Multi-joint robot fault diagnosis system and method based on fractional order neural network

Abstract

The invention relates to the technical field of robots, and particularly discloses a multi-joint robot fault diagnosis system and method based on a fractional order neural network, the fractional order neural network is adopted, and through the synergistic effect of a fractional order convolution layer, a fractional order pooling layer and a fractional order ReLU function, compared with an integer order model, the fault diagnosis efficiency of the multi-joint robot is improved. According to the method, the non-linear and non-stationary feature capture capability of multi-joint vibration signals is remarkably enhanced, early weak fault features can be accurately extracted, the diagnosis accuracy of early faults such as gear tooth micro pitting corrosion and bearing clearance abnormity is improved, and the weak fault recognition capability is enhanced; a temporary decision maker mechanism under a distributed group learning framework retains the advantage of low communication cost, and meanwhile, through joint loss function optimization and dynamic weight distribution, data privacy protection and model training efficiency are considered, the communication cost is reduced, the training period is shortened, and the system performance is improved. The core requirements of industrial scenes for high precision, low time delay and strong privacy of fault diagnosis are met.

Description

Technical Field

[0001] This invention relates to the field of robotics, and in particular to a fault diagnosis system and method for multi-joint robots based on fractional neural networks. Background Technology

[0002] Fractional derivatives are an extension of integer derivatives, their core being the breaking of the limitations of integer-order differentiation / integration, allowing for arbitrary real-number orders. They describe the cumulative effect of the rate of change of a function, rather than merely reflecting the instantaneous rate of change. Since the core of first-order or higher derivatives is the instantaneous limit, while the core of fractional derivatives is integral accumulation (essentially definite integrals), and their mathematical expression is in integral form rather than integer-order difference form, this global integration logic naturally makes them the cumulative result of historical states, thus exhibiting a "memory effect." Because neural networks rely on integer-order derivatives for state updates and gradient calculations, they can only capture instantaneous or local information, making them ill-suited for complex data with long-term dependencies and dynamic evolution. Fractional neural networks, on the other hand, integrate fractional derivatives into the state update, weight optimization, or feature extraction processes. Their memory and nonlocality compensate for this deficiency while retaining the nonlinear fitting ability of neural networks.

[0003] Multi-joint robots, with their multi-degree-of-freedom collaborative motion capabilities, have become core equipment in scenarios such as precision assembly, automotive welding, and electronic processing. The health status of their joint components (such as harmonic reducers, servo motor bearings, and drive shafts) directly determines work efficiency and product quality. Currently, fault diagnosis of multi-joint robots mainly relies on four frameworks: local learning, centralized learning, federated learning, and group learning. However, all of them have technical bottlenecks: local learning requires massive amounts of high-quality data, but the data of each joint of a multi-joint robot is stored in different nodes, and the number of fault samples per node is limited, making it difficult to train a diagnostic model with strong generalization ability; centralized learning requires aggregating data from multiple nodes to a unified platform, which can integrate data resources, but there is a risk of data privacy leakage, and it cannot adapt to the differences in equipment models and working conditions in different factories; federated learning distributes the initial model and collects node training parameters through cloud servers, balancing data privacy and resource utilization, but the central server has too much weight, and repeated parameter transmission leads to high communication costs, and the initial model is prone to retaining redundant information; although group learning reduces communication costs through distributed node interaction, it uses integer-order convolutional neural networks as the core diagnostic model, which is insufficient in capturing the nonlinear and non-stationary features of multi-joint vibration signals, especially in separating the coupling interference generated by the movement of adjacent joints, resulting in low accuracy in diagnosing early weak faults (such as micro-pitting of gear teeth and abnormal bearing clearance). In summary, there is an urgent need for a fault diagnosis system and method that can improve the accuracy of fault diagnosis for multi-joint robots. Summary of the Invention

[0004] This invention provides a fault diagnosis system and method for multi-joint robots based on fractional neural networks, which can improve the accuracy of fault diagnosis for multi-joint robots.

