeling on the time series of working condition modulated vibration time-series feature encoding vector to obtain the hidden state of joint wear time series; and a remaining life prediction module 140, used to predict the remaining life based on the hidden state of joint wear time series to obtain the predicted remaining life.

[0055] The specific operation of each module in the aforementioned neural network-based robot joint wear assessment system has been referenced above. Figures 1 to 4 The description of the neural network-based robot joint wear assessment method is detailed here, and therefore, its repeated description will be omitted.

[0056] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details of the above embodiments are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the specific details described above.

[0057] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the unit division is only a logical functional division, and other division methods may exist in actual implementation. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0058] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the present invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0059] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units stated in a system claim may also be implemented by a single unit through software or hardware.

[0060] Finally, it should be noted that the above description has been given for illustrative and descriptive purposes. Furthermore, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although modifications or equivalent substitutions may be made to the technical solutions with reference to preferred embodiments, they will not depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A neural network-based robot joint wear assessment method, characterized by, The method comprises the following steps: obtaining a robot joint original vibration time sequence and a working condition data time sequence; performing mixed feature extraction based on condition perception feature collaborative mapping on the robot joint original vibration time sequence and the working condition data time sequence to obtain a working condition modulated vibration time sequence feature coding vector, which comprises: inputting the robot joint original vibration time sequence into a sensor feature extraction branch to obtain a joint vibration local time sequence feature sequence; inputting the working condition data time sequence into a working condition feature extraction branch to obtain a working condition local time sequence feature sequence; performing fine-grained feature mapping fusion on the working condition local time sequence feature and the joint vibration local time sequence feature to obtain a working condition-joint vibration fine-grained feature mapping fusion coding vector; performing correlation topological quantization modeling on the feature values at each position in the working condition-joint vibration fine-grained feature mapping fusion coding vector to obtain a set of working condition-joint vibration correlation topological quantization coefficients; based on the set of working condition-joint vibration correlation topological quantization coefficients, performing interactive importance weighted reinforcement on the working condition-joint vibration fine-grained feature mapping fusion coding vector to obtain a working condition modulated vibration time sequence feature coding vector; performing degradation trend time sequence modeling on the working condition modulated vibration time sequence feature coding vector to obtain a joint wear time sequence hidden state; based on the joint wear time sequence hidden state, performing remaining useful life prediction to obtain a predicted remaining useful life.

2. The neural network-based robot joint wear assessment method according to claim 1, characterized in that, The sensor feature extraction branch and the working condition feature extraction branch comprise a one-dimensional convolutional neural network model.

3. The neural network-based robot joint wear assessment method according to claim 1, wherein, The degradation trend time sequence modeling on the working condition modulated vibration time sequence feature coding vector to obtain a joint wear time sequence hidden state comprises: inputting the working condition modulated vibration time sequence feature coding vector into a pre-trained LSTM model to obtain the joint wear time sequence hidden state.

4. The neural network-based robot joint wear assessment method according to claim 1, wherein, The remaining useful life prediction based on the joint wear time sequence hidden state to obtain a predicted remaining useful life comprises: inputting the joint wear time sequence hidden state into an RUL regression head to obtain the predicted remaining useful life.

5. The neural network-based robot joint wear assessment method according to claim 4, characterized in that, The RUL regression head is one or more fully connected layers, the last fully connected layer in the one or more fully connected layers is a linear activation layer with a single neuron, and the predicted remaining useful life is output by the linear activation layer with a single neuron.

6. A neural network-based robot joint wear assessment system, characterized by, The method comprises the following steps: a multi-source time sequence data acquisition module for obtaining a robot joint original vibration time sequence and a working condition data time sequence; The hybrid feature extraction module is configured to perform hybrid feature extraction on the robot joint original vibration time sequence and the working condition data time sequence based on condition-aware feature collaborative mapping to obtain a time sequence of working condition modulated vibration time sequence feature coding vectors, and includes: inputting the robot joint original vibration time sequence into a sensor feature extraction branch to obtain a joint vibration local time sequence feature sequence; inputting the working condition data time sequence into a working condition feature extraction branch to obtain a working condition local time sequence feature sequence; performing fine-grained feature mapping fusion on the working condition local time sequence feature and the joint vibration local time sequence feature to obtain a working condition-joint vibration fine-grained feature mapping fusion coding vector; performing correlation topological quantization modeling on feature values at each position in the working condition-joint vibration fine-grained feature mapping fusion coding vector to obtain a set of working condition-joint vibration correlation topological quantization coefficients; based on the set of working condition-joint vibration correlation topological quantization coefficients, performing interactive importance weighted reinforcement on the working condition-joint vibration fine-grained feature mapping fusion coding vector to obtain a working condition modulated vibration time sequence feature coding vector; The degradation trend time sequence modeling module is configured to perform degradation trend time sequence modeling on the time sequence of the working condition modulated vibration time sequence feature coding vector to obtain a joint wear time sequence hidden state. The residual life prediction module is configured to perform residual life prediction based on the joint wear time sequence hidden state to obtain a predicted residual life.

Images