[0005] To solve the above-mentioned technical problems, this application provides the following technical solution: A fault diagnosis method for multi-joint robots based on fractional neural networks includes the following steps: Step S1: Collect the original vibration data and servo motor control parameters of each joint of the multi-joint robot; denoise the original vibration data to obtain denoised vibration data; standardize the denoised vibration data to convert it into standardized vibration data; calculate the absolute difference between the actual parameters and the rated parameters of the servo motor control parameters to obtain control parameter deviation data; integrate the standardized vibration data and the control parameter deviation data to obtain a standardized multi-source dataset. Step S2: The standardized vibration data in the standardized multi-source dataset is processed by Hilbert transform to obtain an analytical signal; the analytical signal is processed by modulus calculation to obtain an envelope signal; the envelope signal is processed by Fourier transform to obtain a Hilbert envelope spectrum; fault features are extracted from the Hilbert envelope spectrum; the fault features and control parameter deviation data are fused to obtain fused feature data; the fused feature data is divided into training dataset and test dataset according to a preset ratio. Step S3: Construct fractional neural networks on multiple local nodes based on the training dataset as local training models. The fractional neural network includes fractional convolutional layers, fractional pooling layers and fully connected layers. The fractional convolutional layers use fractional convolution operators with a preset order range. The fractional pooling layers use fractional max pooling strategies. The activation function is the fractional ReLU function. Step S4: Local training is performed on the fractional neural network of each node. An improved domain adversarial neural network is introduced. A fractional gradient inversion layer is set between the feature extractor and the domain discriminator. A joint loss function containing fault classification loss and control parameter deviation fitting loss is constructed. The model parameters are optimized through the joint loss function to complete the local training. Step S5: Randomly select a node from multiple nodes as the decision-maker and temporary central server. The decision-maker calculates the weight coefficients of each node using a preset algorithm. The weight coefficients are related to the diagnostic accuracy and control parameter deviation fitting accuracy of the node model. Other nodes, as participants, transmit the locally trained fractional neural network parameters to the decision-maker. The decision-maker combines the weight coefficients of each node to perform weighted fusion of the received parameters to obtain the fused model parameters. In step S6, the decision-maker distributes the fusion model parameters to each participant, and the participants update their local fractional neural networks based on the fusion model parameters. Step S7, determine whether the training stopping criteria have been met: if after a preset number of consecutive rounds of fusion, the diagnostic accuracy of the model on the test dataset does not exceed a preset threshold and the control parameter deviation fitting error does not exceed a preset error value, then proceed to step S8; otherwise, return to step S4. Step S8: Validate the updated fractional neural network using the test dataset, and output the fault type, fault severity, and diagnostic confidence of each joint.

[0006] The basic principles and beneficial effects of this scheme are as follows: This scheme overcomes the limitations of single vibration data diagnosis by simultaneously collecting vibration data and servo motor control parameters. These two types of data reflect the joint's mechanical state and power drive state, respectively, forming complementary feature dimensions and avoiding misjudgments due to incomplete data. The raw vibration data undergoes denoising to remove environmental interference signals, and is then standardized to a uniform magnitude of data with a mean of 0 and a standard deviation of 1, eliminating the interference of data magnitude differences under different joints and operating conditions on model training. The servo motor control parameters accurately capture the deviation between actual operation and rated state through deviation calculation, providing a quantitative reference for fault diagnosis. The integration of these two types of data forms a standardized multi-source dataset that comprehensively covers fault-related features, achieving multi-source data fusion and refined preprocessing, thus solidifying the foundation for accuracy.

[0007] To address the nonlinear and non-stationary characteristics of vibration signals in multi-joint robots, a combination of Hilbert transform, modulus calculation, and Fourier transform is used to extract the Hilbert envelope spectrum from the vibration data. This envelope spectrum is sensitive to impact signals caused by joint faults and can effectively highlight early, subtle fault characteristics. The extracted fault features are then fused with control parameter deviation data to form richer, more discriminative fused feature data. Furthermore, the core model employs a fractional-order neural network. Compared to traditional integer-order models, the adjustable-order operators of fractional-order convolutional layers can flexibly adapt to the complex dynamic characteristics of fault features. Fractional-order pooling layers can finely retain key feature information, and the fractional-order ReLU function enhances the model's ability to fit nonlinear features. The synergistic effect of these three elements improves the feature extraction accuracy for complex faults such as flexible wheel pitting and wave generator misalignment, fundamentally improving the accuracy of fault identification. Therefore, this precise feature extraction and fractional-order model architecture strengthens fault identification capabilities.

[0008] Due to the large motion coupling and varying operating conditions of the joints in a multi-joint robot, domain differences can easily arise in the data from different nodes. This solution introduces an improved domain adversarial neural network (DAN), dynamically adjusting the gradient inversion strength through a fractional-order gradient inversion layer. This allows the model to learn cross-domain general features during training, effectively mitigating the interference caused by domain differences in the data from different nodes. Simultaneously, a joint loss function is constructed, incorporating fault classification loss and control parameter deviation fitting loss. This optimizes fault classification accuracy while constraining the model's ability to fit control parameter deviations, enabling the model to not only identify fault types but also correlate them with the operating state of the power drive system, reducing the probability of misclassification and improving the model's generalization ability under multiple joints and operating conditions. A distributed swarm learning framework is adopted, using random selection of decision-makers and calculation of node weight coefficients for parameter fusion. These weight coefficients are directly linked to the diagnostic accuracy and control parameter deviation fitting accuracy of the node model, giving higher weights to nodes with better diagnostic performance. This ensures that the fused model parameters aggregate the advantages of each node, avoiding performance trade-offs caused by simple averaging. Furthermore, through continuous, preset rounds of iterative training and parameter updates, the model performance is continuously optimized until the fluctuations in diagnostic accuracy and the fitting error of control parameter deviations meet preset standards. This ensures that the final model reaches a stable, high-precision state, enabling accurate determination of the fault type and severity of each joint. In this way, dynamic weight fusion and iterative optimization aggregate globally optimal performance, further improving the fault diagnosis accuracy of multi-joint robots.

[0009] Furthermore, in step S1, the calculation formula for the standardization process is as follows: , Where x represents the standardized vibration data. For noise-reducing vibration data, The mean of the denoised vibration data. The standard deviation of the denoised vibration data; In step S3, the formula for calculating the fractional ReLU function is: , in, x is the output value of the activation function, and x is the input value of the function. It is a fractional-order adjustable parameter; In step S4, the formula for calculating the joint loss function is: , in, For the joint loss function value, For fault classification loss values, To control the parameter deviation fitting loss value, These are the weighting coefficients; In step S5, the formula for calculating the node weight coefficient is as follows: in, Let be the weight coefficient of the i-th node. Let be the diagnostic accuracy of the model for the i-th node. Let be the fitting accuracy of the control parameter deviation of the model at the i-th node. , Assign weighting coefficients; In step S6, the calculation formula for parameter weighted fusion is as follows: , in, These are the parameters for the fusion model, where n is the total number of nodes. Let be the weight coefficient of the i-th node. The parameters of the fractional neural network for the i-th node after the j-th round of training are given.

[0010] The beneficial effects are as follows: standardization ensures the consistency of vibration data and eliminates the interference of data magnitude differences on training; the adjustable parameter α of the fractional ReLU function enhances the adaptability to nonlinear features; the joint loss function balances the fault classification and control parameter fitting objectives through the weight coefficient γ, improving the targeting of model training; the node weight coefficient formula combines the diagnostic accuracy and control parameter fitting accuracy to allocate weights, making parameter fusion more reasonable; in this way, the fault diagnosis accuracy of multi-joint robots is improved.

[0011] Furthermore, in step S1, the collected raw data also includes historical motion trajectory data of each joint of the multi-joint robot; the historical motion trajectory data is interpolated and completed to obtain complete trajectory data; the complete trajectory data is calculated by deviation calculation to solve the Euclidean distance between the actual trajectory and the planned trajectory to obtain trajectory deviation data; the standardized vibration data, control parameter deviation data and trajectory deviation data are integrated to obtain a standardized multi-source dataset; in step S2, the fault features, control parameter deviation data and trajectory deviation data are fused to obtain fused feature data; in step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss and trajectory deviation fitting loss.

[0012] The beneficial effects are as follows: by introducing historical motion trajectory data, the kinematic characteristics of the joint are supplemented, and the trajectory deviation data can reflect the cumulative error of the joint mechanical transmission chain, forming complementary features with vibration and control parameters, which can improve the accuracy of identifying faults that are strongly correlated with motion trajectory, such as poor gear meshing and joint loosening; at the same time, the introduction of trajectory deviation fitting loss strengthens the learning of the dynamic motion characteristics of the joint and improves the accuracy of concurrent fault diagnosis when multiple joints are working together.

[0013] Furthermore, in step S1, the collected raw data also includes real-time operating data of the multi-joint robot, which includes current load parameters, operating speed parameters, and ambient temperature parameters. The real-time operating data is processed by outlier removal to obtain clean operating data. The clean operating data is then normalized to obtain standardized operating data. The standardized vibration data, control parameter deviation data, and standardized operating data are integrated using data fusion methods to obtain a standardized multi-source dataset. In step S2, fault features, control parameter deviation data, and standardized operating data are fused using feature fusion methods to obtain fused feature data. In step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss, and operating parameter adaptation loss.

[0014] The beneficial effects are as follows: by introducing real-time operating data, the robot can be dynamically adapted to its current working condition, improving its sensitivity to real-time variables such as load, speed, and temperature; the fusion of real-time operating data with vibration and control parameters can accurately distinguish between normal operating condition fluctuations and fault characteristics, reducing the misjudgment rate; the addition of operating parameter adaptation loss strengthens the learning ability of real-time operating conditions and improves the diagnostic accuracy in extreme scenarios such as high load and variable speed operation.

[0015] Furthermore, in step S4, during local training, a response time constraint is introduced, setting a threshold for the single-inference response time of a single-node model. A dynamic order adjustment strategy is adopted for the fractional-order convolutional layers of the fractional-order neural network. When the inference response time exceeds the threshold, the order of the fractional-order convolutional operator is reduced by a preset step size until the response time meets the threshold requirement. In step S5, during parameter fusion, the decision-maker prioritizes receiving node parameters that meet the response time standard, and performs lightweight compression processing on node parameters that do not meet the standard before fusion. In step S7, the training stopping criterion, a condition is added that the average response time of the model meets the standard, and the single-inference response time of the model after all nodes are updated does not exceed the preset response time threshold.

[0016] The beneficial effects are as follows: by introducing response time constraints and dynamic optimization strategies, the inference latency that may be caused by the complexity of the fractional-order neural network is prevented from being too long. The synergistic effect of dynamic order adjustment and parameter lightweight compression can shorten the response time of a single inference. The response time meeting the standard is used as a training stopping condition, which ensures that while maintaining high diagnostic accuracy, the latency requirements of real-time fault diagnosis of industrial robots are met, and the fault expansion caused by diagnostic lag is avoided.

[0017] Furthermore, in step S3, when constructing the fractional-order neural network, a precision adjustment is added. Based on the joint type and fault diagnosis priority of the multi-joint robot, the order range of the fractional-order convolution operator and the pooling window size of the fractional-order pooling layer are dynamically configured. In step S4, during local training, the precision adjustment monitors the model's diagnostic precision in real time. When the precision is lower than the preset precision threshold, the order range of the fractional-order convolution operator is expanded and the pooling window is reduced. When the precision is higher than the preset precision threshold and the redundancy exceeds the limit, the order range is reduced and the pooling window is expanded. In step S7, the training stop criterion, it is simultaneously verified whether the diagnostic precision of each joint meets the precision requirements of the corresponding priority.

[0018] The beneficial effects are as follows: by adding precision adjustment and dynamic configuration strategies, an adaptive balance between diagnostic accuracy and model complexity is achieved; differentiated adjustments based on different joint characteristics and diagnostic priorities improve the fault diagnosis accuracy of critical joints, while reducing model redundancy and improving computational efficiency for non-critical joints while ensuring diagnostic reliability; real-time precision monitoring and dynamic parameter adjustment avoid insufficient accuracy or resource waste in complex scenarios due to fixed parameter models, improve flexibility and adaptability, and meet the personalized diagnostic needs of multi-joint robots under different working conditions.

[0019] Furthermore, in step S1, the collected raw data also includes load difference data of each joint of the multi-joint robot; the load difference data is processed by hierarchical calibration to obtain load level label data; the load level label data is processed by one-hot encoding to obtain standardized load label data; the standardized vibration data, control parameter deviation data and standardized load label data are integrated with the standardized load label data to obtain a standardized multi-source dataset; in step S2, the fault features, control parameter deviation data and standardized load label data are integrated with the feature fusion to obtain fused feature data; in step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss and load level adaptation loss.

[0020] The beneficial effects are as follows: by introducing load difference data and building a load adaptation mechanism, the adaptability to different load conditions is improved; the deep integration of load level labels with fault characteristics and control parameters can accurately distinguish the signal differences caused by load changes and faults, and improve the diagnostic accuracy in load fluctuation scenarios; the addition of load level adaptation loss strengthens the learning ability of different load characteristics, and can avoid misjudgment caused by load interference under harsh conditions such as variable load and high load.

[0021] Furthermore, in step S1, the collected raw data also includes drive difference data of each joint of the multi-joint robot; the drive difference data is processed by waveform feature extraction to obtain drive feature data; the drive feature data is processed by standardization to obtain standardized drive data; the standardized vibration data, control parameter deviation data and standardized drive data are integrated with the standardized drive data to obtain a standardized multi-source dataset; in step S2, the fault features, control parameter deviation data and standardized drive data are integrated with the standardized drive data to obtain fused feature data; in step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss and drive difference adaptation loss.

[0022] The beneficial effects are as follows: by introducing drive difference data, the feature dimensions at the joint power source level are supplemented, preventing diagnostic bias caused by ignoring individual drive differences; the multi-dimensional fusion of drive characteristics with vibration and control parameters can accurately identify joint faults caused by drive differences such as drive power imbalance and current waveform distortion, improving the diagnostic accuracy of such faults; the addition of drive difference adaptation loss strengthens the adaptability to different drive characteristics, which can reduce the fluctuation of diagnostic accuracy in scenarios with different drive models and different aging levels of multiple joints.

[0023] A fault diagnosis system for multi-joint robots based on fractional neural networks is used to perform any of the methods mentioned above.

[0024] The basic principles and beneficial effects of the solution are as follows: A complete technical system is constructed, which integrates multi-source data fusion, fractional-order model training, distributed parameter optimization, and accurate diagnostic output. It integrates multi-dimensional data acquisition and processing, and can simultaneously acquire various types of data such as vibration data, control parameters, historical trajectories, and real-time operating status, comprehensively covering joint fault-related feature dimensions and preventing diagnostic bias caused by single data dimensions. The core adopts a fractional-order neural network architecture, combined with an improved domain adversarial neural network and dynamic parameter fusion strategy, which enhances the ability to extract nonlinear and non-stationary fault features. It can also adapt to domain differences and individual differences under multiple nodes and multiple working conditions, improve the accuracy of early weak fault diagnosis, and reduce diagnostic accuracy fluctuations in complex scenarios. At the same time, it incorporates optimization mechanisms such as response time constraints and dynamic accuracy adjustment, taking into account both diagnostic real-time performance and resource utilization efficiency, shortening the single inference response time, and meeting the real-time monitoring needs of industrial robots. Attached Figure Description

[0025] Figure 1 This is a flowchart of Implementation 1 of a fault diagnosis method for multi-joint robots based on fractional neural networks; Figure 2 This is a system block diagram of Embodiment 2 of a multi-joint robot fault diagnosis system based on fractional neural networks. Detailed Implementation

[0026] The following detailed description illustrates the specific implementation method: Example 1

[0027] This embodiment describes a fault diagnosis method for multi-joint robots based on fractional-order neural networks. The method is applied to a 4-axis multi-joint robot used for precision assembly of electronic components. Joints 1-4 of this robot are equipped with harmonic reducers and servo motors. The method aims to accurately diagnose four types of faults: normal operation, pitting of the harmonic reducer flexspline, uneven wear of the wave generator, and looseness of the drive shaft. (See attached diagram.) Figure 1 As shown, the specific implementation process is as follows: Step S1: Multi-source data acquisition and preprocessing. Raw data acquisition includes: acquiring raw vibration data of each joint using a piezoelectric accelerometer; acquiring servo motor control parameters using a servo motor controller; acquiring historical motion trajectory data of each joint using a robot motion control card; acquiring real-time operating data using a torque sensor; acquiring load difference data using a power analyzer; and acquiring drive difference data using an oscilloscope. Wavelet threshold denoising is then applied to the raw vibration data to obtain denoised vibration data. ;right The standardization process is performed using the following formula: , if, mean Standard deviation After calculation, the mean of the standardized vibration data x approaches 0 and the standard deviation approaches 1, thus obtaining the standardized vibration data. Deviation calculations are performed on the servo motor control parameters to determine the absolute difference between the actual operating parameters and the rated parameters. For example, if the actual current is 6.3A and the rated current is 6A, then the control parameter deviation data... Similarly, the rotational speed deviation is calculated to obtain the control parameter deviation data. Cubic spline interpolation is used to complete the historical motion trajectory data. If the trajectory data of joint 2 has three consecutive missing points, interpolation is performed based on five valid coordinate points before and after each missing point to obtain the complete trajectory data. The Euclidean distance between the actual trajectory and the planned trajectory is calculated from the complete trajectory data. For example, if the planned trajectory coordinates are (50, 80) and the actual trajectory coordinates are (50.2, 80.3), then the trajectory deviation data is... The trajectory deviation data is obtained. Outliers are removed from the real-time operation data using a box plot method. For example, an outlier of 45℃ in the ambient temperature data is removed, resulting in clean operation data. Min-max normalization is performed on the clean operation data, mapping the load parameters to the [0,1] interval to obtain standardized operation data. Load difference data is then categorized and labeled. For instance, 20% light load is labeled as Level 1, 50% medium load as Level 2, and 80% heavy load as Level 3, resulting in load level label data. One-hot encoding is performed on the load level label data, with Level 1 encoded as [1,0,0], Level 2 as [0,1,0], and Level 3 as [0,0,1], resulting in standardized load label data. Wavelet packet transform is used to extract features such as energy entropy and peak factor from the current waveform of the drive difference data, resulting in drive feature data. Standardization is then performed on the drive feature data to obtain standardized drive data. Standardized vibration data, control parameter deviation data, trajectory deviation data, standardized operation data, standardized load label data, and standardized driving data are spliced ​​and integrated through feature dimensions to obtain a standardized multi-source dataset with a sample size of 800 and a feature dimension of 12.

[0028] Step S2: Feature Extraction and Dataset Partitioning. Perform a Hilbert transform (using a combination of Fast Fourier Transform and numerical inversion algorithm) on the standardized vibration data in the standardized multi-source dataset to obtain the complex analytic signal; solve for the magnitude of the analytic signal. For example, if the analytic signal is... The modulus is calculated as follows: The envelope signal was obtained; a Fourier transform was performed on the envelope signal to obtain the Hilbert envelope spectrum. Eight fault features, including peak frequency, number of spectral peaks, and spectral energy, were extracted from the Hilbert envelope spectrum. These fault features were then fused with control parameter deviation data, trajectory deviation data, standardized operating data, standardized load label data, and standardized drive data through feature concatenation, resulting in fused feature data with 800 samples and a feature dimension of 20. The fused feature data was divided at a preset ratio of 7:3 to obtain a training dataset of 560 samples and a test dataset of 240 samples.

[0029] Step S3: Fractional Neural Network Construction. Based on the training dataset, fractional neural networks are constructed on three local nodes (Node 1: robot body controller, Node 2: workshop edge computing node, Node 3: factory-level data server), with an added precision adjustment module. The network structure includes two fractional convolutional layers, two fractional pooling layers, and two fully connected layers. The input layer receives fused feature data, and the output layer is the probability distribution of four types of faults. The fractional convolutional layers use fractional convolution operators with orders ranging from 0.5 to 1.1. The first fractional convolution operator has an order of 0.9, a kernel size of 4×1, and 32 kernels; the second fractional convolution operator has an order of 0.8, a kernel size of 6×1, and 64 kernels. The fractional pooling layers employ a fractional max pooling strategy. The first pooling window size is 2×1, and the fractional order is set to 0.85; the second pooling window size is 3×1, and the fractional order is set to 0.75. The activation function used is the fractional ReLU function, with the following formula: ,in, For fractional-order adjustable parameters, joint 1 (core drive joint). Set to 0.95, joint 4 (distal accessory joint). Set to 0.65. The precision adjustment module configures parameters according to joint type and fault diagnosis priority: for joint 1 (high diagnosis priority), the fractional convolution operator order range is set to 0.8-1.1, and the fractional pooling layer pooling window size is set to 2×1; for joint 4 (medium diagnosis priority), the fractional convolution operator order range is set to 0.5-0.8, and the fractional pooling layer pooling window size is set to 3×1.

[0030] Step S4: Local training, including response time constraints and accuracy tuning. An improved domain adversarial neural network is introduced, with a fractional-order gradient inversion layer placed between the feature extractor and the domain discriminator of the fractional-order neural network. The fractional-order parameters of this layer... Set to 0.4, then pass The gradient reversal strength is dynamically adjusted to adapt to the differences in data domains at each node. A joint loss function is constructed, with the following formula: ,in The cross-entropy fault classification loss is used (the loss value is 0.92 in the initial iteration). To control the fitting loss of parameter deviation (calculated using mean square error, with an initial value of 0.65). The trajectory deviation fitting loss is calculated using the mean absolute error, with an initial value of 0.58. The loss is adapted to the operating parameters (calculated using Huber loss, with an initial value of 0.42). The load level adaptation loss is calculated using cross-entropy loss, with an initial value of 0.38. To drive the difference adaptation loss (calculated using mean squared error, initial value 0.32); weighting coefficients. , , , , The joint loss function is minimized using the Adam optimizer (with a learning rate of 0.001). A response time constraint is introduced: the threshold for a single inference response time of a single node model is set to 70ms. During training, if the inference response time of node 3 reaches 82ms, the order of the fractional-order convolution operator is reduced by a preset stride of 0.08 (e.g., from 0.8 to 0.72), and the number of convolution kernels is reduced (e.g., from 64 to 48). The inference response time is then recalculated until it drops to 68ms, meeting the threshold requirement. The accuracy adjustment module monitors the model diagnostic accuracy in real time: if the model diagnostic accuracy of joint 1 is 88%, which is lower than the preset accuracy threshold of 92%, the order range of the fractional convolution operator is expanded to 0.9-1.1, and the pooling window size of the fractional pooling layer is reduced to 1×1 to enhance the feature extraction capability; if the model diagnostic accuracy of joint 4 is 96%, which is higher than the preset accuracy threshold of 90%, and the model parameter redundancy (calculated as 15% through parameter sparsity) exceeds the limit of 10%, the order range of the fractional convolution operator is reduced to 0.5-0.7, and the pooling window size of the fractional pooling layer is expanded to 4×1 to optimize the model complexity.

[0031] Step S5: Parameter Fusion. Using a random number generation algorithm, node 2 is selected as the decision-maker and temporary central server from the three nodes. The decision-maker calculates the weight coefficients of each node using a preset algorithm, with the following formula: ,in, The diagnostic accuracy of the node model (e.g., node 1: 89%, node 2: 93%, node 3: 91%). The fitting accuracy for the control parameter deviations of the node model (e.g., node 1: 86%, node 2: 92%, node 3: 89%). The response time compliance rate of the node model (e.g., node 1: 100%, node 2: 100%, node 3: 100%). For the accuracy adaptation of the node model (e.g., node 1: 88%, node 2: 94%, node 3: 90%); weight allocation coefficients. , , , Suppose that the calculated weight coefficient of node 1... Node 2 weight coefficient Node 3 weight coefficient Nodes 1 and 3, as participants, transmit the locally trained fractional neural network parameters to decision-maker node 2 via TCP / IP protocol. The decision-maker does not need to process node parameters that do not meet the response time threshold; for parameters that meet the threshold, a weighted fusion formula is used for calculation. The formula is as follows: Suppose that node 1 has the weights of the first layer of the fully connected layer. Node 2 Node 3 Then the weight after fusion The parameters of the fusion model are obtained.

[0032] Step S6: Model Update. Decision maker node 2 distributes the fused model parameters to participant nodes 1 and 3 via an encrypted transmission protocol. After receiving the parameters, the participants update the corresponding layer parameters of their local fractional neural network according to the parameter replacement rules. For example, node 1 updates the weight of the first layer of the fully connected layer from 0.45 to 0.49, completing the local model update.

[0033] Step S7: Training Stop Judgment. A preset number of four consecutive fusion rounds is defined. The training stop criteria include: diagnostic accuracy fluctuation ≤ 0.5% on the test dataset, control parameter deviation fitting error ≤ 2%, average model response time ≤ 70ms, and diagnostic accuracy for each joint meeting its corresponding priority requirements. If, after four consecutive fusion rounds, the diagnostic accuracy is 91.2% in round 1, 91.5% in round 2, 91.4% in round 3, and 91.6% in round 4, with fluctuation ≤ 0.5%; control parameter deviation fitting error ≤ 1.7%; average model response time ≤ 65ms ≤ 70ms; and diagnostic accuracy for joint 1 ≥ 93.5% and joint 4 ≥ 94.2% respectively, then the training stop criteria are met, and step S8 is executed; otherwise, return to step S4 to continue training.

[0034] Step S8: Model Validation and Output. The updated fractional neural network is validated using the test dataset. The test data is input into the model to obtain the fault probability distribution for each sample. The fault type is determined based on the maximum probability value, and the severity of the fault is determined by combining the probability values ​​(probability ≥ 95% for minor fault, 85%-95% for moderate fault, and < 85% for severe fault). The diagnostic confidence level, i.e., the maximum probability value, is output. For example, after inputting the test sample of joint 2 into the model, the outputs are: normal probability 0.02, flexspline pitting probability 0.96, wave generator wear probability 0.01, and drive shaft loosening probability 0.01. The diagnostic result is: the fault type of joint 2 is flexspline pitting, the fault severity is minor, and the diagnostic confidence level is 96%. Example 2

[0035] The only difference from Embodiment 1 is that a multi-joint robot fault diagnosis system based on fractional neural networks is provided to execute the fault diagnosis method described in Embodiment 1. The system architecture includes a data acquisition unit, a data processing unit, a model training unit, a parameter fusion unit, and a diagnostic output unit. Each unit interacts with the other via Ethernet. The data acquisition unit includes a piezoelectric accelerometer (used to collect joint vibration data, installed in the housing of each joint harmonic reducer, and connected to a data acquisition card via a shielded cable), a servo motor controller (integrating current and speed acquisition modules, and outputting servo motor operating parameters in real time), a motion control card (with a built-in trajectory data acquisition module, and the sampling interval is configurable), a torque sensor (installed at the joint drive shaft to collect load torque data), a power analyzer (connected to the drive motor power supply line via a current clamp to collect drive current and power data), and a temperature sensor (installed inside the robot body cabinet to collect ambient temperature data). The sampling frequency, range, and other parameters of each sensor and device can be configured through the host computer software to ensure the synchronization and integrity of data acquisition. The data processing unit utilizes an industrial-grade edge computing server with a built-in data preprocessing algorithm module. This module performs operations such as denoising, standardization, interpolation, and feature extraction on the collected multi-source raw data. The model training unit consists of three distributed computing nodes, each equipped with a GPU acceleration module, supporting parallel training of fractional neural networks and capable of simultaneously handling model training tasks for multiple joints. The parameter fusion unit, integrated into node 2, includes a parameter receiving module, a weight calculation module, and a parameter fusion module. The parameter receiving module supports multiple communication protocols such as TCP / IP and MQTT, receiving model parameters transmitted from each training node. The weight calculation module incorporates a node weight coefficient calculation algorithm, dynamically allocating weights based on the diagnostic performance of each node. The parameter fusion module performs weighted fusion calculations and outputs the fused model parameters. The diagnostic output unit includes an industrial touchscreen, an audible and visual alarm device, and a data storage module. The industrial touchscreen displays the fault type, severity, and diagnostic confidence level of each joint in real time. The audible and visual alarm device triggers when the fault confidence level is ≥90%, alerting personnel to handle the issue promptly. This system is compatible with different models of multi-joint robots. By simply replacing the corresponding data acquisition equipment (such as adjusting the sensor installation position and adapting the communication protocol of different servo controllers) and reconfiguring the network parameters through model building software, the system can achieve fault diagnosis, demonstrating strong compatibility and practicality.

[0036] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A fault diagnosis method for multi-joint robots based on fractional-order neural networks, characterized in that, Including the following steps: Step S1: Collect the original vibration data and servo motor control parameters of each joint of the multi-joint robot; denoise the original vibration data to obtain denoised vibration data; standardize the denoised vibration data to convert it into standardized vibration data; calculate the absolute difference between the actual parameters and the rated parameters of the servo motor control parameters to obtain control parameter deviation data; integrate the standardized vibration data and the control parameter deviation data to obtain a standardized multi-source dataset. Step S2: The standardized vibration data in the standardized multi-source dataset is processed by Hilbert transform to obtain an analytical signal; the analytical signal is processed by modulus calculation to obtain an envelope signal; the envelope signal is processed by Fourier transform to obtain a Hilbert envelope spectrum; fault features are extracted from the Hilbert envelope spectrum; and the fault features and control parameter deviation data are fused to obtain fused feature data. The fused feature data is divided into training and testing datasets according to a preset ratio. Step S3: Construct fractional neural networks on multiple local nodes based on the training dataset as local training models. The fractional neural network includes fractional convolutional layers, fractional pooling layers and fully connected layers. The fractional convolutional layers use fractional convolution operators with a preset order range. The fractional pooling layers use fractional max pooling strategies. The activation function is the fractional ReLU function. Step S4: Local training is performed on the fractional neural network of each node. An improved domain adversarial neural network is introduced. A fractional gradient inversion layer is set between the feature extractor and the domain discriminator. A joint loss function containing fault classification loss and control parameter deviation fitting loss is constructed. The model parameters are optimized through the joint loss function to complete the local training. Step S5: Randomly select a node from multiple nodes as the decision-maker and temporary central server. The decision-maker calculates the weight coefficients of each node using a preset algorithm. The weight coefficients are related to the diagnostic accuracy and control parameter deviation fitting accuracy of the node model. Other nodes, as participants, transmit the locally trained fractional neural network parameters to the decision-maker. The decision-maker combines the weight coefficients of each node to perform weighted fusion of the received parameters to obtain the fused model parameters. In step S6, the decision-maker distributes the fusion model parameters to each participant, and the participants update their local fractional neural networks based on the fusion model parameters. Step S7, determine whether the training stopping criteria have been met: if after a preset number of consecutive rounds of fusion, the diagnostic accuracy of the model on the test dataset does not exceed a preset threshold and the control parameter deviation fitting error does not exceed a preset error value, then proceed to step S8; otherwise, return to step S4. Step S8: Validate the updated fractional neural network using the test dataset, and output the fault type, fault severity, and diagnostic confidence of each joint.

2. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 1, characterized in that, In step S1, the calculation formula for standardization is: ， Where x represents the standardized vibration data. For noise-reducing vibration data, The mean of the denoised vibration data. The standard deviation of the denoised vibration data; In step S3, the formula for calculating the fractional ReLU function is: ， in, x is the output value of the activation function, and x is the input value of the function. It is a fractional-order adjustable parameter; In step S4, the formula for calculating the joint loss function is: ， in, For the joint loss function value, For fault classification loss values, To control the parameter deviation fitting loss value, These are the weighting coefficients; In step S5, the formula for calculating the node weight coefficient is as follows: in, Let be the weight coefficient of the i-th node. Let be the diagnostic accuracy of the model for the i-th node. Let be the fitting accuracy of the control parameter deviation of the model at the i-th node. , Assign weighting coefficients; In step S6, the calculation formula for parameter weighted fusion is as follows: ， in, These are the parameters for the fusion model, where n is the total number of nodes. Let be the weight coefficient of the i-th node. The parameters of the fractional neural network for the i-th node after the j-th round of training are given.

3. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 2, characterized in that, In step S1, the collected raw data also includes historical motion trajectory data of each joint of the multi-joint robot; the historical motion trajectory data is interpolated and completed to obtain complete trajectory data; the complete trajectory data is calculated by deviation calculation to solve the Euclidean distance between the actual trajectory and the planned trajectory to obtain trajectory deviation data; the standardized vibration data, control parameter deviation data and trajectory deviation data are integrated to obtain a standardized multi-source dataset; in step S2, the fault features, control parameter deviation data and trajectory deviation data are fused to obtain fused feature data; in step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss and trajectory deviation fitting loss.

4. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 3, characterized in that, In step S1, the collected raw data also includes real-time operating data of the multi-joint robot, which includes current load parameters, operating speed parameters, and ambient temperature parameters. Outlier removal is performed on the real-time operating data to obtain clean operating data. Normalization is then applied to the clean operating data to obtain standardized operating data. Standardized vibration data, control parameter deviation data, and standardized operating data are integrated to obtain a standardized multi-source dataset. In step S2, fault features, control parameter deviation data, and standardized operating data are fused using feature fusion to obtain fused feature data. In step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss, and operating parameter adaptation loss.

5. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 4, characterized in that, In step S4, during local training, a response time constraint is introduced, setting a threshold for the single-inference response time of a single-node model. A dynamic order adjustment strategy is adopted for the fractional-order convolutional layers of the fractional-order neural network. When the inference response time exceeds the threshold, the order of the fractional-order convolutional operator is reduced by a preset step size until the response time meets the threshold requirement. In step S5, during parameter fusion, the decision-maker prioritizes receiving node parameters that meet the response time standard, and performs lightweight compression processing on node parameters that do not meet the standard before fusion. In step S7, the training stopping criterion, a condition is added that the average response time of the model meets the standard, and the single-inference response time of the model after all nodes are updated does not exceed the preset response time threshold.

6. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 5, characterized in that, In step S3, when constructing the fractional-order neural network, a precision adjustment is added. Based on the joint type and fault diagnosis priority of the multi-joint robot, the order range of the fractional-order convolution operator and the pooling window size of the fractional-order pooling layer are dynamically configured. In step S4, during local training, the precision adjustment monitors the model's diagnostic precision in real time. When the precision is lower than the preset precision threshold, the order range of the fractional-order convolution operator is expanded and the pooling window is reduced. When the precision is higher than the preset precision threshold and the redundancy exceeds the limit, the order range is reduced and the pooling window is expanded. In step S7, the training stop criteria, it is simultaneously verified whether the diagnostic precision of each joint meets the precision requirements of the corresponding priority.

7. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 6, characterized in that, In step S1, the collected raw data also includes load difference data of each joint of the multi-joint robot; the load difference data is processed by hierarchical calibration to obtain load level label data; the load level label data is processed by one-hot encoding to obtain standardized load label data; the standardized vibration data, control parameter deviation data and standardized load label data are integrated with the standardized load label data to obtain a standardized multi-source dataset; in step S2, the fault features, control parameter deviation data and standardized load label data are fused with the feature fusion method to obtain fused feature data; in step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss and load level adaptation loss.

8. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 7, characterized in that, In step S1, the collected raw data also includes drive difference data of each joint of the multi-joint robot; the drive difference data is processed by waveform feature extraction to obtain drive feature data; the drive feature data is processed by standardization to obtain standardized drive data; the standardized vibration data, control parameter deviation data and standardized drive data are integrated with the standardized drive data to obtain a standardized multi-source dataset; in step S2, the fault features, control parameter deviation data and standardized drive data are integrated with the standardized drive data to obtain fused feature data; in step S4, the joint loss function includes fault classification loss, control parameter deviation fitting loss and drive difference adaptation loss.

9. The fault diagnosis method for multi-joint robots based on fractional-order neural networks according to claim 8, characterized in that, The fractional pooling layer employs a fractional max pooling strategy, with the first pooling layer having a fractional order of 1. The order of the second pooling fraction is set to 0.

75.

10. A fault diagnosis system for multi-joint robots based on fractional neural networks, characterized in that, The method for performing any one of claims 1-9.

Images